JPS6280704A - Adaptive control system - Google Patents

Abstract

PURPOSE:To obtain high estimation accuracy even if the number of data is small by estimating a parameter relating to a controlled system on the basis of LXN data strings obtained by storing L sample value data and then repeatedly reading out the stored data N times. CONSTITUTION:The finite number of observation data with L length are once stored in a storage device and the stored data are connected N times in the time direction to regard the connected data as one observation data. Ordinary sequential identification algorithm is applied to the observation data. The length L of the data and the number N of connection times of data (the number of repeated times) are determined by the rate of change in characteristics of the controlled system, the amplitude of disturbance and information indicating the degree of continuous excitation of an observation signal to be used as data, i.e. information indicating the width of a frequency band included in the signal.

Description

[Detailed Description of the Invention] [Technical Field of the Invention] The present invention constantly improves control performance by identifying the characteristics of a controlled object whose dynamic characteristics change and adjusting control law parameters based on the results. This paper relates to an adaptive control method that can maintain the optimum performance.

[Technical background of the invention and its problems]

Most adaptive algorithms in conventional adaptive control systems are essentially a combination of a sequential identification algorithm and a control parameter calculation algorithm. Sequential identification algorithms identify models related to the dynamic characteristics of a controlled object in real time from input/output signals, target value signals, etc. of the controlled object, and various types have already been proposed, mainly based on the sequential least squares method. ing.

These Alf+J rhythms were relatively easy and I had notes. That is, compared to a patch-type identification algorithm (hereinafter referred to as an "offline identification algorithm") that processes observation data all at once, the influence of initial errors is large and the estimation accuracy is inferior. This tendency is particularly noticeable when there is little observation data.

Furthermore, for the same reason, the ability of the identification parameters to follow changes in the characteristics of a fixed object is poor, and special measures have been required to improve this.

Furthermore, in these sequential type Al prisms, when unknown disturbances are included in the observation data, the identification parameters fluctuate greatly transiently. Therefore, for example, in a model-normative adaptive control system, which is a typical type of adaptive control, the control parameters are adjusted every time the parameters are updated. The drawback was that the signal was abnormally disturbed. Therefore, it has been difficult to apply adaptive control using sequential identification Alco9 rhythm to a control target of a process system that has many sudden disturbances.

[Purpose of the invention]

The present invention improves the numerous shortcomings of the above-mentioned sequential identification algorithm, achieves good estimation accuracy even with a small amount of data, and provides stable estimation values even when disturbances are mixed in. The purpose of this invention is to provide an adaptive control method that can follow characteristics fluctuations well.

[Summary of the invention]

The present invention is characterized by the following points.

That is, instead of the conventional sequential identification algorithm, a finite number of observation data of length L is once stored in a storage device, and the data is connected N times in the time direction, and the entire observation data is regarded as one piece of observation data. Then, the conventional sequential identification alprism is used for this observation data.

Length of data L1 Number of pieces of data to be connected (number of times of repeated use) N is the characteristic change rate of the controlled object, the amplitude of disturbance,
And view i to use as data! Persistent excitability of the +11 signal (Peralst*ntry Ezaltlng.

It is determined from information 1 indicating the degree of rp-., ie, the width of the frequency band included in the signal.

Furthermore, such identification algorithms are used for adaptive control, and features such as starting and stopping functions for the identification algorithm, functions for resetting the algorithm in conjunction with changes in target values, and changes in characteristics are added to improve characteristics.

〔Effect of the invention〕

According to the present invention, the following effects can be obtained.

1) Since the conventional sequential algorithm is not changed, only the method of handling observation data is changed, the advantages of the sequential algorithm, such as the ability to obtain identification results online and the small amount of total In, can be utilized as is. Therefore, it is easy to switch between the conventional adaptive control alprism and this method.
Continuity can be maintained.

2) High-precision estimation similar to offline identification algorithms is possible with a simple algorithm. This is especially effective when the amount of data is small.

3) In the adaptive control method using this algorithm, since the long data interval is averaged for identification, even if a sudden disturbance is added to the observed data, the identification parameter f will not fluctuate greatly transiently. There are no drawbacks such as instability of the control system.

4) In the adaptive control method using this algorithm, the above 3
) For the same reason, the identification parameters can be stably obtained even if the data length is short, so by shortening the data length, it is possible to improve the followability to characteristic fluctuations of the controlled object.

5) Optimize the adaptive sensitivity of the adaptive control system to the observed data by adjusting the data length and the number of repetitions N according to the control system conditions such as disturbances, characteristic fluctuations, and the amount of information contained in the observed signal. Can be kept in good condition.

6) By importing data while monitoring the p,·, property of the observation signal, stopping data import when the p,·, property becomes weak, and then repeatedly using the stored data. Numerical divergence of the sequential estimation algorithm can be prevented.

[Embodiments of the invention]

FIG. 1 shows a configuration example of an adaptive control system according to an embodiment of the present invention.

The controlled object 1 is controlled by the operation signal u (t). This operation signal (1) is calculated in the manipulated variable calculation unit 2PC based on information on the control t y (t) and the control target value r (t). When the adaptive operation is started, the test signal generating section 3 generates a test signal r (t) consisting of an excit- ing signal. and,
By adding this signal 7(ri) and the target value signal r, (1),
(t) is generated. signal r(t).

u(t), y(t) pass through the filter sunglasser 4, and the discrete time observed signal sequences r(k), u(k), y(
k), data 70 is taken into the control unit 5 and further recorded in the observation signal storage unit 6. The data flow control unit 5 has the role of selecting either the current observation signal or the past observation signal stored in the observation signal storage unit 6 and sending the selected signal to the parameter identification unit 7. The parameter identification unit 7 identifies the unknown parameter θ of the dynamic characteristic model of the controlled object or the entire control system based on this observed signal sequence.

The obtained identification parameter θ is sent to the control parameter calculation unit 8, where a control parameter θC that realizes the optimal control law is calculated from the information of the identification parameter θ. The result is sent to the manipulated variable calculation section 2. The manipulated variable calculation unit 2 adjusts the manipulated variable and the control meter of the control law necessary for the calculation of (1) based on the sent information. The data flow control section 5, the parameter identification section 1, and the control parameter calculation section 8 are controlled by a control system monitoring section 9 in terms of their operation modes 9, starting and stopping, etc.

An example in which the above adaptive control method is applied to a PI control system will be described below.

FIG. 2 is a diagram showing a PI type adaptive control system according to this embodiment.

Let R(s) be the Laplace transform of r(t), u(t), and y(t), respectively.

Expressed as U (s) and Y (s), the controlled object 1 is approximately expressed by the following formula % Formula % On the other hand, the PI controller 2f (
The dynamic characteristics of the included closed-loop system 10 can be expressed as shown in the following equation.

In this example, the closed-loop system dynamic characteristic model of equation (2) is
Estimate from the observed signals r(t) and y(t), and use the results to determine the optimum control meter, that is, the PI constant C8.
A method for determining °C1 is used. /The meter identification unit 7 identifies a model of the following form corresponding to equation (2).

The parameters of the model in equation (3) are estimated using the following procedure.

After passing the target value signal r(t) and the control Sk y(t) through a state variable filter, they are sampled at a period τ to form a signal sequence r(k) = y(k) 0(==Q # l −2
, 3,..." ri is obtained. ・9 lame data identification unit 1,
The following ARMA model is identified using the sequential least squares method.

From equation (4), the coefficient of equation (3) ~I al 1121 as is determined by the following Alco9 rhythm.

In this way, ILOH61+ &2113 is found, υ, and the model of equation (3) is obtained. Is that all NoJ? This is a function of the parameter identification section 7.

Now, the feature of the present invention lies in the sequential algorithm used here.

What is used to identify the model expressed by equation (4) is the usual recursive least squares method as shown in the following equation. Estimated 9
parameter vector θ(k), observation data vector ψ
[Yes]) respectively, θ(k) and [^1. Give..., name, i2.・・・＊”
wya]" (9) 91'(k) minutes (-y(
k-1), -y(k-2), -, y(k-NA), r(
k-1).

-, r(k-NB))" α0 Here, in the present invention, the method of giving the data vector ψ(k) is different from the conventional method.

In the conventional iterative least squares method, ψ(k) becomes 0 at each time.
0, but here first give the data length and the number of repetitions N (L and N are integers), set the time to = 0, and set the initial value of the equation (9) ~ α◆ as follows. Set.

■While applying the first test signal 7(t) to the plant, data y(
k), r(k) are collected into the data flow control unit 5 from time k:O to L+n and stored in the signal storage unit 6, and at the same time, the data vector ψ(k) is stored from time k = n. (While setting as in IF9 formula, calculate the algorithm of αme~α◆ formula, and estimate parameter vector θ(k) (k = n −L +
Find n).

ψ(k) = (-y(k-1), -y(k-2), ・
= , -y(k-NA), r(k-1), -.

r(k-NB))" At time (k=n, n+1. . . -, L+n)=L
+n, the generation of the test signal is stopped, the acquisition of data y(k) and r(k), and the calculation of the estimation algorithm are stopped.

■Second time 7” (L+n) of the previous estimation algorithm α-H.

θ(L+n)', (set 7'(n) and θ(fi) again to return the time to k = n.

Data y(1), r(i) (i
=o.

〜, L+n) (Similar to formula 1119, ψ(k)
Let's create αη~(algorithm of equation 14 again k
= Estimated parameter vector θ by moving from n to n+L
(k) and update matrix F(k).

■After the third time The same process as the second time is repeated until the Nth time.

Such operations have the following meanings:

In other words, in the conventional sequential least squares method, if observation data from k = 0 to L is given, λ(k
)=λ (0<λ<1), the weighting of the data (
The initial value of the degree of emphasis placed on data is a weight of λ1 on 5, as shown in FIG. Therefore, even if an appropriate initial value is given, if the number of data is small, the influence of the error cannot be ignored.

On the other hand, in the method of this embodiment, N pieces of the same data are arranged in the time direction, apparently increasing the number of data N times. Therefore, for example, when N=3, the data weighting is
The result will be as shown in figure (b).

Therefore, the weight of λ3L is applied to the initial value), thereby making it possible to further reduce the influence of errors in the initial value. Tsumashi,
Even if the number of data is small, if the number of iterations N is increased, the accuracy of estimation can be improved since the influence of the initial value, which is a drawback of the sequential algorithm, is small.

The above is an explanation of the identification algorithm used in the parameter identification section 7, which is a feature of the present invention, and its data manipulation method.

Next, based on the identified parameter of equation (3), the control parameter calculation section 8 calculates the PI controller (operated amount calculation section) 20P! The constants C8 and C1 are determined, and the constants are tuned to those values. In this embodiment, the P■ constant is determined by the model matching method as follows.

First, consider matching the closed-loop system dynamic characteristic equation (3) from the target value r(t) to the control amount y(t)t with the reference model of the following equation that provides a desirable response.

At this time, based on the condition that the denominator moment series of equations (4) and (to) are made to match as much as possible from the low-order terms, the following equations for C8 and C1 can be obtained.

First, the estimated equation (4) is transformed into the following equation.

C1 is the PI constant of the PI controller at the time when the model of equation (4) is identified.

From σ thus obtained, a new PI constant is calculated using the following formula.

Then, set the PI constants C8 and C1 of the PI controller to Co
(new) s Change to C1(new).

Finally, the role of the control system monitoring section 9 will be explained below. The main functions here are: 1) Start, stop, and reset the identification algorithm (If)
Identification Al is the data length L1 of the rhythm, and the number of repetitions N is determined and changed 111) The identified parameter θ, the calculated control parameter 0c, etc. are checked and changed. These are executed based on the following rules (promises).

<Starting Alprism> ■Start the algorithm when the operator issues a start command.

(2) When control deviation - (t) = r (t) - y (t) exceeds a certain value for a certain period of time, the engine starts.

Determining the number of repetitions N (stopping the algorithm)> ■ Estimated parameter of the identification algorithm -
Stop repeating when the number of updates reaches .

(9) When the number of repetitions N reaches the maximum number set by the operator, stop the algorithm.

Resetting the identification algorithm> ■When the target value r(t) changes beyond a certain threshold value, that is, when l r(t) Activate the lyse, toshi, and al prism again as in (2) to α-type.

(b) When a change in the characteristics of the controlled object is confirmed, a similar reset and restart of the algorithm are performed.

<Monitoring of p... and determination of data length> After starting the reset Alco9 rhythm, the observed signal r( k), monitor the p... property of y(k).

After starting the algorithm, at time = Lo, J(kKg,
When (c) occurs, (1) If Lo > Lrn1 n (“ml n is the minimum data length set by the operator), then data length L =
As Lo, the generation of test signals and the acquisition of observation data are stopped, and the identification algorithm is repeated for the number of times N = 2.
After the combing estimation is performed.

(It) If Lo<Lrn1n, the test signal r(1
) and continue parameter identification.

Change of data length> Next, the amplitude of the unknown disturbance applied to the controlled object and the rate of change of the characteristic fluctuation are measured as follows.

First, model error t(k) y(k)−〇1k−1)ψ(k)
Monitor □□□ and! ε(k) l > t,
When it becomes @, is there any disturbance applied to the controlled object?
Alternatively, it is detected that there has been a change in the characteristics of the controlled object itself. Immediately after that, when 11(k+Δ]<t, the test signal generator 3 generates a test signal 7(t) in the form of a unit/lupus and adds it to the target value r(t).At this time, (1) If 1ε(k+Δ+1)l>1. again, it is assumed that there has been a characteristic change.At this time, the algorithm is reset, toshi, and identified again, and the obtained parameter is set as on@w.Also, Assuming that the parameter obtained at time -ko before detecting the characteristic change is 0o1d, the characteristic change rate δθ is defined as follows.

(II) If Ig(k+Δ+1)1<ε, then assuming that an unknown disturbance is added to the controlled object, let 8(s) be the amplitude δ of the disturbance.

δ=ε(k) The length of the wire data is changed as follows.

Assuming that the current data length is Lol, (1) When disturbance is confirmed, L=k −L + (1−, )・k2・δ
(to) 1 old (0≦, ≦1, k2) where O is a certain constant) is the new data length. This is intended to maintain parameter estimation accuracy by increasing the data length when the disturbance applied to the controlled object is large.

(II) When characteristic fluctuations are confirmed, calculate the rate of change in plant characteristics δθ according to equation (c), and then L=h1・
Lold + (1-h 1)・h2/δθ
Let 0k ((Kh,<1, h2)O is a certain constant) be the new data length. This is intended to improve the ability of identification parameters to follow changes in characteristics by reducing the data length as the characteristics of the plant change more drastically.

Change of the amplitude of the test signal〉 p of the observed signal, index J(k), disturbance amplitude δ,
When one of the characteristic change rates δθ exceeds a certain threshold and the identification algorithm is operating while incorporating observation data, an exciting test signal r(t ) is generated and added to the target value. Also, J(k).

Adjust the test signal hole according to the respective sizes of δ, δθ, and L as shown in the following formula.

7,=γ. +γ, /J(k)+γ2・δ+γ3・δθ
10γ〆LO group (γOt rl *γ2.γ3.γ4 are certain constants) <Various checks> ■If the model error g(k) (formula) is within the threshold, the estimation of the identification parameters is good. Adopted as.

■Apply upper and lower limiters of the following formula to the PI constant, which is a control parameter.

C0<C8m1n&Raba, CO" COmin □C
If O>COmx, CO=COmax.

If C4<C1m1n, C1=C1m1u・C1>0
If it is 1m8, then C1201mx.

■After tuning the control/main meter, control deviation/(k
If n = r(k) - y(k), then the chaining result is considered good; otherwise, the change in the control 74 parameter is canceled.

Each of these rules can be freely selected and set by the operator. Also, new rules can be added.

The control monitoring unit 9 can finely monitor and adjust the movement of each element of the control system using such a rule group.

In particular, by determining the data length and number of repetitions N necessary for the identification alprism in the present invention from the magnitude of the unknown disturbance, the rate of change in the characteristics of the ground, and the p-e and nature of the observed signal, this can be achieved according to the above-mentioned rules. The effectiveness of the algorithm can be maximized.

Finally, examples of how this PI type adaptive control device is used are shown in Figures 4 to 7.
As shown in the figure. In the PI type adaptive control system having the configuration shown in Fig. 2, when the conventional recursive least squares method is used as the identification algorithm, when a disturbance d(t) is applied to the plant, the 42 meter a of the model of equation (3) is . 8 alt $
The state of each signal of the control system and the identification parameter at that time when the identification value of 121C3 is greatly deviated is shown in FIGS. 4 and 5.

Figure 4 shows the case where there is no disturbance, and the parameters quickly converge to their true values over time. However, when an irregular table disturbance is added as shown in Figure 5, the identification/displacement meter becomes transiently large. It gets messy. In this environment where the parameters are disturbed by disturbances, the response of the control system was compared between the case where the conventional Alprism was used and the case where the algorithm of the present invention was used. The results are shown in Figure 6(a) and
(b) Show K. In (,), the response was not improved after that because the leverage was applied based on the instantaneous meter identification value at the time of chaining, but in (b), the data before chaining Since the chaining was performed based on the /IF parameter identified by repeating the method 10 times, the subsequent response was clearly improved. Note that, as shown in FIG. 7, as the number of times N of repeated use of observation signal data increases, the parameter estimation error becomes smaller, and the remarkable effect of the present invention can be obtained.

[Brief explanation of drawings]

FIG. 1 is a control block diagram showing the basic configuration of an adaptive control system according to an embodiment of the present invention, FIG. 2 is a control block diagram of an embodiment in which the present invention is applied to a PI type adaptive control system, and FIG. The figures show the weighting of observed data in the identification algorithm of the present invention and the conventional sequential least squares method, respectively. FIG. 4 is a diagram showing the situation when there is no disturbance, FIG. 5 is a diagram when disturbance is added, and FIG. 6 is a diagram showing the conventional sequential least squares method, the method of the present invention, and PI type adaptive control using K. FIG. 7 is a diagram showing the difference in response waveforms of the system, and is a diagram showing the relationship between the number of data repetitions and the estimation error of the identification parameter in the present invention. DESCRIPTION OF SYMBOLS 1... Controlled object, 2... Manipulated amount calculation part, 3... Test signal generation part, 4... Filter sampler, 5...
Data flow control unit, 6... Observation signal storage unit, 7...
- Parameter identification section, 8... control parameter calculation section,
9... Control system monitoring unit, 10... Closed loop system, 11.
...Adaptive control device, 12... State variable filter, 13
... Sampler.

Claims (2)

[Claims] (1) Sequential identification means for estimating parameters related to the dynamic characteristics of the controlled object online from sample values of observation signals related to the controlled object, a storage device for storing the sample values, and estimation by the sequential identification means. a control parameter calculation means for calculating a control parameter based on the parameter obtained by the parameter calculation means; a manipulated variable calculation means for calculating a manipulated variable from the target value and the control amount of the controlled object based on the control parameter obtained by the parameter calculation means; In an adaptive control method using a test signal generating means for generating a test signal consisting of a continuous excitatory signal of the controlled object and adding it to the control system,
Adaptive control characterized in that parameters related to the controlled object are estimated using L×N data strings obtained by storing sample value data N times and reading it out repeatedly. method. (2) The data L to be temporarily stored in the storage device and the number of times N to read it repeatedly are determined based on the magnitude of the disturbance applied to the controlled object, the temporal change rate of the dynamic characteristics of the controlled object, and the observation signal used as data. 2. The adaptive control method according to claim 1, wherein the adaptive control method has a function of automatically adjusting according to the degree of sustained excitability.