JP2775968B2 - Process control equipment - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION <Industrial Application Field> The present invention relates to a process control device using a neural network.

<Prior Art> Conventionally, PID control methods have been overwhelmingly adopted at plant sites. In order to adjust the process control system of the plant centered on this PID to the control characteristics as specified, it is necessary to set the parameters (control constants) of the controller to appropriate values according to the characteristic fluctuation of the process to be controlled . In order to adjust the control system, it is necessary to identify a process by some kind of response. There are a method in which the control operation of the controller is stopped in an open loop and a method in which the control operation is performed during closed loop control. However, since the operating conditions of the plant may not be maintained in the open loop, a method that can perform from identification to adjustment during closed loop control is desirable. As a method of adjusting the control system during the closed loop control, more specifically, there is adaptive control as a control method that automatically adjusts online and keeps the performance of the control system in the best state at all times.

Further, adaptive control includes a method called a gain scheduling method or a self-tuning regulator.

FIG. 5 is a conceptual block diagram equivalent to the block diagram of the gain scheduling method described in the basic book of the Automatic Control Handbook (edited by the Society of Instrument and Control Engineers), pages 701 to 703. The environment condition 1 that caused the characteristic fluctuation is determined by the environment measurement mechanism 2, and the measurement result is given to the parameter generator 3. The parameter generator 3 has a table showing the correspondence relationship with the controller parameters, and gives the controller parameters corresponding to the measurement results to the controller 4.

The controller 4 obtains an input (process operation amount) given to the plant 5 based on the target value, the output of the plant 5, and the parameters.

FIG. 6 is a conceptual block diagram for explaining a self-tuning regulator system similar to that described in the Basic Edition of the Automatic Control Handbook (edited by the Society of Instrument and Control Engineers), pages 701 to 703. In this method, an appropriate control strategy is first selected, and a required controller structure is determined on the assumption that plant parameters are known. The identification mechanism 6 receives the input and output of the plant 5,
The output is given to the design calculation mechanism 7. An unknown parameter of the plant is sequentially estimated by the identification mechanism 6 and the design calculation mechanism 7. The parameters obtained by the estimation are given to the controller 4.

In such a block, first, unknown parameters of the plant 5 are sequentially estimated using an appropriate identification method. The result is regarded as a true value, and the parameters of the controller 4 are determined and adjusted online.

<Problems to be solved by the invention> However, the above-described conventional method has the following problems.

The gain scheduling method has a problem that it is necessary to prepare in advance a relation between environmental conditions and parameters of an optimal controller in a certain range in a table.

On the other hand, in the self-tuning regulator method,
Even when the vehicle passes through the same operating point many times, the parameters of the plant are identified and the parameters of the controller are adjusted each time, and there is a problem that there is always a delay.

An object of the present invention has been made in view of the above points, and in a process in which characteristics vary depending on an operating point and an environmental condition, an operating point and a process parameter for each environmental condition are identified during operation. And learning it, so that optimal control is always realized within a certain range without identification, so that the same operating point and environmental conditions
It is an object of the present invention to provide a process control device that does not need to perform identification again when passing.

<Means for Solving the Problems> The present invention for achieving such an object provides data necessary for identification calculation by performing data processing for correctly performing the identification of the process (10). A learning management unit (20) having a function of organizing a given identification calculation result and sending it as data to be learned, and an autoregressive moving average model considered as a process model are represented by the following equations. Identifier (30) that estimates the parameters of by the method of least squares, and the state of process (10)
A neural network (40) that outputs the parameters of the ARMA model and performs learning based on the data provided from the learning management unit (20), and an appropriate PID controller parameter for the ARMA model of the process (10) under a certain standard. And a PID controller (60) for calculating a process operation amount based on a deviation between a target value and an output of the process (10) and an output of the gain setter (50). A process control device, comprising:

Here, u (k) is an input at k time y (k) is an output at k time p, q is an order <Action> In the present invention, in the learning management unit, a process parameter for each operating point or environmental condition during driving. And outputs it as data to be learned.

The identifier estimates the parameters of the ARMA model. The values are then input to a neural network to obtain the ARMA model parameters for the process.

The gain setting section calculates the optimal PID controller gain under a certain standard from the parameters of the ARMA model, and calculates the PID
The controller uses the gain of this calculation result to calculate the manipulated variable from the deviation between the target value and the process output.

<Example> Hereinafter, the present invention will be described in detail with reference to the drawings. FIG. 1 is a configuration diagram showing one embodiment of a process control device according to the present invention. In the figure, 10 is a process, 20 is a learning management unit, 30 is an identifier, 40 is a neural network, 50 is a gain setting unit, and 60 is a PID controller.

The learning management unit 20 receives the input and output of the process 10,
Data processing for correctly performing identification is performed, and the output is passed to the identifier 30 as data for identification calculation.

The details will be described below. Here, the discrete time is k, and the sampling period is T. For the process, the input variable is u (k) and the output variable is y
(K) and the other state measurement variables are xi (k) (where
i = 1, ..., n).

The learning management unit 20 performs data processing for each sample and stores m input / output and measurement data from the past to the present of the process. That is, the following data strings are stored.

u (k), u (k-1), u (k-2), ..., u (km-1) (1) y (k), y (k-1), y (k-2) , ..., y (km-1) (2) xi (k), xi (k-1), ..., xi (km-1) (i = 1, ..., n)
(3) Then, it is determined whether or not these data satisfy the following conditions.

umax-umin ≦ ua where umax and umin are the minimum and maximum values of the above data sequence (1) ymax−ymin ≦ ya where ymax and ymin are the minimum and maximum values of the above data sequence (2) ximax−ximin ≦ xia where ximax and ximin are the minimum value and the maximum value of the data string (3). The data string (1) satisfies the condition of continuous excitation.

If these four conditions are satisfied, the data sequence (1),
(2) the passing to the identifier 30 of the ARMA model of the process identification result parameter _{{a i, ..., a p} , b o, ..., b q} receive.
This and the average value umea of the data strings (1), (2) and (3)
Let n, ymean, ximean (i = 1, ..., n) be one set of process dynamic characteristic data.

This set of process dynamic characteristic data is stored each time it is obtained, and the neural network
40 is given as a teacher signal for learning.

The identifier 30 is an ARMA (Auto R) represented by the following equation (4).
egressive Moving Average) The parameters {a ₁ , ..., a _p , b _o , ..., b _q } of the model are calculated from the input / output data strings (2) and (3) of the process. It is estimated by the well-known least squares method.

Here, p and q are orders. The input / output data given here and the start command of the identifier 30 are obtained from the learning management unit 20.

The neural network 40 has a layered structure as shown in FIG. 2. The input is u (k) for the input of the process, and the output is y for the process.
(K), the measurement variable is xi (k), and the output is that the parameters of the ARMA model of the process are {a ₁ , ..., a _p , b _o , ..., b _q } There are two types.

a) Learning operation Each time the learning management unit 20 obtains a set of process dynamic characteristic data, the learning management unit 20 learns the set of teacher data stored so far. The actual data flow can be obtained by inputting the input in the data set to the neural network 40 and obtaining the output.The output in the data set is given as a teacher signal, and the error in that case can be used. Modify weights with a known back-propagation learning rule.

However, the input in the data set in this case is the average value umean, ymean, ximean instead of u (k), y (k), xi (k) (i = 1, ..., n). (I = 1,..., N) are used.

b) Calculation operation This calculation operation is executed every sample or every several samples. U (k), y (k), xi (k) (i
= 1, ..., n), and outputs the parameters {a ₁ , ..., a _p , b _o , ..., b _q } of the ARMA model of the process, and the gain setting unit 5
Pass to 0.

The gain setting unit 50 controls the ARMA model of the process.
A parameter to the PID controller 60 performing an optimal operation under a certain standard is calculated. For example, an ARMA model is converted to an S-region transfer function, and the parameters of the PID controller that can obtain a waveform that overshoots about 10% are calculated.

The PID controller 60 is a controller that calculates a process manipulated variable by proportional, integral, and derivative operations. Assuming that the target value is yd (k), the proportional gain is kp, the integral time is Ti, and the derivative time is Td, the control deviation is e (k) = yd (k ) -y (k) (5) △ e (k) = e (k) -e (k-1) (6) △ 2 e (k) = △ e (k) - Δe (k−1) (7), from which a process manipulated variable represented by the following equation is output.

u (k) = u (k-1) + kp [△ e (k) + T / Ti ・ e
(K) + Td / T △ ² e (k)] (8) The operation in such a configuration will be described with reference to the operation flow in FIGS. 3 and 4.

(1) Normal control loop (for each sample) FIG. 3 shows the operation flow. The learning management unit 20 captures the current input, output, and measured values. The values are then input to neural network 40 to obtain the ARMA model parameters for the process.

Subsequently, the gain setting unit 50 calculates an optimum PID controller gain under a certain criterion from the parameters of the ARMA model. The PID controller 60 calculates the manipulated variable from the deviation between the target value and the process output using the gain of the calculation result, and operates the process 10 by this.

(2) Learning operation (started every sample or every few samples) FIG. 4 shows the operation flow. The learning management unit 20 always stores the input / output and measurement data strings [(1), (2),
(3)] is stored, and the identifiable condition is determined. next,
If the identifiable condition is satisfied, the process identifier 30 identifies the ARMA model of the process. Further, the average value of the input / output and measurement data is calculated, and the calculated average value and the parameters of the ARMA model are stored as a set.

The set of data and parameters stored as described above is passed to the neural network 40 as teacher data, and learning is performed. In this case, good control cannot be expected at first unless process identification is performed for each operating point and environmental conditions, or unless an appropriate initial state is given to the neural network. It learns process dynamic characteristics for each operating point and environmental condition, and grows to always realize optimal control.

The present invention is not limited to the embodiments. If it is absolutely necessary to identify the process dynamic characteristic, a means for adding an identification signal to satisfy the identifiable condition may be added.

Also, the control norms are not limited to the embodiment, and other norms may be used.

<Effects of the Invention> As described in detail above, the present invention has the following effects.

By learning the dynamic characteristics of the process for each operating point and environmental condition for a process whose characteristics fluctuate depending on the operating point and environmental conditions, use the learning result for the second operating point and environmental upper limit. Thereby, the process dynamic characteristics can be understood without re-identification, and good control can be expected.

Even at operating points and environmental conditions that have passed for the first time, the approximate function of the process can be understood by the interpolation function of the neural network, and good control can be expected.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of a process control device according to the present invention, FIG. 2 is a diagram showing a state of a layered neural network and back propagation, and FIGS. 3 and 4 show an operation flow. FIG. 5 is a conceptual block diagram for explaining the gain scheduling method, and FIG. 6 is a conceptual block diagram for explaining the self-tuning regulator method. 10: Process, 20: Learning management unit, 30: Process identifier, 40: Neural network, 50: Gain setting device, 60: PID controller.

Claims (1)

(57) [Claims] 1. A process control device for a plant centering on PID control, wherein data processing is performed to correctly perform identification of a process (10) to obtain data necessary for identification calculation and to be provided. A learning management unit (20) having a function of organizing the identified calculation results and sending it as data to be learned, and an autoregressive moving average model considered as a process model is represented by the following equation. An identifier (30) for estimating model parameters by the least squares method, and an input of a state of the process (10), an output of an autoregressive moving average model parameter in the state, and a learning management unit (20). Neural network (40) that learns from given data and appropriate PID for the autoregressive moving average model of process (10) A gain setting device (50) for calculating a controller parameter under a certain standard; and a process manipulated variable based on a deviation between a target value and an output of the process (10) and an output of the gain setting device (50). Equipped with a PID controller (60), identify operating points and process parameters for each environmental condition during operation, and learn them. A process control device wherein the control is performed. Where u (k) is input at k time y (k) is output at k time p and q are orders