JPH0367303A - Fuzzy inference method and control method - Google Patents

Abstract

PURPOSE:To improve the accuracy of a fuzzy inference method by performing a fuzzy addition arithmetic operation between the value obtained by inverting the code of the membership value of the inferring result and the result of the fuzzy inference carried out with the knowledge belonging to the success used as a normal rule and inferring the result of the fuzzy addition arithmetic operation. CONSTITUTION:The fuzzy inference means 1 and 2 carry out the fuzzy inference based on the state value x deg.. The means 1 and 2 are described in the rules with which a plant is set in the desirable and undesirable states respectively. The inferring results B deg. and C deg. of the means 1 and 2 mean a fuzzy number respectively. The result B deg. is sent to a weighted average means 4, and the result C deg. is sent to the means 4 via an inverting means 3. The means 4 obtains the weighted average between the result B deg. obtained by a normal rule and the result obtained by inverting the result C deg. obtained by an adverse rule and defines this average as the inferring result y deg.. As a result, the operation is carried out so that the result is more different from the adverse rule as the degree of adaptation is increased. Then the experimental rule obtained through the failures can be utilized.

Description

[Detailed Description of the Invention] Industrial Field of Application First Invention 4i This relates to the improvement of fuzzy inference methods, especially when effective control rules are unknown and when bad control rules are known. The present invention provides a method.

The second invention is a control method for controlling a plant using knowledge, which provides a control method in which knowledge for control can be acquired by oneself through repeated trials, and the dynamic characteristics of the plant are Although it provides an effective control method for unknown cases, the conventional technique is fuzzy control using fuzzy reasoning.In general, the exact characteristics of the plant are complex or unknown, making automation difficult. It is effective for automation in fields where operation was mainly based on the operator's experience.
e n...'' Here, the part following if is called the antecedent, and the part following then is called the consequent. Fuzzy inference will be explained below.

FIGS. 7(a) to 7(g) are diagrams showing a fuzzy inference process. The control rules used here are as follows.

"If Xl is All and x2 is A12, set y to B1""If Xl is Aal and X2 is A22, set y to B2" Here, the first rule is rule I and the others are rule 2. And the state All of each state variable X1, X2
, A+a, Aal, and A22 are represented by the membership values shown in FIGS.

Also, L for operations B1 and B2 corresponding to the consequent part.
It is assumed that the membership values are represented by the outer trapezoids or triangles in FIGS.

The control method when xl and X2 are each obtained by x-1x 21 is as follows. First, calculations related to Rule 1 will be explained. In (a), find the membership value when Xl. Similarly, in (b), X2
Find the membership value when ``.The smaller of the two membership values obtained in this way is adopted.
In the example of FIG. 7, this is the case (b). The membership value obtained at this time becomes the fitness degree ω1 of the rule l for the inputs X1@, x, l. Therefore, the rule B1 is reduced in accordance with the degree of conformity ω1, as shown by the shaded area in FIG. 2(e). Similarly, calculations related to Rule 2 are performed.

In other words, in the same figure (C), the fitness of X11 is
B2
Reduce and transform the rule.

After the calculation of each rule is completed through the above steps, each rule is then integrated. Akira Sunawa As shown in the same figure (g)
The membership values of the two rules B1 and B2 that have been reduced and transformed are superimposed and their union is taken. Although this set is a fuzzy number that gives the manipulated variable y, in reality the manipulated variable y is not a fuzzy number, so we take the center of gravity of the union of the same figure (g) and define the manipulated variable yl. Assuming that the form is B-, the manipulated variable yl can be determined using the following equation.

y'= (J B"ydy) / (f B"dy
) It is possible to realize fuzzy control by operating the plant with the manipulated variable y@ determined in this way, but the problem to be solved by the invention. Therefore, it is necessary to extract only those control laws that have worked effectively from among the knowledge that the operator has through experience.

However, depending on the plant, it is known that certain processes cannot be controlled well or may even worsen.In other words, there are cases where the converse is not necessarily true, that is, there are cases where the control law cannot be written directly. (The conventional fuzzy inference methods cannot handle this, and as a result, it is difficult to control.)

Currently, no new method has been proposed for self-acquiring control knowledge using fuzzy control. Therefore, for example, when there are no experts, it becomes difficult to extract rules and control becomes difficult. There is also.

Means to solve the problemFailureTo solve the problem of not being able to use knowledge
Using empirical knowledge of the content that belongs to failure as a reverse rule,
X, the worst operation is inferred by fuzzy inference using the inverse rule group, and the fuzzy addition operation is performed on the same fuzzy inference result as before and the value with the sign of the membership value of the fuzzy inference result by the inverse rule group inverted. In order to acquire new knowledge, information on error relative to the target value and information on change in error information on plant operating amount information
The amount of change in the absolute amount of error due to the operation result is sequentially stored and it is determined whether the information is similar. If it is determined that the information is similar, the amount of change in the absolute amount of error information due to the plant operation result is stored. Similar information is represented by a value with a large absolute value and stored as a control rule group, and control is executed by determining the manipulated variable by weighted average calculation of the error information for the target value and the control rule group, but it also belongs to an example of failure. The result of fuzzy inference using an inverse rule group made up of a knowledge group can be obtained from a failure example by applying fuzzy inference to the fuzzy inference result using a normal rule group belonging to a success example with a negative fitness. Although information can be used with its weight, the amount of change in the absolute amount of error with respect to the target value gives an evaluation of the control a If the strategy is correct, it corresponds to the one that is correct, and if it is incorrect, it corresponds to the one that is more incorrect. ,
By representing those with a large absolute value of the amount of change in the absolute amount of error, a group of control rules from which features have been extracted can be constructed.

Embodiment FIG. 1 is a block diagram showing an inference configuration of a fuzzy inference method in an embodiment of the present invention.4 Fuzzy inference is performed by fuzzy inference means 1 and 2 based on the state quantity xl. Here, the fuzzy inference means 1 is an inference means described by rules (hereinafter referred to as positive rules) that make the plant desirable, as in the past, and the fuzzy inference means 2
Even if it is an inference method written by a rule that makes the plant undesirable (hereinafter referred to as a reverse rule), the inference method in the two inference methods 1 and 2 is the same as the conventional example, but it is based on the fuzzy inference method 1. Inference result B”,
The inference result C1ζ by the fuzzy inference means 2, even though both are fuzzy numbers, is sent to the weighted averaging means 4, and the inference result C1 is sent to the weighted averaging means 4 via the inversion means 3. The weighted averaging means 4 takes the weighted average of the inference result B'' based on the normal rule and the result obtained by inverting the inference result C- based on the reverse rule as the bundle number and its value as the inference result y-.

Figure 2 is a diagram illustrating the weighted average calculation method.
The inference result is expressed in the form of membership.The inference result B1 based on the correct rule is membership in the positive direction.The inference result C@ is expressed in the form of membership in the negative direction.Therefore, membership in the opposite direction has a negative weight. It can be thought of as having . Therefore, the calculation method C is y@=(1(B"-C")ydy)/(J B"dy
-I C"dy) (However, when L (B"-C") According to the method described above, the higher the degree of fitness of the inverse rule, the more the operation will move away from that rule.
Nanji & Azuma will be able to make use of empirical rules based on failures.A control method according to an embodiment of the present invention will now be described.A schematic configuration thereof is shown in FIG.

The output of the plant 15 is compared with the target value by the comparing means 16 to obtain an error information vector e. The error information vector e is sent to the result memory 11 for storing control results and the strategy determining means 14. In the result memory 11, the error information vector e, the change amount vector of the error information vector △e1
At that time, the plant operation amount vector U and the amount of change in the absolute amount of error ΔIt e II are stored. The amount of change in the absolute amount of error ΔII e IIζ As shown in FIG. 4, it is the difference between the absolute value of the error vector e n obtained this time and the absolute value of the previous error vector e n-1. However, the error vector e is generally an m-dimensional vector.Although it is shown in two dimensions in Figure 4 for convenience, it is possible to judge whether the control method is effective or not. The larger the negative value, the better the control method. Using the extraction means 12 from the result memory 11, features of the result information vector in the result memory 11 are extracted. The extraction method will be described later.The strategy determining means 14 generates a manipulated variable vector based on the data of the extraction result 13 and the error information vector e. is sent to the request △ plant 15 and also to the result memory 11. Even if the control system is configured as described above, the feature extraction method will be explained next.For feature extraction, compare the differences between the extraction results 13 and the results in the result memory 11, and examine the data with the smallest difference.If the absolute amount of error is If there is data with a large absolute amount of change ΔIt e If, the content in the extraction result 13 is rewritten to data with a large absolute amount of change Δll e If in the absolute amount of error.

In other words, even if the data structure of the extraction result 13 is the same as the data structure of the result memory 11, how to find the difference between the results Ct
Error information vector e, variation vector △e1 of error information vector, plant operation amount vector U at that time,
It can be determined as a norm between vectors R each having the amount of change in the absolute amount of error ΔItell as an element.

FIG. 5 shows the space of the vector R in two dimensions for convenience. Assume that the extraction results are indicated by RI, R2, . Find all the norms between Rr and the previous results Even if a very close vector R1 is found as a result, the amount of change in the absolute amount of error between R1 and Rt △If
Compare e II, adopt the result with the larger absolute value, and rewrite the extraction result 13. If you repeat this process, rules with large changes in the absolute amount of error △l e If will accumulate. will be done. That is, after the feature extraction is performed, the operation of the strategy determining means 14 that uses the rule group of the extraction result 13 and the error information vector e will be explained. What we have as a rule group is information about how the error information vector will change if we give a manipulated variable vector when the error information vector has a certain value. Therefore, the actual value of the error information vector is generally Even if the values are different, the norm between the actual error information vector e and the error information vector in each rule is calculated, and the output amount obtained from each rule is weighted according to the inverse ratio of each norm. distribute. Then, the sum of the distributed output values from each rule can be used as the manipulated variable output U.
Therefore, if the actual error information vector force statement is the error information vector of a certain rule, the manipulated variable output can be determined only from the output of the result of that rule.Also, among the rules of extraction result 3, there is a rule/k related to failure, i.e. In this case, deterioration can be prevented by reversing the sign of the weight and calculating the manipulated variable output. If you look at the procedure in another way, it can also be considered as fuzzy inference.In other words, in Figure 6, if the input information is the same in the extracted rules, the membership value is ``1'', and the consequent is not a fuzzy number. It can be considered as a type of fuzzy inference based on real numbers, and this example can also be considered as self-acquisition of knowledge in fuzzy control. II e 11
You can also use the ratio of error information (= e. / e-+). In this case, ζ. Although the absolute value of this is smaller than "1", the plate that has a smaller value Extract. In this way, the control results are accumulated as preferable rules, but in the same way, for those whose absolute value of the ratio is greater than "1" (above), those with a larger value are extracted.In this way, unfavorable rules are also extracted. However, as explained in detail, the first invention can improve the accuracy of fuzzy inference results by making it possible to use empirical rules based on failures, which were difficult to use in the past. It is something that can be done.

In addition, the second invention provides a function for self-acquisition of control knowledge, which has been a problem with conventional fuzzy control, and the effect is large.

[Brief explanation of drawings]

Figure 1 is a block diagram showing a fuzzy inference method in an embodiment of the present invention. Figure 2 is a block diagram showing the concept of weighted average with negative weights. Figure 3 is a block diagram showing a control method in an embodiment of the present invention. Figure 4 shows the concept of the amount of change in the absolute amount of error; Figure 5 shows the concept of similarity between rules; Figure 6 shows the process of determining the manipulated variable; Figure 7 shows the conventional This is a diagram showing the fuzzy inference method. 2... Fuzzy inference method 3... Reversal method
4... Weighted average hand ratio 11... Result memory,
I2... Extraction hand disturbance 13...Extraction power connection 14...strategy decision north 15...Plant.

Claims (5)

[Claims] (1) Empirical knowledge of content belonging to failure is used as an inverse rule, the suitability of the inverse rule is inferred by fuzzy inference using a group of inverse rules, and the sign of the membership value of the inference result is inverted; A fuzzy inference method characterized in that the result obtained by performing a fuzzy addition operation on the result of fuzzy inference using knowledge belonging to success as a positive rule group is used as the inference result. (2) The fuzzy addition operation is the union of the membership value sets of two rule groups, the membership value of the inference result of the inverse rule group is B^■ (y), the membership value of the inference result of the positive rule group Let A^■(y) be y^■=(∫(A
＾■−B＾■)ydy)／(∫A＾■dy−∫B^■d
2. The fuzzy inference method according to claim 1, wherein the inference result is obtained by using an operation such as y) (where, when (A^■-B^■) < 0, it is set as =0.). (3) Error information with respect to the target value, error change information, plant operation amount information, and amount of change in the absolute amount of error due to operation results are sequentially stored and it is determined whether the information is similar or not. If it is determined that this is the case, similar information is stored as a group of control rules with a larger absolute value of the amount of change in the absolute amount of error information due to the plant operation results, and the error information and control rule group for the target value are stored as a representative of the similar information. A control method characterized in that the amount of operation is determined by weighted average calculation and the control is executed. (4) Discrimination of information similarity using error information, error change information,
4. The control method according to claim 3, wherein the control method is performed by using a norm between vector quantities consisting of manipulated variable information and by determining the magnitude of the norm value. (5) Instead of the absolute amount of change in the absolute amount of error information,
Using the ratio of error information change, if the absolute value of the ratio is less than 1, it is represented by information for the case where the absolute value of the ratio is smaller, and if the absolute value of the ratio is greater than 1, the absolute value of the ratio is used. 4. The control method according to claim 3, wherein the information is represented by information when the value is larger than .