JPH04302304A - Nonlinear process controller - Google Patents

Abstract

PURPOSE:To control a process system provided with nonlinear characteristic at high speed and with high accuracy by providing a feedback error compensator to which gain scaling is applied, a first process compensator designed so that process output can follow a process target value, and a second process compensator consisting of a multi-layer type neural network. CONSTITUTION:A series compensator 6 which displays the role of the first process compensator which performs compensation on a nonlinear process 1 by conventional technique, the feedback compensator 7. a neural network disturbance compensator 8, a coupling weight update quantity calculation part 9 for the compensator 8, a network learning controller 10, a feedforward compensation learning error generation filter 11, and a disturbance compensation learning error generating filter 12 are provided. In other words, an error signal from a conventional error compensator is set as a learning error signal, and also, the whole system in which the compensation is applied to a controlled system process by the conventional compensator is recognized as an expansion process system 13, and a compensation circuit setting a parameter signal representing the characteristic of the system as input is provided.

Description

[Detailed description of the invention]

[Object of the invention]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a process control device for controlling a process system having nonlinear characteristics at high speed and with high precision.

0002

[Prior Art] Conventional process control devices for processes in general industrial plants model the target process based on its physical characteristics, and then use a series compensator or feedback compensation system that is designed based on the results. Control is performed using devices such as devices. However, since normal processes are inherently nonlinear systems, when using a linear fixed compensator such as a PID control device, it is not always possible to sufficiently compensate for nonlinear characteristics caused by deviations in the operating operating point. High-speed, high-precision operation is often virtually impossible. Control methods such as a gain scheduling control method (FIG. 4) and an adaptive control method (FIG. 5) have already been proposed as methods for compensating for such deviations from the linear model of the target process.

In the former gain scheduling control method, the gain scheduling device 403
The gain in the transfer function of the controller 402 is adjusted according to the input operating conditions, and the target value r is adjusted using the gain-scheduled transfer function and given to the plant 401 as the manipulated variable u.
The output y is fed back to the controller 402,
It is used to compensate for the deviation from the target value r.

As described above, the gain scheduling control method is a relatively rough compensation method whose primary purpose is to ensure system stability against large changes in the operating point. Therefore, in order to improve operational accuracy, it is necessary to change the fixed compensator parameters appropriately as the operating conditions change.
At this time, it is generally difficult to make use of past operating performance data, and the current situation is that re-tuning is required in the case of new operating conditions.

On the other hand, in the adaptive control method, the reference characteristics of the plant 501 are given as a reference model 503, and the adaptive mechanism 504 controls the controller 502 according to the deviation e between the output yM of this reference model 503 and the output y of the plant 501.
gain scheduling is performed, and the target value r is adjusted through the gain-scheduled transfer function in this way and is given to the plant 501 as the manipulated variable u.

Therefore, this adaptive control method is a time-varying control method that attempts to suppress characteristic fluctuations of the target process by actively adjusting the parameters of the linear compensator.
Its design and operation requires fairly accurate prior knowledge of the system structure of the target process. In general, it is difficult to explicitly identify the dynamic characteristics of industrial processes, so when current adaptive control methods are used in situations where nonlinearity is strongly expressed, not only will sufficient adaptive ability not be obtained, but the system It could also lead to further instability. In order to avoid such instability, methods have been proposed in which a stabilizing loop such as a stability monitoring mechanism is constructed around the adaptive control system, but overall, conventional methods centered on PID control methods etc. At present, performance that surpasses that of other methods has not been achieved.

One of the reasons why these methods do not function satisfactorily is thought to be that their framework is strictly closed in the world of linear systems. Recently, as a means of nonlinear compensation that goes beyond the linear framework, the use of signal processing systems (hereinafter referred to as neural networks) that simplify the operation of human neural networks and simulate their basic performance has been developed. has been proposed (Figure 7). this is,
For example, in the literature (K. Funahashi: On the
Approximate Realization o
f Continuous Mappings byN
neural networks, neural nets
works, Vol1.2, 183/192 (1989
), the theory that a multilayer neural network can realize arbitrary continuous mapping from a dimensional space defined by the number of units in the input layer to a dimensional space defined by the number of output layer units. It is based on rationale.

FIG. 6 shows the configuration of a control system using a neural network that has already been proposed. The control system for nonlinear objects using this configuration is disclosed in the published patents with publication numbers (Heisei 1-228001) and (Heisei 2-54304), and will be briefly described below.

Compensation for the nonlinear process 1 consists of feedback compensation whose main purpose is to maintain stability, and feedforward compensation by the neural network 3 whose purpose is to improve the quick response and high accuracy of the system response. configured. The connection weight update amount calculation unit 4 uses a signal ue obtained by gain scaling the deviation e between the output of the target value generator 5 and the process output by the feedback error compensator 2 as a learning error signal, and calculates the signal ue so that ue → 0. This is to adjust the connection weights of the neural network 3.

Although there are several variations in the configuration of neural networks, FIG. 7 shows the basic configuration. In the same figure, X=[X1
is the output. The whole consists of M layers of neuron columns. Here, the neuron Nk+1 j of the k+1st layer
If we pay attention to (j=1,2,......,nk+1),
Each neuron has all the neurons in the kth layer Nk i (i
=1,2,...,nk) output signal Ok i (i=
1, 2, ... ..., nk ) with weight Wk ik+1 j
are combined with. That is, the input ik+1 j of the j-th neuron Nk+1 j of the k+1-th layer is expressed as follows.

Further, each neuron Nk+1 j receives input ik+1 j and calculates Ok+1 j =f(ik+1
j) is output. Here, the function f(・) is non-decreasing,
A differentiable function, for example, f(x)=1/(1+exp(-x))
(3
) is given by

The characteristics of the neural network are determined by weighting coefficients Wk i k+1 j and θk+1 j . The error back propagation learning method (error back p
ropagation learning method
d) has been proposed in recent years (D.E. Rumelhar
t,G. E. HintonandR. J. William
ms: Learning representation
ns by back-propagating er
rors, Nature, 323-9, 533-536
(1986)). In this method, the weighting coefficient W is modified according to the following formula, for example. Wk-1 i k j =Wk-1 i k j
+ηδk j Ok−1 i
(4) (k=M, M-1,......, 3) (i=1, 2,..., nk-1) (j=1, 2,
... ..., nk ) However, δM j = (y* j −OM j )
f(iM j ) (5)
Here, y* j is the teaching signal, f(・) is the formula (3)
means the differentiation of the function shown in . As this correction is repeated, the weighting coefficient W is learned so that the output Y of the neural network matches the teaching signal Y*. Note that other error backpropagation learning methods are also possible. For an overview of these, see, for example, Corona Corporation Computer Roll No. 24, P53-60 (October 1, 1988
(published on the 0th).

In the above learning process, when the learning is insufficient, the control of the system is performed mainly by feedback compensation, but as the learning progresses (|ue | → small), the control is performed by the feedforward compensation by the neural network 3. Moving on to compensation. In a situation where ue ~0 is achieved, it is equivalent to a configuration in which the feedback loop is broken, and therefore a characteristic P-1 inverse to the input/output characteristic P of the target process 1 is realized in the neural network 3. .

The main feature of this control method is that the final compensator (
That is, there are two points: the framework of the neural network 3) is a nonlinear system, and the compensator is constructed by learning. The former means that the class of compensators has become broader, giving the possibility of higher-performance compensation, and the latter means that the characteristics of the target system (especially nonlinear characteristics) are unknown even if they are unknown. This gives us the possibility of realizing this through the means of learning.

As shown above, the control system configuration using a neural network as a feedforward compensator compensates for the nonlinear characteristics of the target process and is more effective than conventional PID control devices or adaptive control devices. It can be said to be one of the control system architectures that offers the possibility of realizing high-speed, high-precision control.

However, if the above configuration is used as it is for controlling a process system, etc., the neural network will inevitably learn the inverse characteristics of the target process, so the characteristics that should be realized in the neural network are becomes significantly more complex. This means an increase in the size of the network and the learning load, which is likely to make it difficult to put it into practical use.

Furthermore, in the case of the above configuration, since the neural network is trying to learn to a point that far exceeds the limits of the response characteristics of the actuator and far exceeds the tracking performance actually required, There is a danger of falling into a state of overlearning. In other words, in general, target value changes in process systems are often given as step changes, but achieving complete inverse characteristics means constructing a compensator that can perfectly follow such step changes. It means that. However, since such a compensator is physically unfeasible,
Neural networks repeatedly learn to achieve impossible input/output characteristics, resulting in overlearning.

Furthermore, in the case of an actual process, various disturbances are usually applied to the system due to changes in the operating operating point or changes in target values. It is difficult to achieve the desired operational behavior without compensating for such disturbances, but feedforward compensation by realizing the inverse characteristics as described above cannot sufficiently compensate for system disturbances caused by these disturbances. A new control mechanism is required to compensate.

[0019]

[Problems to be Solved by the Invention] In order to solve the above-mentioned problems, the present invention aims to improve the effectiveness of neural networks while retaining the effects of conventional control techniques such as linear compensation and gain scheduling control, for which design know-how has been accumulated. By adopting a control system architecture that can take advantage of its nonlinear compensation ability, a feedforward compensator that improves responsiveness to target value changes and a disturbance compensator that reduces the influence of disturbances are configured using a neural network. It is an object of the present invention to provide a nonlinear process control device that can realize a practical neural network compensator without falling into excessive learning by using an appropriately filtered error signal as a learning error signal.

[Configuration of the invention]

[Means for Solving the Problems] A nonlinear process control device of the present invention includes a feedback error compensator that outputs an error signal subjected to gain scaling for a deviation between a process target value and the process output, and a modeling of a controlled process. The first process compensator is designed such that the process output follows the process target value based on the characteristics, and the error signal from the feedback error compensator is used as a learning error signal, and the first process is applied to the process to be controlled. 1
When the entire system compensated by the process compensator is captured as an expanded process system, a multilayer neural network that inputs a parameter signal indicating the characteristics of the expanded process system and outputs the pre-stage feedforward manipulated variable to the first compensator. and a second process compensator comprising a second process compensator.

The nonlinear process control device of the present invention as set forth in claim 2 further uses the error signal from the feedback error compensator as a learning error signal, and inputs a parameter signal indicating a disturbance to the controlled process, and calculates the disturbance from the process output. A third process compensator is provided, which is a multilayer neural network device that outputs a disturbance compensation operation amount that cancels out the influence of.

The nonlinear process control device of the present invention as set forth in claim 3 further comprises filtering the error signal from the feedback error compensator so that it has a band component desirable for learning the multilayer neural network device. A filter means is provided for supplying the signal as a learning error signal to the multilayer neural network device.

The nonlinear process control device of the present invention as set forth in claim 4 further provides a learning error signal only to the multilayer neural network device constituting the first compensator during a transient response due to a change in the process target value. In a steady state with no change in value, a learning error signal is given only to the multilayer neural network device constituting the second compensator, and the neural network device that performs learning processing is switched and controlled between transient response and steady state. Equipped with network learning control device.

[0024]

[Operation] In other words, the present invention basically treats the entire system in which the nonlinear process, which is the original control target, has been compensated by the prior art using the first compensator as an expanded process system, and this expanded process system. A feed-forward compensator (second It is characterized in that it is incorporated as a third process compensator) and a disturbance compensator (third process compensator) to configure a process control system.

Therefore, according to the process control device of the present invention, the first process compensator performs linear compensation such as improving process vibration characteristics and compensating for dead time, nonlinear compensation for sensor sensitivity characteristics, gain scheduling compensation, etc. The process including the first process compensator is treated as an expanded process system for system characteristics whose characteristics have been improved accordingly, and learning compensation for inverse characteristics and disturbance reduction is performed on this expanded system using a neural network.

When compensating for disturbances, we learn the inverse characteristics of the generation dynamics from the output and state quantities of the process system, not only for observable disturbances but also for unknown disturbances that cannot be observed, and cancel the effects of the disturbances. We will construct a compensation mechanism on a neural network that will do this.

By adopting such a control system configuration, the characteristics to be compensated by the neural network are mainly nonlinear characteristics that cannot be fully compensated for by conventional compensators, and as a result, the load on the neural network compensator is reduced during transients. At the same time, it becomes possible to configure a control system by making use of conventional compensation techniques.

Furthermore, in realizing the learning of the neural network compensator, by using an error signal obtained by band-processing the output signal of the feedback error compensator as the learning error signal, the neural network compensator can be effectively used in the desired frequency band. This makes it possible to train the robot to operate in a similar manner, and avoids excessive learning.

Furthermore, when training the neural networks for feedforward compensation and disturbance compensation,
When the system is in a transient response state due to a change in the target value, the neural network for disturbance compensation is not trained, and only the neural network for feedforward compensation is trained to actively acquire the inverse characteristics of the system and achieve the target value. When the system is in a steady response state with no value changes, only the neural network for disturbance compensation is trained, contrary to the case of transient response, and the learning process is controlled so that it actively acquires dynamics that compensate for the disturbances applied to the system. By incorporating the mechanism into a process control device, it becomes possible to efficiently realize both the target value tracking and disturbance suppression functions.

[0030]

EXAMPLES The present invention will be explained in detail below based on examples. FIG. 1 is a diagram showing an embodiment of a nonlinear process control device according to the present invention. In this embodiment, a nonlinear process 1, a feedback error compensator 2, a neural network feedforward compensator (hereinafter abbreviated as NFF compensator) 3
, the connection weight update amount calculation unit 4, and the target value generation device 5, the sixth
The configuration is the same as the conventional neural network control system shown in the figure, and its operation and functions are also the same. The difference in this embodiment is the presence of a series compensator 6 and a feedback compensator 7, which serve as a first process compensator that performs conventional technology compensation for the process, and a neural network disturbance compensator 8 and its combination. It includes a weight update amount calculation unit 9, a network learning control device 10, a feedforward compensation learning error generation filter 11, and a disturbance compensation learning error generation filter 12. As already explained, compensators in typical process control devices are designed based on modeling characteristics of the process dynamics. Control methods include classical control methods such as PID and I-PD control, as well as modern control methods such as optimal regulators, and gain scheduling aimed at maintaining stability against characteristic fluctuations due to changes in operating operating points. Control etc. are being carried out.

Compensator configurations often used on a practical level include a phase compensator or an integral compensator as the series compensator 6, and a state feedback compensator or a speed feedback compensator as the feedback compensator 7. . Control system design technology centered on such linear compensation has already accumulated a wealth of know-how based on various application examples, and the entire compensated process system (enlarged process system 13) can be compared to the original process alone. It is possible to obtain a product with considerably improved characteristics compared to .

The network learning control device 10 receives the target value r, and when the system is in a transient response state due to a change in the target value r, the neural network of the disturbance compensator 8 does not learn, and the NFF compensator 3 The system learns only the neural network of the disturbance compensator 8 to actively acquire the inverse characteristics of the system, and when the target value r is in a steady response state with no change, only the neural network of the disturbance compensator 8 is trained, contrary to the case of a transient response. , the learning process is controlled to actively acquire dynamics that compensate for disturbances applied to the system.

That is, first, when the target value r changes, the network learning control device 10 stops the learning of the neural network disturbance compensator 8, instructs the learning of the NFF compensator 3, and calculates the learning error through the filter 11, which will be described later. signal ε1
is given to the NFF compensator 3. This NFF compensator 3 receives the target value r (or its multiple differential signal), the process output x from the nonlinear process 1, and the state quantity y to its input layer, and learns the characteristics of the expanded process system 13. go. Therefore, the learning adjustment of this NFF compensator 3 progresses and ue ~
When 0 is achieved, the inverse characteristic Pa −1 of the characteristic Pa of the enlarged process system 13 is learned and realized in the neural network, not the inverse characteristic of the characteristic P of the process l itself. To repeat what has been said above, the expansion process system 13
Since the characteristic Pa of is improved compared to the characteristic P of the original process 1 itself, the burden of learning its inverse characteristic becomes lighter, and as a result, the effect of nonlinear compensation of the system by the NFF compensator 3 is reduced. will increase.

When the transient response state of the system due to the change in the target value r is stabilized and becomes a steady response state, the network learning control device 10 stops learning the NFF compensator and instructs the learning of the neural network disturbance compensator. . At this time,
The neural network disturbance compensator learns the inverse characteristics of the generation dynamics of the observable disturbance d2 and the unobservable disturbance d1 from the process output Adjustments are made so as to output the compensation operation amount ud.

[0035] Both of the above-mentioned neural network learning aims to realize the inverse characteristic, but for example, P
Similar to P-1, a-1 is physically impossible to fully realize (requires future values of the system). Therefore, when the target value r output by the target value generator 5 or the disturbances d1 and d2 change stepwise or sharply, learning is performed with the aim of realizing perfect follow-up characteristics or suppression characteristics for the changes. When you do this, neural network compensator 3
, 8 results in repeated excessive learning to realize a compensator that does not exist.

[0036] When such overlearning occurs, the feedforward operation amount of the neural network compensator becomes oscillatory, and undesirable perturbations remain in the entire process control device. need to be avoided.

Learning error generation filters 11 and 12 are provided for this purpose. Learning error generation filter 1
1 and 12 serve to convert ue, which was conventionally used as a learning error signal, into band-processed signals ε1 and ε2, respectively, and use ε1 and ε2 as new learning error signals to calculate the connection weight update amount. Units 4 and 9 perform learning for each neural network compensator. Here, the filter characteristics of the learning error generation filters 11 and 12 vary depending on the band in which each neural network compensator is desired to operate effectively, but among them, the typical gain - The frequency characteristics are shown in FIGS. 2 and 3.

Generally, in process control, it is impossible to achieve infinitely fast response due to limitations in the response characteristics of actuators, and there is usually an upper limit to the tracking performance actually required. . In other words, this means that there is no need to follow changes in the target value over a certain frequency band. When the feedforward learning error generation filter 11 has a high-frequency cutoff characteristic as shown in FIG.
Since this is performed for the following frequency band components, the NFF compensator 3 is used for frequency components higher than fB of the error signal.
becomes numb. Therefore, unnecessarily reverse characteristic learning, that is, excessive learning is avoided, and stable operation of the entire process control device is possible. Similarly, if the disturbance compensation learning error generation filter 12 is given specific bandpass filter characteristics as shown in FIG.
Since the learning of the suppression function for disturbances in the fB2 band is focused on, it becomes possible to efficiently implement a disturbance compensator when the frequency band of the main disturbance is known.

Although the embodiments of the present invention have been described in detail above, the present invention is not limited to the above embodiments, and the present invention is not limited to the above embodiments. By replacing with various compensation devices or appropriately selecting the filter characteristics of the learning error generation filters 11 and 12, it is possible to realize more advanced nonlinear process control. In that sense, the invention can be modified and implemented as appropriate without departing from the gist.

[0040]Although this specification describes a neural network in which there is no connection that jumps between layers, there is a connection that jumps between layers such as direct connection from an input layer to an output layer via weights. Even if there is, learning is possible using the same method as above. The neural network incorporated in the present invention includes:
General forms including such methods are available.

[0041]

[Effects of the Invention] As described above, since the present invention is a nonlinear process control device as described in the claims, it utilizes the nonlinear compensation ability of the neural network compensator while taking advantage of the compensation effect of conventional control technology. This makes it possible to achieve high-speed, high-precision control of unknown nonlinear processes. In particular, when training a neural network compensator, by appropriately selecting the filter function that generates the learning error signal, the operating band of the neural network compensator can be adjusted and excessive learning can be avoided. By switching the neural network to be learned depending on the response state, it is possible to efficiently learn both the target tracking and disturbance suppression functions on the neural network compensator.

[Brief explanation of drawings]

FIG. 1 is a functional block diagram of a nonlinear process control device according to an embodiment of the present invention.

FIG. 2 is a diagram showing an example of learning error signal generation filter characteristics in the present invention.

FIG. 3 is a diagram showing an example of learning error signal generation filter characteristics in the present invention.

FIG. 4 is a functional block diagram showing the configuration of a conventional gain scheduling control device.

FIG. 5 is a functional block diagram showing the configuration of a conventional adaptive control device.

FIG. 6 is a functional block diagram showing the configuration of a control system using a conventional neural network.

FIG. 7 is a diagram showing one configuration of a neural network.

[Explanation of symbols]

1 Nonlinear process 2 Feedback error compensator 3 Neural network feedforward compensator 4 Connection weight update amount calculation unit 5 Target value generator 6 Series compensator 7 Feedback compensator 8 Neural network disturbance compensator 9 Connection weight update amount calculation unit 10 Network learning control device 11 Feedforward compensation learning error generation filter 12 Disturbance compensation learning error generation filter 13 Expansion process system

Claims (4)

[Claims] Claim 1: A controlled process having nonlinear characteristics;
a target value generator for generating a control target value for the process output; a feedback error compensator for outputting an error signal subjected to gain scaling for a deviation between the process target value and the process output; a first process compensator designed such that the process output follows the process target value based on the process target value, and an error signal from the feedback error compensator as a learning error signal; When the entire system compensated by one process compensator is captured as an expanded process system, a multilayer type inputs a parameter signal indicating the characteristics of the expanded process system and outputs a pre-stage feedforward manipulated variable to the first compensator. a second process compensator comprising a neural network device. 2. A disturbance compensation operation amount that uses an error signal from a feedback error compensator as a learning error signal, and inputs a parameter signal indicating a disturbance to the controlled process to offset the influence of the disturbance from the process output. 2. The nonlinear process control device according to claim 1, further comprising a third process compensator comprising a multilayer neural network device that outputs. 3. Filtering the error signal from the feedback error compensator so that it has band components desirable for learning of the multilayer neural network device, and using the processed signal as a learning error signal for the multilayer neural network. 3. A nonlinear process control device according to claim 1, further comprising filter means for providing a nonlinear process control device to the device. 4. During a transient response caused by a change in the process target value, a learning error signal is provided only to the multilayer neural network device constituting the first compensator, and in a steady state where the process target value does not change, the learning error signal is applied to the second compensator. A claim comprising a network learning control device that switches and controls a neural network device that performs learning processing by giving a learning error signal only to a multilayer neural network device constituting a compensator, between the transient response time and the steady state state. Item 3. The nonlinear process control device according to item 3.