JPH02228763A - Device and method for forecasting - Google Patents

Abstract

PURPOSE:To make forecasting execution efficient, to reflect the experience, knowledge and rule of a designer and to construct the model of satisfactory accuracy by enlarging the width of the forecasting model according to a function approximating method and handling the model besides an independent linear model between parameters. CONSTITUTION:A function approximation expression identifying means 10 is provided to determine the function approximation expression by using data retrieved from a referring data retrieving means 9 and an intermediate forecasting value calculating means 11 is provided to calculate an intermediate forecasting value according to the identified approximation expression, the value of the parameter, correspondent characteristic value and the value of the input parameter. Then, a forecasting value calculation control means 23 is provided to define the intermediate forecasting value as the new characteristic value and to control the intermediate forecasting value calculating means 11 so as to successively repeatedly execute calculation concerning each parameter having difference with the input parameter value. Thus, even in a field for which there is no method excepting for calculating the rough characteristic value according to the intuition of a specialist, for example, in the life forecasting model, etc., of a capacitor, this processing can be effectively utilized and the reliability of the forecasting model can be improved.

Description

[Detailed Description of the Invention] Industrial Application Field The present invention automatically generates laws and mechanisms from unknown models that have not been theoretically elucidated in various phenomena such as chemistry and physics, and predicts the lifespan of parts. The present invention relates to a prediction device for making various predictions such as product demand prediction.

2. Description of the Related Art As an example of a conventional prediction device, a device for predicting the life of a capacitor will be described.

FIG. 2 is a configuration diagram of this capacitor life prediction device.

In Fig. 2, 2 is an input means for inputting the values of multiple types of parameters necessary for predicting characteristic values, 3 is an output means for outputting the prediction results, and 5 is an input means for inputting the values of multiple types of parameters necessary for predicting characteristic values. Data storage means 12 is a data storage means 5 which stores a plurality of data sets each consisting of a value and a characteristic value corresponding to each parameter value group.
Multiple regression model identification means that uses multiple sets of data of parameter value groups and corresponding characteristic values stored in , to find relational expressions between multiple parameters and characteristic values by multiple regression analysis.

13 is predicted value calculation means 201 for calculating a predicted value using the prediction model identified by the multiple regression model identification means 12;
to 205 are control signals and commands.

This is a signal line that indicates data, etc., and predicts the life of the capacitor through the following operations.

First, a plurality of types of capacitor parameter data 201 necessary for life prediction are inputted from the input means 2 and passed to the single regression model identification means 12. Multiple regression model identification means 1
2 refers to the plurality of types of parameter values obtained through experiments or observations and the characteristic values corresponding to each parameter value group in the data 203, and calculates the relational expression between the plurality of types of parameters and characteristic values by multiple regression analysis. Next, the predicted value calculation means 13 receives a plurality of types of parameter data 2 sent from the input means 2.
05 and data 20 regarding relational expressions between multiple types of parameters and characteristic values calculated by the multiple regression model identification means 12
4 to calculate the predicted value.

The result is sent to the output means as data 202.

Displayed and output.

FIG. 4 is a flowchart showing the operation of a conventional capacitor life prediction device.

When multiple types of parameter values are input from the input method (
401), the data is passed to the multiple regression model identification means 12, and it is checked whether data matching the plurality of types of parameters exists in the data storage means.

for example.

Input parameters Voltage = 8.3v Capacity = 47mf Size = 15*20 Material = A substance Electrolyte = B substance Data base Parameter 1. Characteristic value 1〉 Parameter 2. Characteristic value 2><Parameter n, characteristic value n> If the input parameter and parameter l match, the corresponding characteristic value i is returned (402).

If it does not exist, the multiple regression model identification means 12 performs multiple regression on the relational expression between the multiple types of parameters and the characteristic values by referring to the multiple sets of multiple types of parameter value groups and their corresponding characteristic values stored in the data storage means 5. The predicted value calculation means 13 calculates a predicted value using the automatically generated prediction model (403), and the output means displays and outputs the result. (404) The model identification means using single regression analysis will be briefly explained. First, multiple regression analysis is a linear model where parameters are independent.

V = ai * xi + aO where y: Characteristic value XI: Parameter ai: Coefficient applied to each parameter (i = le...m) aO: A plurality of constant terms and the shape of the function are determined and stored in the data storage means. This is a method of identifying a predictive model by calculating unknown constant terms and coefficients using the least squares method so that the relationship between parameters and characteristic values best matches this equation. The characteristic value y is calculated by substituting the input parameters Xl...XmB of the prediction model, and this value is used as the predicted value (404).

Problems to be Solved by the Invention However, the above-described configuration has a problem in that it can only handle linear models in which the dimensions are independent. There is also a neural network model that can handle nonlinear models as another processing method for prediction and inductive reasoning, but because it is a black box model, it is not suitable when you want to use it as a support system for elucidating mechanisms or discovering rules. . For this reason, at present, the theoretical explanation of the 11 models of capacitor life expectancy has not been made, and at least 2
The lifespan was measured through trial experiments lasting over 1,000 hours (approximately 3 months). This requires a lot of time to calculate the lifespan and is also problematic in terms of credibility. Furthermore, since such statistical methods process data all at once, they are not suitable for analyzing the mechanism of one element 9 among parameters, such as the relationship between voltage and characteristic values. For this reason, design work that involves repeated combinations of new materials and changes in the size and capacity of capacitors requires a great deal of time and the skill of the designer.

Also, prediction methods that consist only of statistical methods.

Another disadvantage is that existing important information such as the designer's experience and knowledge and partially elucidated rules cannot be used.

Even if the predictive model is not very accurate.

If the observed data does not change, the predictive model cannot be changed automatically.

In view of this, the present invention has developed a new method called function approximation method in addition to a prediction method using multiple regression analysis, which is a statistical method, and a prediction method using a neural network model.
The range of prediction models has been expanded to handle models other than linear models where parameters are independent, and the efficiency of prediction execution has been improved to reflect the designer's empirical rules, knowledge, and partially elucidated rules. The purpose of the present invention is to provide a prediction device that can construct a model with higher accuracy each time the prediction is executed using a learning function even if the observation data is not changed.

Means for Solving the Problems The first invention provides an input means for inputting values of multiple types of parameters necessary for predicting characteristic values, values of multiple types of parameters obtained by experiment or observation, and the parameter values. data storage means for storing a plurality of data sets consisting of characteristic values corresponding to each group; weighting coefficient storage means for storing weighting coefficients indicating the importance of each of the plurality of types of parameters; approximate data searching means for comparing an input parameter value group with a parameter value group stored in the data storage means while performing weighting using a weighting coefficient, and searching for a parameter value group closest to the input parameter value group; and the approximate data Searching the data storage means for data necessary for determining the relationship between the parameter and the characteristic value by a function approximation method for a parameter that is different from the input parameter value among the approximate parameter value group retrieved by the retrieval means. Reference data search means.

a function approximation formula identifying means for determining a function approximation formula using the data retrieved by the reference data search means; a function approximation formula identified by the function approximation formula identification means, the value of the parameter, and the parameter; an intermediate predicted value calculation means for calculating an intermediate predicted value using the characteristic value corresponding to the value of the input parameter and the value of the input parameter; The prediction device includes predicted value calculation control means for controlling the intermediate predicted value calculation means to sequentially and repeatedly perform calculations for each parameter having a difference from a parameter value.

Further, in the second invention, in the first invention, the prediction execution means is constituted by the approximate data retrieval means, the reference data retrieval means, and the intermediate predicted value calculation control means, and the function approximation formula identification means is configured by the prediction model storage means and the prediction model automatic This is a prediction device characterized by comprising a creation means.

Further, a third invention is a prediction device according to the second invention, characterized in that an interaction processing means for setting data and rules stored in the prediction model storage means and the data storage means is provided.

Further, the fourth invention includes a step of inputting values of multiple types of parameters necessary for predicting characteristic values, and a step of inputting values of multiple types of parameters obtained by experiment or observation and corresponding to each group of parameter values. a data storage step of storing a plurality of data sets consisting of characteristic values; a weighting coefficient storing step of storing weighting coefficients indicating the importance of each of the plurality of types of parameters; and input parameters while being weighted by the weighting coefficients. an approximate data retrieval step of comparing the value group with the parameter value group stored in the data storage step and searching for a parameter value group closest to the input parameter value group; and an approximate parameter value retrieved by the approximate data retrieval step. a reference data retrieval step of retrieving data necessary for determining the relationship between the parameter and the characteristic value using a function approximation method for a parameter that is different from the input parameter value of the group; a function approximation expression identification step in which a function approximation expression is determined using the data retrieved in the search step; an intermediate predicted value calculation step of calculating an intermediate predicted value using the corresponding characteristic value and the value of the input parameter, and the intermediate predicted value calculated by the intermediate predicted value calculation step is used as a new characteristic value and the input parameter value is This prediction method includes a predicted value calculation control step for controlling the intermediate predicted value calculation step to sequentially and repeatedly perform calculations for each parameter having a difference.

The fifth invention also provides human power means for inputting values of multiple types of parameters necessary for predicting characteristic values, control means for controlling the entire device, and values of multiple types of parameters obtained by experiment or observation. and a data storage means for storing a plurality of data sets each consisting of a characteristic value corresponding to each parameter value group.

Weighting coefficient storage means for storing weighting coefficients indicating the importance of each of the plurality of types of parameters.

Predictive model automatic creation means for automatically creating a predictive model by referring to the weighting coefficients and data stored in the data storage means, and a predictive model created by the predictive model automatic creation means is stored. a prediction model storage means; a prediction execution means for executing a prediction by referring to the data, the prediction model, and the weighting coefficient; and an evaluation of the predicted value predicted by the prediction execution means and the evaluation of the prediction model and the weighting coefficient. It is composed of a learning means that performs dynamic changes, and an output means that displays and outputs the prediction result calculated by the prediction execution means, the parameters input from the input means are sent to the control means, and the control means , it is checked whether an appropriate prediction model exists in the prediction model storage means, and if it does not exist, the automatic prediction model creation means is activated, and the automatic prediction model creation means combines the data and the weighting coefficients. Reference and create a predictive model.

The results are stored in the predictive model storage means.

The control means passes control to the prediction execution means.

The prediction execution means executes prediction using the prediction model,
If an appropriate prediction model already exists in the prediction model storage means, the process of the automatic prediction model creation means is omitted, and the prediction execution means is directly activated to execute prediction using the already existing prediction model. The prediction execution means returns the prediction result to the control means, the control means passes control and the prediction result to the learning means, the learning means evaluates the predicted value, and if the evaluation result is incorrect, the The weighting coefficients are dynamically changed by the learning algorithm, and the control means again passes control to the automatic prediction model creation means, and the automatic prediction model creation means creates the prediction model again by referring to the weighting coefficients changed by the learning algorithm. The prediction device is characterized in that the prediction execution means has a learning function that performs prediction execution with reference to the updated prediction model.

Further, in a sixth invention, in the fifth invention, the predictive model storage means is constituted by a predictive model storage means for storing the predictive model and a cluster analysis result storage means for distributing the predictive model for each class, and the learning means is configured to store the predictive model by storing the predictive model for each class. This is a prediction device characterized by comprising an evaluation means and a weighting coefficient dynamic change means.

According to the first invention, a data set consisting of the values of a plurality of types of parameters obtained through experiments or observations and characteristic values corresponding to each parameter value group is stored in the data storage means, and the data sets of the plurality of types are stored in the data storage means. A weighting coefficient indicating the importance of each parameter is stored in a weighting coefficient storage means, and the approximate data storage means weights the input parameter value group and the parameter value group stored in the data storage means while performing weighting using the weighting coefficient. The reference data search means searches for a parameter value group that is closest to the input parameter value group, and the reference data search means searches for the parameters and characteristics of the parameters that are different from the input parameter values among the approximate parameter value groups searched by the approximate data search means. Retrieve the data necessary to determine the relationship with the value using the function approximation method from the data storage means.

The function approximation formula identification means determines the function approximation formula using the data retrieved by the reference data search means, and the intermediate predicted value calculation means uses the function approximation formula identified by the function approximation formula identification means and the parameter. and the characteristic value corresponding to the value of the parameter and the value of the input parameter, and the predicted value calculation control means uses the intermediate predicted value calculated by the intermediate predicted value calculation means as a new characteristic value. , by controlling the intermediate predicted value calculation means to perform repeated calculations in sequence for each parameter that has a difference from the input parameter value, and by repeating this series of processes for the number of differences, the characteristics corresponding to multiple types of parameter values to be predicted can be calculated. Value calculations can be performed based on approximate data,
In addition to improving calculation accuracy and efficiency, since a function approximation formula is applied to approximate data in a recursive manner, a predictive model based on a nonlinear combination of parameters that cannot be expressed by multiple regression analysis can be identified. Furthermore, if the recursive action of the function approximation formula is output to the output means, it becomes possible to analyze the mechanism of the model.

Moreover, the second invention comprises a prediction execution means by an approximate data search means, a reference data search means, an intermediate predicted value calculation means, and a predicted value calculation control means, and the function approximation formula identification means is configured by a prediction model storage means and a prediction model automatic The method consists of a creation means and clearly separates the means for only performing predictive execution, the means for automatically constructing a predictive model, and the predictive model storage means for storing the predictive model, so in the case of predictive execution that does not involve updating the predictive model. Since it is possible to omit the processing of the automatic prediction model creation means and the prediction model storage means is separated,
The prediction model can be easily modified by the interaction processing means.

In addition, the third invention provides an interactive processing means to enable the user to modify, process, add, etc. the predictive model and data that stores the predictive model, and improve the user's predictive model and parameters. Experience and knowledge regarding operations and partially elucidated rules can be reflected.

Further, the fifth invention is a learning method comprising dynamically changing the prediction result evaluation function 1 weighting coefficient by a learning algorithm, dynamic clustering of the data stored in the data storage means, and reconstructing the prediction model. By realizing this function, even if observation data is not updated, a more accurate prediction model can be constructed with each prediction execution.

EXAMPLE Hereinafter, a first example of the prediction device of the present invention will be described with reference to the drawings, taking a capacitor life prediction device as an example.

In FIG. 1, 1 is a capacitor life prediction device.

2 is an input means through which the user inputs values of multiple types of parameters necessary for predicting characteristic values; 22 is a parameter analysis means for analyzing a group of parameter values inputted by the input means 2; and 5 is an input means obtained by experiment or observation. data storage means for storing a plurality of data sets each consisting of a plurality of parameter values and characteristic values corresponding to each parameter value group; e is a weight indicating the importance of each of the plurality of parameter values; The weighting coefficient storage means 7, which stores coefficients, compares the input parameter value group with the parameter value group stored in the data storage means 5 while performing weighting using the weighting coefficient, and selects the one closest to the input parameter value group. Approximate data search means 8 searches for a group of parameter values, difference parameter detection means 8 detects a parameter different from the input parameter among the group of approximate parameter values searched by the approximate data search means 7, 9 searches by the approximate data search means reference data retrieval means for retrieving from the data storage means 5 data necessary for determining the relationship between the parameter and the characteristic value by the function approximation method for a parameter that is different from the input parameter value among the approximate parameter value group; , 10 is a function approximation formula identification means for determining a function approximation formula using the data searched by the reference data search means 9, and 11 is a function approximation formula identified by the function approximation formula identification means 10 and the relevant parameter. Intermediate predicted value calculation means 23 calculates an intermediate predicted value based on the characteristic value corresponding to the value of the parameter and the value of the input parameter; As a value, predicted value calculation control means controls the intermediate predicted value calculation means to perform repeated calculations in sequence for each parameter that is different from the input parameter value, and 3 is the final prediction calculated by the intermediate predicted value calculation means. This is an output means for outputting the results.

Signal lines 101 to 116 indicate control signals, commands, data, etc., and the life span of the capacitor is predicted by the following operations.

First, a group of parameter values 101 of a capacitor whose life is to be predicted is inputted from the input means 2 and passed to the parameter analysis means 4. The parameter analysis means 22 uses the control signal 10
2 activates the approximate data search means 7. The activated approximate data retrieval means 7 retrieves data 103 and weighting coefficients 10 prepared in advance in the data storage means 5 consisting of parameters and characteristic values obtained from past experiments.
4 to search for data closest to the input parameters.

Next, the difference parameter detection means 8 detects parameters having different values between the input data 105 and the data 106 searched by the approximate data search means 7. Next, the reference data search means 9 generates data (parameter type) 10 regarding the difference parameter detected by the difference parameter detection means 8.
7, a data set in which only the value of the relevant parameter differs and the values of other parameters are the same is divided into weighting coefficient data 11
0 and the data 108 stored in the data storage means 5 are searched. Next, the function approximation equation identifying means 10 uses the data set 112 searched by the reference data search means 9 to automatically identify a function approximation equation representing the relationship between the parameter and the characteristic value. The method of this automatic identification will be described later. Next, the intermediate predicted value calculation means 11 uses the function approximation formula 113 identified by the function approximation formula identification means 10, the value of the approximation parameter, the characteristic value corresponding to the value of the parameter, and the value of the input parameter to make an intermediate prediction. Calculate the value. This calculation brings the predicted value corresponding to the input data one step closer. The predicted value calculation control means 23 sets the intermediate predicted value calculated by the intermediate predicted value calculation means 11 as a new characteristic value in the intermediate predicted value calculation means 11 with data 114 and 115, and sets the intermediate predicted value calculated by the intermediate predicted value calculation means 11 as a new characteristic value, and sets the intermediate predicted value calculated by the intermediate predicted value calculation means 11 as a new characteristic value, and sets the intermediate predicted value calculated by the intermediate predicted value calculation means 11 as a new characteristic value. The intermediate predicted value calculating means 11 is controlled to perform calculations similar to those described above for other parameters. This predicted value calculation control means 12
similarly controls the intermediate predicted value calculation means 11 to sequentially repeat calculations for all other parameters that are different from the input parameter value.

The output means 3 displays the result as data 116 when the intermediate predicted value calculation means 11 obtains the calculation result for the final parameter.

FIG. 3 is a flowchart showing the operation of this embodiment.

When a parameter value group is input from the input means 2 (301
), the parameter analysis means 22 checks whether data matching the parameter value group exists in the data storage means 5 (302).

For example, let the values of the five input parameters be voltage = 16V, size = 8, foil material = C3, electrolytic paper = M1, electrolyte = 21, and the following 10 data sets are stored in the data storage means 5. Suppose there is.

Note that the storage format in the data storage means 5 is
Species parameters and characteristic values>.

Sample 1: <8.310 03 MI Zl, 8
>Sample 2: <Ei, 31G O3M2 z,
1.18) Sample 3: <I[i12゜503M2Z1,
8) Sample 4: <18 1Ei C3M2Z1,8
>Sample 5: <35 8 02 old Zl, 3.4>Sample 6: <35 12.5 C2M2 Zl, 3.
3) Sample 7: <35 18 02 M2 Zl,
3 Sample 8: <50 8 02 old Zl, 3.3) Sample 9
: <50 12.502 M221.3.2> Sample 10: <50 1Ei C2M2Z1.2.8) The weighting coefficient of each parameter is as follows.

Voltage Size Foil Electrolytic paper Electrolyte weight 1 12 2 2 In this example,
Since there is no parameter matching the input parameter in the data storage means 5, the process advances to the next step (303).

Since the approximate data search means 7 refers to the data storage means 5 and the weighting coefficient storage means 6 and searches for data close to the input parameter value, the data of the sample 1 is retrieved (303
).

The method of calculating the degree of closeness is to check the input parameter value and the parameter value stored in the data storage means 5, add the weighting coefficients corresponding to the parameters with different parameter values, and select the minimum value. Let the data be the approximate data.

The difference parameter detection means 8 examines the approximate data and the value of the input parameter and detects that there is a difference between the two parameters, voltage and magnitude (304).

The reference data search means 9 searches for data samples 2 and 4 necessary for interpolating the first difference voltage from the approximate data to the voltage of the input data (305).

Note that in this example, only two reference data are retrieved because a linear function is used in the function approximation formula.For example, if approximation is to be performed using a quadratic equation, three reference data must be selected. Must be.

The function approximation equation identifying means 10 refers to this data and finds equations regarding voltage and characteristic values (306).

The intermediate predicted value calculation means 11 changes the characteristic value by changing the voltage term from approximate data to input data using the obtained function approximation formula (307).

The predicted value calculation control means 23 controls the processing to be repeated until the parameter part of this series of processed approximate data matches the input data, and if they match, the output means 3 displays and outputs the result (308).

The function approximation method is a method of deriving the relationship between one element of the parameters and the characteristic value, but the second method shows that by repeating this method for the number of elements, it is possible to identify a predictive model that is not a linear combination of parameters. This will be explained using an example.

Prediction model D7) using multiple regression analysis The general formula is expressed as follows using parameters (xl, x2+ + r x n) and characteristic value y.

y=f L (xi)+f2 (x2)+e*
・+fn(xn) Also, when expressed using the simplest model form, y:a 1xl+a2x2+s fist e + a
.

becomes.

Next, if we use the function approximation method according to the present invention.

y:fl(Xl)xf2(X2)*o...*fn(xn
), it is possible to identify it even if it is not a linear combination of parameters. However, in the embodiment of the present invention, fi(xi) is limited as follows.

When the relationship between xi and y is directly proportional, fi (xf) = aixi + bixi and when the relationship between bixi and y is inversely proportional, fi (xi) = 1/(aixi + bi) is used.

First, the method for identifying +fi will be briefly explained.

xL x2+ ” ” ”+ XI-
L

y=A* (aixi+b 1) =cixi+di (however, c i=A*a i, d i=A*
b i) or.

7=A*1/ (a ix i+b) =1/(cixi+di) (However, + c i=a t/As d L = b
t/A).

This is to identify the function approximation formula for y and xt.

The reference data retrieval means searches for two appropriate data, and the function approximation equation identification means refers to it to obtain +CI+di and identifies the function approximation equation fi that flashes on y and xi.

Next, based on the identified fi, the intermediate predicted value calculation means calculates an intermediate predicted value using the following formula.

If y and xi are directly proportional, ynb' = (xnb i' /xnb i)
* (Ynb-d i) +d When iy and xi are inversely proportional, Ynb' = ynb* (xnb i*c i+d i
)/(xnbi'*ci+di) However, r ynb' is the intermediate predicted value of the characteristic value when the approximate data is brought one step closer to the input parameter, and +Vn
b is the intermediate predicted value of the current approximate data, xnbi is the element of the i-th parameter of the approximate data + x
nbi' is the i-th element of the input parameter.

FIG. 9 illustrates a mechanism for identifying a function approximation formula and calculating an intermediate predicted value in the present invention. When the approximate data of the input parameter is retrieved, data necessary for interpolating the difference (xi IT
+ r Xnl+ 7+) (however, i=1, 2)
is searched by the reference data search means. Next, by 1ndex of the difference detected by the difference parameter detection means, +
If 1ndex is 1, add 1ndex to 902
If is 2, it becomes 903.

If 1ndex is n, control is passed to 904. 902, 903, and 904 to which control is passed are the above-mentioned C11
dl is calculated, intermediate predicted values are calculated by intermediate predicted value calculation means 906, 907, and 908 based on the characteristic values of the approximate data 905, and this series of processing is repeated until the approximate parameters match the input parameters. The predicted value calculation control means performs control, and if they match, calculates the predicted value in 909.

In addition, such a prediction method using the function approximation method uses input means for inputting multiple types of parameters necessary for capacitor life prediction as input means for inputting multiple types of parameters necessary for prediction. A data base consisting of values of multiple types of parameters of a capacitor and characteristic values corresponding to each parameter value group is stored in a data base consisting of values of multiple types of parameters of the target to be predicted and their characteristic values. Replace the weighting coefficient storage means that stores weighting coefficients indicating the importance of each of the plurality of types of parameters with the weighting coefficient storage means that stores the weighting coefficients that indicate the importance of each of the plurality of types of parameters to be predicted. Therefore, it goes without saying that the present invention can be used not only for a capacitor life prediction device but also for a membrane prediction device and an inductive reasoning device.

Regarding the relationship between inductive reasoning and prediction, prediction in a broad sense means elucidating a model from an unknown model and calculating prediction results using that model, so it can be considered equivalent to inductive reasoning. It's fine.

Further, as processing methods for prediction and inductive reasoning, there are statistical methods using multiple regression analysis and neural network models.

Next, a third embodiment of the prediction device of the present invention will be described with reference to the drawings.

In FIG. 5, 1 is a prediction device, and 2 is an input means for inputting a plurality of types of parameters necessary for prediction.

Reference numeral 4 denotes a control means having a control function for the entire device; 5 stores a plurality of data sets each consisting of values of a plurality of types of parameters obtained through experiments or observations and characteristic values corresponding to each group of parameter values; data Φ base storage means;
15 is a predictive model automatic creation means for automatically creating a model by referring to the data; 18 is a predictive model storage means for storing data regarding the predictive model automatically created by the predictive model automatic creation means; Prediction execution means for performing prediction execution; 3 is output means for outputting prediction results;
501 to 510 are control signals and commands.

This is a signal line that indicates data, etc., and prediction is performed by the following operations.

First, a plurality of types of parameter data 501 of objects to be predicted are inputted from the input means 2 and passed to the control means 4. The control means 4 checks the data 510 to see if a suitable model exists in the predictive model storage means 16. If an appropriate prediction model exists, the control means 4 does not pass control to the automatic prediction model creation means 15, but activates the prediction execution means 14 with the control signal 502, and the activated prediction execution means 14 performs experiments and tests. A predicted value is calculated by referring to data 504 regarding a plurality of types of parameter value groups obtained through observation, characteristic values corresponding to each parameter value, and data 506 regarding a predictive model, and the results are sent to the output means 3 as data 508. The output means 3 displays and outputs it. In addition, if an appropriate prediction model does not exist, the control means 4
The control signal 503 activates the predictive model automatic creation means.

The activated predictive model automatic creation means 15 refers to the contents stored in the data storage means 5 as data 505, automatically creates a predictive model, and uses the results as data 50.
7 and stored in the predictive model storage means 16. moreover,
The control means 4 activates the prediction execution means 14 and performs the same processing thereafter.

The feature of this embodiment is that the flow of the prediction function is clearly separated into a prediction execution means, a prediction model automatic creation means, and a prediction model storage means.

As a result, there is no need to start the automatic prediction model creation means unless it is necessary to change the prediction model due to changes in observation data or weighting coefficient data, which not only improves the efficiency of prediction execution processing, but also allows users to Processing such as correction, processing, addition, etc. of the prediction model can be easily realized.

Next, a fourth embodiment of the prediction device of the present invention will be described with reference to the drawings.

In FIG. 6, 1 is a prediction device, and 2 is an input means for inputting a plurality of types of parameters necessary for prediction.

Reference numeral 4 denotes a control means having a control function for the entire device; 5 stores a plurality of data sets each consisting of values of a plurality of types of parameters obtained through experiments or observations and characteristic values corresponding to each group of parameter values; database storage means;
15 is a predictive model automatic creation means that automatically creates a model by referring to the data; 16 is a predictive model storage means that stores data regarding the predictive model automatically created by the predictive model automatic creation means; 3 is an output means for outputting a prediction result; 17 is a prediction execution means for executing a prediction; Interactive processing means for interactively modifying types, values, and characteristic values by users, 601 to 6
15 is a control signal.

This is a signal line that indicates commands, data, etc., and prediction is performed by the following operations.

First, a command 601 for executing prediction is inputted from the input means 2 and passed to the control means 4. The control means 4 analyzes the command and checks the data 615 to see if an appropriate model exists in the predictive model storage means 16. If an appropriate prediction model exists, the control means 4 does not pass control to the automatic prediction model creation means 15, but activates the prediction execution means 14 with a control signal 602, and the activated prediction execution means 14 performs experiments or tests. Data 604 regarding parameters and characteristic values obtained through observation and data 606 regarding prediction models
A predicted value is calculated with reference to , and the result is sent to the output means 3 as data 608, and the output means 3 outputs it for display. If no suitable prediction model exists, the control means 4 activates the automatic prediction model creation means using a control signal 603.

The activated predictive model automatic creation means 15 automatically creates a predictive model by referring to the contents stored in the data storage means as data 605, and stores the result in the predictive model storage means 16 as data 607. Further, the control means 4 activates the prediction execution means 14 and performs the same processing thereafter. Further, an interactive processing command 601 is input from the input means 2.

When the command is passed to the control means 4, the control means 4 analyzes the command and converts the contents stored in the data storage means 5 and the prediction model into data 611. Correction fist processing is performed using data 612°, data 613, and data θ14.

The feature of this embodiment is that the prediction function includes not only a predictive model automatic creation means but also an interactive processing means, which enables various models automatically constructed, such as models using heavy regression analysis. , models obtained using algorithms such as function approximation methods and neural network models, and by modifying and processing these algorithm models, and using this algorithm model as part of a user-defined model, users can Because it is possible to build a flexible model that reflects the experience, knowledge, and partially elucidated rules, it can be used not only as a prediction device but also as a support device for model elucidation.

Next, a fifth embodiment of the prediction device of the present invention will be described with reference to the drawings.

First, in Fig. 7, 1 is a capacitor life prediction device;
2 is an input means through which the user inputs multiple types of /f parameters necessary for prediction; 4 is a control means having a control function for the entire device; and 5 is a group of multiple types of parameter values obtained through experiments or observations and their values. a database base storage means for storing characteristic values corresponding to each parameter value group; 6 a weighting coefficient storage means for storing weighting coefficients indicating the importance of each of the parameters; 15 for the data storage means; Predictive model automatic creation means for automatically creating a model by referring to stored data; 19 is data stored in the data storage means for improving the accuracy of the predictive model activated by the predictive model automatic creation means; Cluster analysis means to classify.

18 is a cluster analysis result storage means for storing the results of the cluster analysis calculated by the cluster analysis means; 14 is a prediction execution means for executing a prediction by referring to the data in the data storage means, the prediction model, and the weighting coefficient. , 20 is a prediction result evaluation means for evaluating the predicted value predicted by the prediction execution means, and 21 is a prediction result evaluation means for dynamically changing the content of the weighting coefficient by a learning algorithm if the result evaluated by the prediction result evaluation means is bad. weighting coefficient dynamic changing means; 3 is an output means for displaying and outputting the prediction results calculated by the prediction execution means; 701 to 724 are signal lines indicating control signals, commands, data, etc., and the prediction is performed by the following operations. Is going.

First, a command 701 for executing prediction is inputted from the input means 2 and passed to the control means 4. The control means 4 analyzes the command and checks the data 724 to see if a suitable model exists in the predictive model storage means 16. If a suitable prediction model does not exist, the control means 4 activates the automatic prediction model creation means with the control signal 70θ, and the activated automatic prediction model creation means 15 uses the control signal 70θ to activate the automatic prediction model creation means 15.
At step 22, the cluster analysis means 19 is activated, and the cluster analysis means 19 refers to the data 713 and the data 718 regarding the weighting coefficients to carry out the classification process of the data stored in the data storage means in order to improve the accuracy of the prediction model. Perform cluster analysis.

The results are stored as data 717 in the cluster analysis storage means, and the cluster analysis means 19 returns control to the predictive model automatic creation means using a control signal 721. Next, the predictive model automatic creation means 15 generates data 714, data 716 regarding the results of cluster analysis, and data 7 regarding the weighting coefficients.
automatically create a prediction model with reference to 19 and store the result in the prediction model storage means 16 as data 715,
The automatic prediction model creation means 15 passes control to the control means 4 using a control signal 707, and the control means 4 activates the prediction execution means 14 using a control signal 702. In addition.

If an appropriate prediction model exists, the control means 4 does not pass control to the automatic prediction model creation means 15, and outputs the control signal 702.
The prediction execution means 14 is directly activated. The activated prediction execution means 14 collects data 709 on parameters and characteristic values obtained through experiments and observations, data 710 on prediction models, and data 711 on the results of cluster analysis.
and data 71 regarding weighting coefficients of multiple types of parameters.
Calculate the predicted value with reference to 2 and use the result as data 70.
3 to the control means, the control means 4 activates the prediction result evaluation means 20 with a control signal 704 to evaluate the prediction result, the prediction result evaluation means 20 evaluates the prediction value, and if the prediction result is good, returns the result to the control means 4 as data 708.

If the prediction result is bad, the prediction result evaluation means 20 activates the weighting coefficient dynamic change means 21 with a control signal 723.

The activated weighting coefficient dynamic changing means 21 changes the content of the weighting coefficient using a learning algorithm and stores the result in the weighting coefficient storage means 6 as data 720. Furthermore, the 1-weighting coefficient dynamic changing means 21 activates the predictive model automatic creation means 15 using a control signal 722. The activated predictive model automatic creation means 15 activates the cluster analysis means 19, and the activated cluster analysis means 19 stores data based on the weighting coefficient data changed in the above processing. The data stored in the means is reclassified, and a predictive model is rebuilt using the reclassified clusters. A prediction execution process is performed using the reconstructed prediction model, and the prediction execution means 14 returns the prediction result to the control means 4 as data 703.

The control means 4 passes the prediction result to the output means 3 as data 705, and the output means 3 outputs it for display.

FIG. 8 is a flowchart showing the operation of this embodiment.

When a parameter is input from the input means (801), the control means 4 checks whether data matching the parameter exists in the data storage means 5.

for example.

Input parameters Voltage = 16v Size = 8 Foil material = C3 Electrolytic paper = M1 Electrolyte = 21 The data format of the data storage means is comprised of the following 10 data as 5 parameters and characteristic values. shall be.

Sample 1: <8.310 03 old Zl, 8) Sample 2
: <Ilt, 3 + [i C3M2 Zl, 1G
>Sample 3: <16 12.5 C3M2 Zl,
8) Sample 4: <IG 16 C3M2Z
1,8) Sample 5: <35 8 02 old Z1. ．． 3.
4>Sample 8: <35 12.502 M2 Zl
, 3.3>Sample 7: <35 111i C2
M2Z1.3) Sample 8: <50 8 02
MI Zl, 3.3>Sample 9: <50 12.5
02 M2 Zl, 3.2>Sample 10: <50 1
6 02M2Z1, 2.8) Also, the weighting coefficient of each element of the parameter is as follows.

Voltage Size Foil Electrolytic Paper Electrolyte Weight 1122 2 does not exist in the data storage means 5, so proceed to the next process.

When the prediction device is activated for the first time, the prediction model does not yet exist, so the automatic prediction model creation means is activated. The automatic prediction model creation means uses the cluster analysis means and the weighting coefficients to classify the data stored in the data storage means (802).

For example, assuming that the dissimilarity used for cluster analysis is the addition of weighting coefficients corresponding to types with different parameter values, the dissimilarity calculation for sample 1 and sample 2 is as follows.

Weight <1 1 2 2 2> Sample 1: <8.3 10 03 MI Zl, 8
) Sample 2: <8.3 111i C3M2
Because Zl, IG>.

Dissimilarity = 1 (weight of size) + 2 (weight of electrolytic paper) = 3 In this case, the parameter that matches the input parameter The result is:

1st class: Sample 1: <Ilf, 310 03 MI Zl,
8 >Second class: Sample 2: <[i, 3111i C3M2 Zl,
1G > Sample 3: <IG 12.503 M2
Zl, 8) Sample 4: <l1li 1G
03M2Z1.8 >Third class: Sample 5: <35 Sample 8: <50 02 MI Zl, 3.4) 02 MI Zl, 3.3> Fourth class: Sample 6: <35 Sample 7: <35 Sample 9 : <50 Sample to: <50 12.5 IG 12.5 M2ZI.

M2 Zl.

M2 Zl.

M2 Zl.

3.3>3>3.2>2.8> A predictive model is created for each clustered class (803). However, if the amount of data is small as in this example, it may not be possible to create class models for all classes.

Next, prediction execution is performed using this prediction model (
804). The algorithm of the prediction execution means calculates the class to which the plurality of input parameter values belong by the above-described dissimilarity calculation, and calculates the predicted value using the prediction model of the approximate class. If a predictive model of an approximate class does not exist, a predictive model of the next closest class is used. The prediction result evaluation means evaluates the predicted value calculated by the prediction execution means using a test algorithm or the user's empirical evaluation criteria (805).

If the evaluation result is good, the prediction result is displayed and output by the output means and the process ends (807). However, if the evaluation result of the predicted value is bad, the weighting coefficient is dynamically adjusted by the supervised learning method using the learning algorithm of the Berceptron model. Make changes (8
0B).

Although not specified in this embodiment, the weighting coefficients also change when the degree of dissimilarity between the input multiple types of parameter values and the members belonging to the approximate class is greater than the evaluation criterion (=2) described above. It is also possible to activate the target change means to improve the accuracy of the model. Next, the 1 weighting coefficient changing means activates the predictive model automatic creation means again, and performs clustering based on the contents of the changed weighting coefficients (802).

A new prediction model is created (803)L, prediction is executed using the model (804), prediction evaluation is performed (805), and the results are displayed and output.

The prediction device according to the present invention has a learning mechanism in which the prediction model becomes more precise as the prediction data increases, and the prediction model is adjusted according to the state of the prediction result. Even if you do not do this, the prediction model becomes more precise as the predictions are executed, making it possible to make fresh predictions forever.

Effects of the Invention The present invention is a method that can be widely used in addition to conventional statistical methods to elucidate models and rules, and it is difficult to elucidate models because conventional methods could only handle linear models with independent parameters. Fields where there is no other way than to obtain characteristic values through experiments or to obtain rough characteristic values based on the intuition of experts 9
For example, it increases the possibility of model elucidation when there is no linear relationship between parameters that often appear in capacitor life prediction models or chemical/physical laws. In addition, since it is not a black box model like a neural network model, it can be effectively used for mechanism analysis, making it possible to measure lifespan without trial experiments, increasing the reliability of predictive models, and designing new materials and products. It can be used for a wide range of purposes, including demand forecasting.

In addition, since the flow of the prediction function is clearly separated into the prediction execution means, the automatic prediction model creation means, and the prediction model storage means, predictions can be made as the data stored in the data storage means or the weighting coefficients are changed. There is no need to start the automatic prediction model creation means unless it is necessary to change the model, which not only improves the efficiency of prediction execution processing, but also makes it easier for users to modify, process, add, etc. to the prediction model. This is possible.

In addition, the prediction function includes not only a predictive model automatic creation means but also an interactive processing means, so various models can be automatically constructed, such as models based on multiple regression analysis, models based on function approximation, neural network models, etc. By modifying and processing these algorithmic models, or by using this algorithmic model as part of a user-defined model, the user's empirical rules, knowledge, and partially elucidated Because it is possible to construct a flexible model that reflects the laws and regulations, it can be used not only as a simple prediction device but also as a support device for model elucidation.

The prediction device according to the present invention not only refines the prediction model as the amount of prediction data increases, but also has a learning mechanism that adjusts the prediction model according to the state of the prediction result even if the amount of prediction data does not increase. Therefore, it is possible to refine the prediction model each time the prediction is executed.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a capacitor life prediction device according to an embodiment of the present invention, FIG. 2 is a block diagram of a conventional capacitor life prediction device, and FIG. 3 is an operation of a capacitor life prediction device according to an embodiment of the present invention. FIG. 4 is a flowchart showing the operation of a conventional capacitor life prediction device.
FIG. 5 is a block diagram of a capacitor life prediction device according to another embodiment of the present invention, and FIG. 6 is a block diagram of a capacitor life prediction device according to yet another embodiment of the present invention. FIG. 7 is a block diagram of a capacitor life prediction device in another embodiment of the present invention, FIG. 8 is a flowchart showing the operation of the capacitor life prediction device in the same embodiment, and FIG. 9 explains the mechanism of the function approximation method. This is an auxiliary figure for. DESCRIPTION OF SYMBOLS 1... Prediction device main body, 2... Input means, 3... Output means, 4... Control means, 5... Data storage means,
θ...Weighting coefficient storage means, 7... Approximate data retrieval means. 8... Difference parameter detection means, 9... Reference data search means, 10... Function approximation equation identification means, 11... Intermediate predicted value calculation means, 12... Multiple regression model identification means. 13... Prediction value calculation means, 14... Prediction execution means. 15... Predictive model automatic creation means, 16... Predictive model storage means, 17... Dialogue processing means, 18...
Cluster analysis result storage means, 19... Cluster analysis means, 20... Prediction result evaluation means, 21... Weighting coefficient dynamic change means, 22... Parameter analysis means. 23...Predicted value calculation control means. Name of agent: Patent attorney Shigetaka Awano and one other person

Claims (6)

[Claims] (1) Consists of an input means for inputting values of multiple types of parameters necessary for predicting characteristic values, values of multiple types of parameters obtained through experiments or observations, and characteristic values corresponding to each group of parameter values. data storage means for storing a plurality of data sets; weighting coefficient storage means for storing weighting coefficients indicating the importance of each of the plurality of types of parameters; and input parameter values while being weighted by the weighting coefficients. approximate data retrieval means for comparing the parameter value group and the parameter value group stored in the data storage means and searching for a parameter value group closest to the input parameter value group; and an approximate parameter value group retrieved by the approximate data retrieval means. reference data retrieval means for retrieving data necessary for determining the relationship between the parameter and the characteristic value by a function approximation method for a parameter that is different from the input parameter value, and the reference data retrieval means; a function approximation formula identifying means for determining a function approximation formula using the data retrieved by the means; meson measurement value calculation means for calculating a meson measurement value from the characteristic value and the value of the input parameter; A prediction device comprising predicted value calculation control means for controlling the meson measurement value calculation means to sequentially and repeatedly perform calculations for each parameter. (2) In claim 1, the prediction execution means is constituted by the approximate data search means, the reference data search means, the meson measurement value calculation means, and the predicted value calculation control means, and the function approximation formula identification means is constituted by the prediction model storage means and the prediction model. A prediction device characterized by comprising an automatic creation means. (3) The prediction device according to claim 2, further comprising an interaction processing means for setting data and rules stored in the prediction model storage means and the data storage means. (4) A step of inputting values of multiple types of parameters necessary for predicting characteristic values, and data consisting of values of multiple types of parameters obtained through experiments or observations and characteristic values corresponding to each group of parameter values. a data storage step of storing a plurality of sets; a weighting coefficient storing step of storing a weighting coefficient indicating the importance of each of the plurality of types of parameters; and storing input parameter value groups and the data while weighting by the weighting coefficients. an approximate data retrieval step that searches for a parameter value group closest to the input parameter value group by comparing the parameter value groups stored in the step; and an input parameter of the approximate parameter value group retrieved by the approximate data retrieval step. a reference data retrieval step for retrieving data necessary for determining the relationship between the parameter and the characteristic value using a function approximation method for a parameter having a difference in value from the data storage step; a function approximation formula identification step for determining a function approximation formula using data; a function approximation formula identified in the function approximation formula identification step; a value of the parameter; and a characteristic value corresponding to the value of the parameter; an intermediate predicted value calculation step of calculating an intermediate predicted value based on the value of the input parameter, and the intermediate predicted value calculated by the intermediate predicted value calculation step is used as a new characteristic value for each parameter that is different from the input parameter value. A prediction method comprising a predicted value calculation step of sequentially repeating the intermediate predicted value calculation step. (5) Input means for inputting the values of multiple types of parameters necessary for predicting characteristic values, control means for controlling the entire device, and values of multiple types of parameters obtained through experiments or observations and their parameter values. data storage means for storing a plurality of data sets consisting of characteristic values corresponding to each group; weighting coefficient storage means for storing weighting coefficients indicating the importance of each of the plurality of types of parameters; Prediction comprising automatic prediction model creation means for automatically creating a prediction model by referring to weighting coefficients and data stored in the data storage means, and a prediction model created by the automatic prediction model creation means. a model storage means; a prediction execution means for executing a prediction by referring to the data, the prediction model, and the weighting coefficient; and evaluation of a predicted value predicted by the prediction execution means and movement of the prediction model and the weighting coefficient. The method comprises a learning means for changing the target, and an output means for displaying and outputting the prediction result calculated by the prediction execution means, the parameters input from the input means are sent to the control means, and the control means: It is checked whether an appropriate prediction model exists in the prediction model storage means, and if it does not exist, the automatic prediction model creation means is activated, and the automatic prediction model creation means refers to the data and the weighting coefficients. to create a prediction model, the results are stored in the prediction model storage means, and the control means:
Control is passed to the prediction execution means, the prediction execution means executes prediction using the prediction model, and if an appropriate prediction model already exists in the prediction model storage means, the process of the automatic prediction model creation means is omitted. , the prediction execution means is directly activated to execute prediction using an already existing prediction model, the prediction execution means returns a prediction result to the control means, and the control means controls the learning means. The prediction result is passed, the learning means evaluates the predicted value, and if the evaluation result is bad, the weighting coefficient is dynamically changed by a learning algorithm, and the control means again passes control to the prediction model automatic creation means. , the automatic prediction model creation means constructs the prediction model again by referring to the weighting coefficients changed by the learning algorithm, and the prediction execution means has a learning function to perform prediction execution by referring to the updated prediction model. A prediction device characterized by: (6) In claim 5, the predictive model storage means comprises a predictive model storage means for storing the predictive model and the cluster analysis result storing means for distributing the predictive model for each class, and the learning means comprises the predictive result evaluation means and the weighting means. A prediction device comprising coefficient dynamic changing means.