JPH0550002B2 - - Google Patents

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a recursive fuzzy inference device that monitors various industrial processes and infers parameter values appropriate for the industrial process. .

[Conventional technology]

Figure 5 shows the operating principle of a conventional fuzzy inference device, as shown, for example, in ``Fuzzy System Theory and Fuzzy Control'' (Kiyoharu Asai) in the November 1986 issue of ``Labor Saving and Automation'', pages 61-66. It is an explanatory diagram,
In the figure, 1 and 2 are inference rules, 3 and 4
are the feature quantities input to the fuzzy inference device, and here they are the control deviation e of the control system and the rate of change Δe of the control deviation, respectively. Further, 5 and 6 are membership functions of the antecedent part of the rule 1, 7 is a membership function of the consequent part of the rule 1, 8 and 9 are membership functions of the antecedent part of the rule 2, and 10 is a membership function of the antecedent part of the rule 2. is the membership function of the consequent of . Further, 11 is a membership function obtained by combining the membership functions 7 and 10, and 12 is an inference value obtained by taking the center of gravity from this membership function 11. In this example, the fuzzy inference device outputs it as the rate of change Δu of the manipulated variable. be done.

Further, FIG. 6 is a block diagram showing an example of a conventional fuzzy inference device based on such an operating principle.
weighting means for evaluating the degree of validity of the antecedent part from the input feature values 3 and 4 and weighting the membership function of the consequent part based on the degree of validity;
Reference numeral 4 denotes a synthesis means for synthesizing the membership functions weighted by this weighting means 13, and 15 determines an inference value 12 from the membership functions synthesized by this synthesis means 14, and calculates the rate of change Δu of the manipulated variable. This is a means for determining the inferred value to be output.

Next, the operation will be explained. Here, rule 1
``Feature amount 3 (control deviation e) is slightly negative,
And if feature quantity 4 (change rate Δe of control deviation) is slightly deviated from the positive, inference value 12 (operated quantity Δu)
It means "to shift just a little bit," and of this, the "if..." part is called the antecedent part,
The part after that is called the consequent part mentioned above.
Therefore, the membership function 5 in the antecedent part of Rule 1 defines "a set of control deviations with a slight negative deviation", and the membership function 6 defines "a set of change rates of control deviations with a slight positive deviation". are doing.

Now, if the actual value of the control deviation as the feature quantity 3 input to the weighting means 13 is e _{0 ,} as shown in FIG. A certain degree is evaluated as "0.8" by the membership function 5, and a degree to which the value Δe ₀ is "the rate of change of the control deviation that is slightly deviated from the positive" is evaluated as "0.7" by the membership function 6. be done.
Of these two evaluation values, the lower value "0.7" is adopted and becomes the degree of fulfillment of the antecedent part of the rule 1. Furthermore, the membership function 7 of the consequent part of this rule 1 means "to shift the amount of operation by a small amount", and the membership function 7 is 0.7 times as large as the value of the probability of holding the antecedent part. weighted.

This is exactly the same for rule 2, and the actual value e ₀ of the control deviation of the input feature quantity 3
and the actual value of the rate of change of control deviation of feature quantity 4 Δe ₀
The degree of validity of the antecedent part is evaluated based on , and the membership function 10 of the consequent part is weighted 0.5 times based on the value of the degree of validity "0.5". The membership functions 7 and 10 weighted in this manner are input to the synthesizing means 14 and synthesized to obtain a composite membership function 11. Furthermore, this composite membership function 11 is input to the inference value determining means 25 to calculate the center of gravity, and as a result, the rate of change Δu ₀ of the manipulated variable is outputted as the inference value 12 from the fuzzy inference device.

As described above, in the 3-fuzzy inference device, multiple rules work simultaneously, the consequent part is weighted according to the probability of each antecedent part being true, and a balanced value as a whole is output as the inference value. .

FIG. 7 is a block diagram showing the configuration of a control system to which the above-described conventional fuzzy inference device is applied as a control device. In the figure, 40 is the target value (r) given from outside the control system, 41 is the controlled variable (y) output from the process 36, 3 is the control deviation (e) between the target value 40 and the controlled variable 41, and 100 1 is a differential element which inputs the control deviation (e) 3 and outputs the rate of change (Δe) 4 of the control deviation; 101 is a fuzzy inference device; the whole of FIG. 6 corresponds to this.

12 is an inference value output from the fuzzy inference device 101, which in this application example is the rate of change of the manipulated variable (Δu). Reference numeral 102 is an integral element that inputs the rate of change Δu of the manipulated variable and outputs the manipulated variable x43 to the process 36.

Figure 8 is shown, for example, on page 16 of "Feedback Control" by Masaru Nagata (published by Ohmsha, 1971).
This is a block diagram of a PID control device. In the figure, 103 is a format element that inputs the control deviation (e) 3 and outputs it as is, and 104 inputs the control deviation (e) 3 and obtains the integral value of 1/T _I Integral element that outputs times 10
5 inputs the control deviation (e) and calculates the rate of change of the control deviation.
Differential element that outputs T _D times, 106 is formal element 1
03, the sum of the outputs of the integral element 104 and the differential element 105 is input, and the signal obtained by multiplying this sum by K _C is the manipulated variable.
(y) This is a gain element output as 41.

FIG. 9 shows the PID control device 107 shown in FIG. 8 above.
FIG. 2 is a block diagram showing the configuration of a control system when the system is applied as a control device. In this case, the PID control device 107 inputs the control deviation (e) 3 and controls the manipulated variable (y) 41.
is output to process 36.

[Problem that the invention seeks to solve]

As described above, conventional control systems are configured using a fuzzy inference device or a PID control device as a control device.

However, in either case, tuning of the control device is required due to the nature of the process 36 to be controlled. Therefore, when a fuzzy inference device is used as a control device, it is necessary to tune each membership function 5 to 10 in FIG. 5, and when a PID control device is used as a control device, each coefficient in FIG. Tuning of K _C , T _I , and T _D is required.

Therefore, if you are trying to configure an auto-tuning control system that has a mechanism to automatically perform the above-mentioned tuning, it is recommended to use the
Applying a PID control device, its respective coefficients K _C , T _I , T _D
For tuning, it is conceivable to use a fuzzy inference device suitable for automating operator know-how.

However, the quantities that are input to the fuzzy inference device during tuning are the instantaneous values of the signal, and are not the control bias (e) and its rate of change (Δe), which can always be input.
Features (Si) that indicate the state of the control system are extracted intermittently by observing signal trends for a certain period of time, such as the divergence tendency of control deviations, the magnitude of control deviations, and the degree of follow-up of control variables with respect to changes in target values. It is.

Figure 10a shows the control deviation e or its rate of change Δe
Fig. 10b shows the time course of the value of the feature quantity (Si) indicating the state of the control system.

The above feature quantity (Si) is usually Si = 0, and is 0 only when the state of the control system is extracted from the signal trend.
<Si≦1. This value of 0 is not a value that truly means 0, but is a value for convenience until the feature quantity (Si) is extracted next time.

Therefore, in the conventional fuzzy inference device, if the feature quantity (Si) is inputted intermittently, the antecedent parts 5, 6, 8, and 9 cannot be evaluated, and therefore inference is impossible. Furthermore, even if inference is possible when all of the features (Si) are input at the same time, the inference only occurs at that moment, and even if only a part of the features (Si) is input, I'll waste it.

Furthermore, since the values of the coefficients Kc, T _I , and T _D of the PID control device are constant and change only slowly, the inferred values should naturally converge.
However, conventional fuzzy inference devices react only to instantaneous values and do not accumulate past inferences, so this convergence cannot be achieved.

That is, it has been impossible to configure an autotuning control system by simply combining the conventional fuzzy inference device, which is the first prior art, and the PID control device, which is the second prior art.

This invention was made to solve the above-mentioned problems, and can be applied not only to autotuning control systems, but also to other applications, such as when input values have meaning only intermittently. Obtain a fuzzy inference device that can be used for inference even when all input values are not aligned, and that can perform inference with convergence when the value to be inferred is constant or only changes slowly. The purpose is to

[Means for solving problems]

The fuzzy inference device according to the present invention is configured to synthesize not only the membership function of the consequent of each rule but also the previous composite membership function when synthesizing the composite membership function.

[Effect]

The fuzzy inference device of this invention synthesizes a membership function that includes not only the current but also the past functions of each rule, making it possible to perform inferences similar to those in the case of many rules working at the same time. The inference value can take continuous values, and a learning effect occurs in the composite membership function.
It also enables convergent inference.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of a recursive fuzzy inference device according to the present invention, FIG. 2 is a block diagram showing an example of its application to control gain tuning of a process control controller, and FIG. It is an explanatory diagram showing the principle of operation. In each of these figures, 20 and 21 are inference rules, 22 to 25
are the feature quantities input to the recursive fuzzy inference device, 26 and 27 are membership functions of the antecedent part of the rule 20, 28 are membership functions of the consequent part, and 29 and 30 are the antecedent part of the rule 21. , and 31 is the membership function of its consequent. Furthermore, 32 is a composite membership function representing the dissatisfaction synthesized last time, 33 is a composite membership function that is weighted according to the change in process characteristics to the previous composite membership function 32 representing this dissatisfaction.
4 is a composite membership function obtained by combining the membership function 28 of the consequent part of rule 20, the membership function 31 of the consequent part of rule 21, and the weighted previous membership function 33. Further, 35 is an inference value obtained from this composite membership function 34, which in this example is output from the recursive fuzzy inference device as a control gain Kc.

Further, 36 is a process to be controlled, 37
is a control device, such as a PID control device, which controls this process 36, and 38 is a control device 37, such as a PID control device.
39 is a feature extractor that supplies features 22 to 25 to the fuzzy inference device 38, and 40 is a control The target value (r) given from outside the system, 41 is the process 36
The controlled variable (y) output from the control device 37, 42 is the change (e) between the target value 40 and the controlled variable 41, which is input to the control device 37, and 43 is the manipulated variable (x) given from the control device 37 to the process 36. ), 44 is the process characteristic change amount sent from the process 36 to the fuzzy inference device 38.

Furthermore, 45 is a weighting means for evaluating the degree of establishment of the antecedent part from the feature values 22 to 25 inputted for each of the rules 20 and 21, and weighting the membership function of the consequent part based on the degree of establishment of the antecedent part. , 4
6 synthesizes each membership function 28, 31 weighted by this weighting means 45 and the previous combined membership function 33 to obtain each inference value 3.
A composition means 47 obtains a new composition membership function 34 representing the degree of dissatisfaction of 5 (for example, gain K);
5, a delay means 48 delays the new composite membership function 34 synthesized by the composition means 46, and 49 has strong nonlinearity, so the control characteristics vary depending on the operating point. For example, in flow rate control where only the opening and closing of a valve can be controlled, a signal that cannot be directly controlled, such as fluid temperature, which affects the viscosity of the fluid and changes the control characteristics, is used to control the process characteristics. A characteristic change evaluation means receives input from the process 36 as the quantity 44 and outputs a weighting coefficient for forgetting the past according to the degree of change in the control characteristic. The previous composite membership function 32 delayed by the delay means 48 is multiplied by the weighting coefficient value output from this characteristic change evaluation means 49 to perform weighting, and then sent to the composition means 46 as a composite membership function 33. This is a multiplication means for feedback. In FIG. 1, the double line represents the fuzzy amount, and the single line represents the non-fuzzy amount.

Next, the operation will be explained. The purpose of inference in this recursive fuzzy inference device is to process 3
The control gain Kc is tuned by monitoring the feature quantities 22 to 25 of No. 6. Therefore,
First, the feature amounts 22 to 25 will be specifically shown. That is, the feature quantity 22 is the divergence tendency Sa of the deviation (e) 42, the feature quantity 23 is the magnitude Sb of the deviation 42, and the feature quantity 24 is the degree of follow-up of the control quantity (y) with respect to the change in the target value (r) 40.
Sc, the feature quantity 25 is the magnitude Sd (=Sb) of the deviation 42. At this time, rule 20 is ``If deviation (e) 42
This means that if the tendency for divergence is large and its absolute value is also large, it can be determined that the current control gain Kc is too large. Here, if the actual values of the feature quantities 22 to 25 input to the weighting means 45 are Sa ₀ , Sb ₀ , Sc ₀ , and Sd ₀ , it is determined in the rule 20 whether this value Sa ₀ is "large" or not. , whether the value Sb ₀ is "large" is determined by the membership function 26 of the antecedent part of the rule 20.
and 27. In the example shown in Figure 3, the respective evaluation values are "0.4" and "1.0", and the lower value "0.4" of both evaluations is adopted as the probability of the antecedent part of the rule 20 being satisfied. be done.

Furthermore, the consequent part of this rule 20 defines a fuzzy set of "too large control gain Kc", and the current control gain Kc≧Kc ₀ means that " Too big"
It can be said. Therefore, the probability of establishment of this antecedent part is “0.4”
The membership function 2 is weighted according to
Create 8. This is exactly the same for rule 21, and the consequent is determined based on the degree of validity evaluated by evaluating the actual values Sc ₀ and Sd ₀ of the input features 24 and 25 using the membership functions 29 and 30 of the antecedent. A membership function 31 is created. In this example, the membership function 31 is a function in which all values are "0".

The composition means 46 synthesizes the membership functions 28 and 31 weighted in this manner and the composition membership function 33 obtained by weighting the previous composition membership function 32 to which a delay of one iteration was given by the delay means 48 and the evaluation weighting means 51. and infer a new composite membership function 34,
The result is:

This inference result is as the lowest gain of dissatisfaction.
The lowest gain of the new composite membership function 34 is set, and this is the one used as the control gain.

Here, FIG. 4 is a flowchart showing the flow of the operation. The composite membership function 34, which was synthesized by the synthesis means 46 in the previous iteration and sent out as an input for the next iteration (step ST8), is sent to the delay means 48 and is
A delay for the iteration is given, resulting in the membership function 32 (step ST1). In addition, apart from this, the feature quantities 22 to 25 are added to the weighting means 45.
, the process characteristic change amount 44 from the process 36
are respectively input to the characteristic change evaluation means 49 (step ST2), and the weighting means 45 evaluates the input feature quantities 22 to 25 using membership functions 26, 27, 29, and 30 of the antecedent parts of rules 20 and 21, The membership functions 28 and 31 of the consequent part are created based on the obtained degree of establishment (step
ST3).

Further, the characteristic change evaluation means 49 evaluates the degree of characteristic change of the process 36 from the input process characteristic change amount 44, sends the evaluation value to the multiplication means 50, and calculates the previous composite membership function delayed by the delay means 48. 32 is multiplied by this evaluation value to perform weighting, thereby obtaining a weighted composite membership function 33 (step ST4).

The membership functions 28 and 31 of the consequent part of each of the rules 20 and 21 and this weighted composite membership function 33 are input to the composition means 46 and are combined to generate a new composite membership function 34 (step ST5).

Here, a union operation is used for this composition operation. Therefore, since this composite membership function 34 is the union of the fuzzy set of ``too large control gain Kc'' and the fuzzy set of ``too small control gain Kc'', it ends up being the fuzzy set of ``unsatisfactory control gain Kc''. can be understood.

Particularly in the case of autotuning, it is often difficult to know how to change the current value, and it is also often unclear whether there is a better value than the current value. In such a case, the only evaluations are dissatisfaction caused by the current value being too large and dissatisfaction caused by the current value being too small. Generally, in such a case, it is appropriate to express dissatisfaction with a consequent, and the fuzzy inference device according to the present invention can perform fuzzy inference using this consequent.

The generated composite membership function 34 is input to the inferred value determining means 47, and the inferred value determining means 47
The fuzzy set membership function 34 determines the control gain Kc ₀ as the inference value 35 by selecting the value for which the membership function 34 of the fuzzy set is the lowest, and the recursive fuzzy inference device 38 output to the control device 37 (step
ST6). Specifically, the control gain that gives the lowest value of the composite membership function 34 may be selected.

This is because in the fuzzy inference device according to the present invention in which dissatisfaction is expressed by the consequent part, the composite membership function always has a concave shape, and the lowest value thereof is the control gain with the least dissatisfaction.

Next, a decision is made to terminate the operation (step
ST7), step the process if you want to continue the operation.
Return to ST8 and use the composite membership function obtained in step ST5 as input for the next iteration.

In addition, although the number of inference rules is two in the above example, the number may be three or more, and the number of inputs and outputs, and the number of stages of conditions in the antecedent part of each rule may also be arbitrary. Can be set to

Further, in the above embodiment, a case has been described in which the present invention is applied to control gain tuning in a control device for process control, but the present invention may also be applied to estimation of other parameters, and the same effects as in the above embodiment can be obtained.

〔Effect of the invention〕

As described above, according to the present invention, the composition membership function obtained in the previous composition operation is fed back and used again for the composition of the current membership function, so that the feature quantity {Si}
As, normally Si = 0, but only when a specific phenomenon occurs, inference is possible even if 0<Si≦1 is used, and moreover, the inferred value can take continuous values. Alternatively, if the parameters to be estimated are constant or change only slowly, it is possible to obtain inferred values with convergence.
Furthermore, when the target value changes frequently, feature quantities can be obtained one after another, and new information can always be prioritized.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of a recursive fuzzy inference device according to an embodiment of the present invention, and FIG. 2 is a block diagram showing an example of its application to control gain tuning of a process control control device. FIG. 3 is an explanatory diagram showing the operating principle, FIG. 4 is a flowchart showing the flow of its operation, FIG. 5 is an explanatory diagram showing the operating principle of the conventional fuzzy inference device, and FIG. 6 is its block diagram. Fig. 7 is a block diagram showing the configuration of a control system when a conventional fuzzy inference device is applied as a control device, Fig. 8 is a block diagram showing the configuration of a PID control device, and Fig. 9 is a block diagram showing the configuration of a PID control device. FIG. 10 is a block diagram showing the configuration of the control system when applied as an apparatus, and is a diagram showing the time course of the control deviation or its rate of change and the time course of the feature quantity. 20 and 21 are rules, 22 to 25 are features, 3
5 is an inference value, 36 is a process, 38 is a fuzzy inference device, 45 is a weighting means, 46 is a synthesis means, 47
48 is a delay means, and 51 is an evaluation weighting means. In addition, in the figures, the same reference numerals indicate the same or equivalent parts.

Claims (1)

[Claims] 1 The antecedent part is formed using a membership function that takes a value between the first value and the second value, and has multiple rules each describing the unsatisfactoriness of the control parameter value depending on the state of the feature quantity of the process. For each rule, a weighting means evaluates the degree of validity of the antecedent part from the feature values that occasionally appear in the process, and weights it according to the degree of validity to create a membership function of the consequent part; a delay means for outputting a composite membership function; an evaluation weighting means for weighting the previous composite membership function according to the degree of change in characteristics of the process; and a composite membership function output from the evaluation weighting means and each of the rules. A combining means combines the membership function of the consequent part to obtain a new combined membership function and supplies it to the delay means, and selects a value that minimizes the value of the new combined membership function and uses it as an inference value. A fuzzy inference device comprising an inference value determining means for determining and outputting a control parameter.