OF BOND
GRAPH
MODELLING
IN CONTROL
A DISSERTATION SUBMITTED TO THE DEPARTMENT UNIVERSITY OF THE OF MECHANICAL OF GLASOW REQUIREMENTS OF ENGINEERING
OF THE
IN FULFILLMENT
DEGREE
OF PHILOSOPHY
Copyright
Abstract
This thesis presents new aspects of bond graph modelling in control, where established control theory is used for closed loop bond graph representations. In particular, the physical model
based framework of bond graph modelling addresses Backstepping Control and Energy Shaping in Stabilisation Control.
Even though
methodologies are quite different on analytical allow for closed loop bond graph models.
Concepts of passivity and the portHamiltonian the thesis. Various detailed examples
leading bond of graphs play a role throughout structure impart the essential results.
Summary
framework
for the
have witnessed tremendous advancement in open loop modelling ever since the
inception of this graphical modelling technique by Professor Henry Paynter in 1968. On the other hand, bond graphs in control, or closed loop bond graph models, have not received the same level of research commitment This thesis contributes loop representations, novel and virtually compared to aspects of open loop systems modelling.
focusing by bond design in on closed graphs control new aspects of where the idea of applying bond graphs for closed loop modelling is The thesis does not control concepts to The physical model
nonexistent
present new control theoretical results in any way but applies wellknown find closed loop bond graph representations for stabilisation problems.
based character of general bond graph models is shown to be suitable for the control strategies of Backstepping Control. Backstepping control within the bond graph framework is shown to be a case of exact backstepping by which the closed loop dynamics is put into portHamiltonian suitable choice of variables. Consequently, a bond graph representation The physical modelling form through a of the closed loop Control, Model Matching Control and Energy Shaping in Stabilisation
additive bond graph is referred to as the virtual actuator. It is known that backstepping control is a recursive design technique to obtain a closed loop Lyapunov function; however, the geometric structure of the closed loop itself is generally not an immediate design goal. This thesis, on the other hand, aims at closed loop bond graph representations loop portHamiltonian wellknown Lyapunov dynamics as an explicit arguments are implicitly backstepping by having closed
11
Model matching
tracking
of prescribed
trajectories
of
some desired dynamic model. This thesis shows that such prescribed models can be chosen as bond graph models that are structurally requirements of the Model Matching "close" to the plant to satisfy certain solvability Tracking control through (bi)causal but the
Problem (MMP).
feedback designs has not appeared in the current literature. mechanism of such is linked with various ideas of center
This thesis argues that the bond graph based MIM manifold theory and output regulation problems.
this is that the closed loop error dynamics can be described by the plant bond graph such that additive bond graph elements can be used for closed loop stabilisation. Energy shaping in stabilisation control, as considered in this thesis, addresses feedback designs structure and resistive elements of
the plant. It is shown that bond graphs can be used to find the closed loop energy function that attains feedback passivation with respect to the natural output. closed loop energy function Most importantly, the
known follows be beforehand but from a "power balneed not instead of modifying the energy
by means of the power balancing method alone, the Interconnection (IDAPBC) Control Passivity Based ment interconnection by modifying
and damping assignment is shown to allow for bond graph representations the junction structure and the dissipative elements of the plant bond graph. and damping structures are therefore guided by bond designs generally require the solution of
The desired closed loop interconnection graph topological first order partial analytically. considerations. differential
Since IDAPBC
iii
Preface
When I learned about bond graphs, in the year 1998, someone once said to me that bond graph modelling looked like a "black art": A collection of arcane, unpublished, techniques developed for a particular application and mostly adhoc
by but I believe description in do 2005, this the this surprised now, not remark; year quite of bond graph modelling is completely unjustified... Bond graphs look intriguing is radically different for the first time, because the graphical seen when topology
clever, for it reinforces ones confidence in the modelling process, where aspects of constraint dynamics and algebraic loops have virtually no obscurities. Furthermore, bond graphs are
based on energy concepts to specifically accommodate the systematic modelling of multidisciplinary physical systems. However, even though bond graph modelling is wellknown by
small group of professionals only. Also, instead of being a collection of adhoc techniques, the bond graph language is highly structured and rich in literature.
Now that my three years of graduate research have come to an end, I can say that bond graphs do embody certain elements of a "black art" after all: The graphical topology of bond graph models appears mystical at first, but a closer look reveals a spellbinding cleverness. By writing structure and
this thesis, I have tried to uncover some new secrets of bond graph
modelling in control design, hoping that what captivated my thoughts has been put in clear writing for everyone to read.
Dustin Vink
'Free OnLine
Dictionary
of Computing
iv
Acknowledgments
Research takes time and funding, so I hereby thank Dr. D. Ballance and Professor P. Gawvthrop for the opportunity Unfortunately, and financial support to enjoy three years of Ph. D. research.
the three years of graduate research were sufficient to produce the required this time frame. Needless to say, I wish to this thesis.
thank the Dutch government for its financial support while writing
Finally, my family and friends have been extremely supportive during my days as a graduate student. Thank you all.
I
Contents
i ii iv
I.
Preliminaries
1
2 2
1. The Art of Bond Graph Modelling ..................................... 1.2. Bond Graphs and Block Diagrams 1.2.1. 1.2.2. NonCausal Bond Graphs ... 1.1. Introduction
...
.3 .4 .8 . 11 12 13
..... . ... 1.3. Bond Graphs as PortHamiltonian Systems 1.3.1. 1.3.2. Basic Facts on PortHamiltonian Network Interconnections;
.... Systems
Dissipation ......
......
15 19
2.1. Introduction
19 20 20
............................... Recursive Lyapunov Design ........................ Closed Loop PortHamiltonian Dynamics Control
.... ....
....... ....
. ... .......
22 25
... ...... ...... Some Facts on Model Matching Problems Remarks on Output Regulation
. .... ... . .
25 33 35
2.4. Stabilisation
vi
...... . ...... . ... . ....... ... 2.6. Interconnection and Damping Assignment ..... . ...... . ... 2.7. Concluding Remarks . . ...... . ... . ...... ... ......
35 36 38
II.
39
40 40 40 41 51 65 68 72 74 74 . ...... ..... ...... . ... . ... . ...... . ... . ... .... ..... . ...... ..... . ... ... . .. ... ... 76 87 91 92 93 104
3. Backstepping
... . ..... ... ..... ..... 3.2. Backstepping Control in the Physical Domain 3.2.1. 3.2.2. 3.2.3. SingleInput Systems: Examples Systems ... . .....
3.1. Introduction
. ...... ... . ... ... 3.3. Bicausal Bond Graphs in Backstepping Control ...... 3.4. Conclusions ... . ......... ..... ... ..... 4. Model Matching Control
4.3. Specification
......... ..... 4.4. General Cases of Model Matching .... .... 4.4.1. A Class of Implicit Systems ... ......... . ..... . ......... Concluding Remarks ................................ 4.4.2. Examples
4.5.
106
106 108 108 111 122 ........ . .. . ... ... . 122 128 129 135
Introducing
............. Compatibility
....... Designs
5.5. Conclusion
vi'
6. Conclusions and Future Research ........................................ 6.2. Backstepping Control ..... 6.3. Model Matching Control .... 6.4. Energy Shaping in Stabilisation 6.5. Future Research 6.1. Review
137 137 . .... ... . ...... .......... ... ... . . . 137 138 139 140
..................................
vin
List of Figures
system . ...
......... ....
. ... ...
4 5
system . .....
1.3. Power continuous elements ... ......... . .... ...... ..... 1.4. Causal bond graph .... ...... ... . ... ...... ......... 1.5. Basic causal propagation through causal strokes ........ . ........ 1.6. Integral causality for storage elements 1.7. Bicausal propagation ...................... ....
. .
5 8 9 9
of effort and flow .... . ... ......... 1.8. Examples of bicausal propagation on junctions .................. 1.9. Vector bond graph without dissipation ... . .... . ... . .... ...
11 11
... ...
. . . .
13 14 41 42
Example 3.1 Rl of element resistive ..... . .... 3.2. Dynamics (3.7) of Example 3.1 ........ ..... ....
. ...
. ...
......
...
... . ...
.
. . . ..
43
44 45 45 46
system of Example 3.2 . .... ...... bond graph of Example 3.2 Massspringdamper ......... ... Target closed loop system of Example 3.2 ......... . .... . ..... Bond graph virtual actuator of Example 3.2 ........ ..... .... Closed loop massspringdamper of Example 3.2 ..... .......
. ...
47 48 48 49 49 52 52 54 54
of Example 3.3 ... . . ............ . ...... 3.10. Bond graph system of Example 3.3 . ..... .... .... ...... .... 3.11. Target closed loop of Example 3.3 ....... . ... ...... .... .... 3.12. Target closed loop bond graph of Example 3.3 . ...... .. ......... 3.13. Cascaded C element of Proposition 3.1 . ... ... . ...... ... . .... 3.14. Closed loop Ccascaded system of Proposition 3.1 . ... . ........... 3.15. Cascaded I element of Corollary 3.2 ........................ 3.16. Closed loop Icascaded system of Corollary 3.2 ... .. . ... ..... .... 3.9. Massspringdamper
ix
3.17. Closed loop Ccascaded system of Corollary 3.3 ..... .. .... ...... 3.18. Closed loop Icascaded system of Corollary 3.3 ... ....... ... ..... 3.19. Repeated linear cascaded elements of Proposition 3.5 . .... .. . ...... 3.20. Target cascaded closed loop of Proposition 3.5 ... . ...... . ....... 3.21. Extended backstepping junction of Corollary 3.7 .... ..... .. ...... 3.22. Compound element backstepping of Example 3.6 . ........ .. ...... 3.23. Compound backstepping bond graph of Example 3.6 ..... .. .... .... 3.24. Twoinput system of Exercise 3.7 . .... .. . ...... ....... . ... 3.25. Target closed loop of Exercise 3.7 ... . ... ... . ....... ... . ... 3.26. Extended cascaded bond graph of Proposition 3.9 ..... .... .. ..... 3.27. Bicausal extended cascaded C and I pattern of Proposition 3.9.. . ...... 3.28. Closed loop dynamics with bicausal approach of Proposition 3.9. . .... .. 3.29. Generic Ci backstepping junction ..... 3.30. Backstepping with bicausal assignment; Example 3.8 .... 3.31. Bicausal backstepping towards u2; Example 3.9 ..... ... 3.32. Bicausal backstepping towards ul; Example 3.9 ........ ... configuration .......................... of Proposition 3.9 . ....... ......... . ....... .... ... .. .. ..
55 55 58 58 61 62 62 66 66
69 69 70 70 71 72 72
4.1. Plant input/output 4.2. Model input/output 4.3. 4.4. 4.5. 4.6. 4.7. 4.8.
76 77
.. .. ... . .. .. ... .. .. ...
configuration ......................... Causal inversion of P. ..... . ... .... . ..... . ....... Simple massspring plant of Example 4.2 ..... ........ ...... Simple massspring model of Example 4.2 . ...... . ... . ... ... Plant bond graph of Example 4.2 . ....................... Model bond graph of Example 4.2 . ....... ......... . ..... Bicausal inversion of the plant of Example 4.2 . .... . ... . ......
78 80 80 81 81 82 83 83 84 84
85
mechanical plant of Example 4.3 ..... . ... . ...... Model with nonlinear spring element of Example 4.3 ... ..... ..... Plant bond graph of Example 4.3 ..... ..... . ...... ...... Model bond graph of Example 4.3 .... . ... . ...... ....... Bicausal plant inversion of Example 4.3 ...................... Induced closed loop bond graph of Example 4.3 ..... ..... .... pairs . ... . ... .... pairs.. . ... . ....
... .. ..
86 87 87 88
4.15. Plant bond graph with collocated input/output 4.16. Inverse plant bond graph ........ 4.17. Plant bond graph with noncollocated
..........
input/output
pairs.
......
...
88 89 89 94 95 96 99 99 100 102 110 110 112 114 116 117 120 125 125 127 127 128
of Example 4.4 .... . ..... ..... ... ....... 4.20. Bicausal RCcircuit of Example 4.4 . ... . ..... ..... . ..... ... 4.21. Inverted pendulum of Example 4.5 ..... ... ... .... . ......... 4.22. Pendulum bond graph with Amultipliers of Example 4.5. ........... 4.23. Simple pendulum model bond graph with multipliers of Example 4.5. ... 4.24. Frictionless slider of Example 4.6 .......................... 4.25. Frictionless slider with Amultiplier of Example 4.6 ...... ...... ... 4.26. Frictionless slider with LCAP of Example 4.6 ..... . ..... . .... ... 4.27. Model slider with LCAP of Example 4.6 .... .... .......... ... . 5.1. Power flow oy for passive feedback with SCAP of Example 5.1. ........ 5.2. Power flow uy for passive feedback with LCAP of Example 5.1. ... ..... 5.3. Conceptual bond graph based power balancing with SCAP. . ...... ... 5.4. Basic MIMO bond graph based power balancing of Example 5.2. . ...... 5.5. TORA physical configuration of Example 5.3 ... ...... ..... . .... 5.6. TORA Lagrangian assigned bond graph of Example 5.3 ... .......... 5.7. Power balancing not applicable to slider of Example 5.4 ....... ... ... 5.8. Dissipative system of Example 5.5 ......................... 5.9. Energy shaping compatible with bond graph topology; Example 5.5. ..... 5.10. Energy shaping impeded by modulation of Example 5.6 ... .......... 5.11. Underactuated bond graph of Example 5.6 ........... ... . ..... 5.12. Conceptual representation of energy shaping ...................
.....
.........
. ...
130
130 131 131 133
ball bond graph of Example 5.7 . .............. 5.15. Energy shaping compatible with bond graph topology; Example 5.7. ..... 5.16. Gyration and damping assignment; Example 5.7 . ... .... ..... . ... 5.17. Nonobvious additive damping; Example 5.7 .......... ..... . ...
xi
Preliminaries
1.1. Introduction
of
Technology (MIT) in Cambridge, Massachusetts, USA, and this lead to the first book [Pay6l] published on bond graphs. In addition D. L. Margolis to Henry Paynter's own work, his former Ph. D. accelerated [Kar00].
students D. C. Karnopp,
contributed
But others, too, picked up bond graph modelling and published a wide variety of textbooks [B1u82], [Bor00], [Bor04], [Bre92b], [Ce191], [Dix74], [Gaw96], [Tho99] that describe both bond graph theory and various applications. In parallel to the ongoing bond graph research, it be
came clear that the systematic modelling approach offered by bond graphs rendered software implementation possible, where bond graph simulation simulation [Ros74] started to emerge packages bond graph topolog
Today, a variety of software solutions are available that offer graphical design with advanced symbolical systems [Dyn04], [MTT04], and numerical simulation [BV04]. engines for complex mul
environments tidisciplinary
It is safe to say that the art of bond graph modelling has been subjected to extensive research last four decades. However, it is relatively the topics on a wide variety of over difficult to
compile a compact list of bond graph references that provide an adequate overview of bond graph theory and its applications. This can be partially attributed to the fact that bond
the journals,
for contemporary
issues [Bre9l],
[Gaw02] on bond graphs have appeared that present various states of affairs. and references
Regardless of the topic, the reader is simply referred to the above literature therein on past and current research pertaining to the bond graph language.
recalled
in the standard
literature with
[Gaw96], respect
some small
modifications
and conventions.
reproductions
their connection with bond graph models [Go102], [Go103]. These developments cannot be called standard portHamiltonian by any means and the identification of bond graph models as a class of textbooks at this time of writing. for various
Finally, having presented the various modelling aspects of bond graphs, the thesis rationale and objective can be outlined constructively. It can be argued, for example, that certain
aspects of physical model based control can be assisted by means of the closed loop bond graph representation, further where unification of modelling and control methods in the physical of
domain with a systematic modelling framework may lead to an improved understanding physical model based control problems.
modelling framework can be argued to be the standard graphical modelling tool for systems and control in both academia and industry. For example, consider the block diagram in Figure 1.1 with states xi, constants a, C, ml, m2 and r to be multiplied, Even though block diagrams are straightforward, before the block diagram can be drawn. causal information, X2 and x3, some
f integrates the ingoing signal. and where it is required that causal relations are known
showing that the modeller must actively derive all causal relations for
/1
Y2
system.
representation
identify system signals as completely separate entities but uses generalised energy and power considerations instead. More precisely, the bond graph identifies a natural and "flow" respectively, pairing of two
signals denoted as e and f, called the "effort" yields generalised power. Therefore,
notions of energy and power are meaningful, Furthermore, and this is important,
such as multidisciplinary
causality, which is absent in the block diagram framework. make a clear distinction ment statements. between a :=b and b :=a,
which are referred to as causal assigndepicted and subjected to actively putting such
To elaborate the most basic aspects of bond graphs in more detail, consider the noncausal bond graph in Figure 1.2 that represents the same electricalmechanical system of Figure 1.1.
The efforts and flows of a bond graph, e1 and flows fj, are always associated with a "bond" that is drawn as a harpoon shaped arrow, hence the name bond graph. When ejf3 > 0, the power flow is in the direction of the bond arrow.
12
X2
11
mlxl mlxl
m2
ml
a x2
XI
GY M1
M2
SS2
2 Vx3
Lth3
C
system.
R
Figure 1.2.: Noncausal
fi
e3
g(X)
TFZ
12
e4
fs
t(x)
fa
Figure 1.3.: Power continuous elements. In this thesis, efforts are drawn to the left or above the vertical and horizontal tively, whereas flows are drawn to the right or below the bonds. The power variables ej and fj of each bond can be collected into the pair (ej, fj) readily derived from Figure 1.2 as (ui, yl), (th1,x1/ml), and are bonds respec
flows contained in such pairs are referred to as conjugate power variables. Now, suppose that all bonds connected to a1 or 0 element point outward or inward, then the power balance and 1junctions Zejfj i The relation (1.1) expresses power continuity of 0junctions and 1junctions, but correct That is to say is defined as
=0.
(1.1)
signs must be accounted for when subsets of bonds have alternate directions.
that either the inward or outward bond direction must be designated as being positive when evaluating the power balance (1.1). In addition to 0junctions and 1junctions, TF, the power continuous gyrator, GY, and the
encountered in
bond graph models. The gyrator maps efforts into flows and flows into efforts, whereas the transformer is defined as mapping efforts into efforts and flows into flows. These elements relationships in multidisciplinary engineering systems.
gyrator, MGY,
and modulated
Nonetheless, the standard noncausal definitions for gyrators and transformers in Figure 1.3 are adopted and given by e2  9(x)fi
e4 t(x)e3
= 0,
= 0,
el  g(x)f2 f3t(x)f4=0,
(1.2)
g(x) and t(x) may depend on state space coordinates xEXC gyrators
RI.
is not strictly
does not induce a loss of generality in any way. Power continuity is indeed guaranteed, regardless of the modulation,
e1f1
e3f3
= 9(x)f2f1
= e3t(x)f4
(1.3)
= e4 f4"
Now, standard
but the power balance is likewise satisfied for all outer bond pairs of which represents an arbitrary network interconnection of bonds, junc
tions, gyrators and transformers. The O junctions and 1junctions do not only induce the power balance (1.1) with respect have some additional rules. For
to all those bonds connected to them, but these junctions instance, by taking (1.1) into account, a single I junction 1 whereas a single 0junction 0= ==:> induces ej =ei fj = fz
O, =
(1.4)
Eft=0.
(1.5)
(1.6)
It
relationships
is readily observed that the minus sign in the second relation alternate bond direction.
[Bus98]. in models roles many physical crucial Now that basic notions of bonds, 0junctions, 1junctions, TF and GY components have
been briefly recalled, the definitions of the SS, C, I and R components that are to terminate the outer bonds of bond graphs are given as follows. First, the source sensor, SS, component [Gaw96] described in represents an element that is associated with power supply; howas ever, the reader should note that the conventional effort source, Se, and flow source, Sf, often found in the literature will not be used in this thesis. Instead, the Se and Sf source elements losing generality of bond graph modelling
Second, the C and I elements are storage elements and represent the storage of physical energy, such as kinetic and potential associated with realvalued defines the maps C: X  l in the mechanical domain. These elements are usually X, so that one
and I: X > R. The bond graph framework uses these energy with such C or I
functions to define the states of the system, where the bonds terminated elements have constitutive constitutive defined as C= e(t)  e(0) Jf relationships relationships
for efforts and flows associated with storage elements are typically t (s) ds =0
t If (t) f (0) J 0
The bond graph in Figure 1.2, for example, shows that one could write ej(t) = ij(t)/mj jE {1,2}, so that the constitutive
ff (t)
(1.7) e(s) ds = 0.
with
fj
relations as in (1.7), one often defines an energy function relations are derived. For example, in = xj/mj. defined
for the storage elements from which the constitutive Figure 1.2 one would define the functions Ij(x) Therefore, the constitutive relationships
12
SS1 \
Yi
II mixt
11
1/ X1 ml ,
GY
al
Ml X2 m2 , Z'X3 xs Y2
SS2
mlxl
RC
Figure 1.4.: Causal bond graph Finally, energy dissipation phenomena are implemented through R elements that represent have constitutive relation
various resistive effects, where the associated signal pairs (ej, fj) ships that are to satisfy ej(t) fj(t) >0 for all t>0,
since the power flow is positive and outgoing. associated with asymptotic stabilisation stabilising
control purposes.
1.2.2.
The block diagram in Figure 1.1 shows the causality of variables by means of ingoing and outgoing arrows, where the causality had already been established before the block diagram was drawn. However, modelling may require alternative causal patterns with respect to
system inputs, thereby rendering the block diagram of limited interest for causal analysis. Now, the bond graph in Figure 1.2 offers 'a noncausal which the causal dynamics can be derived. representation of the system from rela
tionships into assignment statements, the bond graph uses causal strokes and their junction causality to propagate computational graph. causality of all power variables throughout the bond
To that end, consider the system in Figure 1.4 of which each bond has been augstroke, which induce assignment statements that lead to rules of these strokes in terms of
the equations of motion; see Figure 1.5 for the graphical strong causality on 0junctions indicate the computational and 1junctions.
direction of the efforts and flows: The flow is always directed away
from the stroke whereas the effort is always directed towards the stroke, thereby offering a systematic, graphical mechanism for causal computation.
0 f
f/
CI
Figure 1.6.: Integral causality for storage elements To arrive at the equations of motion, the noncausal bond graph is first assigned the causality of source elements for which the causality is propagated using the rules depicted in Figure 1.5. Subsequently, the preferred integral causality as depicted in Figure 1.6 is imposed and propagated, where the relations (1.7) evidently become t C e(t) = e(0) +Jf 0 t I=f (t) =f (0) +J e(s) ds. (s) ds (1.9)
0
The opposite of integral causality is referred to as derivative causality and is associated with the reversal of causal strokes in Figure 1.6, leading to f (t) = e(t) for C components and e(t) =j (t) for I components. So by taking the above considerations into account, it is seen
that the bond graph in Figure 1.5 yields the equations of motion
ar xl = x2 xl  Ul m2 ml a1 th2 = ml X1  cx3  U2 1 X3 = X2, m2
1.10)
which are indeed identical to the dynamics derived from the block diagram in Figure 1.1. Causal assignment is systematic causal propagation Also, structural but may require additional attention in some cases, since
properties
1. The Art of Bond Graph Modelling The causality Causality assignment as briefly Procedure described is sometimes referred to as the Sequential [Ros87], [van94]. This procedure implies that
Assignment
(SCAP)
and where a largest set of C and I eleIt must be remembered that the stan
dard bond graph language imposes state variables defined by C and I elements, but these states may not yield efficient models in certain cases. Indeed, alternative ment procedures, such as the Lagrangian shown to be capable of offering additional graph induced dynamics [Kar83], [Mar02]. Causality Assignment Procedure causal assign(LCAP), have
freedom to manipulate
the structure
of bond
[Bir9O], [Dij91], [Gaw95a], [Gaw92], [Hog87], [Jos74], [Lam97] and references therein.
flows associated with bonds has proven to be unnecessarily restrictive in some cases [Gaw95a]. This can be attributed to the fact that propagation of causality with the single causal stroke So by setting the causality On the
mechanism implies opposite conjugate effort and flow directions. of one bond signal fixes the propagation other hand, power continuity direction
the causal stroke mechanism can be generalised to the cases where the conjugate effort and flow have identical computational Causality of efforts on 0junctions directions. and flows on 1junctions need not be compromised when
the single causal stroke is abandoned and where the conjugate efforts and flows attain individual causal strokes instead. Doing so increases the number of causal configurations for
bond graph models and has shown to offer an additional elling purposes [Gaw00], [Gaw03]. In particular,
quite useful for "causal inversion" problems [Ngw96] whereby the input/output inverted through the (bi)causal stroke mechanism, if possible.
Figure 1.7 shows the possible bicausal propagation of conjugate effort and flow pairs, where it is seen that the computational in a bicausal context. direction of both the effort and flows are rendered independent remains and that
unchanged, and Figure 1.8 shows that efforts retain strong causality on 0junctions flows retain strong causality on 1 junctions.
10
1. The Art of Bond Graph Modelling Te f yf Figure 1.7.: Bicausal propagation of effort and flow. e
Figure 1.8.: Examples of bicausal propagation Bicausality for causal inversion problems is particularly
instructive
ure 1.4 it is seen that the single causal stroke mechanism selects the bond signal yj as the output of SS elements. Now, provided no causal conflict occurs, causal inversion is then
achieved by moving some or all causal strokes to the other end of the bond. On the other hand, such input/output inversion is not suitable for nonconjugate input/output pairs. To (UI) Y2) pair
see this, consider Figure 1.4 once again and suppose one defines the input/output
by setting u2 =0 and by ignoring the output yl, but it is readily understood that the single stroke mechanism cannot be used to causally invert the pair (Ui, y2) in such a scenario. The reader is referred to [Ngw99a], [Ngw0la], in physical systems modelling. [NgwOlb] for further accounts on causal inversion
Systems
Even though the bond graph language offers a structured motion, the underlying mathematical nontrivial.
framework
to derive equations of
structure of bond graph induced dynamics are relatively representation of a class of bond graph models is dynam
ics [Mar94] of bond graphs was recognised at a later stage later [Mas92], [Mas95]. Also, the portHamiltonian framework presented in [Dal97], [Sch96] offers a detailed geometric frame
11
1. The Art of Bond Graph Modelling 1.3.1. Basic Facts on PortHamiltonian Systems
systems are introduced in the context of
by have input that conserving systems physical and output ports which they connect energy to the external environment. an important The authors then show that the coordinate representation systems takes the form of
subclass of portHamiltonian
x=
y= where x= (xl,... EX x0) ,
(1.11)
are independent
is
derivatives
of the smooth energy function H: X + R. The funcand represents the physical energy stored by the system. matrix and defines power continuous network matrix R(x) = RT (x) is the dissipation
interconnections, structure
that incorporates
resistive effects. The port space of the system is represented by Clearly, the
the matrix g(x) and where uE IR' are inputs and where yE 1R' are outputs. portHamiltonian (1.11) system satisfies H(x) = KT(x)R(x)K(x) < yTU yTU, 
(1.12)
which shows that the product yTu expresses the power injected into or extracted from the (1.11). system As argued in [Da197], the system (1.11) is called portHamiltonian lowing. It is possible to define a bilinear, 'antisymmetric by considering the fol
functions defined as {F, G} (x) = JZ; (x) (x) (x), x 8x xi j with F, G: X > R. Jacobi's identity This bracket operation (1.13)
th = J(x)K(x).
Therefore, the portHamiltonian by including the dissipation
the input matrix g(x).
(1.14)
system (1.11) can be said to generalise the system (1.14) R(x) and the input/output port interaction by means of
matrix
12
X1 Ki K2 C
X2
Figure 1.9.: Vector bond graph without 1.3.2. Network Interconnections, Dissipation and Ports
dissipation.
framework
as presented Fig
in [Da197], [Sch96], [SchOOb], [Go102] is achieved by briefly considering the following. bond 1.9 the shows notation ure vector [Bre92a], [Bre95] of a system without
dissipative
elements, where it is seen that systems inputs can be any combination Define the respective inputs and outputs u= nates x= (xl, x2), the tangent vector x= H(x). (ul) u2) and y=
of the Hamiltonian
Note that K, (x) and K2 (x) are column vectors of partial derivatives
with respect to xl and x2 respectively. It can be shown that the network interconnections of power continuous bond graph elein [Ros7l] it was
is itself power continuous [KarOO]. Furthermore, structure JS(x) represents a linear map.
These considerations
then imply that the causal assigned bond graph in Figure 1.9 leads to the relation x_ K(x) = J(x) (x) gT g(x) D(x) g(x) of suitable dimensions. Next K(x)
B(x)
(1.15)
Iyu
for some matrices J(x), D(x) observe that power continuity of JS(x)
KT(x)x+uTy=
[KT(X)
uT
] B(x)
K(x)
= 0.
(1.16)
holds for all energy functions H(x) = BT(X), function H(x) or y=0.
lows that (1.16) must satisfy B(x) Note that by (1.16) an arbitrary
The system (1.15) represents an energy conserving portHamiltonian relatively large class of systems it will be the case that D(x) = 0.
13
R2
f2
yi ssl ul JS(x)
th1 I Ki K2
SS2
u2
Y2
IC
x2
Figure 1.10.: Vector bond graph with dissipation. To include dissipation phenomena to arrive at a larger class of bond graph induced portHamiltonian dynamics requires the following considerations. Define the vectors of inputs and outputs as
(fl, (el, Rz Rl f2). Then suppose the the and elements as u, = e2) and y,. = sociated with vector bond graph in Figure 1.10 yields the dynamics x y= Yr which defines an antisymmetric structure. In many practical cases it is possible to model dissipation phenomena by considering the J(x) gT (x) gr (x) g(x) D(x) bT (x) gr(x) b(x) 0 K(x) u ur (1.17)
mapping
simple linear relation ur = Sy,  with S= ST > 0; this implies that yTUr >0 and energy is therefore extracted from the system. From (1.17) it follows that the portHamiltonian dynamics (1.17) can be rewritten as
x
y with A(x)
J(x)  R(x)
(x) gT AT(x) 
K(x)
u
(1.18)
R(x) = RT(x) >0 and U(x) = UT (x) > 0. The portHamiltonian dynamics (1.11) are seen to be contained in the bond graph induced large class of multidisciplinary =0 and U(x) = 0. As new physical model
(1.11) be for form in the thesis, the of seen models allow a relatively will based control framework [Ort02b].
14
1. The Art of Bond Graph Modelling Nov, the general bond graph induced dynamics (1.18) can be written form, being xI([ y J(x) gT(x) ._ [J(x) 9(x)) D(x) K(x) R(x) AT (x) A(x)) U(x) K(x) u (1.19) in a more compact
LJ
R(x)1
u
induces the structure matrix matrix Y(x) = JT(x) and where
structure JS(x)
the power flow of all bonds in Figure 1.10 one finally obtains dH(x)_[KT(x)
uT ] K(x) R(x) u
y Tu <
YT
U,
(1.20)
which clearly shows that the stored energy depends on the supply rate yT u. The following final remark is in order. Section 1.2.2 pointed out that it may not be possible to have all C and I elements in the preferred integral causality without inducing causal conflicts in the junction insert additional structure JS(x). To overcome such causal problems, it is always possible to
so implies that the outputs of such source element are to be zero [Mar02]. In the case where such additional portHamiltonian SS elements are necessary, the bond graph can be shown to induce implicit dynamics [SchOOa]. This will not be further elaborated.
and Objective
It is safe to say that the open loop modelling capabilities of bond graphs are wellunderstood. Indeed, as briefly presented in Section 1.2, the graphical aspects of bond graphs have been subjected to significant research efforts, where the more geometric portHamiltonian tion of Section 1.3 can be argued to have contributed network modelling of physical systems. On the other hand, bond graph modelling in control cannot claim to have reached the level of research commitment and sophistication comparable to the modelling aspects of bond graphs. to a further understanding descripof the
Nonetheless, a wide variety of compelling bond graph considerations in various control designs have appeared over the years [Kar79], [Bar77], [Gaw95b], [Hog85], [Jun01], [Rob95], [Yeh99].
15
be quite successful in addressing a myriad of control problems. analytic approach, there are valuable notions of "Control
that attempt to use certain properties of the physical system to aid the controller design. For example, the property laws [OrtOlj. Therefore, instead of analytical approaches, the objective of this thesis is to use established More precisely, this thesis is primarily to facilitate conof physical stored energy can often be used to derive certain feedback
cerned with closed loop bond graph representations for control purposes. Most importantly,
the portHamiltonian
is certainly
[Gaw95b], [NgwOla], but this thesis "rebundles" not new modelling of closed loop dynamics.
The main control methods considered in this thesis are (1) Backstepping Matching Control and (3) Energy Shaping in Stabilisation Control.
Control,
(2) Model
are described in separate chapters that present new views and developments.
paragraphs summarise the rationale behind these feedback designs in further detail.
Backstepping
Control
The backstepping method uses virtual control variables and recursive purposes and is thoroughly documented in the nonlinear
used to design backstepping controllers by defining additive elements that impose the virtual control law. Furthermore, in [Gaw0l] it is even recognised that certain exact backstepping (bi)causal inversion as outlined of in
In contrast to the existing works on bond graph based backstepping, backstepping can be used to design a closed loop portHamiltonian the plant portHamiltonian model interpretation.
16
jectory of some desired model or exosystem can be said to have a long history in the control literature, [Hui94] for a summary of the topic. is the to where reader referred of input/output Two imporinversion, and
(2) concepts of dynamic disturbance decoupling. As argued in Section 1.2.2, the bond graph language can be used to invert the input/output [Gaw95a], mechanism such that the application certain Model Matching Problems (MMP). dymap by means of the (bi)causal assignment of a causal bond graph tool is available for
This thesis explores the bond graph based MMP by prescribing closed loop input/output namics with a bond graph model. In particular, will be shown to allow for additional
closed loop bond graph representations in certain cases. understanding of what the underlying principles are
of the MMP as considered in this thesis. Such developments have been absent in the current bond graph literature.
Control
concerns itself with feedback laws that induce closed loop passivity energy function and supply rate [Byr9l].
language provides any tools that render the passive feedback design constructive degree. This thesis shows that the junction structure
loop energy functions that induce feedback passivity with respect to the natural output. relevance of this result can be attributed to the fact that the energy function considerations.
need not be
"guessed" but that it follows from power continuity The portHamiltonian Interconnection framework
in [OrtO2a], [OrtO2b]. This control method addresses feedback designs that can be associated with the shaping of the Hamiltonian The solvability of an IDAPBC and the modification of structure and damping matrices.
design, however, is dependent on first order partial differential of basic IDAPBC designs can be matrices is
depicted with the bond graph language, where the modification prescribed by the desired junction
of the structure
17
have the common goal of using the closed loop bond graph for stabilisation differently,
the presented control strategies are certainly different on the analytical level, but
the closed loop dynamics obtained with those methods will allow for bond graph models. It must be noted that aspects of bond graphs in control as presented in the thesis are not (non)linear Instead, define the to rigid procedures. meant ous control methods render generalisations facilitate unforeseen problems. difficult, systems framework and the varishould be retained to
so that flexibility
In conclusion, this thesis shows that the above three control that can aid the control
strategies have proven to allow for valuable physical interpretations design of multidisciplinary systems modelled with bond graphs.
18
2.1. Introduction
is to say that generic feedback strategies generally apply signal theoretic techniques through considerable collections physical interpretations of abstract of controlled mathematical dynamics. methods without concerning itself with
control designs
can be very effective and systematic for a wide variety of (robust) control problems, in order to explore bond graph representations structural for closed loop dynamics it can be argued that more
designs should impose closed loop dynamics that allows for an associated bond graph model. For example, as outlined in Section 1.3, in case the closed loop is to be represented by means of a bond graph, this would imply that closed loop portHamiltonian dynamics should be an
explicit design goal. Therefore, the aim of this chapter is to address feedback design methods that allow for structural and physical considerations in order to attain closed loop dynamics purposes. control [Kri95] is recalled and shown chosen virSec
with an associated bond graph model for stabilisation The chapter is organised as follows. Backstepping
dynamics
by judiciously
are possible.
ond, model matching control as presented in [Hui94] is shown to allow for closed loop bond graph models by borrowing problems. Finally, certain ideas of center manifold through theory in output regulation
stabilisation
control
energy shaping is presented in terms of and damping assignment proceand damping structures
19
2. Fundamentals
2.2. Backstepping
Control
larger, the to subsystems, are connected obtain complex models. objects, smaller decomposition such of into subsystems, suppose that some particular
bilised by placing a virtual actuator at some desired location but for which no regular control is available. the virtual Then, intuitively, actuator one could try to find suitable feedback control that imposes stabilises the subsystems "between" the actual
actuated subsystem.
The above conceptual control problem can be addressed by means of a systematic backstepping design, where a suitable variable is designated as the virtual control that represents the physical location at which the virtual feedback law for the virtual through actuator is to be connected. By imposing a suitable
each recursive design step uses Lyapunov arguments to guarantee (global) convergence of trajectories. Interested readers are referred to the
and asymptotic
works [Kri95], [Isi99], [Kha92], [Sep97] and references therein for a comprehensive treatment on backstepping control designs.
2.2.1.
The backstepping
methodology
th = f(x, )
e=
u (x, (0,0) (x) 1R2 IR f Let be a smooth realvalued function, ) V 0. E x where and = positive definite and proper, and suppose there exists a static feedback law = v*(O) = 0, such that llxil >0= DV(x) f (x, v*(x)) < 0.
(2.1)
which is with
v*(x),
(2.2)
20
IIxII + KI >0
==>
(2.3)
Proof. The point of departure is to recognise that can be viewed as a virtual which a stabilising function v*(x) (2.2) is satisfied. that exists such v*(x),
control for
through towards u, define the global change of variable z== z=uObserve that the feedback u= v*(x) + x= z=, implying that passive stabilisation f(x, v*(x) + z) v*(x). yields the system f(X, v*(x) + z)
(2.6)
can now be used to stabilise the zdynamics. in the form _ f(x, v*(x)) +p(x, , z)z
To this
(2.7)
W(x, Z) = V(x) +1 z2 = V(x) +1 [e  v*(x)]2, and observe that dtW (x, z) = DV (x) f (x, v*(x)) + DV(x)p(x, z)z + zp.
(2.8)
(2.9)
Thus, by taking the control p= for some c>0, it follows that DV(x)p(x, cz z), (2.10)
IIXII + Izl>0d
The control that globally asymptotically
(2.11)
(2.12) 0
21
fo(x, ei)
i=fi
2=
n=
fn(X, e1l...
' n)
+9n(X,
1)...
'en)u
where xER
[Isi99].
The triangular
structure
shows that
Lemma 2.1 can be applied to the upper two systems by viewing fi as the virtual control and by identifying 6 as a regular control. Note that both f, (x, , ) and g, (x, 1) can be eliminated Define for each step i
by feedback provided gi(x, ei) is nonzero on the domain of interest. the change of variables zi = e:  vz i(X, and observe that the recursive application i,..., eii)
reached. At step i the closed loop Lyapunov function is given by Wi(x, zi,..., Zi) =V(X)+ 2 z Ezk.
k=1
(2.15)
Note that the control (2.12) is based on exact cancellations to render (2.11) fulfilled, so that Lemma 2.1 is commonly referred to as exact backstepping [Isi95]. However, when model parameters are not precisely known, but known to exist within certain bounds, it is readily seen that exact backstepping cannot be applied. In case of parameter uncertainties, the control
problem is then to be addressed from an alternative standpoint, inputtostate stability and smallgain
tainties will not be elaborated and all developments in the thesis are in the context of exact backstepping designs.
2.2.2.
Dynamics
stepping design, and this will proof to be quite valuable for bond graph based backstepping as presented in the thesis. It is interesting antisymmetry to note that [Kri95] mentions the possible designs, but the author has not found dynamics.
22
_
sin(x +)
(2.16)
u. stabilised. Towards
arctan(x)
+ x,
(2.17)
x +1 z=uChoose the Lyapunov function (x, z) W wt which implies the control u= v*(x)  cz v*(x).
IL
x
cos(z)  sin(z) lz
+1
(2.18)
(2.8) as W (x, z) = (1/2) (x2 +Z2 ) and write x2 cos(X2 sin(z)  x2z +u+ lx v*(x)J (2.19)
x2 +1+z
x2  x2zCos(Z+ lx
sin (Z)
(2.20)
for some c> 1/2 and where v is the new control. Set d= (x, z) _x 72 cz2 + zv < 
min{1, c(c
2dtW x x2 t1
+1
)z2 +2 v2 2
(2.21)
the equilibrium
to (2.18) and conclude that the closed loop system can be written 1xx _+ xx x +1 cos(z)  sin(z) TI zx form. cos(z)  sin(n) GT +1 z C
101
v,
(2.22)
which is of portHamiltonian
23
2. Fundamentals
In view of robust control, it is generally preferred to majorise nonlinearities implies that the control (2.20) is to be robustified
dtu'(x,
Izl
IX2z
COS(Z)
I+
Izlly*(x)I
lu z sill (Z)
sin(z) I
VIXT _1 (2.23)
}1]
for some function b(x) > (1/2)'x2 control u=L2+2 Global asymptotic stability
x2+l+cJz+zx
+v.
(2.24)
of the equilibrium
z
2x+1
0
point of the above example is that backstepping designs offer flexibility to for systems
The important
(2.9) render negative, such that robustness can be addressed quite systematically (2.1). form However, this flexibility the of is required to have a specific structure.
will not be used in the thesis since the closed loop More precisely, the following dynamics. control v*(x) yields the corollary shows that
(2.26)
V (x), and where
J(x) = JT (x) and R(x) = RT (x) >0 exists a control u(x, 6), with u(0,0),
such that the closed loop takes the portHamiltonian J(x)  R(x) p(x, z) S(x, z), c z) and z=v*(x).
= (x, z) pT (1/2)z2, (x, ST V(x) W(x, + where z) = z) = DW(x, Proof. In view of (2.7), take the control u= v*(x)  pT (x, z)K(x)
(2.27)
 cz,
(2.28)
24
2. Fundamentals
(2.27). (2.2) form Note does loop dynamics the that that the takes closed not observe and hold, since llxll> 0= KT (x) f (x, v*(x)) = KT (x)R(x)K(x) < 0, (2.29) Q
which is merely nonpositive. Thus, going back to the system (2.18) of Example structure 2.1, observe that
the interconnection
IlxiI
(aV\2
ax 
7= +i<o.
2
(2.30)
steppping design would seem to have favorable consequences for bond graph considerations. In particular, bond graphs have been shown to represent a class of portHamiltonian sys
tems [Go102], [Go103], so that closed loop dynamics attained through exact backstepping can indeed be given an associated bond graph model.
Control
Model Matching
Problem (MMP).
In
to the existing MMP theory of [Hui94], this section shows that various concepts of [Isi95] are valuable for an understanding
of closed loop dynamics in the MMP. For example, if the tracking error is asymptotically regulated to zero then this implies the existence of a maximal submanifold on which output matching is fulfilled. characterisation (locally) controlled invariant
of the underlying mechanism with regards to the physical model based MMP addressed in the
has been in the thesis, this mechanism not explicitly as considered where current bond graph literature.
2.3.1.
Problems
The following developments can be found in the works [Hui94], [Hui92]. Consider the plant
P of the form
P: th = y= f(x) h(x). + gj(x)ui (2.31)
25
and yE 1Rp. The vector fields f (x) and gj (x) and the function Then consider the model M of the form x fG) + 9k (x) k
M:
2.32
where xE
Likewise, the vector fields 1(.;v) and gk(x), Notice, in particular, that rn <m
function /a(te) are assumed to be real analytic. both y and y are of equal dimension. Then consider the controller Q described by
Qz= u=
where zE'
(2.33)
M and the controller Q, the nonlinear MMP can now be described as follows. Definition 2.1 (Model Matching Problem). ([Hui92]) Consider the plant P, model M
X (xp, Rn IRf. Find X integer v, an moo, E x and a point o) neighborhoods of xo and of an . V, Q, defined that the open subset V of ', and a map F: XxX such compensator on  VxU, renders the difference y(x, F(x, J5),t)  y(:i5,t) independent of zi for all t>0 the trajectory initialised at x. of y(t) initialised and all (x, J) EXxX. (x, F(x, :t)) at The output (2.34) y(x, F(x, :f), t) denotes of y(t) 0
E:
1th1=11(x)l+Ii(x)]u x f(x) 0L
_+0
11k
9k()
(2.35)
h(am). h(x) e= The extended output e will be referred to as the tracking error and is to be asymptotically regulated to zero. Now, if the model inputs ii are now seen as measurable disturbances then solvability Theorem of the MMP is readily formulated. 2.3. ([Hui92]) The MMP is solvable for (M, P) if and only if the nonregular
26
2. Fundamentals
on Physical Model Based Control will become clear through the Decoupling Prob
The proof of this theorem is omitted here, but its implications developments. upcoming lem (DDDP)
with disturbance
[Hui92]. the compensator of matching would imply 11h(x(t)) h(x(t))Il <M for a
and observe that output and for some N>0. dynamics algorithm
characterises a class of systems that can be encountered in the MMP. Lemma x0 EX 2.4. [Hui94] Consider a square plant P and a square model M. be given. If the decouplings matrix A(x) of P has full rank for x= Let xo EX and
is solvable around (x0, xo) if and only if ri < r2 (i = 1, m). , ... The relevance of this lemma can be explained by the following. DDDP with disturbance In [Hui92], the nonregular that is capable of
plant inputs in the time derivative dr t[h(x) for some r>0. h(x)] =0 
(2.36)
controls, meaning that some controls are set to zero in order for this intrinsic iidependence to be removed. However, this chapter addresses MMPs that have a physical model based "close" to the plant. As a
character of which the prescribed models will be structurally result, the relative degree condition ri
< ri, where ri and rz are the relative degrees of and this is important, the
the plant and model respectively, is readily satisfied. Furthermore, thesis does not explicitly
assume that both the plant and model are square as per Lemma 2.4. be for in < rj will effect all physical systems r=
27
2. Fundamentals
Since bond graph representations are the main theme, the affine plant (2.31) is now specialised to the system
x=
[J(x)  R(x)]K(x)
+ gj(x)uj
(2.37)
(x), hj yj =
more general outputs, and not the collocated outputs yj = gj(x)K(x), of systems to be considered. This will become clear later in the thesis. There are some key aspects of the MMP significant importance. and its associated nonregular
can be used to solve the nonregular DDDP [Hui92], where it must be noted that disturbance decoupling does not address stability of possible internal dynamics. Second, if output matchimposes attractivity of the submanifold on which
ing is to be achieved then the controller output matching is fulfilled. the following example. Example XI 2
_+
2.2. Consider the linear plant P of the form 0010 00 1 1 1 1 1 0 (1 + pi)xi (1 + /12)x2 (1 + 113)x3 10 0
ul +
1
U2
x3
10
4
1
1
1
00
(1 + fc4)X4
00
(2.38)
Yi Y2
X3 X4
output y=
small 4) represents physical parameters with nominal value a=0. ... , consider the model M described by the Brunovsky canonical form
0 0
P
0 0
1
(2.39)
yj =,
where 1<j<2 and (vi _ x4 i 4)
28
on Physical Model Based Control tracking design, because the input iij Let rZ and rz denote the argument yields ri < rZ for by plant inputs.
for gj [Nij90].
The MMP for the extended system (2.35) can now be constructively the constrained dynamics algorithm
addressed by means of
= 0,
.1
(2.40)
Zo =1 (x, c) E R4 x R8 So(x, )= 01, .; . so that So (x, x) = Bo (x, x) + Ao (x, x) dt for all (x, x) E Zo. This yields Bo (x, x) = x2 xl + X2  '5511 i2 x2 It is seen that Ao(x, j) has constant rank ro =1 (so
Eu
=0
(2.42)
(2.43)
giving d)o(x, ) = Ro(x, )Bo(x, ) = x2  Z = 0. Now use I)o(x, l) to extend the constraint (2.40) as (2.45)
Si(x,
)=
So (X, ) _
1X3
xi
X4  xi X2 i i2
(2.46)
with sl = so  ro =1.
(2.47)
29
u=0
(2.48)
The constrained dynamics algorithm terminates since rank A, (x, :i) =m=2,
(2.50)
u2
xi
+. t2 + 3
solves the nonregular DDDP with disturbance measurement. Therefore, as per Theorem 2.3, the MMP is solvable and (2.50) renders Z* controlled invariant for (x(0), :T(0)) E Z*. Looking back on the steps taken, it is clear that the decoupling process does not address of Z* is to be further analysed.
Doing so leads to the conclusion that the decoupling control (2.50) does in fact regulate the difference h(x)  h(x) to zero. To see this, define the "error" variables
e2=X2+x2, e3=X3Xi, e4=X4xi' (2.51
of Z* is confirmed by writing
th1=xl+e32X2
the dynamics
0 =1 1
1 1 00
1 0
e2 e3 e4
(2.52)
xl
12 x2 
x21
be new control inputs, then the feedback (2.50) can now be written
7) _ 7)2X2 i2
ul u2
= =
r) +. ti Xi
+ 2 + 2 + Wi
(2.54)
30
I
2. Fundamentals on Physical Model Based Control
Next defining the error variable el = x1r7, so that by applying (2.54) one obtains the closed loop portHamiltonian dynamics
ei
e2 e3 e4 'bi
/2
00
00 1 0 e3el
e2
10 1 1 1
1
el 1
0
1 0
+ 10 Wl+
0 1
W2
e2
e3
(2.55)
00
e4
00
(01,02) =
are new collocated outputs for feedback and damping structures as well as
the input vector fields are identical to those of the plant. Finally, note that the stability of the closed loop is not compromised by small parameter
perturbations controller
is stabilising
hold in the tracking generally e(t) will not such case and =0 0 0 objective will therefore not be attained. convergence property Stability is an important for the MMP considered here. not obvious
of the zerodynamics
requirement
(2.37) form have intrinsic stable internal dynamics that is compatible the whether systems of with the constraint hj(x) following 0. The proposition = shows that internal stability for
th = [J(x)  R(x)]K(x)
yj = he(x), where DTH(x) antisymmetric hj(x) =0 K(x) = for some smooth, positive
+ 9j (x)uj
(2.56)
definite function
H(x).
Let J(x)
be
be positive (semi)definite,
such that
dtH(x)
(2.57)
31
2. Fundamentals
on Physical Model Based Control with yj(x) =0 is stable since ft(x) < 0. On the
are not collocated, it follows that (2.58) cannot be inferred. Indeed, con
Im Im 0
Qlxl
Im Im (2.59
jQ2x2
positive definite.
The system is of
the form (2.56) but does not have collocated outputs. u= Qixi renders the submanifold x2 =0
It is readily checked that the control The internal dynamics is satisfies (2.60) 0
given by xl = Qlxl
H*(x)
=x
is thus immediate.
on unstable zerodynamics
with noncollocated
systems are often not the quantities to be from a control point of view. For exam
controlled,
ple, in the mechanical domain it is often the case that positions are to be controlled, whereas velocities are the collocated outputs in the portHamiltonian ical to define positions as the outputs example, such noncollocated input variables are identically for control purposes. framework. It is therefore logfor
Model
Inversion
Instead of going through the constrained dynamics algorithm (2.38) inverse the of sider nominal plant given as = U1 =
U2 =
yl ul +yi+yi+y2+V1 77
yl+y2+y2+v2.
(2.61)
32
2. Fundamentals
Next observe that the feedback (2.54) can be obtained by enforcing the relations yl = y2 = i.
Therefore, in terms of bond graphs, if the (bi)causal plant inverse exists and the decoupling feedback can be
fulfilled, degree disturbance is < then ri the condition rti relative found by enforcing the relation h(s) = h(x).
dynamics Z*, that the compute so constrained algorithm not will in the search for this constrained manifold.
2.3.2.
Remarks on Output
Regulation
Theory
The MMP is closely linked with the output regulation problem as presented in [Isi9O], where regulation is achieved when trajectories converge to a center manifold containing the origin.
e=
where xEX, x= E X, uE II8t and f (0,0,0) 0, s(0) =0 = and h(0,0) The 0. model = merely
has eigenvalues on the imaginary Problem" Definition is formulated 2.2 (Full as follows. Information
Regulation
Output
Regulation).
) such that system (2.62), find, if possible, a mapping a(x, .; 1. the equilibrium x=0 of x=f is asymptotically (x, 0, a(x, 0)) (2.63)
stable in the first approximation. VCXx9 (0,0) of such that for each initial a(x,. t) satisfies (2.64) condition
Wi(t)) = 0.
0
Define A= regulation Dy f (0,0,0) and B=D. f (0,0,0), then the solvability requirement of the above problem is given by the following theorem.
33
on Physical Model Based Control Output Regulation Problem is solvable if and co(x) and u= c(am), with
(A, B) is stabilisable and there exists mappings x= if the pair only V(O) =0
9 both defined 0, in and c(0) = a neighborhood of the origin, satisfying the conditions d d P() =f (co(w), , c(am)) (2.65)
(2.63) K is Indeed, K the the a suitable gain. gain renders origin of asymptotically where stable in the first approximation, thereby guaranteeing the existence of a center manifold since cp(x)
Ds(O) merely has eigenvalues on the imaginary axis. Thus, on the center manifold, x= (2.65) is satisfied since u= the and relation Now, to demonstrate the similarities a(co(x), x) = c(x).
(2.65), con
(2.54) (2.55) Example from 2.2 that sider and observe and one can define (n, ) Wi X; 'q
) X_ P(rl,
(rl ) P2
'P3(Th 74(777) )
msz
1 1
(2.67)
and
(i ) l % xi +
u=
z
+2
(2.68)
+. t2 ix3
Then, evidently, the nominal system (2.38) satisfies 0010 dT x) =00 P(7l, 1 0 1 1 1 1 00 1 0 cpl(i, ) <p2(7lx) (71 W3 , x) X) W4(1), +0 10 00 1 ci(7l, )+ 0 1 (2.69) C2(77)t)
d2e2, dl d2 0, the >0 > and and giving control el) w2 with = 
dl u= c(ii, 1)+ 0
0 d2
dl 00
[x  cp(71, 0]. ;
(2.70)
34
2. Fundamentals
has been shown to be closely related to the Full Information In conclusion, the 1VIINIP Regulation
Problem found in [Isi95], and this fact is used to characterise the bond graph
2.4. Stabilisation
Control through
Energy Shaping
The concept of passivity can be safely argued to be a pillar trol [Wi172], [Des75], [Byr9l], where the notion of rendering
of systems theory and conthe system passive through The idea that control can be
feedback has shown to be effective [Ort89], [Ort98], [Str98]. associated with energy storage and dissipation
the passivity framework to the subject of physical systems modelling. This section recalls some basic aspects of control subsequently through feedback passivation that will
Feedback passivation requires where this thesis shows that in certain atten
the knowledge of some suitable closed loop storage function, the bond graph junction cases. In addition, structure
portHamiltonian
tion [OrtOOc], [B1a02], [OrtO2b], and this type of feedback is generally referred to as Interconnection and Damping Assignment Passivity Based Control on IDAPBC (IDAPBC). Fundamental facts
is based on structural
Consider the affine control system th =f y= where xEX Definition (x) + 9(x)' h(x), (2.71)
and u, yE IRt, and where f (0) =0 and h(0) = 0. 2.3 ([SchOOb]). The system (2.71) is said to be passive if there exists a function such that yT (s)u(s) ds (2.72)
(X(0)) <J
35
2. Fundamentals
There are two properties of passive systems that are of interest. Let (2.71) be passive with a (x), function definite V then smooth positive storage 1. for u=0 the point x=0 is stable by V (x) = DV(x)f(x) < 0, <0 for u*(x)
2. for y(x) =0
compatible with y(x) = 0. Therefore, passivity with respect to a positive definite storage function and zerodynamics Since stability stability (2.71). the of system system is somewhat restrictive, it can be of interest to implies zeroinput
of the zeroinput
(2.71) to are render an unstable system passive by means Towards that end, consider the smooth feedback
of feedback, hence feedback passivation. u= (x)v, 6(x) invertible +, with cx(x) , x=f y=
(2.73)
Suppose the above closed loop system is passive with smooth positive definite V (x), then from the above two properties it follows that the zerodynamics i= for v*(x) f(x) + g(x)u*(x) f(x) = + g(x)a(x) are stable, where (2.74)
+ g(x)/3(x)v*(x)
y(x) [SchOOb]. Hence, 1(x)[u*(x) if the syscompatible a(x)] with =0 = , tem (2.71) is to be rendered passive by means of feedback passivation then the zerodynamics under feedback.
In case the system has been rendered passive with respect to the positive definite storage function V (x), it is readily seen that with v= ry, for some r>0, the asymptotic stability
2.6. Interconnection
and Damping
Assignment
[OrtO2b], [OrtO2a], [OrtOOc]. has been in system studied various papers are a main theme of the thesis seems to
Because closed loop bond graph representations indicate possible bond graph interpretations
in IDAPBC.
36
2. Fundamentals
x= [J(x)  R(x)]K(x)
(x)K(x), y= 9T where xEX and u, yE R', and KT (x) = DH(x)
 g(x)u
(2.75)
H: X  R.
(2.76)
(x) and R0(x) = R (s) >0 are the assigned interconnection a(x) +w be a smooth feedback such that + [J3(x)  Rs(x)]Ifa(x) = g(x)a(x). form
(x)]K(x) Ra, 
(2.78)
It is readily verified that the closed loop has the portHamiltonian [J8(x) R3(x)]K3(x) = Ys = 9T (x)Ks(x), showing that the feedback a(x) yields a portHamiltonian connection and damping structures.  g(x)w
(2.79)
Rs(x)]Ka(x) 0, = 
(2.80)
of g(x). The relation (2.80) represents a set of first that are to be satisfied simultaneously by the
(PDE's)
function Ka,(x). Clearly, when (2.80) can be solved for some function Ha,(x) then the control is given as
(x)s(x)]'gT [gT
= U.
(2.81)
function H3(x) need not be "guessed" but follows from (2.80). That is to say that the
IDAPBC method characterises all possible energy function Ha(x) that can be assigned.
37
hold that
Ka(xe) = K(xe), (xe) In K3 0. addition, ensuring = one must have D2Hs(xe) > 0, (local) is that minimum xe a strict so to assign a nonzero equilibrium Even though IDAPBC of the energy function H5(x). (2.83 In this way it is possible (2.82
or "forced" equilibrium
For example, there are no clear guidelines in regard to the choice of the stabilisation conditions can
Furthermore,
[OrtO2b] the as authors of point out, no explicit of (2.80) for choices of Ja(x) and Ra,(x). example of bond graph representations
of IDAPBC
is presented, where the bond graph topology interconnection and damping structures.
2.7. Concluding
Remarks
Three control methods have been presented that are largely based on structural to allow for closed loop bond graph considerations.
considerations
exhaustive and other control strategies may well exist that allow for structural design goals. Furthermore,
and the interested reader is referred to the various citations and references therein for further details. It is interesting to note that backstepping, model matching and energy shaping have major differences on the analytical level, yet all three control methods allow for structural design
dynamics for example. This has not been addressed in the of a particular set of control methods for
38
Part II.
Bond Graphs for Closed Loop Dynamics
39
3. Backstepping Control
3.1. Introduction
literature
of backstepping
of this chapter is the detailed presentation of bond graph dynamics. Some results on bond graphs [Gaw04] on
in backstepping control have appeared in [Yeh99] and in the works of [Gaw0l], virtual actuators. However, these papers do not address the portHamiltonian multiinput
framework and
do not address certain nonlinear cases. In addition, considerable attention in the bond graph literature.
a sufficient number of open questions remain on bond graph based backstepping the various results in this chapter.
3.2. Backstepping
Control
control. Bond graph tools in backstepping as introduced by [Yeh99] will be addressed through detailed examples and subjected to portHamiltonian (bi)causal approach to backstepping a "shortcut" considerations. Furthermore, the novel
[Yeh0l] [Yeh02] backstepping design See for in to method a certain cases. and
further developments on backstepping control in the physical domain. The explicit association of portHamiltonian stepping design is an important tion 2.2 is selfcontained contribution dynamics with the closed loop through a backof the chapter. Backstepping theory of Sec
and forms the basis for the all developments, showing the clear
parallels between existing theory and the closed loop bond graph considerations.
40
3. Backstepping
Control I
c 1 ssk
Figure 3.1.: Virtual x
r 
Yo
L____________J
exam
ples are used to introduce fundamental bond graph arguments in the context of physical model based backstepping. The first example shows a simplest backstepping design conceivable, layactuators and stabilising functions in a bond graph context. The control problem that
is can be addressed with bicausal bond graphs; the actual bicausal bond graph approach will not be addressed until later sections. Nov, because the first two examples are onestep designs, the third example addresses a multiplestep be presented in considerable detail here. Example 3.1. Consider the bond graph of a massspring system in Figure 3.1 with the design taken from [Yeh99], which will
(3.1)
actuator defined as a
Then consider the smooth function yo(x) as the output resistive Rl element.
through a backstepping
control design. By ignoring the dashed region in Figure 3.1, causal analysis of the bond graph yields the system
th =
k
x 7n +1 (3.2)
k = yo(x) + kz,
(3.3)
41
3. Backstepping C
Control
k01 SS
R1
Figure 3.2.: Dynamics (3.7) of Example 3.1. yielding the xdynamics t =f (x, Z) =f (x, (x) + z) = (x) + kz. go 1 go (3.4)
Hence the desired effect of the virtual and proper Lyapunov function
x2 +1  1),
(3.5)
+1
+x mx +1
Since Rl is assumed to be globally resistive implies that xyo(x) >0 the xdynamics the control u= globally asymptotically (1/k)yo(x) + stable for z=0.
Now, the following conveys a key aspect of the ideas of this chapter: Observe that (3.7) has the bond graph representation depicted in Figure 3.2, which is seen to be identical to the resistive element, where u and are to be interchanged
plant bond graph with the virtual with It and z respectively. Further stabilisation
of (3.7) is readily achieved by replacing the SS element with a linear (k/f)z for some positive damping
constant r, for example. Doing so yields the closed loop bond graph in Figure 3.3 of which the elements are defined as
I(x) =1(
x2 +1
1),
C(z) =1 kz2.
(3.8)
42
3. Backstepping
Control I
C
z
R.2
A11/
9o
Rl
Figure 3.3.: Closed loop bond graph of Example 3.1. Finally, take W(x, z) = V(x) + (1/2); z2 and conclude that XI } Izl >0 dtW(x, z) _ m z +1Tz<0, yx2 (3.9) stable because W (x, z) 0
(x, (0,0) is globally asymptotically that the origin which shows z) = is positive definite and proper. Some important
observations on behalf of Example 3.1 can now be made. First, the bond as mentioned in [Yeh99], which is a form if backstepping is expressed in
in [Yeh99] for the backstepping design. Third, the change of variable (3.3) is not written the conventional form as defined in Lemma 2.1, but it can be derived naturally graph in Figure 3.1 by considering that go cannot be placed at the 1junction the control variable u. So, it is intuitively
plausible that k should "carry" the term yo + kz control. Finally, observe that the closed and R2 components.
This emphasises the idea that the controller should induce physical, closed loop dynamics by emulating plant interaction with another physical system [Sha9l].
Remark 3.1. From now on, all bond graph elements that are part of the backstepping design are overlined as demonstrated in Figure 3.3. This should separate and clarify those parts of
the bond graph that belong to the open loop plant. The main point so far is that backstepping in the physical domain can be facilitated the stabilising control as the output by viewing
location, where it should be noted that single bond graph components are the simplest virtual actuators that can be defined. However, there is no reason to restrict so that the stabilising the virtual actuator
solely to single bond graph components, controller states and controller inputs.
function
may depend on
43
3. Backstepping Control
kr umd
Figure 3.4.: Simple massspringdamper In [Gaw0l], the authors address a setpoint with a single control input.
As a result, the backstepping design is shown to yield a dynamic simple modelling arguments.
3.2. ([Gaw01]) Consider the physical system depicted in Figure 3.4 and its bond in Figure 3.5. The element definitions CO = 2k2, I(x) are given by R=r. (3.10)
graph representation
21x2,
is to find the velocity u such that the closed loop dynamics behaves actuator
like the system depicted in Figure 3.6, where it is a velocity input to the virtual
and where d is a constant disturbance force acting on the mass m; the velocity of mass m is the system output w conjugate to d. The dashed region in Figure 3.6 represents the virtual actuator of which the bond graph is depicted in Figure 3.7, where the virtual elements have the definitions
1
bond graph
Ci(Ji) =2
ixi(1
12
+ 2xi),
It should be noted that the C1 storage element has the characteristic instead of the linear characteristic the restoring
deployed in [GawOl]; using such a hardening spring allows for larger excursions. Since the junction structure
topology is independent bond graph based virtual storage and dissipation Without
of the bond graph element definitions actuator offers some design flexibility
functions.
loop further is it that the analysis, plausible closed system will meet the setpoint any for consider a step velocity p to the right such that the virtual actuator
control objective,
force until the step velocity is reached. Next observe that the closed of Figure 3.6, where the R element
44
3. Backstepping
Control
CI
SS1
01
k2
L 
r
J
Figure 3.6.: Target closed loop system of Example 3.2. The backstepping design starts with the simple causal analysis of Figures 3.5 and 3.7, yielding the model
ked
k erm 1xu (3.12)
= =
=
o  =2x2 r
1x1
+ o
+ 1) 2) + K'2;
as defined in Section 2.2.1 is not a function of the states x. To see this, observe that the bond graphs in Figures 3.5 and 3.7 are interconnected the backstepping method gives the virtual control to satisfy the constraint iio = yo such that
k_
kz + yo
v*
(3.14)
45
3. Backstepping
Control
ss 71
0 ll 2
C1
Figure 3.7.: Bond graph virtual The dynamic nature of the virtual of Section 2.2.1, since the virtual actuator
C2
actuator of Example 3.2.
control is generally taken to be a static function design can still be applied unmodified
where the
appearance of controllers states J1 and x2 does not alter the backstepping procedure. Next use (3.14), (3.12) and (3.13) to write 10010 X2 th z00 0 1 1/r 1 10 01 10
l(xl  xl k2x2 x/m
0 0 0
klr + yo/k
100
100 001 010 (3.15) it u d
kz
By considering the damper r in the target system of Figure 3.6, choose the feedback U=which induces the closed loop 0010 kl(xl + xi) 10 e+! yo+z, (3.16)
x2
th
0
1
1/r
1
10
01
k2t2
x/m
10
01d
jc
(3.17)
z00
1
1/r
kz
00
of Figure 3.8.
+ 2kz2+
(3.18)
46
3. Backstepping Control CI
,zx
r
 d
1o0 uo 1w;

SS
SS2
I
x2
C1
L 
C2
j J
of Example 3.2.
 wd.
(3.19)
Thus, the closed loop is passive with respect to the supply rates V)p and wd. The feedback is found by reversing the change of coordinates (3.14). Q
It must be noted that Lyapunov arguments have not been used to obtain the control in the Examples 3.1 and 3.2. Instead, the closed loop Lyapunov function is implicitly contained in
the bond graph based backstepping design by retaining the form of the Hamiltonian. Even though Examples 3.1 and 3.2 are onestep designs, multistep designs are realised in an
analogous manner. The example found in [Yeh99] will now be presented to show a twostep design in detail, where the causal path between the virtual actuator and control readily shows the bond signals that are to be transformed. Example 3.3. ([Yeh99], adapted) Consider the massspringdamper system depicted in
2ml
1
cl(ef)
=I2,
C2(x2) =
2m2 Ik2xz.
X3,
I3(S1)
=1
23
X2
(3.20)
R=r. d to be forces
It is emphasised that this examples considers the control u and disturbance instead of velocities.
47
3. Backstepping
k1
Control
k2 r
ml
m2%VDd
3d
of Example 3.3.
C2 12 13
x3
X1
SS2

m1 2
0)1I0
Figure 3.10.: Bond graph system of Example 3.3. Suppose the system in Figure 3.9 is to attain through analytical feedback on u. a particular
IdIw
R ssl
attenuation level problem in an
disturbance attenuation
fashion, the context of the chapter demands that the closed loop dynamics is [Sha9l]. system
Consider the physical system in Figure 3.11, where the control objective is to find an appropriate (dynamic) with this system. feedback u such that the closed loop dynamics from d to w is associated Clearly, the controller is to induce closed loop dynamics with the bond elements are defined as R2 = rz, R3 = f3 (3.21)
The most characteristic step of the backstepping design considered here is the choice of virtual controls kll backstepping and e2/ml depicted in Figure 3.10. Thus, without any further analysis, the
graph in Figure 3.12 it is readily seen that the efforts of the virtual cannot be imposed by the regular control u, so that a backstepping
Thus, as a first step in the design, the reasoning from the first two examples would suggest that the bond signal kill should "carry" the effort imposed by the virtual actuator composed
48
3. Backstepping
Q
Control
k2 T
h%1 ml
172 m2 m3
T3
xxx
A , rl)
Il
z2
Cl
ji 1 xg
I2
C2
X1
I3
C2
\101
1/
xi R3 R2 C1
dw SS,
Now, the backstepping design starts with the causal analysis of Figure 3.10 to obtain the
system
X1 2 =
010 1 0
11 m2 ml
Lx3
2= Note that the xdynamics application of Corollary
loci
2.2 possible. Then, in accordance with Figure 3.12, the first change
of variable is found to be
rl 11 =x3m2 kli + klzi = v =
is portHamiltonian
By invoking LaSalle's theorem [Kha92], the point (t1, x) _ (0,0) is globally asymptotically (3.23), In stable. view of write 11* Z2  voi il =X3 +m2 ml (3.24)
49
Control
1
ml 2=
kl
zl + vo + ml z2
vl =
mlkl
TZ
+ mlvo. . L
(3.25)
m2
The backstepping design is completed by evaluating the time derivatives v and vi and by substituting the definitions for zl and z2. Collecting the results, it is now readily verified that
xl
X2
0
0
0
0
0
0.
0
0
1
0
0
0
0
1
klxl
k2x2
0
0
1 X2 X3
zl
0 =0 1
0
0 0 0
0
0 1 0
0
1 1/r 1
0
0 1 rl
1
0 0 1
1/r2
0 0 0
1
1 0d. 0
0
z2
w =
0
xl/m3.
1
1
r3
z2/ml
0
(3.28)
Just as in the first two examples, the above backstepping of the virtual normally virtual controls v and vi through
controls in view of Lemma 2.1, but these Instead, it is more intuitive to take
to some extent.
these bond signals a;,C are readily selected from the causal
path connecting the control u. The following important observation can be made on the influence of the disturbance d
depicted in Figure 3.12. Even though it would seem obvious that the target closed loop can be attained through a backstepping respect to the disturbance more difficult to address. design, the relative degrees of the virtual controls with
50
3. Backstepping
Control
The relative degree problem can be loosely explained by observing that vQ in (3.23) depends on x3 and (3.25). depends by 1 that In turn, the control u has a so vi on x2, x3 and
dependency on xi because of (3.27). It can now be concluded that the closed loop does take the form (3.28), but this will not be the case when the R element in Figure 3.10 is virtual as
well. In such a scenario the backstepping design would require an additional step, rendering u directly dependent on d. But disturbances are generally assumed to be unknown, so that the disturbance cannot be removed by feedback. In such case, the closed loop will not allow for of Figure 3.12 and the disturbance would feed through at the
To clarify this problem with a simple counter example, consider the system JE = = It is readily seen that for z=+x and u= = z= d XU. loop becomes the +z closed x x+zd d. x (3.30)
(3.29)
addresses various observations and conditions for such designs to be applicable. material to be presented has certainly results presented here explicitly
been inspired by the work of [Yeh99] and [Gaw0l], the formulation to define the control Onestep and
objective and to give the closed loop an associated bond graph representation. multistep Proposition designs are considered. 3.1. Consider the singleinput
functions be Let 1 the to to the the represents virtual actuator connected realvalued junction. H(x) and H(x) be smooth, positive definite and proper, where xE 1R and 7r E R7,0. Define . the cascaded element as C() = (1/2)ae2 for some a>0.
51
3. Backstepping Control
E: H(x) uloy 0
1
u y
11 k01
ho
H(x) E: /
3.1.
Suppose the systems E and 2 are explicit bond graph models with the inputoutput (uo, yo), (moo, yo) E IR XR respectively, where y= y(x)
pairs
) (dynamic) feedback law , such that the closed loop allows the bond graph a smooth u(x, representation of Figure 3.14, where C(z) = (1/2)az2 is given as (3.31) and R=r.
z z2 < 0,
(3.32)
U(x, x)
Proof. Since the systems E and Eo are explicit bond graph systems, it follows that E can be given the form x= [J(x) R(x)]K(x) +g(x)uo (3.33)
(x)K(x), Yo =9T where KT (x) = DH(x). Likewise, Eo admits the portHamiltonian representation
(3.34)
52
3. Backstepping The interconnection the form ae = yo(xo, Jo(x)) +az The system in Figure 3.13 then becomes t x
,= 9T (x)Ic(x)
Control
(3.35)
g(x)9
(x)
K(x) Ko() I
g(x) 0 az
Jo(5)  Ro(J5)
(3.36)
Hence the control 1_azv u= induces the closed loop of the form J(x)  R(x)  g(x)bo(.; )gT(x)
= 9o(.t)gT gT (x) (x)
T
(3.37)
9(x)9
Jo(2)
(. t)
g(x)
0
K(x)
Ko(io)
az (3.38)
Ro(t) 0
(x)]K(x) +K
(3.39)
Because W (x, , Wie)is positive definite and proper, the origin is globally stable since U(x, xe) is assumed to be positive (semi)definite. the origin, consider the set P= {(x, z, i) : U(x, x) =a z2 = 0}. To investigate the global asymptotic stability of
(3.40)
Let Po be the largest subset of P that is invariant under the dynamics (3.38), then by LaSalle's Theorem [Kha92] the origin is globally asymptotically It should be noted that J(x) and R(x) stable if Po = {0}. Q
the class of systems E can be enlarged by observing that both modulated with . Observe that the structure
t) positive (semi)definite. and where U(x, , . However, observe that if the change of variables (3.35) is to remain valid then it must hold matrix J(x, ) remains trivially antisymmetric that go = yo(x, yo(x)) does not depend on e, thus dg0/dd = 0, which is clearly satisfied for the relation g(x, ) = g(x). R(x, e) = R(x). From now it will be assumed that J(x, ) = J(x) and
53
3. Backstepping Control
TE:
H(x) Yo uo 
SS
0' lo
Po : Ho(.; V)
3.2.
E: H(x) Yo uo
0 y
20 : Ho (.; 0 C)
Figure 3.16.: Closed loop Icascaded system of Corollary 3.2. Corollary 3.2. The Icascaded system depicted in Figure 3.15 allows for the bond graph
representation
of Figure 3.16 by some smooth feedback u(x, , a). analogues to Proposition 3.1, but (3.32) evidently reads xO)  ra2z2 G 0, U(x,. t). (3.41) Q
Proof. Entirely
W (X, z, ) = U(x, dt
for all nonzero (x, z, x) and positive (semi)definite The following corollary
Figures 3.14 and 3.16 can be generalised to more complicated systems. Corollary 3.3. Consider the systems in Figures 3.13 and 3.15. Then there exists a smooth ), with = , such that the respective closed loops in Fighas the states
ures 3.17 and 3.18 are attained. xl E II8"1 and the input/output
Proof. From Figures 3.17 and 3.18 define the control in (3.36) as u= yl(.: i1, az)  v* = 9i (xl)Ki (:TI) + bl (J1)az v*. (3.42)
54
3. Backstepping
Control
CE:
H(x)
Figure 3.17.: Closed loop Ccascaded system of Corollary I T yo uo El : Hl(: i1) Y1 iij I1/0 UO go E: H(x)
1 20 : Ho(x) /I 3.3.
Figure 3.18.: Closed loop Icascaded system of Corollary Then rewrite the portHamiltonian **0 **00 x1
zjL 9T
00
(x) 0
J1(1)  R1(xl)
91 ( 1)
9i(x1)
41 (i') JL
Ki(1)
az
So far, the C and I elements are taken as simple quadratic so can be attributed
to the fact that onestep backstepping designs as defined by Lemma 2.1 function of the form W(x) (1/2)z2. V(x) + = From this
standpoint, if the plant dynamics already has the simple quadratic storage (1/2)ae2 associated with aC (3.35) I then the or element, change of variable simply interchanges the role of 6
with z. Thus, such a backstepping design renders the closed loop Lyapunov function identical to the plant Hamiltonian; however, due to the change of variables, one must be aware that it is Furthermore, the
not possible to associate physical energy with the closed loop Hamiltonian. closed loop bond graph represents a physical system in conceptual sense. Now, it would seem that Proposition change of variable z=z= (1/a). ()  v*(x) 3.1 and Corollaries 3.2 and proposition
to the
55
ditions apply except that C(6) = F(6) function F(t; ) satisfying F"(6)
>0 for all 6. Then the closed loops in Figures 3.17 and 3.18
are attainable by smooth feedback u(x, J,6). Proof. Put )() = F'() and observe that the plant has the form
[J(x)  R(x)]K(x)
_ gT (x)K(x)  u.
+g(x)A()
(3.44)
Now, in the same fashion as (3.35), define the change of variable A(e) = yo(o, yo(x)) + az, (3.45) actuator output go Because F(e) implies = 0.
which reflects the idea that the bond signal A() is to "carry" the virtual
and where az renders the C element into a simple quadratic storage function. is positive definite and satisfies F"() >0 for all , it follows that A() =0
The change of variable (3.45) is seen to yield the system [J(x) R(x)]K(x) yo(x)) +9(x)az  9(x)yo(. 7co, 1. ()K()  u + t  yo(x)),
(3.46)
aA/(e)[9T
from which to derive the control
yo(xo,
(3.47)
C()
)= F( ae arctan(e) a =
>0
ln(E2 + 1),
for all and a>0.
(3.48)
The plant is
readily given as
a arctan(C) x +I rra x (3.49) U.
56
In this case, observe that the change of variable because the dynamics x= a arctan(x) is readily simple, a must be
stable. Although
1 a arctan(x + z) = a arctan(x) +zJ0 _ a arctan(x) + p(x, z)z, where p(x, z) is smooth. This procedure will induce portHamiltonian of Figure 3.3. arctan(x + s)
Ltz
dt (3.50)
as per Corollary
2.2
control, identify
(3.45) bond logical is the a graph context and yields the virtual variable choice within a arctan() __+ 1+ az,
(3.47) Rl be in Figure linear. 3.2 is Then the to resistive element where chosen reads U=x+ (2 + 1) x x2 F 1 2+1 ri rlx am(x2 + 1)3/2 am x2 +1 ] az F2 (3.52)
x2
1m
which induces the closed loop rl 1 Global asymptotic stability 1 1/r2 x/(m x+ az 1)
(3.53)
follows immediately.
of Proposition
3.1 and Corollaries 3.2 and 3.3 is readily possible for shows
systems having a cascading sequence of C and I elements. The following proposition the recursive application of Corollary 3.3 that encompasses Proposition
Just as in the onestep design, the closed loop retains the plant structure to which additional bond graph elements are added that represent stabilising step a new coordinate is introduced provide arguments for stabilising dynamics. More precisely, for each
elements.
57
3. Backstepping
Control
Il
2T
C1
11
E: H(x)
Uo Yo 
a1 u ...
01
fy
20 : Ho(. to)
3.5.
1zi
1p1 'Li 91 2
uo
yo
20 : Ho(xo)
92
E2 : H2(x2)
El : Hl(. t1)
3.5.
3.5. Consider the cascaded system depicted in Figure 3.19, where E and Eo are The input/output pairs are
explicit bond graph systems defined on ][8n and IR'O respectively. (no, u), (o, go) E II8 x IR with yo = yo(x) and y0 = y0(.t, ). H(x) H(x) and
(3.54)
Then there exists a feedback u(x, t, t; ), with .: (xo, ) (ei, ), loop bond Ca,. that the i' the and such closed admits graph represen= = .. .. tation of Figure 3.20. The explicit bond graph systems Ei are defined on l[8"=, for i>1, and where aZ >0 for all i. pairs (icy, g) E ]R x IR with yz = yj(Jz, Uz).
Proof. Depending upon which element terminates the sequence, the control u in Figure 3.19 is either an effort or a flow. Now, the first step of the design is analogue to Proposition and the first change"of variable is therefore defined as a1 1= Yo(xo, yo(x)) + alzl thus zi = yo(x) + a22  v. (3.56) = vo =1 go(xo, yo(x)), (3.55) 3.1
al
58
3. Backstepping
Control
By considering the target dynamics of Figure 3.20, it is seen that the second change of variable
becomes
a26 = yi Then, for j>2, 1, (J alzl) +, &0 I azzz = vi =2 1, [yi (.; v aizl) v] (3.57)
where
ai+1. 7+1 = aj1vj2 +'Uj1 y. Y(xj, ajz7) + a. 9+1zj+1
of elements in
Figures 3.19 and 3.20 for which the C and I elements are swapped. Corollary 3.6. Let the elements in Proposition
C~'j(6j1) = Fj(aji),
3.5 be defined as
Ik(ak) = Gk(6k)i
(3.60)
>0
Fj(e2j_1)
and G'k(e2k) >0 for all l;Z. Then the closed loop in Figure 3.20 can be attained by smooth feedback.
Proof. The first step of the design starts with the relation Ai (ei) _ 9(x, yielding the z1dynamics yo (x)) +alzi, (3.61)
zl =i
Ai(ei)[yo(x)
+ A2(6)] + iYo(xoiYO(x)).
control
(3.62)
To further enforce the target dynamics in Figure 3.20, define the virtual A2( 2) = Jo(x) + AI(i) so that the z2dynamics
z2
(3.63)
becomes
o(x)] +)%3(e3)1 + l) 
al
2i(i)[2(
2)[ill
ai
2Ai(Zi)[A2(e2)
 yo(x)]
+2
Eo(x) al1
+
+ yi(xl,
alzl)]
(3.64)
59
3. Backstepping
Select the expression for X3(3) that "cancels" the righthand the proper target dynamics. Unfortunately,
tational purposes, but it is readily seen that the recursive process does yield the closed loop portHamiltonian dynamics of Figure 3.20.0
The recursive scheme for nonlinear elements can be clarified by considering Example 3.3, but assume that the C1 and Il elements are nonlinear Proposition Example 3.5. 3.5. Consider Figure 3.10 and define C1(1) = F1(1) and I1(e2) = Gl(2), where elements that satisfy the conditions of
the actual definitions of C1(e1) and I1(62) will be omitted to avoid some algebra, but observe that (3.22) now reads x1 2 : = 010 1 1/r 1 x1/m3 k2x2 +0 01 Al (ei) 0d
X3
1= 2 =
0
y0(x) A1(ei)
1
+, \2(e2) u.
x3/m2
10
(3.65)
The design can now be based on (3.61) and (3.63) for which al
kl, 1/ml a2 = =
and
A(b) _ u.
As mentioned earlier, the class of systems depicted in Figure 3.19 of Proposition
0
3.5 are
However, it is possible to enlarge this class of systems by allowing explicit bond graph of the cascaded pattern of C and I elements. For C element, but
it is understood that the same arguments equally holds for a quadratic I element. Corollary (q) be an explicit bond graph model with the input/output pair 3.7. Let E, : ., 7 (u,,, yn) E IR x JR with y,, = yq(ii, u, ). Suppose E,, is connected to aO junction of some quadratic Ci element as depicted in Figure 3.21. Then the recursive backstepping procedure of Proposition 3.5 can be applied if Eq is inputtostate stable.
60
3. Backstepping
Ci
ibi
Control
ai F
ai
u17 y,1 E, : H, 7 7(71) Figure 3.21.: Extended backstepping junction of Corollary 3.7.
Proof. First observe that the bond graph in Figure 3.21 yields the system z = a2izi + ai+let+i  y,7(i, ajez)
[In(k) Rn(i)]Kn(ij) = 
+9n(ii)ajej.
(3.66)
(77)Kn(7l) + bn(il)aLe=. yT/ = 9,7 The target dynamics of Figure 3.20 are attainable
ai+16i+1 = ai1v 1 vi [ai1vi2 + 2+ vi 1ji(xZ,
(3.67)
Uz 1vi (xi a{zi) + y1(7], aji j)],
ai+l
thereby removing the influence of y, from the target dynamics. 7 the internal rqdynamics must remain stable for arbitrary
bounded input u, 7.
(3.66) (3.67) Take the usual change of variable j = zi + vz that with reads 1, so
zi = ailzii  yi(x2, aizi) + az+lzi+1
(3.68)
(3.69)
Consider the physical system depicted in Figure 3.22 and its bond graph in Figure 3.23, where the various elements are defined as
representation
I(x) = 21x2,
Ci(f) =1 k1 2, 22
Ca(ll) =
1k27]2,
R= r.
(3.70)
61
3. Backstepping
Control
k2 kl
md
SS11 uw
01Id
SS2
Rk
1 jC2
11
Figure 3.23.: Compound backstepping bond graph of Example 3.6. As in example 3.2, suppose that the virtual mass m with the connection constraint actuator in Figure 3.7 is to be connected to the 3.7 it is still possible to
zi = w, then by Corollary
attain the closed loop of Figure 3.6 through a backstepping design. The point of departure is straightforward causal analysis to obtain the system
th =
_ i=
kl d 11
'k2
rr
Xi+
kl
T ki
,
k2
u+ T77
(3.71)
(3.66) dynamics the where are clearly recognised. Then by recalling Example 3.2, take the change of variable
1 vo =1
yo(x,. i)
(3.72)
62
Control
x=
z=xz
7=
(3.74)
It is clearly seen that the linear 71dynamics are driven by v + z, so that the inputtostate stability property it trivially fulfilled. More precisely, observe that u,, = ki (v + z) and that
The presented results on bond graph based backstepping did not yet address possible disturbances entering the subsystem E: H(x); that backstepping however, it has been shown by Examples 3.2 and 3.3 attenuation problems. Now, the presented
theory can "in principle" is to have a particular degrees are instrumental Proposition E: H(x)
be applied unmodified, but caution must be taken if the closed loop The following proposition shows that relative
desired representation.
is denoted as d and the conjugate output as w. Suppose that the cascaded pattern of quadratic Then loop Figure 3.20 in the C and I elements define the state variables (1i closed ... , k). d. has degree be to if >k ro respect yo(x) a relative with attained can Proof. Recall that each system Ej : Hj (x1) with j>0 of the form (3.76) is an explicit portHamiltonian
system
Set j=0
tr"
dtryo(oJo(x))
(3.77)
63
3. Backstepping
This implies that (3.77) does not depend on d for 0<r<k. seen to be true for k=1 since zi = yo(x)
which write does not depend on d by (3.77).
+ a2e2  v,
(3.59) in the following steps, set k=2
(3.78)
and
To apply
(3.79)
1 f d2
a2
dt2v0  dtyl(:
11 alzi)J
,
(3.80)
t2v0
d2
yl(1, aiz1)]
1`d
(d3
2 a2Z2)J  dty2( , .
d_
(3.82)
It is clear that z2 depends on at most d2v/dt2 and that z3 depends on at most d3v/dt3. The . above procedure can be continued by backsubstitution of previously defined virtual controls and their time derivatives. In addition, the various time derivatives of Pj(xi, ajzz) can be
(3.76) through resolved and the relations zi = z  vz 1. Then, by (3.77) and ro > k, it is seen that the closed loop in Figure 3.20 is attainable since zk depends on at most dkvo/dtk, proving , that the disturbance d does not enter the zdynamics. Q Various techniques for bond graph based backstepping have been presented, but this section is by no means exhaustive and further extensions of the various developments are readily
conceivable. For example, the class of systems suitable for bond graph based backstepping can be further enlarged by certain modulations of the systems Ej: Hj(xi). In any case, it
is safe to say that valuable problems have been presented to allow for such extensions to be developed in the future. Bond graph based backstepping as considered here addresses the case in which no derivative causalities are induced by the bond graph topology. It can be argued that such bond graph models belong to a relatively somewhat limited. small subset of models, rendering the backstepping method
Future research could look into bond graph based backstepping in case of
64
Control
This section briefly shows how the results of the previous sections can be applied to multiinput systems, but no attempt singleinput For simplicity, is made to develop additional results since the techniques for the
case. system
3j=
anj1Snj1
UJ,
where xE
Rh and i = (ai.....
nj).
branch j c IP"i, where the application end, the design starts with the relations
aiei = yon (Jo1,yon(x)) + aizi, where the recursive scheme applies to all relation (3.84) in parallel fashion.
(3.84) Depending on
the various values nj, the backstepping design may lead to certain controls uj to depend on other controls ui, hence the lowest dimensional branch 3 is to be resolved first. Example 3.7. Consider the twoinput C2(i) =1k system in Figure 3.24 with the element definitions (i)2, (2)2' I1( z) =1 2m2 12(X) =1 x2. 2m (3.85)
C1( i) =1 22 i(i)2,
The control objective is to impose the closed dynamics with the bond graph representation
of
Figure 3.25. Even though it may seem that the control problem is considerably more complex than the singleinput without modification. actuator is given as a single resistive R elecase, the backstepping design for singleinput systems can be applied
ment, where the causal path from this resistive element to both controls emanates from a common 1junction. In such a scenario it is possible to fictitiously virtual if the replace element with outputs gol and 902
actuator
65
3. Backstepping Il Uz Cl ix
Control I2 C2 l
SS1
11/01/0
11
Tb
1 0 R1
iR
Figure 3.25.: Target closed loop of Exercise 3.7. to the corresponding signals kjQ and ki 1. Now, causal analysis yields the dynamics
x= i 2
1
k'Z
nisi 
= =
=1
(3.86)
 ul
U2,
which are of the form (3.83) with J(x) = R(x) = 0, gl(x) with (3.84), define the change of variables
kii r = 2x+kizi r
In accordance
2klx Im 2k2mX i
(3.87)
kg 2 = + 2m yielding zl
zl
11 =
=X'nom
+'n1
a
v01 (3.88)
M case, the target dynamics in Figure 3.25 shows that one can choose
u2  v02.
klzl rl
+ vpl +
11 m2 z2 = vii = mi rl
kkzl
+ m2v1
(3.89)
66
Control
(3.90)
The z2dynamics
m2
z2.
(3.92) systems
The design technique considered so far is identical to the procedure for singleinput
Also, it is readily checked that the expanbranches instead. that two considered must except sion of the control ul requires the definition that nl =2 for u2. This is readily explained by observing
and n2 = 1. By applying both controls to the plant in Figure 3.24, the target representation 1 10
r2 0
r 1
10 1/Fi
1
x/m kizi
z2/m2
(3.93)
100 stable.
1/r2
kizi
Note that simple quadratic C1, C2, I, elements have been considered so far, but the structure of the portHamiltonian dynamics (3.83) shows that Proposition 3.4 can be used for branches
having nonlinear elements. In such cases, the relationships All (ai) r2mx+kizi
(3.87) read
(3.94) =+
so that (3.95) zi Next enforce the relation
'\z( z) 1 = mx + Zmx mx i O1 m2z2J 1+ (3.96)
J_x
kiz1,
=i(,
\2)1(E2)
mx
+r
kiziJ
(3.97)
67
Control
i)
[(x)'(ei)[_kef
ul] rT11]
m
x]
2m+
m+
T1 i
kl i11
(3.98)
Control
used the bond graph to define the plant and the closed loop (xi) for j>0. In
target dynamics, which comprised the explicit bond graph models Ej: i particular, the characteristic
change of coordinates to induce a closed loop bond graph identical to the plant with additional elements. It can be shown that some of the design steps can be performed through the sole application of bicausality [GawOl]. More precisely, instead of manually going through the
various changes of variables, bicausal assignment can be used to immediately backstepping controller without the explicit introduction of new variables.
derive the
Most of the material presented will use the examples of Section 3.2.1 and the various results of Section 3.2.2, thereby facilitating noted that in [Gaw0l] tual Actuators" comparisons with the bicausal approach. It should be with Vir
In [Gaw0l], it is shown that bond graph based backstepping is possible through the (bi)causal inversion mechanism, but the paper does not address these ideas in great detail other than through the motivational therefore to complement Example 3.2 in Section 3.2.1. The main impetus of this section is [Gaw0l] by showing that the bicausal approach does indeed yield problems. Furthermore, it will be shown
imposed, where the C and I elements are assumed to be quadratic. structures these ideas.
68
3. Backstepping
Il
21
C1
jI
...
asks
I1
al 11
0 a2a
Ir
ii
Yo
', /
20:
Ho(x0)
Y77 u, 7z r,
u17 y77I
a33 U """
1 a2
y0
Y'I unZ
E712 : Hg2(n2)
Figure 3.27.: Bicausal extended cascaded C and I pattern of Proposition Proposition 3.9.
Consider the bond graph in Figure 3.26 and the bicausal bond graph in (3.54) and that
Figure 3.27. Suppose that the cascaded elements have the definitions
(Ti, Yni U71.+ uni 11 (7 yr,, uni = 1 +y (T?, u, ) i
(3.99)
Then the bicausal bond graph yields a smooth feedback u(x, , rte) that induces the closed loop dynamics of Figure 3.28.
Proof. Consider the change of variable aiel = yo + alzl = alvo + alzl, giving the z1dynamics zi = yo  y7'(n', aii) + a26  v * i? + y,, a22 y, aiv0*) _ yo l(7)1, aizi) i(7)l,  0. Suppose that the original portHamiltonian structure is to be retained and that no stabilising control a22 then becomes (3.102) (3.101) (3.100)
69
3. Backstepping
Control E: H(w)
I1
i2
ci
1z1
0
u77 2
uo
yo
yo
yn2
u17 y771
Ent
H712
('/)2,
(17 E271 : H, 71
with Ci bicausal approach of Proposition 3.9.
Figure
3.28.:
0
ai+1ei+1
'Y7?
a2vZ 1
ailvi2
7)
Recall that alv = Jo and consider the bicausal bond graph in Figure 3.27. Bicausal analysis then shows that the bond signal a22 coincides with (3.102) for z2 = 0. Thus, for a onestep design it follows that the bicausal approach coincides with the backstepping troller. Continuing (3.102) with yields the z2dynamics
(712, + vl y, a2e2) a3e3 72
con
z2 = aiei
(3.103)
= alzi therefore
a3 3=
(3.104)
Set z3 =0 and observe that bicausal analysis yields the virtual control (3.104). The recursive process continues for j>2
zj = ajlj1
as
+ a3e3 vj1
(,  yrli / , ajj)
= ajizji
(r, ynj
ajzj)  ajivy2
The virtual controls are of the form aj+lej+l W, vji + + y,, aj1v2 aivv1) = j + aj+izj+i. (3.106)
70
3. Backstepping
Control
SS1 U0
Now, the bicausal bond graph does not introduce new variables, so that it follows that (3.106) must impose zj+l =0 if the virtual controls aj+ij+l are to coincide with the bicausal
mechanism. Further causal analysis of Figure 3.27 yields the generic C2 backstepping junctions depicted in Figure 3.29, where it should be noted that the indexing does not conform to (3.60) but merely reflects the relationships the 0junctions. Finally, between the various bond signals incident to
Of course, the generic I2 has the same bond signal relationships. control ai+lj+l is now seen to coincide with (3.106) for zj+l = 0, but
the virtual
is must be noted that the obtained closed loop portHamiltonian bicausal approach still require the variables zj for their definition. It can be concluded that the exposition in [Gaw0l]
backstepping controllers; however, the bicausal bond graph does not introduce new variables to define the dynamics Ej: Hj(x3), thereby restricting the class of systems. That is to say but does not provide tools
that the bicausal inversion mechanism retains the plant structure to define further stabilising dynamics without Example 3.8. The controller
(3.16) of Example 3.2 is obtained by means of the bicausal To see this, observe that y7(,q, u,, ) = u,, /r and therefore the
(3.99) is satisfied. Bicausal analysis then yields the control 1 ryo+ 1. kyo, (3.107)
where Proposition
3.9 confirms the closed loop in Figure 3.8. It can be concluded that the for the backstepping design but for which the 0
bicausal bond graph provides a "shortcut" closed loop bond graph representation
71
3. Backstepping
Control
C2
i
21
SS, 1L*011
V11
i
klv*
01 71
Figure 3.32.: Bicausal backstepping towards ul; Example 3.9. In [Gaw0l], only the singleinput case is considered, however, the multiinput system (3.83)
shows that bicausality can be applied for such multiinput complements [Gaw0l] by applying Proposition
scenarios. The following example case and shows that the original
3.9 to a multiinput
the closed loop does not comprise any stabilising structure Example of the system.
3.9. Consider, once again, the system in Figure 3.24 and the associated bicausal and
bond graphs in Figures 3.31 and 3.32. Clearly, u2 must be found first since 2=1. The control u2 is immediate from Figure 3.31, being
U2 = '602,
(3.108)
(3.90) for coincides which with z1 = 0. Then, from Figure 3.32, the control ul becomes
ul = klvol
so that ul coincides with (3.92) for z2 = 0.
vil1 
(3.109)
3.4. Conclusions
This chapter contributes certain results on bond graph based backstepping control that were inspired by the works [Yeh99], [YehOl], [Yeh02], [GawOl] and references therein. The novelty of the presented results are the applications of judiciously for closed loop bond graph representations. chosen virtual control laws to allow
72
3. Backstepping
Control
Crucial to all developments is that backstepping control is capable of inducing bond graph based dynamics provided that the virtually actuated subsystem has certain portHamiltonian proper
properties. More precisely, the fact that bond graphs have intrinsic portHamiltonian
ties [Go102] clearly shows the connections between backstepping and bond graph modelling. This observation readily materialises by taking the virtually actuated system as a bond
design can be made to retain the original portHamiltonian which to "step back". These ideas show that such an
and effective; however, as with all backstepping designs, new the virtual control and regular control have to be introwith respect to the states of the because of
duced. Consequently, the closed loop is portHamiltonian virtually actuated subsystem and the new coordinates. actuated trajectories
The bond graph based backstepping method shows to allow for certain disturbances to enter the virtually actuated subsystem, provided a relative degree condition is fulfilled. If such
relative degree conditions are not satisfied, then the backstepping controller may depend on the disturbance and its time derivatives. This would be problematic since disturbances are
generally assumed not to be measurable. Multiinput systems show to be solvable by means of the singleinput case, but no further sysit
results have been elaborated since the singleinput tems. Even though such extensions to multiinput
should be observed that the regular controls are likely to have different relative degrees with respect to the virtual controls. This implies that regular controls with the lowest relative
degrees have to be resolved first. Finally, the bicausal approach does indeed yield a class of backstepping controllers, but the approach only works on a smaller class of systems since new variables are not part of the controller design. Consequently, stabilising dynamics remain somewhat difficult case, multiinput to define
from within
backstepping
73
4.1. Introduction
The (non)linear
Model Matching
Problem
(MMP)
(2.31) affine plant and model (2.32). However, it can be of interest to specialise these affine systems to certain physical models, where both P and M describe physical behavior such that the MMP differently, incorporates a form of "physical equivalence" as outlined in way that attaches explicit in [Sha9l]. Put
the input/output
method would represent the physical design objective instead of a sole signal theoretic approach.
This section presents applications of the bond graph based MMP for which the control objective has a physical interpretation. here is as follows. The main ingredients of a typical application considered
Then, once P and M are defined, the natural passive outputs of the bond graphs are often not to be controlled as mentioned in Section 2.3.1. Consequently, redefinition of such passive
outputs may be required, where the model M will be defined to contain a copy of the plant and be "close" to the plant in structural condition to be satisfied. The key aspects of this chapter are not about strict design steps such as bond graph based backstepping. This chapter shows that the closed loop bond graph representation may be sense. This will allow a necessary relative degree
inferred, in some cases, from the plant in accordance with the MMP theory presented in Section 2.3. Furthermore, controller through bicausal bond graphs are used whenever possible to find the decoupling
matching of the plant and model outputs must be found in a conventional manner.
74
4. Model Matching
Control
There may exist certain MMP scenarios that need not be efficiently solved through a bicausal inversion mechanism. Indeed, the problem that could hinder such application of bicausality
is that the Standard Causality Assignment Procedure (SCAP) does not efficiently model the plant bond graph, or that plant outputs are difficult to define through SS components. More
precisely, it is known that SCAP can yield unnecessary complex dynamics due to the "inefficient" selection of state variables. To remedy such problems, alternative causality assignment be can used to simplify procedures the dynamics significantly [Mar02]. Causality For example, in the Assignment Proce
Lagrangian domain it is the that the often case mechanical dure (LCAP)
dynamics, to efficient obtain second order a more method which can provides into first order form. However, LCAP is difficult to use in the bi
be readily transformed
are causally assigned with SCAP. Also, there are no available results on this matter in the current literature. This chapter will address the above considerations in more detail.
Even though the bicausal bond graph mechanism will be used for inversion purposes whenever possible, certain outputs may not be readily modelled with SS components such that the This can occur when outputs appear nonlinearly As a result, the SS component bicausal
in the model or when they are functions of state variables. is not a suitable solution for extracting the output
inversion not applicable due to the absence of suitable SS components that define the required output. These issues will be elaborated in later sections. often restrict the applicability of feedback
domain of interest. To render the MMP solvable, it is quite natural to address the linearised MMP instead. The bond graphs for P and M do not change, but their induced dynamics is linearised around some point of interest. It will be shown that the closed loop bond graph is then merely associated with input/output dynamics of the prescribed model.
It must be acknowledged that the MMP considered here is nothing new in itself, but the specific application novel contribution. regulation of physical considerations through bond graph modelling can be seen as a Furthermore, and this holds for all bond graph based MMPs, the output foundations that
problem addressed in Section 2.3.2 provides control theoretical in the bond graph literature.
75
SS vi 2V SS \
y2 Figure 4.1.: Plant input/output
SS Ep: H(x) Ul
U2
SS
configuration.
4.2. Virtual
Actuation
of Input/Output
Dynamics
actuation
in input/output
The input/output
controlled
dynamics. Such control objective can be specified with a model that comprises an exact copy of the plant, where additive dynamics represents the external dynamics. The plant is a bond graph model with a subset of passive outputs ignored. standard bond graph model. Consider the plant P in Figure 4.1, where u= input and output variables have the dimensions
1 il ... 'v 1 1), 2= 11 (v 2'... 'v 2 2)
(ul, u2), y=
uP, ), ul),
y1 = (yi,..., y2 = (yi,...,
y2 yP2)
(4.1)
The superscript
and dissipative elements are contained in Ep. Next consider the model M depicted in Figure 4.2, where zi = (ii1, ice) and dimensions of the model inputs and outputs are 1 = (i, ... 1 1), 2 = (i, ... The model Hamiltonian p2), y1 = (yi' ... , YP1) (yi y2 = '... , y2). function fl: X Bond R. * (yl, 92). The
(4.2)
76
4. Model Matching
Control
_ Em : H(x)
u2
The bond graph based MMP objective considered in this section can be clarified as follows. Suppose that the plant P has dynamics of the form th _ yj = [J(x, Ii) R(x,. u)]K(x, 9jT(x, p)K(x, !z), u)  9i(x, lu)uj4i(x, fi)v= (4.3)
where xE 1R" are the states associated with the storage elements, and where the system inputs and outputs are u, yE 1f and vE R'. Physical parameters of the plant are = (p1, /tk) ... ,
x2
sT (x2)
J(jZ)  R(.t2)
k(j2)
(4.4)
by the x2dynamics
of the model. To implement the idea that the (u, y)dynamics as a partial
solution to the nonregular DDDP with disturbance measurement. is derivable from explicit portHamiltonian
Figure 4.1. So, for the following developments, it is assumed that the bond graph of P yields explicit portHamiltonian dynamics that take the form (4.3) when the conjugate outputs to obtained from bond graphs that have no
v are ignored. Such explicit systems are typically storage elements in derivative causality.
By definition of P and M it follows that the relative degree condition rZ < rt is fulfilled. solvability
The
of the MMP considered here now depends on whether the bicausal bond graph in independent of U. be can
Figure 4.3 exists and that it = is is fulfilled to render the difference y9
The following example shows the basic concept on how the bond graph based MIM defined to have virtual actuation in an input/output sense.
77
4. Model Matching
Control y
1
SS
SS
4.1. Consider the linear plant P, with all parameters unity, of the form X1 0 1 0 x1 10
x2 x3
=10 010
y=Xl,
1
XZ X3
0U0v 01
(4.5)
xl
X2 X3
X4
0 1
1 0
d
0 1
a 0
X1 x2 x3 bx4 
1 0 0 0
0 100 000
(4.6)
y=
xi, It is seen that M contains an exact copy of the plant actuated by the x4dynamics. decoupling problem, the difference xl be is to now J1
where a and b are model parameters. and that the xldynamics is virtually
To independent that end, consider the plant inverse of (4.5) given by U. rendered of v= where dry/dtr relation u= 2y('  y(3) uu(2) + w, By enforcing y=9 (4.7) it is found that the
is is indeed required to attain decoupling, giving the control v= c(:t, ) +w= abt4 + alb(x2 ) + abt4 w. 
(4.8)
Z* _ {(x, x) :x
cp(i) =
X2 2g
X2 + abi4 x3 a 2bil
0}. =
(4.9)
78
(4.10)
Fel
e2 =10
0 1
010
0
1
Fe1l
e2 e3 0w,
0
(4.11) 1
Le3
and observe that this system has the structure is now achieved through the passive feedback u= for some damping constant r>0. The general setup of the MMP
(4.12)
0
considered in this section is to consider bond graphs that
induce systems of the form (4.3), after which the plant bond graph is copied and extended with additional storage and dissipative elements to define the MMP objective. To solve the
MMP then requires that the bicausal inverse exists and that it is well defined. If the causal inverse exists, then the virtual relation y=y is substituted actuation should be achieved when u= is is imposed and the
into the plant inverse. is the first step in all bond graph based MMPs in of the constrained dynamics algo
rithm, which is often sufficient in finding the map cp(x), or cp(, q, x) in case of internal dynamics rj. These maps are used for the feedback control with the structure of (2.66) to stabilise
Z* if possible. It is interesting to note that the bicausal bond graph is not a requirement in the bond graph based MMP, because the constrained dynamics algorithm yields Z*, which, in turn, leads to the decoupling control. The following example extends the basic Example 4.1 by using bond graph modelling arguments to define the MMP. Example 4.2. Consider the simple massspring system in Figure 4.4 and its associated
Note that all input variables u, v and is denote forces and that all
79
Figure 4.5.: Simple massspring model of Example 4.2. The bond graphs of both system are depicted in Figure 4.6 and Figure 4.7. From causal analysis it immediately XI x2 =10 follows that the plant P has the form 0 1 0 1 xl/ml kx2 0 10 u0v
(4.13) 01
3
y_
010
xi/mi
x3/m2
and where the model M is described by :t1 0 1 0 1 00 QQ 1 0 1/ml kJV 20  13/m2 kx4 0 0 (4.14) 1
X2 x3 X4
10 01 10
9= 1mi
physical interpretation:
Feedback
on v should impose convergence of the plant and model outputs, thus Iy(t)  y(t)l j 0, and the influence of model inputs on the extended output yy Since the relative degree condition r<f in Figure 4.8 yields the inverse plant v= (ml + m2)y(l) u k2 u(2) m2 y(3). (4.15) is removed by the relation u=U. the bicausal bond graph
is fulfilled by construction,
80
4. Model Matching \ Ii C
Control I2
xl
x3
SS1
I
/
17
SS2
s
j
I7
01
and u=,
m2k21
yielding
m2k
v=
c(, ) +w=
m,iii
X2 +k
km
/)
X4
km l
+ w.
(4.16)
As mentioned in Section 2.3.1 on model inversion, the bicausal bond graph does not yield the output matching submanifold an important Z*, so that the constrained dynamics algorithm remains to be
tool in the search for this submanifold for all MMP problems. The constrained
XI
ZxcpGx) = x2
X3
J2+ Ix4
m2k X3 mlkxl _
_ 0})
(4.17)
as the
portHamiltonian
system
ei e2 e3 =10 01
= e3/m2"
1
0 1 0
(4.18)
1
81
Ii
I2
T
1l /1
x2
x3
SS1
NI
01)1r)
SSZ
Figure 4.8.: Bicausal inversion of the plant of Example 4.2. Observe that this systems has the bond graph topology of Figure 4.6 for which SS1 is removed. Passive stabilisation of the origin e=0 is possible by terminating re3 for some r>0. the SS2 with a linear Q
In [Vin03], the bond graph based MMP and its virtual but no clear answer is given on how the attractivity
actuation interpretation
is presented The
answer to this question is readily found by the various considerations ing that no bond graph operations exists that or x= "automatically"
), and considerable analysis remains necessary regarding the description cp(r, of Z*. case is not intrinsically different from a singleinput
stabilisation
The multiinput
ing example taken from [Vin03] presents such a multiinput the considerations of Section 2.3.2 on the Full Information
Problem are
not mentioned in that paper. Example 4.3. Consider the mechanical system depicted in Figure 4.9 and its associated
forces The Figure 4.10. in inputs are plant and u2 applied to the masses ml and ul model m2 respectively, and where the controls v are velocities. By passivity, the plant outputs
Figure 4.10 incorporates a copy of the plant and adds further dynamics through a nonlinear spring characteristic k(x) to be defined later. The bond graphs of the plant P and model M
are depicted in Figure 4.11 and 4.12. The control v imposes the desired dynamics whereas the inputs are known and the condition u= should render the difference yy in an input/output independent of . In particular, to implement virtual
actuation
82
4. Model Matching
ki u1i yl
Control
'u2, Y2 2
e vl r
ml
Q v2
m2
r
Figure 4.10.: Model with nonlinear spring element of Example 4.3. The design proceeds with the causal analysis of the plant bond graph in Figure 4.11, giving the dynamics
X1
2 : th3 X4 Yi
Y2
0010
00 1 0 1 1 1 r 00 1 0
klxl
k2x2 x3/ml x4/m2
0010
00 10 0100 ui U2 01 r 0 vl V2
x3/m1
L
X2
X3 =
x4/m2
(4.19)
The model bond graph in Figure 4.12 then induces the system 00100 00
1 1
k1 100 1
r
10
00
k2X2
x3/m1 10
00
i
R2
X4
1
00
1
4/m2
01
(4.20)
X5
yi 92
00010
x3/m1 x4/% 2
sinh(J5)
00
To solve the disturbance decoupling problem, observe that the relative degree condition rz < ri is satisfied and that the bicausal bond graph in Figure 4.13 shows no causal conflicts.
83
I1
c2
x2 X4
I2
1o
x1 V1 u1
1k
y1
o
11)2 U2 Y2
cl
ss3
ss1
Ss4
552 .
12
0
x4
11k
C3
U1 1
ii2
92
C1
SS1
SS2
Figure 4.12.: Model bond graph of Example 4.3. The inverse is readily found to be r _ vi
V2

(mlyil)
+ m2y21) + kii
+ ui + u2) (4.21)
=
=
yi +
y2yl+
(miyil)
2y22'+
where wl and w2 are new controls for stabilisation relations y=y and u= is yield the control TJ =qr v1
V2
kl
ki 7Ti rrr
=
=1
kl
(4.22)
Even though the bicausal approach gives the decoupling controller ner, the submanifold
in straightforward
man
Z* is needed to derive the feedback of the form u= zi) + K[x  cp(rj,x)]. c(Ti, d, (4.23)
84
4. Model Matching
Control
Ii
x3
C2
T2
I2
4
T
k1 L o
IV1
1k
U Y1
o
1v2 u2
1
Y2
ILIth1
C1
Ss3
ssl
SS4
SS2
Figure 4.13.: Bicausal plant inversion of Example 4.3. The constrained dynamics algorithm now yields
x1
z*_{(77,x,x): xw(77, )=
X2
X3 X4
t2
+
2
sinh(X5)
x3 x4
=0}
(4.24)
It is important
(locally)
controlled
invariant
submanifold in the usual sense, because it has been "extended" for convenience. Next define the error e=xsystem of the form
e1
e2 e3
0010
00 1 1 =1 r 1 0
k1e1
k2e2 e3/m1
10
01 r 0 W1 W2
e4
01
0
klel
1
00
e4/m2
r01[ W1
00
(4.25)
+ re3/m1
02
k2e2
00
W2
Observe that the closed loop has the bond graph topology of Figure 4.14: The edynamics allows for a bond graph topology that is identical to the plant with u=0. closed loop passive outputs Furthermore, the
but can be
readily included from the bond graph in Figure 4.14. The submanifold to be attractive for w=0 due to the dissipative element of the plant. K[x  cp(rl,x)] for some suitable gain K.
of Z* is possible through w=
85
4. Model Matching
Control
Ii
T3
C2
C2 eq
12
100
el
O1
wl
'tI2
w2
C1
S53
SS4
Figure 4.14.: Induced closed loop bond graph of Example 4.3. So far it has been shown that the physical model based MMP can be defined in a manner that can be represented through bond graphs. To that end, the system is assumed to be of the form (4.3), which represents an explicit portHamiltonian system of which a subset of portHamiltonian More precisely, can of the bond graphs in
passive outputs are ignored. The question rises whether a class of implicit systems can be used in the MMP scenario considered in this section. configuration
Figures 4.1 and 4.2 for which P and M have derivative causalities? It is intuitively plausible that bond graph models with derivative causalities can "in principle"
be used in the same MMP setup as depicted in Figure 4.1 and 4.2. That is to say that the bond graphs in Figure 4.1 and 4.2 merely depict a certain input/output configuration but
do not show the bond graph topology itself, which may or may not have dependent storage elements. However, the presence of derivative causalities poses additional the associated implicit dynamics of the bond graph. Furthermore, will be more difficult difficulties in regard
to derive.
In view of the above considerations, the MMP scenario of this section requires further research on the topic of derivative causalities. It is expected that the scenario can be applied to
causalities provided the causal inverse exists and that However, the closed loop error dynamics can These issues will not be
are satisfied.
elaborated any further and a more general bond graph based MMP with dependent storage elements will considered in later sections.
86
4. Allodel Matching
Control yl SS
Ep: H(x)
SS
ul uZ y2
pairs.
EPl : H(x)
U SS y2
4.3. Specification
Based Inversion
based inversion"
is introduced.
This bond
graph approach is very closely related to the bond graph based IMP theory considered in this chapter.
design but addresses a (bi)causal inversion process for which certain aspects can be extended to controller design in terms of the MMP. Specification based inversion can be outlined briefly by considering the collocated plant bond graph and its causal inverse in Figures 4.15 and 4.16. The model is equipped with a copy of the plant and has the same input/output be depicted. configurations, so its bond graph need not
Causal plant inversion is used to find the control that achieves the required dynamics by imposing the constraint y=y, provided the relative degree
model input/output
Now, for more generality, the MMP scenario in this section need not stay with the collocated [Ngw99b], based because the noncollocated inversion in case of specification case is concep
tually identical to the collocated case. More precisely, the model is equipped with an exact copy of the plant and has the same input/output the bond graphs of the noncollocated By imposing the constraint y=g configuration. Figure 4.17 and 4.18 depict (u', u2) and y= (y1, y2).
that solves the MMP, provided relative degree conditions on model inputs are satisfied.
87
Figure 4.17.: Plant bond graph with noncollocated SS ul SS uZ EP1: H(x) 0 y2 Figure 4.18.: Bicausal bond graph with noncollocated To show a possible class of dynamics suitable for specification i yl 0
71
/71
SS
input/output
pairs.
plant P described by
i= wj yj where xE R' = =q [J(x, fi) (x, R(x, JL)]K(x, )  9i u)uj9jT(x, M)K(x, ), (x, fi)K(x, lR'. li), The ktuple (x, qi )v (4.26)
It = (ti, ... , Ick) denotes the physical parameters of the plant. Then to follow the noncollocated scenario based on Figure 4.17, for and u, y, v, wE example, let the outputs w be ignored and let v=0.
x1 J(: i1, ji) R(1, ) S(, 2)
ST (2)
J(2)  R(x2)
K(z)
0
(4.27)
fu) denotes the prescribed ... , plant parameters. Lemma 2.4 can, but need not, be invoked in this particular case. The ptuple = The following example is taken from [Ngw99b] and shows the bond graph based MMP of a
(l,
Suppose the model has been chosen to comprise an exact copy of the plant only, implying that the control objective is solely based on parametric modifications. As a result, the bond
graph topology of the model is identical to the plant and need not be depicted here.
88
R2
C2
T
SS1 111/0
11
//
00
/1
SS2
of Example 4.4. C2 R2
T2
SS1
1A0
00
SS2
of Example 4.4.
11
rl
1
r2

r2
_1r2
1 xl Cl
1x2
c2
1 rl
p
u
x2
1
r2
(4.28)
1 y=X2, C2 (4.26). form The passive output w= is the of which Suppose the model is chosen with the parameters 11
T1 _ r2
xi/(rlcl)
1 r2
1_ xl cl
1
_ rl ,u
1
T2
1
r2
1
C2
X2
x2
(4.29)
1
C2
(4.27) to with x2 = 0. which conforms Clearly, the plant is not dynamically extended through the model M and the control objective
is solely expressed in terms of parametric modifications. The relative degree condition r<r has no causal conflict.
u=
is trivially
89
purposes.
0}.
(4.31)
C2
Ccirl
C2r2
+ C2 C2f2
C2rl
ClC2rlT2 C22
C2T2 C2T2
clc2rlr2\ C1C2f2 +
(4.30) (4.32) As by substituting is is identical to a new control. expected, where w Stability of Z* is addressed by defining the error e=x1111 ei
11 e2 rl r2 r2 r2 r2 Cl l0 C2 e2
y=y.
1
rl w, (4.33)
for w=0.
Further stabilisation
K(x  cp(x)) with some suitable gain K. Note that the closed loop has the bond graph
of [Ngw99b] can be extended to an MMP design. Even though the methodology presented through a explicit this MIM
SISO system, it is readily argued that for VIIMO bond graphs possible provided no causal conflicts occur. However, as of dependent storage elements renders the design of of the closed form for
scenario is conceptually
difficult. bond thereby graph considerations analysis, rendering the MIMO and implicit case in greater detail.
90
4. Model Matching
Control
Matching
for the
MMP, where it was shown that for certain MMP setups it is possible to use the bicausal approach to find the control that renders the difference yg independent of U. Furthermore, it by
was seen that the closed loop error dynamics allowed for a bond graph representation means of output regulation arguments such that closed loop stabilisation passivity arguments.
could be based on
Even though the previously presented MMP scenarios represent a relevant class of systems, it can be argued that their applicability configuration. can be limited due to the specialised input/output of the For
MMP, a larger class of systems that bond graphs can generate should be considered.
example, bond graph models with derivative causalities can be considered, leading to a large class of implicit systems. Also, less specialised input/output configurations of the plant P
and model M can further enlarge the class of MMPs considerably. Because derivative causalities often occur in a wide variety of bond graph models, it can be argued that such systems represent an important set of dynamic systems. Therefore, this
section addresses bond graph models that are allowed to have storage element in derivative causality. However, in view of Theorem 2.6, the implicit systems context renders the
closed loop bond graph representations of the error dynamics of limited interest. This can be explained by the fact that implicit dynamics must be reduced to explicit dynamics. Nonethe
less, even though the closed loop bond graph representation of the error dynamics may not be available, the closed loop input/output by construction. As already mentioned, the following sections address a more general bond graph based MMP with no specific input/output configuration. In addition, the model IVI need not contain a "close" to the plant in order of bicausal bond with dynamics remain to have a bond graph representation
copy of the plant, but the model will be chosen to be structurally to satisfy the relative degree condition ri < rj. Furthermore,
the application
graphs will not be considered for system inversion purposes due to various difficulties respect to nonstandard output definitions. Also, the application of alternative
causal assign
ment procedures to deal with derivative causalities renders the application mechanism rather difficult.
(bi)causal the of
91
Control
by the following: First, the Standard Causality Assignment Procedure (SCAP) may yield unnecessary complex dynamics due to the "automated" is possible to consider alternative Causality Assignment selection of state variables. To remedy such problems, it causality assignment procedures, such as the Lagrangian denoted as ALCAP [Mar02]. However, it
should be noted that (bi)causal bond graphs have not been developed for such alternative causal assignment procedures. Second, it must be recognised that general bond graph models need not be feedback linearisable on the domain of interest. Therefore, if feedback linearisability is not suitable or instead. This
applicable for the system at hand, one could consider the linearised MMP scenario will be demonstrated Third, if nonstandard in detail later on.
input/output
configurations are considered, the use of SS components "contrived" bond graph mainly
to extract the desired output variable need not be possible without modelling.
because the bond graph does not define the output variable through the SS source component. As a result, such nonstandard output definitions should not be modelled with bond graphs
4.4.1.
A Class of Implicit
Systems
portHamiltonian
p) R(x', it)
plant P of the form 0 00 9(xl, p) q(xl, Iz b(xl, bT (xl, ) 0A ) li) K' (x 1, /j,) K2 (X2, p) u (4.34)
x2 y 0
0 9T(xl, IL)
00 I112
4T(x1, IL)
AE R"2 and u, yE Rm. Let tC be a vector of physical parameters. where x1 E IR711, x2 E 1[n2, This type of system is typically obtained from a bond graph in which each dependent storage directly at the depenin
element is forced into integral causality through a Lagrange multiplier dent storage element. Note, in particular,
92
for feedback for for be bicausal bond which, example, which graphs can used and scenarios linearisation is feasible. The approach taken here is to consider procedures similar to ALCAP The model will be defined to be structurally are fulfilled. In case feedback
linearisation is not feasible, it will be shown that the linearised MMP can be attempted, where the bond graph definition of the plant and model remain unmodified. show instructive The following examples
(4.34). form the systems of Example 4.5. Consider a pendulum mounted on a horizontally be given by (x } x4) + mgl sin(xl). (4.35) decart moving, massless
The moment of inertia about its centre of mass G is denoted as I, the mass of the pendulum is in, the distance from the hinge to G is denoted as 1. The variable xl =0 angular position counter clockwise from the horizontal, x4 are the respective vertical and horizontal u is a velocity. tl(xl) =I cos(xi) denotes the
momenta of G.
grange multipliers
to be "workless", implying that their conjugate velocities are to be nullified accordingly. Next observe that the plant bond graph does not show a simple SS element that extracts the pendulum angle xi as a system output. In view of (4.35), the reason for doing so is that in xl, so that the angle xl cannot be
by bicausal inclusion Consequently, the SS the approach element. of such simply extracted is not readily applicable for this system due to the absence of this SS element. This is not believed to be a great problem, since the MIM is solved analytically once the plant and
defined, been have the to desired the where plant and model simply added are outputs model definitions once the causal analysis is completed. So, bicausal assignment can add a certain bond but it is design, to the systematic not a crucial aspect and graph convenience approach based MNIP the model physical considered here. of
93
4. Model Matching it
Control
Figure 4.21.: Inverted pendulum From Figure 4.22 it is a straightforward resentation X1 42 0 1 1 r 0 0 0 0 0 0 0
of Example 4.5.
portHamiltonian
rep
0 l sin(xi)
l cos(xi)
x3 X4
y
0 = 0
0
0 0
0
0 0
0
0 0
0
0 0
0
1 0
0
0 1
1
0 0
0 0
l cos(xl) 1 sin(xi)
1 0
0 1
0 1
0 0
0 0
Al A2
(4.36)
coefficient of the hinge, and where Al and A2 are Lagrange multipliers manifold defined by the last two
equations of (4.36). Clearly, the above system is of the form (4.34). The implicit the multipliers representation of the plant can be turned into an explicit system by eliminating plant
xZ Y= xi,
I+
ml2x2
I+
ml2
(g cos(x1) + sin(x1)A)
(4.37)
sys
tem (4.36) is readily reducible to the explicit system (4.37), but it cannot be guaranteed that the reduction of the implicit (4.34) system yields an explicit portHamiltonian model.
94
4. Model Matching
Control
I1
Iz
x2
xl
1k11
TF ti(xi)
I/
SS2
TF
00
u SS1 13
SS3
t2(Xl)
pendulum at the angle x1 = it/2, which is the upright position. that the pendulum on the moving cart is structurally
This observation is entirely subjective but suggests a suitable that can be associated with the moving pendulum.
The bond graph topology of the simple pendulum is depicted in Figure 4.23 and is seen to be similar to the plant, where the modulations t1) =l cos(J1) and t2(t1) = l sin(: t1). are tl (. angle :f1 does not
No SS element is used to extract an output variable since the pendulum appear conveniently in the model. Causal analysis of Figure 4.23 yields the implicit portHamiltonian
system
X1
x2 X3 X4 9 =
0
1 0 0 0
10000
r 00001 00000 10000 00 1
0
l t1) cos(x1) sin(. l 0 1 1
Ifl(t)
K2() K3(x) K4(x) , (4.38)
0 0
0 0
000 1
X
0 00 0 a2
95
4. Model Matching
Control
I1
C
II
I2
'
R 
11
k1I/
TF tl(x)
0I

SS2
SS1
TF  i2(x)
x4
9S3
I3
Figure 4.23.: Simple pendulum model bond graph with \multipliers , The upright choosing k> an attractive equilibrium position x= (ir/2,0,0,0)
of Example 4.5. by
mgl. This will guarantee the the simple pendulum has the upright equilibrium point.
position as
X2 y=
I+ 1.
ml2
x2
I+
ml2
+ k(x1 2
n) + ii)
(4.40)
Before the MMP can be addressed, observe that the absence of a relative degree for sin(xi) is an important characteristic of the plant and implies that all control authority
=0
is lost when
is possible only for points bounded away from xi = nir, which implies that it is possible to asymptotically track the angle of the simple pendulum around the upright position. to observe from properties
On a side note, the absence of a relative degree is something that is difficult general bond graphs of arbitrary
standard control theory should be applied to establish the relative degree property.
96
Control
decoupling problem simple, consider the linearising control 1u [rx2 + gmll cos(xi)] 2 (4.41)
I in(xl)
Ixt
w.
(4.42)
The submanifold
Z* = {(x, i):
Lx2]
0}. =
(4.43)
Then take kl, k2 >0 and define e=xw=c(; t, )+K[xco(w)] rI+ m12 X2 I+
with
(4.44) k2e2. 
Some observations on behalf of the above example can be made. For instance, the inverted pendulum shows that the absence of a welldefined relative degree need not compromise the
MMP objective, provided that the relative degree exists on the domain of interest. However, it is certainly conceivable that the structure of the system does not sustain a welldefined cannot be addressed in a manner
that has been portrayed so far. The upcoming example shows that the linearised MMP can be considered if the nonlinear plant dynamics are not suitable for feedback linearisation. Furthermore, resentation multiplier the implicit of the inverted portHamiltonian pendulum dynamics (4.34) shows to be a suitable rep
of the Lagrange
is readily possible. On the other hand, the upcoming example shows that such im
plicit dynamics can yield unnecessary complex dynamics and that the Lagrangian causality assignment procedure LCAP [Mar02] is preferable instead. However, in view of such alter
native causality procedures, the bicausal bond graph mechanism is of limited interest since it is mainly applied to bond graphs that are causally completed with SCAP. In fact, bicausal assignment has not yet been reported for LCAP procedures. or other alternative causality assignment
97
Control
4.6. Consider a sliding mass along a straight, slender rod that pivots around its centre of mass depicted in Figure 4.24. The control input u is a torque around
its centre of mass. Let the plant energy function be defined as H(x) where xl =0 and x3 = r. = mgx3 sin(xi) + (x4 + + x5), x2 2I 2m (4.45)
linear momenta, denoted as x4 and x5i are in the x and y directions respectively. Elementary kinematic analysis based on x= X3 cos(xi) and y= X3 sin(xi) explains the modulations in
(4.46)
t3(x1) = x3 sin(xl), t4(xi) = COS(x1).
portHamiltonian
system
XI
x2 x3 X4 X5 =
0
1 0 0 0
1
0 t3/t4 0 0
00000
t3/t4 0 1/t4 00001 00 1/t4 000 000 tl/t4 1 (t1t3 t2t4)/t4 
K, (x)
K2 (T) 1f3(X) K4(x) K5(X)
y
0
0
0 (t2t4
1
 tit3)/t4
00000
0 tl/t4 1 00
u
A
(4.47) The Amultiplier be "workless". systematically, has a nullified conjugate velocity, imposing the requirement that it is to form
the structure of (4.47) yields a model of unnecessary complexity: the horizontal of the system is better suited
momentum x4 is a state variable while the physical structure for polar coordinates. The occasions where implicit portHamiltonian
dynamics (4.34) yield complex reduced dyALCAP instead, where multipliers are
namics can sometimes be remedied by considering introduced if one is not able to find a minimal
[Mar02].
Figure 4.26, where the source elements SS1 and SS2 define the generalised coordinates. It is important to note that ALCAP considered here reduces to LCAP since a minimal set
of generalised coordinates has been found, which are xl =B multipliers are not needed.
98
4. Model Matching y
Control
x5
C2
TF
tl(Xi)
0A
SS2
TF t4(xl)
II
0 \I
TF t2(xl)
I2
TF  t3(xl)
1I7
SS1
C i
X2
Ii
Figure 4.25.: Frictionless slider with Amultiplier The modulations in Figure 4.26 are defined as ta(xi) t3(xl) = sin(xi), = x2 sin(xl), t2(xi)
of Example 4.6.
= x2cos(x1),
(4.48)
99
SS2 0 C2 xz
11
TF ti (x, )
13
TF t4(xl)
TF t2(x1)
I2
l0

TF \ta(xi)
I1
SS1
Cl \I1

11
Figure 4.26.: Frictionless slider with LCAP of Example 4.6. By applying LCAP, the equations of motions are obtained through the summation at the onejunctions of efforts
to which SS1 and SS2 are connected. This yields the dynamics xl x2 = (2mx2x thl + mgx2 cos(xl) + u)
I+
mx2a
(4.49)
x2xi g sin(xl).
By defining zi = xi, z2 =& 1) z3 = x2 and z4 =d 2i the plant takes the first order form
Z1 = Z2
z2 z3
z4
=1 =
=
I +mz3 Z4
z3z2 g
sin(zl)
=Z3, z3 has been added to incorporate the requirement that the distance of On a side note, observe that
the slider with respect to the hinge point is to be controlled. it is by no means trivial system (4.50). how y=
100
of
y(2) y(3)
= =
(4.51)
where dry/dtr
= y() and by having applied the linearising control u= (I + mz3)v. COs(zl) mgz3 2mz2z3z4 (4.52) at the points
Observe that the structure of the system strongly impedes feedback linearisation z2 = Z3 = 0, thereby rendering the tracking of prescribed trajectories very difficult. More precisely, the condition z2 =0
the hinge, meaning that the slider is to be positioned away from the hinge point for tracking purposes. In conclusion, the feedback linearisation of the full nonlinear case is abandoned
and a linearised MMP design around the origin is considered instead. Towards that end, let A= Dz f (0,0) and B= Dj (0,0), where (4.50) is written as z=f as .z= Az + Bu, thus (z, u)
Z2 =I
z3 = Z4
gin
zgu
(4.53)
9Z1 Z3.
so that Theorem 2.6 can be invoked to show that the Full Problem is solvable whenever (2.65) can be fulfilled. of M in Figure 4.27. It is now attempted and suitable dissipation. to
Hence, the
dynamics is effectively modelled with the model bond graph. let the Hamiltonian
+ 21x1 22+
= mg3 sin(1)
at
101
02 2
C2\
TF
l(, 1)

/I3
TF t4(xl)
TF t2(xl)
12
\0
TF\ t3 (t 1)
1 xi
SS
Cl
I1/
11
k2
r2 _
(4.55)
xl Let zl = X1 1, z2 =
Z1 =
x2i first form that the takes the x2 z4 order and so model z3 = =
z2
z2
=12
(2m22z3z4 I  mz3
+ mgz3
cos(z1)
+ k1z1 + riz2
+ R)
z3
Z4
Z4
= =z3i
102
4. Model Matching
Control
1(0,0) and B= Du f (0,0), where 1(0,0) Az z= Then Define A= D2; 0. + B and write =
Z1 = Z2
z2 Z2
Z3
z4
=
=
=
k1IZ1zl
Z4 Z4
rlZ2 Iz2%2_
1gm I I z3Iu
(4.57)
T2_ mZ4 _ ` Z3 m
9Z1
/=
z3.
At this stage both the plant and model have been linearised and the linear MMP can now be addressed, where the constrained dynamics algorithm Doing so yields
z1+
r2
zg
Z*.
2z3` rZ4 mg
r2k2 z_ I
mg k2 mg 
Z*={(z,
_  z2 il 
r2
mzgs
m2g
z4
=: 0}.
13
Z4
Z4
(4.58) In view of (2.65), the control c(z, ) is readily found by considering the relationship
(4.59)
r2I m2
zZ +
53 gm3
3\
+
Finally, e=z
J E4++W.
(4.60)
the convergence of the tracking error can be assessed through dynamics has the cp(z), which
el = e2
e2 =I
e3 =
gm
e4
1 e3  w
(4.61)
e4 = Most importantly,
gel.
(4.53). (4.61) has Stabilisation that the the same structure as plant observe for w= K[z  cp(2)] = Ke and 0
through spectral assignment readily solves the linear IMP some suitable gain K.
103
characteristics and difficulties of more generalised bond graph based MMPs that can be briefly summarised as' follows. Example 4.5 shows that the model can be chosen to be structurally the implicit context, but it is believed that such model definitions dependent on the plant. Furthermore, "close" to the plant in are difficult to formalise
an inversion process showed not to be possible, since the output by means of an SS component. definitions As a result, it can be
to extract output
must be added to the plant dynamics outside bond graph modelling steps
the bond graph framework, thereby avoiding certain "contrived" that would increase the complexity.
Example 4.6 touches on the issue of alternative causal assignment procedures and the solution to a linearised VIVIP due to the lack of feedback linearisability. slider proved not to be particularly efficient model that simplified can be said to be highly It followed that the frictionless yielded a more
the MMP.
The choice of the causal assignment procedure so that any explicit or elaborated guidelines on the here. Also, the
formalised
of the slider did lead to the linearised MMP, where the bond Therefore, the closed loop as prescribed by the
4.5. Concluding
Remarks
This chapter showed that bond graph modelling physical model based MMPs.
of certain
the controller design in this chapter: 1. The nonregular Dynamic Disturbance ment [Hui92], [Hui94] . 2. The Full Information Output Regulation Problem [Isi9O], [Isi95]. Decoupling Problem with disturbance measure
104
Control
con
cepts is valuable, since it formalises controller design steps that were not documented in the bond graph literature. Because the model bond graph is based on the plant bond graph, this produces control designs that moderately redefinition alter the plant input/output dynamics. Thus, instead of a complete model, a physical model of the
of plant input/output
similar to the plant can be used in the control design and to render the solvability MMP more likely. The MMP stabilisation for explicit
systems can be based on the ideas of Theorem 2.6 for closed loop the plant bond graph.
However, the closed loop error dynamics and its associated bond graph representation more difficult tion of implicit to find for implicit
systems, since, depending on the method used, the reducdynamics need not yield portHamiltonian dynamics for reduced implicit dynamics.
dynamics to explicit
Nonetheless, the closed loop input/output have the bond graph representation Feedback linearisation linearised MMP
systems remains to
graph, on the other hand, need not change for the linear model since the linearised dynamics is derived from the nonlinear bond graph dynamics.
105
5.1. Introduction
As mentioned in Section 2.5, the notion of feedback passivation can be loosely referred to as the process of rendering the closed loop passive with respect to some energy function [Byr9l]. feedback function through output and
(non)linear be field have in in they to that shown successful a vast quite various control and control problems. For example, the widely used backstepping approach provides a relatively systematic framework plication for both stabilisation and tracking control through the recursive apOther interesting examples
[Ort98], [Ort00c], [OrtO2a] [OrtO2b]. These papers and works such as plant can be transformed into another portHamiltonian sys
the above examples of feedback passivation belong to a small subset of control problems that literature However, be in the terms theory. sheer of can addressed of passivation volume passivity techniques in control renders it virtually works that provide a concise overview of the topic. on
impossible to even begin to cite certain The reader is therefore referred to the
above cited papers and references therein for further details on feedback passivation. It is wellknown that bond graph modelling is based on energy concepts [KarOO]. In
deed, the C and I elements represent energy storage, the R elements denotes energy dissipation, and SS elements represent energy supply/extraction. to generate a class of portHamiltonian the second order EulerLagrange Bond graphs have been shown
[Go103] but are also capable of producing systems an alternative causal assignment bond graph charac
scheme [Kar77], [Kar83], [Bre94], [Mar02]. In view of such fundamental teristics, this chapter explores bond graph modelling in stabilisation
developments Interconnection Some ideas and on energy shaping and passivation. ing Assignment (IDAPBC) [Ort00a], [OrtO2b] are also considered.
106
Control
from balancing a bond graph perspective, where connections with the work presented power in [OrtOOa], [OrtOOb] are pointed out. It will be shown that power balancing through bond graph considerations implicit is applicable to a class of bond graphs that can produce explicit dynamics. The particular and
portHamiltonian
the case that some desired closed loop storage function is "guessed" for feedback passivation purposes [Ort98]. Examples of power balancing control will be presented to show that the structure considerations. and Damping
The latter part of the chapter explores bond graph modelling in Interconnection Assignment Passivity Based Control in [OrtO2b]. (IDAPBC) of portHamiltonian
systems as presented
However, IDAPBC
Even though it does not belong to the overall objective of for these first order of a
PDEs to be solvable at all. A detailed example shows the bond graph interpretation basic IDAPBC
design, where the bond graph topology is shown to aid the choice of desired and damping structures. methodology cannot
that bond graph modelling can be used to define IDAPBC level by means of the graphical representation of IDAPBC
of the closed loop. However, the complexity techniques to find the control.
This chapter is by no means exhaustive and the various developments merely present a subset of bond graph aspects in feedback passivation control. In particular, considerable analysis bicausal
remains necessary for all feedback passivation designs considered here. In addition, assignment will not be used since feedback passivation linearisation of input/output behavior as a design goal.
107
Control
5.2. Stabilisation
through
Power Balancing
This section addresses control through power balancing, where it is shown that this control strategy can be characterised by the junction structure topology. Closed loop energy functions are shown to be almost immediately decomposition. derivable when the bond graph allows for a certain of power balancing can from the
on the associated energy function; for is positive definite based on between the is given
the bond graph alone. Of course, the I and C storage components distinguish types of energy considered, but no analytical whatsoever. information
As a result, aspects of bond graph modelling in any sort of energy shaping are structure
more analytical in nature. This section, on the other hand, shows that the junction can provide qualitative information
to remain passive with respect to the natural output. The bond graph based power balancing method is a special case in which the junction ture topology provides information on the type of energy function struc
with feedback. Nov, if the plant energy is known, and the bond graph can be decomposed into two subsystems, then it may be the case that the energy function of one subsystem is a suitable energy function for stabilisation purposes. The graphical identification of ingoing
and outgoing power flows of some subsystems is shown to be crucial for the power balancing method. Various examples are presented that impart the results.
5.2.1.
Introducing
Power Balancing
system
= [J(x)  R(x)]K(x)
y= where KT(x) tonian = DH(x) 9T (x)K(x),
 g(x)u
(5.1)
(5.2)
108
law. Suppose there exists an HQ(x) such that H[(x) where a(x) is smooth. H(x) = + HQ(x) < yT (x)u + yT (x)a(x), (5.3)
u= a(x) +w yields
H8(x) < yT(x)w. Hence, the control u= shaped Hamiltonian a(x) +w (5.4)
H3(x). for example, if H8(x) is positive ST > 0, asymptotically {x : y(x) = 0} can only
Standard passivity theory can be invoked to assess stability, definite at the equilibrium stabilises the equilibrium be xe [Sep97]. xe then the control w= xe if trajectories
Sy, with S=
The following elementary example shows how a power rate of the form (5.3) can be derived by using the bond graph junction Example 5.1. structure.
Consider the bond graph in Figure 5.1, which is causally assigned with H: X f R be simply given as kx2 +1 +1 x1 x3. =1 2m1 2 2m2 structure
X3x3.
(5.5)
m2
But this clearly suggests that one can choose HQ(x) =2 for some gain c>0. Ixe m x3J M2 . (x) = coy, (5.7)
of
the form (5.3) can be derived from Figure 5.2 by simply considering all power flows through external bonds. Doing so yields H5(x) = y(u so that the passive control is of the form u= immediately  co'), cu +w=
(5.8)
be is to + seen w, which ckx2
109
Control I2
x3
ss
\I
l/
01
Figure 5.1.: Power flow ay for passive feedback with SCAP of Example 5.1. Ii C 12
SSIN xl
: 01 2/1552
Figure 5.2.: Power flow cry for passive feedback with LCAP of Example 5.1. The important step so far is the isolation of the bond graph subsystem associated with the of the form 1YQ(x) = y'(x)a(x). consider the bond graph
C and 12, which allows for the factorisation If one prefers the Lagrangian
causally assigned with LCAP in Figure 5.2. Note that the passive output remains conjugate to u, thus y= To make the design more explicit bi. L(x, x) =
so that by causal analysis of Figure 1=k x2
As usual, by setting zl = Xl,
as (5.9)
2mlthi
+ 2m2t2 
(5.10)
z2 = b1 and
z3 = x2,
system
turns
into
the
first
order form
Z1
i2
=
=
Z2
k
(zl 
zg)
u
'nl
'nl
(5.11)
= =
Z4 k M2
(zl
is to be applicable,
Lagrangian
 z3)2.
(5.12)
110
Control
Completely analogous to (5.2), define the shaped energy as E3(z) = E(z) + E,, (z). Then, as in the SCAP case, the relevant power balance can be derived directly and is found to be 6y = k(zl  Z3)(Z2  Z4) + m2z44 This clearly suggests to choose Ea(z) =C [m2Z4 + k(zi  z3)2] = Ea(z) = c. Qy, (5.14) (5.13) from the bond graph
which shows that the shaped associated energy must satisfy E3(z) = y(u The control is therefore u= It is important co, +w=  ca). (5.15)
ck(zl  z3) + w. of LCAP does not change the power balance structures is independent of the
Some remarks can be made on behalf of the above example. First, consider Figure 5.1 and suppose that the bond graph subsystem comprising element connected to the 1junction an additional the C and 12 elements has a resistive
HQ(x). be flow imposed on associated with the resistive element will power
Second, the storage functions associated with the 11, I2 and C elements are not coupled. As will be seen in the following section, the storage elements cannot be arbitrarily coupled if
5.2.2.
Defining
the Power Balance Method example of bond graph based power balancing, a more general
case of power balancing can now be considered. To that end, consider the explicit MIMO system depicted in Figure 5.3, where bond graph storage and dissipative elements are contained Hamiltonians bond The Ei has block Eo but the storage graph elements only. where within Ho(x) and Hl() and El. denote the associated energy functions of the bond graph subsystems Eo
111
Control
Eo: Ho(x)
112
SSi
E1: Hl()
20 Qu /1 2
SS2
Figure 5.3.: Conceptual bond graph based power balancing with SCAP. ', ), (VI, ) Now, for convenience, define ui = (ui, (o', ) uni yj yn, and ai = 0.3 = ... , ... , ... , {1,2}, (ul, (yl, (Q1, jE and set with u= u2), y= y2) and o= Q2). Even though power continuity of bond graph junction structures is independent of causality, the power balance
method considered here is readily defined in terms of standard causal assignment. The point of departure is the definition of the plant Hamiltonian in Figure 5.3, given by
(5.16)
(5.17)
structure can be modulated by all variables. As in (5.2), define the shaped )= H(x, ) + H,, (x, and observe that from (5.17) it follows This yields the (in)equality (5.18)
function H8(x,
that one can choose Ha (x, ) = cHl (e) for some c>0.
Note that power balancing considered here assumes that the subsystems Eo and El do not have coupled Hamiltonians. Thus, the factorisation Hl() = yTO. is obtained when the power
flow yT5 is external to the system El. Some important conditions for power balancing must be mentioned at this stage. First, supequilibrium that is (locally) globally asymptotically stable
Sy(x, ) and S=
tion problem is to regulate y(z, ) to y(ze, fie) =0 implies that y(ze, e)a(ze, fie) = 0. It follows that the power flow associated with the control must be zero at the equilibrium [OrtOOa] balancing energy method of also mentions this particular be no power flow at the equilibrium. restriction (xe, fie). The that there can
112
Control
Therefore, the above considerations show that the class of systems suitable for power balancing is limited, since various stabilisation regulation problems exist that cannot be solved by asymptotic since the systems
are assumed to be passive with respect to the output loop, it must hold that the systems are (weakly) minimum [Byr9l]. [Sep97], sible
problems with respect to the natural output, but the class of systems suitable for this type of control is limited. The following example clarifies the conceptually defined MIMO in Figure 5.3. Example 5.2. Consider the 2input bond graph model depicted in Figure 5.4. The point power balancing as depicted
of departure is to recognise that the system can be decomposed into the form as depicted in Figure 5.3. Suppose the components Ii and Cj for ij suppose that the plant Hamiltonian takes the form do not share coordinates and
H(x,
where p= PT and Q= 1 J2 00 0001 100
2 0 1
)=Ho(x)+H1()=
1xTPx+26TQ6,
(5.19)
P12X1+ p22X2 1
10
ul U2
q111 + g12e2
g12e1 + g222
(5.20)
Yi Y2
p11x1 +p12X2
jL
p12X1+ p22X2 structure in Figure 5.4 it follows that the power balancing control This control can be verified by
(5.21)
Note that the power balance method does not provide a of the energy function by means of the gain c in this shows that P and Q are to
case. More precisely, since the gain is assumed to satisfy c>0 be positive definite if Hs is to be positive definite.
113
Control
Il
zl12
C2
I2
C1
z2
SS1 \
U1
I1 yi
l/

Y2
011Q
I2
>0
J u2
SS2
bond graph based power balancing of Example 5.2. Sy, with S= ST > 0, can
By assuming that P and Q are positive definite, the feedback w= be used to control the asymptotic convergence. Finally,
the requirement
to have passivity
with respect to the natural output limits the coordinate dependency of the additive energy, for it is seen that DHa(x, e)9j =0y= 9T Ids (x, e) = 9T K(x), (5.22) can therefore not 0
where gj are the columns of the input matrix of (5.20). The xcoordinates be used in the definition of Ho,.
Provided
addressed within
possible, to find the bond graph subsystems Eo and E1 with the added assumption that El has no dissipation. In practice, the general way to proceed is by identifying the junctions
step can be completed, it should be checked whether the plant Hamiltonian representation of Figure 5.3, which may or may not be possible.
can be written as the sum of two suitable functions, then the MIMO (weakly) is the provided system minimum phase.
Implicit
Systems:
Dependent
Storage Elements
various storage elements, it can be argued that not allowing dependent storage elements is rather restrictive in bond graph modelling. As a result, it is of great importance to address
derivative causalities in the context of power balancing if this method is to be applicable to a larger, relevant class of systems.
114
Control
storage elements that are manifested as derivative causalities. to eliminate the dependent storage elements is possible Indeed, multipliers are "workless" by definition, which
implies that the associated power flow is zero such that it cannot have any contribution. However, with standard causality, the insertion of various multipliers may yield complicated
dynamics since the state variables selected by SCAP may not be suitable for the problem at hand. Even though multipliers are capable of effectively addressing certain constraint dynamics, the multipliers, denoted as LCAP, appears to be the
best option for power balancing as considered here. It must be recalled, however, that LCAP is applicable if and only if a minimal set of generalised coordinates are available [Mar02]. In case this minimal set of coordinates has been found, the dependent storage elements are readily circumvented with LCAP. The second order dynamics obtained with LCAP are subsequently transformed into first order form, where the associated energy E of the Lagrangian is used Ea,
as a storage function.
which is to be derived from the bond graph by means of the junction SCAP case, the power flow yTQ is found at the 1junctions velocities. The Lagrangian
Similar to the
approach is quite effective in addressing the derivative causalities of bond (5.18) requires further as
graph models; however, the next example shows that the inequality generalisation.
More precisely, as in. (5.2), define the shaped energy of the Lagrangian
E3(z) = E(z) + Ea,(z) and suppose there exists an additive energy E3(z) such that ES(z) < yTp(z, where p(x, u) is smooth. Then by invoking u), Function Theorem (5.23) [Abr88] it is
the Implicit
Thus, as per (5.23), general bond graph based power balancing can be said to address passivation problems for which a suitable factorisation attained by means of an additive energy function. with respect to the natural output can be
115
2;

iI1,2
physical configuration
of Example 5.3.
example has been taken from [Jan96] and shows how the Lagrangian the
be balancing bond in in tool can a graph modelling power control. In particular, approach function be "guessed" but the junction energy need not as such, associated additive hints at a candidate additive energy function. Example Oscillations 5.3. In [Jan96], the authors consider the problem of controlling of a Rotational Actuator (TORA)
structure
the Translational
Now, the authors of [Jan96] present the design of various that are shown to stabilise the
cascade and feedback passivation controllers for the TORA system globally.
It is pointed out by [Jan96] that the passivation control designs cannot be since the closed loop storage function is to be "guessed" in some way.
said to be constructive
This example shows that bond graph based power balancing control can provide a way of designing a passivation controller in a relatively constructive The structure of the TORA fashion. that can oscillate without a
damping in a horizontal plane, thus the effect of gravity is not considered. On the platform, rotating eccentric mass is actuated by a DC motor and its motion
is the torque applied to the eccentric mass to stabilise the system globally equilibrium. Since the kinematic relations of the system will certainly
an efficient model.
[Mar02] be AHCAP that of can applied, but the dynamics tend to become more comnoted in Figure The 5.5 are designated as xl = x, which represents coordinates generalised plex. the position of the platform, and where the pendulum angle is designated as x2 = 0.
116
Control
cl
ssl
2
Il
J0 11
x1
j SS2
\ I3
TF t2(x2)
U2
TF tl(x2)
o q
I2
C2
Figure 5.6.: TORA Lagrangian assigned bond graph of Example 5.3. (x2) LCAP, tl with where =1 cos(x2) and
t2(x2) =l sin(x2). The Lagrangian is given as the kinetic energy T(x, th) minus the potential (x), V thus energy L(x, x) =T(x, where a= x) V(x) 2a1 = +cos(x2)xix2+ 2ry2 2kxi, (5.25)
ml + m2i /3 = m21 and ry = m212. By following the depicted causality yields the
+=0.
(5.26)
Next set zl = xl, Z2 = 1 and z3 = x2i Z4 = x2, and write the first order dynamics
z1 = Z2 'y(/3sin(z3)z4 kzl) 8COS(z3)
z2
z3
Z4
2 cost
kzl) +a
(z3)
ay
(5.27)
Z4
Cos(z3)( sin(z3)z4 a7 Q2 COS2(z3)
Q2 COS2(z3)
a'yu
117
5. Energy Shaping in Stabilisation In [Jan96], the first passivation controller, u= Cz3 + rz3 for some gains c>0 of z= (0,0,0,0).
Control
denoted by the authors as "P2", is chosen to be This control is shown to achieve global it is seen that this
and r>0.
stabilisation controller
can be obtained by considering the power balance for 02y only, which represents of additive potential energy with respect to the angle of the pendulum. The
the contribution
power flow v2y implies that one can (but need not) choose the additive E,, (z) = (c/2)z3. Even though this controller achieves stability
associated energy
of the origin, the authors beyond a certain limit. to improve the settling is chosen to
time of xl
They argue that the energy can be shaped in the x1coordinates time; however, the authors do not disclose how the additive attain the x1dependency of the feedback passivation controller.
energy function
In view of power balancing, observe that the dependence of the controller on the xlcoordinate can be achieved constructively by considering the power flow uly. structure + kx11 (5.29) as More precisely, consider
the power balance equations induced by the junction "W U2Y = m244 + mixlxi
__ ax2
av Y
step in the design, because it becomes possible to select a candidate additive energy function yielding a factorisation in the natural (5.23). form In particular, the output y of the above
power balance equations lead to the choice Ea(z) =2 =2 where Cl >0 [kzi + m2(z2 +l cos(z3)z4)2 + mizi] [kzi + 2b cos(z3)z2z4 C2 + z3
(5.30)
+ az2 + Cz4 COS2(z3)] +2 z3,
the bond graph it now follows that this additive energy must satisfy
E.
which, in turn, implies that
= z4(a l+ Qz),
(5.31)
E5(z) _ z4p(z, u). By invoking u= implicit function arguments, it is readily seen that
118
5. Energy Shaping in Stabilisation Indeed, the control satisfying (5.33) takes the form
Control
u=
a(z,
w)
(2
QZC1 [Qycikzl
aY
3tclkzl
COS3(z3) +Q2C2z3
COS2(z3)
F
g2yclz4  , ayc2Z3
sin(z3)]
Cos(z3)
+ [2 cos2(z3)
a^yIw
(5.34
a(z, w) and
the system (5.27) is feedback passive with storage function E5(z). Global asymptotic stability of z=0 is achieved by the further control w= ry with r>0. (V/x2)y. since Qty =
It is important
0
As mentioned earlier, power balancing cannot be used for systems that are not (weakly) minimum phase [Byr9l]. For instance, there are occasions where a proper bond graph de
composition exists, but where the additive energy function cannot render the shaped energy function positive definite at the desired equilibrium. system. Example junction 5.4. Consider the bond graph in Figure 5.7 of the frictionless structure of this system is seen to induces derivative slider. Since the The following example addresses such a
As usual, the Lagrangian is the kinetic energy minus the potential energy and takes the form L(x, x) = 2mx2 2mx22 2Ithi th1 + +  mgx2 sin(xl). (5.35
With zl = xl, Z2 = xl and z3 = x21 Z4 = x2 the associated energy E(z) of the Lagrangian readily follows to be
E(z) = Set y= th and identify 2mz4 2mz3z2 2Iz2 + +mgz3sin(zl). 2+ (5.36)
Ea = y(Qi + ca + 73).
(5.37)
119
Control
C2
TF ti(xi)
13
TF t4(x1)
TF t2(x1)
Q3
120
TF
t3(x1)
J1x
SS1
Cl
1J
l1f

Il
Figure 5.7.: Power balancing not applicable to slider of Example 5.4. At this stage it must be noted that the C1 and C2 storage elements are coupled, so that the power flow my must be taken into account. The additive energy can be described by
[mz4
EQ(z) =
+ 2mz3z2
+ mgzgsin(zl)J
+2
zi,
(5.38)
for some controller gains cl >0 and c2 > 0. However, it is of no use continuing the design since the shaped energy E3(z) cannot be rendered positive definite at z=0. the minimum phase property Furthermore, since
the frictionless slider cannot be stabilised by feedback passivation with respect to the natural passive output y. 0
The above example shows that the power balancing precludes systems with unstable internal dynamics. But, as mentioned earlier, the bond graph does not provide any information energy function to assessinternal stability. does not guarantee that asymptotic alone. Therefore, the existence of the factorisation on the (5.23)
stabilisation
Future research could address bond graph based power balancing that allows for passive output factorisations for certain nonminimum phase systems.
alternative
120
Control
It has been shown that bond graph based power balancing can be used to obtain passivation controllers with respect to the natural output. Most importantly, output the junction structure
passivity
is retained when
selecting the additive energy function novel in the bond graph literature
passivity in closed loop. Moreover, power balancing is applicable regardless of the bond graph complexity, but the proper decomposition Natural output may be difficult to find, if possible at all. For exam
ple, admissible equilibria that can be stabilised are limited due to the fact that such equilibria are to be compatible with the zero output, thus y=0. implies that the systems must be (weakly) minimum Furthermore, natural output passivity is to be possible.
phase if stabilisation
In other words, natural output passivity in closed loop requires that the internal dynamics is at least marginally stable.
Power balancing as presented here can be applied to bond graph models with dependent storage elements. For example, the Lagrangian istence of dependent storage elements without Indeed, power continuity of the junction approach has shown to circumvent compromising the ex
structure
is independent
by construction.
use LCAP, it is required that a minimal set of generalised coordinates can be found. Future research could address other causal assignment schemes in power balancing control, such as )HCAP or )LCAP, since the application of Lagrange multipliers does not require a minimal
set of generalised coordinates, thereby avoiding the search for such coordinates for complex systems of interconnected bond graph submodels.
Finally, it may be of interest to see whether power balancing can be extended to include closed loop passivity with respect to different outputs, so that the class of systems suitable phase
for bond graph based power balancing can be enlarged to include certain nonminimum systems. For example, the stabilisation with respect to the natural problem may be solved. output,
of the frictionless slider was shown not to be possible but by modifying the closed loop passive output this
121
Control
5.3. Control
by Interconnection
and Damping
Assignment
In the previous section it has been shown that bond graph based power balancing can be used to derive an additive energy function for control purposes, and where the natural output remains the closed loop passive output. power flow considerations, representation In particular, the controller design is solely based on
where it is important
This section explores bond graph modelling aspects and interpretations called Interconnection explicit portHamiltonian and Damping Assignment Passivity
Based Control
[OrtO2b]. developed in The main feature of a IDAPBC systems as with respect to some shaped energy of the
design is that the closed loop remains portHamiltonian function. But, in addition, IDAPBC
interconnection
and damping structures through suitable feedback. design allow for bond graph representations, but the
solution to the PDEs associated with IDAPBC such conceptual representations an IDAPBC
have merit and are helpful for bond graph models for which
design is considered. Even though bond graph models have been identified as a systems [Gol02j, the aspects of bond graphs in IDAPBC This section presents an introductory designs
been addressed.
account on IDAPBC
5.3.1.
Structure
Compatibility
The power balancing method in Section 5.2 shapes the energy of systems in way that renders the closed loop passive with respect to the natural system output. Indeed, the sole purpose But
of power balancing is to render the time derivative of the shaped energy nonpositive.
instead of finding additive energy functions that retain passivity with respect to the natural output, it can be of interest to characterise all additive storage functions that are compatible with the original bond graph topology. Section 2.6 but where interconnection This section follows the various arguments found in and damping structure assignment is not considered structure
at this stage. As will be seen, energy shaping that is compatible with the junction often leads to passivity with respect to a different output.
122
Control
Following the arguments of [OrtO2b] and Section 2.6, consider an explicit portHamiltonian system of the form
x=
y= where DH(x) function
[J(x)  R(x)]K(x)
9T (x)K(x),
 g(x)u
(5.39)
H8(x) = H(x) + Ha(r), where Ha(x) is the additive Hamiltonian to be associated with feedback control. a(x) +v such that
(5.40) To that
(5.41)
 g(x)v
(5.42)
(x)KK(x)" Ys = 9T The additive function Ha,(x) can therefore be used, in principle, to assign some desired (local) minimum to the shaped energy H3(x), so that stabilisation is possible by rendering xe the Sys for S= ST >0 and
Z=
{x EX:
I<(x)[R(x)
+9(x)SYT(x)]K3(x)
= 0}.
(5.43)
a
Now, suppose H5(x) is positive definite at the desired equilibrium LaSalle argument [Kha92] it can be shown that asymptotic only trajectory contained in Z is the equilibrium xe.
stabilisation
If (5.41) is to hold for all Ha,(x) then is must also hold that
91(x)[J(x)
where g1(x) is a full rank left annihilator
but is to simultaneously
Suppose (5.44) can be solved for some H,,,(x), then the control is obtained with [9T (x)9(x)]19T (x)[J(x) R(x)]Ka(x) = u. (5.45)
123
Control
Even though the above energy shaping method is conceptually to (5.44) is a crucial step in the design. The following
straightforward,
the solution
proposition
provides a necessary
5.1. Let X be a smooth manifold with local coordinates (xl...... IR j=1, with > Suppose the function m. ... ,
H: X 4R simulta
neously satisfies the first order PDEs Fi (xi, pi) = 0, alllax with pi =
(5.46)
(5.47)
{", is Poisson bracket T*X. the where standard on "} Proof. Let F, G: T*X R H(xi) and suppose >
satisfies
F(xi, Pi) = 0,
with pi = aH/xi.
G(xz, pi) = 0,
(5.48)
o, +p + = =a x ax ax ap oxi
Multiplying the first relation with aG/p, the second relation with F/pi, the second relation from the first yields IF, G} axi Opi 5 Pz x, _0. aF BG 8F 8G
o9F
(5.49)
and subtracting
(5.50) 0
is merely implicitly
of [Car65] and gives a necessary condition that is to be satisfied if (5.46) is to be at all solvable for some function H(x=). Note that no explicit conditions on the solvability 5.1 is reported here for maximum clarity. of (5.46) can be
The following example shows that energy shaping as per (5.41) implies compatibility the plant bond graph topology.
with
124
Control
C
x3
I2
xl
SS
1f
/R
0T fla
SS I 1 I/
K2a 00
011R
I(3a
Figure 5.9.: Energy shaping compatible with bond graph topology; Example 5.5. Example simplicity, 5.5. Consider the bond graph of a dissipative suppose the Hamiltonian H(x) system depicted Figure 5.8. For
2XT
where the constants Il and 12 are inductances and where C is a capacitor. yields the simple linear dynamics x1 x2 X3 y= =10 01 xl/Il, 0 1 0 1 rl xl/Il x2/C x3/I2 0u 0 1
(5.52)
where rl >0 is a resistive constant associated with the R element. The admissible equilibria of (5.52) take the form
xe = (4, x 23 ,x)=
(Ilue/ri,
Cu',
I2ue/r1),
(5.53)
attainable through feedback, observe that the condition (5.41) can be represented as the bond graph in Figure 5.9. Thus for energy shaping to be compatible with the plant bond graph, the contribution of the additive energy is to "cancel" internally the tangent vector due to underactuation, which can
be represented by nullifying
in Figure 5.9.
125
5. Energy Shaping in Stabilisation The PDEs to be simultaneously Figure 5.9, and doing so yields
Control
Ha  rl
49X3
{F1, (5.54) F2} In fact, 0. does exist and takes the to that that observe solution confirm = the form
(5.55
at xe to the shaped
function
in x
xe. To
2I1
2C
212 '
(5.56)
where Cl >0 is a controller gain. Observe that with the above choice of 0 it follows that the shaped Hamiltonian takes the form
(5.57)
Cl
Finally, the control that attains the desired energy shaping is obtained from (5.45), or from the bond graph in Figure 5.9 as u= v= aHa/x2 + v. Standard output convergence. feedback of the form 0
As shown by the above example, energy shaping that is compatible with the junction structure induces first order PDEs by considering the summations property of underactuation can be imposed by nullifying of zero and onejunctions. the tangent vector x. The
126
Control
I2
xl
x2
SS
1l
p1
GY
a(x)
of Example 5.6. 12
0K
T la
Kea 0O
KT 3a
SS\
{11/01
GY a(x) Figure 5.11.: Underactuated In case the junction bond graph of Example 5.6.
structure is smoothly modulated by x, the relation (5.47) readily provides for such modulations to be admissible, and this argument will also be
a necessary condition
used when interconnection The following energy shaping. Example modulation explicitly, 5.6.
and damping structures are modified through feedback control. structure modulations in
Consider the bond graph depicted in Figure 5.10, where a(x) is a smooth It is not needed to evaluate (5.44)
8HaaHa x1 x3
Ua a(x) By introducing
_0
axl
H" +=0.
(5.58)
0X2
0. pl J = 3 (5.59) example, it must hold 0
the functions F, (x, p) = p1  p3 and F2 (x, p) =a (x)pl + p2 it follows that {Fl, F2} R 8a_aal l
Therefore, if energy shaping is to be at all solvable for this particular that a=a(xl+x3).
127
Control
CI
81a
xi
SS 2 SS
00
axa
axe
Ri
JS(x) R2
of energy shaping.
portHamiltonian but the solution dynamics can be
is to be obtained can
analytical
be conceptually JS(x)
structure
may impede
energy shaping,
Interconnection
and Damping
Assignment
Instead of energy shaping alone, the IDAPBC for the modification systems through of interconnection
methodology
and damping structures of explicit portHamiltonian applications of this theory have been method enlarges the designs often lead to
feedback control.
Some instructive
reported in [OrtO2a] and [OrtOOc]. Even though the general IDAPBC class of portHamiltonian systems for stabilisation problems, IDAPBC
a set of first order PDEs that need to be solved. However, solving the associated PDEs of an IDAPBC design need not imply that the control problem can be attained [OrtO2a]. More
precisely, finding a mere solution to the PDEs does not guarantee the control objective can be attained. Extensive theoretical research on the solvability of IDAPBC type designs and
the associated PDEs is still ongoing, see [B1a02] and references therein for an authoritative treatment of the topic. based on [OrtO2b]
The energy shaping method as described in Section 5.3.1 is completely and is equivalent to the case for which no additive interconnection considered.
In this section, on the other hand, mere energy shaping is complemented with of the interconnection and damping structures through feedback.
possible modifications
128
Control Designs
TF, GY and R components. The insertion of such additive elements induces the Ja,(x) and R,, (x) matrices of (2.77). It must be noted that it is by no means attempted to define formal procedures and formalisations. Instead, the applicability of bond graph aspects in IDAPBC designs is presented by and generalisations that design
means of a detailed example, which readily allows for formalisations can be committed
can be used effectively for certain stabilisation closed loop interconnection Example 5.7. In [OrtOl],
and damping structures. the authors consider a magnetic levitated of this system is given as
H(x) + x3 +m9x2,
2I(x2)xl
2m
(5.60)
where xl is the flux linkage of the coil, and where x2 and x3 are the vertical displacement and momentum of the ball respectively. for some physical constants y' >0 The inductance of the coil is given by I(x2) _ 'yl/(72 x2) and rye > 0. It is readily understood that this system
can be modelled with a bond graph depicted in Figure 5.14, which shows a complete lack of structural interconnection between the electrical and mechanical energy domains.
The control objective is to stabilise the ball at some desired vertical position through voltage control on u. To this end, causal analysis yields the explicit portHamiltonian xl r 00 (Y2 xi  x2)/'Yl 1 dynamics
x2
x3
=001
0 1 0
gm  xl/(27i) x3/m
0u 0
(5.61)
71xi(72
129
Control
l9
Figure 5.13.: Magnetic levitating
IC
xi x2
SS \I11
Ia
R1
Figure 5.14.: Magnetic levitating The point of departure for stabilisation additive energy is to be compatible ball bond graph of Example 5.7.
mere energy shaping is possible, it can be expected that the induced PDEs are less complex than a more general IDAPBC design.
From (2.80), or from the bond graph in Figure 5.15, it is readily found that
( Ha
which shows that energy shaping is possible for the x1coordinate the Hessian D2H3(x)
is evaluated, it becomes clear that the shaped energy H, (x) cannot Ho, = Ha,(xl). Hence, it appears that an problem.
130
Control
R
Figure 5.15.: Energy shaping compatible with bond graph topology; Example 5.7.
IC
1th2
SS 

GY
a
I2
R1
R2
design commences by considering the plant bond graph with added gyration
damping R2 Observe Figure damping depicted in 5.16. that the gyrator element and as and lack The is bond through from a of coupling resolved perspective: choice a graph are a natural (constant) gyrator and where linear damping in the mechanical domain may improve asymptotic convergence. Note that the resistance R2 is not considered in [Ort0l]. At this point the IDAPBC design is conceptually clear, but it remains to be verified whether
the closed loop can be attained through feedback. To that end, take
a
rl Ra(X) =000,
00 (5.64) r2
OH4
x3 Ha
aaXl 
=0
= 0.
Ha
99x2
5Ha
r2x3
(5.65)
r2
mx3{IYlxl(72x2)
131
Control
Fi (x, p) F2(x)P)
(5.66) +a .
7i
x1('(2  x2),
from which it follows that {F1, F2} = r2/m. setup is to be at all solvable.
Thus r2 =0
8x2
Iyi
which is readily solved using the method of characteristics explicitly, define the initial data
r={(x1,
Ha=0(T)}y
(5.68)
and verify that the additive energy function Ha(x) takes the form Ha(m) _f where 0 is an arbitrary initial
differentiable
(2:
xi a/
(5.69)
function.
x2 = 0) Ha = 0(7")}i
which would yield the additive energy Ha(x) _ (xi + ax2) + aYzxixz '71 72x2 +az xlx2 a 271 27 a2 x2. 67i (5.71)
Clearly, the simplest solution is (5.69) and seen to yield the shaped energy HS(x) = H(x) + Ha(x) _ where q5(xl/a+x2) q be defined as
a xi +x2) (=x1 11e211ee ci + X2  xl =2  x2) (xl  mg + x2 'xl  x2 (5.73)
(5.72)
DH, (x') = 0.
(5.74)
132
Control
IC
1
Ra
XZ
ss\
Ys
GY a
13I
This proves the extremal assignment at xe. Moreover, for all cl > 0, the Hessian satisfies
I ci/a2
D2H5(xe) =
+ xi/(a'Yi) ci/a 00
cl/a Cl
which show that xe is a strict minimum of H5(x). and takes the form U= ax3+(r+ri)i9Ha m
(5.76)
design shows that the bond graph in Figure 5.16 offers a relatively natural system. However, as mentioned before, the which may
way of defining the closed loop portHamiltonian bond graph can only be used conceptually or may not be attainable.
dampfurther To its is though of additive explore choices a natural choice. appearance even ing, consider the bond graph in Figure 5.17 and observe that the R2 element has been moved to a 0junction. The interconnection by damping given structures are now and
00 Ja(x) =000,
a Ra(X) =0
rl
00 1/r2 0, (5.77)
a00000
where a>0, be PDEs 0. The induced to that > >0 need solved are and r2 ri
r2
aHQ + gm ax2
r2
12
2r2y1 (y2 xl =0
aHQ
aHQ
aX2
(5.78)
X2) = 0.
xl ryl
133
5. Energy Shaping in Stabilisation Invoking Proposition 5.1, define the functions Fi (x, p) F2(x, p) 1 = = P3 r2p2 gm r2 +
Control
12 2r271 xi (5.79)
a aPl  p2 + xl('Y2
 x2),
(5.69), leads (5.78) then to the additive energy the which admits solution of (ixi H,, (x) _ q5 Q+x2
1
+ r2 rx3)
xi (y2  X2)
2+ Yl
x3
(5.80)
2a
 Mg
r2 (lx
Xe 1
x2
2x3
1x2
i x3axixz2x3). 2
111l
(5.81)
thus DH3(xe)
Hessian satisfies D2H3(xe) >0 for all values of cl > 0. The control now reads u= (r+r1)i +aH3 + aX3+ixl('y2x2)+w. (5.82)
0
The above example shows an introductory basic IDAPBC deployment of bond graph representations for a design, where the following observations can be made. First, it is seen that can be used to choose interconnection and damping sense This is
electrical and mechanical domains. The insertion of a GY component remedies this shortage of coupling and a allows a force to be applied to the ball. Second, in view of damping structures, within additive damping can be based on available junctions
the bond graph, where Figure 5.16 shows that it is quite natural to place a resistive in the mechanical domain. On the other hand, as depicted in insertion of a new junction is
bond graph
arguments.
134
A final remark is in order, namely, it must be remembered that IDAPBC applies to explicit portHamiltonian not induce derivative causalities.
dynamics, so that the associated bond graph should For example, the bond graph of the frictionless slider in
a small class of systems that do not produce derivative causalities in bond graphs.
5.5. Conclusion
The first half of the chapter introduced a novel method for bond graph based power balancing control, which addresses stabilisation the natural output. wellknown through feedback that retains passivity with respect to technique must meet Crucial
detectability
to the power balancing method has been shown to be a proper decomposition bond graph into two subsystems with uncoupled Hamiltonians,
of the plant
the natural outputs as system inputs is assumed to have no resistive elements. In case such bond graph decomposition can be found, it has been shown that the additive energy function associated with the subsystem that has the natural output
Power balancing as presented in this chapter can be scaled and allows for bond graphs that will either induce explicit or implicit applicability can be attributed dynamics when causally assigned with SCAP. This broad structures remain power
application of the Lagrangian assignment procedure circumvented derivative causalities without changing the power continuity as detectability property. In any case, however, system requirements such remain in effect for all bond graph models, irrespective
of the causal configuration. The latter part of the chapter addressed basic bond graph interpretations signs. Unlike power balancing, the closed loop bond graph representation of IDAPBC de
of an IDAPBC
design has been shown to be more conceptual in nature and does not provide a means to find the control. This can be attributed to the fact that IDAPBC designs often lead to a set of
135
5. Energy Shaping in Stabilisation Even though the bond graph aspects of IDAPBC shown that the control objective can be graphically the design to some degree. In particular,
Control
designs are mostly conceptual, it has been depicted, thereby increasing insight into and damping structures have been of GY, TF and R components. design requite
interconnection
shown to be represented by the insertion and modification However, the shaped interconnection
mains to be chosen by the designer, thereby rendering flexible but less constructive.
the IDAPBC
methodology
136
6.1. Review
This thesis presented the following control design methodologies for bond graph based control purposes: (1) Backstepping in Stabilisation Control. Control, (2) Model Matching Control, and (3) Energy Shaping
no similarities on analytical
levels, it has been shown that these methods are capable of addressing closed loop bond graph representations. wellunderstood, More precisely, it has been argued that open loop bond graph modelling is
whereas closed loop bond graph modelling is not. Therefore, the impetus of control design methods capable of addressing Important aspects systems
of these structural
in relation to bond graph models. In conclusion, this thesis has shown that the above control design methods allow for structural found, thereby contributing with bond graphs. design goals such that associated bond graphs can be
6.2. Backstepping
Control
control design has been addressed from a bond graph perspective, where the
functions were defined as additive bond graph models and referred to as virtual Most importantly, this thesis showed that explicit portHamiltonian through an exact backstepping judiciously closed loop
Through
chosen virtual
shown that the plant bond graph topology can be retained such that passive stabilisation can be interpreted by means of the plant bond graph. Furthermore, the bond graph based
backstepping approach readily showed to be applicable in a nonlinear context with suitable virtual control laws, thereby complementing the existing literature on the mere linear case.
137
6. Conclusions and Future Research Backstepping is known for its use of new coordinates that are introduced that recursively at
without
closed loop dynamics, hence for the closed loop bond graph representation. bicausal inversion mechanism has been shown not to facilitate each step due to the absence of these new coordinates. Most of the existing literature systems. on bond graph based backstepping the further
addresses singleinput
of multiinput
systems having the required interlaced structure can be applied. The multiinput
based backstepping
be applied to each branch. The (bi)causal inversion mechanism has shown to be applicable to such multiinput systems.
Control
tracking controllers by means of bicausal bond graphs has previously This thesis complemented some known results by means of the (MMP) of prescribed model trajectories instead of the common
of "arbitrary"
reference trajectories.
be used to define prescribed reference models, where the closed loop input/output should match the input/output behavior of the prescribed model.
The main result of the bond graph based MMP has shown to be the underlying mechanism of the MMP design. Indeed, this thesis has shown that bond graph modelling for the physical model based MMP implicitly relies on the theory of dynamic disturbance decoupling and
concepts of output regulation. For certain MMP scenarios it has been shown that the "error dynamics" ated bond graph representation, graph arguments. In particular, so that passive stabilisation allows for an associ
manifold considerations whereby the tracking error of state variables can be found by means of a submanifold on which output matching occurs.
138
such as derivative
state variables.
be preferable for the MMP design. Towards that end, the Lagrangian causality assignment procedure did not yield explicit graph representation For certain nonlinear portHamiltonian dynamics, so that an associated bond
for the error dynamics could not be found. be based MMP has bond to the graph shown not solvable scenarios, mechanism. As a result, the linearised MMP was
the linearisation
operating point.
Control
has been presented from a bond graph perspective, where the bond has shown to be capable of identifying feedback passive control
structures
balancing" and was used to derive the closed loop storage function from the junction instead of a predefined storage function. zerodynamics,
feedback for is a general requirement all passivation design with respect which Power balancing has been shown to be suitable for multiinput
bond graphs and for bond graphs having derivative causalities. Interconnection and Damping Assignment Passivity Based Control (IDAPBC) has been
IDAPBC of solution
designs, however, is known to depend on first order partial differential in IDAPBC are applicable on conceptual but
139
The thesis has identified three control methods that allow for structural of closed loop bond graph representations. or "geometric" uncertainties.
feedback designs often lack certain robustness margins with respect to model In particular, the robustness of exact backstepping and model matching conperformed in these
designs. For example, recall that this thesis focused on exact backstepping, on exact cancellations. Furthermore,
which is based
expected to be relatively low due to the explicit feedback linearisation dynamics. Even though it can be argued that linearisation mathematical dynamics.
context, real physical dynamics will generally deviate from prescribed model performance often deteriorates significantly In view of controller or may even and
As a result, controller
commissioning
robust control of bond graph based feedback designs is the most important advances of modern robust
control theory for linear systems suggest that future work on bond graph based controllers should be restricted to linear systems first.
140
Bibliography
[Abr88]
Abraham,
R., J. E. Marsden.
Manifolds,
vol
ume 75 of Applied mathematical sciences. SpringerVerlag, [Bar77] Barnard, B. W., and P. D. Dransfield. Predicting
draulic control system using bond graphs. ASME Journal of Dynamics Systems, Measurement and Control, 99(1): 18,1977. [Bir9O] Birkett, S. H., and P. Roe. The mathematical foundations of bond graphs  iv.
matrix representations and causality. Journal of the Franklin Institute, 327(1): 109128,1990. [B1a02] Blankenstein, G., R. Ortega, and A. J. Van Der Schaft. The matching conditions based control. International Journal of
of controlled lagrangians and idapassivity Control, 75(9), 2002. [B1u82] Blundell, A. Bond Graphs for Modelling
Engineering
Systems.
Ellis Horwood
Publishers, 1982.
[Bor00] Borutzky, W. Bondgraphen Eine Methodologie Modellierung zur in der Simulationstechnik. multidisziplinSCS European
rer Systeme, volume 6 of Fortschritte Publishing [BorO4] Borutzky, House, 2000. W. Bond Graphs A
DyHouse,
research. Journal of Systems and Control Engineering 328(56), 1991. [Bre92a] Breedveld, P. Proposition for an unambiguous vector bond graph notation. ASME
141
Bibliography [Bre92b] Breedveld, P., and G. DauphinTanguy. 1992. graph elements in physical systems theory. Journal of Bond Graphs for Engineers. Elsevier
[Bre94]
Breedveld, elimination
P.
Multibond
New York, 1998. [BV04] [Byr9l] Control Lab Products By. 20Sim. URL: http: //wNvw. 20sim. com/, 2004. Passivity, feedback equivalence, and Transactions
Byrnes, C. I., A. Isidori, and J. C. Willems. the global stabilization on Automatic of miminum
Control, 36: 12281240,1991. C. Calculus of Variations and Partial Differential series in mathematical Equations of the 1965.
[Car65]
Caratheodory, First
Order. HoldenDay
[Ce191] [Da197]
Cellier, F. E. Continuous System Modeling. SpringerVerlag, Dalsmo, M., and A. J. van der Schaft. A Hamiltonian physical systems. Technical report, Universiteit
[Des75]
[Dij9l]
Dijk,
Simulation
Zero
Order Causal Paths  I. Classification of ZeroOrder Franklin Institute, [Dix74] Dixhoorn, 328(5/6): 959979,1991.
J. van, and F. Evans. Physical Structure in Systems Theory: Network Acadamic Press, 1974. Economics. and
URL: http: //www. dynasim. se/, 2004. Differential Equations, volume 19 of Graduate Studies in
Evans, L. C. Mathematics.
Partial
American Mathematical
Society, 1998.
142
Bibliography [Gaw92] Gawthrop, P. J., and L. Smith. Causal augmentation loops. Journal of the Franklin Institute, of bond graphs with algebraic
329(2): 291303,1992.
[Gaw95a] Gawvthrop, P. J. Bicausal bond graphs. In Cellier and Granda, pages 8388,1995. [Gaw95b] Gawthrop, P. J. Physical model based control: A bond graph approach. Journal
Gawthrop, P. J., D. Ballance, and D. Vink. Bond graph based control with virtual actuators. In Simulation in Industry. ESS, 2001. Special issue on bond graphs. Proceed
[Gaw021
Control Engineering 216(1), 2002. [GawO3] Gawthrop, P., and D. Palmer. A bicausal bond graph representation of operational amplifiers. Proceedings of the IMechE Part I: Journal of Systems and Control
Engineering, 217: 4958,2003. [GaNv04] Gawtrop, P. Bond graph based control using virtual actuators. Proceedings of the
IMechE Part I: Journal of Systems and Control Engineering, 218: 251268,2004. [Go102] Colo, G. Interconnection and Simulation. [Go103] Structures in PortBased Modelling: Tools for Analyisis
2002. for
Colo G., A. J. van der Schaft, P. Breedveld, and B. M. Maschke. Hamiltonian rnulation of bond graphs, pages 12. Lecture Notes for the Euro/Geoplex School. SpringerVerlag, 2003. control: An approach to manipulation: Part i
Summer
[Hog85]
Hogan, N.
Impedance
theory.
ASME Journal of Dynamics Systems, Measurement and Control, (1): 17,1985. [Hog87] Hogan, N. Modularity and causality in physical systems modelling. ASME Journal
143
Bibliography [Hui92] Huijberts, H. J. C. A nonregular solution of the nonlinear dynamic disturbance to a complete solution of the nonlinear on Control and Optimization, 30: 336
decoupling problem with an application model matching problem. 349,1992. [Hui94] Huijberts, SIAM Journal
H. J. C. Dynamic Feedback in Nonlinear Synthesis Problems. Centrum 1994. and Simulation (ICBGM),
voor Wiskunde en Informatica, [ICB03] International 2003. [Isi90] Isidori, A., and C. I. Byrnes.
Output
regulation
Transactions on Automatic [Isi95] [Isi99] [Jan96] Isidori, A. Nonlinear Isidori, A. Nonlinear Jankovic,
Control, volume 35, pages 131140,1990. 3 edition, 1995. London, 1999. Cascade and
SpringerVerlag, TORA
M., D. Fontaine,
P. V. Kokotovic.
Example:
passivitybased
Transactions
ogy, 4(3): 292297,1996. [Jos74] Joseph, B. J., and H. R. Martens. The method of relaxed causality in the bond
graph analysis of nonlinear systems. ASME Journal of Dynamics Systems, Measurement and Control, 96(1): 9599,1974. [Jun01] Junco, S. Lyapunov second method and feedback stabilization graphs. In Proceedings of the ICBGM, [Kar77] Karnopp, 2001. ASME directly on bond
Journal of Dynamics Systems, Measurement and Control, 99(4): 300306,1977. [Kar79] Karnopp, D. C. Bond graphs in control: Physical state variables and observers.
308(3): 219234,1979.
bond graph causal patterns and equation formulataASME Journal of Dynamics Systems, Measurement
144
Bibliography [KarOO] Karnopp, D. C., D. L. Margolis., and R. C. Rosenberg. System Dynamics: and Simulation [Kha92] [Kri95] Khalil, Kristic, of Mechatronic Systems. John Wiley, 3nd edition, 2000. New Jersey, 2 edition, 1992. Nonlinear and Adaptive Control Modeling
H. K. Nonlinear
Systems. MacMillan,
M., I. Kanellakopoulos,
and P. Kokotovic.
73(2): 143173,1997.
D. H. D., Berg, Jordan M., and W. P. Dayawvansa. Loss of structurally implies loss of stability. In IEEE Transactions on Automatic
stable regulation
SpringerVerlag,
Measurement and Control, 124:457463,2002. [Mas92] Maschke, B. M., A. J. van der Schaft, P. C. Breedveld. formulation of network dynamics: Nonstandard An intrinsic Hamiltonian
Transactions
on Circuits
Systems I: Fundamental
42(2): 7382,1995.
[MTT04] MTT.
http: //mtt. sourceforge. net, 2004. [Ngw96] Ngtivompo, R. F., S. Scavarda, and D. Thomasset. Inversion of linear time
invariant siso systems modelled by bond graph. Journal of the Franklin Institute,
333B(2): 157174,1996.
[Ngw99a] Ngwompo, R. F., and S. Scavarda. Dimensioning bicausal bond graphs. Journal of Simulation problems in system design using
1999.
145
Bibliography [Ngwv99b] Ngwompo, R. F., F. Roger, and P. J. Gawthrop. Bond graph based simulation Journal of
2001.
[NgwOlb] Ngwompo, R. F., S. Scavarda, and D. Thomasset. Physical modelbased inversion in control systems design using bond graph representation 2: applications. part 
Proceedings of the IMechE, Part I: Journal of Systems and Control Engineering, 215(12): 105112,2001. [Nij90] Nijmeijer, H., and A. J. van der Schaft. 1990. properties for stabilization of cascaded nonlinear systems. Nonlinear Dynamical Control Systems.
SpringerVerlag, [Ort89]
27: 423424,1989. A., Nicklasson, and P. J., SiraRamirez, Systems. SpringerVerlag, 1998. H. PassivityBased
[Ort98]
Ortega, R., and A. J. van der Schaft. Energy shaping revisited. In IEEE Conference on Control Applications, pages 123126,2000. passivitybased control. In Proceed
[Ort00b]
Control Conference, volume 2, pages 12651270,2000. of underactuated mechanical systems using and NJ,
[OrtOOc]
Workshop on Lagrangian
Methods in Nonlinear
2000.
[Ort01] Ortega, R., A. van der Schaft, I. Mareels, and B. Maschke. Putting in control. [OrtO2a] IEEE Control Systems Magazine, 21(2): 1833,2001. and G. Blankenstein. Stabilisation of a energy back
Transactions
146
Bibliography [Ort02b] Ortega, R., A. J. van der Schaft, and B. Maschke. ing assignment passivitybased Automatica, [Pay6l] 38(4): 585596,2002. Analysis and Design of Engineering Systems. MIT Press, CamInterconnection Hamiltonian and dampsystems.
control of portcontrolled
Paytner, H. M.
bridge, Massachusetts, USA, 1961. [Rob95] Roberts, D. W., D. Ballance, and P. J. Gawvthrop. Design and implementation bond graph observer for robot control. 1457,1995. [Ros7l] Rosenberg, R. C. Statespace formulation for bond graph models of multiport sysof a
tems. ASME Journal of Dynamics Systems, Measurement and Control, 93(1): 3540,1971.
[Ros74]
[Ros87]
[Sch96]
Schaft, A. J. van der, M. Dalsmo, and B. Maschke. the network representation of energyconserving
Mathematical
structures
in
on Decision and Control, pages 201206. IEEE, 1996. [Sch00a] Schaft, A. J. van der. Implicit 418,2000. [SchOOb] Schaft, A. J. van der. SpringerVerlag, [Sep97] 2000. Constructive Nonlinear control. L2gain and Passivity Techniques in Nonlinear Control. portcontrolled Hamiltonian systems. SICE, 39: 410
[Sha9l]
Journal of the Franklin Institute, 328(5/6): 697721,1991. [Son95] Sontag, E. D. On characterizations of the inputtostate stability property. Sys
147
Bibliography [Str98] Stramigioli, S., B. Maschke, and A. J. van der Schaft. Passive output feedback
[van94]
van Dijk, J. On the role of bond graph causality in modelling mechatronic systems. PhD thesis, University of Torente, The Netherlands, 1994.
[Vin03]
Vink,
and V. T. Coppola.
Decision Control, Proc. IEEE Conference In 33rd eccentric on and oscillating rotor. pages 40244029,1994. [Wi172] Willems, J. C. Dissipative dynamical systems, parts I and II. Archive for Rational Mechanics and Analysis, 45: 321393,1972. [Yeh99] Yeh, T. Backstepping control in the physical domain. In Proceedings of the American Control Conference, 1999. [Yeh0l] Yeh, T. J. Controller synthesis using bondgraphs. Control Conference, pages 47654770, Arlington, [Yeh02] Yeh, T. J. Controller In Proceedings of the American 2001.
tional Journal of Systems Science, 33: 11611177,2002. [Zha98] Zhao, J., and K. Kanellakopoulos. disturbance attenuation. Flexible backstepping design for tracking and Control,
International
148