JP2020177542A - State variation detection device and program for state variation detection - Google Patents

Abstract

To realize verification and correction performed on prediction that is made by state variation detection using a prediction model created based on teaching data, or change of a result.SOLUTION: A state variation detection device comprises: prediction model creation means 31 that creates a prediction model based on teaching data; predetermined range importance explanatory variable searching means 32 that determines, in the prediction model, a predetermined range importance explanatory variable that is an explanatory variable with lower importance than that of a preset threshold; teaching data geometric distance calculation means 33 that determines the geometric distance of the teaching data by using the predetermined range importance explanatory variable; analysis object data geometric distance calculation means 34 that determines the geometric distance of analysis object data; and state variation detection means 35 that detects a variation in state of the analysis data based on the distribution of the teaching data geometric distance and the distribution of the analysis object data geometric distance.SELECTED DRAWING: Figure 4

Description

The present invention relates to a state change detection device and a state change detection program.

For example, when a state change such as an abnormality is detected by measuring the current, voltage, and vibration of a machine, the abnormality is generally determined based on the past experience of a person or the like. That is, the threshold value of the state change is set by a person, and it is almost impossible for a machine having no abnormal data to perform the state change.

Patent Document 1 discloses a system for predicting the arrival status of the radio wave when an abnormal propagation of the radio wave occurs. This system uses the meteorological information importing means for capturing meteorological information and the meteorological information captured by this meteorological information capturing means to at least information on temperature at each altitude, atmospheric pressure at each altitude, and humidity at each altitude. It is provided with a meteorological data extraction means for extracting one as meteorological data. Further, regarding the arrival of the radio wave on the day when the weather data is extracted by the weather data extraction means, a prediction function for predicting the abnormal propagation of the radio wave calculated based on the measurement data accumulated in advance is accumulated. The meteorological data extracted by the meteorological data extracting means is input to the accumulated prediction function to determine the predicted value which is the occurrence probability of the abnormal propagation of the radio wave.

Patent Document 2 describes an acquisition process for acquiring time-series data of a sensor possessed by the machine-equipment for each predetermined production unit of the product from a machine-equipment capable of producing the product, and a plurality of divisions of the time-series data. The division process for generating the divided data group divided into data, the first additional information which is additional information indicating that there is no failure is associated with the predetermined normal quality portion included in the divided data group, and the divided data group A quality abnormality detection method including a setting process for unconditionally associating a second additional information indicating that there is a failure with respect to the latest production amount included in the above is disclosed.

This quality abnormality detection method further corresponds to the input data by a grouping step of forming two groups in which the normal quality component and the latest production component are mixed, and learning using one of the two groups. In the generation step of generating the prediction model for predicting the additional information, in the prediction result of the prediction model obtained by using the other of the two groups as the input data, the additional information is the second additional information. It has a calculation process for calculating the predicted probability, and determines the quality abnormality of the product based on the above probability.

Further, Patent Document 3 discloses a data prediction method for predicting the future based on past achievements. In this data prediction method, the prediction model includes a first prediction model that outputs a prediction value for at least one feature quantity on the prediction target date, and a prediction value output from this first prediction model as input factors. It is composed of a second prediction model that outputs a predicted value for each predetermined time on the prediction target day.

Then, in the data prediction method of Patent Document 3, a prediction model construction means for constructing a prediction model using the collected closest actual data and past actual data, and a prediction input data for input to the constructed prediction model for prediction. The prediction execution means for obtaining the prediction value, the prediction error calculation means for calculating the prediction error or the model error from the collected nearest actual data and the prediction value, and the prediction value based on the prediction error or the model error. The correction coefficient or correction amount to be corrected is calculated, and the axial correction predicted value is obtained.

On the other hand, in a random forest, a state change detection device that indicates a normal case by "1" and an abnormal case by "2" by the boundary line of the region as shown in FIG. 1 (A) is known. Has been done. That is, in this state fluctuation detection device, the region where the value in the vertical axis direction is 160 or more and the value in the horizontal axis direction is smaller than 45 is all normal "1", and in the region where the value in the vertical axis is smaller than 160, the horizontal axis If the value in the direction is smaller than 30, it is a normal "1", and if the value in the horizontal axis direction is 30 or more, it is an abnormal "2". Further, the region where the value in the vertical axis direction is 160 or more and the value in the horizontal axis direction is larger than 45 is abnormal "2" when the value in the vertical axis direction is smaller than 260, and the value in the vertical axis direction is 260 or more. If there is, it becomes a normal "1". In FIG. 1A, the plain white region is a normal “1” and the satin-patterned region is an abnormal “2”.

When the above-mentioned areas are divided into abnormal and normal judgments, it is possible to realize a state change detection device by machine learning that makes predictions by branching by a classifier (for example, in a tree structure). In such a state change detection device, for example, when the analysis target data D as shown in FIG. 1 (B) is generated, according to the state change detection device for the region of FIG. 1 (A), the vertical direction It corresponds to a region where the value is 160 or more and the value in the horizontal axis direction is smaller than 45, and is classified as a normal "1" as shown in the figure. However, since the value in the vertical axis direction is approximately 10,000, it is far from the normal vertical axis value 400 and the like as shown by the arrow, and it is clearly anomalous "2" for analysis target data. Yes, there is a misclassification.

Further, in the random forest state change detection device using the decision tree-structured classifier as shown in FIG. 2, branching is performed at the points indicated by circles. In FIG. 2, the explanatory variables used at the branch are shown as A, B, C, D, and E. For example, the explanatory variable A is the age, the explanatory variable B is the body weight, the explanatory variable C is the height, the explanatory variable D is “sex”, and so on. In such a state change detector, less important explanatory variables do not affect the results. Specifically, if the device results in health and unhealth, it can be said that the "sex" of the explanatory variable D is unlikely to affect the result. In all the explanatory variables A, B, C, D, and E, the explanatory variables that do not affect the result are set as the explanatory variables of low importance, and the result does not change even if the value of the explanatory variable fluctuates greatly. Is. In this case, the values are randomly exchanged for each of the explanatory variables A, B, C, D, and E to calculate how much error occurs and determine the importance.

It was found that some examples shown in FIG. 1 may not be properly classified, and also in the example shown in FIG. 2, explanatory variables with low weight have a low effect on the results. Although the probability of an error is low, the accuracy of the entire determination device may be reduced.

In Patent Document 4, a two-dimensional plane is considered as a space for acquiring signals, and the data used for machine learning and identification is multidimensional information having a three-dimensional structure in total of distribution information and spectral information in the plane. The procedure for machine learning and identification when using this multidimensional information is essentially the same as when using spectral data. Further, in this case, it is possible to acquire a plurality of feature quantities suitable for describing a data pattern and then use them as feature vectors for machine learning and identification processing, as a typical example of feature quantities. Is stated to have volume, curvature, spatial gradient, HLAC (higher-order local autocorrelation), and the like.

Then, it is also possible to select the feature amount used for machine learning in advance. In this case, it is stated that, for example, the Mahalanobis distance may be calculated and the feature amount used for identification may be selected. The larger the Mahalanobis distance, the easier it is to identify, so it is possible to preferentially select features that increase the Mahalanobis distance between each group of interest.

That is, what is described in Patent Document 4 states that when selecting a feature amount that facilitates machine learning and identification processing, a feature amount that increases the Mahalanobis distance is selected.

Further, in column 0117 of Patent Document 5, "When a decision tree is used as the above-mentioned predetermined algorithm, the degree of relevance (importance or contribution) of each feature amount (explanatory variable) is determined by a random forest method or the like. (It may be called) can be calculated. Specifically, when creating a decision tree, each feature amount (explanatory variable) so as to reduce the impureness indicated by the Gini coefficient, cross entropy, etc. Therefore, the node (branch) is selected from the above. Therefore, the amount of decrease in purity when each feature amount (explanatory variable) is selected as the node can be used as the importance of the feature amount (explanatory variable). In step S102, the importance may be used to select a feature amount effective for predicting the occurrence of an abnormality from a plurality of types of feature amounts. " Shows selection by importance as described with reference to FIG.

Japanese Unexamined Patent Publication No. 2005-315753 JP-A-2018-147406 Japanese Unexamined Patent Publication No. 2004-94437 Japanese Unexamined Patent Publication No. 2015-52581 Japanese Unexamined Patent Publication No. 2018-116545

The present invention can be used for verification, correction, change of result, etc. performed on the prediction performed in the state fluctuation detection using the prediction model created based on the teacher data, and the degree of misclassification in the above prediction model can be determined. Provided are a state change detection device and a state change detection program which can be reduced and can cover the decrease in accuracy.

The state change detection device according to the present embodiment is a prediction model creating means for creating a prediction model for state change detection based on teacher data, and a predetermined range importance explanatory variable which is an explanatory variable of importance lower than a preset threshold. The predetermined range importance explanatory variable search means for obtaining the above in the prediction model, the teacher data geometric distance calculation means for obtaining the geometric distance of the teacher data using the predetermined range importance explanatory variable, and the geometry of the teacher data. The analysis target data geometric distance calculation means for obtaining the geometric distance of the analysis target data using the average value of each explanatory variable used for calculating the scientific distance and the inverse matrix of the dispersion covariance matrix, and the teacher data geometry. It is characterized by comprising a state change detecting means for detecting a state change of the analysis data based on the distribution of the scientific distance and the distribution of the geometric distance of the data to be analyzed.

Explanatory drawing for showing the method when state change detection is performed by a random forest. The figure which shows an example of the decision tree used when the state change detection is performed by a random forest. The block diagram of the computer system which realizes the state change detection apparatus which concerns on embodiment of this invention. The functional block diagram of the state change detection apparatus which concerns on embodiment of this invention. The figure which shows an example of the teacher data used for the state change detection apparatus which concerns on embodiment of this invention. The Mahalanobis distance of the teacher data shown in FIG. 5 and the figure showing this by a histogram. The Mahalanobis distance of the data to be measured and the figure showing this by a histogram. The flowchart which shows the operation of the state change detection apparatus which concerns on embodiment of this invention. The figure which shows the importance of the explanatory variable obtained by the operation of the state change detection apparatus which concerns on embodiment of this invention, the ratio of the importance, and an example of the cumulative value thereof. This is a diagram in which the distribution map of the Mahalanobis distance of the teacher data and the distribution map AC of the Mahalanobis distance of the analysis target data are arranged in the vertical direction with the horizontal axes aligned with each other, and is a state change detection device according to the embodiment of the present invention. Explanatory drawing for explaining the method of state change detection by.

Hereinafter, embodiments of the state change detection device and the state change detection program according to the present invention will be described with reference to the accompanying drawings. In each figure, the same components are designated by the same reference numerals and duplicate description will be omitted. The state change detection device according to the embodiment of the present invention can be configured by, for example, a personal computer, a workstation, or other computer system as shown in FIG. This computer system operates as a state change detection device by controlling each part based on a program or data stored in the main memory 11 or read into the main memory 11 by the CPU 10 and executing necessary processing. is there.

An external storage interface 13, an input interface 14, a display interface 15, and a data input interface 16 are connected to the CPU 10 via a bus 12. A program such as a state change detection program and an external storage device 23 in which necessary data and the like are stored are connected to the external storage interface 13. An input device 24 such as a keyboard as an input device for inputting commands and data and a mouse 22 as a pointing device are connected to the input interface 14.

A display device 25 having a display screen such as an LED or an LCD is connected to the display interface 15. Sensors 26-1, 26-2, ..., 26-m for obtaining measurement data are connected to the data input interface 16. This computer system may be provided with other configurations, and the configuration of FIG. 3 is only an example.

In the above, in the CPU 10, each means and the like shown in FIG. 4 are realized by the state change detection program in the external storage device 23. That is, the prediction model creating means 31, the predetermined range importance explanatory variable search means 32, the teacher data geometric distance calculating means 33, the analysis target data geometric distance calculating means 34, and the state change detecting means 35 are realized. Further, the teacher data T is stored in the external storage device 23. The prediction model creating means 31 creates the prediction model 30 using the teacher data.

The prediction model 30 predicts the objective variable from the explanatory variables by machine learning. Here, as a machine learning algorithm, a random forest of pattern mining can be mentioned, but in addition to this, branching is performed by a classifier (for example, in a tree structure) such as a classification tree or a regression tree. It is possible to adopt algorithm by machine learning that makes predictions.

The predetermined range importance explanatory variable search means 32 obtains a predetermined range importance explanatory variable which is an explanatory variable of importance lower than the threshold value set in advance in the prediction model 30. The predetermined range importance explanatory variable search means 32 obtains the importance for each acquisition source of the explanatory variable of the teacher data, and determines the explanatory variable of the acquisition source having the importance of a value lower than the predetermined threshold value in the predetermined range importance. Obtained as an explanatory variable. Here, the acquisition source is referred to as an acquisition source when explanatory variables are acquired / collected using several sensors. Further, when the explanatory variables are acquired / collected using several devices, each device is referred to as an acquisition source. In other words, the acquisition source is used in the sense that it is the source for acquiring the objective variable and the explanatory variable.

For example, as shown in FIG. 5, the teacher data includes the data of the objective variable to be obtained by the sensor A, the data of the first explanatory variable obtained by the sensor B, and the second explanation obtained by the sensor C. From the variable data, the data of the third explanatory variable obtained by the sensor D, the data of the fourth explanatory variable obtained by the sensor E, and the data of the fifth explanatory variable obtained by the sensor F. It is composed. The number of data acquisitions (the number of data in the vertical direction in FIG. 5) is arbitrary. In the present embodiment, it is assumed that the acquisition sources of the explanatory variables having the importance of a value lower than the predetermined threshold value are the sensor E and the sensor F, and the predetermined range importance explanatory variables are the sensor E and the sensor of FIG. It becomes the value described in the column of F.

The teacher data geometric distance calculation means 33 obtains the geometric distance of the teacher data by using the predetermined range importance explanatory variable obtained above. The Mahalanobis distance can be used as the geometric distance of the teacher data.

The analysis target data geometric distance calculation means 34 uses the mean value of each explanatory variable used for calculating the geometric distance of the teacher data and the inverse matrix of the variance-covariance matrix to analyze the geometric distance of the analysis target data. Is what you want. The Mahalanobis distance can be used as the geometric distance of the data to be analyzed.

As a result of using the mean value of each explanatory variable used to calculate the geometric distance of the teacher data and the inverse of the variance-covariance matrix when calculating the geometric distance of the data to be analyzed as described above. The geometric distance from the data mean of the teacher data to the data to be analyzed can be obtained.

The state change detecting means 35 detects the state change of the analysis data based on the distribution of the teacher data geometric distance and the distribution of the analysis target data geometric distance. Specifically, it is assumed that the Mahalanobis distance (MD) of the teacher data is obtained as shown in FIG. 6 (A), and when this is made into a histogram, it becomes as shown in FIG. 6 (B). Further, as shown in FIG. 7 (A), the Mahalanobis distance (MD) of the data to be analyzed is obtained, and when this is made into a histogram, it is assumed that the result is as shown in FIG. 7 (B). The analysis data is data collected by the sensors B to F, which are the acquisition sources of the explanatory variables of the teacher data shown in FIG. 5, as targets for state fluctuation detection, and are not shown, but are values as shown in FIG. (Of course, since it is an actually measured value, it is not usually the same as that in FIG. 5).

In the distribution of Mahalanobis distance (MD) in the teacher data, the maximum value is 4.00, and no larger value appears. Therefore, it can be concluded that the analysis target data in which the Mahalanobis distance (MD) larger than the maximum value of the Mahalanobis distance (MD) of the teacher data appears has a state change. Then, it can be concluded that the degree of state variation is large when the Mahalanobis distance (MD) value is held at a position where the distance from the maximum Mahalanobis distance (MD) of the teacher data is large. Therefore, if the distribution range of the Mahalanobis distance (MD) of the teacher data is normal, it can be concluded that the degree of abnormality is large at the position where the distance from the maximum value of the Mahalanobis distance (MD) of the teacher data is large.

That is, the state change detecting means 35 obtains the distribution range of the teacher data geometric distance, and detects the state change of the analysis data based on the ratio of the distribution of the analysis target data geometric distance beyond this distribution range. can do. In the above operation, the state fluctuation detecting means 35 obtains the distribution range of the teacher data geometric distance, and the analysis data is based on the degree to which the distribution of the analysis target data geometric distance is separated from this distribution range. It shows that the state change can be detected.

Since the state change detection device configured by the above means and the like executes the processing operation according to the flowchart shown in FIG. 8, the operation will be described with reference to this flowchart.

First, the prediction model 30 is created using the teacher data (STEP 1). The teacher data is the data already illustrated, and is composed of the objective variable data to be obtained by the sensor A and the explanatory variable data obtained from the sensor B by the sensor F. Thus, the prediction model 30 is created by the prediction model creating means 31 using the teacher data T of FIG. The prediction model 30 created here assumes that prediction is performed by a random forest, but if the prediction is performed by machine learning by branching by a classifier (for example, in a tree structure), it is not a random forest. It may use the method of.

Next, the importance of the explanatory variables in the prediction model 30 created above is acquired (STEP2). Since the acquisition sources of the explanatory variables are sensors B to F, the importance value is obtained for each acquisition source. As the method of determining the importance, a known method such as a method used for obtaining an appropriate decision tree when creating a prediction model of a random forest can be used.

The value of importance for each acquisition source obtained in STEP 2 above is shown in FIG. 9 (A). In the present embodiment, following STEP2, a predetermined range importance explanatory variable belonging to the importance of a preset range is obtained in the prediction model (STEP3). For this purpose, FIG. 9B shows a ratio obtained from each acquisition source based on the importance value, and the ratio value is accumulated from the side with the lower importance value to obtain the cumulative value. In the present embodiment, a cumulative value of "1.0" is adopted as the threshold value, and an explanatory variable of the acquisition source in a range smaller than this "1.0" is obtained. The cumulative value "1.0" as the threshold value is only an example, and is appropriately changed depending on the system to be used and the situation in which it is used. In the present embodiment, from FIG. 9B, the acquisition source is the explanatory variable corresponding to the explanatory variables of the sensor E and the sensor F.

Next to STEP3, the Mahalanobis distance, which is the geometric distance of the teacher data, is obtained using the predetermined range importance explanatory variable obtained above (STEP4). Here, the target for obtaining the Mahalanobis distance is the explanatory variables excluding the objective variable in the teacher data. The Mahalanobis distance for the teacher data obtained above is shown as MD in FIG. 6 (A), and when this is used as a frequency distribution graph, it is as shown in FIG. 6 (B).

Following STEP4, the Mahalanobis distance, which is the geometric distance of the data to be analyzed, is obtained (STEP5). In this case, the Mahalanobis distance of the data to be analyzed is obtained by using the mean value of each explanatory variable used for calculating the geometric distance of the teacher data and the inverse matrix of the variance-covariance matrix. As a result, the geometric distance from the data average of the teacher data to the analysis target data can be obtained.

Although the data to be analyzed is not shown as described above, it is assumed that the values are based on the values shown in FIG. On the other hand, the Mahalanobis distance for the data to be analyzed is shown as MD in FIG. 7 (A), and when this is used as a frequency distribution graph, it is as shown in FIG. 7 (B). The Mahalanobis distance and its distribution for the data to be analyzed are obtained for each acquisition source.

Following STEP 5, the state change of the analysis target data is detected based on the distribution of the Mahalanobis distance which is the teacher data geometric distance and the distribution of the Mahalanobis distance which is the analysis target data geometric distance (STEP 6). The distribution of the Mahalanobis distance, which is the above-mentioned teacher data geometric distance, is as shown in FIG. 6 (B). The distribution of the Mahalanobis distance, which is the geometric distance of the data to be analyzed, is as shown in FIG. 7 (B).

The horizontal axis is aligned with the distribution TC of the Mahalanobis distance of the teacher data shown in FIG. 6 (B) and the distribution AC of the Mahalanobis distance of the data to be analyzed shown in FIG. 7 (B) in the vertical direction. When arranged side by side, it is as shown in FIG. As is clear from drawing the vertical line segment L passing through the maximum Mahalanobis distance T-MAX of the teacher data, the Mahalanobis distance of the data to be analyzed is in the region of a value larger than the maximum Mahalanobis distance T-MAX of the teacher data. Also exists.

Assuming that the distribution TC of the Mahalanobis distance of the teacher data is in the range of no state change (or normal), the area of the distribution of the Mahalanobis distance of the data to be analyzed that exceeds the vertical line segment L (the area to the right of L in the figure). ) Is said to have state fluctuations (or abnormalities).

The state change (or abnormality) depends on whether or not the difference between the maximum Mahalanobis distance T-MAX of the teacher data and the maximum Mahalanobis distance A-MAX of the analysis target data is equal to or greater than a predetermined threshold value. Can be determined. In addition, there is a state change based on whether or not the ratio (also called the abnormality rate) of the region exceeding the vertically extending line segment L (the region on the right side of the line segment L in the figure) to the entire region exceeds a predetermined threshold value (it can be said to be an abnormality rate). Alternatively, it is possible to determine (abnormal occurrence) or the like. In this way, the state change detecting means 35 obtains the distribution range of the teacher data geometric distance, and the state change of the analysis data is based on the ratio of the distribution of the analysis target data geometric distance beyond this distribution range. Is detected.

Further, the state is based on the difference between the average value of the Mahalanobis distance distribution TC of the teacher data and the average value of the Mahalanobis distance distribution AC of the analysis target data, and whether or not the ratio of the above two averages exceeds a predetermined threshold value. It can be determined that there is a fluctuation (or an abnormality has occurred). Since the above judgment is performed for each acquisition source, it is determined that the data to be analyzed has a state change (or an abnormality has occurred) by performing statistical processing such as taking a majority vote of the judgment results obtained for each acquisition source. be able to.

In the above, the alternative judgment of whether there is a state change (or abnormality) or no state change (or normal) is made, but some threshold values are set and the state change (or abnormality) is performed. It is also possible to judge the possibility of the above in three stages of large, medium and small, and in multiple stages with more stages such as large possibility, large possibility, small possibility, and small possibility.

In the above embodiment, since the predetermined range importance explanatory variable search means 32 indicates that the explanatory variable of importance lower than the preset threshold value is obtained, the explanatory variable of low importance such as a random forest is used. It is expected that a conclusion different from that of the state change detection device that is not used can be obtained, and the result of this embodiment can be used to correct the result of a random forest or the like, and the accuracy of state change detection can be improved. it can. For example, the threshold value for the determination of the present embodiment is set to be high so that it is unlikely that there is a state change (or abnormality), and the determination result that there is a state change (or an abnormality) can be obtained in the present embodiment. If so, a method such as giving priority to the judgment of the present embodiment can be adopted. The state change detection device by random forest, which is a device for obtaining the result to be corrected, is the device described in Japanese Patent Application No. 2018-40531 filed by the applicant of the present application, the device described in Japanese Patent Application No. 2019-55609, and the present invention. The apparatus according to the application 2019-55615 can be mentioned.

Further, in the above embodiment, the predetermined range importance explanatory variable search means 32 seeks an explanatory variable having an importance lower than a preset threshold value, but it is used as a confirmation of the result by a random forest or the like. You can also. In that case, the predetermined range importance explanatory variable search means 32 obtains the explanatory variables excluding the explanatory variables of importance lower than the preset threshold value, and may obtain the geometric distance using the explanatory variables. good. Alternatively, the predetermined range importance explanatory variable search means 32 obtains an explanatory variable of importance between a second threshold value lower than a preset first threshold value and larger than the first threshold value (that is, importance level). (Find the explanatory variable in the middle of), and use this to find the geometric distance.

Further, in the above embodiment, as a method for obtaining the geometrical distance, a method called the well-known MT method for obtaining the Mahalanobis distance is used. However, recently, in order to solve the problems of the MT method, the MTA method, the T (3) method (RT method), and the like have been proposed. In this embodiment, the above-proposed MTA method and T (3) method (RT method) may be adopted.

10 CPU
11 Main memory 12 Bus 13 External storage interface 14 Input interface 15 Display interface 16 Data input interface 22 Mouse 23 External storage device 24 Input device 25 Display device 26-1 to 26-m Sensor 30 Prediction model 31 Prediction model creation means 32 Predetermined range Importance explanatory variable search means 33 Teacher data geometric distance calculation means 34 Analysis target data Geometric distance calculation means 35 State fluctuation detection means

Claims (14)

A predictive model creation method that creates a predictive model that predicts the objective variable from the explanatory variables based on the teacher data,
A predetermined range importance explanatory variable search means for obtaining a predetermined range importance explanatory variable belonging to a preset range importance in the prediction model, and
A teacher data geometric distance calculation means for obtaining the geometric distance of the teacher data using the predetermined range importance explanatory variable, and
An analysis target data geometric distance calculation means for obtaining the geometric distance of the analysis target data using the mean value of each explanatory variable used for calculating the geometric distance of the teacher data and the inverse matrix of the variance-covariance matrix. ,
A state change detecting apparatus comprising: a state change detecting means for detecting a state change of the analysis data based on the distribution of the teacher data geometric distance and the distribution of the analysis target data geometric distance. The state change detecting means obtains the distribution range of the teacher data geometric distance, and detects the state change of the analysis data based on the ratio of the distribution of the analysis target data geometric distance beyond this distribution range. The state change detection device according to claim 1. The state change detecting means obtains the distribution range of the teacher data geometric distance, and detects the state change of the analysis data based on the degree to which the distribution of the analysis target data geometric distance deviates from this distribution range. The state change detection device according to claim 1 or 2, wherein the state change detection device. The state change detection device according to any one of claims 1 to 3, wherein the prediction model created by the prediction model creation means is for performing prediction by machine learning by branching with a classifier. .. The state change detection device according to any one of claims 1 to 4, wherein the prediction model created by the prediction model creation means predicts by a random forest. The teacher data geometric distance calculation means obtains the Mahalanobis distance of the teacher data as the geometric distance of the teacher data.
The state change detection according to any one of claims 1 to 5, wherein the analysis target data geometric distance calculation means obtains the Mahalanobis distance of the analysis target data as the geometric distance of the analysis target data. apparatus. The state change detection device according to any one of claims 1 to 6, wherein the predetermined range importance explanatory variable search means obtains an explanatory variable having an importance lower than a preset threshold value. Computer,
Predictive model creation means that creates a predictive model that predicts the objective variable from the explanatory variables based on the teacher data.
A predetermined range importance explanatory variable search means for obtaining a predetermined range importance explanatory variable, which is an explanatory variable of importance lower than a preset threshold value, in the prediction model.
Teacher data geometric distance calculation means for obtaining the geometric distance of teacher data using the predetermined range importance explanatory variable,
An analysis target data geometric distance calculation means for obtaining the geometric distance of the analysis target data using the mean value of each explanatory variable used for calculating the geometric distance of the teacher data and the inverse matrix of the variance-covariance matrix.
A state change detection program characterized by functioning as a state change detection means for detecting a state change of the analysis data based on the distribution of the teacher data geometric distance and the distribution of the analysis target data geometric distance. Using the computer as the state change detecting means, the distribution range of the teacher data geometric distance is obtained, and the state change of the analysis data is determined based on the ratio of the distribution of the analysis target data geometric distance beyond this distribution range. The state change detection program according to claim 8, wherein the program is made to function to detect. As the state change detecting means, the computer obtains a distribution range of the teacher data geometric distance, and determines the state change of the analysis data based on the degree to which the distribution of the analysis target data geometric distance is separated from this distribution range. The program for detecting state fluctuations according to claim 8 or 9, wherein the program is for detecting. The state according to any one of claims 8 to 10, wherein the prediction model created by the computer as the prediction model creation means branches by a classifier to perform prediction by machine learning. Fluctuation detection program. The state change detection program according to any one of claims 8 to 11, wherein the prediction model created by the computer as the prediction model creation means predicts by a random forest. The computer is made to function as the teacher data geometric distance calculation means to obtain the Mahalanobis distance of the teacher data as the geometric distance of the teacher data.
Any one of claims 8 to 12, wherein the computer functions as a means for calculating the geometrical distance of the data to be analyzed so as to obtain the Mahalanobis distance of the data to be analyzed as the geometric distance of the data to be analyzed. The program for detecting state fluctuations described in the section. Any one of claims 8 to 13, wherein the computer functions as the predetermined range importance explanatory variable search means to obtain an explanatory variable of importance lower than a preset threshold value in the prediction model. The program for detecting state fluctuations described in the section.