JPH02299002A - Adaptive controller - Google Patents

Abstract

PURPOSE:To easily update and apply a stabilization control gain by using a method which uses a neural network constituted so that outputs of respective units in an intermediate unit layer are inputted even to the intermediate unit layer itself. CONSTITUTION:When the control circuit 4 of a model identification control part 1 outputs a reference signal generation signal S1 to a reference signal generating circuit 5, the circuit 5 inputs the reference signal REF to a controlled system 3 and a model identifying circuit 2. At this time, the system 3 and circuit 2 outputs a response signal RES and a model response signal RESM, whose difference ERR is inputted to the learning/deciding circuit 6 of the circuit 1. The circuit 6 when judging that the signal ERR is sufficiently small outputs a learning end signal S3 to the circuit 4. When it is decided that the signal ERR is still larger, on the other hand, a model correction signal COR is outputted to the circuit 2 and the internal state of the circuit 2 composed of the neural network consisting of plural unit layers is so corrected that the signal ERR becomes small.

Description

[Detailed Description of the Invention] [Objective of the Invention] (Industrial Application Field) The present invention provides control for a controlled system having unknown dynamics by applying a control gain that minimizes a certain evaluation function. In this case, the control gain relates to an adaptive controller that is updated at any time based on the dynamics of the controlled system identified at regular intervals.

(Prior art) The theory of an optimal regulator that determines the control gain so as to minimize the quadratic evaluation function not only makes the closed-loop system containing the optimal controller stable, but also reduces the gain margin and phase margin. is large. That is, it has been highly evaluated as a control method due to its robustness, and has been put into practical use in a wide range of areas.

By the way, in order to configure an optimal controller for a controlled system and obtain the above properties, it is necessary that the dynamics of the controlled system be strictly given as a specific linear differential equation.

For this reason, when the dynamics of the controlled system are actually unknown or when the dynamics change over time, a method is used to identify the dynamics of the controlled system using a controlled object identification method, for example. . Such a method is called a controlled object identification method, and various methods have been proposed in the past.

(Problems to be Solved by the Invention) However, with these conventional control object identification methods, there are limits to the control object systems to which they can be applied, the procedures are complex, and the method has to rely on skilled engineers. In addition, it was difficult to constantly repeat identification online.

The present invention was made to solve such problems, and it constantly performs system identification for various controlled systems with unknown dynamics and updates the stabilization control gain that minimizes a certain evaluation function. The objective is to provide an adaptive controller that can be easily implemented.

[Configuration of the Invention] (Means for Solving the Problems) In order to achieve the above-mentioned objects, the adaptive controller of the present invention has the following features:
For controlled systems with unknown dynamics,
The controlled object identification unit identifies the model at regular intervals, determines a control gain to minimize a predetermined evaluation function for the obtained model, and controls the controlled object system using the control gain. In the adaptive controller, the control object identification section is configured to include a neural network composed of a plurality of unit layers, and the output of each unit of the intermediate unit layer is input to the intermediate unit layer itself as well as the lower unit layer. A model identification circuit configured to send a predetermined reference input signal to the controlled system, record a system response signal from the controlled system to the basic human input signal, and send the model identification circuit to the model identification circuit. The reference human input signal is sent out, a model response signal from the model identification circuit is input to the reference input signal, and the internal state of the model identification circuit is determined based on the error between the model response signal and the system response signal. and a model identification control unit that performs correction to reduce the error, and by repeatedly performing the correction for a plurality of types of the reference input signals, the internal state of the model identification circuit realizes the dynamics of the controlled system. This is how it was done.

(Function) In the present invention, as a control object identification method, a method using a neural network configured such that the output of each unit in the intermediate unit layer is input not only to the lower layer but also to the intermediate unit layer itself is used. Therefore, it is possible to easily realize an adaptive controller that constantly performs system identification for various systems to be controlled having unknown guidance dynamics, and updates and applies a stabilizing control gain that minimizes a certain evaluation function.

(Example) Hereinafter, an example of the present invention will be described in detail based on the drawings.

FIG. 1 is a block diagram for explaining the configuration of a control system employing an adaptive controller according to an embodiment of the present invention.

In the figure, 10 indicates the control system as a whole,
As an operation, two functions are executed in parallel. This can be easily achieved using a multi-CPU configuration. That is, the first function is a control function that controls the controlled system 3, which is the controlled object, by the adaptive controller 11, and the second function is a control function in which the dynamics of the controlled system 3 is identified by the model identification control unit 1, and the model This is a model identification function that forms a model of the controlled system 3 in the identification circuit 2.

In the control system 10, the first function, ie, the control function, continues in real time regardless of whether the second function, ie, the model identification function, is activated or not.

That is, the control human power U output from the adaptive controller 11 is input to the controlled object system 3, and the model identification circuit 2
is input. Here, while the model identification function is in operation, the reference input signal REP output from the model identification control section 1 is superimposed on the control human power U. The controlled system 3 outputs a response output y to the input control human power U, and the response output y is fed back to the adaptive controller 11 .

(Thus, the controlled system 3 and the adaptive controller 11 constitute a closed loop system.

FIG. 2 is a block diagram showing the configuration of the above-mentioned adaptive controller 11 in detail.

As shown in the figure, in the adaptive controller 11, the response output y input from the controlled system 3 is given to the state synthesis circuit 12. The state synthesis circuit 12 estimates the state of the controlled system 3 from the response output y and the model state signal xM output from the state model circuit 13, and inputs the estimation result to the optimal controller 14 as an observed state signal.

Here, the state synthesis circuit 12 obtains the observed state signal M as follows. That is, i out of r-dimensional response output y
(o≦i≦“) dimension is the same as the control target system 3
If it represents the state of , then use this i-dimensional response output y as it is for the i-dimensional part of the observed state signal Ma. Regarding the observed state signal of the remaining dimensions, the state model circuit 13
Apply the model state signal xM from . The state model circuit 13 is composed of parameters (given from the gain configuration circuit 15) that indicate the dynamics of the controlled system identified by the model identification function of the control system 10 at regular intervals, and is connected to the controlled system 3. By applying the same control human power U, a model state signal xM that is considered to represent the state of the controlled system 3 is output.

A control gain built into the optimum controller 14 is given from a gain configuration circuit 15 at regular intervals.

The optimum controller 14 converts the input Ap1 state signal M through the control gain K into control human power U and outputs it.

The gain configuration circuit 15 calculates a stabilizing optimal control gain that minimizes a certain quadratic evaluation function from parameters (φ, B) representing the dynamics of the controlled system 3 obtained from the model identification circuit 2 at regular intervals. At the same time as calculating K, the state model φB of the controlled system 3 is calculated,
The optimal controller 14 and the state model circuit 13 are updated at regular intervals.

Thus, with the configuration of the adaptive controller 11 as described above, adaptive optimal control that can continuously perform optimal control is realized even when the guineamics of the controlled system 3 change.

Here, the theory for realizing adaptive optimal control will be described below.

That is, the dynamics of the controlled system obtained from the model identification circuit 2 is expressed as follows: )(j+1s*φxt+But=(1)Y
Suppose that it is expressed in a discrete form such as t −Cxt −(2). Here, x t+1 (1=0
．． 1.2. ) is an n-dimensional vector representing the state of the controlled system 3 at t=t+iXΔt, ut is an m-dimensional vector representing the input to the controlled system 3, and y
t is an r-dimensional vector representing the output of the controlled system 3, and φ, B, C.

are constant coefficient matrices of nXn%nxm and rXn dimensions, respectively. At this time, the following equation, J-Σ(x' t Qxt +u' t Rut )-
(3) Consider a quadratic evaluation function J expressed by IIO. Here, Q is a semi-fixed nxn dimensional matrix, and R is a fixed mxm dimensional matrix, which are appropriately selected by the operator. Now (φ
, B) is reachable, that is, rank [B, φB−
, φII-IB] = n - (4) and (Q), φ
) is observable, i.e. rank[(Q')', φ'(Q')'
, −−−−・−・・(φ')"(to Q)'
)-n... If (5), the control input 11 t that minimizes the quadratic evaluation function J expressed by equation (3) is unique, and the following equation, ut --Kxt
-(6)K-(R+B'SB)"'B'
SΦ ... is given by (7). Here, K is mX
It is a stabilization control gain expressed by a matrix of n, and S
is the following discrete Riccati equation, S-Q+Φ' SΦ- Φ' SB (R+B'SB)"B' SΦ...(
8) is an nXn matrix representing the only constant symmetric solution of . That is, the adaptive controller 11 has a stabilizing control gain K expressed by equation (7) in the internal adaptive controller 14, and manually controls the input state xt so as to satisfy equation (6). Convert and output.

Incidentally, when the stabilization control gain expressed by equation (7) is used, the control system 10 not only becomes stable as a whole, but also has so-called robustness such as low sensitivity characteristics and a large gain margin. It is known to have excellent properties as a control system.

The model identification function, which is the second function of the control system 10, is activated at regular intervals to update the model identification circuit 2 having the dynamics of the controlled system 3 as a model.

At this time, the model identification control unit 1 causes the model identification circuit 2 to realize the dynamics of the controlled system 3 in the following manner.

That is, in FIG. 3 showing the configuration of the model identification system when the model identification function is operating, the control circuit 4 that starts the model identification function operation at regular intervals receives the reference signal generation signal S.
1 to the reference signal generation circuit 5, the reference signal generation circuit 5 inputs the reference signal REF to the controlled system 3 and also to the model identification circuit 2. At this time, the controlled system 3 and the model identification circuit 2 each receive a response signal I? which is a response to the reference signal REP. ES and model response signal 17ESM, and the response signal R
The difference ERR between ES and model response signal RESM is input to the learning/judgment circuit 6.

When the learning/judgment circuit 6 receives the learning/judgment start signal S2 from the control circuit 4, it evaluates the input error signal ERR. If it is determined that the error signal ERR is sufficiently small, the model identification circuit 2 determines that learning has been completed and outputs a learning completion signal S3 to the control circuit 4. On the other hand, if it is determined that the error signal ERR is still large, the model correction signal COR is output to the model identification circuit 2,
For example, the internal state of the model identification circuit 2 is corrected using a backpropagation method.

The model identification circuit 2 is constituted by a neural network as shown in FIG. 4, for example.

That is, the input layer has m neurons 11, ν12°...
・、Si1. m, n neurons in the middle layer 21. ν2
2,..., shi2n, one new range 31 in the output layer
．． In a three-layer network with ν32, ..., ν3rs, the m-dimensional learning reference signal REPX(rr +
, =-, rf'i.

..., rf'a), the r-dimensional model response signal RESM (031, -=, 03, -, 03r
) is output.

The input layer neuron 11 (1≦l≦I) outputs the manually generated learning reference signal rf'l as it is 01j01i=r
fi (1≦I 5m) .=(9) and sends it to the intermediate layer new engines 21 ν 22 . . . , ν 2 n.

The middle layer neuron 2j (1≦j≦mouth) is the input net
As 2j, the receiver takes the following equation, (1≦j≦n)...(10).

In equation (10), W2jli (1≦i≦m) is the weight for the output 011 of the neuron logic 11 in the input layer, and W2j2 (1≦j≦n) is the weight for the output 011 of the neuron logic 11 in the input layer.
This is the weight for the output 021 of 1.

Generally, since W2j21≠0, the intermediate layer neuron 2j (1≦j≦n) receives as input not only the input layer but also the output of the intermediate layer including itself. In addition, the output 02j (
1≦j≦n) is the following formula, 02j-f 2j(net2N
=・ It is expressed as (11). Here, f 2j (·) represents, for example, the s1gmoid function.

The output layer neuron 3k (1≦to≦“) is the input net
3 as the following formula, net3 is one ΣW3 is 2j02j (1≦≦r)−(
12) 3 for I to receive.

In Equation (12), 2j (1≦j≦n) in W3 is a weight for the output 02j of the new logic 2j in the intermediate layer. -Output 03k (l≦
≦r) is the following formula, 03k −f3k(net3k) −(
1B). Here, f3k(・) is, for example, 51
This is the g1oid function.

Now, vector x, x”, us ySF2 (・)
, F3 (·) and the matrix φ, BSC are defined as follows.

x = [021, -102j, -..., 02n)'
eR"...(14) (Hereafter, margin) X" 1ml+ [net 21. -1net 2j, -...,ne
t 2n] ' E Rn... (15) u = [rf'L,..., rl'i...
・, r['m]'-REFIER"
-(16)y -[031,-...
,03,-...,03r]'-RESM(Rr
... (17) F2
(・) − [f21(-), −f2j(-) −・・
, f 2n (・) ko ′... (
18) F3 (・) − [fll(−), −=f3k(−) ・・・
, f 3r (・) ko ′...
(19) (hereinafter referred to as margin) (hereinafter referred to as margin) At this time, the following relationship is obtained. However, X5
The F suffix t attached to uSY indicates the value at time t, and similarly, t+1 indicates the time 10 Δt (Δt is a constant)
shows the value in

X11t+1-mφXi +I3 ut
+++ (23)xt −F2 (x' t)
−(24)y t −F 3 (Cx
If t)-(25)F2 (.) and F3 (.) are linear functions, equations (23), (24), and (25) are equivalent to state equations of a discrete-time system. F2 (・)
, F3 (·), for example, if the 51gff1oid function is selected, it can be assumed that it is sufficiently linear as long as it is close to the bias point, so the model form circuit 2 can eventually become a model of a discrete time system.

That is, by giving the reference signal REF to the model identification circuit 2 in time series, the response signal RES can be obtained in time series, and by learning using the pack propagation method, for example, model identification can be performed. The internal state of the circuit 2 will realize the dynamics of a discrete-time system corresponding to the reference signal REF and the response signal RES. It seems appropriate to use an M-sequence function suitable for expressing the system as the reference signal REF.

Here, the elements that determine the internal state of the model identification circuit 2 are matrices expressed by equations (18) to (22), but these are the model correction signals COI? sent from the learning/judgment circuit 6. amended based on. In backpropagation, which is a typical correction method, the model correction signal C
ot? represents the squared error.

Incidentally, it is known that a neural network with the configuration shown in Figure 4 can represent a wider range of systems than a feedforward neural network, and it is also possible to achieve the same result using a feedforward neural network. However, it is known that a large number of neurons, which are constituent units, are required.

As described above, according to the present invention, as a controlled object identification method, a method using a neural network configured such that the output of each unit in the intermediate unit layer is input not only to the lower layer but also to the intermediate unit layer itself is used. This enables adaptive control that constantly identifies the controlled object and uses the identified model to apply a stabilizing control gain that minimizes a quadratic constraint function to various controlled systems with unknown dynamics. can be easily realized.

In the above embodiment, the combination of the response output y and the model state signal xM of the state model circuit 13 was used as the observed state signal of the controlled system 3 input to the optimal controller 14. Alternatively, the state X may be estimated based on the response output y and the control human power U using an observer.

[Effects of the Invention] As described above, according to the present invention, as a control object identification method, a neural network configured such that the output of each unit in the intermediate unit layer is manually input not only to the lower layer but also to the intermediate unit layer itself is used. This method makes it easy to create an adaptive controller that constantly performs system identification for various controlled systems with unknown dynamics and updates and applies stabilizing control gains that minimize a certain evaluation function. It is possible.

[Brief explanation of the drawing]

Figure 1 is a block diagram showing the overall configuration of a control system to which the optimal controller according to the present invention is applied, Figure 2 is a block diagram showing the configuration of the control system in model identification mode, and Figure 3 explains the model identification function. FIG. 4 is a block diagram showing the detailed structure of the model identification circuit. 1...Model identification control unit 2...Model identification circuit 3...Controlled system 10...Control system 11...Adaptive controller 12...State synthesis circuit Applicant Toshiba Corporation Attorney Patent attorney Satoshi Suyama - Figure 1

Claims (1)

[Claims] For a controlled system having unknown dynamics,
The controlled object identification unit identifies the model at regular intervals, determines a control gain to minimize a predetermined evaluation function for the obtained model, and controls the controlled object system using the control gain. In the adaptive controller, the control object identification section is configured to include a neural network composed of a plurality of unit layers, and the output of each unit of the intermediate unit layer is input to the intermediate unit layer itself as well as the lower unit layer. a model identification circuit configured to send a predetermined reference input signal to the controlled system, store a system response signal from the controlled system to the reference input signal, and send the model identification circuit to the model identification circuit; Sending out a reference input signal, inputting a model response signal from the model identification circuit in response to the reference input signal, and determining the internal state of the model identification circuit based on the error between the model response signal and the system response signal. and a model identification control unit that performs correction so that the reference input signal decreases, and by repeatedly performing the correction on a plurality of types of the reference input signals, the internal state of the model identification circuit realizes the dynamics of the controlled system. An adaptive controller characterized by: