Академический Документы
Профессиональный Документы
Культура Документы
152
World Academy of Science, Engineering and Technology 1 2005
basic idea of the accumulated GA is to learn the fuzzy inference based on the error e(k) and the error rate 'e(k). GS_FPID
system rule by rule until the learning error is reduced to a range controller can thus be considered as an adaptive controller with
of tolerance. The parameters learned in the previous rules are gain scheduling. It continuously adjusts the gains of fuzzy PD
accumulated and updated along with the parameters in the and PI controller based on error and error rate. The gain
current rule. scheduling mechanism is designed to be the one with fuzzy
inference scheme.
II. GAIN SCHEDULED FUZZY PID CONTROLLER For TSK model, the output of FLC in (2)-(4) is calculated
The structure of proposed GS_FPID controller consists of as:
m
both fuzzy PI controller and fuzzy PD controller. Both of fuzzy
controllers are weighted through scheduled gains, which are
D( k ) ¦ rDi ( ai1e( k ) ai 2 'e( k )), (9)
i 1
also determined by fuzzy logic inference. At each sampling m
instance k, the model of proposed GS_FPID controller is
defined as
E( k ) ¦ rEi ( ai3 e( k ) ai4 'e( k )), (10)
i 1
u(k) = D(k)'u(k) + E(k)u(k-1) (1) m
where D(k), E(k) and 'u(k) are defined by fuzzy logic inference. 'u( k ) ¦ rui ( ai5 e( k ) ai6 'e( k )), (11)
As shown in Fig. 1, given that the reference input is denoted by i 1
r(k), the error is defined as e(k) = r(k) – c(k) and the error rate where
'e(k) = e(k) – e(k-1). The gains D(k), E(k) and the increments of W Di
rDi m
, (12)
controller output 'u(k) can be respectively determined by the
following fuzzy rules: ¦W D
i 1
i
R1: if e(k) is Ai1 and 'e(k) is Ai2
then D(k) = ai1e(k)+ai2'e(k); (2) W Ei
rEi m
, (13)
R2: if e(k) is Ai3 and 'e(k) is Ai4
then E(k) = ai3e(k)+ai4'e(k); (3) ¦W E
i 1
i
j = 1…6, are fuzzy sets for e(k) and 'e(k), respectively. For a
The degrees of firing (DOF) in (12)-(14) are respectively
fuzzy PD controller, the controller output is defined as
defined as:
u(k) = 'upd(k) (5)
W Di Ai1 ( e( k )) Ai 2 ( 'e( k )), (15)
where 'upd(k) is inferred by fuzzy logic as in (4) based on e(k)
and 'e(k). The output of a fuzzy PI controller can be defined as W Ei Ai 3 ( e( k )) Ai 4 ( 'e( k )), (16)
u(k) = u(k-1) + 'upi(k) (6) W ui Ai 5 ( e( k )) Ai6 ( 'e( k )), (17)
where 'upi(k) is inferred by fuzzy logic as in (4) based on e(k) where the notation denotes the operation of minimization.
and 'e(k). In addition to improving system’s relative stability, The FLCs in (9)-(11) can be viewed as PI-like controllers
fuzzy PD controller is utilized for increasing the speed of since
response while fuzzy PI controller aims to reduce steady state D ( k ) k PD e( k ) k ID 'e( k ) , (18)
error. Therefore, if the GS_FPID controller is designed to be a
composite of both fuzzy PD and PI controllers, it has both E ( k ) k PE e( k ) k IE 'e( k ) , (19)
functionalities of these two fuzzy controllers. The proposed 'u( k ) k Pu e( k ) k Iu 'e( k ) , (20)
fuzzy PID controller is essentially the weighting adjusted fuzzy m
PD and PI controller. Therefore,
u(k) = D1(k)'upd(k) + E (k)(u(k-1) + 'upi(k))
with k PD ¦ rD a
i 1
i i1 , (21)
153
World Academy of Science, Engineering and Technology 1 2005
m fuzzy rule. In fact, the first and second fuzzy rules are both
k Iu ¦r ui ai 6 . (26) learned by GA in the same learning process. In order not to
i 1 waste the efforts that has been spent on learning the previous
Substituting (18)-(20) into (1) yields fuzzy rule, the parameters of previously learned fuzzy rule are
u( k ) (k PD e( k ) k ID 'e( k ))( k Pu e( k ) k Iu 'e( k )) taken as the initial conditions for the GA so that the learning
(k PE e( k ) k IE 'e( k ))u( k 1 ) process can be greatly improved. Similarly, in the third
learning process, the parameters to be learned include the ones
k PD k Pu e 2 ( k ) ( k PD k Iu k Pu k ID )e( k )'e( k ) (27) in the first, second and third rules. As the learning processes go
k ID k Iu ( 'e( k ))2 k PE e( k )u( k 1 ) on, the learning error in each learning process is a decreasing
sequence given that the GA in each learning process converges.
k IE 'e( k )u( k 1 ). Denote the GA applied in the i-th learning process by GAi
E( k )
and assume that it takes Gi generations for GAi to converge.
E(k)
Furthermore, assume that L chromosomes are employed in
z 1
every generation of GA. Let dijg be the parameters of the i-th
e(k ) 'u (k ) y( t ) fuzzy rule to be learned, which is implemented as part of the
z 1 'u (k ) u (k ) g-th chromosome of GA’s gene pool in the j-th generation, g =
'e(k )
1…L, j = 1…Gi. That is to say, dijg is the additional set of
D (k )
parameters to be learned in GAi. If the set Dijg denotes GAi’s
D (k )
whole set of parameters implemented in the g-th chromosome
of the j-th generation, then the total number of parameters to
learned is accumulated process by process, i.e.,
Fig. 1 Structure of GS_FPID controller Dijg = [D(i-1)jg, dijg], g = 1…L, j =1…Gi. (28)
It is shown in (27) that the output of GS_FPID controller, u(k), Note that for the initial learning process, D1jg = d1jg. Referring
p p
is no longer like a traditional PID controller, which is a linear to (2)-(4), let Lijg and Cijg respectively be the width and center
function of e(k), 'e(k) and u(k-1). Instead, it is a complex p
of triangular membership functions of fuzzy sets Aip , a ijg be
recurrent function of e2(k), e(k)'e(k), ('e(k))2, e(k)u(k-1) and
'e(k)u(k-1). Referring to (21)-(26), GS_FPID thus has more the p-th parameter in the consequent of fuzzy rule, which are
degrees of freedom adapting with these influences to the implemented in the g-th chromosome of GAi’s j-th generation,
controller output. With the additional degrees of freedom, it p = 1…6. Then, dijg can be defined as
will be shown in this paper that GS_FPID performs well for dijg = [ L1ijg ,..., L6ijg ,C ijg
1 6
,..., C ijg 1
, aijg 6
,..., aijg ]. (29)
controlling the higher order systems as well as the systems with Let y( k ; Dijg ) be the sampled system output due to the
time delays.
fuzzy PID controller associated with Dijg. For GAi, the fitness
III. LEARNING OF GS_FPID CONTORLLER function associated with the chromosome Dijg is defined as the
integration of square error (ISE):
The membership functions of antecedent fuzzy set Aij, i = N1
1…m, j = 1…6, for input variables e(k) or 'e(k) of FLC in f ijg ¦ U1 ( r( k ) y( k ; Dijg ))2
(2)-(4) are defined as symmetric triangular functions k 1
(30)
P Aij ( x j ; Lij ,Cij ) , of which Lij and Cij respectively denotes the N
2
width and center of the triangular membership function. For
¦ U 2 ( r( k ) y( k ; Dijg ))
k N 1 1
GS_FPID controller, GA is applied to learn the parameters of where N denotes the total number of samples to be evaluated.
input membership functions as well as the coefficients of the In order to make the designed GS_FPID controller hold good
first order FIR filter in the consequent of each fuzzy rule. The transient response, (30) allows the flexibility of assigning the
parameters to be learned are expressed as binary strings and first N1 samples different weightings from the rest of the
cascaded as a chromosome. The number of fuzzy rules in FLC samples. Within N samples in total, N1 samples are given the
for D, E and 'u are assumed to be the same in GS_FPID weighting U1 while the rest of them are given the weighting U2.
controller. However, the number of fuzzy rules is not known a The accumulated GA tends to adjust the parameters of
priori. In order to learn the number of fuzzy rules as well as the GS_FPID controller minimizing the ISE. The best
parameterizations of each fuzzy rule, the fuzzy rule is chromosome learned by GA is determined by searching all
iteratively learned and accumulated. In other words, a single possible chromosomes minimizing fijg. The elitist method is
fuzzy rule for D, E and 'u, respectively, is first utilized. GA is utilized to choose parent chromosomes to mate and generate
then applied to learn the parameterizations in both antecedent children chromosomes. In order to increase GA’s searching
and consequent parts of the rule. As the learning of GA efficiency, the parameters learned in the previous learning
converges, a second rule is added based on the previously
process are given as the initial conditions. If Di*1 denotes the
learned fuzzy rule continuing the learning process. The second
learning process is not restricted to simply learn the second
154
World Academy of Science, Engineering and Technology 1 2005
best chromosome learned by GAi-1, then the best chromosome e(·) and 'e(·), GS_FPID controller adjusts controller
in the j-th generation of GAi can be determined as parameters by fuzzy inference.
Arg min Assume that the variations of system dynamics with
Dij* (f ) * (31) operating conditions are known. If q samples of variations are
Dijg , g 1..L ijg Di 1
taken, then, with n = 1…q, the system outputs can be
In order to guarantee the fitness function can be reduced
represented as
generation by generation, the best chromosome in the current
generation is passed to the next generation. With the elitist y n ( t ) d ( x(t),u( t ),t ) 'd n ( x(t),u( t ),t ), (33)
method for choosing parent chromosomes, the variations of where x(·) is the vector of state variables, u(·) is the input, d ( )
fitness values for GAi is a non-increasing sequence. If is a fixed system function and 'd n ( ) is due to the n-th
f ij* denotes the fitness function associated with Dij* , the best variation of system dynamics. With self tuning of fuzzy gains
chromosome determined by GAi can thus be determined as D(k) and E(k), robustness is embedded in the GS_FPID
Arg min controller. As the dynamics of controlled system varies,
Di* * ( f* ) (32) GS_FPID controller is able to accommodate the variations of
Dij , j 1..Gi ij
system dynamics with self tuning of fuzzy gains D(k) and E(k).
Since the sequence of f ij* is a nonincreasing sequence, Di* is In order to learn the GS_FPID controller for gain scheduling,
the best chromosome learned in the final generation, the fitness function in (30) can be modified accommodating the
i.e., D*i D*iG i . In (31), although 6ui parameters are to be variations of system dynamics. Let y n ( k ; Dijg ) be the sampled
learned in the i-th learning process, 6u(i-1) best parameters system output due to n-th variation of system dynamics in
association with the parameters of GS_FPID controller, Dijg.
learned in the previous learning process, denoted by Di*1 , can
Similar to (30), the fitness function associated with the
be taken as the initial values in the initial generation of GAi. In chromosome Dijg can be modified as
addition, 6u(i-1) parameters in each of the other (L-1) q N1
chromosomes in the initial generation of GAi can also be
*
f ijg ¦ U1 ( ¦ ( r( k ) y n ( k ; Dijg ))2
obtained by mating Di n 1 k 1
1 with the other randomly generated (34)
N
chromosomes. 2
In order to determine the appropriate number of fuzzy rules
U2 ¦ ( r( k ) y n ( k ; Dijg )) )
k N 1 1
for D, E and 'u, GAs with increasing size of chromosomes are where q denotes the number of known system variations. The
iteratively applied to learn the parameters in each fuzzy rule. best chromosome minimizing (34) by GAi is obtained as in (31)
For each learning process, a new fuzzy rule is added associated and (32). Referring to (34), the best GS_FPID controller
with the previously learned fuzzy rules. Since the sequence learned by the accumulated GA is able to adaptively change the
f ij* , j = 1…Gi, is a non-increasing sequence, if the set of gains adjusting the weightings of fuzzy PI and fuzzy PD
controllers to accommodate the variations of system dynamics.
parameters Dij* cannot further reduce the fitness function, the
convergence of fitness values will stagnate. Within each GA, V. COMPUTER SIMULATION
the convergence of fitness values is monitored. If the fitness In this section, the accumulated GA are iteratively applied
values do not change for certain number of generations, and the to design the GS_FPID controller for the higher order, variable
fitness value is still larger than a threshold, a new GA is structure or time delayed system. The fitness function is
applied. designed in the form of ISE. For each GA, 80 chromosomes
are utilized in each generation. For the fitness function in (30),
IV. ROBUSTNESS OF GS_FPID CONTROLLER N1 = 50, U1 = 5, U2 = 1;
In some applications, it is known a priori how the
dynamics of a controlled system change with the operating Example 1. Assume that the system to be controlled is with
conditions of the system. The controller can thus be designed variable system dynamics as following:
to change the parameters accommodating the variations of 2s 1
M (s) 2
, (36)
system dynamics based on the measurements of system s as 1
operating conditions. This type of feedback control scheme is where a = 0.9, 1.2, and 1.5 is known a priori. The sampling
called gain scheduling.[8] The proposed GS_FPID controller interval is set to be 0.01 second and 1500 samples are taken for
can also be applied to solving the gain scheduling problems. In evaluation in the fitness function. The step responses due to
order to design suitable scheduling parameters based on system GS_FPID controller with 2 fuzzy rules are shown in Fig. 2. It is
operating conditions, traditional controller with gain shown in Fig. 2 that the GS_FPID controller is able to
scheduling requires system dynamics. However, GS_FPID accommodate the variations of system dynamics.
requires no information of system dynamics. The operating
conditions for the GS_FPID controller simply come from the Example 2. Assume that the system to be controlled is with
measurements of error e(·) and error rate 'e(·). On the basis of high order time delay variable system dynamics as following:
155
World Academy of Science, Engineering and Technology 1 2005
156