JP6881207B2 - Learning device, program - Google Patents

Description

The present invention relates to, for example, a learning device for learning normal sounds and abnormal sounds from operating sounds of a machine or the like, a learning device for learning normal data and abnormal data of a machine or the like, a learning method, and a program.

It is important from the viewpoint of business continuity to detect a failure of a machine or the like before the failure or quickly after the failure. In order to save labor, there is a field called anomaly detection that detects an "abnormality" that is a deviation from the normal state by using an electric circuit or a program using a sensor or the like. In particular, a sensor that uses a sensor that converts sound into electricity, such as a microphone, is called abnormal sound detection.

Conventionally, data (normal data) such as sonic shapes that have not occurred and are considered to be in a normal state are collected, and a regression model or the like is used to derive a distribution that the normal data follows, and normal / abnormal. A typical technique is to determine normality / abnormality by using the small generation probability and the magnitude of regression error for an unknown sample as the degree of abnormality. For example, there is a conventional technique (Non-Patent Document 2) that uses a reconstruction probability and a reconstruction error of a variational autoencoder (Non-Patent Document 1).

When the abnormal sound detection system is actually operated, it may be possible to collect a very small amount of abnormal data other than normal data in case of failure. However, since the amount of normal data and abnormal data is significantly biased, it is difficult to simply use the abnormal data as a discrimination problem for learning. As a technique for solving this, there is a technique for performing manifold learning using both data (Non-Patent Document 3).

Kingma, Diederik P., and Max Welling. "Auto-encoding variational bayes." ArXiv preprint arXiv: 1312.6114 (2013). An, Jinwon, and Sungzoon Cho. Variational Autoencoder based Anomaly Detection using Reconstruction Probability. Technical Report, 2015. Du, Bo, and Liangpei Zhang. "A discriminative metric learning based anomaly detection method." IEEE Transactions on Geoscience and Remote Sensing 52.11 (2014): 6844-6857.

Since the variational auto-encoder outputs the probability distribution, it is possible to detect anomalies using the inter-distribution distance scale by reflecting the ambiguity information of the estimation result. Variational autoencoders have such merits, but no method has been proposed to introduce a probability distribution for anomalous data.

Therefore, an object of the present invention is to provide a learning device that can be used for learning by introducing a probability distribution related to abnormal data.

The learning device of the present invention includes a learning unit. The learning unit uses data based on normal sound and data based on abnormal sound as learning data, encodes data based on normal sound into the probability distribution of the first latent variable using a variational auto-encoder, and is based on abnormal sound. Learn an encoder that encodes data into a probability distribution of a second latent variable that is different from the probability distribution of the first latent variable, and a decoder that decodes the encoded result.

According to the learning device of the present invention, a probability distribution related to abnormal data can be introduced and used for learning.

The figure which illustrates the probability distribution of the 1st latent variable and the probability distribution of a 2nd latent variable. The block diagram which shows the structure of the learning apparatus of Example 1. FIG. The flowchart which shows the learning operation of the learning apparatus of Example 1. The flowchart which shows the abnormality detection operation of the learning apparatus of Example 1. The figure explaining the encoder and the decoder which the learning apparatus of Example 1 learns. The figure which shows the encoding result of the data based on a normal sound and the data based on an abnormal sound.

Hereinafter, embodiments of the present invention will be described in detail. The components having the same function are given the same number, and duplicate explanations will be omitted.

In the learning apparatus according to the first embodiment below, different probability distributions for normal and abnormal are set for the latent variable prior distribution in the variational auto-encoder learning. For example, for normal data, there is a constraint that gathers at a predetermined origin in the latent space (becomes dense at the predetermined origin and its surroundings), and for abnormal data, it moves away from the predetermined origin (sparse at the predetermined origin and its surroundings). Introduce constraints to learn encoders and decoders. An abnormality can be determined by applying an encoder to unknown test data (detection target data) and calculating a scale related to the proximity of the output latent variable to a predetermined origin. In addition, by using the latent variable space of the variational autoencoder to calculate the degree of anomaly, it is possible to use the difference and ratio of KL divergence to the probability distributions of normal and abnormal, and the variation of the estimation result of the latent variable can be used. Since it can be taken into consideration, improvement in abnormality determination accuracy can be expected.

<Definition of anomalous distribution>
Consider a space in which both normal and abnormal data such as device operating sounds and their latent expression vectors exist as points. For example, if an acoustic feature is used, one sample can be expressed as one point in the Euclidean space having the same dimension as the number of dimensions of the acoustic feature.

From the viewpoint that everything other than the normal pattern is abnormal, it is assumed that all normal samples exist inside the finite normal region D in this space, and that the abnormal samples exist only in places other than the normal region D. ..

_{Consider a uniform distribution U D} (z) on the normal domain D as a distribution representing the degree of normality of the data z. This is just normal if the data is contained within region D. That is, U _D (z) = c _D > 0: Normal. c _D is the probability density value within the definition of uniform distribution. The abnormal, uniform distribution U _{D -} having a positive density locations D ^¯ other than the region D by using a _{(z), U D¯ (z} ) = c D¯ = α (c D -U D ( z))> 0: It is natural to define it as anomalous (α is an appropriate positive constant).
But,

There is no such probability density function U _D￣ (z) because is a contradiction. Therefore, in order to satisfy this condition, another normal distribution R (z) is used.

Is instead a probability density function that represents an anomalous sample. Y is a normalization constant for making the integral with respect to z 1. Here, since it is difficult to derive an analytical solution with a uniform distribution, the uniform distribution U _D (z) is replaced with a normal distribution with a mean 0 variance 1, and the normal distribution R (z) is also a mean 0 variance s ² With a normal distribution of

Is shown as an example of an abnormal distribution (see the broken line in FIG. 1). The value of s is a hyperparameter that is usually determined experimentally. In the example of FIG. 1, the normal distribution p (z) is a normal distribution with a mean 0 variance 1 and has a characteristic of being dense in and around the origin O, and the anomalous distribution p ^￣ (z) is the origin O. It has the characteristic of being sparse in and around it.

The configuration of the learning device 1 of this embodiment will be described with reference to FIG. As shown in the figure, the learning device 1 of this embodiment includes an AD conversion unit 11, a preprocessing unit 12, a normal data storage unit 12a, an abnormal data storage unit 12b, a detection target data storage unit 12c, and the like. It includes a learning unit 13, an encoder / decoder storage unit 13a, an abnormality degree calculation unit 14, and an abnormality determination unit 15. Hereinafter, the operation of each configuration requirement will be described with reference to FIGS. 3 and 4.

<AD conversion (S11-1)>
First, the AD conversion unit 11 AD (analog-digital) converts the sound wave shape of a normal sound (for example, a normal operation sound of a machine) at an appropriate sampling frequency, and prepares quantized waveform data (S11-1). The AD conversion unit 11 AD-converts the sound wave form of an abnormal sound (for example, an abnormal operation sound of a machine) in the same manner as the normal sound (S11-1).

<Data preprocessing (S12-1)>
The pre-processing unit 12 performs feature extraction processing that expands AD-converted normal sound data and abnormal sound data in multiple dimensions by using concatenation of a plurality of samples, discrete Fourier transform, filter bank processing, etc., and data averaging. , Perform preprocessing such as calculating the variance and normalizing the range of values (S12-1). The preprocessing unit 12 is not an indispensable constituent requirement. Normal sound data and abnormal sound data may be treated as data in which one-dimensional values are arranged in chronological order without preprocessing.

The data based on the normal sound after the execution of step S12-1 is called normal data, and the data based on the abnormal sound after the execution of step S12-1 is called abnormal data. When step S12-1 is omitted, the data based on the normal sound after the execution of step S11-1 is referred to as normal data, and the data based on the abnormal sound after the execution of step S11-1 is referred to as abnormal data. The normal data is stored in the normal data storage unit 12a. The abnormal data is stored in the abnormal data storage unit 12b.

<Variational autoencoder learning (S13)>
The learning unit 13 uses normal data and abnormal data as training data, encodes the normal data into the probability distribution (normal distribution) of the first latent variable using the variable auto encoder, and encodes the abnormal data into the first latent variable. Learn an encoder that encodes into a probability distribution (abnormal distribution) of a second latent variable different from the probability distribution of (S13) and a decoder that decodes the encoding result (S13).

The following is an example of learning an encoder / decoder using a variational autoencoder. In the following, the case where the dimension of the latent variable is one-dimensional will be mainly described, but the same can be said for multiple dimensions by taking the sum of each dimension of the optimization criterion. A normal variational autoencoder is used for normal data. As a result, the learning unit 13 has an encoder 131 (q) having a parameter φ that converts the data x (which may be a vector or a scalar) into the distribution N (z; μ _x , σ ² _{x) of the latent variables corresponding to x.} (z | x; φ)) and the decoder 132 (p (x | z; θ)) having the parameter θ that reconstructs the latent variable z into x are learned (see FIG. 5). For the sake of simplicity, _{the subscript x of μ x} and σ ² x will be omitted hereafter.

The optimization criterion is the marginal likelihood maximization criterion of the data x in the model with the parameter θ.

Is. On the other hand, the learning unit 13 has a variational lower limit.

Was obtained, and the distribution q (z | x; φ) was approximated by the empirical distribution of the number of samples L from this distribution.

Optimize instead. The specific optimization criterion is an equation that expresses the amount of Kullback-Leibler information for the probability distribution (normal distribution p (z)) of the first latent variable of the reconstruction error term and encoder q (z | x; φ). Sum

Is. The first half can be regarded as a reconstruction error, and the second half can be regarded as a latent variable constraint term. The coefficient C _n is a term for balancing the two terms. The learning unit 13 executes learning of the encoder and the decoder using this as a learning reference for normal data. On the other hand, the above-mentioned anomalous distribution p ^￣ (z) is substituted for the prior distribution in the variational autoencoder equation. That is, the equation expressing the Kullback-Leibler information amount for ^{the probability distribution (abnormal distribution p ￣} (z)) of the second latent variable of the encoder q (z | x; φ) is

Will be. In order to decompose the addition in the log of the above equation, we introduce an approximation that uses only the terms up to the first order of the Taylor expansion result, and solve this.

Will be. As a standard in optimization,

Will be. At that time, the coefficient C _a was introduced as a similarly balancing term. In addition, in order to decompose the addition in the above log, an approximate expression using terms up to the nth order (n is a natural number of 2 or more) of the Taylor expansion result may be used. The learning unit 13 executes learning of the encoder and the decoder using this as a learning reference for abnormal data. When the latent variable z is multidimensional, if p ^￣ (z) is independent of each dimension, it can be the sum of the above optimization criteria in each dimension. Therefore, let J be the number of dimensions of the latent variable.

May be used as a reference in the case of multiple dimensions. Learning may be performed using only a part of the terms connected by the addition of the optimization criteria, or each term may be arbitrarily multiplied by a constant to change the balance of weights between the terms. Further, the constant coefficient may be removed.

When the number of dimensions of a latent variable is large, it is assumed that the part near the center is difficult to appear due to the spherical concentration phenomenon. In such a case, the Gauss prior distribution, which has a larger variance than the normal distribution, may be used as the anomalous distribution. That is, the abnormal distribution

age,

You may also learn about abnormal data. The normal data and the abnormal data may be learned alternately at any number of times, and when the stochastic gradient method is used, a batch may be created and learned for each normal or abnormal, or the gradient method may be used. For each error calculation, the error for normal and abnormal data may be calculated, and the average may be taken as the total error. Further, in consideration of the bias in the amount of normal data and abnormal data, each gradient may be multiplied by a constant.

<Abnormality calculation (S14)>
The abnormality degree calculation unit 14 performs detection target data (AD conversion in S11-2, preprocessing in S12-2), which is data whose normality or abnormality is unknown and is target data for detecting an abnormality. The posterior probability distribution q (z│x; φ) = N (z; μ) of the latent variable obtained by executing and encoding the data stored in the data storage unit 12c to be detected by the encoder of the variational autoencoder. , Σ ² ) is used to calculate the degree of anomaly (S14).

Multiple methods can be considered for calculating the value (abnormality or normality) proportional to the degree of abnormality. As shown in FIG. 6, the normal data encoding result 131a appears in the normal distribution (broken line in the figure), and the abnormal data encoding result 131b appears outside the normal distribution (inside the abnormal distribution). In step S14, the degree of abnormality is calculated by utilizing this feature.

That is, the abnormality calculation unit 14 uses only the estimated average vector μ of the latent variable, and for example, (1) the probability distribution of the first latent variable of the estimated average vector of the latent variable obtained by encoding the detection target data (1) First likelihood which is the likelihood for (normal distribution) (2) Likelihood of the estimated average vector of the latent variable obtained by encoding the detection target data with respect to the probability distribution (abnormal distribution) of the second latent variable. A second likelihood (3) The degree of anomaly may be calculated using the likelihood ratio of the first likelihood and the second likelihood. Further, the anomaly degree calculation unit 14 uses both the latent variable estimated average vector μ and the estimated variance vector σ ² as an inter-distribution distance scale.
(4) The first KL divergence, which is the amount of Kullback-Leibler information for the probability distribution (normal distribution) of the first latent variable of the latent variable obtained by encoding the detection target data.

(5) The second KL divergence, which is the amount of Kullback-Leibler information for the probability distribution (abnormal distribution) of the second latent variable of the latent variable obtained by encoding the detection target data.

(6) Ratio of the first KL divergence to the second KL divergence

The degree of anomaly may be calculated using any scale using the posterior probability distribution q (z│x; φ) such as. Operations that do not change the order, such as taking a negative logarithm, or complete reversal of the order may be used. In addition, the above conditions such as the value of the lower limit of variation and the reconstruction error due to encoding and decoding may be combined by addition, etc., or the method of introducing the probability distribution for abnormal data is applied only to model learning, and the degree of abnormality is applied. You may use only the reconstruction error in the calculation of.

<Abnormality judgment (S15)>
The abnormality determination unit 15 determines an abnormality from the degree of abnormality calculated in step S14 by using the method used in the abnormality detection method (S15). For example, the abnormality determination unit 15 may perform an abnormality determination based on whether or not the degree of abnormality exceeds a predetermined threshold value as a determination criterion.

<Modification example>
In Example 1 described above, a learning device for learning an encoder / decoder used for detecting an abnormal "sound" by paying attention to a normal "sound" and an abnormal "sound" has been disclosed. Is not limited to sound. The subject of determination of the present invention covers all data that can be determined to be normal or abnormal. Therefore, the first embodiment is modified, and the learning unit uses normal data and abnormal data (not limited to sound data) as learning data, and uses a variable auto encoder to use normal data as a first latent variable. An encoder that encodes into the probability distribution (normal distribution) of, and encodes abnormal data into a probability distribution (abnormal distribution) of a second latent variable that is different from the probability distribution of the first latent variable, and a decoder that decodes the encoding result. It may be configured to learn. The learning device according to the modified example can be applied to anomaly detection targeting arbitrary sensor data such as temperature, pressure, displacement, and traffic data such as network communication volume other than sound.

<Effect>
In the abnormal sound detection using the variational autoencoder, the abnormality can be detected with high accuracy by learning using both normal and abnormal data. At that time, the computational complexity of model learning hardly increases because only the constraint term is changed.

<Supplement>
The device of the present invention is, for example, as a single hardware entity, an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected, and a communication device (for example, a communication cable) capable of communicating outside the hardware entity. Communication unit to which can be connected, CPU (Central Processing Unit, cache memory, registers, etc.), RAM and ROM which are memories, external storage device which is a hard disk, and input units, output units, and communication units of these. , CPU, RAM, ROM, has a connecting bus so that data can be exchanged between external storage devices. Further, if necessary, a device (drive) or the like capable of reading and writing a recording medium such as a CD-ROM may be provided in the hardware entity. A physical entity equipped with such hardware resources includes a general-purpose computer and the like.

The external storage device of the hardware entity stores the program required to realize the above-mentioned functions and the data required for processing this program (not limited to the external storage device, for example, reading a program). It may be stored in a ROM, which is a dedicated storage device). Further, the data obtained by the processing of these programs is appropriately stored in a RAM, an external storage device, or the like.

In the hardware entity, each program stored in the external storage device (or ROM, etc.) and the data necessary for processing each program are read into the memory as needed, and are appropriately interpreted, executed, and processed by the CPU. .. As a result, the CPU realizes a predetermined function (each configuration requirement represented by the above, ... Department, ... means, etc.).

The present invention is not limited to the above-described embodiment, and can be appropriately modified without departing from the spirit of the present invention. Further, the processes described in the above-described embodiment are not only executed in chronological order according to the order described, but may also be executed in parallel or individually as required by the processing capacity of the device that executes the processes. ..

As described above, when the processing function in the hardware entity (device of the present invention) described in the above embodiment is realized by a computer, the processing content of the function that the hardware entity should have is described by a program. Then, by executing this program on the computer, the processing function in the above hardware entity is realized on the computer.

The program describing the processing content can be recorded on a computer-readable recording medium. The computer-readable recording medium may be, for example, a magnetic recording device, an optical disk, a photomagnetic recording medium, a semiconductor memory, or the like. Specifically, for example, a hard disk device, a flexible disk, a magnetic tape, or the like as a magnetic recording device is used as an optical disk, and a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), or a CD-ROM (Compact Disc Read Only) is used as an optical disk. Memory), CD-R (Recordable) / RW (ReWritable), etc., MO (Magneto-Optical disc) as a magneto-optical recording medium, EP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. as a semiconductor memory Can be used.

Further, the distribution of this program is performed, for example, by selling, transferring, or renting a portable recording medium such as a DVD or a CD-ROM on which the program is recorded. Further, the program may be stored in the storage device of the server computer, and the program may be distributed by transferring the program from the server computer to another computer via a network.

A computer that executes such a program first, for example, first stores a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. Then, when the process is executed, the computer reads the program stored in its own recording medium and executes the process according to the read program. Further, as another execution form of this program, a computer may read the program directly from a portable recording medium and execute processing according to the program, and further, the program is transferred from the server computer to this computer. Each time, the processing according to the received program may be executed sequentially. In addition, the above processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition without transferring the program from the server computer to this computer. May be. The program in this embodiment includes information to be used for processing by a computer and equivalent to the program (data that is not a direct command to the computer but has a property of defining the processing of the computer, etc.).

Further, in this form, the hardware entity is configured by executing a predetermined program on the computer, but at least a part of these processing contents may be realized in terms of hardware.

Claims (3)

The data based on the normal sound and the data based on the abnormal sound are used as training data, and the data based on the normal sound is encoded into the probability distribution of the first latent variable by using the variable auto encoder, and the data based on the abnormal sound is used. Kullback for the probability distribution of the encoder for encoding the probability distribution of the different second latent variables and the probability distribution of the first latent variable, a decoder for decoding the encoded result, the pre-Symbol encoder, the second latent variables - an expression representing a Leibler information amount resulting from Taylor expansion only an approximation equation used terms up to first order, as the learning reference for data based on the abnormal sound, the learning device for learning. The data based on the normal sound and the data based on the abnormal sound are used as training data, and the data based on the normal sound is encoded into the probability distribution of the first latent variable by using the variable auto encoder, and the data based on the abnormal sound is used. Kullback for the probability distribution of the encoder for encoding the probability distribution of the different second latent variables and the probability distribution of the first latent variable, a decoder for decoding the encoded result, the pre-Symbol encoder, the second latent variables -Leibler information amount representing the Taylor expansion as a result of n-th order equation (n is a natural number of 2 or more) an approximate expression using only terms up, as learning reference for data based on the abnormal sound, the learning device learns .. A program that causes a computer to function as the learning device according to claim 1 or 2.

Images