Model forming device, model forming method, and program

Abstract

To provide technology that can form a highly accurate abnormality diagnostic model even with a small amount of target learning purpose data.SOLUTION: A model forming device according to one aspect of the present disclosure is a model forming device for forming a first abnormality diagnostic model which performs abnormality diagnosis from data belonging to a first domain, the model forming device comprising: a calculation unit that calculates a value of a specific parameter which represents a characteristic of the first abnormality diagnostic model using a first data set that belongs to the first domain; and a forming unit that forms the first abnormality diagnostic model by changing a value of a specific parameter among a plurality of parameters included in a second abnormality diagnostic model constructed from a second data set belonging to a second domain different from the first domain, to a value calculated by the calculation unit.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a model creation device, a model creation method, and a program. [Background technology]

[0002] There are known anomaly diagnosis techniques for monitoring and controlling various plants, facilities, equipment, devices, etc. In these anomaly diagnosis techniques, normal operation data collected from the target is often used as learning data to create an anomaly diagnosis model using a machine learning technique. For example, Non-Patent Document 1 describes a technique called multivariate statistical process control, which uses normal operation data of the target to model its normal state. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Manabu Kano, "Multivariate Statistical Process Control", Internet<URL:http: / / manabukano.brilliant-future.net / research / report / Report2005_MSPC.pdf> Summary of the Invention [Problem to be solved by the invention]

[0004] However, with conventional anomaly diagnosis techniques that use machine learning techniques, it is not possible to create an accurate anomaly diagnosis model until a sufficient amount of learning data has been collected.

[0005] The present disclosure has been made in consideration of the above-mentioned points, and aims to provide a technology that can create an anomaly diagnosis model with high accuracy even when the amount of target learning data is small. [Means for solving the problem]

[0006] A model creation device according to one aspect of the present disclosure is a model creation device for creating a first abnormality diagnosis model that performs abnormality diagnosis from data belonging to a first domain, and includes: a calculation unit that calculates a value of a specific parameter that represents a characteristic of the first abnormality diagnosis model using a first data set belonging to the first domain; and a creation unit that creates the first abnormality diagnosis model by changing the value of a specific parameter among a plurality of parameters included in a second abnormality diagnosis model constructed from a second data set belonging to a second domain different from the first domain, to the value calculated by the calculation unit. Effect of the Invention

[0007] Even if the target learning data is small, a highly accurate anomaly diagnosis model can be created. [Brief description of the drawings]

[0008] [Figure 1] FIG. 2 is a diagram for explaining an example of creating an abnormality diagnosis model by the proposed method according to the first embodiment. [Diagram 2] 1 is a diagram illustrating an example of a hardware configuration of an abnormality diagnosis model creation device according to a first embodiment. [Diagram 3] 1 is a diagram illustrating an example of a functional configuration of an abnormality diagnosis model creation device according to a first embodiment. [Figure 4] 5 is a flowchart showing an example of an abnormality diagnosis model creation process according to the first embodiment. [Diagram 5] FIG. 1 is a diagram illustrating an example of an application of multivariate statistical process control. [Figure 6] FIG. 13 is a diagram showing an example of an abnormality diagnosis result based on an abnormality diagnosis model after parameters have been changed. [Figure 7] FIG. 13 is a diagram illustrating an example in which a separation forest is applied. [Figure 8] FIG. 11 is a diagram illustrating an example of a functional configuration of an abnormality diagnosis model creation device according to a second embodiment. [Figure 9]13 is a flowchart showing an example of an abnormality diagnosis model creation process according to the second embodiment. [Figure 10] FIG. 13 is a diagram illustrating an example of a functional configuration of an abnormality diagnosis device according to a third embodiment. [Figure 11] 13 is a flowchart showing an example of an abnormality diagnosis process according to a third embodiment. [Figure 12] FIG. 13 is a diagram illustrating an example of a functional configuration of an abnormality diagnosis model creation device according to a fourth embodiment. [Figure 13] 13 is a flowchart showing an example of an abnormality diagnosis model creation process according to the fourth embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Hereinafter, first to fourth embodiments of the present invention will be described in detail with reference to the drawings. Note that the first to fourth embodiments can be combined with each other as long as they are not inconsistent.

[0010] [First embodiment] First, a first embodiment will be described. Hereinafter, a case where an abnormality diagnosis model is created for performing abnormality diagnosis on various plants, various facilities, equipment, devices, etc. will be described. In the following, the target for which an abnormality is diagnosed by the abnormality diagnosis model will be called an "abnormality diagnosis target."

[0011] In the first embodiment, it is assumed that an abnormality diagnosis model for another abnormality diagnosis target that is at least similar to the abnormality diagnosis target for which an abnormality diagnosis model is to be created already exists. Hereinafter, the other abnormality diagnosis target for which an abnormality diagnosis model has already been created is referred to as "abnormality diagnosis target S". In addition, the domain to which the driving data of the abnormality diagnosis target S belongs is referred to as the source domain, and the abnormality diagnosis model created from these driving data is referred to as M. S On the other hand, the abnormality diagnosis target for which we want to create an abnormality diagnosis model is called "abnormality diagnosis target T." Also, the domain to which the operation data of abnormality diagnosis target T belongs is called the target domain.

[0012] When you want to create an abnormality diagnosis model for an abnormality diagnosis target T, S Examples of situations in which this already exists include the situations shown in (1) and (2) below.

[0013] (1) Abnormality diagnosis model M S When you have already performed an abnormality diagnosis on an abnormality diagnosis target S using the above, and you want to perform an abnormality diagnosis on an abnormality diagnosis target T that is a different equipment or a new equipment similar to the abnormality diagnosis target S, (2) When you want to perform an abnormality diagnosis on an abnormality diagnosis target T after repairs or modifications have been made to an abnormality diagnosis target S In both (1) and (2) above, the characteristics of the driving data, such as the distribution and the range of possible values, may differ between the abnormality diagnosis target S and the abnormality diagnosis target T, so the domains of the driving data of the abnormality diagnosis target S and the driving data of the abnormality diagnosis target T may differ. For this reason, the abnormality diagnosis model M S It is not possible to diagnose an abnormality of the abnormality diagnosis object T using the abnormality diagnosis model M S When performing abnormality diagnosis on an abnormality diagnosis object T using the above, it is not possible to perform the abnormality diagnosis with sufficient accuracy.

[0014] Therefore, it is necessary to create a new abnormality diagnosis model using the operation data of the abnormality diagnosis target T, but in order to create an accurate abnormality diagnosis model, a sufficient amount of operation data is required as learning data for the target domain. Therefore, until a sufficient amount of learning data is collected, abnormality diagnosis of the abnormality diagnosis target T cannot be performed (or even if abnormality diagnosis can be performed, the accuracy will be insufficient).

[0015] Therefore, the source domain anomaly diagnosis model M S We propose a method (hereinafter referred to as the proposed method related to the first embodiment) for creating an abnormality diagnosis model that can accurately diagnose abnormalities in an abnormality diagnosis target T even when only a small amount of learning data for the target domain is available by utilizing the above.

[0016] <Proposed method according to the first embodiment> Hereinafter, the creation of an abnormality diagnosis model by the proposed method according to the first embodiment will be described with reference to FIG. 1. FIG. 1 is a diagram for explaining an example of the creation of an abnormality diagnosis model by the proposed method according to the first embodiment. Hereinafter, the abnormality diagnosis model M S The training dataset used to create S Let D be a training dataset consisting of a small amount of training data in the target domain. T 1 Let D be a training dataset that consists of a sufficient amount of training data in the target domain. T 2 In addition, the abnormality diagnosis model has a parameter for specifying or determining its characteristics (i.e., a parameter whose value is determined by learning using learning data), and when the parameter θ of a certain abnormality diagnosis model M is explicitly stated, it is written as M(θ). Furthermore, the parameter θ = (θ (1) , ,θ (N) ) The characteristics of the anomaly diagnosis model include, for example, the position, shape, size, orientation, representative point (for example, center, center of gravity, median of each variable, etc.), area, volume, etc., when the anomaly diagnosis model (or the learning data or its feature amount data) is drawn in a space to which the learning data (or the feature amount data representing its features) belongs.

[0017] When there is a small amount of training data for the target domain In this case, as shown in Figure 1(a), the learning dataset D T 1 to the parameter θ to be changed T (i) (Step 1-1). Here, the parameter to be changed θ T (i) is the training data set D T 1 The parameter θ calculated from (i) Here, i ∈ {1, ,N} is set in advance.

[0018] Next, as shown in Fig. 1(a), the abnormality diagnosis model MS (θ S ) parameter θ S =(θ S (1) , ,θ S (N) ) the parameter θ to be changed S (i) The parameter θ to be changed T (i) (steps 1-2).

[0019] As shown in Fig. 1(a), the abnormality diagnosis model M S (θ T ') (where θ T ':=(θ S (1) , ,θ T (i) , ,θ S (N) ) is created as an abnormality diagnosis model for the abnormality diagnosis target T (step 1-3). That is, the abnormality diagnosis model M S Parameter θ S θ included in S (i) θ T (i) The abnormality diagnosis model M was changed to S (θ T ') is created as an abnormality diagnosis model for the abnormality diagnosis object T.

[0020] When sufficient learning data for the target domain has been collected In this case, as shown in Figure 1(b), a learning dataset D for the target domain is generated by a known machine learning method. T 2 From the abnormality diagnosis model M T (θ T ) (where θ T :=(θ T (1) , ,θ T (N) )) (step 2-1).

[0021] As described above, in the proposed method according to the first embodiment, when the amount of learning data for the target domain is small, the change target parameter θ T (i) After calculating, the existing abnormality diagnosis model M S Parameter θ S θ included in S (i) θ T (i) Updated abnormality diagnosis model M S (θ T On the other hand, if sufficient learning data for the target domain is collected, the abnormality diagnosis model M T (θ T In this way, in the proposed method according to the first embodiment, when the amount of learning data for the target domain is small, the existing anomaly diagnosis model M S By utilizing this, we developed an anomaly diagnosis model M S (θ T This creates an anomaly diagnosis model M that can accurately diagnose anomalies in the anomaly diagnosis target T even if the learning data for the target domain is small. S (θ T ') can be obtained.

[0022] Note that a small amount of learning data means that when a new anomaly diagnosis model is constructed from the learning data, the model cannot perform anomaly diagnosis with a desired degree of accuracy, whereas a sufficient amount of learning data means that when a new anomaly diagnosis model is constructed from the learning data, the model can perform anomaly diagnosis with a desired degree of accuracy.

[0023] Hereinafter, an abnormality diagnosis model creating device 10 that creates an abnormality diagnosis model for an object T to be diagnosed by the proposed method according to the first embodiment will be described.

[0024] <Hardware Configuration Example of the Abnormality Diagnosis Model Creation Device 10 According to the First Embodiment> An example of a hardware configuration of the abnormality diagnosis model creation device 10 according to the first embodiment will be described with reference to Fig. 2. Fig. 2 is a diagram showing an example of a hardware configuration of the abnormality diagnosis model creation device 10 according to the first embodiment.

[0025] 2, the abnormality diagnosis model creation device 10 according to the first embodiment includes an input device 101, a display device 102, an external I / F 103, a communication I / F 104, a RAM (Random Access Memory) 105, a ROM (Read Only Memory) 106, an auxiliary storage device 107, and a processor 108. Each of these hardware components is connected to each other via a bus 109 so as to be able to communicate with each other.

[0026] The input device 101 is, for example, a keyboard, a mouse, a touch panel, a physical button, etc. The display device 102 is, for example, a display, a display panel, etc. Note that the abnormality diagnosis model creation device 10 does not necessarily have to include at least one of the input device 101 and the display device 102, for example.

[0027] The external I / F 103 is an interface with an external device such as a recording medium 103a. Examples of the recording medium 103a include a CD (Compact Disc), a DVD (Digital Versatile Disk), an SD memory card (Secure Digital memory card), and a USB (Universal Serial Bus) memory card.

[0028] The communication I / F 104 is an interface for connecting the abnormality diagnosis model creation device 10 to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily stores programs and data. The ROM 106 is a non-volatile semiconductor memory (storage device) that can store programs and data even when the power is turned off. The auxiliary storage device 107 is a non-volatile storage device (storage device) such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive) in which programs and data are stored. The processor 108 is various types of arithmetic devices such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit).

[0029] 2 is merely an example, and the hardware configuration of the abnormality diagnosis model creation device 10 is not limited to this. For example, the abnormality diagnosis model creation device 10 may have a plurality of auxiliary storage devices 107 and a plurality of processors 108, may not have some of the hardware shown in the figure, or may have various hardware other than the hardware shown in the figure.

[0030] <Functional Configuration Example of the Abnormality Diagnosis Model Creation Device 10 According to the First Embodiment> An example of the functional configuration of the abnormality diagnosis model creating device 10 according to the first embodiment will be described with reference to Fig. 3. Fig. 3 is a diagram showing an example of the functional configuration of the abnormality diagnosis model creating device 10 according to the first embodiment.

[0031] As shown in Fig. 3, the anomaly diagnosis model creation device 10 according to the first embodiment includes a change target parameter calculation unit 201, a parameter change unit 202, and a model construction unit 203. These units are realized, for example, by a process in which one or more programs installed in the anomaly diagnosis model creation device 10 are executed by a processor 108 or the like. The anomaly diagnosis model creation device 10 according to the first embodiment also includes a model storage unit 204 and a learning data set storage unit 205. These storage units are realized, for example, by a storage area of ​​the auxiliary storage device 107 or the like. At least one of the model storage unit 204 and the learning data set storage unit 205 may be realized by a storage area of ​​a storage device communicably connected to the anomaly diagnosis model creation device 10.

[0032] The parameter change calculation unit 201 calculates a learning data set D for the target domain. T 1 Using the above, the parameter θ T (i) Here, the parameter θ (1) , ,θ (N) Which of the parameters is to be changed is set in advance. (1) , ,θ (N) Two or more of the parameters may be set as the change target parameters.

[0033] The parameter change unit 202 changes the abnormality diagnosis model M S (θ S ) parameter θ S =(θ S (1) , ,θ S (N) ) the parameter θ to be changed S (i) The parameter θ to be changed T (i) This changes the parameter θ T '=(θ S (1) , ,θ T (i) , ,θ S(N) ) is obtained.

[0034] In addition, the parameter change unit 202 changes the abnormality diagnosis model M S (θ T ') is stored in the model storage unit 204 as an abnormality diagnosis model for the object T to be diagnosed.

[0035] The model construction unit 203 is a learning dataset D for the target domain. T 2 Using this, the abnormality diagnosis model M T (θ T ) (where θ T =(θ T (1) , ,θ T (N) )) from scratch.

[0036] In addition, the model construction unit 203 constructs an abnormality diagnosis model M T (θ T ) is stored in the model storage unit 204 as an abnormality diagnosis model for the object T to be diagnosed.

[0037] The model storage unit 204 stores the abnormality diagnosis model. That is, when the amount of learning data for the target domain is small, the model storage unit 204 stores the abnormality diagnosis model M S (θ S ) and the abnormality diagnosis model M S (θ T On the other hand, when sufficient learning data for the target domain is collected, the model storage unit 204 stores the abnormality diagnosis model M T (θ T ) is stored.

[0038] The learning dataset storage unit 205 stores a learning dataset for the anomaly diagnosis model. That is, when the amount of learning data for the target domain is small, the learning dataset storage unit 205 stores the learning dataset D T 1On the other hand, when sufficient learning data for the target domain is collected, the learning data set storage unit 205 stores the learning data set D T 2 Here, each time learning data for the target domain (i.e., normal operation data of the abnormality diagnosis target T) is acquired, the learning data is stored as a learning data set D T 1 A certain amount of training data is added to the training dataset D T 1 The dataset after adding to is the training dataset D T 2 That is, D T 1 ⊂D T 2 It is.

[0039] In addition, the learning data set storage unit 205 stores an abnormality diagnosis model M S (θ S ) used to construct the training dataset D S is also stored.

[0040] <Abnormality diagnosis model creation process according to the first embodiment> The abnormality diagnosis model creation process according to the first embodiment will be described with reference to Fig. 4. Fig. 4 is a flowchart showing an example of the abnormality diagnosis model creation process according to the first embodiment. Here, steps S101 to S104 in Fig. 4 are executed when the amount of learning data for the target domain is small, and steps S105 to S107 are executed when a sufficient amount of learning data for the target domain has been collected.

[0041] In the following, the learning data of the target domain is not sufficiently collected, and the learning data set storage unit 205 contains only the learning data set D T 1 is stored.

[0042] The parameter change calculation unit 201 calculates a learning data set D for the target domain. T 1 is acquired from the learning dataset storage unit 205 (step S101).

[0043] The change target parameter calculation unit 201 uses the learning data set D acquired in step S101 above. T 1 Using the above, the parameter θ T (i) is calculated (step S102).

[0044] For example, when there are two or more change target parameters, the change target parameter calculation unit 201 calculates each of the two or more change target parameters. Specifically, when there are two change target parameters, the change target parameter calculation unit 201 calculates the change target parameters θ T (i) ,θ T (i') (where i ≠ i' and i, i' ∈ {1, , N}) are calculated. The same applies when the number of parameters to be changed is three or more.

[0045] The parameter change unit 202 changes the abnormality diagnosis model M stored in the model storage unit 204. S (θ S ) parameter θ S =(θ S (1) , ,θ S (N) ) the parameter θ to be changed S (i) The parameter θ to be changed T (i) (step S103). As a result, the changed parameter θ T '=(θ S (1) , ,θ T (i) , ,θ S (N) ) is obtained.

[0046] In addition, when two or more change target parameters are calculated in the above step S102, the parameter change unit 202 changes the parameter θ S =(θ S (1) , ,θ S (N) ) to the two or more change target parameters calculated in step S102. Specifically, for example, when there are two change target parameters, the parameter change unit 202 changes the parameter θ S The parameter θ to be changed is included in S (i) ,θ S (i') The parameter θ to be changed calculated in step S102 above T (i) ,θ T (i') The same applies when the parameter to be changed is 3 or more.

[0047] The parameter change unit 202 changes the abnormality diagnosis model M S (θ T The abnormality diagnosis model M is stored in the model storage unit 204 as an abnormality diagnosis model for the abnormality diagnosis target T (step S104). S (θ T The abnormality diagnosis of the abnormality diagnosis object T is performed using the abnormality diagnosis model creation device 10. The abnormality diagnosis may be performed by the abnormality diagnosis model creation device 10 or by a device different from the abnormality diagnosis model creation device 10.

[0048] Hereinafter, sufficient learning data for the target domain is collected, and the learning data set storage unit 205 stores the learning data set D T 2 is stored.

[0049] The model construction unit 203 is a learning dataset D for the target domain. T 2is acquired from the learning dataset storage unit 205 (step S105).

[0050] The model construction unit 203 is a learning dataset D for the target domain. T 2 Using this, an anomaly diagnosis model M T (θ T ) (where θ T =(θ T (1) , ,θ T (N) The model construction unit 203 constructs a new anomaly diagnosis model M by, for example, multivariate statistical process control (MSPC), which is classified as one of the unsupervised learning techniques. T (θ T ) can be created. However, this is just one example, and the model construction unit 203 can employ any method as long as it is a method that can construct an abnormality diagnosis model by unsupervised learning using normal operating data of the target for abnormality diagnosis as learning data. Specific examples of such methods include, in addition to MSPC, isolation forest, LOF (Local Outlier Factor), One Class SVM, and INNE (isolation using Nearest Neighbor Ensemble), etc.

[0051] The model construction unit 203 constructs the abnormality diagnosis model M T (θ T ) is stored in the model storage unit 204 as an abnormality diagnosis model for the abnormality diagnosis object T (step S107). T (θ T ) is used to perform abnormality diagnosis of the abnormality diagnosis object T. Note that the abnormality diagnosis may be performed by the abnormality diagnosis model creation device 10, or may be performed by a device different from the abnormality diagnosis model creation device 10.

[0052] <Example of the first embodiment> As an example, the anomaly diagnosis model M in the case where multivariate statistical process control is applied as a method of unsupervised learning when constructing the anomaly diagnosis model in the first embodiment will be described below. S (θ T ') and the anomaly diagnosis model M when applying separation forest S (θ T ') will be explained.

[0053] Example 1-1: Application of multivariate statistical process control In this case, the abnormality diagnosis model M S The parameter θ S =(μ S ,E S ,W S ,n PCS ) where μ S is the training dataset D S is the mean value vector of the training data that constitutes E S is a diagonal matrix whose diagonal elements are eigenvalues ​​corresponding to each principal component adopted in principal component analysis used in multivariate statistical process control techniques. S is a matrix whose column vectors are the eigenvectors corresponding to each of these eigenvalues. PCS is the number of principal components used in the principal component analysis used in multivariate statistical process control. S is also called the eigenvector matrix.

[0054] At this time, the parameter μ T is the training dataset D T 1 and in step S103 of FIG. S μ T As a result, the abnormality diagnosis model M S (θ T ') (where θ T '=(μ T ,E S ,W S ,n PCS )) is obtained.

[0055] As an example, when each learning data of the source domain and the target domain is represented as (x1, x2, x3), S (θ S ) and the abnormality diagnosis model M S (θ T The relationship between the mean vector μ S is the abnormality diagnosis model M S (θ S ) is the vector that represents the center of the region represented by the mean vector μ S Let μ be the mean vector T When changed to abnormality diagnosis model M S (θ T ') is a vector whose center is the mean vector μ T So, the abnormality diagnosis model M S (θ S ) is translated (i.e., μ T -μ S The mean vector μ T is the training dataset D for the target domain. T 1 is the mean value vector of each learning data constituting

[0056] Here, the above abnormality diagnosis model M S (θ T An example of an abnormality diagnosis result by the abnormality diagnosis model M' is shown in FIG. 6. The example shown in FIG. 6 is a result of 50 normal operation data items with sample numbers "1" to "50" and 50 abnormal operation data items with sample numbers "51" to "100" being used to derive the abnormality diagnosis model M S (θ T The results are shown in Fig. 1. The abnormality index value was the Q value, and the threshold value was set to th=3.

[0057] As shown in Fig. 6, the above abnormality diagnosis model M S (θ T ') shows that there are no false positives and few missed detections, achieving highly accurate anomaly diagnosis.

[0058] Example 1-2: When Isolation Forest is applied Isolation forest is a technique in which a training data set is downsampled into multiple data sets (hereinafter also referred to as sub-data sets), and then for each sub-data set, random variables are selected and domain division is performed according to the boundary values ​​of the variables, and domain division is performed until only one training data point is included in each domain. In isolation forest, as a result of this domain division, an anomaly diagnosis model (hereinafter also referred to as a sub-model) is created for each sub-data set, which is expressed as a tree structure having a branch node that is a conditional expression relating to the magnitude relationship between the variable and the boundary value of the variable. In addition, during anomaly diagnosis, an ensemble of each sub-model is used to determine whether an anomaly diagnosis target is abnormal or not. Isolation forest is also called, for example, "Isolation Forest".

[0059] Below is the learning dataset D S Let D be the subdataset of S (k) (k=1, ,K), subdataset D S (k) The sub-model created from M S (k) (θ S (k) ) and the parameter θ S (k) The training dataset D S A vector P representing the representative points of each training data S This representative point vector P S The parameter to be changed is the representative point vector P S For example, the training data set D S The vector may be a mean vector or a center of gravity vector of each piece of learning data constituting the vector, or a median vector formed of the medians of each variable.

[0060] At this time, the representative point vector P T is the training dataset D T1 , and each parameter θ S (k) P included in S are P T As a result, the abnormality diagnosis model M S (k) (θ T ' (k) ) (where θ T ' (k) is the parameter θ S (k) P included in S P T (changed to ).

[0061] As an example, the training data of the source domain and the target domain are represented as (x1, x2), and the training data set D S There are three subdata sets D S (1) ,D S (2) ,D S (3) When the submodel M is downsampled to S (k) (θ S (k) ) and submodel M S (k) (θ T ' (k) ) is shown in Figure 7. As shown in Figure 7, the representative point vector P S Let P be the representative point vector T If you change to each sub-model M S (k) (θ T ' (k) ) is for each sub-model M S (k) (θ S (k) ) to P T -P S The result is a translation of the

[0062] <Modification of the first embodiment> A modification of the first embodiment will now be described.

[0063] <<Variation 1>> Training Dataset D S may be a standardized dataset with each variable having a mean of 0 and a standard deviation of 1. Similarly, the training dataset D for the target domain T 1 and D. T 2 may be a data set that has been standardized so that each variable has a mean of 0 and a standard deviation of 1.

[0064] <<Variation 2>> In the above Example 1-1, the abnormality diagnosis model M S (θ S ) is translated into the abnormality diagnosis model M S (θ T In addition to the parallel movement, the abnormality diagnosis model M S (θ S ) is stretched or rotated to create the abnormality diagnosis model M S (θ T Here, the expansion and contraction means that the abnormality diagnosis model M S (θ S ) is considered, each side of the rectangular parallelepiped is stretched or contracted (i.e., the anomaly diagnosis model M S (θ S ) is to expand or contract the size of the abnormality diagnosis model M S (θ S ) (note that these axes form a Cartesian coordinate system).

[0065] - Parallel translation and stretching (part 1) Abnormality diagnosis model M S The parameter θ S =(μ S ,E S ,W S ,n PCS At this time, in step S102 of FIG. 4, the parameters to be changed are μ T1 , E T2 and W T2is calculated, and in step S103 of FIG. S μ T1 , E S E T2 , W S W T2 As a result, the abnormality diagnosis model M S (θ T ') (where θ T '=(μ T1 ,E T2 ,W T2 ,n PCS )) is obtained.

[0066] Here, we use multivariate statistical process control techniques to generate a training dataset D for the target domain. T 1 The parameters (μ T1 ,E T1 ,W T1 ,n PCT1 ) In this case, E T2 and W T2 is calculated as follows:

[0067] Σ S =sqrt(diag(W S E S W S τ )) Σ T1 =sqrt(diag(W T1 E T1 W T1 τ )) [E T2 ,W T2 ]=Svd(Σ T1 -1 Σ S W S E S W S τ Σ S Σ T1 -1 ) where τ is the transposition, diag(·) is the diagonalization operation, Svd(·) is the operation of performing singular value decomposition, and sqrt(·) is the operation of taking the square root of each element value. Singular value decomposition generally produces a total of three outputs: a singular value matrix and two singular value vector matrices, but in this case, since the matrix to be operated on is square, positive definite, and symmetric, the two singular value vector matrices are the same, and two outputs, a singular value matrix and a singular value vector matrix, are obtained. Note that the eigenvalue matrix and eigenvector matrix obtained by performing eigenvalue decomposition on the matrix to be operated on may also be output. In this case, however, it is assumed that the eigenvalues ​​are sorted in descending order.

[0068] ・Translating and rotating (part 1) Abnormality diagnosis model M S The parameter θ S =(μ S ,E S ,W S ,n PCS At this time, in step S102 of FIG. 4, the parameters to be changed are μ T1 , W T1 is calculated, and in step S103 of FIG. S μ T1 , W S W T1 As a result, the abnormality diagnosis model M S (θ T ') (where θ T '=(μ T1 ,E S ,W T1 ,n PCS )) is obtained.

[0069] - Parallel translation and stretching (part 2) Training Dataset D S and D. T 1 is assumed to be standardized. In this case, the abnormality diagnosis model M S The parameter θ S =(μ S ,Σ S ,G S ,V S ,n PCS) where Σ S is a diagonal matrix whose diagonal elements are the standard deviations of each variable, G S is a diagonal matrix whose diagonal elements are the eigenvalues ​​corresponding to each principal component adopted in the principal component analysis used in the multivariate statistical process control method, V S is a matrix whose column vectors are the eigenvectors corresponding to the eigenvalues.

[0070] At this time, in step S102 of FIG. 4, the parameters to be changed are μ T1 , Σ T1 is calculated, and in step S103 of FIG. S μ T1 , Σ S Σ T1 As a result, the abnormality diagnosis model M S (θ T ') (where θ T '=(μ T1 ,Σ T1 ,G S ,V S ,n PCS )) is obtained. Note that Σ T1 is the training dataset for the target domain D T 1 It is a diagonal matrix whose diagonal elements are the standard deviations of each variable.

[0071] ・Translating and rotating (part 2) Training Dataset D S and D. T 1 is assumed to be standardized. In this case, as in the above "Translation and contraction (part 2)", the abnormality diagnosis model M S The parameter θ S =(μ S ,Σ S ,G S ,V S ,n PCS )

[0072] At this time, in step S102 of FIG. 4, the parameters to be changed are μ T1 , ΣT2 ,G T2 ,V T2 is calculated, and in step S103 of FIG. S μ T1 , Σ S Σ T2 , G S G T2 , V S V T2 As a result, the abnormality diagnosis model M S (θ T ') (where θ T '=(μ T1 ,Σ T2 ,G T2 ,V T2 ,n PCS )) is obtained.

[0073] Here, we use multivariate statistical process control techniques to generate a training dataset D for the target domain. T 1 The parameters (μ T1 ,Σ T1 ,G T1 ,V T1 ,n PCT ) In this case, Σ T2 , G T2 and V T2 is calculated as follows:

[0074] [E T1 ,W T1 ]=Svd(Σ T1 -1 V T1 G T1 V T1 τ Σ T1 ) Σ T2 =sqrt(diag(W T1 E S W T1 τ )) [G T2 ,V T2 ]=Svd(Σ T2 -1 W T1 E S WT1 τ Σ T2 -1 ) In addition, E T1 is a diagonal matrix whose diagonal elements are the eigenvalues ​​corresponding to each principal component adopted in the principal component analysis used in the multivariate statistical process control method when standardization is not performed, and W T1 is a matrix in which the column vectors are the eigenvectors corresponding to the eigenvalues ​​when standardization is not performed.

[0075] <Modification 3> In the above-mentioned modified example 2, an anomaly diagnosis model M constructed by a multivariate statistical process control method is S (θ S ) is translated and scaled, or translated and rotated, but the anomaly diagnosis model M constructed by any other unsupervised learning method (e.g., separation forest, etc.) S (θ S ) is translated and scaled, or translated and rotated to create the abnormality diagnosis model M S (θ T ') can also be used. In this case, the training data set D S and D. T 1 Using this, the abnormality diagnosis model M S (θ S ) is translated and scaled, or translated and rotated, and the translation is applied to the anomaly diagnosis model M S (θ S ) parameter θ S can be applied to

[0076] <Summary of the First Embodiment> As described above, when the amount of learning data for the target domain is small, the abnormality diagnosis model creation device 10 according to the first embodiment selects the change target parameter θ T (i) After calculating, the existing abnormality diagnosis model M S Parameter θ S θ included in S (i) θ T(i) Updated abnormality diagnosis model M S (θ T On the other hand, if sufficient learning data for the target domain is collected, the abnormality diagnosis model M T (θ T ) is newly constructed. In this way, in the abnormality diagnosis model creation device 10 according to the first embodiment, even if the learning data of the target domain is small, an abnormality diagnosis model M capable of accurately diagnosing an abnormality in the abnormality diagnosis object T is constructed. S (θ T Therefore, it is possible to start the abnormality diagnosis of the abnormality diagnosis object T without waiting until the learning data of the target domain is sufficiently collected, and it is possible to prevent, for example, a situation in which an abnormality occurring in the abnormality diagnosis object T is overlooked.

[0077] [Second embodiment] Next, a second embodiment will be described. In the second embodiment, a learning dataset D T 1 Using this, we create a training dataset D for the source domain. S After transforming the distribution of, the transformed training dataset D S From the abnormality diagnosis model M S (θ T The abnormality diagnosis model creating device 10 for constructing the abnormality diagnosis model (model A) will now be described.

[0078] <Hardware Configuration Example of the Abnormality Diagnosis Model Creation Device 10 According to the Second Embodiment> An example of the hardware configuration of the abnormality diagnosis model creating device 10 according to the second embodiment may be similar to that of the abnormality diagnosis model creating device 10 according to the first embodiment, and therefore a description thereof will be omitted.

[0079] <Functional Configuration Example of the Abnormality Diagnosis Model Creation Device 10 According to the Second Embodiment> An example of the functional configuration of the abnormality diagnosis model creation device 10 according to the second embodiment will be described with reference to Fig. 8. Fig. 8 is a diagram showing an example of the functional configuration of the abnormality diagnosis model creation device 10 according to the second embodiment. Note that, hereinafter, differences from the first embodiment will be described, and descriptions of functions similar to those of the first embodiment will be omitted.

[0080] As shown in Fig. 8, the anomaly diagnosis model creating device 10 according to the second embodiment includes a data conversion unit 206 and a model construction unit 203. These units are realized, for example, by a process in which one or more programs installed in the anomaly diagnosis model creating device 10 are executed by the processor 108 or the like. The anomaly diagnosis model creating device 10 according to the second embodiment also includes a model storage unit 204 and a learning data set storage unit 205. These storage units are realized, for example, by a storage area of ​​the auxiliary storage device 107 or the like. At least one of the model storage unit 204 and the learning data set storage unit 205 may be realized by a storage area of ​​a storage device communicably connected to the anomaly diagnosis model creating device 10.

[0081] The data conversion unit 206 converts the learning data set D T 1 Using this, we create a training dataset D for the source domain. S The distribution of is transformed. Below, the transformed training dataset D S D S→T Let us express it as follows.

[0082] The model construction unit 203 uses the learning data set D S→T Using this, the abnormality diagnosis model M S (θ T The model construction unit 203 also constructs an abnormality diagnosis model M S (θ T ') is stored in the model storage unit 204.

[0083] <Abnormality diagnosis model creation process according to the second embodiment> The abnormality diagnosis model creation process according to the second embodiment will be described with reference to Fig. 9. Fig. 9 is a flowchart showing an example of the abnormality diagnosis model creation process according to the second embodiment. Here, steps S201 to S204 in Fig. 9 are executed when the amount of learning data for the target domain is small, and steps S205 to S207 are executed when a sufficient amount of learning data for the target domain has been collected. Note that steps S205 to S207 are similar to steps S105 to S107 in Fig. 4, respectively, and therefore description thereof will be omitted.

[0084] In the following, the learning data of the target domain is not sufficiently collected, and the learning data set storage unit 205 contains only the learning data set D T 1 is stored.

[0085] The data conversion unit 206 converts the learning data set D S and the target domain training dataset D T 1 and are acquired from the learning dataset storage unit 205 (step S201).

[0086] The data conversion unit 206 converts the learning data set D T 1 Using this, we create a training dataset D for the source domain. S D S→T (Step S202). As an example, the learning data set D S A case where the distribution of is translated and expanded / contracted, and a case where the distribution is translated and rotated will be described.

[0087] - Translation and stretching Training Dataset D S Let μ be the mean vector of the training data that constitutes S , training dataset D S The diagonal matrix Σ is the diagonal element of which is the standard deviation of each variable of the training data that constitutes SSimilarly, the learning dataset D T 1 Let μ be the mean vector of the training data that constitutes T , training dataset D T 1 The diagonal matrix Σ is the diagonal element of which is the standard deviation of each variable of the training data that constitutes T Let us assume that.

[0088] At this time, each learning data x S ∈D S The training data x is transformed as follows: S→T ∈D S→T Convert to.

[0089] x S→T =Σ T Σ S -1 (x S -μ S )+μ T This results in a training dataset D T 1 A training data set D with the same distribution center and the same standard deviation for each variable. S→T is obtained.

[0090] Translation and rotation Training Dataset D S The eigenvector matrix when the principal components are analyzed is W S , training dataset D S Let μ be the mean vector of the training data that constitutes S Similarly, the learning dataset D T 1 The eigenvector matrix when the principal components are analyzed is W T , training dataset D T 1 Let μ be the mean vector of the training data that constitutes T Let us assume that.

[0091] At this time, each learning data x S ∈D S The training data x is transformed as follows: S→T ∈DS→T Convert to.

[0092] x S→T =W T τ W S (x S -μ S )+μ T This results in a training dataset D T 1 A training dataset D with the same distribution center and the same distribution direction (e.g., the direction of the (hyper)elliptical region that approximates the distribution) as S→T is obtained.

[0093] The model construction unit 203 uses the learning data set D obtained in step S202 above. S→T Using this, an anomaly diagnosis model M S (θ T At this time, the model construction unit 203 constructs a learning data set D S→T In addition, the training dataset D T 1 The abnormality diagnosis model M S (θ T ') may be constructed.

[0094] The model construction unit 203 constructs the abnormality diagnosis model M S (θ T ') is stored in the model storage unit 204 (step S204).

[0095] <Summary of the second embodiment> As described above, when the amount of learning data in the target domain is small, the abnormality diagnosis model creation device 10 according to the second embodiment uses the learning data to convert the learning data in the source domain so that the distribution of the learning data in the source domain becomes the distribution of the learning data in the target domain. This makes it possible to build an abnormality diagnosis model with high accuracy even when the amount of learning data in the target domain is small.

[0096] [Third embodiment] Next, a third embodiment will be described. In the third embodiment, when the amount of learning data for the target domain is small, the anomaly diagnosis model M S (θ T Instead of constructing an anomaly diagnosis model M S (θ S An abnormality diagnosis device 20 for diagnosing an abnormality using the above-mentioned method will now be described.

[0097] <Hardware configuration example of the abnormality diagnosis device 20 according to the third embodiment> An example of the hardware configuration of the abnormality diagnosis device 20 according to the third embodiment may be similar to that of the abnormality diagnosis model creation device 10 according to the first embodiment, and therefore a description thereof will be omitted.

[0098] <Functional configuration example of the abnormality diagnosis device 20 according to the third embodiment> An example of the functional configuration of the abnormality diagnosis device 20 according to the third embodiment will be described with reference to Fig. 10. Fig. 10 is a diagram showing an example of the functional configuration of the abnormality diagnosis device 20 according to the third embodiment.

[0099] As shown in FIG. 10, the abnormality diagnosis device 20 according to the third embodiment includes a data conversion unit 207, an abnormality diagnosis unit 208, and a model construction unit 203. These units are realized, for example, by a process in which one or more programs installed in the abnormality diagnosis device 20 are executed by the processor 108 or the like. The abnormality diagnosis device 20 according to the third embodiment also includes a model storage unit 204 and a learning dataset storage unit 205. These storage units are realized, for example, by a storage area of ​​the auxiliary storage device 107 or the like. At least one of the model storage unit 204 and the learning dataset storage unit 205 may be realized by a storage area of ​​a storage device communicably connected to the abnormality diagnosis device 20.

[0100] The data conversion unit 207 converts the learning data set D T1 and the source domain training dataset D S The diagnostic target data x in the target domain is converted using the above. Hereinafter, the converted diagnostic target data x is referred to as x T→S Let us express it as follows.

[0101] The abnormality diagnosis unit 208 is an abnormality diagnosis model M S (θ S ) to the diagnostic data x T→S Abnormalities are diagnosed from the

[0102] In addition, the abnormality diagnosis unit 208 detects the abnormality diagnosis model M T (θ T ) is constructed, the anomaly diagnosis model M T (θ T ) is used to diagnose abnormalities from the diagnosis target data x.

[0103] The model construction unit 203 is a learning dataset D for the target domain. T 2 Using this, the abnormality diagnosis model M T (θ T ) from scratch.

[0104] In addition, the model construction unit 203 constructs an abnormality diagnosis model M T (θ T ) is stored in the model storage unit 204 as an abnormality diagnosis model for the object T to be diagnosed.

[0105] The model storage unit 204 stores the abnormality diagnosis model. That is, when the amount of learning data for the target domain is small, the model storage unit 204 stores the abnormality diagnosis model M S (θ S On the other hand, when sufficient learning data for the target domain is collected, the model storage unit 204 stores the abnormality diagnosis model M T (θ T ) is stored.

[0106] The learning dataset storage unit 205 stores a learning dataset for the anomaly diagnosis model. That is, when the amount of learning data for the target domain is small, the learning dataset storage unit 205 stores the learning dataset D T 1 On the other hand, when sufficient learning data for the target domain is collected, the learning data set storage unit 205 stores the learning data set D T 2 is stored.

[0107] In addition, the learning data set storage unit 205 stores an abnormality diagnosis model M S (θ S ) used to construct the training dataset D S is also stored.

[0108] <Abnormality Diagnosis Processing According to Third Embodiment> The abnormality diagnosis process according to the third embodiment will be described with reference to Fig. 11. Fig. 11 is a flowchart showing an example of the abnormality diagnosis process according to the third embodiment. Here, steps S301 to S303 in Fig. 11 are executed when the amount of learning data for the target domain is small.

[0109] The data conversion unit 207 acquires the given diagnostic target data x (step S301). Note that the diagnostic target data x is operating data of the target domain.

[0110] The data conversion unit 207 converts the diagnostic object data x acquired in the above step S301 into x T→S Specifically, the learning data set D S The distribution of is used as the training dataset D T 1 Using the inverse transformation of the transformation to convert the diagnostic data x into the distribution of x T→S As an example, let us consider the learning dataset D S We will explain the transformations that translate and scale the distribution of and the transformations that translate and rotate it.

[0111] - Translation and stretching Training Dataset D S Let μ be the mean vector of the training data that constitutes S , training dataset D S The diagonal matrix Σ is the diagonal element of which is the standard deviation of each variable of the training data that constitutes S Similarly, the learning dataset D T 1 Let μ be the mean vector of the training data that constitutes T , training dataset D T 1 The diagonal matrix Σ is the diagonal element of which is the standard deviation of each variable of the training data that constitutes T Let us assume that.

[0112] In this case, the diagnostic target data x is converted to x T→S Convert to.

[0113] x T→S =Σ S Σ T -1 (x-μ T )+μ S This is the training dataset D S The distribution of is used as the training data set D T 1 This is the inverse transformation of the transformation for converting the target domain diagnostic data x into a distribution with the same center and standard deviation of each variable as the distribution of . Therefore, by the above inverse transformation, the diagnostic target data x in the target domain can be converted into the diagnostic target data x in the source domain. T→S It is possible to convert it into

[0114] Translation and rotation Training Dataset D S The eigenvector matrix when the principal components are analyzed is W S , training dataset D S Let μ be the mean vector of the training data that constitutes S Similarly, the learning dataset D T 1The eigenvector matrix when the principal components are analyzed is W T , training dataset D T 1 Let μ be the mean vector of the training data that constitutes T Let us assume that.

[0115] In this case, the diagnostic target data x is converted to x T→S Convert to.

[0116] x T→S =(x-μ T )W S τ W T +μ S This is the training dataset D S The distribution of is used as the training data set D T 1 This is the inverse transformation of the transformation for converting the target domain diagnostic data x into a distribution whose center is the same as that of the source domain diagnostic data x and whose orientation (e.g., the orientation of the (hyper) elliptical region that approximates the distribution) is the same as that of the target domain diagnostic data x. T→S It is possible to convert it into

[0117] The abnormality diagnosis unit 208 uses the abnormality diagnosis model M stored in the model storage unit 204. S (θ S ) to obtain the diagnostic object data x obtained in step S302. T→S From this, an abnormality diagnosis model M S (θ T '), the anomaly diagnosis model M S (θ S ) can be used to perform accurate abnormality diagnosis.

[0118] When sufficient learning data for the target domain has been collected, steps S105 to S107 in FIG. 4 are executed, and the abnormality diagnosis model M T (θ T) is constructed and saved. Thereafter, every time the target domain diagnosis data x is given, the anomaly diagnosis model M T (θ T ) is used to perform an abnormality diagnosis on the diagnosis target data x.

[0119] <Summary of the third embodiment> As described above, when the amount of learning data in the target domain is small, the abnormality diagnosis device 20 according to the third embodiment converts the diagnosis target data using the inverse transformation of the transformation that makes the distribution of the learning data in the source domain the distribution of the learning data in the target domain. This makes it possible to perform accurate abnormality diagnosis using the abnormality diagnosis model in the source domain even when the amount of learning data in the target domain is small.

[0120] [Fourth embodiment] Next, a fourth embodiment will be described. In the fourth embodiment, an abnormality diagnosis model creating device 10 will be described, which constructs a plurality of abnormality diagnosis models for each predetermined time span (or period) until sufficient learning data for a target domain is collected, and uses the abnormality diagnosis model with the highest abnormality diagnosis accuracy for abnormality diagnosis in the next time span.

[0121] Hereinafter, the index of a certain time span (or period) is n, and the time span (or period) of index n is expressed as time span n (or period n). Index n is updated as n←n+1 every time the time span (or period) elapses. In addition, a set of driving data of the target domain collected in time span n (or period n) is referred to as a data set D T 1 (n).

[0122] <Hardware Configuration Example of the Abnormality Diagnosis Model Creation Device 10 According to the Fourth Embodiment> An example of the hardware configuration of the abnormality diagnosis model creation device 10 according to the fourth embodiment may be similar to that of the abnormality diagnosis model creation device 10 according to the first embodiment, and therefore a description thereof will be omitted.

[0123] <Functional configuration example of the abnormality diagnosis model creation device 10 according to the fourth embodiment> An example of the functional configuration of the abnormality diagnosis model creating device 10 according to the fourth embodiment will be described with reference to Fig. 12. Fig. 12 is a diagram showing an example of the functional configuration of the abnormality diagnosis model creating device 10 according to the fourth embodiment.

[0124] As shown in Fig. 12, the anomaly diagnosis model creation device 10 according to the fourth embodiment includes a division unit 209, a model construction unit 203, and a model evaluation unit 210. These units are realized, for example, by a process in which one or more programs installed in the anomaly diagnosis model creation device 10 are executed by the processor 108 or the like. The anomaly diagnosis model creation device 10 according to the fourth embodiment also includes a data set storage unit 211 and a model storage unit 204. These storage units are realized, for example, by a storage area of ​​the auxiliary storage device 107 or the like. At least one of the data set storage unit 211 and the model storage unit 204 may be realized by a storage area of ​​a storage device communicably connected to the anomaly diagnosis model creation device 10.

[0125] The division unit 209 divides the data set D T 1 (n-1) is used as the learning dataset D T 1 (n-1;1) and evaluation dataset D T 1 Split into (n-1;2).

[0126] The model construction unit 203 uses the learning data set D T 1 Using (n-1;1), multiple anomaly diagnosis models M S (θ T ';n-1,m) (m = 1, ,M), where M is the total number of anomaly diagnosis models.

[0127] In addition, the model construction unit 203 uses a learning dataset D for the target domain. T 2 Using this, the abnormality diagnosis model M T(θ T ) is newly constructed. Furthermore, the model construction unit 203 newly constructs an abnormality diagnosis model M T (θ T ) is stored in the model storage unit 204 as an abnormality diagnosis model for the object T to be diagnosed.

[0128] The model evaluation unit 210 uses the evaluation data set D T 1 Using (n-1;2), each abnormality diagnosis model M S (θ T The model evaluation unit 210 evaluates the abnormality diagnosis accuracy of the abnormality diagnosis model M ′; n−1, m) that is evaluated to have the highest abnormality diagnosis accuracy. S (θ T ';n-1,m') (m'∈{1, ,M}) is stored in the model storage unit 204. This abnormality diagnosis model M S (θ T ';n-1,m') is used for abnormality diagnosis in time span n (or period n).

[0129] The data set storage unit 211 stores a data set including a learning data set for the abnormality diagnosis model. That is, the data set storage unit 211 stores a data set D T 1 (n) is stored. If the index of the time span (or period) when sufficient operation data of the target domain is collected is n=N, for example, D T 2 =D T 1 (1)∪∪D T 1 (N).

[0130] In addition, the data set storage unit 211 stores an abnormality diagnosis model M S (θ S ) used to construct the training dataset D S is also stored.

[0131] The model storage unit 204 stores the abnormality diagnosis model. That is, when the driving data of the target domain is not sufficiently collected as learning data, the model storage unit 204 stores the abnormality diagnosis model M S (θ S ) and the abnormality diagnosis model M S (θ T On the other hand, when sufficient driving data of the target domain is collected as learning data, the model storage unit 204 stores the abnormality diagnosis model M T (θ T ) is stored.

[0132] <Abnormality diagnosis model creation process according to the fourth embodiment> The abnormality diagnosis model creation process according to the fourth embodiment will be described with reference to Fig. 13. Fig. 13 is a flowchart showing an example of the abnormality diagnosis model creation process according to the fourth embodiment. Here, steps S401 to S405 in Fig. 13 are repeatedly executed every time a time span n (or period n) elapses while sufficient operation data of the target domain has not been collected. In the following, steps S401 to S405 in Fig. 13 will be described assuming that a time span n-1 (period n-1) has elapsed and n has been updated to n←n+1.

[0133] The division unit 209 divides the data set D T 1 (n-1) is acquired from the dataset storage unit 211 (step S401).

[0134] The division unit 209 divides the data set D acquired in step S401 above. T 1 (n-1) is used as the learning dataset D T 1 (n-1;1) and evaluation dataset D T 1 (n-1; 2) (step S402). T 1 90% of the data in (n-1) is used as the learning dataset D T1 (n-1;1), the remaining 10% of the data is used as evaluation dataset D T 1 For example, the evaluation data set D T 1 Each operation data included in (n-1;2) is labeled with the correct answer (normal / abnormal) of the abnormality diagnosis result.

[0135] The model construction unit 203 uses the learning data set D obtained in step S402 above. T 1 Using (n-1;1), multiple anomaly diagnosis models M S (θ T ';n-1,m) (m=1,...,M) (step S403). For example, the model construction unit 203 constructs the following anomaly diagnosis model M S (θ T ';n-1,1)~M S (θ T ';n-1,4).

[0136] M S (θ T ';n-1,1): Abnormality diagnosis model M S (θ S ) M S (θ T ';n-1,2): Anomaly diagnosis model M S (θ S ) is translated into an anomaly diagnosis model M S (θ T ';n-1, 3): Anomaly diagnosis model M S (θ S ) is translated and scaled to create an anomaly diagnosis model M S (θ T ';n-1, 4): Anomaly diagnosis model M S (θ S ) translated and rotated anomaly diagnosis model However, these are just examples, and other abnormality diagnosis models are also available as abnormality diagnosis models M S (θ T ';n-1,m) may also be constructed.

[0137] The model evaluation unit 210 uses the evaluation data set D T 1 Using (n-1;2), each abnormality diagnosis model M S (θ T In other words, the model evaluation unit 210 evaluates the anomaly diagnosis accuracy of each anomaly diagnosis model M S (θ T ';n-1,m) for each of the evaluation datasets D T 1 We evaluate how accurately we can predict the labels assigned to each driving data contained in (n-1;2).

[0138] The model evaluation unit 210 evaluates the abnormality diagnosis model M that is evaluated to have the highest abnormality diagnosis accuracy. S (θ T ';n-1,m') (m'∈{1,...,M}) is stored in the model storage unit 204 (step S405). As a result, in the time span n (or period n), this anomaly diagnosis model M S (θ T ';n-1,m') will be used for abnormality diagnosis.

[0139] When sufficient operation data of the target domain has been collected as learning data, steps S105 to S107 in FIG. 4 are executed, and the abnormality diagnosis model M T (θ T ) is constructed and saved. From now on, the abnormality diagnosis model M T (θ T ) is used to perform abnormality diagnosis on the abnormality diagnosis object T.

[0140] <Summary of the Fourth Embodiment> As described above, the abnormality diagnosis model creation device 10 according to the fourth embodiment uses the learning data set D T1 (n-1;1) to multiple anomaly diagnosis models M S (θ T ';n-1,m)(m=1,...,M), we construct the anomaly diagnosis model M with the highest anomaly diagnosis accuracy. S (θ T ';n-1,m') is used for abnormality diagnosis in the next time span n (or period n). This makes it possible to perform accurate abnormality diagnosis even when sufficient operation data for the target domain is not collected.

[0141] The present invention is not limited to the above-described embodiments specifically disclosed, and various modifications, changes, combinations with known technologies, etc. are possible without departing from the scope of the claims. [Explanation of symbols]

[0142] 10. Anomaly diagnosis model creation device 20 Abnormality diagnosis device 101 Input Device 102 Display device 103 External I / F 103a Recording media 104 Communication I / F 105 RAM 106 ROM 107 Auxiliary storage 108 processors 109 Bus 201 Parameter calculation unit to be changed 202 Parameter change section 203 Model Construction Department 204 Model Memory Section 205 Learning Data Set Storage Unit 206 Data Conversion Department 207 Data Conversion Department 208 Abnormality Diagnosis Unit 209 Split section 210 Model Evaluation Department 211 Data Set Storage Unit

Claims

1. A model creation device for creating a first abnormality diagnosis model for performing abnormality diagnosis from data belonging to a first domain, comprising: a calculation unit that calculates a value of a specific parameter that represents a characteristic of the first abnormality diagnosis model by using a first data set that belongs to the first domain; a creating unit that creates the first abnormality diagnosis model by changing a value of a specific parameter among a plurality of parameters included in a second abnormality diagnosis model constructed from a second data set belonging to a second domain different from the first domain to a value calculated by the calculating unit; A model creation device having:

2. The calculation unit is Calculating a representative point of data included in the first data set as the value of the specific parameter; The creation unit is 2. The model creating device according to claim 1, wherein the first abnormality diagnosis model is created by changing a representative point of data included in the second data set, among the plurality of parameters, to a representative point calculated by the calculation unit.

3. 3. The model generating device according to claim 2, wherein the representative point is any one of a point representing an average value of the data, a point representing the center of gravity of the data, and a point representing a median value of the values ​​of each variable constituting the data.

4. 4. The model generating device according to claim 1, wherein the number of data included in the first data set is small compared to the number of data included in the second data set.

5. 5. The model creating device according to claim 4, wherein the second abnormality diagnosis model is a machine learning model in which a normal state is modeled by any one of a method of multivariate statistical process control, a separation forest, LOF, a one-class SVM, and INNE.

6. the second domain is a domain to which data of the abnormality diagnosis target belonged before the abnormality diagnosis target was repaired or modified, The model creating device according to claim 1 , wherein the first domain is a domain to which data of the object to be diagnosed of an abnormality belongs after the object to be diagnosed of an abnormality is repaired or modified.

7. the second domain is a domain to which data of a first abnormality diagnosis target belongs; 2 . The model creating device according to claim 1 , wherein the first domain is a domain to which data of a second abnormality diagnosis target similar or of the same type as the first abnormality diagnosis target belongs.

8. The calculation unit is Using the first data set and the second data set, in addition to the representative point, a parameter value representing a magnitude of the first abnormality diagnosis model is calculated as the specific parameter value; The creation unit is 3. The model creating device according to claim 2, wherein the first abnormality diagnosis model is created by changing a value of a parameter representing a magnitude of the second abnormality diagnosis model, among the plurality of parameters, to a value calculated by the calculation unit.

9. The calculation unit is Using the first data set and the second data set, in addition to the representative point, a parameter value representing a direction of the first abnormality diagnosis model is calculated as the specific parameter value; The creation unit is 3. The model creating device according to claim 2, wherein the first abnormality diagnosis model is created by changing a value of a parameter representing a direction of the second abnormality diagnosis model, among the plurality of parameters, to a value calculated by the calculation unit.

10. an evaluation unit that evaluates an abnormality diagnosis accuracy of the first abnormality diagnosis model; and an abnormality diagnosis unit that performs an abnormality diagnosis from data belonging to the first domain, The creation unit is Creating a plurality of the first abnormality diagnosis models; The abnormality diagnosis unit is 10. The model creating device according to claim 8, wherein the abnormality diagnosis is performed using the first abnormality diagnosis model that is most highly evaluated by the evaluation unit, among the plurality of first abnormality diagnosis models.

11. a model creation device for creating a first abnormality diagnosis model for performing abnormality diagnosis from data belonging to a first domain, a calculation step of calculating a value of a specific parameter representing a characteristic of the first abnormality diagnosis model by using a first data set belonging to the first domain; a creating step of creating the first abnormality diagnosis model by changing a value of a specific parameter among a plurality of parameters included in a second abnormality diagnosis model constructed from a second data set belonging to a second domain different from the first domain to a value calculated by the calculating step; A model creation method that performs the following:

12. a model creation device for creating a first abnormality diagnosis model for performing abnormality diagnosis from data belonging to a first domain, a calculation step of calculating a value of a specific parameter representing a characteristic of the first abnormality diagnosis model by using a first data set belonging to the first domain; a creating step of creating the first abnormality diagnosis model by changing a value of a specific parameter among a plurality of parameters included in a second abnormality diagnosis model constructed from a second data set belonging to a second domain different from the first domain to a value calculated by the calculating step; A program that executes the following.

Images