Abnormal data generation device, abnormal data generation method, and program

Abstract

To easily generate a large amount of abnormal data.SOLUTION: According to an aspect of the present disclosure, there is provided an abnormal data generation apparatus including a learning unit configured to learn, using normal data acquired from a target device, a model including a compressor configured to compress a dimension of input data and a restorer configured to restore a dimension of output data of the compressor, a compression unit configured to compress the dimension of the normal data using the learned compressor, an estimation unit configured to estimate a normal region representing a region in which the compressed normal data is included, and a generation unit configured to generate data not included in the normal region randomly or in a grid pattern. And a restorer that generates abnormality data of the target device by restoring the dimension of the data.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present disclosure relates to an abnormal data generation device, an abnormal data generation method, and a program. [Background technology]

[0002] Generally, measurement data collected from plants, facilities, etc. contains a large amount of normal data, and a small amount of abnormal data, or even no abnormal data. For this reason, when creating an abnormality diagnosis model using machine learning technology, it may not be possible to achieve sufficient abnormality diagnosis performance. To address this issue, a technology is known that generates abnormality data using simulations based on physical models (Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2019-133212 Summary of the Invention [Problem to be solved by the invention]

[0004] However, since the technology described in Patent Document 1 uses simulation, it is not possible to easily generate a large amount of abnormal data.

[0005] The present disclosure has been made in consideration of the above points, and aims to make it possible to easily generate a large amount of abnormal data. [Means for solving the problem]

[0006] An abnormal data generation device according to one aspect of the present disclosure includes a learning unit that uses normal data acquired from a target device to learn a model composed of a compressor that compresses the dimensions of input data and a restorer that restores the dimensions of output data from the compressor; a compression unit that compresses the dimensions of the normal data using the trained compressor; an estimation unit that estimates a normal area that represents an area that includes the normal data after compression; a generation unit that generates data that is not included in the normal area in a random or grid pattern; and a restoration unit that generates abnormal data of the target device by restoring the dimensions of the data using the trained restorer. [Effects of the Invention]

[0007] A large amount of abnormal data can be easily generated. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 2 is a diagram illustrating an example of a hardware configuration of the abnormal data generating device according to the present embodiment. [Figure 2] FIG. 2 is a diagram illustrating an example of a functional configuration of an abnormality data generating device according to the present embodiment. [Figure 3] 10 is a flowchart illustrating an example of an abnormality data generation process according to the present embodiment. [Figure 4] FIG. 10 is a diagram showing an example of estimation of a normal region. [Figure 5] FIG. 10 is a diagram illustrating an example of generating data that is not included in the normal area. DETAILED DESCRIPTION OF THE INVENTION

[0009] An embodiment of the present invention will be described in detail below with reference to the drawings. In the following embodiment, an abnormality data generation device 10 will be described that can easily generate a large amount of abnormality data for equipment such as a plant or facility (hereinafter also referred to as "target equipment").

[0010] In the following, we assume that sufficient normal data is available for the target device, and define the set of normal data as D={x (i) |i=1,...,I}, where x (i) is the i-th normal data, and I is the total number of normal data. Also, each normal data x (i) is assumed to be expressed in N-dimensional vector form.

[0011] <Example of Hardware Configuration of Abnormal Data Generation Device 10> An example of the hardware configuration of the abnormal data generation device 10 according to this embodiment will be described with reference to Fig. 1. Fig. 1 is a diagram showing an example of the hardware configuration of the abnormal data generation device 10 according to this embodiment.

[0012] 1, the abnormality data generation device 10 according to this 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 pieces of hardware is connected to each other via a bus 109 so as to be able to communicate with each other.

[0013] 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 abnormal data generation device 10 does not necessarily have to include at least one of the input device 101 and the display device 102, for example.

[0014] 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.

[0015] The communication I / F 104 is an interface for communicating with other devices and equipment. 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), an SSD (Solid State Drive), or a flash memory. The processor 108 is one of various arithmetic devices such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit).

[0016] 1 is an example, and the abnormal data generation device 10 may have other hardware configurations. For example, the abnormal data generation device 10 may have multiple auxiliary storage devices 107 and multiple processors 108, or may have various types of hardware other than the hardware shown in the figure.

[0017] <Example of functional configuration of abnormal data generation device 10> An example of the functional configuration of the abnormal data generation device 10 according to this embodiment will be described with reference to Fig. 2. Fig. 2 is a diagram showing an example of the functional configuration of the abnormal data generation device 10 according to this embodiment.

[0018] As shown in FIG. 2 , the abnormal data generation device 10 according to this embodiment includes a learning unit 201, a compression unit 202, a normal region estimation unit 203, a data generation unit 204, a restoration unit 205, and an output unit 206. These units are implemented, for example, by a processor 108 or the like executing one or more programs installed in the abnormal data generation device 10. The abnormal data generation device 10 according to this embodiment also includes a normal data storage unit 207, a model storage unit 208, and an abnormal data storage unit 209. Each of these storage units is implemented, for example, by a storage area of ​​the auxiliary storage device 107 or the like. However, at least one of the normal data storage unit 207, the model storage unit 208, and the abnormal data storage unit 209 may be implemented by a storage area of ​​a storage device (e.g., a storage device included in a database server) or the like communicatively connected to the abnormal data generation device 10.

[0019] The learning unit 201 uses the normal data set D stored in the normal data storage unit 207 to learn the model ML stored in the model storage unit 208. The model ML is a machine learning or statistical model composed of a compressor (encoder) and a decompressor (decoder), and is learned so that the result of compressing and decompressing the dimensions of given input data reproduces the original input data.

[0020] As the model ML, a variational autoencoder (VAE), an autoencoder (AE), a PCA (Principal Component Analysis), etc. can be used, but below we will assume a variational autoencoder. A variational autoencoder is a type of generative model that can estimate the probability of given data using latent variables, and is trained to maximize the variational lower bound (ELBO).

[0021] Below, we will define the decoder (restorer) of the variational autoencoder as p θ (x|z), encoder (compressor) q φLet it be (z|x). Here, θ is a parameter of the decoder, and φ is a parameter of the encoder. Also, z is data in the form of a vector of dimension M (where M < N), which is called a latent variable or a latent vector, etc.

[0022] The compression unit 202 uses the learned encoder q φ (z|x) to compress the dimension of each normal data x included in the normal data set D (i) . That is, the compression unit 202 samples (generates) the latent vector z φ according to q (i) (z|x (i) 0>). As a result, a latent vector set Z = {z (i) |i = 1, ···, I} is obtained.

[0023] The normal region estimation unit 203 estimates the region containing all the latent vectors z included in the latent vector set Z in the latent vector space as the normal region. (i)

[0024] The data generation unit 204 randomly generates latent vectors that are not included in the normal region in the latent vector space. In other words, the data generation unit 204 randomly generates latent vectors that exist outside the normal region in the latent vector space. Hereinafter, the set of latent vectors that exist outside the normal region in the latent vector space is denoted as Z' = {z' (j) |j = 1, ···, J}. Here, J is the number of generated latent vectors.

[0025] The restoration unit 205 uses the learned decoder p θ (x|z) to restore the dimension of each latent vector z' included in the latent vector set Z'. (j) That is, the restoration unit 205 samples (generates) the data x' θ according to p (j) (x|z' (j) ). Since this data x' (i) exists outside the normal region in the latent vector space, it is highly likely to be abnormal data of the target device. Therefore, these data x' (j)By regarding x as abnormal data, the set of abnormal data D'={x' (j) |j=1,···,J} is obtained.

[0026] The output unit 206 outputs the abnormal data set D'={x' (j) |j=1, . . . , J} is output (saved) to the abnormality data storage unit 209.

[0027] The normal data storage unit 207 stores the normal data set D. The model storage unit 208 stores the model ML. The abnormal data storage unit 209 stores the abnormal data set D′ saved by the output unit 206.

[0028] <Abnormal data generation process> An example of the abnormal data generation process according to this embodiment will be described with reference to Fig. 3. Fig. 3 is a flowchart showing an example of the abnormal data generation process according to this embodiment.

[0029] The learning unit 201 acquires the normal data set D stored in the normal data storage unit 207 (step S101).

[0030] The learning unit 201 uses the normal data set D acquired in step S101 to learn the model ML stored in the model storage unit 208 (step S102). That is, the learning unit 201 learns the encoder q φ (z|x) and decoder p θ (x|z). Note that the variational autoencoder can be trained using known optimization methods to maximize the variational lower bound.

[0031] The compression unit 202 uses the trained encoder q φ Using (z|x), each normal data x included in the normal data set D (i) (Step S103). (i) |i=1,···,I} is obtained.

[0032] The normal region estimation unit 203 estimates all latent vectors z included in the latent vector set Z in the latent vector space. (i) The normal region is estimated as a region including all latent vectors z (i) ∈Z, but it is preferable that the hypervolume of the normal region (volume for M=3, area for M=2, length for M=1) is as small as possible. That is, ideally, all latent vectors z (i) It is preferable to define the normal region as the region that includes ∈Z and has the smallest hypervolume.

[0033] As an example, a Delaunay diagram (also called a Delaunay triangulation) for the latent vector set Z can be used as the normal region. The Delaunay diagram for the latent vector set Z is shown in Figure 4. In the example shown in Figure 4, a Delaunay diagram E for the latent vector set Z is obtained, and this Delaunay diagram E can be used as the normal region.

[0034] The normal region may be the convex hull of the latent vector set Z. In this case, the algorithm for finding the convex hull of the latent vector set Z may be, for example, the gift-wrapping method, Graham scan, quick hull, divide-and-conquer method, monotone chain, incremental convex hull algorithm, Kirkpatrick-Seidel algorithm, or Chang's algorithm.

[0035] Furthermore, for example, when M≦3, the latent vector set Z may be visualized as a scatter plot, and the normal region may be estimated by the user encircling the latent vector set Z with their freehand, etc. Alternatively, the normal region may be estimated from the latent vector set Z using unsupervised learning techniques such as PCA and Isolation Forest.

[0036] The data generation unit 204 randomly generates latent vectors that are not included in the normal region in the latent vector space (step S105). For example, the data generation unit 204 randomly generates latent vectors in the latent vector space (particularly, randomly generates latent vectors in a predetermined region that also includes the region surrounding the normal region), and then deletes those latent vectors that are included in the normal region. As a result, a set of latent vectors that exist outside the normal region in the latent vector space, Z'={z' (j) |j=1,...,J} is obtained. The latent vector z' that exists outside the normal region in the latent vector space (j) An example of this is shown in Figure 5. In the example shown in Figure 5, the latent vector z' falls outside the normal region represented by the Delaunay diagram E for the latent vector set Z. (j) (j=1, ,J) is generated. Note that the data generating unit 204 may, for example, generate latent vectors that become lattice points within the latent vector space (especially within a predetermined region including the region surrounding the normal region), and then delete the latent vectors included in the normal region from among those latent vectors.

[0037] The reconstruction unit 205 receives the trained decoder p θ Using (x|z), each latent vector z' included in the latent vector set Z' is (j) (Step S106). (j) |j=1,···,J} is obtained.

[0038] The output unit 206 outputs (stores) the abnormal data set D' obtained in the above step S106 to the abnormal data storage unit 209 (step S107).

[0039] <Summary> As described above, the abnormal data generation device 10 according to this embodiment trains a model ML consisting of a compressor and a restorer, and then uses a normal data set to estimate normal regions within the dimension-reduced space. The abnormal data generation device 10 according to this embodiment then randomly generates data that is not included in the normal region within the dimension-reduced space, and then restores the data to the original space using the restorer. This makes it possible to easily generate any number of abnormal data.

[0040] Therefore, by using the abnormality data generation device 10 according to this embodiment, even when there is little or no abnormality data for the target device, a large amount of abnormality data can be easily generated, making it possible to create an accurate abnormality diagnosis model using supervised learning. Note that the abnormality data generation device 10 according to this embodiment may further include a functional unit (e.g., a functional unit called an "abnormality diagnosis model creation unit" or the like) that creates an abnormality diagnosis model using such a supervised learning technique.

[0041] <Modification> In the above embodiment, the abnormal data set D' is generated using the model ML configured with a compressor and a decompressor, but the abnormal data set D' may be generated without performing dimensionality compression and restoration. Specifically, the abnormal data set D' may be generated by the following steps 1 to 3.

[0042] Step 1: The normal region estimation unit 203 estimates a normal region from the normal data set D (i.e., estimates a normal region in an N-dimensional vector space). Note that the normal region estimation method may be the same as that of step S104 in FIG. 3.

[0043] Step 2: The data generation unit 204 randomly generates data that is not included in the normal region estimated in step 1 above. That is, similar to step S104 in FIG. 3, the data generation unit 204 randomly generates data in the N-dimensional vector space (particularly, randomly generates data in a predetermined region including the region surrounding the normal region), and then deletes the data that is included in the normal region from that data. Note that the data generation unit 204 may, for example, generate data that become lattice points in the N-dimensional vector space (particularly, in a predetermined region including the region surrounding the normal region), and then delete the data that is included in the normal region from that data.

[0044] Step 3: The output unit 206 treats the data generated in step 2 above as abnormal data and outputs (stores) an abnormal data set D′ made up of this abnormal data to the abnormal data storage unit 209.

[0045] The present invention is not limited to the above-described specifically disclosed embodiments, and various modifications, changes, and combinations with known technologies are possible without departing from the scope of the claims. [Explanation of symbols]

[0046] 10. Abnormal data generator 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 Learning Department 202 Compression section 203 Normal region estimation part 204 Data Generation Unit 205 Restoration Department 206 Output section 207 Normal data storage unit 208 Model Memory Unit 209 Abnormal data storage unit

Claims

1. a learning unit that uses normal data acquired from a target device to learn a model that includes a compressor that compresses the dimensions of input data and a restorer that restores the dimensions of output data from the compressor; a compression unit that compresses the dimensions of the normal data using a trained compressor; an estimation unit that estimates a normal area representing an area including normal data after compression; a generating unit that generates data that is not included in the normal area randomly or in a grid pattern; a restoration unit that restores the dimensions of the data using a trained restorer to generate abnormal data of the target device; An abnormality data generating device having the above configuration.

2. The abnormal data generating device according to claim 1 , wherein the model is a variational autoencoder, an autoencoder, or a principal component analysis.

3. The normal region is 2. The abnormal data generating device according to claim 1, wherein the abnormal data generating device is a convex hull or a Delaunay diagram including the compressed normal data, or a region estimated from the compressed normal data by an unsupervised learning method.

4. The generation unit 4. The abnormal data generation device according to claim 1, wherein data is generated randomly or in a grid pattern within a predetermined area that includes the normal area within the compressed space, and data that is included in the normal area is deleted from the data, thereby generating data that is not included in the normal area.

5. an estimation unit that estimates a normal area representing an area including normal data acquired from the target device using the normal data; a generating unit that generates data that is not included in the normal area as abnormal data randomly or in a grid pattern; An abnormality data generating device having the above configuration.

6. a learning procedure for learning a model including a compressor for compressing the dimensions of input data and a restorer for restoring the dimensions of output data of the compressor, using normal data acquired from the target device; a compression procedure for compressing the dimensions of the normal data using a trained compressor; an estimation step of estimating a normal area representing an area including normal data after compression; a generation procedure for generating data not included in the normal area randomly or in a grid pattern; a restoration procedure for restoring the dimensions of the data using a trained restorer to generate abnormal data of the target device; The abnormal data generation method is performed by a computer.

7. an estimation step of estimating a normal area representing an area including normal data acquired from the target device using the normal data; a generation step of generating data that is not included in the normal area as abnormal data randomly or in a grid pattern; The abnormal data generation method is performed by a computer.

8. a learning procedure for learning a model including a compressor for compressing the dimensions of input data and a restorer for restoring the dimensions of output data of the compressor, using normal data acquired from the target device; a compression procedure for compressing the dimensions of the normal data using a trained compressor; an estimation step of estimating a normal area representing an area including normal data after compression; a generation procedure for generating data not included in the normal area randomly or in a grid pattern; a restoration procedure for restoring the dimensions of the data using a trained restorer to generate abnormal data of the target device; A program that causes a computer to execute the following.

9. an estimation step of estimating a normal area representing an area including normal data acquired from the target device using the normal data; a generation step of generating data that is not included in the normal area as abnormal data randomly or in a grid pattern; A program that causes a computer to execute the following.

Images