Академический Документы
Профессиональный Документы
Культура Документы
ITI GmbH
Webergasse 1
01067 Dresden, Germany
{andreas.abel,tobias.naehring}@iti.de
1 Introduction
2 Periodic Steady-State Simulation
In many engineering disciplines frequency-domain
methods play an important role. Powertrain engi- 2.1 Application to Modelica Models
neers for instance not only exploit transient simula-
tions, but to a large extend assess the behaviour of The main area of application for the nonlinear peri-
their systems based on the natural frequencies, the odic steady-state simulation in SimulationX is the
resulting vibration models, and also in terms of vibration analysis of powertrains.
steady state results, which show vibrations under The example Modelica model in Fig. 1 is an adaption
stationary conditions resulting from the uneven and from [4] p. 246 with some added damping and cylin-
multi-order excitation of the driveline by the engine. ders including oscillating masses and driven by some
Other engineering domains and tasks also require typical combustion engine cylinder pressure.
frequency-domain approaches.
The Modelica Association 383 Modelica 2008, March 3rd − 4th , 2008
A. Abel, T. Nähring
The steady-state for a range of mean rotational the calculated harmonic torque components in the
speeds of the engine has to be computed. The oscilla- mass-damper spring are shown in Fig. 2.
tion time period is determined by the engine speed In SimulationX the initial conditions corresponding
and the periodicity of the excitation over the crank to the results of the periodic steady-state simulation
angle. can be calculated and used to initialize a successive
Engine Damper transient simulation.
4-Cylinder Engine
Fly Wheel
The Modelica Association 384 Modelica 2008, March 3rd − 4th , 2008
Frequency-Domain Analysis Methods for Modelica Models
The Modelica Association 385 Modelica 2008, March 3rd − 4th , 2008
A. Abel, T. Nähring
where xˆ[k ] is the k -th frequency component of the shifting f to the right-hand side (14) is transformed
state space vector (we use a hat x̂ or ( x ) to denote
^
by the FFT into the equation
complex amplitudes). Since ~ x is real xˆ[k ] is the ((∂ 1 ) (
f ) * δ xˆ + j ω (∂ 2 f ) * (Iδ xˆ ) +
^ ^
) (16)
complex conjugate of xˆ[ − k ] . Thus, the values of x̂ + (∂ 2 f ⋅x ' ) δω + (∂ 3 f ) δ xC = − fˆ
^ ^
The Modelica Association 386 Modelica 2008, March 3rd − 4th , 2008
Frequency-Domain Analysis Methods for Modelica Models
for the unknown Newton correction (δ xˆ , δω, δ xC ) size is implemented in SimulationX. A short outline
in the frequency domain. Together with (9) and (10) of this algorithm shall conclude this subsection.
written as Only at the starting value rStart and the end
δ xˆi P [0] = 0; δ xˆi R [0] = 0 (17) value rStop of the interval for the reference quan-
this system formally determines the Newton correc- tity xiR the full system (13) is solved. At intermedi-
tion in the frequency domain completely. ate points for xiR the last equation determining the
With the number of n (2 N + 1) + 2 real unknowns value of the reference quantity is removed resulting
the system is rather large and the convolution opera- in
tor in (16) causes large fill-in of the system matrix
making direct solving infeasible in real-world appli- X F(X ) = 0
cations. Therefore, the iterative GMRES algorithm is
used instead (see e.g. [5]). This method only requires F ( X ) = F (X P ( k ) )
δX ⊥(k )
the evaluation of the left-hand side of (16) for X P( k )
known (δ xˆ , δω, δ xC ) . This also makes it possible X (k )
to replace the frequency-domain convolutions in (16) δX ||(k −1)
by the cheaper corresponding time-domain products X ( k −1) xiR
in (14) (together with the therefore needed FFT- rStart rStop
operations). GMRES only works well with an appro-
priate pre-conditioner. Thus, one must be able to Fig. 4: Curve tracing algorithm (see text for details)
roughly solve systems with the left-hand side of (16) ⎛ fˆ ( xˆ , ω , xC )⎞
fast. For this end the block-diagonal preconditioner F ( X ) = 0 with F ( X ) := ⎜⎜ ⎟
⎟ (19)
is used (see e.g. [6]). This approximates the convolu- ⎝ xˆ iP [ 0 ] ⎠
tions by only retaining the mean value component and with the unknowns collected in X := ( xˆ , ω , xC ) .
of (∂ k f )^ : Since (19) has one scalar equation less than un-
((∂1 f )^ * δ xˆ )[k ] ≈ (∂1 f )^ [0] ⋅ δ xˆ[k ] knowns it formally defines a solution curve (see also
((∂ 2 f )^ * ( Iδ xˆ ))[k ] ≈ (∂ 2 f )^ [0] ⋅ kδ xˆ[k ] (18) upper branch in Fig. 4) instead of a single point.
Given the last solution point X (k −1) on the solution
The so approximated system (16) can be solved fre-
quency-component wise. curve and the tangent direction δ X ||(k −1) of the solu-
If the dynamical system is linear then the Jacobi- tion curve in that point a prediction
ans ∂ 1 f , ∂ 2 f are constant in time and the corre- X P (k ) = X (k ) + s δ X ||(k −1)
sponding higher spectral components in the convolu- for the new solution point is computed. Thereby, the
tions (e.g. (∂ 1 f ) [k − l ] with k − l ≠ 0 in (15)) are
^
step size s is chosen in dependence of the estimated
zero. In this case ` ≈ ´ in (18) can be replaced by `=´ curvature of the solution path, the estimated distance
and the approximations are exact. For increasing of X P (k ) to the solution path, and the local conver-
nonlinearities the higher spectral components of gence behaviour of Newton's algorithm (for details
∂ 1 f , ∂ 2 f omitted in the preconditioner gain influ- see [2]). In the predicted point a new estima-
ence, the approximations become more coarse. In tion δ X ⊥(k ) for the tangent vector is computed as
general one can say that with stronger nonlinearities the solution of the system
D F (X P (k ) )δ X ⊥( k ) = 0,
the number of GMRES iterations per Newton step
and the number of Newton-iterations increase.
If the local Newton method does not converge fast (δ X ( ) )
|| k −1 T
⋅ δ X ⊥(k ) = 1.
enough then the Newton-algorithm with backward- This is not the tangent direction to the solution curve
error minimization via backtracking (see [1] and [7])
is applied. For a better numerical condition the states
but to the curve defined by F ( X ) = F X P (k ) (see ( )
Fig. 4). Nevertheless, these curves and their tangents
are automatically scaled during the computation.
are supposed to be close to each other. The Newton
In section 2.3 we will give an example of a nonlinear correction for the computation of the next solu-
system with a turning point in its frequency re-
tion X (k ) of (19) is then carried out in the affine
sponse. To make the computation of such points pos-
sible a curve tracing algorithm with variable step- plane with X P (k ) as origin and δ X ⊥(k ) as normal
The Modelica Association 387 Modelica 2008, March 3rd − 4th , 2008
A. Abel, T. Nähring
5
turning points in the frequency characteristic. The
curve tracing algorithm implemented in SimulationX
makes the computation of such kind of frequency 0
characteristics possible.
−5
The simple mechanical system of Fig. 5 is a torque
excited spring-mass-oscillator. The frequency of the
sinusoidal torque source is chosen as the reference −10
quantity and swept between 0.2 Hz and 0.7 Hz . 0 60 120 180 240 300 360
Since this reference quantity is a parameter and not a Phase of Oscillation / degree
variable SimulationX chooses it automatically as Fig. 7: Angular speed curves for the two possible stable
compensation parameter as well. The phase of the limit cycles of the nonlinear spring-mass-system at
sine oscillator is the period variable with period 2π . excitation frequency 0.405 Hz represented over
The quadratic term added to the spring characteristic phase.
makes the system nonlinear in such a way that it We kept this example simple to demonstrate that
shows a turning point in the frequency characteristic even very basic nonlinear systems may have fre-
(see Fig. 6). quency responses with turning-points. More compli-
sqr
cated examples can be found in [8], and [9].
tau
to mechanical systems. As an example the periodic
amplitude=1
J=1
d=0.1 steady state simulation is applied to a Modelica
Fig. 5: Nonlinear Spring-Mass-system model for an active electronic pass-band filter (see
Fig. 8). The reference and compensation quantity in
In the interval from 0.397 Hz to 0.426 Hz the fre-
this example is the frequency of the sinusoidal
quency characteristic is multi-valued. That corre-
source vin and its phase is the phase variable.
sponds to multiple periodic limit cycles at those exci-
tation frequencies.
The Modelica Association 388 Modelica 2008, March 3rd − 4th , 2008
Frequency-Domain Analysis Methods for Modelica Models
rc1 6
vc1
C=6.4e-9
R=250e3
rb1
c2
V R=20e3 5
q2
power…
+
R=2.5e3 C=6.4e-9
-
R=1e3
3
re2
vin
The Modelica Association 389 Modelica 2008, March 3rd − 4th , 2008
A. Abel, T. Nähring
energy calculation block which SimulationX uses in earized system. Any result variable of the model may
order to compute the energy distribution. be used as the system output. SimulationX has a spe-
cial class of signal inputs that may be open even for
model RotInertiaEnergyBlock the top-level model. Those inputs may be used for
the input-output-analysis. In Fig. 11 a cut-out of the
import M=Modelica.Mechanics;
powertrain from Fig. 1 is shown where a torque
extends M.Rotational.Inertia; source with such an input has been added. The input-
Mechanics.Rotation.CalcEnergyBlock eb; output behaviour is described by the frequency re-
equation sponse function and the pole-zero diagram of the
eb.dom = w; system.
eb.T = J*a;
end RotInertiaEnergyBlock;
The modification of the Type SpringDamper is
similar. For a demonstration the (rotational and
translational) masses and spring-dampers in the
powertrain from Fig. 1 have been substituted by the
modified types. The distribution of energy calculated Fig. 12: Element linSysAnaOpenInput in the example
by SimulationX for the eigenmode at 1.6664 Hz is from Fig. 1 with open input for the input-output-
shown in Fig. 12. In practical applications such rep- analysis
resentations show the engineer which masses, Fig. 13 and Fig. 14 show the pole-zero plot and the
springs, and dampers dominate the behaviour in cer- frequency characteristic, resp., for the powertrain
tain eigenmodes of the system, so he can take sys- from Fig. 1 with the torque at the first cylinder as
tematic countermeasures to avoid unwanted oscilla- input (Fig. 12) and the torque in the engine damper
tions. as output.
Up to now these blocks are not documented and only
used for the internal element libraries of Simula-
tionX. But this may change in future.
The Modelica Association 390 Modelica 2008, March 3rd − 4th , 2008
Frequency-Domain Analysis Methods for Modelica Models
physical domains since it is purely equation-based. • Autonomous Systems: The ansatz used for the
Only very little knowledge of the system is required harmonic balance also bears the potential for the
from the user. Two mechanical examples and one simulation of autonomous systems. The required
from electronics were given in the paper. randomization of the start values for the harmonic
balance could be implemented.
• Detection of Stable/Unstable Limit Cycles: Up to
now there is no automatic discrimination of the sta-
ble and unstable branches in the nonlinear frequency
response computed via harmonic balance. This can
be implemented by an eigenvalue analysis of the
monodromy matrix of the computed limit cycles.
References
[1] J. E. Jr. Dennis and Robert B. Schnabel:
Numerical Methods for Unconstrained Opti-
mization and Nonlinear Equations. SIAM
1996.
Fig. 14: Frequency response of the system in Fig. 1; [2] E. L. Allgower and K. Georg: Numerical
top: amplitude, bottom: phase Continuation Methods: An Introduction.
Furthermore, we discussed methods for the small- Springer-Verlag, 1990.
signal analysis in the current operating point (result- [3] http://www.simulationx.com
ing from a transient or equilibrium computation). [4] H. Dresig and F. Holzweißig: Maschinendy-
Beside pole-zero plots and frequency response func- namik. 5th ed., Springer-Verlag Berlin, 2004.
tions also some remarks about the deviation- and
[5] A. Meister: Numerik linearer Gleichungssys-
energy distribution analysis for oscillation modes
teme. Vieweg-Verlag, Wiesbaden, 2005.
were given. They are especially useful for the me-
chanical engineer to detect the powertrain elements [6] Ognen J. Nastov: Methods for Circuit Analy-
which participate in selected oscillation modes. sis. PHD-theses, Massachusetts Institute of
Technology, 1999.
• Behaviour Description In Frequency Domain:
In future it is planned to include a behavioural de- [7] U. Feldmann, U. A. Wever, Q. Zheng, R.
scription in frequency domain (e.g., for modeling of Schultz, and H. Wriedt: Algorithms for Mod-
dynamic stiffness) for the periodic steady state simu- ern Circuit Simulation. AEÜ, Vol. 46 (1992),
lation as well as for the frequency response computa- No. 4.
tion, which was one main argument for the harmonic [8] A. Al-shyyab and A. Kahraman: Non-linear
balance method to be preferred over the shooting dynamic analysis of a multi-mesh gear train
method (see e.g. [11] for a short introduction and using multi-term harmonic balance method:
further references). One major reason for the fre- period-one motions. Journal of Sound and
quency domain description not yet being imple- Vibration, 284 (2005) 151-172.
mented in SimulationX is that Modelica currently [9] Wen-I Liao, Tsung-Jen Teng, and Chau-
still lacks a standardized way for computations with Shioung Yeh: A method for the response of
complex numbers (even if some steps in this direc- an elastic half-space to moving sub-Rayleigh
tion have already been taken, see e.g. [10]). point loads. Journal of Sound and Vibration
• Event Iterations: Event iterations are already em- 284 (2005) 173-188.
bedded into the harmonic balance algorithm. But [10] Peter Aronsson at al.: Meta Programming
there remains still some work for the treatment of and Function Overloading in OpenModelica.
time-discrete variables in special cases. Modelica 2003, November 3-4, 2003.
• Improved Convergence for Strongly Nonlinear [11] Kenneth S. Kundert: Introduction to RF
Systems: As long-term objective the convergence Simulation and Its Application. IEEE Journal
speed of the harmonic balance for strongly nonlinear of Solid-State Circuits, Vol. 34, No. 9, Sep-
systems can be improved by time domain precondi- tember 1999.
tioners (see [6]).
The Modelica Association 391 Modelica 2008, March 3rd − 4th , 2008