JPH09101803A - Control knowledge generation method and apparatus - Google Patents

Abstract

(57) [Abstract] [Problem] If the automatic control system cannot exhibit the desired performance unless manual correction is added due to disturbance or the like, the control knowledge necessary to reduce or eliminate the burden of manual correction is provided. It is possible to automatically generate by using the learning function. SOLUTION: Time-series data relating to an operation amount to the plant and a control amount of the plant when trial-controlling the plant is acquired, and the quality of the control state is evaluated based on the acquired time-series data, and a good evaluation is obtained. Time series data (DA
TA-g) and time-series data (DAT)
A-b) is extracted from the time-series data, and the control amount is bad based on the extracted time-series data with good evaluation (DATA-g) and time-series data with bad evaluation (DATA-b). Then, the operation amount for correcting to a good state is obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control knowledge generation method and apparatus suitable in the field of plant control,
Especially, if the automatic control system cannot achieve the desired performance unless manual correction is added due to external disturbances, etc., the learning function is used to automatically control the control knowledge necessary to reduce or eliminate the burden of manual correction. The present invention relates to a control knowledge generation method and device capable of generating a control knowledge.

[0002]

2. Description of the Related Art There are various cases in which humans are forced to intervene in the operation of plants. One of the reasons is
Since the plant to be controlled operates under various conditions, it is difficult to completely entrust control to a PLC (Programmable Logic Controller), a temperature controller, a computer and the like.

Such a case is remarkably observed in, for example, the operation of a batch reactor (hereinafter referred to as a jacket tank) which is frequently used in a chemical plant. The operation of the jacket tank relies heavily on the operating ability of the worker due to the difficulty of automatic control. That is,
When the polymerization reaction is carried out in the jacket tank, the generation of the heat of polymerization promotes the activation of the reaction, which accelerates the generation of the heat of polymerization.
Controlling the reaction is not easy. In addition, various materials may be added during the manufacturing process, and in such a case, there is a disturbance element in the process that the reaction state changes with the addition, so from this point However, controlling the reaction is not easy.

Therefore, an intermediate parameter is used as the setting parameter of the PID controller used for automatic control of this type of jacket tank, and the operator can manually correct the parameter to ensure controllability. . Under such circumstances, the psychological burden on the worker is enormous, and the quality of the manufactured product depends on the skill of the worker, so that the variation in quality is inevitable.

By the way, as such an automatic operation method of a jacket tank, it has been proposed to build a model in advance for each product type and perform model reference control (for example, "System / Control / Information Vol. 35, No. 35"). No. 3 (1991) pp. 23-28 "," Journal of the System Control Information Society Vol. 7 No. 11p "
p. 448-460 1994 ").

In this method, it is necessary to set the chemical reaction and the structure of the heat balance of the reactor to create a model with the desired accuracy. However, in the batch type reactor, the number of products manufactured per reactor reaches several tens to several hundreds, and the number of man-hours is enormous if all models are created. This is model reference control
Since the control performance depends largely on the accuracy of the model, it is necessary to examine the precise heat balance structure for each product, and the number of model parameters also increases, which increases the difficulty of parameter estimation work. This is because Therefore,
Although it is possible to apply the model reference control to a line that manufactures a very small number of products, it was practically impossible to introduce it to a line that has many products.

On the other hand, in Japanese Patent Laid-Open No. 5-127705, operation history data such as operation content and plant value before operation is converted into operation knowledge data for outputting operation content in response to input of plant value, Based on this operation knowledge data and the current plant value, the operation amount of a skilled operator is inferred.
A monitoring man-machine device is disclosed that is adapted to automate operations during non-steady-states by executing it, and a learning method of converting such operation history data into operation knowledge data can be applied to jacket tank control. For example, unlike the case of model reference control, complicated model reference creation work is considered unnecessary. However, in the water level control specifically described in the publication, the valve is fully opened when the amount of water in the tank becomes empty (0), or fully closed when the amount of water in the tank becomes full (100). As mentioned above, the assumption is that a unique relationship between the manipulated variable and the controlled variable is always established (water injection starts when the valve is opened and water injection stops when the valve is closed). Therefore, the method cannot be immediately applied to process control of a jacket tank, where the relationship between the manipulated variable and the controlled variable may fluctuate.

[0008]

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned technical background, and its object is to provide an automatic control system that is not intended unless manual correction is added due to disturbance or the like. To provide a control knowledge generation method and device capable of automatically generating control knowledge necessary for reducing or eliminating the burden of manual correction when performance cannot be exhibited, by using a learning function. It is in.

[0009]

The invention according to claim 1 of the present application is a control system for controlling a plant by adding an operation amount corresponding to the state of the plant to the plant, when the trial control of the plant is performed. A time-series data acquisition step of acquiring time-series data relating to an operation amount to the plant and a control amount of the plant, a control state evaluation step of evaluating quality of a control state based on the acquired time-series data, and the control state A time-series data extraction step of extracting time-series data (DATA-g) for which a good evaluation was obtained and time-series data (DATA-b) for which a bad evaluation was obtained in the evaluation step; Time-series data with good evaluation (DATA-g) extracted by the time-series data extraction step and time-series data with poor evaluation (DA-g) Characterized by comprising a knowledge data generation step of generating a knowledge data for determining the manipulated variable to correct a good state control amount based on A-b) from bad.

The invention according to claim 2 of this application is a control system for controlling a plant by adding the manipulated variable according to the state of the plant to the plant, and the manipulated variable to the plant when the plant is trial-controlled. Time-series data acquisition step of acquiring time-series data regarding the control amount of the plant, a control state evaluation step of evaluating the quality of the control state based on the acquired time-series data, and good evaluation in the control state evaluation step A time-series data extraction step of extracting the obtained time-series data (DATA-g) and the time-series data (DATA-b) for which a bad evaluation has been obtained from the time-series data, and the time-series data extraction step. The extracted time series data with good evaluation (DATA-g) and the time series data with bad evaluation (DA
Based on TA-b), a knowledge data generating step of generating knowledge data for obtaining a manipulated variable for correcting the control amount from a bad state to a good state, and the plant using the knowledge data generated. A time-series data re-acquisition step of controlling and re-acquiring time-series data regarding the control amount and the operation amount, and evaluating the control state indicated by the re-acquired time-series data, and when the evaluation value is good in the evaluation. If it is better than the series data (DATA-g),
Re-generating the knowledge data by using the re-acquired time-series data instead of the time-series data (DATA-g) which has been evaluated as good until then.

According to a third aspect of the present application, in the invention according to the first or second aspect, the knowledge data is a time-series data (DATA-g) with a good evaluation when the control amount. Series data and time series data with poor evaluation (D
Data of the difference between the control amount in ATA-b) and the time-series data, the operation amount in the time-series data with good evaluation (DATA-g) and the operation amount in the time-series data with bad evaluation (DATA-b). It is a fuzzy rule created based on the data of the difference between and.

According to a fourth aspect of the present application, in the invention according to any one of the first to third aspects, the plant is a system including a jacket tank and a temperature control device, and the control is performed. The amount is an internal temperature and an external temperature of the jacket tank, and the operation amount is an external temperature set value.

The invention according to claim 5 of this application is a control system for controlling a plant by adding the manipulated variable according to the state of the plant to the plant, and the manipulated variable to the plant when the plant is trial-controlled. Time-series data acquisition means for acquiring time-series data regarding the control amount of the plant, control state evaluation means for evaluating the quality of the control state based on the acquired time-series data, and good evaluation in the control state evaluation means The time-series data extracting means for extracting the obtained time-series data (DATA-g) and the time-series data (DATA-b) for which a bad evaluation is obtained from the time-series data, and the time-series data extracting means. The extracted time series data with good evaluation (DATA-
knowledge data generation means for generating knowledge data for obtaining an operation amount for correcting the control amount from a bad state to a good state based on g) and time-series data with bad evaluation (DATA-b). It is characterized by doing.

The invention according to claim 6 of this application is a control system for controlling a plant by adding the manipulated variable according to the state of the plant to the plant, and the manipulated variable to the plant when the plant is trial-controlled. Time-series data acquisition means for acquiring time-series data regarding the control amount of the plant, control state evaluation means for evaluating the quality of the control state based on the acquired time-series data, and good evaluation in the control state evaluation means The time-series data extracting means for extracting the obtained time-series data (DATA-g) and the time-series data (DATA-b) for which a bad evaluation is obtained from the time-series data, and the time-series data extracting means. The extracted time series data with good evaluation (DATA-
knowledge data generating means for generating knowledge data for obtaining an operation amount for correcting the control amount from a bad state to a good state based on g) and time-series data with bad evaluation (DATA-b); The time series data re-acquisition means for re-acquiring the time series data regarding the controlled variable and the manipulated variable by controlling the plant using the acquired knowledge data, and evaluating the control state indicated by the reacquired time series data. , Time-series data whose evaluation value is good in the evaluation (DATA-g)
In a better case, the re-acquired time series data is used in place of the time series data (DATA-g) which has been evaluated as good until then, and means for regenerating knowledge data. A control knowledge generation device characterized by:

According to a seventh aspect of the present application, in the invention according to the fifth or sixth aspect, the knowledge data is the time when the control amount in the time-series data (DATA-g) having a good evaluation. Series data and time series data with poor evaluation (D
Data of the difference between the control amount in ATA-b) and the time-series data, the operation amount in the time-series data with good evaluation (DATA-g) and the operation amount in the time-series data with bad evaluation (DATA-b). It is a fuzzy rule created based on the data of the difference between and.

The invention according to claim 8 of this application is the invention according to any one of claims 1 to 3, wherein the plant is a system including a jacket tank and a temperature control device. The amount is an internal temperature and an external temperature of the jacket tank, and the operation amount is an external temperature set value.

[0017]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the method and apparatus of the present invention will be described in detail below with reference to the accompanying drawings. First,
First, the outline of the method of the present invention will be explained together with the algorithm adopted therein. In the method of the present invention, in a control system for controlling a plant by adding a manipulated variable according to the state of the plant to the plant, as the first step, the manipulated variable to the plant and the plant when trial controlling the plant The time series data relating to the control amount is acquired (see step 101 in FIG. 1). For example, in the case of the process control for the jacket tank described above, this trial control is performed until the acquired data reaches the specified number of data while manually correcting the operation amount in response to the influence of various disturbances. (FIG. 1, step 102)
reference).

Next, the quality of the control state is evaluated based on the time-series data acquired in this manner, and thereafter, when the time-series data (DATA-g) for which good evaluation is obtained and the bad evaluation are obtained. One series data (DATA-b) each is extracted from the time series data (see step 103 in FIG. 1). FIG. 2 shows the time series data Yg of the control amount and the time series data Ug of the operation amount that constitute the time series data (DATA-g) for which a good evaluation is obtained among the plurality of sets of time series data acquired. And the time series data Yb of the controlled variable and the time series data Ub of the manipulated variable that constitute the time series data (DATA-b) for which bad evaluation is obtained. Since the specific evaluation method varies depending on the content of the plant to be controlled, it will be described later in detail in relation to the control of the jacket tank.

Next, the extracted time series data with good evaluation (DATA-g) and time series data with bad evaluation (D
Based on ATA-b), knowledge data for obtaining an operation amount for correcting the control amount from a bad state to a good state is created (see step 104 in FIG. 1). Here, as the above-mentioned knowledge data, it is preferable that the time-series data (Y) of the controlled variable in the time-series data (DATA-g) with good evaluation is used.
g) and the data (Yg-Y) of the difference between the time series data (Yb) of the controlled variable in the time series data (DATA-b) with poor evaluation.
b), the operation amount (Ug) in the time series data with good evaluation (DATA-g) and the time series data with bad evaluation (DATA-).
Data (Ug-Ub) of the difference from the operation amount (Ub) in b)
It is considered to be a fuzzy rule created based on. FIG. 3 shows the control amount difference data (Yg-Yb) and the operation amount difference data (Ug-Ub) described above, respectively.

Here, obtaining the knowledge data for obtaining the operation amount for correcting the above-mentioned control amount from the bad state to the good state, in other words, obtaining the good control result (Yg, U
g) and a bad control result (Yb, Ug) and a difference component between the bad control result and a good control result (hereinafter referred to as a correction function F ()). () Is expressed as Equation (1). Further, the operation amount (Ub) when this correction function F () and a bad control result are obtained.
The fact that a good control result (Yg) (where Yg is the target trajectory) is obtained by using and is expressed by the mathematical expression (2). F (Yg-Yb) = Ug-Ub ... Numerical formula (1) Yg = PLANT (Ug) = PLANT {Ub + F (Yg-Yb)} ... Numerical formula (2) In addition, each of the said Numerical formulas (1) and (2). The meanings of the symbols are as follows. Yg: Control amount time series data with good control result Yb: Control amount time series data with bad control result Ug: Operation amount time series data with good control result Ub: Operation amount time series data with poor control result PLANT ( U): Time series data of the response when the time series (U) of the manipulated variable is given to the controlled object Yg = PLANT (Ug) which is the first expression of Expression (2)
Means that if the controlled object does not change at all, the highest level control result can always be obtained by giving Ug. Further, the correction function F () is considered to have the ability to correct a bad control result into a good control result. further,
When the controlled object changes and the response to Ug deviates, it is expected to correct the manipulated variable Ug so that the correction function F () approaches a better control result.

Thereafter, by controlling the plant using the obtained knowledge data, the next production will be performed (see step 105 in FIG. 1). Here, the control amount (Yg) when a good control result is obtained is corrected to a better waveform, and the corrected function (F ()) having the corrected waveform (Yg ') as the target trajectory is good. Obtaining a better control result by using the manipulated variable (Ug) when a different control result is obtained is expressed by Equation (3). Yn = PLANT (Ug + F (Yg′−Yn)) ≦ Yg (Equation 3) In addition, the meaning of each symbol in the above Equation (3) is as follows. ≦: The left side has the same or better control result than the right side. Yn: The control amount obtained this time Yg ′: The Yg corrected to a better controllability In the above formula (3), the good control result ( Yg)
When the control result is changed to a better control result (Yg ') and the degree of change is small, Yg and Yg' are calculated by the correction function F ().
It is expected that the intermediate performance of will be obtained. Next, an example in which the control knowledge generation method and device according to the present invention are applied to a temperature control system for a jacket tank will be described.

In the hardware configuration diagram shown in FIG. 4, the jacket tank (A) is a double-structured container for generating a chemical reaction, and the heat medium is circulated in the outer shell to adjust the temperature of the inner shell. It is done like this. In the following description, the temperature inside the inner shell is [inner temperature] and the temperature inside the outer shell is [outside temperature].
Shall be called. The motor (B) serves as a drive source for stirring the reactant in the inner shell with a stirring fan.

The existing system (C) is a general term for existing temperature control systems attached to the jacket tank, and these are generally DCS (Distributed Control System).
Is built by. In the existing system (C), a controller 1 (D) for controlling the internal temperature and a controller 2 (E) for controlling the external temperature
And a process control section (F) are included, and these functions are realized by software on the DCS.

Here, the controller 1 (D) for controlling the internal temperature is
It has a function of calculating the set value of the outside temperature from the measured value of the inside temperature and the set value of the inside temperature, and outputting this to the corrector [G] for the set value of the outside temperature. In addition, controller 2 (E) for adjusting the outside temperature
Is for calculating the valve openings of steam and cooling water from the measured value of the external temperature and the set value of the external temperature. Further, the process control section (F) operates information relating to product types such as product type, lot number, and production amount.

The corrector (G) for the set value of the external temperature is a part in which the control knowledge control method and device according to the present invention are realized, and is composed of a personal computer, a PLC or the like in terms of hardware. FIG. 5 shows the functional configuration of the corrector (G) for the external temperature set value.
Shown in As shown in the figure, the data acquisition function (G1), the control knowledge creation function (G2), the automatic control function (G3), the control data ( It includes a database (DB) that stores the operation amount and the control amount time series data obtained by executing the control and the control knowledge for each product type. Data acquisition function (G1)
Is for obtaining control data from an existing temperature control system. The control knowledge creation function (G2) includes four functions including a target performance setting function, a performance evaluation function, a knowledge creation function, and a knowledge correction function.

The target performance setting function is for the user to set the target value and the gain for each control index, and may be registered as fixed values in advance, and is shown in FIG. You may make it input from a setting screen. The performance evaluation function is for calculating the control index of the acquired control data. The knowledge creation function is for creating control knowledge using the acquired control data. The knowledge correction function is to update the control knowledge using the control data obtained by controlling the plant using the created control knowledge.

The database (DB) is for holding the acquired control data and the created control knowledge. The automatic control function (G3) calculates the set value of the external temperature with reference to the created control knowledge, and its configuration is specifically shown in FIG.

Next, the respective functions (G1) to (G3) constituting the above described outside temperature set value corrector (G) will be described in more detail. An example of the data acquired by the data acquisition function (G1) is shown in FIG. 7, and the range of those data is shown in FIG.
Each is shown in the table. As is clear from these figures, in the data acquisition function (G1), the set value of the internal temperature, the set value of the external temperature obtained from the existing control system, the current value of the internal temperature, the current value of the external temperature, the type of product type , Lot identification, manufacturing quantity, manufacturing time (unit: time), contact signal indicating manufacturing start / end state, and external temperature set value calculated by the external temperature set value compensator. We handle it,
These data are input / output at a specified timing, and each has a specified range.

In the performance evaluation function (G2), the following 7
Performance is evaluated using one of the indicators. That is, in the performance evaluation function (G2), an evaluation value (performance evaluation value) is obtained by linearly combining these seven indexes, and the smaller the evaluation value, the better the control result. Overshoot amount (OVER) This is the difference between the maximum control amount and the control target value in the section from when the control amount first reaches the control target value to when the slope first becomes negative. The initial value is -1.0. Undershoot amount (UNDER) This is the difference between the minimum value of the control amount and the control target value in the section from the end of the calculation section of the overshoot amount to the first positive slope. The initial value is the control target value × -1 Maximum value during settling (MAX) The difference between the maximum control value and the control target value in the section from the end of the undershoot amount calculation section to the change of the control target value. is there. Minimum value at settling (MIN) This is the difference between the minimum value of the control amount and the control target value in the section from the time when the maximum value at the settling time is first calculated until the control amount target value changes. Hunting frequency (HUMP) This is the number of times the upper / lower limit values at the time of settling were exceeded in the interval from the end of the undershoot amount calculation interval to the change of the control amount target value. Error area ratio (ERROR) After the controlled variable reaches the defined ratio (START) of the control target value,
It is calculated from the following formula in the section until the control ends. Error area ratio = Σ (Control amount-Control target value) ² / Section time Rise time (STAND) From when the control amount reaches a predetermined value (predefined: constant) until it reaches the control target value first It's time. The performance evaluation value is calculated by the following mathematical expression (4). Performance evaluation value = Σ (target value-index value) ² × gain Equation (4) The target value and gain for each index are set in advance. The index that could not be calculated is set to [0], and the formula (4) is calculated.

The control knowledge creating function (G2) creates control knowledge in the following procedure. Select the one with the smallest performance evaluation value (DATA-g). The performance evaluation value at that time is α. Performance evaluation value of α + δ or more and minimum (DA
Select TA-b). It should be noted that δ is a constant and is set in advance. The reason that the maximum performance evaluation value is not selected as DATA-b is that data will not be registered when an error occurs in the system including the plant.
If the system is in the normal range, and the control state is worse than the best control state, but not too bad
Register it as DATA-b. By doing so, the correction of the operation amount by the control knowledge becomes appropriate. FF knowledge = U _DATA-g . When calculating the FB knowledge, the correction function F () is calculated in the following procedure. i) Calculate the difference component β. β = Yg−Yb OUT = Ug−Ub ii) β derivative (dβ), integral (iβ), second derivative (dd)
β) Calculate the component. In addition, as input signals β, dβ, d
Assume that the teacher data consists of dβ, iβ and OUT as an output signal. That is, the input to the correction function F () is β, d
β, ddβ, iβ, and the output is Ug-Ub. iii) Describe the correction function F () by fuzzy rule inference according to the following procedure. Note that the MF number (odd number) for the input / output signal is set in advance. 1) Obtain each maximum / minimum value from input / output data. 2) MF is defined by equally dividing the maximum value and the minimum value by the defined number of MFs. 3) Calculate the degree of coincidence between the input / output data and the corresponding MF. 4) A set of MFs having the highest degree of coincidence is created as a rule. The product of the degree of coincidence is registered as the weight of the rule. 5) The MFs having the same degree of coincidence have a smaller MF when the data is close to the maximum value and a larger MF when the data are close to the minimum value.
Is adopted. 6) Execute 3, 4, 5 for all input / output data. 7) When rules with the same antecedent / consequent part occur, the weight is updated to the one with the higher degree of coincidence. 8) When the same rule occurs in the antecedent part, the one with the smaller weight is deleted. 9) After the creation is completed, change the width of the MF of the consequent part to 0.
(Singleton).

As a specific example, a case of two inputs and one output in which the deviation (β) and the deviation differential (dβ) are input and the operation amount difference (Ug-Ub) is output is shown below. Input: x1 Number of defined MF: 3 Maximum value: 10 Minimum value: -10 x2 Number of defined MF: 3 Maximum value: 10 Minimum value: 0 Output: y Number of defined MF: 3 Maximum value: 5 Minimum value: -5

The MF defined in this case is shown in FIG. 9 and FIG.
0 is shown. Input / output data A: x1 ＝ 7.50, x2 ＝ 5.00, y ＝ 5.00 → A11 (x1) ＝ 0.00, A12 (x1) ＝ 0.25, A13 (x1) ＝ 0.75 A21 (x2) ＝ 0.00, A22 (x2) ＝ 1.00 , A23 (x2) = 0.00 Ay1 (y) = 0.00, Ay2 (y) = 0.00, Ay3 (y) = 1.00 → IF x1 = A13 & x2 = A22 THEN y = Ay3 Weight = A13 (x1) ・ A22 (x2) ) ・ Ay3 (y) ＝ 0.75 B: x1 ＝ 7.50, x2 ＝ 2.50, y ＝ 1.25 → A11 (x1) ＝ 0.00, A12 (x1) ＝ 0.25, A13 (x1) ＝ 0.75 A21 (x2) ＝ 0.50, A22 (x2) ＝ 0.50, A23 (x2) ＝ 0.00 Ay1 (y) ＝ 0.00, Ay2 (y) ＝ 0.75, Ay3 (y) ＝ 0.25 → IF x1 ＝ A13 & x2 ＝ A22 THEN y ＝ Ay2 Weight ＝ A13 (x1 ) .A22 (x2) .Ay2 (y) = 0.28125 The antecedent part of the rule A created by the input / output data A and the rule B created by B are the same. In this case, rule A has a larger weight, so rule B is deleted.

As another specific example, the deviation (β), the deviation derivative (dβ), the deviation second-order differentiation (ddβ), and the deviation integral (iβ) are input, and the manipulated variable difference (Ug-Ub) is output. The data structure of the fuzzy rule in the case of 4-input 1-output is shown in FIG. 11, and examples of the rules are shown in FIGS.

In the data structure shown in FIG. 11, the number "10" following "; rule definition" is the number of rules, and the following symbol string "N: N: *: P; OUT1".
Means IF IN0 = N and IN1 = N and 1 IN2 = * and IN3 = P then OUT = OUT1. However, "*" is Don't Care
Means

Next, the knowledge correction function will be described. The knowledge is modified by two types of modification methods (step 150 below) according to the control performance of the control result obtained by using the knowledge.
5 or step 1506) is used properly (FIG. 15).
reference).

That is, when the process is started in FIG. 15, the plant is controlled by using the generated knowledge data, and accordingly, the time series data regarding the control amount and the operation amount are reacquired (step 1501). ). Then, the control state indicated by the reacquired time series data is evaluated (step 1502). Here, if the evaluation result of this time is better than the specifications required for the plant (YES in step 1503), the next production is directly performed (step 150).
7). On the other hand, if the evaluation result this time is worse than the required specifications (step 1503 NO) and the obtained control performance is worse than the control performance of the best data in the past (step 1504 NO), the obtained control data is FB knowledge of control knowledge is created using the best control data in the past (step 1506). On the other hand, if the obtained control performance is equivalent to the control performance of the best data used at the time of knowledge creation (YES in step 1504), the control amount data of the obtained control data is changed to be better in terms of accuracy. (Step 1505). After that, the changed data is used as the reference trajectory for the next control. The above changing method is performed by detecting the portion where the control amount reaches the target value for the first time and correcting the hunting component in the settling range according to the equation (5). Y1 = control amount + deviation × θ (Equation 5)

The automatic control function (G3) has a structure shown in FIG.
F) and a feedback unit (FB) are included. The processing flow of this automatic control function is shown in FIG.

In the figure, when the process is started, a type name setting process (step 1701) and a control knowledge reading process (step 1702) are performed, and then a sensor signal is being read (step 1703). It waits for the activation signal to be turned on (step 1704). In this state, when the activation signal is turned on (step 17
04 YES), FF operation amount reading processing (step 1
705), target (Y REF) read processing (step 1
706) and the calculation processing of the FB operation amount (step 170).
After sequentially performing 7), a manipulated variable (= FF manipulated variable + FB manipulated variable) is calculated based on them (step 1707) and applied to the plant (step 1709). After that, the abnormal process (steps 1705 to 1709) is repeated.

[0039]

According to the present invention, the control required for reducing or eliminating the burden of manual correction when the automatic control system cannot exhibit the desired performance unless manual correction is added due to disturbance or the like. Knowledge can be automatically generated using a learning function.

[Brief description of the drawings]

FIG. 1 is a flowchart showing an outline of control knowledge creation processing.

FIG. 2 is a graph showing a relationship between a control result and an operation amount.

FIG. 3 is a graph showing a difference sentence between a case where the control state is good and a case where the control state is bad for each of the control amount and the operation amount.

FIG. 4 is a block diagram showing a hardware configuration of a jacket tank temperature control system.

FIG. 5 is a block diagram showing a functional configuration of a corrector for an external temperature set value.

FIG. 6 is a diagram showing an example of a setting screen for setting a target performance.

FIG. 7 is a table showing a list of input / output signals in the data acquisition function.

FIG. 8 is a table showing a list of signal domains in the data acquisition function.

FIG. 9 is a graph showing an initial shape of a defined membership function.

FIG. 10 is a graph showing the final form of the defined membership function.

FIG. 11 is a diagram showing a data structure of an FB knowledge correction function.

FIG. 12 is a diagram showing an example of a fuzzy rule that constitutes control knowledge.

13 is a graph illustrating the meaning of the rules shown in FIG.

FIG. 14 is a graph explaining the meaning of the rules shown in FIG.

FIG. 15 is a flowchart showing processing at the time of operation using control knowledge.

FIG. 16 is a block diagram showing a structure of an automatic control unit.

FIG. 17 is a flowchart showing the contents of an automatic control function.

[Explanation of symbols]

 A Jacket tank B Motor C Existing system D Controller for internal temperature control 1 E Controller for external temperature control 2 F Process control unit G Compensator for external temperature setting device G1 Data acquisition function G2 Control knowledge creation function G3 automatic control Function DB database

Claims (8)

[Claims] 1. In a control system for controlling a plant by adding a manipulated variable according to the state of the plant to the plant, time-series data relating to the manipulated variable to the plant and the controlled variable of the plant when trial controlling the plant is acquired. Time series data acquisition step, a control state evaluation step of evaluating the quality of the control state based on the acquired time series data, and time series data (DATA-g) for which good evaluation was obtained in the control state evaluation step. ) And a time series data (DATA-b) for which a bad evaluation is obtained from the time series data, and a time series data of good evaluation extracted by the time series data extraction step (DATA DATA-g) and time-series data with bad evaluation (DATA-b), the control amount is changed from bad to good. Control knowledge generation method characterized by comprising: a knowledge data generation step of generating a knowledge data for determining the manipulated variable to correct the. 2. A control system for controlling a plant by adding a manipulated variable according to the state of the plant to the plant, and acquiring time series data relating to the manipulated variable to the plant and the controlled variable of the plant when trial controlling the plant. Time series data acquisition step, a control state evaluation step of evaluating the quality of the control state based on the acquired time series data, and time series data (DATA-g) for which good evaluation was obtained in the control state evaluation step. ) And a time series data (DATA-b) for which a bad evaluation is obtained from the time series data, and a time series data of good evaluation extracted by the time series data extraction step (DATA DATA-g) and time-series data with bad evaluation (DATA-b), the control amount is changed from bad to good. A knowledge data generating step of generating knowledge data for obtaining a manipulated variable for correction to, and controlling the plant using the generated knowledge data to reacquire time series data relating to the controlled variable and the manipulated variable. A time-series data re-acquisition step, and a control state indicated by the re-acquired time-series data is evaluated, and the evaluation value is a time-series data (DA) with good evaluation.
If it is better than TA-g), the step of regenerating knowledge data by using the re-acquired time-series data instead of the time-series data (DATA-g) which has been evaluated as good until then. And a control knowledge generation method. 3. The knowledge data includes time series data of a controlled variable in time series data of good evaluation (DATA-g) and time series data of a controlled variable in time series data of bad evaluation (DATA-b). Of the difference between the operation amount in the time series data with good evaluation (DATA-g) and the operation amount in the time series data with bad evaluation (DATA-b), and a fuzzy The control knowledge generation method according to claim 1 or 2, wherein the control knowledge is a rule. 4. The plant is a system including a jacket tank and a temperature control device, the controlled variable is an internal temperature and an external temperature of the jacket tank, and the manipulated variable is an external temperature set value. The control knowledge generation method according to any one of claims 1 to 3, which is characterized. 5. In a control system for controlling a plant by adding a manipulated variable according to the state of the plant to the plant, time series data regarding the manipulated variable to the plant and the controlled variable of the plant when trial controlling the plant is acquired. Time-series data acquisition unit, a control state evaluation unit that evaluates the quality of the control state based on the acquired time-series data, and time-series data (DATA-g) for which a good evaluation is obtained in the control state evaluation unit. ) And time-series data (DATA-b) for which a bad evaluation is obtained, from the time-series data, and time-series data with good evaluation extracted by the time-series data extraction means (DATA-b). DATA-g) and time-series data with poor evaluation (DATA-b) to adjust the control amount from a bad state to a good state. Control knowledge generating apparatus characterized by comprising a knowledge data generating means for generating a knowledge data for determining the amount of a. 6. A control system for controlling a plant by adding a manipulated variable according to the state of the plant to the plant, and acquiring time series data relating to the manipulated variable to the plant and the controlled variable of the plant when the plant is trial-controlled. Time-series data acquisition unit, a control state evaluation unit that evaluates the quality of the control state based on the acquired time-series data, and time-series data (DATA-g) for which a good evaluation is obtained in the control state evaluation unit. ) And time-series data (DATA-b) for which a bad evaluation is obtained, from the time-series data, and time-series data with good evaluation extracted by the time-series data extraction means (DATA-b). DATA-g) and time-series data with poor evaluation (DATA-b) to adjust the control amount from a bad state to a good state. Knowledge data generating means for generating knowledge data for obtaining a quantity, and time series data re-acquiring means for controlling the plant using the generated knowledge data and re-acquiring time series data concerning a control quantity and an operation quantity And the control state indicated by the reacquired time series data is evaluated, and the evaluation value is the time series data (DA) with good evaluation.
Means for regenerating knowledge data by using the reacquired time-series data instead of the time-series data (DATA-g), which has been evaluated as good until then, when it is better than TA-g). And a control knowledge generation device. 7. The knowledge data includes time series data of a controlled variable in time series data of good evaluation (DATA-g) and time series data of a controlled variable in time series data of bad evaluation (DATA-b). Of the difference between the operation amount in the time series data with good evaluation (DATA-g) and the operation amount in the time series data with bad evaluation (DATA-b), and a fuzzy 7. The control knowledge generation device according to claim 5, wherein the control knowledge generation device is a rule. 8. The plant is a system including a jacket tank and a temperature control device, the controlled variable is an internal temperature and an external temperature of the jacket tank, and the manipulated variable is an external temperature set value. The control knowledge generation device according to any one of claims 5 to 7.