JP2000242497A - Method for assisting structuring of fuzzy inference model and fuzzy inference method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To structure a fuzzy inference model in a short time with high precision by numerically judging the validity of input data when the fuzzy inference model is made to learn by using the statistic indexes of those input data. SOLUTION: Learning data used for fuzzy inference are inputted first (S11). Then static calculation is performed so as to check linearity (S12). As a statistic index, the correlation coefficient of input and output data is normally obtained. The correlation coefficient is a numeral between -1 and +1; as the correlation coefficient is closer and closer to ±1, the linearity of the input data is higher and higher and the validity of the structured fuzzy inference model is higher and higher. Namely, the validity of the input data when the fuzzy inference model is made to learn is numerically judged by using those input indexes in the assisting method for structuring the fuzzy inference model of a system which performs inference and prediction by the fuzzy inference model learnt and structured by using specific input data and output data.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a fuzzy inference model construction support method and a fuzzy inference method applied to an information processing field for performing various inferences, predictions, controls, diagnoses, and the like using a fuzzy inference model.

[0002]

2. Description of the Related Art Fuzzy theory can be applied to non-linear systems in which "a human-like ambiguous judgment can be expressed by a computer", "knowledge of a skilled person can be expressed in the form of rules". , Etc., are widely applied in various fields such as control and diagnosis. On the other hand, neural networks have been widely applied to fields such as control and diagnosis, as in fuzzy theory, because of their features such as "has learning ability" and "can be applied to nonlinear systems." However, the neural network has a disadvantage that the output value is almost impossible to explain due to the black box inside.

[0003] In recent years, researches have been conducted to fuse the two to take advantage of their respective advantages. That is, by giving the fuzzy inference model a learning ability, the black box property of the neural network is eliminated. However, most of the research focuses on making the fuzzy inference model capable of learning, and little research has been done on explaining why inferences and predictions were obtained. Is the current situation. Also,
There are very few applications, most of which are simple function fittings.

[0004]

The performance of a fuzzy inference model depends more on the quality of input information than on a learning algorithm. If the input data is selected as data having a strong relationship with the output, a good fuzzy inference model can be constructed. However, if data that has no relation to the output is input, a good fuzzy inference model cannot be constructed even if learning is repeated. There is no research on how to select input data, and even today, operators and control engineers make decisions empirically, by trial and error, and it takes a long time to construct a fuzzy inference model. It was not always possible to construct a fuzzy inference model.

As another problem, the reason for inferring the output values (inference values such as predicted values and diagnostic results) of the fuzzy inference model (why such output values were obtained as a result of fuzzy inference) is explained in an easy-to-understand manner. There is still no way to do this. A fuzzy inference model differs from a neural network in that the internal calculation process is clear. However, the calculation method is complicated when compared with the linear system, and it is difficult for a general operator who is not a fuzzy engineer to understand why such a result is obtained.

Accordingly, the present invention provides a method for supporting the construction of a fuzzy inference model which determines the validity of input data used for learning a fuzzy inference model and constructs the fuzzy inference model with high accuracy and in a short time. Along with
It is intended to provide a fuzzy inference method capable of easily explaining the inference result to a general operator.

[0007]

Although fuzzy inference is applicable to nonlinear systems, it is often the case that higher linearity of input data indicates better inference results. In the inventions described in claims 1 and 2, utilizing this principle, the linearity of the input data used for learning is determined using a statistical index, and the validity of the fuzzy inference model is determined. Here, as the statistical index, attention is paid to the correlation coefficient of input / output data and the like.

FIG. 1 is a flowchart showing a processing procedure according to the first and second aspects of the present invention. Hereinafter, the contents will be described. (1) Learning data input (S11) Learning data used for fuzzy inference is input.

(2) Statistical index calculation (S12) Statistical calculation is performed to check the linearity. Usually, a correlation coefficient of input / output data is obtained as a statistical index. The correlation coefficient is a numerical value from −1 to +1 and the correlation coefficient is ± 1.
It can be said that the closer to, the higher the linearity of the input data, and the higher the validity of the constructed fuzzy inference model.

[0010] Therefore, the invention according to claim 1 is a fuzzy inference model in a fuzzy inference model construction support method in a system for performing inference / prediction using a fuzzy inference model learned and constructed using predetermined input data and output data. The validity of input data at the time of learning is numerically determined by using the statistical index of these input data. Further, as described in claim 2, it is preferable to use a correlation coefficient of input / output data as the statistical index.

Next, the invention described in claim 3 will be described. According to the first and second aspects of the present invention, the linearity is determined from the statistical index of the input data. However, it is known that the fuzzy inference shows a good result even for input data that is linear in a section such as a polygonal line or non-linear input data such as a curve. It is difficult to determine the validity of non-linear input data such as a broken line or a curve using only the first and second aspects of the present invention. In addition, numerical expressions are not suitable for operators who are not familiar with statistics. Therefore, in the invention of claim 3, the validity of the input data can be visually and intuitively determined.

FIG. 2 is a flowchart showing a processing procedure according to the third aspect of the present invention. Hereinafter, the contents will be described. (1) Learning data input (S21) Learning data used for fuzzy inference is input.

(2) Data correlation display (S22) A correlation diagram of each input data and output data is displayed. The displayed correlation diagram is a two-dimensional diagram in which one of the input data is represented on the x-axis and the output data is represented on the y-axis.
One example is a three-dimensional diagram in which one is represented on the x and y axes and the output data is represented on the z axis. By looking at this correlation diagram, it is possible to visually and intuitively judge the validity of the input data and, consequently, the validity of the fuzzy inference model.

According to a third aspect of the present invention, there is provided a method for supporting the construction of a fuzzy inference model in a system for performing inference and prediction using a fuzzy inference model learned and constructed using predetermined input data and output data. Is determined by visually displaying the correlation between the input data and the output data.

Next, the invention described in claim 4 will be described. Fuzzy inference is based on multiple linguistic rules, such as "Rule 1: If xx, then XX" and "Rule 2: If xx, then □□". Obtained using such linguistic rules. The present invention uses this principle to explain the reason for inference by indicating to the operator which rule was used and how much. This makes it possible to clearly explain the reason for fuzzy inference even to a general operator who is not a fuzzy engineer.

FIG. 3 is a flowchart showing a processing procedure according to the present invention. Hereinafter, the contents will be described. (1) Data input (S31) Data to be predicted / diagnosed is input to the fuzzy inference model.

(2) Calculating the degree of conformity (S32) The degree of conformity of each rule is calculated, and it is verified how much each rule affects the output value (inference value).

(3) Display (S33) The ratio of the influence of each rule on the output value is displayed. If there are many rules, displaying all the rules makes it difficult to understand. Therefore, usually, only rules having a high influence are sorted and displayed in order from the rule having a high contribution. This makes it possible to clarify the reason why the inference value is obtained from the contents of the rule.

Accordingly, a fourth aspect of the present invention is a system for performing inference / prediction using a fuzzy inference model, and calculates a degree of conformity of a plurality of rules when inference is performed by giving predetermined input data to the fuzzy inference model. Then, the degree of influence of each rule on output data is displayed.

[0020]

Embodiments of the present invention will be described below. In this embodiment, the inflow into the dam (for example, the prediction of the amount of water flowing into the dam one hour later) was performed by fuzzy inference. As a fuzzy inference structure, a simplified inference was used, and a learning algorithm constructed a fuzzy neural network using a gradient method. FIG. 4 shows a conceptual diagram of a prediction method using a fuzzy inference model.

In FIG. 4, the fuzzy inference model 10 includes, as input data, a current inflow, a current discharge, a difference (change) in inflow from a predetermined time ago, and a difference in discharge (change) in the same manner. Similarly, the difference (change) of the rainfall is given, and the inflow after one hour is obtained as the predicted value as the output data. Here, the inflow and discharge are those of the dam.

First, a first embodiment of the present invention will be described. In order to predict the inflow amount one hour after flowing into the dam, a data set including two types of input / output data (case 1 and case 2) shown in FIGS. 5A and 5B was prepared as learning data. Which of the data sets is better data is determined by using the first and second aspects of the present invention.

The statistical indices of these data are shown in FIG.
(A) and (b) show. According to FIGS. 6A and 6B,
Since the correlation coefficient of Case 2 is closer to ± 1 than Case 1 for each of the discharge difference, the inflow difference, and the rainfall difference of the input data, it can be determined that the input data has high linearity. In other words, learning using the data set of Case 2 can perform better learning than using Case 1.

FIG. 7 shows the prediction results (average absolute value error) for each case. In case 2, the error was smaller and a better prediction result was obtained. FIG. 8 shows a prediction result when the data set of Case 2 is used in comparison with an actual value, and the predicted value almost coincides with the actual value.
That is, according to the first and second aspects of the present invention, it is possible to select a data set having high validity as input data.

Next, a second embodiment of the present invention will be described. In this embodiment, in the data set of Case 2 shown in FIG. 5B, unnecessary input data is removed by using the first and second aspects of the present invention to improve the prediction accuracy. That is, as is clear from FIG. 6B, the correlation coefficient between the inflow amount absolute amount and the discharge amount absolute amount is close to 0, and the correlation between the input and output data is low. Therefore, in this embodiment, a new data set (Case 3) is created by excluding the data of the absolute inflow amount and the absolute amount of the discharge amount from the data set of Case 2, and learning is performed using this data set, thereby performing fuzzy inference. The model was built.

FIG. 9 shows a data set of Case 3. Input data in this data set is an inflow difference, a discharge difference, and a rainfall difference. FIG. 10 shows a prediction error when using the data set of Case 2 and a prediction error when using the data set of Case 3 in the first embodiment. From FIG. 10, it can be seen that a good prediction result can be obtained by learning using the data of Case 3 in which inappropriate data (data with low correlation) has been removed from the input data. That is, according to the first and second aspects of the present invention, the validity (in other words, inappropriateness) of the input data can be evaluated.

Next, a first embodiment of the present invention will be described. In the first and second embodiments of the first and second aspects of the present invention, appropriate or inappropriate input data used for learning is numerically determined by a statistical index. It is desirable for an operator who is not familiar with the system to be able to judge the validity of input data more visually and intuitively. In the first embodiment of the present invention, it is verified that the validity of input data can be visually and intuitively determined.

In the data sets of Case 1 and Case 2 described above, the data having the largest difference in the correlation coefficient was the data of the discharge difference. Therefore, on the assumption that the level of the correlation becomes visually clear in the data of the discharge flow difference,
By applying the invention of claim 3 to the data of the discharge difference, it is verified that the data in case 2 is better than the data in case 1.

FIG. 11 shows the correlation between the discharge rate difference as input data and the inflow rate difference as output data for each of the data sets in Case 1 (FIG. 11A) and Case 2 (FIG. 11B). It shows the relationship. From the figure, it is determined that in case 1, the relationship between input and output data is not linear or curved, and is unclear, whereas in case 2, the relationship between input and output data is almost linear and the correlation is determined to be high. Can be. As described above, according to the third aspect of the present invention, the correlation between the discharge amount difference and the inflow amount difference can be visually and intuitively determined to be higher in case 2 and more appropriate. This makes it possible to easily select input data to be used for constructing a fuzzy inference model regardless of the statistical and numerical proficiency of the operator.

Next, a second embodiment of the present invention will be described. In this embodiment, it is visually verified that, among the data sets of Case 2, the absolute inflow amount and the absolute discharge amount are inappropriate data for constructing the fuzzy inference model. FIGS. 12A to 12C show the correlation between each input data of the data set of Case 2 and the inflow amount difference which is the output data.
(E).

As can be seen from FIGS. 12 (a) to 12 (e),
The inflow amount absolute amount and the discharge amount absolute amount are not represented by a shape such as a straight line, a broken line, or a curve, and have large variations. Such data is more likely to be inappropriate as input data than linear data such as the discharge flow difference in FIG. Therefore, also in this embodiment, the validity of the input data can be visually evaluated and determined.

Next, an embodiment of the present invention will be described. In this embodiment, the reason for the inference based on the fuzzy inference, that is, the reason why the inference result was obtained, will be described using the inflow prediction of the dam as an example.

First, a fuzzy inference model for predicting the inflow one hour later was constructed from the rainfall in the upstream area, the past inflow actual value, the upstream dam discharge, and the like. Figure 1 shows input / output data
3 is shown. That is, the input data is the current dam inflow, the current upstream dam discharge, the difference of the dam inflow from 10 minutes ago, the difference in the dam discharge from 1 hour ago, the average rainfall for 15 hours up to now, and It is the difference of the average rainfall for 15 hours up to 3 hours before, and the output data is the inflow of own dam one hour ahead.

FIG. 14 shows a membership function (antecedent membership function) of fuzzy inference.
The upstream dam discharge flow, inflow difference, own dam discharge difference, and rainfall difference are set as shown in the figure. Here, since the number of rules is equal to the number of combinations of the membership functions, the number of rules is multiplied by 162 (= 2 × 3 ×
3 × 3 × 3).

FIG. 15 shows the result of dam inflow prediction. FIG. 15A is substantially the same as FIG.
15B shows the output value of the fuzzy inference according to the present embodiment, and FIG. 15C shows the prediction error. FIG. 16 shows the input data at point A in FIG. FIG. 17 shows inference reasons (prediction reasons) when inference is performed using these input data.

The following is clear from FIG. Note that the NO. 1 to NO. Rule 8 is selected from among 162 rules by paying attention to the degree of matching. The rule with the highest conformance is NO. The rule is that the inflow is small, the outflow is small, the inflow difference is medium, the outflow difference is medium, and the rainfall difference is large (the effect on the predicted value (degree of conformity) is 22). %). In addition, NO. 2 to NO. Little by little, it complied with the 7 rules of 8. Note that the list of rules and conformance as shown in FIG. 17 is displayed to the operator through a display or the like.

As described above, according to the present embodiment, the reason for inferring the reason why the inflow amount as the prediction result was derived and the reason for the prediction can be clarified by a combination of fuzzy rules. The reason for inference and the reason for prediction based on the empirical knowledge of the person can be shown more objectively, and can greatly contribute to the automation of the reasoning and prediction work.

[0038]

As described above, according to the first to third aspects of the present invention, input data used for learning a fuzzy inference model can be determined numerically or visually. Conventionally,
Since the input data was selected by the operator by trial and error,
There are drawbacks in that it takes a long time to select the most appropriate input data, and there is a case where inappropriate input data is selected, so that the accuracy is not improved. The present invention overcomes that shortcoming,
Even relatively inexperienced operators can select good input data. Therefore, a highly accurate fuzzy inference model can be easily constructed as compared with the related art.

In particular, the first and second aspects of the present invention determine the validity of input data numerically using a statistical index. Since the determination can be made numerically in this way, it is suitable for automation by a computer. The invention of claim 3 is:
Since the validity of the input data is visually expressed, even an operator or a control engineer who is not familiar with statistics can intuitively select the optimum input data. Further, it is possible to determine a curve or a broken line input / output relationship, which is difficult to determine in the first and second aspects of the present invention.

According to the fourth aspect of the present invention, it is possible to easily explain to the operator the reason why the output value of the fuzzy inference (inference value such as a predicted value and a diagnosis result) is obtained. Usually, an operator who performs a prediction operation makes a prediction based on an empirical rule of “XX because it is XX”. Similarly, when showing the reason for the forecast in an operation log, etc.,
XX and predicted. "Wrote an empirical rule. Conventional fuzzy inference models predict and predict
Although reasoning has been performed, no attempt has been made to display and explain the reason for the reasoning. However, according to the invention of claim 4,
The reason why the inference value is derived can be described in the same format as that of the operator. For this reason, it is possible to automate a prediction task that is difficult to automate in a computer by the same method as that performed by the operator.

Another effect of the invention of claim 4 is as follows.
It is easy to improve the fuzzy rules. Conventionally, when a large prediction error occurs, it has been difficult to determine whether the cause is in the prediction model or due to unique data. According to the invention of claim 4, since the reason for inference can be displayed, if there is an error, if the reason for inference is the same as the operator, it can be estimated that the data is peculiar, and if it is different from the operator, It can be estimated that there was a problem in the structure of the inference model. Then, when there is a problem in the inference model structure, it is possible to easily improve the fuzzy rules to be more accurate by adding a new rule of the operator.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a processing procedure of the invention described in claims 1 and 2;

FIG. 2 is a flowchart showing a processing procedure of the invention described in claim 3;

FIG. 3 is a flowchart showing a processing procedure of the invention described in claim 4;

FIG. 4 is a conceptual diagram of a prediction method using a fuzzy inference model.

FIG. 5 is an explanatory diagram of a data set according to the first embodiment of the present invention.

FIG. 6 is an explanatory diagram of a statistical index according to the first embodiment of the present invention.

FIG. 7 is a diagram showing a prediction error in the first embodiment of the first and second aspects of the present invention.

FIG. 8 is a diagram showing, in the first embodiment of the first and second aspects of the present invention, a prediction result when a data set of case 2 is used, in comparison with an actual value.

FIG. 9 is an explanatory diagram of a data set according to the second embodiment of the present invention.

FIG. 10 is a diagram showing prediction errors in the first and second embodiments of the first and second aspects of the present invention.

FIG. 11 is a diagram showing a correlation between input and output data for case 1 and case 2 in the first embodiment of the third invention.

FIG. 12 is a diagram showing a correlation between input and output data for case 2 in the second embodiment of the third invention.

FIG. 13 is an explanatory diagram of input / output data of fuzzy inference in the embodiment of the fourth invention.

FIG. 14 is an explanatory diagram of a membership function of fuzzy inference in the embodiment of the present invention.

FIG. 15 is an explanatory diagram of a prediction result and an error in the embodiment of the fourth invention.

FIG. 16 is an explanatory diagram of input data at point A in FIG. 15;

FIG. 17 is an explanatory diagram of a reason for prediction in the embodiment of the invention of claim 4;

Claims (4)

[Claims] 1. A method for supporting a fuzzy inference model in a system for performing inference and prediction using a fuzzy inference model learned and constructed using predetermined input data and output data. A method for supporting the construction of a fuzzy inference model, characterized in that the validity is numerically determined using the statistical index of these input data. 2. The fuzzy inference model construction support method according to claim 1, wherein the statistical index is a correlation coefficient of input / output data. 3. A method for supporting the construction of a fuzzy inference model in a system for performing inference / prediction using a fuzzy inference model learned and constructed using predetermined input data and output data. A method for supporting construction of a fuzzy inference model, characterized in that the validity is visually determined by judging the correlation between the input data and the output data. 4. A system for performing inference / prediction using a fuzzy inference model, wherein a predetermined input data is given to a fuzzy inference model to calculate a degree of conformity of a plurality of rules, and output data of each rule is calculated. Fuzzy inference method, characterized by displaying the degree of influence on the fuzzy.