JPH09146610A - Multivariable nonlinear process controller - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a robust and efficient process control system. SOLUTION: This process controller is structured as a nonlinear PID feedforward type controller for many variables equipped with a decoupling type external loop controller 10 which performs integrating operation, a neural network multivariable internal loop PD controller 12 which is connected in series with the controller 10, a downstream controller 14 which is connected downstream in series and coupled with a process 16, and a restraint processor 18. An internal loop neural network controller is trained so that future process behavior attains optimum performance when training is repeated. An external loop controller compensates process variation, unmeasured disturbance, and an error in modeling.

Description

Detailed Description of the Invention

[0001]

FIELD OF THE INVENTION This invention relates to industrial process control, and more particularly to a method and apparatus for controlling multivariable non-linear dynamic industrial processes.

[0002]

Industrial automation has been aimed at finding optimal ways to control industrial processes to meet quality as well as manufacturing requirements. However, most of today's industrial processes are complex and require a large number of controlled variables with delays, delayed interactive dynamics and non-linearities. Various process control technologies have been created to handle such complex industrial processes.

State-of-the-art process control techniques rely on monitoring one or more aging characteristics of a process to
It seeks the optimum operation of the process and adjusts the operating parameters of the process. In order to calculate optimal operating parameters, in particular taking into account setpoint changes, system dynamics and turbulence, these prior art techniques use plant process models to predict future behavior of the system. Further, a technique of incorporating this model or a part thereof into a controller structure is also known. The accuracy of these techniques depends on the ability to create accurate dynamic models of the process. In addition, in the case of a process having an uncertainty that cannot be accurately or easily modeled, it may not be possible to generate such a model.

[0004] In recent years, neural networks have been attracting attention as a powerful means for modeling complicated processes. This is because neural networks have the inherent ability to approximate multivariable nonlinear functions. Neural networks also have the advantage of not requiring a complete and accurate understanding of the process. Neural networks, which display processes, are trainable and have the ability to learn by example. Neural networks are also capable of handling delay amounts, ie representing dynamic systems.

Although the application of neural networks in the field of process control has only recently begun,
Various neural network control systems have already been developed. As such a system, for example,
Control systems that use neural networks in a well established model predictive control framework are mentioned. In this type of control system, a controller that employs a process model is usually used in order to obtain an operation amount for bringing a process to a target value. Also,
Process feedback is provided through the process-model mismatch signal provided to the setpoint. As a result, unmodeled perturbations are compensated. This mismatch signal is
6 illustrates the difference between a process output and a modeled process output generated by a neural network of processes.

The controller consists of a neural network model and an optimizer. The neural network model is used to predict the influence of the possible manipulated variable trajectory on the process output with respect to the future time trajectory by considering the current and previous process input and output values. Further, the optimizing device uses this information to select the value of the manipulated variable so that the process output optimally follows the set value and satisfies the predetermined binding force set.

[0007]

There are some limitations to this type of process control system. The first limitation is
In a process where delay is dominant, unmeasured load disturbance cannot be effectively dealt with. Using the model error feedback allows the system to handle time-dominated processes, but unless some additional feedback is provided,
In this way, non-self-regulating processes or open-loop instability processes cannot be stabilized. There is no proportional or derivative feedback, only quasi-integral feedback behavior through the process-model mismatch signal. Further, the output locus of the controller is optimized for each control interval in order to obtain the next change in the manipulated variable. Since this optimization requires substantial calculation time,
The output update interval of the controller becomes large.
This interval adds more dead time to the process dead time and increases the maximum achievable control error for unmeasured load changes.

The present invention seeks to provide a robust and efficient process control system which is capable of accounting for the aforementioned limitations. More specifically, a robust optimal multivariable nonlinear control system adapts non-self-regulating processes and pure dead-time processes, requires no on-line optimization, is compensated to prevent disturbances due to measurement loads, and The aim is to provide a system to combat turbulence not measured with feedback gain.

[0009]

A method and apparatus for use in a robust process control system is provided for efficient control of multivariable nonlinear dynamic processes using neural network controllers, as described below. I do.

The process control system uses a pseudo-inverse neural network model of the process as an inner-loop multivariable PD (proportional-derivative) controller connected in series with a decoupled outer-loop controller that performs an integral operation. By combining in this way, a multivariable nonlinear PID (proportional-integral-derivative) feedforward type controller is realized. So that future process behavior can achieve optimal performance when trained repeatedly
Train the inner loop neural network PD controller. The outer loop controller compensates for unmodeled process variations, unmeasured perturbations, and modeling errors by adjusting the inner loop target value. The last part of the integration operation is disturbed by a constraint management scheme (planning) that stops the integration operation of the controller if a limit is detected in the downstream path. Control quantity and operation quantity are paired (paired)
The operation is useful for ascertaining the downstream path associated with a particular outer loop controller and is used for ascertaining the effective delay of a closed inner loop.

Four embodiments embodying such a process control system will be described in detail below. In the first and second embodiments, the process controller uses a multivariable nonlinear proportional-integral-derivative feedforward controller combined with a constraint management scheme. Here, the neural network is used as an inner loop PD controller coupled with an outer loop integral controller. The inner loop PD controller produces the change in manipulated variable required to bring the process to the desired set level based on the measured value of the process and its derivative. The inner loop PD controller also produces an optimal prediction time for the process used by the outer loop integral controller that performs the integral operation. In the first control embodiment, the inner loop PD controller and the outer loop controller simultaneously input process feedback from a controlled variable sampled at a predetermined measurement rate. In contrast, in the second control embodiment, the inner loop controller inputs the process feedback from the estimator sampled at the faster measurement rate. This allows a faster response to process changes. On the other hand, the integration controller inputs feedback from the control amount and performs integration operation to adjust the set value to compensate for unmodeled error, unmeasured disturbance, and the like.

In the third and fourth embodiments, the neural network controller is used as a PD feedforward controller without an outer loop integral controller or constraint management scheme. In a third embodiment, the controller inputs process feedback from a controlled variable and responds to process changes at a predetermined measurement rate. Further, in the fourth embodiment, the controller inputs feedback from the estimated amount, and realizes quicker response / control to process changes.

[0013]

BEST MODE FOR CARRYING OUT THE INVENTION In order to further clarify the constitution and operation of the present invention described above, the present invention will be described below based on some preferred embodiments. Before describing the embodiments, the symbols used in the following description will be briefly described.

T: An integer representing a predetermined time step of the process. r (t): Outer loop set value-an input signal representing the target value of the process output at time step t. r ^* (t): Inner loop setpoint-the input signal adjusted by the outer loop controller to compensate for unexplained process dynamics, perturbations, and modeling errors. u (t): manipulated variable-a vector of a signal that indicates the amount or condition that changes at the time step t to change the value of the controlled variable. y (t): controlled variable-a vector of process output signals representing the measured process output at time step t. y '(t): derivative of the controlled variable-a vector of signals representing the rate of change of the measured process output at time step t. w (t): estimator-a vector of process output signals representing an estimate of the measured process output at time step t.

W '(t): derivative of the estimator-a vector of signals indicating the rate of change of the estimated process output with time at time step t. v (t): measured load-a vector of signals representing a known disturbance of the measured quantity occurring at time step t. v '(t): derivative of measured load-vector of signals indicating rate of change of measured load signal at time step t. Outer 1 (t): Optimal prediction time-corresponding control amount y _i (t + ^ k
_i (t)) is a real vector that represents the relative time that is most sensitive to the change in the manipulated variable. In this specification, the terms delay time, dead time, and optimum prediction time are used synonymously. This vector is ^ k (t)
Notation.

[0016]

[Outside 1]

D (t): unmeasured load turbulence-a signal representing the turbulence caused by noise or load changes as an unmeasured quantity at time step t, where the above symbols are used in the text or drawings of this specification, these The symbol indicates a vector of signals. For example, u (t) is the vector u of the manipulated variable signal.
Indicates ₁ (t) ... u _n (t) (where i = 1, ..., n).

An embodiment of the process control system using the neural network PD controller will be described below. This controller is off-line and is described in detail in the first and second embodiments of the "multivariable nonlinear control method and apparatus" filed on the same day as the present invention,
Be trained.

The neural network controller is trained to generate, for each control loop, an optimal prediction time and the manipulated variable used to quickly bring the process to the desired level of operation. The optimum prediction time is equal to the effective response time of the controlled variable with respect to the set value. This predicted time represents the future time when the maximum change in the controlled variable will occur in response to a corresponding small change in the controlled variable. Further, the optimum prediction time represents a situation in which a minimum manipulated variable change is required in order to make the future controlled variable reach its target value or set value. If a time other than the optimum prediction time is used, the change in the manipulated variable becomes larger, resulting in an overcorrected state, and the behavior vibrates, and in many cases becomes unstable.

The neural network acts as a proportional-derivative (PD) controller. The neural network inputs a proportional term including a controlled variable or an estimated amount and a measured load amount, and a differential term including a differential value of the controlled variable or the estimated amount and a differential value of the measured load amount.

In the first and second embodiments detailed below, the neural network includes an inner loop P coupled to an outer loop integral controller and a constraint management scheme.
Used as a D controller. The inner loop PD controller produces the change in manipulated variable required to bring the process to the desired setpoint level based on the measured value of the process and its derivative. Furthermore, the inner loop PD
The controller generates an optimal prediction time for the process used by the outer loop integral controller with integral action. In the first control embodiment, the inner loop PD controller and the outer loop controller simultaneously input process feedback from a controlled variable sampled at a predetermined measurement rate. On the other hand, in the second control embodiment, the inner loop controller inputs the faster process feedback from the estimator obtained at the faster sampling rate. This allows a faster response to process changes. On the other hand, the integration controller inputs feedback from the control amount and performs integration operation to adjust the set value to compensate for unmodeled error, unmeasured disturbance, and the like. The last part of the integration operation is disturbed by a constraint management scheme (planning) that stops the integration operation of the controller if a limit is detected in the downstream path.

In the third and fourth embodiments, the process control system comprises only a neural network PD feedforward type controller and no outer loop integral controller or constraint management scheme. In a third embodiment, the controller inputs process feedback from a controlled variable and responds to process changes at a predetermined measurement rate. On the other hand, in the fourth embodiment, the controller inputs feedback from the estimated amount, and realizes quicker response / control to process changes.

FIG. 1 shows a process control system as a first embodiment according to the present invention. In FIG. 1, a process control system includes an outer loop integral type controller 10,
A PD neural network controller 12 connected in series with the controller 10, and a PD controller 12
Controller 14 that is connected in series downstream of (downstream) and is coupled to process 16
And a constraint processor 18, and is constructed as a PID controller. The downstream controller 14 may be configured to include well-known control elements such as valves, actuators and the like. The outer loop controller 10 has an outer loop setpoint r (t) 20 and a controlled variable y (t) 22 fed back from the process, and in addition to, but not in all cases, by the neural network controller 12. Optimal predicted time generated
^ k (t) 28 is input. Outer loop controller 10
Adjusts the outer loop setpoint r (t) 20 to produce the inner loop setpoint r ^* (t) 24. The outer loop controller 10 is a general-purpose computer, a microprocessor,
Or, The Foxboro Company
It is realized by a special purpose controller such as a 760 series controller. On the other hand, the neural network controller 12 is software that runs on the CPU.
It is realized as hardware composed of electronic circuits, or partially as a software and partially as a hybrid system of hardware.

The inner loop controller, that is, the neural network controller 12, has an inner loop set value r ^* (t) 24, a controlled variable y (t) 22 fed back from the process, and a derivative of the controlled variable with respect to time. y ′ (t) 30, the measured load amount v (t) 38, and the time-dependent derivative v ′ (t) 36 of the measured load amount are input. Further, the inner loop controller 12 outputs the manipulated variable u (t) 26 and the optimum predicted time ^ k (t) 28. The manipulated variable u (t) 26 is given to the downstream controller 14 and output as the value 42 transferred to the process 16.

When the downstream controller 14 limits any one of the manipulated variables 26 (for example, the full open / close valve), a logic signal 34 indicating this limitation is output from the downstream controller 14 to the restraint processor 18. Given. The restraint processor 18 uses the signal 4
It responds to this limiting action in a predetermined way, selected from among a number of ways based on 4. A signal 32 indicative of the response to the limiting operation is output from the constraint processor 18 to the appropriate outer loop controller. Input of the manipulated variable change required to bring the process 16 to the desired set level 42, the measured load v (t) 38 and the unmeasured disturbance d.
Based on the input of (t) 40, the process 16 performs continuous processing. The controlled variable y (t) 22 measured at an appropriate measurement speed is transmitted to the inner loop controller 12 and the outer loop controller 10. The functions of these elements will be described in detail later.

The outer loop integral type controller is composed of n outer loop controllers 10 (C ₁ , ... C _n ), and each outer loop controller corresponds to a control amount and an operation amount of each pair. The outer loop controller 10 adjusts the outer loop set value r (t) 20 to set the inner loop set value 24r ₁ ^* (t).
... r _n ^* (t) is output. The outer loop controller 10 includes an outer loop set value r (t) 20 and a control amount y (t) 2
2 and are input. Furthermore, the optimum prediction time ^ k (t) output from the neural network controller 12
28 may be used to determine the integration time constant used in the outer loop integration type controller. By using the optimal prediction time, the integral controller is automatically tuned without the need for manual tuning operation,
It has the advantage that it can be adapted to the dynamic conditions of the process.

The outer loop controller 10 may perform either a pure integration operation, a PI (proportional-integral) operation, or another type of integration operation. For example, the i-th PI operation implementation is mathematically expressed as:

R ^* _i (t) = x _i (t) + .2 [r _i (t) -y _i (t)] (1) where the integral term is represented by x _i (t), and Calculated according to the formula. In case of initialization, x _i (0) = y _i (0) (2) If there is a restriction in the downstream (downstream), then x _i (t + 1) = x _i (t) or x _i (t + 1) _i (t) = x _i (t-1) (3) Otherwise, x _i (t + 1) = x _i (t) + (r _i (t) -x _i (t)) / (.3 * ^ k (t)) (4)

When the outer loop controller 10 adopts only the pure integral operation, the i-th outer loop integral operation is mathematically expressed as follows. r ^* (t) = r ^* (t-1) + [(r (t) -y (t)) * (1 / (2 * ^ k (t) -1))] (5)

In any case, the outer loop controller 1
From 0, the inner loop set value r ^* (t) 24 is output,
It is transmitted to the neural network controller 12.

In the neural network 12, in addition to the inner loop set value 24, the measured load amount 38, the measured load amount derivative 36, the controlled variable 22, and the controlled variable derivative 30.
Is input. By feeding back these measured values, a feedforward type proportional-derivative (PD) controller is realized. The neural network controller 12 determines the optimum manipulated variable 26 based on these inputs.
And the optimum prediction time 28. The optimum prediction time 28 is
Is transmitted to the outer loop controller 10, or
It is thrown away as it is. On the other hand, the selected value of the manipulated variable is transmitted to the downstream controller 14 or directly transmitted to the process 16.

Next Downstream Controller 14
Is used when it is necessary to adjust the manipulated variable in the control system. When either the neural network controller 12 or the downstream controller 14 sets the operation amount to the threshold value or a value higher than the threshold value, the logic signal 34 representing the limited operation amount is output. For example, if a control element such as a valve is completely open and the controller requirements cannot be met,
There are restrictions. The constraint processor 18 uses this logic signal 3
4 and receives a restraint signal 32 for instructing a predetermined restraint operation.
Output to the outer loop controller 10 is started. The predetermined restraint action to be performed is transmitted via the restraint signal 32 to the outer loop controller having the corresponding limited manipulated variable.

FIG. 2 is a flow chart showing the steps performed by the constraint processor 18. First, in step 50, the logic signal 34 is tested to determine if the corresponding manipulated variable or its downstream (downstream) path signal is restricted (ie, is the manipulated variable or signal greater than or equal to a threshold). Is determined). If the signal is restricted, select one of the three possible options for constraint management operation (step 5).
4). In this case, signal 44 selects a particular option. Which option is selected depends on the type of application. If the first option is chosen, then in step 56 the controller's signal 3 is signaled to mark the relevant manipulated variable and in step 58 to temporarily freeze the integral action of the corresponding controller.
Set 2. If you chose the second option, step 6
At 0, the controller signal 32 is set so that the corresponding controller setpoint r (t) is equal to the measured value y (t).
Set. For example, if u _i (t) is the upper limit value, r _i with respect to the corresponding outer loop controller C _i (t) = y _i
Set to (t). If the third option is chosen, in step 62 the controller is re-initialized to the measured value of the integral term of the corresponding controller, i.e. to achieve x _i (t) = y _i (t). Signal 32 is set. If the manipulated variable is not limited, the process proceeds to step 52, and it is confirmed whether or not the manipulated variable was previously limited. If so, then at step 64, the controller signal 32 is set to restart the integral operation of the corresponding controller.

As shown in FIG. 1, each outer loop controller outputs a new inner loop setting value 24.
The neural network controller 12 receives these inner loop settings 24 and calculates a new manipulated variable 26. In the next downstream (downstream) controller 14, further control operation is performed, and the value 42 obtained by adjusting the operation amount is output. Process 16
Receives the value 42 and adjusts its operating parameters. The measured load 38 and the unmeasured turbulence 40 are continuously input to the process 16.

The operation in the first control embodiment is as follows.
It is. At a predetermined time interval, the controlled variable y (t) 22 and
The measurement load v (t) 38 is detected and measured by an appropriate measuring device.
Is determined. The measured value of the controlled variable 22 is
・ Network controller 12 and outer loop controller
It is fed back to LA 10. Normally, the
Process characteristics include flow rate, pressure, temperature, and fluid chemistry.
Composition, etc. The derivatives of these signals are
By measuring with a measuring instrument or with a processor.
It is obtained by arithmetic. The obtained differential signal is
It is sent to the network controller 12. neural
-The network controller 12 can quickly input signals.
In response, the manipulated variable u (t) 26 and the optimum predicted time ^ k (t)
28 and are output. In the initial stage, the adjustment set value r
^*(T) 24 is set to be equal to the set value r (t)
Is done. Neural network controller 12?
The manipulated variable u (t) 26 output from
It is input to the controller 14. On the other hand, the outer loop controller
The roller 10 performs a slower integral control operation and adjusts
Set value r * (T) 24 is output. This adjustment set value 24
In the next control cycle
Used by the trolley 12. Downstream
The controller 14 executes the control operation and adjusts the operation amount.
42 to process 16. Manipulated variable or its
Limit if the downstream (downstream) volume is limited
An operation signal 34 is output to the restraint processor 18.
You. The restraint processor 18 determines the appropriate restraint action.
Then, the determined motion is transmitted through the restraint signal 32 to an appropriate external rule.
Output to the loop controller. Following this whole process
Repeat in the measurement cycle of.

FIG. 3 is a block diagram showing a second embodiment of the process control according to the present invention. As shown in FIG. 3, the physical configuration of the second control embodiment is the same as that of the first embodiment. The process control system includes an outer loop integral type controller 10, a PD neural network controller 12 connected in series with the controller 10, and a P neural network controller 12.
It is constructed as a PID controller comprising a downstream controller 14 connected in series downstream of the D controller 12 (downstream) and coupled to the process 16 and a constraint processor 18. The control amount y (t) 22 measured at predetermined time intervals is input to the outer loop controller 10 as process feedback.

The difference from the system of the first embodiment lies in the measured value of the process fed back to the inner loop controller (neural network controller) 12. In the second embodiment, the estimated quantity w (t) 46 measured at a shorter time interval than the measurement of the controlled quantity and its derivative w ′.
(T) 48, the measured load amount v (t) 38, and its derivative v ′ (t) 36 are input to the inner loop controller 12 as process feedback. The estimated amount is an estimated approximate value of the control amount. This estimator is usually measured continuously without detailed analysis. Taking a distillation column as an example, the controlled amount is, for example, the chemical composition of the mixture, and the estimated amount is the tray temperature at which the chemical composition can be estimated. The tray temperature is sampled continuously, whereas the controlled variable sampling interval is significantly open due to the need for detailed analysis. By giving the estimated amount as an input value to the inner loop controller 12, the control speed of the process can be made faster than that of the integral type controller, and as a result, it becomes possible to quickly respond to the process change.

The operation of the system of the second embodiment will be described below. The estimated amount w (t) 46 is detected and measured by an appropriate measuring device at a higher sampling rate. The measured estimator w (t) 46 is used as the inner loop controller 12
Is input to Also, the derivative w '(t) 48 of the estimator
Are obtained by measurement using an appropriate measuring device or by calculation using a processing device and sent to the inner loop controller 12. Further, the measured load amount v (t) 38 and the measured load amount derivative v ′ (t) 36 are similarly measured at a higher sampling rate and input to the inner loop controller 12. The inner loop controller 12 calculates the manipulated variable u (t) 26 and the optimum predicted time ^ k (t) 28 based on these input values. Manipulated variable u (t) 26
Is sent to the downstream controller 14 and is output as the value 42 affecting the process 16. On the other hand, the optimum prediction time ^ k (t) 28 is transmitted to the outer loop controller 10. The inner loop controller 12 takes in the value of the adjustment set value r ^* (t) calculated in the previous cycle by the outer loop controller 10, and continuously performs this processing.

Simultaneously with the above operation, at a predetermined measurement interval,
The controlled variable y (t) 22 is detected / measured by an appropriate measuring device and fed back to the outer loop controller 10. Here, the control amount is measured at a sampling rate slower than the sampling rate of the estimated amount. Further, the derivative y ′ (t) 30 of the controlled variable is obtained by measurement using an appropriate measuring device or by calculation using a processing device, and is given to the outer loop controller 10. Further, the optimum prediction time ^ k (t) 28 obtained in the previous cycle and output from the inner loop controller 28 is input to the outer loop controller 10. The outer loop controller 10 executes the integral control operation using the input optimum prediction time, obtains the integral time constant, and further adjusts the set value r ^*.
(T) 24 is generated and output to the inner loop controller 12.

FIG. 4 is a block diagram showing a third embodiment of the process control system according to the present invention. In the third embodiment, the neural network controller performs only PD control operation. This type of control system has a simpler structure and is suitable for processes that do not require integral control action. As shown in FIG. 4, the control system comprises a neural network PD controller 12 and a downstream controller 14 serially connected downstream (downstream) of the PD controller 12 and coupled to the process 16. The neural network PD controller 12 has a measured load amount v (t) 38.
And the derivative v ′ (t) 36 of the measured load amount, the outer loop set value r (t) 20, the control amount y (t) 22, and the derivative y ′ (t) 30 of the control amount. Is entered. neural·
The network controller 12 executes PD control based on these input values, and generates the manipulated variable u (t) 26 and the optimum predicted time ^ k (t) 28. Manipulated variable u (t) 2
The output of 6 is the downstream controller 14
On the other hand, the output of the optimum prediction time ^ k (t) 28 is not used in this embodiment and is discarded as it is. The downstream controller 14 controls the operation amount u.
(T) 26 is processed and the value 42 is output. Process 1
The value 42 is used to adjust the process 16 so that 6 reaches the desired setpoint. Further, the process 16 continuously includes the measured load v (t) 38 and the unmeasured disturbance d
(T) 40 is input. At a predetermined measurement interval, the controlled variable y (t) 22 is detected and measured by an appropriate measuring device,
It is fed back to the neural network controller 12. As described above, the process control system according to the present embodiment has a feedforward type proportional-derivative (P
D) Acts as a controller.

FIG. 5 is a block diagram showing the process control system of the fourth embodiment according to the present invention. As shown in FIG. 5, the configuration of the fourth embodiment is similar to that of the control system of the third embodiment. That is, in the process control system of the fourth embodiment, the neural network PD controller 12 that generates the operation amount u (t) 26 and the operation amount u
A downstream controller 14 that outputs a value 42 based on the input of (t) 26 and a process 16. The output value 42 adjusts the process 16 so that the process 16 reaches the desired set point. The difference from the third embodiment is that the neural network controller 1
2 is in the process feedback input. In the fourth embodiment, the estimated amount w (t) 46 detected and measured by an appropriate measuring device at a time interval shorter than the control amount measurement.
, The derivative w ′ (t) 48 of the estimated amount, and the measured load amount v
The (t) 38 and the measured load derivative v ′ (t) 36 are input to the neural network controller 12 as process feedback. Such fast feedback enables the controller of this embodiment to quickly respond to process changes.

The operation of the control system of the fourth embodiment will be described below. The estimated amount w (t) 46 is detected and measured by an appropriate measuring device at a higher sampling rate. The measured estimator w (t) 46 is input to the neural network controller 12. Further, the derivative w ′ (t) 48 of the estimated amount is obtained by measurement using an appropriate measuring device or by calculation using a processing device, and is sent to the neural network controller 12.
Further, the measured load amount v (t) 38 and the measured load amount derivative v ′ (t) 36 are similarly measured at a higher sampling rate and input to the neural network controller 12. The neural network controller 12 determines the manipulated variable u based on these input values.
(T) 26 and the optimum prediction time ^ k (t) 28 are calculated.
The manipulated variable u (t) 26 is sent to the downstream controller 14 and output as a value 42 that affects the process 16. On the other hand, the optimum prediction time ^ k (t) 28 is
It is not used in this embodiment and is discarded as it is. The neural network controller 12 continuously performs this processing by using the input of the set value r (t) 20.

The configurations of the above-described embodiments can be modified and changed to suit various types of processes. For example, in processes that do not require a derivative, the derivative of the controlled variable and the derivative of the measured load may be omitted. This is preferably used, for example, in a process in which dead time is dominant or a process in which large measurement noise exists.

Furthermore, in all the above-described embodiments, the function of the neural network, including the function as the controller, can be realized by another data processing structure. Examples of such structures include, but are not limited to, non-linear function creation routines or characterizers, fuzzy logic processors, look-up tables, estimator logic, and a series of suitable values for the signals generated. A processor using the formula of

[0045]

As described in detail above, the present invention provides a robust process control system for controlling a multivariable nonlinear dynamic process using a neural network PD controller.

The process control system of the present invention has the following advantages. The first advantage is that it is effective in dealing with unmeasured load disturbances and that robust control can be realized in pure dead-time processes or non-self-controlled processes. This is due to the direct proportional-derivative feedback from the process to the controller. Furthermore, the actual control measurement value is
Being fed back into the network, the overall process control system is more effective in compensating to prevent turbulence and more effectively dealing with unmeasured turbulence due to feedback gain.

Second, since the neural network controller is trained and optimized off-line, it does not require on-line computational overhead. Third, the neural network controller is given an input value independently for the proportional term and the derivative term. The input values of the proportional term are the control amount or the estimated amount, the inner loop set value, and the measured load amount. The input values of the differential term are the derivative of the controlled variable and the derivative of the measured load amount. These input values essentially give the controller different proportional and derivative gains for the proportional and derivative terms, respectively, and as a result, the controller is responsible for changes in setpoints and disturbances in the measured load. It is possible to respond to both quickly and accurately without making an overshooting response.

The use of estimators also allows the inner loop controller to respond more quickly to process changes. By the control based on the predicted time that requires the minimum change of the manipulated variable, it is possible to prevent the process vibration and the overshoot and realize the stable control. In addition, by using the optimal prediction time to determine the time constant used by the outer loop controller, the integral controller is automatically tuned without the need for manual tuning operations to allow process dynamics. Can be adapted to the conditions.

[Brief description of the drawings]

FIG. 1 shows, as a first embodiment of the present invention, a process control system in which a neural network inner loop PD controller is operated by a controlled variable measurement and is connected in series with an outer loop controller and a constraint management scheme. Block Diagram.

FIG. 2 is a flow chart showing steps in a restraint management scheme used in the process control systems of the first and second embodiments.

FIG. 3 is a block diagram showing a process control system in which a neural network inner loop PD controller is operated by an estimator and is connected in series with an outer loop controller and a constraint management scheme as a second embodiment of the present invention. .

FIG. 4 is a block diagram showing a process control system consisting of only a neural network PD controller, which is operated by a controlled variable measurement value, as a third embodiment of the present invention.

FIG. 5 is a block diagram showing a process control system including only a neural network PD controller, which is operated with an estimated amount, as a fourth embodiment of the present invention.

[Explanation of symbols]

 10 Outer Loop Controller 12 Neural Network Controller 14 Downstream Controller 16 Process 18 Constraint Processor

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical display location G06F 15/18 550 G06F 15/18 550E // G05B 11/36 501 G05B 11/36 501Q

Claims (29)

[Claims] 1. An apparatus for controlling a process having a process input and a process output including a controlled variable, wherein the process changes the process with respect to the controlled variable according to a manipulated variable, the apparatus comprising: a). Measuring means for measuring a value of any one of the process outputs to generate a process output signal indicating the measured value; and b) setting for inputting a set value indicating a target value of the controlled variable. Value means, c) control means connected to the measuring means and the process for generating a predetermined value of the manipulated variable as a function of the set value and the process output signal, in accordance with the optimum predicted time signal. Control means comprising a non-linear function creation routine trained to calculate a predetermined value of the manipulated variable, (wherein the optimum predicted time signal corresponds to a change in the set value; The effective response time of the process is substantially indicated, and the predetermined value of the manipulated variable indicates the change of the process necessary for moving the control amount toward the set value that changes depending on the optimum prediction time) d ) An actuator means connected to the control means and the process for applying a predetermined value of the manipulated variable to the process. 2. An apparatus for controlling a process having a process input and a process output including at least one controlled variable, the process varying the process with respect to the controlled variable in response to at least one manipulated variable. The device comprises: a) measuring means for measuring at least one value of the process output and generating at least one process output signal indicative of the measured value; and b) indicating a target value of the controlled variable. Set value means for inputting at least one set value; c) connected to the measuring means and the process, as a function of the set value and the process output signal, a predetermined value of the manipulated variable and at least one A control means for generating an optimum predicted time, which is pre-trained to determine a predetermined value of the manipulated variable and the optimum predicted time. Control means using a nonlinear function creation routine, wherein (wherein the optimum prediction time substantially indicates an effective response time of the process with respect to a change in the set value, and the predetermined value of the manipulated variable is the optimum prediction time). The change of the process required to obtain the value of the control amount so as to approach the set value that changes with time is shown.) D) Integral control according to the process output signal, the set value, and the optimum predicted time. And integral control means for calculating internal setpoints to compensate for process variations and outputting said internal setpoints for use by said control means, where: E) Actuator means connected to the control means and the process for applying the manipulated variable to the process, Device, characterized in that it comprises. 3. An apparatus for controlling a process having a process input and a process output including at least one controlled variable and at least one estimator, the process comprising:
A first for varying a process with respect to said controlled variable in response to at least one manipulated variable, said device: a) measuring the value of the process output and generating at least one controlled variable indicative of the measured value; Measuring means; b) second measuring means for measuring the value of the process output for approximating the controlled variable to generate at least one estimated variable indicative of the measured value; and c) the control. Set value means for inputting at least one set value indicating a target value of the quantity; d) a predetermined value of the manipulated variable as a function of the set value and the estimated quantity, connected to the measuring means and the process. Control means for generating a value and at least one optimal prediction time, a non-linear function generating routine pre-trained to determine the predetermined value of the manipulated variable and the optimal prediction time. (Wherein the optimum predicted time substantially indicates the effective response time of the process with respect to the change in the set value, and the predetermined value of the manipulated variable changes depending on the optimum predicted time). The change of the process necessary for obtaining the value of the control amount approaching the set value is shown) e) Integral control is performed according to the control amount, the set value, and the optimum predicted time to compensate for the process change. An integral control means for calculating an internal set value for outputting the internal set value so that it can be used by the control means; (wherein, the optimum prediction time is used for obtaining an integral time constant) F) Actuator means connected to the control means and the process for applying the manipulated variable to the process. 4. The measuring means comprises means for generating a differential process output signal representing a time derivative of the process output or the estimator, and the control means comprises the differential process output signal or the estimator. 4. The apparatus according to claim 1, wherein the operation amount is calculated as a function of, and PD (proportional-derivative) control is performed. 5. Further comprising means for inputting a measured value of the process input, wherein the control means calculates the manipulated variable as a function of the process input and performs additional feedforward control.
The device according to any one of claims 1, 2, and 3, wherein: 6. An apparatus according to claim 5, wherein said measuring means measures a process input representing a measured load disturbance or a derivative of the measured load disturbance with respect to time. 7. Further, integral control is performed according to the process output signal, the set value, and an integration time constant, an internal set value is calculated to compensate for a process change, and the internal set value can be used by the control means. The apparatus according to claim 1, further comprising integral control means for outputting the internal set value. 8. The control means generates the optimum prediction time in order to obtain the integration time constant used by the integration control means.
The described device. 9. The apparatus according to claim 2, wherein the integral control means is selected from a PID controller, an integral controller, and a PI controller. 10. The integration control means further comprises means for detecting a limit of the manipulated variable, and performs control according to the limit of the manipulated variable.
The apparatus according to any one of 3 and 5. 11. The apparatus according to claim 1, wherein the process output includes an estimated amount that approximates the control amount. 12. The apparatus according to claim 1, wherein the controlled variable comprises a measured process output. 13. The device is a multivariable controller,
The process output signal includes a plurality of process output signals, the manipulated variable includes a plurality of manipulated variables, the set value includes a plurality of set values, the optimum prediction time includes a plurality of optimum prediction time, The device according to claim 1, characterized in that 14. A multivariable non-linear controller for controlling a process comprising a process output including at least one controlled variable and an actuator influencing the process, comprising: a) inputting a process output signal indicative of at least one process output. B) means for inputting at least one set value signal indicating a target value of the controlled variable, c) depending on the process output signal and the set value signal,
A control means having a non-linear function generating routine for outputting at least one predetermined value of the manipulated variable affecting the process, wherein the control means includes i) the process output signal, the set value, and an integration time constant. And performing at least one adjusted set point indicating the result of the integral control as a function of, and ii) a nonlinear function creation routine as a function of the process output signal and the adjusted set point. And outputting an optimum prediction time representing a response time of the controlled variable with respect to the adjusted set value, (wherein the integration for performing the integration control is performed using the optimum prediction time). A time constant is calculated, and the non-linear function creation routine is configured to set a predetermined value of the manipulated variable that affects the control amount that changes depending on the optimum prediction time. Iii) Transmitting a predetermined value of the manipulated variable to an actuator that affects the process, a multivariable nonlinear controller. 15. A multivariable non-linear controller for controlling a process comprising a process output including a controlled variable and an estimated variable and an actuator that affects the process, comprising: a) inputting at least one process output signal. Means, b) means for generating at least one estimated quantity that approximates the control quantity, c) means for inputting at least one set value indicating a target value of the control quantity, and d) Control means comprising a non-linear function creation routine for outputting at least one predetermined value of the manipulated variable affecting the process in accordance with the process output signal and the set value, the control means comprising: i) Integral control is performed using the process output signal, the set value, and the integral time constant, and at least one adjusted value indicating a result of the integral control. Outputting a constant value, ii) driving a non-linear function creation routine in response to the estimated amount and the adjusted set value to represent an optimal prediction representing the response time of the controlled variable to the adjusted set value. Outputting a time, (wherein, the optimum prediction time is used to calculate the integration time constant for performing the integration control, and the nonlinear function creation routine is changed depending on the optimum prediction time. Outputting a predetermined value of the manipulated variable that influences the controlled variable so as to approach the set value) iii) transmitting the predetermined value of the manipulated variable to an actuator that affects the process. Characteristic multi-variable nonlinear controller. 16. The means for inputting the at least one process output signal further comprises means for calculating and generating at least one derivative of the process output signal or the estimator with respect to time. 16. The controller according to claim 14, wherein the predetermined value is generated as a derivative of the process output signal or the estimated amount by using a function creation routine. 17. A means for inputting at least one measured load quantity representing a disturbance of the measured load, wherein the predetermined value is generated as a function of the measured load quantity by using the non-linear function creating routine. The controller according to claim 14 or 15, characterized in that. 18. Means for inputting said at least one measured load further comprising means for calculating and generating at least one differential measured load representing a time derivative of said measured load. The controller according to claim 14 or 15, wherein the non-linear function creating routine is executed according to a function including the differential measured load amount. 19. The PI for the step of performing the integral control.
16. The controller according to claim 14, wherein the controller is selected from a D controller, an integral controller, and a PI controller. 20. The step of performing the integral control further comprises a step of detecting a limit of each manipulated variable and performing an integral control according to the limit of the manipulated variable. The controller according to any one of 1. 21. The apparatus according to claim 1, wherein the non-linear function creating routine is a neural network. 22. A method for controlling a process comprising a process input, a process output including at least one controlled variable, and an actuator influencing the process, the method utilizing a non-linear function generating routine. And calculating at least one manipulated variable used by the actuator to influence the controlled variable, the method comprising: a) measuring the process output to generate at least one process output signal indicative of the measured value. Generating; b) inputting at least one set value indicating a target value of the controlled variable; c) driving the non-linear function generating routine as a function of the output signal and the set value to perform the operation. Outputting a predetermined value of the quantity, (wherein the non-linear function creation routine is based on an optimal prediction time, Training in advance to determine a predetermined value of the cropping amount, the optimum predicted time indicates a response time of the controlled variable with respect to the set value, and the predetermined value of the manipulated variable changes according to the optimum predicted time. Indicating a change in the process that affects the controlled variable to approach a value) d) transmitting the manipulated variable to the actuator and applying the manipulated variable to the process. How to characterize. 23. The step of measuring the process output comprises the step of calculating at least one derivative of the process output with respect to time to generate at least one differential process output signal indicative of the derivative. 23. The method of claim 22, wherein driving a non-linear function creation routine comprises outputting the manipulated variable according to a function that includes the differentiated process output signal. 24. The step of measuring the process output comprises the step of measuring the load disturbance of the process to generate at least one measured load quantity indicative of the load disturbance. 24. The method of claim 23, wherein driving comprises outputting the manipulated variable according to a function that includes the measured load. 25. The step of measuring the load disturbance comprises the step of calculating a time derivative of the at least one measured load to generate at least one differential measured load indicative of the derivative. 23. The method according to claim 22, wherein the step of driving the non-linear function creation routine comprises the step of outputting the manipulated variable according to a function including the differential measured load amount. 26. The method of claim 23, wherein the process output is at least one controlled variable indicative of a measured process output or at least one estimated variable approximating the controlled variable. 27. The non-linear function creation routine is a neural network.
3. The method according to 3. 28. A system for controlling a process comprising a process input, a process output including a plurality of controlled variables, and an actuator for influencing the process, wherein a) a plurality of used by the actuator for influencing the process. The first control means for generating a predetermined value of the operation amount of, i) measuring the disturbance of the load of the process, generating a plurality of measured load quantities indicating the disturbance of the load, further, the time of the measured load disturbance Determining a derivative with respect to, and generating a plurality of differential measured loads representative of the derivative, ii) measuring a plurality of process outputs and generating a plurality of process output signals indicative of the measured process outputs, Further, determining a derivative of the process output signal and generating a plurality of differential process output signals representing the derivative, iii) the operation. Used by the first control means to determine a predetermined value of the quantity, inputting a plurality of adjusted set values indicating the value of the controlled quantity, iv) the adjusted set value, and the measurement Generating a predetermined value of the manipulated variable and a plurality of optimum prediction times as a function of the load quantity, the differential measurement load quantity, the estimated quantity, and the differential estimated quantity; Indicates the response time of the controlled variable with respect to the adjusted set value, and the predetermined value of the manipulated variable affects the controlled variable so as to approach the adjusted set value that changes according to the optimum prediction time. A) a first control means comprising: v) transmitting a predetermined value of the manipulated variable to the actuator and applying it to the process; b) generating the adjusted set value. Let first In the control means, i) a step of providing a desired set value signal indicating a target value of the process and storing the set value signal, and ii: a controlled variable indicating the measured value by measuring the process. And iii) determining whether or not the manipulated variable is limited,
Generating a logic signal indicating the result, iv) integral control as a function of the desired set value, the controlled variable, the logic signal, and the optimum prediction time output by the first control means. Calculating the adjusted setpoint by performing the step of generating the adjusted setpoint used by the first control means to generate the manipulated variable affecting the process, A second control means comprising :. 29. The system of claim 28, wherein the process output signal is a controlled variable indicative of a measured process output or an estimated variable approximating the controlled variable.