Anomaly detection device and computer program product
Through the combination of adaptive batch standardization and standardized flow, the trade-offs on data collection cost and accuracy in exception detection are solved, and through pairwise data domain transformation technology, high-precision abnormality detection and stability in a large number of same machines are achieved. study.
Patent Information
- Application Number
- CN201980052478.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2018-11-07
- Filing Date
- 2019-07-04
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2039-07-04
AI Technical Summary
The existing technology has a trade-off between data collection or learning cost and accuracy in abnormal detection, especially in the monitoring scenarios of a large number of same type of machines, and it is difficult to achieve high-precision abnormal detection. At the same time, there is learning in the domain transformation technology of no paired data. Instability issues.
Adopting the probability distribution learning method based on adaptive batch standardization, a common model is generated by generating normal sounds collected from multiple machines, and adjusting the model based on a small amount of adaptive learning data to achieve high-precision anomaly detection; at the same time, standardized flow and self-directed Adapt to the paired-free data domain transformation technology combined with batch standardization, and perform domain transformation stably.
High-precision abnormality detection in a large number of the same type of machines is realized, which reduces the cost of data collection and learning, and improves the reliability and efficiency of learning through stable domain transformation technology.
Smart Images

Figure CN112567460B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to anomaly detection technology and domain transformation technology. Background Art
[0002] For example, if a large-scale manufacturing machine or molding machine installed in a factory stops operating due to a fault, it will cause great obstacles to the business. Therefore, it is necessary to monitor the operating status on a daily basis and respond immediately when an abnormality occurs. As a method to deal with this problem, there is a method in which the management of the business machine regularly dispatches maintenance personnel to the site and the maintenance personnel confirm the wear of parts. However, this method consumes a lot of expenses (for example, labor costs or moving costs) and labor, so it is difficult to implement it on all business machines or factories. Therefore, as an alternative, there is a method of monitoring the operating status on a daily basis based on the operating sound received by the microphone installed inside the machine. Specifically, the operating sound is analyzed, and an alarm is issued when a sound considered abnormal (abnormal sound) is detected. The technology that uses sound to determine whether the machine being monitored is in a normal state or in an abnormal state is called abnormal sound detection technology.
[0003] In the abnormal sound detection technology, the type and detection method of abnormal sound are set for each type or individual of the machine, which is costly. Therefore, it is necessary to automatically design rules for detecting abnormal sound. As one of the solutions, abnormal sound detection based on statistical methods is widely known (Non-patent Document 1). The abnormal sound detection based on statistical methods can be roughly divided into abnormal sound detection with a teacher and abnormal sound detection without a teacher. In abnormal sound detection with a teacher, a large amount of normal sound and abnormal sound are collected as learning data, and the recognizer is learned to maximize the recognition rate (the ratio of normal sound to abnormal sound that can be recognized). On the other hand, in abnormal sound detection without a teacher, only a large amount of normal sound is collected as learning data, and the probability distribution related to the occurrence of normal sound (hereinafter referred to as normal model) is learned. And, using the normal model, when it is determined that the newly collected sound (that is, the sound that becomes the object of abnormal detection) is similar to the normal sound (high likelihood), it is determined to be normal, and when it is determined to be dissimilar (low likelihood), it is determined to be abnormal.
[0004] In the application field of industry, it is difficult to collect a large amount of learning data of abnormal sounds, so unsupervised abnormal sound detection is often used. In addition, in the application field of industry, for a large number of machines of the same type, it is sometimes desirable to set each machine as the object of abnormality detection. For example, it is sometimes desirable to monitor the abnormal sounds of a large number of servers in a data center. In this case, an abnormality detection system is applied to each server separately, but it is assumed that the distribution of the sound emitted from each server is subtly different due to the installation location or assembly error. Therefore, as a method of applying an abnormality detection system to such a large number of machines of the same type, the following method can be considered.
[0005] (1) Using normal sounds collected from some machines, a normal model common to these machines is learned. Then, the common model is used to detect abnormalities in all machines.
[0006] (2) Using the normal sounds collected for each machine, a normal model different for each machine is learned, and the abnormality detection of each machine is performed using each model.
[0007] The method (1) does not learn for each machine, so even if the number of machines to be monitored increases, there is no need to collect learning data or learn models, which can reduce the cost of data collection or learning. However, on the other hand, it is impossible to understand the subtle differences in the normal sound generation of each machine, so it may not be possible to detect abnormalities with high accuracy. In addition, the method (2) only uses the normal sound obtained from each machine for learning, so it is expected to generate a high-precision normal model. However, on the other hand, it is necessary to collect learning data for each machine and learn the model, so there is a problem that the cost of data collection or learning is generated as the number of machines to be monitored increases.
[0008] Next, domain conversion is described. Domain conversion refers to a technique for converting data in a certain domain into data in a domain different from the domain. Here, the data to be converted is, for example, an image or sound. For example, as described in reference to non-patent document 1, an "image of a landscape photo" is converted into an "image of a landscape painting", or an "image of a horse photo" is converted into an "image of a zebra photo".
[0009] (Reference Non-Patent Literature 1: Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros, "Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks", arXiv:1703.10593v5, https: / / arxiv.org / abs / 1703.10593v5)
[0010] In order to realize domain conversion, a data converter from domain D to domain D' can be created. In order to create such a data converter, the following method can be considered: for example, a large number of pairs of "landscape photos" and "landscape paintings" realistically depicted therefrom are collected as learning data, and a neural network is used to learn the conversion from "landscape photos" to "landscape paintings". Hereinafter, the structure of learning the data converter using pairs of data from two domains is called domain conversion with paired data. Domain conversion with paired data has the following advantages: using pairs of data from two domains, which are input and the correct answer to the input, as learning data, it is possible to construct the data converter relatively easily through learning. However, a large amount of learning data has to be collected. In the example of "landscape photos" and "landscape paintings" just now, it is necessary to first collect "landscape photos" and then create "landscape paintings" realistically depicting "landscape photos" (for example, by commissioning a painter to create them). In addition, in the example of "horse photos" and "zebra photos", it is difficult to take photos with the same composition, so it is impossible to collect learning data in reality.
[0011] Therefore, in recent years, in order to solve the problems related to learning data collection, a framework that can learn data transformers without using paired data has been proposed. This framework is called domain transformation without paired data. In domain transformation without paired data, data from domain D and data from domain D' are used to learn a data transformer that transforms data from domain D to data from domain D'. Here, the data from domain D and the data from domain D' used in learning do not need to be paired. Therefore, even for transformations between domains such as "photos of horses" and "photos of zebras" where it is difficult to collect paired data, a data transformer can be learned.
[0012] As an example of domain transformation without paired data, a method such as StarGAN described in Non-Patent Document 2 has been proposed.
[0013] [Prior art literature]
[0014] [Non-patent literature]
[0015] [Non-patent document 1] Ide Tsuyoshi and Sugiyama Masaru, “Abnormality detection and change detection (Abnormality detection and change detection)”, Kodansha, pp. 6-7, 2015.
[0016] [Non-patent document 2] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, Jaegul Choo, "StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation", IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2018, pp.8789-8797, 2018. Summary of the invention
[0017] Problems to be solved by the invention
[0018] First, the first subject will be described.
[0019] As described above, in the methods (1) and (2), there is a problem of the trade-off between the cost involved in data collection or learning and the accuracy of abnormality detection. Therefore, as a third method, the following method can also be considered.
[0020] (3) Using normal sounds collected from some machines, a normal model common to these machines is learned. Next, using normal sounds collected from each machine, a normal model different for each machine is adaptively learned based on the common model. And, using each adaptive normal model, abnormality detection is performed for each machine.
[0021] Although the method (3) may be able to solve the above trade-off problem, in order to actually implement such a method, it is necessary to collect a small amount of normal sounds from each machine that is the target of abnormality detection, and it is necessary to be able to use the normal sounds to efficiently learn a normal model that is adaptive to each machine based on a common normal model. However, such a method has not been developed so far.
[0022] This is the first topic.
[0023] Next, the second problem will be described.
[0024] In StarGAN of non-patent document 2, a method such as Generative Adversarial Networks (GAN) described in reference non-patent document 2 is used to achieve domain transformation without paired data, but there is a problem of unstable learning.
[0025] (Reference non-patent literature 2: Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, BingXu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio, "Generative Adversarial Nets", Advances in Neural Information Processing Systems 27 (NIPS2014), 2018.)
[0026] This is the second topic.
[0027] Therefore, the present invention aims to provide a highly accurate anomaly detection technology that suppresses the cost required for normal model learning, and a domain transformation technology that enables stable unpaired data learning.
[0028] Means for solving problems
[0029] One embodiment of the present invention includes: an abnormality degree estimation unit, which estimates the abnormality degree representing the degree of abnormality of the machine based on the sound emitted from the abnormality detection target machine (hereinafter referred to as the abnormality detection target sound) based on a first probability distribution representing the distribution of normal sounds emitted from one or more machines different from the abnormality detection target machine and the association with the normal sound emitted from the abnormality detection target machine (hereinafter referred to as the normal sound for adaptive learning).
[0030] One embodiment of the present invention is a probability distribution learning device, comprising: a learning unit, based on normal sounds emitted from one or more machines different from the abnormality detection target machine (hereinafter referred to as learning normal sounds), learning a first probability distribution representing the distribution of normal sounds emitted from one or more machines different from the abnormality detection target machine, wherein a variable x of the first probability distribution q1(x; θ) is a variable representing input data generated based on normal sounds emitted from one or more machines different from the abnormality detection target machine, and the variable x is obtained by using a transformation f i (i=1, ..., K, K is an integer greater than 1, relative to the transformation f i There exists an inverse transformation f i -1) and the latent variable z0 is represented as x=f K (f K-1 (…(f1(z0))…)), taking q0(z0) as the probability distribution of the latent variable z0, the probability density q1(x;θ) of the input data x is calculated by using the latent variable z0=f1 corresponding to the input data x. -1 (f2 -1 (…(f K -1 The probability density q0(z0) of (x))…)) is calculated for the transformation f i At least one transformation among (i=1, ..., K), whose inverse transformation is adaptive batch normalization.
[0031] One embodiment of the present invention includes: an abnormality degree estimation unit, which estimates the abnormality degree representing the degree of abnormality of the machine based on the sound emitted from the abnormality detection target machine (hereinafter referred to as the abnormality detection target sound) according to the association between a first autoencoder that restores normal sound emitted from one or more machines different from the abnormality detection target machine and the normal sound emitted from the abnormality detection target machine (hereinafter referred to as the normal sound for adaptive learning).
[0032] One embodiment of the present invention is an autoencoder learning device, comprising: a learning unit, which learns a first autoencoder for restoring normal sounds emitted from one or more machines different from the abnormality detection target machine based on normal sounds emitted from one or more machines different from the abnormality detection target machine (hereinafter referred to as normal sounds for learning), wherein the first autoencoder is constructed as a neural network including an AdaBN layer for performing adaptive batch normalization calculations.
[0033] One embodiment of the present invention includes: an abnormality degree estimation unit, which estimates the abnormality degree representing the degree of abnormality of the machine based on the sound emitted from the abnormality detection target machine (hereinafter referred to as the abnormality detection target sound) based on the collection of normal sounds emitted from one or more machines different from the abnormality detection target machine and the association with the normal sound emitted from the abnormality detection target machine (hereinafter referred to as the normal sound for adaptive learning).
[0034] One embodiment of the present invention is a data transformation device, comprising: a latent variable calculation unit, which calculates latent variables based on input data corresponding to domain data of a first domain; and an output data calculation unit, which calculates output data corresponding to domain data of a second domain based on the latent variables, the latent variable calculation unit uses a specific function with an inverse function (hereinafter referred to as the first function) to perform calculations, and the output data calculation unit uses a specific function with an inverse function (hereinafter referred to as the second function) to perform calculations, and the first function and the second function are derived from a specific function that transforms the latent variable z0 into the variable x.
[0035] Effects of the Invention
[0036] According to the present invention, the cost required for normal model learning can be suppressed and high-precision abnormality detection can be performed. In addition, according to the present invention, domain conversion that enables stable unpaired data learning can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 This is a block diagram showing an example of the structure of the probability distribution learning device 100.
[0038] Figure 2 This is a flowchart showing an example of the operation of the probability distribution learning device 100 .
[0039] Figure 3 It is a block diagram showing an example of the structure of the probability distribution adaptive learning device 200.
[0040] Figure 4 This is a flowchart showing an example of the operation of the probability distribution adaptive learning device 200 .
[0041] Figure 5 This is a block diagram showing an example of the structure of the abnormality detection device 300.
[0042] Figure 6 This is a flowchart showing an example of the operation of the abnormality detection device 300.
[0043] Figure 7 2 is a block diagram showing an example of the configuration of abnormality degree estimation section 320 .
[0044] Figure 8 : is a flowchart showing an example of the operation of abnormality estimation section 320 .
[0045] Fig. 9 4 is a block diagram showing an example of the structure of the autoencoder learning device 400.
[0046] Fig.10 : is a flowchart showing an example of the operation of the autoencoder learning device 400.
[0047] Fig.11 : is a block diagram showing an example of the structure of the autoencoder adaptive learning device 500.
[0048] Fig.12 : is a flowchart showing an example of the operation of the autoencoder adaptive learning device 500.
[0049] Fig.13 This is a block diagram showing an example of the structure of the abnormality detection device 600.
[0050] Fig.14 This is a flowchart showing an example of the operation of the abnormality detection device 600.
[0051] Fig.15 2 is a block diagram showing an example of the configuration of abnormality degree estimation section 620 .
[0052] Fig.16 2 is a flowchart showing an example of the operation of abnormality degree estimation section 620 .
[0053] Fig.17 This is a block diagram showing an example of the structure of the probability distribution learning device 1100.
[0054] Fig.18 This is a flowchart showing an example of the operation of the probability distribution learning device 1100 .
[0055] Fig.19 This is a block diagram showing an example of the structure of the data conversion device 1200.
[0056] Fig. 20 This is a flowchart showing an example of the operation of the data conversion device 1200.
[0057] Fig.21 2 is a diagram showing a state of processing performed by latent variable calculation section 1220 and output data calculation section 230 .
[0058] Fig. 22 This is a block diagram showing an example of the structure of the data conversion device 1300.
[0059] Fig.23 This is a flowchart showing an example of the operation of the data conversion device 1300.
[0060] Fig.24 This is a diagram showing an example of a functional configuration of a computer for realizing each device in the embodiment of the present invention. DETAILED DESCRIPTION
[0061] Hereinafter, the embodiments of the present invention will be described in detail. In addition, the same numbers are attached to the structural units having the same functions, and the repeated description is omitted.
[0062] <Memorization>
[0063] _ (underscore) indicates a subscript. For example, x y_z It means that yz is a superscript of x. y_z Represents y z is the subscript for x.
[0064] First, the technical background of the first to third embodiments of the present invention and each embodiment will be described.
[0065] <Technical Background>
[0066] An embodiment of the present invention provides a framework for abnormality detection that can be applied to multiple machines of the same type. Specifically, a framework based on the method of (3) above is provided. Based on the hypothesis that the difference in the distribution of normal sounds emitted from each machine is expressed by the statistics of the feature quantity related to the sound, by making the average of the feature quantity as a primary statistic and the variance as a secondary statistic consistent between different machines, normal models of different machines can be derived from one model.
[0067] First, the conventional technology used in the embodiment of the present invention will be described.
[0068] Abnormal sound detection without a teacher
[0069] Abnormal sound detection is a task that determines whether the condition of the monitored object machine that emits the sound (observation signal) that is the target of abnormality detection is normal or abnormal. Here, in the input data x generated from the observation signal, for example, the logarithmic amplitude spectrum ln|X of the observation signal can be used as follows: t,f |A vector as elements.
[0070] [Number 1]
[0071] x:=(ln|X t-Q,1 |,ln|X t-Q,2 |,…,ln|X t-Q,F |,
[0072] ln|X t-Q+1,1 |,ln|X t-Q+1,2 |,…,ln|X t+Q,F |) T …(1)
[0073] Here, t={1, ..., T} and f={1, ..., F} represent a time index and a frequency index, respectively. Also, Q represents the number of past and future frames considered in the input.
[0074] The input data x is not limited to the above-mentioned example, and other feature quantities may be used as input data generated from an observation signal.
[0075] Next, the abnormal sound detection based on density estimation is described. The input data generated from the normal sound (hereinafter referred to as the normal sound input data) is assumed to be generated according to the probability distribution p(x). First, a probability distribution q(x; θ) with parameter θ is designed. Next, a set {x} of N normal sound input data generated according to the probability distribution p(x) is used. i} i=1N , find the parameter θ that makes q(x;θ) closest to p(x) * , q(x;θ * ) is set as the approximate distribution of p(x). And, when the input data of the sound as the abnormality detection object is input, the abnormality degree A(x; θ) associated with the input data is calculated by the following formula, for example: * ).
[0076] [Number 2]
[0077] A(x;θ * )=-lnq(x;θ * )…(2)
[0078] Finally, the determination result R is obtained by, for example, the following formula.
[0079] [Number 3]
[0080] R=H(A(x;θ * )-φ)…(3)
[0081] Here, the threshold is a specific constant, and H(·) is a step function that returns 1 if the argument is non-negative and returns 0 if the argument is negative. When R=1, the machine that emits the sound that is the target of the abnormality detection is judged to be abnormal, and when R=0, it is judged to be normal. That is, if the abnormality degree A(x; θ * ) is greater than the pre-set threshold If it is large, it is considered abnormal.
[0082] As described above, in the abnormal sound detection based on density estimation, it is necessary to (1) design the probability distribution q(x; θ) and (2) determine the parameter θ. In the design of the probability distribution q(x; θ) in (1), a normalizing flow can be used. In addition, in the determination of the parameter θ in (2), for example, a loss function (objective function) can be used, which is L(θ) = -∑logq(x i ;θ)(that is, for the set of normal sound input data {x i} i=1 N Learning with the gradient descent method of (the sum of the log-likelihoods and negative log-likelihoods).
[0083] Normalizing Flow
[0084] Normalization flow is a method of finding a distribution that approximates the probability distribution p(x) associated with data generation.
[0085] The following is an explanation.i (z)} i=1 K Suppose there are K transforms with inverse transforms (where f i (z): R D →R D , R is a set of real numbers, and D is an integer greater than 1). In addition, f i -1 (z)(i=1,...,K) is set to f i The inverse transform of (z).
[0086] In the normalized flow, for a set of N input data {x i} i=1 N , there are corresponding latent variables {z 0,i} i=1 N , input data x i It can be considered that by using K transformations {f i (z)} i=1 K and the following formula of the latent variable z0 of x, for the corresponding latent variable z 0,i The transformed data.
[0087] [Number 4]
[0088] x=f K (f K-1 (…(f1(z0))…))…(4)
[0089] That is, for i=1, ..., K, the following equation holds.
[0090] [Number 5]
[0091] x i =f K (f K-1 (…(f1(z 0,i ))…))…(4)′
[0092] In addition, it is sometimes expressed as z1=f1(z0), z2=f2(z1), ..., x=f K (z K-1 ).
[0093] Furthermore, assuming that the latent variable {z 0,i} i=1 N For example, a probability distribution q0(z0) that is easy to sample using Monte Carlo is generated based on an isotropic Gaussian distribution. i} i=1 NThe probability distribution q(x;θ) (x is a variable representing input data) used can be expressed as follows.
[0094] [Number 6]
[0095]
[0096] Here, z0=f1 -1 (f2 -1 (...(f K -1 (x))...))). In addition, {θ i} i=1 K is the same as the transformation {f i (z)} i=1 K The corresponding parameters, θ = [θ1 T ,θ2 T , ..., θ K T ] T .
[0097] In addition, the probability distribution q0(z0) is not limited to the distribution that is easy to perform Monte Carlo sampling, as long as it is a distribution that is easy to perform rigorous probability density estimation. As an example of a distribution that is easy to perform rigorous probability density estimation, a probability distribution p(x) that satisfies the following conditions can be cited.
[0098] (Condition 1) For a probability distribution p(x), there exists R D A non-negative function g(x)(≥0) on D It becomes p(x)=g(x) / ∫g(x)dx.
[0099] (Condition 2) For the function g(x), it is easy to calculate ∫g(x)dx.
[0100] As a function satisfying condition 2, Gaussian distribution can be cited. On the other hand, as a function not satisfying condition 2, g(x)=exp(sin(x)-x 2 ).
[0101] In the normalization flow, the input data set {x i} i=1 N To learn the parameter θ of the probability distribution q(x;θ). Then, using the parameter θ obtained by learning * The probability distribution q(x;θ * ), approximate the original probability distribution p(x) associated with data generation.
[0102] In the normalized stream, various transformations can be used as transformations {f i (z)} i=1 K For example, batch normalization or leaky ReLU (Rectified Linear Unit: normalized linear function) described in reference non-patent document 3 can be used. In addition, the following linear transformation described in reference non-patent document 4 can also be used.
[0103] [Number 7]
[0104] f(z)=LUz…(6)
[0105] Here, L,U∈R D×D are lower triangular matrices and upper triangular matrices respectively. This transformation can be achieved by taking the absolute value of the product of the diagonal elements of L and U (that is, |Π i=1 D L ii U ii |) to calculate the absolute value of the Jacobian matrix Therefore, it has the characteristic that the probability density q(x; θ) of the input data x can be easily calculated (the calculation cost of the probability density q(x; θ) of the input data x can be suppressed) (refer to formula (5)).
[0106] (Refer to non-patent document 3: S. Ioffe, C. Szegedy, "Batch normalization: accelerating deep network training by reducing internal covariate shift", ICML 2015, 2015.)
[0107] (Refer to non-patent document 4: J.Oliva, et al., “Transformation Autoregressive Networks”, ICML 2018, 2018.)
[0108] Below, we simply normalize the batch BN:x→y(x, y∈R D ) is explained. In batch normalization BN, adjustments are made so that the set of input data {x i} i=1 N After the average of the elements of each dimension of becomes 0 and the variance becomes 1, scale transformation and offset transformation are performed. Specifically, y is calculated by the following formula i =BN(x i ).
[0109] [Number 8]
[0110]
[0111] Here, γ and β are scale transformation parameters and offset transformation parameters, respectively, and are parameters to be learned. In addition, ε is a non-negative real number, which is set to a positive real number when it is desired to avoid division by zero, and is set to zero when it is not necessary to avoid division by zero.
[0112] In addition, since the scale transformation parameter γ and the offset transformation parameter β are made explicit, BN(·) is sometimes expressed as BN γβ (·).
[0113] Furthermore, the K transformations do not all need to be of the same type, and thus, for example, some types of transformations may be combined, such as performing batch normalization on the transformation f1(z) and performing linear transformation on the transformation f2(z).
[0114] AdaBN: Adaptive Batch Normalization
[0115] Domain adaptation refers to a technique that, when the distribution of learning data used in model learning is different from the distribution of test data to be processed using the learned model, adjusts the learned model based on the difference in distribution so that the accuracy of the processing using the learned model is not reduced. Here, the set of learning data and the set of test data are domains, and sometimes they are each a learning domain and a test domain.
[0116] There are various domain adaptation methods that can be combined with a deep neural network (DNN). Here, adaptive batch normalization is described (refer to non-patent document 5). Adaptive batch normalization is a method that calculates the mean and variance in batch normalization and adjusts the mean and variance for each domain (refer to formulas (7a) to (7d)). In other words, it is assumed that the calculation based on formulas (7a) to (7c) is performed on the data of the same domain. When actually testing, for the set of input data {x i} i=1 N Calculate the statistics (mean and variance) and use them to output the processing result y through equations (7c) and (7d) i In addition, when the transformation is adaptive batch normalization, it is sometimes expressed as AdaBN:x→y(x, y∈R D ).
[0117] (Refer to non-patent document 5: Y. Li, et al., “Revisiting Batch Normalization For Practical Domain Adaptation”, ICLR 2017, 2016.)
[0118] The following describes anomaly detection in an embodiment of the present invention. First, a problem setting for anomaly detection in an embodiment of the present invention is described. Next, a specific structure of anomaly detection in an embodiment of the present invention using the above-mentioned prior art is described.
[0119] 《Problem Setting》
[0120] The problem to be solved is to "use a common normal model, i.e., a first probability distribution, learned by using a large number of normal sounds obtained from a plurality of machines, and a small number of normal sounds obtained from a machine to be detected as an abnormality, to learn a normal model, i.e., a second probability distribution, that can be used for a machine to be detected as an abnormality, and use the second probability distribution to perform abnormality detection based on the sound emitted by the machine to be detected as an abnormality". Thus, the following data is processed.
[0121] (1) Learning data: normal sounds emitted from one or more machines different from the abnormality detection target machine, and it is assumed that a large amount of data can be prepared. Since these sounds are used in learning, they are called normal sounds for learning. In addition, their collection is called a learning domain. In addition, the machine for collecting normal sounds is preferably a machine of the same type as the abnormality detection target machine.
[0122] (2) Adaptive learning data: These are normal sounds emitted from the abnormality detection target machine. It is assumed that only a small amount can be prepared. Since these sounds are used in adaptive learning, they are called normal sounds for adaptive learning. In addition, their collection is the test domain described later.
[0123] (3) Test data: This is the sound emitted by the abnormality detection target device, and the device is judged to be normal or abnormal based on the sound. Therefore, this sound is called the abnormality detection target sound. In addition, its collection is called the test domain.
[0124] Hereinafter, the stage of learning using normal sound is called the learning stage, the stage of adaptive learning using normal sound is called the adaptive learning stage, and the stage of abnormality detection based on the abnormality detection object sound is called the testing stage (abnormality detection stage).
[0125] Key Points
[0126] In the embodiment of the present invention, in order to adaptively learn the second probability distribution with a small amount of adaptive learning data and a low amount of computation, adaptive batch normalization is introduced into the normalization stream. Specifically, for the K transformations {f i (z)} i=1 K At least one transformation f i (z), its inverse transformation f i -1 (z) is the adaptive batch normalization. In addition, in the calculation of the adaptive batch normalization, the scale transformation and the offset transformation, that is, the calculation of equation (7d) can also be omitted. If another expression is used, it can also be expressed as the inverse transformation f i -1 (z) is adaptive batch normalization with γ=1, β=0.
[0127] "Concrete Structure"
[0128] The specific structure is described below.
[0129] (1) Learning stage
[0130] First, a neural network is described that transforms input data x generated from normal sounds (that is, learning data) emitted from one or more machines different from the abnormality detection target machine into a latent variable z0 (~q(z0)) that is considered to be generated according to the probability distribution q0(z0). Here, the neural network using five transformations {f i (z)} i=1 5 That is, z0 is z0=f1 -1 (f2 -1 (f3 -1 (f4 -1 (f5 -1 (x))))) is obtained.
[0131] The above five transformations {f i (z)} i=1 5 It is defined by the following formula. In addition, for convenience, the transformation f is represented instead of i And its inverse transformation f i -1 (where z4=f5 -1 (x), z3=f4 -1 (z4), z2=f3 -1 (z3), z1=f2 -1 (z2), z0=f1 -1 (z1)).
[0132] [Number 9]
[0133]
[0134] Here, L2, L5∈R D×D is a lower triangular matrix with diagonal components of 1, except for all elements of the upper triangular part L 2,ij , L 5,ij (i≥j) is the learning target parameter (that is, parameter θ2 or parameter θ5). D2, D5∈R D×D is a diagonal matrix with diagonal components D 2,ij , D 5,ij (i=j) is the learning target parameter (that is, parameter θ2 or parameter θ5). U2, U5∈R D×D is an upper triangular matrix with diagonal components of 1, except for the lower triangular part, all elements U 2,ij , U 5,ij (i≤j) is a learning object parameter (that is, parameter θ2 or parameter θ5). In addition, α3 (≥0) is a parameter of LeakyReLU, which can be set as a hyperparameter or as a learning object parameter (that is, parameter θ3) (when α3 is set as a learning object, it is called a parametric ReLU (Parametric ReLU) (see non-patent document 6)). In addition, AdaBN γ_4β_4 (·) and AdaBN γ_1β_1 (·) is the above-mentioned adaptive batch normalization, and γ1, β1, γ4, β4 are learning object parameters (that is, parameter θ1 or parameter θ4).
[0135] (Refer to non-patent document 6: K. He, et al., "Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification", ICCV 2015, pp. 1026-1034, 2015.)
[0136] In addition, the transformation {f i (z)} i=1 5 The absolute values of the Jacobian matrix formulas are each calculated by the following formula (where x=f5(z4), z4=f4(z3), z3=f3(z2), z2=f2(z1), z1=f1(z0)).
[0137] [Number 10]
[0138]
[0139] Here, s4' is the standard deviation of z4 (corresponding to the input data x generated from the learning data), δ is the number of elements below zero in z3 (corresponding to the input data x generated from the learning data), and s1' is the standard deviation of z1 (corresponding to the input data x generated from the learning data). In addition, the absolute value of the matrix formula of the Jacobian for the transformations f4 and f1 is The absolute value of the Jacobian matrix when using inference (that is, when using a learned model) rather than when using learning is recorded.
[0140] In addition, as described above, the probability distribution q0(z0) is set to a probability distribution that is easy to perform rigorous probability density estimation. For example, if the probability distribution q0(z0) is set to a Gaussian distribution N(0, I) with a mean of 0 and a variance of the unit matrix I, the probability distribution q0(z0) is expressed by the following formula.
[0141] [Number 11]
[0142]
[0143] Therefore, it can be seen that by setting q1(x; θ) as the probability distribution of the input data x generated from the learning data and using equation (5), it is possible to obtain the probability distribution of the latent variable z 0,i The probability density q0(z 0,i ) Calculate the input data x i The probability density q1(x i ;θ).
[0144] Next, the learning method of the parameter θ is described. As with the usual neural network learning, the loss function L(θ) can be used, for example, by gradient descent, momentum SGD (Stochastic Gradient Descent), ADAM (Adaptive Moment Estimation), or a combination thereof. When using a normalized flow, the average of the negative log-likelihood defined by the following formula is often used as the loss function L(θ).
[0145] [Number 12]
[0146]
[0147] In addition, in the above learning, a mini-batch learning method can be used, which uses a set of learning data called a mini-batch as a unit. Here, a mini-batch refers to a plurality of learning data randomly selected from all learning data. The value of the loss function L(θ) is calculated for each mini-batch.
[0148] (2) Adaptive learning stage
[0149] Next, the adaptive learning method for generating the probability distribution q2(x;θ) of the input data x based on the normal sound emitted by the abnormality detection target device (that is, the adaptive learning data) is described. -1 (x), z3=f4 -1 (z4), z2=f3 -1 (z3), z1=f2 -1 (z2), z0=f1 -1 (z1), the learning can be performed through the following process. First, according to the set of input data {x' i} i=1 M Calculate {z' 4,i} i=1 M (z'4, i =f5 -1 (x' i )). Next, find {z' 4,i} i=1 M '. Finally, m and s in formula (7c) are 2 Use the calculated mean and variance to replace. Similarly, according to the input data set {x' i} i=1 M Calculate {z' 1,i} i=1 M (z' 1,i =f2 -1 (f3 -1 (f4 -1 (f5 -1 (x' i )))))) Next, find {z' 1,i} i=1 M Finally, m and s in formula (7c) are 2 Replace with the found mean and variance.
[0150] (3) Testing phase
[0151] As the abnormality detection method, for example, the method described in "Abnormal Sound Detection Without Instruction" can be used.
[0152] "Effect"
[0153] By introducing adaptive batch normalization into the normalization flow, the following effects can be achieved.
[0154] (1) The difference between the distribution of learning data and the distribution of test data can be adjusted, and the decrease in the accuracy of abnormality detection in the test domain can be suppressed.
[0155] Specifically, if adaptive batch normalization is used in the final layer, the parallel shift in the axial direction and the scale deviation of the data distribution between domains can be corrected. In addition, when the difference in the distribution of data between different domains is expressed by parallel shift and axial scaling, in principle, as long as adaptive batch normalization is introduced to the first layer, even if the parameter θ learned in the learning domain is applied to the test domain as is, anomaly detection in the test domain can be performed with high accuracy.
[0156] (2) The second probability distribution can be adaptively learned with low computational effort.
[0157] As described above, the processing required for adaptive learning of the second probability distribution is basically just to calculate the statistics in adaptive batch normalization for the adaptive learning data. Therefore, adaptive learning can be performed with a lower amount of calculation than the case of repeating normal learning, and can also be performed online depending on the situation.
[0158] In addition, as in the above example, when a linear transformation is introduced into the normalized stream, the following effects can be obtained by performing LU decomposition or LDU decomposition on the matrix W corresponding to the linear transformation in advance.
[0159] (3) The computational cost of probability density is suppressed, and the learning cost is suppressed.
[0160] In the normalized flow, the matrix of the Jacobian of the linear transformation f needs to be calculated one by one. Therefore, when the matrix W is kept in the form of LU decomposition or LDU decomposition, the calculation of the matrix |W| takes O(k times the number of W). 3 ) is calculated. However, when W=LU or W=LDU is maintained in the form of LU decomposition or LDU decomposition, |W| can be obtained by |W|=|LU|=|L|×|U| (that is, the product of all the diagonal components of L and the diagonal components of U) or |W|=|LDU|=|L|×|D|×|U| (that is, the product of all the diagonal components of L, the diagonal components of D, and the diagonal components of U), so the calculation can be performed very quickly.
[0161] <First embodiment>
[0162] In a situation where there are two or more machines of the same type, consider detecting an abnormality in one of the machines (set as the abnormality detection target machine). To this end, first, based on normal sounds emitted from one or more machines different from the abnormality detection target machine (hereinafter referred to as normal sounds for learning), learn a probability distribution representing the distribution of normal sounds emitted from one or more machines different from the abnormality detection target machine (hereinafter referred to as the first probability distribution) (learning stage). Next, using normal sounds emitted from the abnormality detection target machine (hereinafter referred to as normal sounds for adaptive learning), adaptively learn a probability distribution representing the distribution of normal sounds emitted from the abnormality detection target machine (hereinafter referred to as the second probability distribution) based on the first probability distribution (adaptive learning stage). And, based on the sound emitted from the abnormality detection target machine (hereinafter referred to as the abnormality detection target sound), determine whether the machine is abnormal (testing stage (abnormality detection stage)).
[0163] The probability distribution learning device 100 learns the first probability distribution based on the normal sound for learning. The probability distribution adaptive learning device 200 adaptively learns the second probability distribution based on the first probability distribution using the normal sound for adaptive learning. The abnormality detection device 300 determines whether the device is abnormal based on the abnormality detection target sound.
[0164] Below, reference Figure 1-2 The probability distribution learning device 100 will be described. Figure 1 It is a block diagram showing the structure of the probability distribution learning device 100. Figure 2 is a flowchart showing the operation of the probability distribution learning device 100. Figure 1 As shown, the probability distribution learning device 100 includes an input data generating unit 110, a latent variable estimating unit 120, a loss function calculating unit 130, a parameter updating unit 140, a convergence condition determining unit 150, and a recording unit 190. The recording unit 190 is a structural unit that appropriately records information required for the processing of the probability distribution learning device 100. The recording unit 190 records, for example, a parameter θ of the first probability distribution q1(x; θ) as a learning object. As the initial value of the parameter θ, for example, a value generated by using a random number is recorded.
[0165] The probability distribution learning device 100 is connected to a normal sound recording unit 910 for learning. Normal sounds for learning prepared in advance are recorded as learning data in the normal sound recording unit 910. As described above, it is preferable to prepare as many normal sounds for learning as possible in advance.
[0166] according to Figure 2 The operation of the probability distribution learning device 100 is described. The input data generating unit 110 generates a normal sound for learning based on the input. i (i=1, ..., N), generate input data xi (i=1, ..., N) (S110). For example, as shown in equation (1), a normal sound s for learning is generated. i The logarithmic amplitude spectrum of the vector is taken as the element, and it is set as the input data x i In addition, it is also possible to use a voice feature quantity other than the above as the normal voice for learning. i Generated input data x i .
[0167] The latent variable estimation unit 120 uses the parameter θ of the first probability distribution q1(x; θ) to estimate the input data x generated in S110. i (i=1,…,N), estimation and input data x i The corresponding latent variable z0, i (i=1, . . . , N) (S120) The parameter θ used here is a value being learned.
[0168] Here, the variable x of the first probability distribution q1(x;θ) to be learned is a variable representing input data generated from normal sound emitted from one or more devices different from the abnormality detection target device, and the variable x is set by using the transformation f i (i=1,…,K,K is an integer greater than 1,for transformation f i There exists an inverse transformation f i -1 ) and the latent variable z0 is represented as x=f K (f K-1 (…(f1(z0))…)).
[0169] Thus, with the input data x i The corresponding latent variable z 0,i Given by the following formula.
[0170] [Number 13]
[0171]
[0172] In addition, the latent variable z 0,i (i=1, ..., N) is assumed to be generated according to the probability distribution q0(z0) of the latent variable z0. The probability distribution q0(z0) has the property that a strict probability density estimation is easy.
[0173] Therefore, the probability distribution q1(x;θ) can be expressed by the following equation using the probability distribution q0(z0) (see equation (5)).
[0174] [Number 14]
[0175]
[0176] Here, θ i is the transformation f i The corresponding parameters are θ = [θ1 T ,θ2 T ,…,θ K T ] T .
[0177] The latent variable estimation unit 120 can be implemented by using a neural network NN that calculates the latent variable z0 based on the input data x. In addition, the calculation formula is as follows.
[0178] [Number 15]
[0179]
[0180] That is to say, the neural network NN takes the input data x as input and transforms it according to the inverse transformation f K -1 , f K-1 -1 , …, f2 -1 , f1 -1 The calculation is carried out in sequence, and the latent variable z0 is finally output.
[0181] For the above transformation f i At least one transformation f among (i=1, ..., K) i_0 (where i0 is an integer satisfying 1≤i0≤K), let its inverse transformation f i_0 -1 This is called adaptive batch normalization. In this way, a small amount of adaptive learning normal sounds can be used to adaptively learn the second probability distribution q2(x;θ) (see "Key Points").
[0182] In addition, it can also be set as, for the transformation f i At least one transformation f among (i=1, ..., K) i_1 (where i1 is an integer satisfying 1≤i1≤K), its inverse transformation f i_1 -1 It is a linear transformation, and the matrix corresponding to the linear transformation is expressed as the product of the lower triangular matrix and the upper triangular matrix, or the product of the lower triangular matrix, the diagonal matrix and the upper triangular matrix. In this way, the calculation of the probability density required for learning the first probability distribution q1(x;θ) (the calculation in the loss function calculation unit 130 described later) can be performed at low cost (refer to (3) of "Effect").
[0183] As a specific transformation f i In the example of (i=1, . . . , K), for example, K=5 can be set, and five transformations represented by equations (8a) to (8e) can be used as inverse transformations.
[0184] The loss function calculation unit 130 calculates the loss function based on the latent variable z estimated in S120. 0,i (i=1, ..., N), calculate the value of the loss function L(θ) used in the optimization of the parameter θ of the first probability distribution q1(x; θ) (S130). The loss function L(θ) can be defined as the average of the negative log-likelihoods as in equation (11). At this time, it is necessary to calculate the input data x i The probability density q1(x i ;θ), input data x i The probability density q1(x i ; θ) can be obtained by using the input data x i The corresponding latent variable z 0,i The probability density q0(z 0,i ) to calculate. For example, when the probability distribution q0(z0) is a Gaussian distribution N(0, I), the latent variable z 0,i The probability density q0(z 0,i ) can be calculated by the following formula:
[0185] [Number 16]
[0186]
[0187] Therefore, we can use formula (5)', according to the latent variable z calculated above. 0,i The probability density q0(z 0,i ) Calculate the input data x i The probability density q1(x i ;θ).
[0188] The parameter updating unit 140 updates the parameter θ of the first probability distribution q1(x;θ) so as to optimize (minimize) the value of the loss function L(θ) calculated in S130 (S140). In updating the parameter θ, for example, a gradient descent method may be used.
[0189] The convergence condition determination unit 150 determines the convergence condition that is set in advance as the end condition of the parameter update. When the convergence condition is satisfied, the first probability distribution q1(x; θ) using the parameter θ updated in S140 is output. When the convergence condition is not satisfied, the processing of S110 to S140 is repeated (S150). As the convergence condition, for example, a condition such as whether the number of executions of the processing of S110 to S140 has reached a specific number of times can be used. When outputting, the parameter θ updated in S140 (this parameter is also referred to as the learning completion parameter) and the inverse transformation f can also be output. K -1 (x;θ K ), fK-1 -1 (z K-1 θ K-1 ), …, f2 -1 (z2;θ2), f1 -1 (z1;θ1).
[0190] In addition, the structural unit including the input data generating unit 110, the latent variable estimating unit 120, the loss function calculating unit 130, the parameter updating unit 140, and the convergence condition determining unit 150 is referred to as the learning unit 105. That is, the learning unit 105 is a structural unit that learns the first probability distribution based on the normal sound for learning.
[0191] According to the invention of this embodiment, it is possible to learn the first probability distribution q1 (x; θ) that can easily adjust the difference between the distribution of learning data and the distribution of test data.
[0192] <Modification>
[0193] The probability distribution learning device 100 has been described as learning a first probability distribution q1(x; θ) based on normal sounds emitted from one or more machines different from the abnormality detection target machine, i.e., normal sounds for learning. However, it may also be configured as follows: the number of one or more machines different from the abnormality detection target machine is set to W (W is an integer greater than 1), and the first probability distribution q1 representing the distribution of normal sounds emitted from each of the W machines is learned. (1) (x;θ),…,q1 (W) (x; θ). In this case, by inputting data x (equivalent to a small batch) i The latent variable estimation unit 120 performs the latent variable estimation process by calculating the mean and variance in the adaptive batch normalization according to the input data generated from the normal sound for learning issued by the same machine (i=1, ..., N). That is, in the calculation in the adaptive batch normalization, the mean and variance of W groups are used instead of the mean and variance of 1 group. Here, the learned parameter θ is in the W first probability distributions q1 (1) (x;θ),…,q1 (W) (x;θ) are common, so they are group 1.
[0194] <Second embodiment>
[0195] Below, reference Figure 3-4 The probability distribution adaptive learning device 200 is described. Figure 3 It is a block diagram showing the structure of the probability distribution adaptive learning device 200. Figure 4 2 is a flowchart showing the operation of the probability distribution adaptive learning device 200. Figure 3As shown, the probability distribution adaptive learning device 200 includes an input data generating unit 110, a parameter updating unit 240, an output unit 250, and a recording unit 190. The recording unit 190 is a structural unit that appropriately records information required for the processing of the probability distribution adaptive learning device 200. The recording unit 190, for example, pre-records the parameter θ (that is, the learned parameter) of the first probability distribution q1 (x; θ) learned by using the probability distribution learning device 100. The learned parameter becomes the initial value of the parameter θ of the second probability distribution q2 (x; θ). In addition, the variable x of the second probability distribution q2 (x; θ) is a variable representing input data generated based on the normal sound emitted from the abnormality detection target machine.
[0196] The probability distribution adaptive learning device 200 is connected to the adaptive learning normal sound recording unit 920. The adaptive learning normal sound recording unit 920 records the adaptive learning normal sound prepared in advance as the adaptive learning data. As described above, the adaptive learning normal sound can be prepared in advance in a smaller amount than the learning normal sound.
[0197] according to Figure 4 The operation of the probability distribution adaptive learning device 200 is described. The input data generating unit 110 generates a normal sound s′ for adaptive learning based on the input. i (i=1, ..., M), generate input data x' i (i=1, . . . , M) (S110) Here, the number M of normal sounds for adaptive learning is basically an integer not exceeding the number N of normal sounds for learning, but may be an integer exceeding the number N of normal sounds for learning.
[0198] The parameter updating unit 240 uses the input data x′ generated in S110 i (i=1, ..., M), update the parameter θ of the second probability distribution q2(x; θ) (S240). Specifically, for the transformation f whose inverse transformation is adaptive batch normalization i_0 , the mean and variance to be used in its calculation, using the input data x' i (i=1, ..., M) to calculate the inverse transformation f i_0 -1 The input data z' i_0,i (=f i_0-1 -1 (f i_0-2 -1 (…(f K -1 (x' i ))…))) can be updated by the average and variance. That is, m and s in formula (7c) 2 Using input data z'i_0,i The mean and variance of (i=1,…,M) are used for replacement.
[0199] The output unit 250 outputs the second probability distribution q2(x;θ) using the parameter θ updated in S240 (S250). In addition, when outputting, the parameter θ updated in S240 (this parameter is also referred to as a learned parameter), the inverse transformation f K -1 (x;θ K ), f K-1 -1 (z K-1 θ K-1 ), …, f2 -1 (z2;θ2), f1 -1 (z1;θ1).
[0200] The structural unit including the input data generating unit 110, the parameter updating unit 240, and the output unit 250 is referred to as the adaptive learning unit 205. That is, the adaptive learning unit 205 is a structural unit that adaptively learns the second probability distribution from the first probability distribution using the normal sound for adaptive learning.
[0201] According to the invention of this embodiment, the second probability distribution q2 (x; θ) can be learned by adjusting the difference between the distribution of learning data and the distribution of test data.
[0202] <Modification>
[0203] In the probability distribution learning device 100, W first probability distributions q1 are learned (1) (x;θ),…,q1 (W) In the case of (x; θ), there is only one set of learned parameters θ. The probability distribution adaptive learning device 200 performs adaptive learning using this one set of parameters.
[0204] <Third Embodiment>
[0205] Below, reference Figure 5-6 The abnormality detection device 300 will be described. Figure 5 It is a block diagram showing the structure of the abnormality detection device 300. Figure 6 is a flow chart showing the operation of the abnormality detection device 300. Figure 5As shown, the abnormality detection device 300 includes an input data generation unit 110, an abnormality degree estimation unit 320, an abnormality determination unit 330, and a recording unit 390. The recording unit 390 is a structural unit that appropriately records information required for processing of the abnormality detection device 300. The recording unit 390 records, for example, in advance the parameter θ (that is, the learned parameter) of the second probability distribution q2 (x; θ) learned by using the probability distribution adaptive learning device 200.
[0206] That is, the abnormality detection device 300 performs abnormality detection using the second probability distribution q2(x;θ) using the learned parameters as the learned second probability distribution.
[0207] according to Figure 6 The operation of the abnormality detection device 300 will be described. The input data generating unit 110 generates input data x based on the abnormality detection target sound s as input (S110).
[0208] The abnormality estimation unit 320 estimates the abnormality indicating the degree of abnormality of the machine according to the input data x generated in S110 based on the learned second probability distribution q2(x;θ) (S320). Figure 7-8 The abnormality degree estimation unit 320 will be described. Figure 7 2 is a block diagram showing the structure of abnormality estimation section 320 . Figure 8 is a flowchart showing the operation of the abnormality estimation unit 320. Figure 7 As shown, the abnormality estimation unit 320 includes a latent variable calculation unit 321 and an abnormality calculation unit 322 .
[0209] according to Figure 8 The operation of the abnormality estimation unit 320 is described. The latent variable calculation unit 321 calculates the latent variable z0 corresponding to the input data x based on the input data x generated in S110 (S321). Specifically, the calculation can be performed using a neural network in which the parameters of the neural network NN are set to the learned parameters of the second probability distribution.
[0210] The abnormality calculation unit 322 calculates the abnormality A(x;θ) related to the input data x based on the latent variable z0 calculated in S321 (S322). The abnormality can be calculated by the following equation, for example.
[0211] [Number 17]
[0212] A(x; θ)=-lnq2(x; θ)…(2)′
[0213] When the probability distribution q0(z0) is a distribution for which a strict probability density estimation is easy, the probability density q2(x;θ) of the input data x can be calculated by the following formula.
[0214] [Number 18]
[0215]
[0216] When calculating Formula (5)", for example, when the probability distribution q0(z0) is a Gaussian distribution N(0, I), and the inverse transformation is the five transformations represented by Formulas (8a) to (8e), Formula (10) or Formulas (9a) to (9e) can be used.
[0217] The abnormality determination unit 330 generates a determination result indicating whether the machine is abnormal based on the abnormality A(x;θ) estimated in S320 (S330). For example, using equation (3), a determination result indicating abnormality is generated when R=1, and a determination result indicating normality is generated when R=0.
[0218] That is, the abnormality detection device 300 can be said to include: an abnormality degree estimation unit 320, which estimates the abnormality degree indicating the degree of abnormality of the machine according to the sound emitted from the abnormality detection target machine (abnormality detection target sound) based on the first probability distribution indicating the distribution of normal sounds emitted from one or more machines different from the abnormality detection target machine and the association with the normal sound emitted from the abnormality detection target machine (normal sound for adaptive learning). An example of the association is a second probability distribution indicating the distribution of normal sounds emitted from the abnormality detection target machine obtained by updating the first probability distribution using the normal sound for adaptive learning.
[0219] According to the invention of this embodiment, by performing abnormality detection using the second probability distribution that adjusts the difference between the distribution of learning data and the distribution of test data, it is possible to perform highly accurate abnormality detection. In other words, it is possible to suppress a decrease in abnormality detection accuracy.
[0220] In addition, in the first to third embodiments, a method of calculating the abnormality A(x; θ) using a standardized flow is described, but the abnormality can also be obtained using other statistical models. For example, an autoencoder (AE: autoencoder) described in reference non-patent document 7 can also be used. An autoencoder is a combination of an encoder (encoder) and a decoder (decoder).
[0221] (Reference Non-Patent Document 7: Y. Koizumi, S. Saito, H. Uematsu, Y. Kawachi, and N. Harada, "Unsupervised Detection of Anomalous Sound based on Deep Learning and the Neyman-Pearson Lemma," IEEE / ACM Transactions on Audio, Speech, and Language Processing, Vol. 27-1, pp. 212-224, 2019.)
[0222] Hereinafter, the technical background of the fourth to sixth embodiments of the present invention and each embodiment will be described.
[0223] Technical Background
[0224] When using an autoencoder, the abnormality can be calculated by the following formula.
[0225] [Number 19]
[0226] A(x;θ)=||xD(E(x,θ E ), θ D )|| 2 …(13)
[0227] Here, ||·|| represents the L2 norm, E and D represent the encoder and decoder respectively, and θ E and θ D They represent the parameters of the encoder E and the decoder D respectively. That is, θ={θ E ,θ D}.
[0228] Both the encoder E and the decoder D can be configured as a neural network. In this case, for example, θ is learned so that the reconstruction (restoration) error of the learning data of the normal sound is minimized.
[0229] [Number 20]
[0230]
[0231] Here, x i is the learning data of the i-th normal sound, and N is the number of samples of the learning data of the normal sound.
[0232] In order to realize adaptive learning using an autoencoder, the encoder E and the decoder D, or one of them, may be configured as a neural network using adaptive batch normalization (AdaBN). That is, adaptive batch normalization is used in the calculation process of the encoder E and the decoder D, or one of them. For example, instead of configuring the encoder E as E(x; θ E )=W2[σ(W1x+b1)]+b2, and the encoder E can be constructed as a neural network that calculates the following equation with the AdaBN layer inserted.
[0233] [Number 21]
[0234] E(x,θ E )=W2{AdaBN[σ(W1x+b1)]}+b2…(15)
[0235] Here, W1 and W2 represent weight matrices, b1 and b2 represent bias vectors, and σ represents an activation function. The AdaBN layer is a layer that performs AdaBN (Adaptive Batch Normalization) calculations, such as in equation (8b) or equation (8e).
[0236] <Fourth Embodiment>
[0237] In a situation where there are two or more machines of the same type, consider detecting an abnormality in one of the machines (set as the abnormality detection target machine). To this end, first, based on the normal sound emitted from one or more machines different from the abnormality detection target machine (hereinafter referred to as the normal sound for learning), learn an autoencoder (hereinafter referred to as the first autoencoder) that restores the normal sound emitted from one or more machines different from the abnormality detection target machine (learning stage). Next, using the normal sound emitted from the abnormality detection target machine (hereinafter referred to as the normal sound for adaptive learning), adaptively learn an autoencoder (hereinafter referred to as the second autoencoder) that restores the normal sound emitted from the abnormality detection target machine based on the first autoencoder (adaptive learning stage). And, based on the sound emitted from the abnormality detection target machine (hereinafter referred to as the abnormality detection target sound), determine whether the machine is abnormal (testing stage (abnormality detection stage)).
[0238] The autoencoder learning device 400 learns the first autoencoder based on the normal sound for learning. The autoencoder adaptive learning device 500 adaptively learns the second autoencoder based on the first autoencoder using the normal sound for adaptive learning. The abnormality detection device 600 determines whether the device is abnormal based on the abnormality detection target sound.
[0239] Below, reference Figure 9-10 The autoencoder learning device 400 is described. Fig. 9It is a block diagram showing the structure of the autoencoder learning device 400. Fig.10 4 is a flowchart showing the operation of the autoencoder learning device 400. Fig. 9 As shown, the autoencoder learning device 400 includes an input data generating unit 110, a restored input data estimating unit 420, a loss function calculating unit 430, a parameter updating unit 440, a convergence condition determining unit 450, and a recording unit 490. The recording unit 490 is a structural unit that appropriately records information required for the processing of the autoencoder learning device 400. The recording unit 490 records, for example, a parameter θ of the first autoencoder that is a learning object. As the initial value of the parameter θ, for example, a value generated by using a random number is recorded.
[0240] The autoencoder learning device 400 is connected to a normal sound recording unit for learning 910. Normal sounds for learning prepared in advance are recorded as learning data in the normal sound recording unit for learning 910. As described above, it is preferable to prepare as many normal sounds for learning as possible in advance.
[0241] according to Fig.10 The operation of the autoencoder learning device 400 is described. The input data generating unit 110 generates a normal sound s for learning based on the input. i (i=1,...,N), generate input data x i (i=1, . . . , N) (S110).
[0242] The restored input data estimation unit 420 uses the parameter θ of the first autoencoder to estimate the input data x generated in S110. i (i=1, ..., N), for the input data x i The corresponding restored input data y i (i=1, . . . , N) is estimated (S420). The parameter θ used here is a value being learned.
[0243] The restored input data estimation unit 420 can use the input data x i Calculate the restored input data y i In addition, the calculation formula is as follows.
[0244] [Number 22]
[0245] y i =D(E(x i ,θ E ), θ D )…(16)
[0246] This neural network is the first autoencoder (called neural network NN).
[0247] Here, θ={θE ,θ D}(where θ E and θ D Each represents a parameter of the encoder E and a parameter of the decoder D). In addition, at least one of the neural network constituting the encoder E and the neural network constituting the decoder D is set to include an AdaBN layer. The AdaBN layer refers to a layer that performs the calculation of AdaBN (adaptive batch normalization). That is, the neural network NN includes an AdaBN layer.
[0248] The loss function calculation unit 430 calculates the restored input data y estimated in S420 according to i (i=1, ..., N), calculate the value of the loss function L(θ) used in optimizing the parameter θ of the first autoencoder (S430). The loss function L(θ) can be set to the average of the abnormality defined by the following formula, for example.
[0249] [Number 23]
[0250]
[0251] The parameter updating unit 440 updates the parameter θ of the first autoencoder so as to optimize (minimize) the value of the loss function L(θ) calculated in S430 ( S440 ). In updating the parameter θ, for example, a gradient descent method may be used.
[0252] The convergence condition determination unit 450 determines the convergence condition set in advance as the end condition of the parameter update, outputs the parameter θ updated in S440 when the convergence condition is satisfied, and repeats the processing of S110 to S440 when the convergence condition is not satisfied (S450). As the convergence condition, for example, a condition such as whether the number of execution times of the processing of S110 to S440 has reached a specific number of times can be used.
[0253] In addition, the structural unit including the input data generating unit 110, the restored input data estimating unit 420, the loss function calculating unit 430, the parameter updating unit 440, and the convergence condition determining unit 450 is referred to as the learning unit 405. That is, the learning unit 405 is a structural unit that learns (the parameters θ of) the first autoencoder based on the normal sound for learning.
[0254] According to the invention of this embodiment, it is possible to learn the first autoencoder that can easily adjust the difference between the distribution of learning data and the distribution of test data.
[0255] <Fifth embodiment>
[0256] Below, reference Figure 11-12 The autoencoder adaptive learning device 500 is described. Fig.11is a block diagram showing the structure of the autoencoder adaptive learning device 500 . Fig.12 is a flow chart showing the operation of the self-encoder adaptive learning device 500. Fig.11 As shown, the autoencoder adaptive learning device 500 includes an input data generating unit 110, a parameter updating unit 540, an output unit 550, and a recording unit 490. The recording unit 490 is a structural unit that appropriately records information required for the processing of the autoencoder adaptive learning device 500. The recording unit 490 records, for example, in advance the parameters θ (that is, the learned parameters) of the first autoencoder learned by using the autoencoder learning device 400. The learned parameters become the initial values of the parameters θ of the second autoencoder.
[0257] The autoencoder adaptive learning device 500 is connected to the adaptive learning normal sound recording unit 920. The adaptive learning normal sound recording unit 920 records the adaptive learning normal sound prepared in advance as the adaptive learning data. As described above, the adaptive learning normal sound can be prepared in advance in a smaller amount than the learning normal sound.
[0258] according to Fig.12 The operation of the self-encoder adaptive learning device 500 is described. The input data generating unit 110 generates the adaptive learning normal sound s' according to the input. i (i=1, ..., M), generate input data x' i (i=1, . . . , M) (S110) Here, the number M of normal sounds for adaptive learning is basically an integer not exceeding the number N of normal sounds for learning, but may be an integer exceeding the number N of normal sounds for learning.
[0259] The parameter updating unit 540 uses the input data x′ generated in S110 i (i=1, ..., M), the parameters θ of the second autoencoder are updated (S540). Specifically, for the AdaBN layer included in the first autoencoder (neural network NN) as a layer for calculating adaptive batch normalization, the mean and variance used in its calculation are updated according to the input data x' i (i=1,…,M) The restored input data y' calculated i The mean and variance of can be updated.
[0260] The output unit 550 outputs the parameter θ updated in S540 ( S550 ).
[0261] The structural unit including the input data generating unit 110, the parameter updating unit 540, and the output unit 550 is referred to as the adaptive learning unit 505. That is, the adaptive learning unit 505 is a structural unit that adaptively learns the second autoencoder from the first autoencoder using the normal sound for adaptive learning.
[0262] According to the invention of this embodiment, the second autoencoder can be learned by adjusting the difference between the distribution of learning data and the distribution of test data.
[0263] <Sixth embodiment>
[0264] Below, reference Figure 13-14 The abnormality detection device 600 is described. Fig.13 It is a block diagram showing the structure of the abnormality detection device 600. Fig.14 600 is a flowchart showing the operation of the abnormality detection device 600. Fig.13 As shown, the abnormality detection device 600 includes an input data generating unit 110, an abnormality estimation unit 620, an abnormality determination unit 630, and a recording unit 690. The recording unit 690 is a structural unit that appropriately records information required for processing of the abnormality detection device 600. The recording unit 690, for example, pre-records the parameter θ (that is, the learned parameter) of the second autoencoder learned by using the autoencoder adaptive learning device 500.
[0265] That is, the abnormality detection device 600 performs abnormality detection using the second autoencoder that uses the learned parameters as the learned second autoencoder.
[0266] according to Fig.14 The operation of the abnormality detection device 600 will be described. The input data generating unit 110 generates input data x based on the abnormality detection target sound s as input (S110).
[0267] The abnormality estimation unit 620 estimates the abnormality indicating the degree of abnormality of the machine based on the input data x generated in S110 based on the learned second autoencoder (S620). Figure 15-16 The abnormality estimation unit 620 will be described. Fig.15 is a block diagram showing the structure of abnormality estimation section 620 . Fig.16 is a flowchart showing the operation of the abnormality estimation unit 620. Fig.15 As shown, the abnormality degree estimation unit 620 includes a restored input data calculation unit 621 and an abnormality degree calculation unit 622 .
[0268] according to Fig.16The operation of the abnormality estimation unit 620 is described. The restored input data calculation unit 621 calculates the restored input data y corresponding to the input data x based on the input data x generated in S110 (S621). Specifically, the calculation can be performed using a neural network in which the parameters of the neural network NN are set to the learned parameters of the second autoencoder.
[0269] The abnormality calculation unit 622 calculates the abnormality A(x;θ) associated with the input data x based on the restored input data y calculated in S621 (S622). The abnormality can be calculated by, for example, equation (13).
[0270] The abnormality determination unit 630 generates a determination result indicating whether the machine is abnormal based on the abnormality A(x;θ) estimated in S620 (S630). For example, using equation (3), a determination result indicating abnormality is generated when R=1, and a determination result indicating normality is generated when R=0.
[0271] That is, the abnormality detection device 600 can be said to include: an abnormality degree estimation unit 620, which estimates the abnormality degree indicating the degree of abnormality of the machine based on the first autoencoder that restores the normal sound emitted from one or more machines different from the abnormality detection target machine and the normal sound emitted from the abnormality detection target machine (normal sound for adaptive learning). An example of the association is a second autoencoder that restores the normal sound emitted from the abnormality detection target machine, which is obtained by updating the first autoencoder using the normal sound for adaptive learning.
[0272] According to the invention of this embodiment, by using the second autoencoder that adjusts the difference between the distribution of learning data and the distribution of test data to perform abnormality detection, high-precision abnormality detection can be performed. In other words, a decrease in abnormality detection accuracy can be suppressed.
[0273] Next, the technical background of the seventh to ninth embodiments of the present invention and each embodiment will be described.
[0274] Technical Background
[0275] The embodiment of the present invention uses a normalizing flow to learn a data transformer for domain transformation without paired data. The normalizing flow has the characteristic that it is easier to learn than GAN, and as a result, it can learn more stably than the previous GAN-based domain transformation technology without paired data (StarGAN).
[0276] Hereinafter, the conventional techniques used in the embodiments of the present invention will be described.
[0277] Normalizing Flow
[0278] Normalization flow is a method of finding a distribution that approximates the probability distribution p(x) associated with data generation.
[0279] {f i (z)} i=1 K Suppose there are K transforms with inverse transforms (where f i (z): R D →R D , R is a set of real numbers, D is an integer greater than 1, and K is an integer greater than 1). In addition, f i -1 (z)(i=1,...,K) is set to f i The inverse transform of (z).
[0280] In the normalized flow, for a set of N input data {x i} i=1 N , there are corresponding latent variables {z 0,i} i=1 N , input data x i It can be considered that by using K transformations {f i (z)} i=1 K The latent variable z0 of x is given by (21) to the corresponding latent variable z 0,i The data obtained by transformation.
[0281] [Number 24]
[0282] x=f K (f K-1 (…(f1(z0))…))…(21)
[0283] That is, for i=1, ..., K, the following equation holds.
[0284] [Number 25]
[0285] x i =f K (f K-1 (…(f1(z 0,i ))…))…(twenty one)'
[0286] In addition, it is sometimes expressed as z1=f1(z0), z2=f2(z1), ..., x=f K (zK-1 ).
[0287] Furthermore, assuming that the latent variable {z 0,i} i=1 N For example, a probability distribution q0(z0) that is easy to sample using Monte Carlo is generated based on an isotropic Gaussian distribution. i} i=1 N The probability distribution q(x;θ) (x is a variable representing input data) used can be expressed as follows.
[0288] [Number 26]
[0289]
[0290] Here, z0=f1 -1 (f2 -1 (...(f K -1 (x))...))). In addition, {θ i} i=1 K is the same as the transformation {f i (z)} i=1 K The corresponding parameters, θ = [θ1 T ,θ2 T , ..., θ K T ] T .
[0291] In addition, the probability distribution q0(z0) is not limited to the distribution that is easy to perform Monte Carlo sampling, as long as it is a distribution that is easy to perform rigorous probability density estimation. As an example of a distribution that is easy to perform rigorous probability density estimation, a probability distribution p(x) that satisfies the following conditions can be cited.
[0292] (Condition 1) For a probability distribution p(x), there exists R D A non-negative function g(x)(≥0) on D It becomes p(x)=g(x) / ∫g(x)dx.
[0293] (Condition 2) For the function g(x), it is easy to calculate ∫g(x)dx.
[0294] As a function satisfying condition 2, Gaussian distribution can be cited. On the other hand, as a function not satisfying condition 2, g(x)=exp(sin(x)-x 2 ).
[0295] In the normalization flow, the input data set {x i} i=1 N The parameters θ of the probability distribution q(x; θ) are learned. Then, the probability distribution q(x; θ) using the parameters (referred to as learned parameters) θ obtained through learning is used to approximate the original probability distribution p(x) related to data generation.
[0296] In the normalized stream, various transformations can be used as transformations {f i (z)} i=1 K For example, batch normalization or leaky ReLU (Rectified Linear Unit: normalized linear function) described in reference non-patent document 3 can be used. In addition, the following linear transformation described in reference non-patent document 4 can also be used.
[0297] [Number 27]
[0298] f(z)=LUz…(23)
[0299] Here, L,U∈R D×D are lower triangular matrices and upper triangular matrices respectively. This transformation can be performed by the absolute value of the product of the diagonal components of L and U (that is, |Π i=1 D L ii U ii |) Calculate the absolute value of the Jacobian matrix Therefore, it has the characteristic of being able to easily calculate the probability density q(x; θ) of the input data x (in other words, being able to suppress the calculation cost of the probability density q(x; θ) of the input data x) (refer to formula (22)).
[0300] Below, we simply standardize the batch BN: x--y(x, y∈R D ) is explained. In batch normalization BN, adjustments are made so that the set of input data {x i} i=1 N After the average of the elements of each dimension of becomes 0 and the variance becomes 1, scale transformation and offset transformation are performed. Specifically, y is calculated by the following formula i =BN(x i ).
[0301] [Number 28]
[0302]
[0303] Here, γ and β are scale transformation parameters and offset transformation parameters, respectively, and are parameters to be learned. In addition, ε is a non-negative real number, which is set to a positive real number when it is desired to avoid division by zero, and is set to zero when it is not necessary to avoid division by zero.
[0304] In addition, since the scale transformation parameter γ and the offset transformation parameter β are made explicit, BN(·) is sometimes expressed as BN γβ (·).
[0305] The K transforms used in the normalization flow do not all need to be of the same type. Thus, for example, some types of transforms may be combined, such as performing batch normalization on the transform f1(z) and performing a linear transform on the transform f2(z).
[0306] AdaBN: Adaptive Batch Normalization
[0307] Domain adaptation refers to a technique that, when the distribution of learning data used in model learning is different from the distribution of test data to be processed using the learned model, adjusts the learned model based on the difference in distribution so that the accuracy of the processing using the learned model is not reduced. Here, the set of learning data and the set of test data are domains, and sometimes they are each a learning domain and a test domain.
[0308] There are various domain adaptation methods that can be combined with a deep neural network (DNN). Here, adaptive batch normalization is described (refer to non-patent document 5). Adaptive batch normalization is a method that calculates the mean and variance in batch normalization and adjusts the mean and variance for each domain (refer to formulas (24a) to (24d)). In other words, it is assumed that the calculation based on formulas (24a) to (24c) is performed on the data of the same domain. When actually testing, for the set of input data {x i} i=1 N Calculate the statistics (mean and variance) and use the statistics to output the processing result y through equations (24c) and (24d) i In addition, when the transformation is adaptive batch normalization, it is sometimes expressed as AdaBN:x→y(x, y∈R D ).
[0309] AdaFlow
[0310] AdaFlow is a method that introduces adaptive batch normalization into the normalization flow. Specifically, for the K transformations {f i(z)} i=1 K At least one transformation f i_0 (z), its inverse transformation f i_0 -1 (z) is a method such as adaptive batch normalization. In addition, in the calculation of adaptive batch normalization, the scale transformation and offset transformation, that is, the calculation of formula (24d) can also be omitted. If another statement is used, it can also be expressed as an inverse transformation f i_0 -1 (z) is adaptive batch normalization with γ=1, β=0.
[0311] If AdaFlow is used for learning, probability distributions of multiple domains can be generated from a learned model. In addition, data transformation between multiple domains is also possible.
[0312] The following is a detailed description of AdaFlow. AdaFlow is a neural network that transforms input data x generated from learning data into latent variables z0 (~q(z0)) generated according to probability distribution q0(z0). i (z)} i=1 5 That is, z0 is z0=f1 -1 (f2 -1 (f3 -1 (f4 -1 (f5 -1 (x))))) is obtained.
[0313] The above five transformations {f i (z)} i=1 5 It is defined by the following formula. In addition, for convenience, the transformation f is represented instead of i And its inverse transformation f i -1 (where z4=f5 -1 (x), z3=f4 -1 (z4), z2=f3 -1 (z3), z1=f2 -1 (z2), z0=f1 -1 (z1)).
[0314] [Number 29]
[0315]
[0316] Here, L2, L5∈R D×D is a lower triangular matrix with diagonal components of 1, except for all elements of the upper triangular part L2,ij , L 5,ij (i≥j) is the learning target parameter (that is, parameter θ2 or parameter θ5). D2, D5∈R D×D is a diagonal matrix with diagonal components D 2,ij , D 5,ij (i=j) is the learning target parameter (that is, parameter θ2 or parameter θ5). U2, U5∈R D×D is an upper triangular matrix with diagonal components of 1, except for the lower triangular part, all elements U 2,ij , U 5,ij (i≤j) is a learning object parameter (that is, parameter θ2 or parameter θ5). In addition, α3 (≥0) is a parameter of LeakyReLU, which can be set as a hyperparameter or as a learning object parameter (that is, parameter θ3) (when α3 is set as a learning object, it is called a parametric ReLU (Parametric ReLU) (see non-patent document 6)). In addition, AdaBN γ_4β_4 (·) and AdaBN γ_1β_1 (·) is the above-mentioned adaptive batch normalization, and γ1, β1, γ4, β4 are learning object parameters (that is, parameter θ1 or parameter θ4).
[0317] In addition, the transformation {f i (z)} i=1 5 The absolute values of the Jacobian matrix formulas are each calculated by the following formula (where x=f5(z4), z4=f4(z3), z3=f3(z2), z2=f2(z1), z1=f1(z0)).
[0318] [Number 30]
[0319]
[0320] Here, s4' is the standard deviation of z4 (corresponding to the input data x generated from the learning data), δ is the number of elements below zero in z3 (corresponding to the input data x generated from the learning data), and s1' is the standard deviation of z1 (corresponding to the input data x generated from the learning data). In addition, the absolute value of the matrix formula of the Jacobian for the transformations f4 and f1 is The absolute value of the Jacobian matrix when using inference (that is, when using a learned model) rather than when using learning is recorded.
[0321] In addition, as described above, the probability distribution q0(z0) is set to a probability distribution that is easy to perform rigorous probability density estimation. For example, if the probability distribution q0(z0) is set to a Gaussian distribution N(0, I) with a mean of 0 and a variance of the unit matrix I, the probability distribution q0(z0) is expressed by the following formula.
[0322] [Number 31]
[0323]
[0324] It can be seen that by setting q1(x; θ) to the probability distribution of the input data x generated from the learning data and using equation (22), it is possible to obtain the probability distribution of the latent variable z 0,i The probability density q0(z 0,i ) Calculate the input data x i The probability density q1(x i ;θ).
[0325] Next, the learning method of the parameter θ is described. As with the usual neural network learning, the loss function L(θ) can be used, for example, by gradient descent, momentum SGD (Stochastic Gradient Descent), ADAM (Adaptive Moment Estimation), or a combination thereof. When using a normalized flow, the average of the negative log-likelihood defined by the following formula is often used as the loss function L(θ).
[0326] [Number 32]
[0327]
[0328] In addition, in the above learning, a mini-batch learning method can be used, which uses a set of learning data called a mini-batch as a unit. Here, a mini-batch refers to a plurality of learning data randomly selected from all learning data. The value of the loss function L(θ) is calculated for each mini-batch.
[0329] <Seventh embodiment>
[0330] The probability distribution learning device 1100 learns the model parameters θ of the neural network, which will be based on the domain D of P types j The input data x generated by the data (j=1, ..., P) (hereinafter referred to as domain data) is transformed into a latent variable z0 (~q(z0)) generated according to the probability distribution q0(z0). Using the model parameter θ, the probability distribution q(x; θ) of the input data x can be obtained from the probability distribution q0(z0) (refer to formula (22)). Here, the domain D jSet to include N j Therefore, if the number of domain data contained in the sum of P types of domains is set to N, then N = Σ j N j .
[0331] Below, reference Figure 17-18 The probability distribution learning device 1100 is described. Fig.17 It is a block diagram showing the structure of the probability distribution learning device 1100. Fig.18 1 is a flowchart showing the operation of the probability distribution learning device 1100. Fig.17 As shown, the probability distribution learning device 1100 includes an input data generating unit 1110, a latent variable estimating unit 1120, a loss function calculating unit 1130, a parameter updating unit 1140, a convergence condition determining unit 1150, and a recording unit 1190. The recording unit 1190 is a structural unit that appropriately records information required for the processing of the probability distribution learning device 1100. The recording unit 1190 records, for example, a parameter θ of the probability distribution q(x; θ) of the input data x. As the initial value of the parameter θ, for example, a value generated by using a random number is recorded.
[0332] The probability distribution learning device 1100 is connected to a learning domain data recording unit 1910. In the learning domain data recording unit 1910, domain data s prepared in advance are recorded. i (i=1, ..., N) and is used to identify the domain data s i The set of domain identifiers j of the domains included in i , j) as learning data. In addition, (s i , j) is also called learning domain data.
[0333] according to Fig.18 The operation of the probability distribution learning device 1100 will be described. In S1110, the input data generating unit 1110 generates the learning domain data (s i , j)(i=1,…,N,s i is the domain data, j is the data containing s i domain identifier of the domain) as input, according to the domain data s i , generate input data x i (i=1, ..., N), output input data x i The method of generating input data from domain data may be any method as long as the domain data can be restored from the input data. For example, when the domain data is an image or sound, as a method based on the domain data s i Generated input data x i, it is sufficient to use a specific feature quantity that can restore the image or sound that is the extraction source.
[0334] In S1120, the latent variable estimation unit 1120 converts the input data x generated in S1110 into i (i=1, ..., N) and contains domain data s i The domain identifier j of the domain is used as input, and the parameters θ of the probability distribution q(x; θ) are used to calculate the domain identifier j according to the input data x. i (i=1,…,N), for the input data x i The corresponding latent variable z 0,i (i=1, ..., N) is estimated and outputted. Note that the parameter θ used here is a value being learned.
[0335] Here, the variable x of the probability distribution q(x;θ) represents the domain D according to P types. j The variable x is set to be the input data variable generated by the domain data by using the transformation f i (i=1,…,K,K is an integer greater than 1,for transformation f i There exists an inverse transformation f i -1 ) and the latent variable z0 is represented as x=f K (f K-1 (…(f1(z0))…)).
[0336] Thus, with the input data x i The corresponding latent variable z 0,i Given by the following formula.
[0337] [Number 33]
[0338]
[0339] In addition, the latent variable z 0,i (i=1, ..., N) is assumed to be generated according to the probability distribution q0(z0) of the latent variable z0. The probability distribution q0(z0) has the property that a strict probability density estimation is easy.
[0340] Therefore, the probability distribution q(x;θ) can be expressed by the following equation using the probability distribution q0(z0).
[0341] [Number 34]
[0342]
[0343] Here, θ i is the transformation f i The corresponding parameters are θ = [θ1 T ,θ2 T,…,θ K T ] T .
[0344] The latent variable estimation unit 1120 can be implemented by using a neural network NN that calculates the latent variable z0 based on the input data x. In addition, the calculation formula of the latent variable z0 is as follows.
[0345] [Number 35]
[0346]
[0347] That is to say, the neural network NN takes the input data x as input and transforms it according to the inverse transformation f K -1 , f K-1 -1 , …, f2 -1 , f1 -1 The calculation is carried out in sequence, and the latent variable z0 is finally output.
[0348] For the above transformation f i At least one transformation f among (i=1, ..., K) i_0 (where i0 is an integer satisfying 1≤i0≤K), let its inverse transformation f i_0 -1 It is called Adaptive Batch Normalization.
[0349] In addition, it can also be set as, for the transformation f i The transformation f contained in (i=1, ..., K) i_1 (where i1 is an integer satisfying 1≤i1≤K), its inverse transformation f i_1 -1 It is a linear transformation, and the matrix corresponding to the linear transformation is expressed as the product of a lower triangular matrix and an upper triangular matrix, or the product of a lower triangular matrix, a diagonal matrix, and an upper triangular matrix.
[0350] As a specific transformation f i In the example of (i=1, ..., K), for example, K=5 can be set, and five transformations represented by equations (25a) to (25e) can be used as inverse transformations.
[0351] In addition, the latent variable estimation unit 1120 is used to transform f for which the inverse transform is adaptive batch normalization. i_0 , calculated in each domain in the inverse transform f i_0 -1 The mean and variance used in the calculation of i (i=1, ..., N) calculates the inverse transformation f i_0 -1 The input data z i_0,i (=fi_0-1 -1 (f i_0-2 -1 (…(f K -1 (x i ))…))) of the mean and variance, and in advance (z i_0,i , j) (where j is the input data x i The domain data of the generated source i The domain identifier of the domain is recorded in the recording unit 1190. i_0 -1 The domain D of the input data j The mean and variance of i_0,j and i_0,j 2 In addition, calculate the average m i_0,j and variance s i_0,j 2 The structure is as described later, and is a convergence condition determination unit 1150.
[0352] In S1130, the loss function calculation unit 1130 calculates the potential variable z estimated in S1120 0,i (i=1, ..., N) as input, according to the latent variable z 0,i (i=1, ..., N), calculate the value of the loss function L(θ) used in the optimization of the parameter θ of the probability distribution q(x; θ) and output it. The loss function L(θ) can be defined as the average of the negative log-likelihoods as shown in equation (28). At this time, it is necessary to calculate the input data x i The probability density q(x) of (i=1,…,N) i ;θ), input data x i The probability density q(x i ; θ) can be obtained by using the input data x i The corresponding latent variable z 0,i The probability density q0(z 0,i ) to calculate. For example, when the probability distribution q0(z0) is a Gaussian distribution N(0, I), the latent variable z 0,i The probability density q0(z 0,i ) can be calculated by the following formula:
[0353] [Number 36]
[0354]
[0355] Therefore, using formula (22), we can calculate the latent variable z 0,i The probability density q0(z 0,i ) Calculate the input data xi The probability density q(x i ;θ).
[0356] In S1140, the parameter updating unit 1140 takes the value of the loss function L(θ) calculated in S1130 as input, updates the parameter θ of the probability distribution q(x; θ) so as to optimize (minimize) the value of the loss function L(θ), and outputs it. In the updating of the parameter θ, for example, the gradient descent method can be used.
[0357] In S1150, the convergence condition determination unit 1150 determines the convergence condition that is previously set as the end condition of the parameter update, and outputs the probability distribution q(x; θ) using the parameter θ (the parameter is referred to as the learned parameter) updated in S1140 when the convergence condition is satisfied. At this time, the convergence condition determination unit 1150 uses (z recorded in S1120) i_0,i , j)(i=1,…,N), calculate the inverse transformation f i_0 -1 The domain D of the input data j_0 The average m i_0,j and variance s i_0,j 2 (j=1, ..., P) and output. On the other hand, if the convergence condition is not satisfied, the convergence condition determination unit 1150 repeatedly performs the processing of S1110 to S1140. As the convergence condition, for example, a condition such as whether the number of executions of the processing of S1110 to S1140 reaches a specific number of times can be used. In addition, the parameter θ and the inverse transformation f updated in S1140 can also be output. K -1 (x;θ K ), f K-1 -1 (z K-1 θ K-1 ), …, f2 -1 (z2;θ2), f1 -1 (z1; θ1). Next, we will transform f i_0 -1 The domain D of the input data j The average m i_0,j and variance s i_0,j 2 , called according to the domain D j Statistics calculated from domain data.
[0358] According to the invention of this embodiment, it is possible to learn the domain D according to P types. j The probability distribution q(x;θ) of the input data x generated by the domain data of . By using AdaFlow based on the normalized flow for learning, it is possible to stably perform learning without paired data.
[0359] <Eighth Embodiment>
[0360] The data conversion device 1200 uses the learned model parameters θ learned in the probability distribution learning device 1100 and the inverse transformation f i_0 -1 The domain D of the input data j The average m i_0,j and variance s i_0,j 2 (j=1,…,P), the domain D j_0 The domain data is transformed into domain D j_1 The domain data of (where j0, j1 are integers satisfying 1≤j0, j1≤P, and j0≠j1). j_0 It is called the transformation source domain, and the domain D j_1 It is called the transformation destination region.
[0361] Below, reference Figure 19-20 The data conversion device 1200 will be described. Fig.19 This is a block diagram showing the structure of the data conversion device 1200. Fig. 20 1 is a flow chart showing the operation of the data transformation device 1200. Fig.19 As shown, the data transformation device 1200 includes an input data generating unit 1110, a latent variable calculating unit 1220, an output data calculating unit 1230, a domain data generating unit 1240, and a recording unit 1190. The recording unit 1190 is a structural unit that appropriately records information required for processing by the data transformation device 1200. The recording unit 1190, for example, records in advance the parameter θ of the probability distribution q(x;θ) learned by using the probability distribution learning device 1100 (that is, the learned parameter θ). In addition, the recording unit 1190, for example, records in advance the inverse transformation f i_0 -1 The domain D of the input data j The average m i_0,j and variance s i_0,j 2 (j=1,...,P).
[0362] according to Fig. 20 The operation of the data transformation device 1200 is described below. In S1110, the input data generating unit 1110 receives the transformation source domain D as input. j_0 The domain data s and its domain identifier j0 (called the transformation source domain identifier) are taken as input, and according to the domain data s, the input data x is generated and output.
[0363] In S1220, the latent variable calculation unit 1220 takes the input data x generated in S1110 and the transformation source domain identifier j0 as input, and uses the learned parameters θ and the inverse transformation f i_0 -1 The domain D of the input data j_0 The average m i_0,j_0 and variance s i_0,j_0 2 , based on the input data x, calculate and output the latent variable z0 corresponding to the input data x. The latent variable z0 corresponding to the input data x uses the transformation f used in the probability distribution learning device 1100 i (i=1, ..., K) is calculated by the following formula.
[0364] [Number 37]
[0365]
[0366] At this time, using the learned parameters θ and the inverse transformation f i_0 -1 The domain D of the input data j_0 The average m i _ 0,j_0 and variance s i_0,j_0 2 The latent variable calculation unit 1220 is different from the latent variable estimation unit 1120 of the probability distribution learning device 1100 in this point.
[0367] In S1230, the output data calculation unit 1230 receives the conversion destination domain identifier j1, which is the identifier of the conversion destination domain, and the latent variable z0 calculated in S1220 as input, and uses the learned parameter θ and the inverse transformation f i_0 -1 The domain D of the input data j_1 The average m i_0,j_1 and variance s i_0,j_1 2 , based on the latent variable z0, the output data x' corresponding to the latent variable z0 is calculated and output. The output data x' corresponding to the latent variable z0 uses the transformation f used in the probability distribution learning device 1100 i (i=1, ..., K), calculated by the following formula.
[0368] [Number 38]
[0369] x′=f K (f K-1 (…(f1(z0))…))…(21)″
[0370] At this time, using the learned parameter θ and the inverse transformation f i_0-1 The domain D of the input data j_1 The average m i_0,j_1 and variance s i_0,j_1 2 In addition, the neural network that implements the output data calculation unit 1230 may be a neural network that takes the output of the neural network NN that implements the latent variable calculation unit 1220 as input and takes the input of the neural network NN as output.
[0371] The processing performed by the latent variable calculation unit 1220 and the output data calculation unit 1230 is as follows: Fig.21 shown. Fig.21 The five transformations {f i (z)} i=1 5 , from input data to latent variables, and from latent variables to output data. Here, f1 -1(j_0) , f1 (j_1) For using domain D j_0 The average m i_0,j_0 and variance s i_0,j_0 2 Function, f4 -1(j_0) , f4 (j_1) For using domain D j_1 The average m i_0,j_1 and variance s i_0,j_1 2 function.
[0372] In S1240, the domain data generating unit 1240 takes the output data x' calculated in S1230 as input, and generates a domain D as a conversion destination based on the output data x'. j_1 The transformation of the data is completed and the domain data s' is output.
[0373] In addition, the structural unit including the latent variable calculation unit 1220 and the output data calculation unit 1230 is referred to as the data conversion unit 1205. That is, the data conversion unit 1205 is a structural unit that converts the input data generated from the domain data of the conversion source domain into the output data that is the generation source of the domain data of the conversion destination domain.
[0374] According to the invention of this embodiment, data can be converted between domains.
[0375] <Ninth embodiment>
[0376] In the seventh and eighth embodiments, description has been given of domain data using P types of domains, with P being an integer greater than or equal to 1. Here, description will be given of the data conversion device 1300 in the case where P=2.
[0377] Below, reference Figure 22-23 The data conversion device 1300 will be described. Fig. 22 It is a block diagram showing the structure of the data conversion device 1300. Fig.23 1 is a flow chart showing the operation of the data transformation device 1300. Fig. 22 As shown, the data transformation device 1300 includes an input data generating unit 1310, a latent variable calculating unit 1320, an output data calculating unit 1330, a domain data generating unit 1340, and a recording unit 1190. The recording unit 1190 is a structural unit that appropriately records information required for the processing of the data transformation device 1300. The recording unit 1190 records, for example, in advance the parameter θ of the probability distribution q(x; θ) learned by using the probability distribution learning device 1100 (that is, the learned parameter θ).
[0378] Hereinafter, the transformation source domain is referred to as the first domain, and the transformation destination domain is referred to as the second domain, which are represented by D1 and D2 respectively. i_0 -1 The average m of the first domain D1 of the input data i_0,1 and variance s i_0,1 2 Similarly, in the output data calculation unit 1330, the inverse transformation f is set in advance. i_0 -1 The average m of the second domain D2 of the input data i_0,2 and variance s i_0,2 2 .
[0379] according to Fig.23 The operation of the data transformation device 1300 is described. In S1310, the input data generating unit 1310 receives the domain data s of the first domain as input, generates input data x based on the domain data s, and outputs the input data. Hereinafter, the input data is also referred to as input data corresponding to the domain data of the first domain.
[0380] In S1320 , the latent variable calculation unit 1320 takes the input data x generated in S1310 as input, uses the learned parameter θ, calculates and outputs the latent variable z0 corresponding to the input data x based on the input data x.
[0381] In S1330 , the output data calculation unit 1330 takes the latent variable z0 calculated in S1320 as input, uses the learned parameter θ, and calculates and outputs the output data x′ corresponding to the latent variable z0 based on the latent variable z0 .
[0382] In S1340, the domain data generating unit 1340 takes the output data x' calculated in S1330 as input, generates and outputs domain data s' of the second domain based on the output data x'. Hereinafter, the output data is also referred to as output data corresponding to the domain data of the second domain.
[0383] In addition, the structural unit including the latent variable calculation unit 1320 and the output data calculation unit 1330 is called a data conversion unit 1305. That is, the data conversion unit 1305 is a structural unit that converts input data corresponding to domain data of the first domain into output data corresponding to domain data of the second domain.
[0384] According to the above description, it can be said that the latent variable calculation unit 1320 uses a function with an inverse function that transforms the latent variable into input data (hereinafter referred to as the first function) for calculation, and the output data calculation unit 1330 uses a function with an inverse function that transforms the latent variable into output data (hereinafter referred to as the second function) for calculation. In addition, the first function and the second function are derived from a specific function that transforms the latent variable z0 into the variable x. The specific function is a function obtained by using the sum set of the first domain and the second domain. More specifically, the specific function is learned by using domain data that is an element of the sum set of the first domain and the second domain as learning data, thereby using the transformation f i (i=1,…,K,K is an integer greater than 1,for transformation f i There exists an inverse transformation f i -1 ) is expressed as x = f K (f K-1 (…(f1(z0))…)) (where, for the transformation f i At least one transformation f among (i=1, ..., K) i_0 (1≤i0≤K), its inverse transformation f i_0 -1 is a function obtained by adaptive batch normalization). In addition, the first function is a statistic calculated by using the domain data of the first domain included in the sum set (specifically, the inverse transformation f i_0 -1 The average m of the first domain D1 of the input data i_0,1 and variance s i_0 , 1 2 ) is derived from the above-mentioned specific function, and the second function is a statistic calculated by using the domain data of the second domain included in the sum set (specifically, the inverse transformation f i_0 -1 The average m of the second domain D2 of the input data i_0,2 and variance s i_0,2 2) is a function that is exported from a specific function.
[0385] [Application example: Application to the anomaly detection problem with teachers]
[0386] The data transformation device 1300 can be applied to the problem of taught anomaly detection. Here, taught anomaly detection refers to a framework for performing anomaly detection by learning an anomaly detection model based on a large amount of normal data and a small amount of abnormal data and using the anomaly detection model.
[0387] The more abnormal data there are, the more accurate the model can be learned, but it is difficult to collect a large amount of abnormal data. Therefore, multiple domains that are desired to be used as abnormality detection objects are prepared, as much data as possible is collected, and the model parameter θ is learned based on the data using the probability distribution learning device 1100. In addition, the abnormal data of the first domain is transformed into abnormal data of the second domain different from the first domain using the data conversion device 1300. In this way, the number of abnormal data that is difficult to collect can be increased by artificially creating abnormal data, and as a result, a more accurate abnormality detection model can be learned.
[0388] For example, the first domain is set as a set of sounds emitted from one or more machines different from the abnormality detection target machine, and the second domain is set as a set of sounds of the abnormality detection target machine, and the parameter θ is learned using the probability distribution learning device 1100. Next, the sound of the abnormality detection target machine is generated as the domain data of the second domain based on the abnormal sound emitted from one or more machines different from the abnormality detection target machine as the domain data of the first domain using the data conversion device 1300. The generated sound of the abnormality detection target machine is expected to be an abnormal sound. And, the generated sound of the abnormality detection target machine is used as learning data to learn the abnormality detection model of the second domain.
[0389] In addition, it is considered that most of the sounds emitted are normal sounds. When learning, it is preferred to use as much data as possible, so regardless of normal sounds or abnormal sounds, the sounds of the first domain can be used for learning.
[0390] According to the invention of this embodiment, data can be mutually converted between domains. As a result, for example, abnormal data used in abnormality detection model learning can be efficiently generated.
[0391] <Experimental Results>
[0392] The set of "landscape photos" and the set of "landscape paintings" are used as domains, and Glow described in reference non-patent document 8 is used as the architecture of the normalized flow. The neural network AdaFlow is used to learn by replacing Glow's activation normalization with AdaBN. Image data of 400 landscape photos and landscape paintings are collected respectively, and they are used to make the neural network AdaFlow learn and perform data transformation. Specifically, "landscape photos" are transformed into "landscape paintings", or "landscape paintings" are transformed into "landscape photos".
[0393] (Refer to Non-Patent Literature 8: Diederik P. Kingma, Prafulla Dhariwal, "Glow: Generative Flow with Invertible 1x1 Convolutions", arXiv:1807.03039, https: / / arxiv.org / abs / 1807.03039)
[0394] Through this experiment, that is, using the domain transformation of unpaired data of AdaFlow, it was confirmed that image data with good quality was generated.
[0395] <Supplement>
[0396] Fig.24 The processing in each of the above-mentioned devices can be implemented by causing the recording unit 2020 to read a program for causing the computer to function as each of the above-mentioned devices, and operating the control unit 2010, the input unit 2030, the output unit 2040, etc.
[0397] The device of the present invention, for example, as a single hardware entity, has an input unit that can be connected to a keyboard, etc., an output unit that can be connected to a liquid crystal display, etc., a communication unit that can be connected to a communication device (such as a communication cable) that can communicate with the outside of the hardware entity, a CPU (Central Processing Unit, which can also have a cache memory or registers, etc.), a RAM or ROM as a memory, an external storage device as a hard disk, and a bus connected in a manner that enables data interaction between these input units, output units, communication units, CPU, RAM, ROM, and external storage devices. In addition, as needed, a device (drive) that can read and write recording media such as CD-ROMs, etc., can also be set for the hardware entity. As a physical entity with such hardware resources, there are general-purpose computers, etc.
[0398] In the external storage device of the hardware entity, the program required to realize the above-mentioned functions and the data required for the processing of the program are stored (not limited to the external storage device, for example, the program may be stored in advance in a ROM which is a read-only storage device). In addition, the data obtained by the processing of these programs is appropriately stored in the RAM or the external storage device.
[0399] In the hardware entity, each program stored in an external storage device (or ROM, etc.) and the data required for processing the 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 specific function (each structural element represented by the above-mentioned ... units, ... components, etc.).
[0400] The present invention is not limited to the above-mentioned embodiments and can be appropriately changed within the scope of the purpose of the present invention. In addition, the processing described in the above-mentioned embodiments is not only performed in a time series according to the order of the records, but can also be set to be performed in parallel or individually according to the processing capacity or needs of the device performing the processing.
[0401] As described above, when the processing functions in the hardware entity (device of the present invention) described in the above embodiment are implemented by a computer, the processing contents of the functions that the hardware entity should have are described by a program. And, by executing the program by a computer, the processing functions in the above hardware entity are implemented on the computer.
[0402] The program describing the processing content can be recorded in advance on a recording medium that can be read by a computer. As a recording medium that can be read by a computer, for example, any medium such as a magnetic recording device, an optical disk, an optical magneto-optical recording medium, a semiconductor memory, etc. can also be used. Specifically, for example, a hard disk device, a floppy disk, a magnetic tape, etc. can be used as a magnetic recording device, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only Memory), a CD-R (Recordable) / RW (Rewritable), etc. can be used as an optical magneto-optical recording medium, MO (Magneto-Optical Disc), etc. can be used, and an EEP-ROM (Electronically Erasable and Programmable-Read Only Memory) can be used as a semiconductor memory.
[0403] In addition, the program can be circulated by, for example, selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM on which the program is recorded. Furthermore, the program can also be circulated by storing the program in advance in a storage device of a server computer and forwarding the program from the server computer to other computers via a network.
[0404] For example, a computer that executes such a program first temporarily stores the program recorded in a portable recording medium or the program forwarded from a server computer in its own storage device. In addition, when executing the process, the computer reads the program stored in its own storage device and executes the process according to the read program. In addition, as another execution method of the program, the computer can also directly read the program from the portable recording medium and execute the process according to the program, and then each time the program is forwarded from the server computer to the computer, the process according to the acquired program can be executed one by one. In addition, it can also be configured as follows: without forwarding the program from the server computer to the computer, the so-called ASP (Application Service Provider) type service that realizes the processing function only through its execution instructions and result acquisition is used to execute the above-mentioned process. In addition, in the program in this method, it is set to include information that is used for processing by the electronic computer and is equivalent to the program (although it is not a direct instruction to the computer, it has the nature of data that specifies the processing of the computer, etc.).
[0405] In this embodiment, a hardware entity is configured by executing a specific program on a computer. However, at least a part of the processing contents may be realized in hardware.
[0406] The above-mentioned description of the embodiments of the present invention is presented for the purpose of illustration and description. It is not meant to be comprehensive, nor is it meant to limit the invention to the strict form disclosed. It can be deformed or changed (variation) according to the above teachings. The embodiments are selected and expressed in order to provide the best example of the principles of the present invention, and to enable those skilled in the art to utilize the present invention in a manner suitable for well-thought-out practical use using various embodiments or adding various variations. All such deformations or changes fall within the scope of the present invention determined by the appended claims interpreted in accordance with the scope fairly and legally given fairly.
Claims
1. An abnormality detection device, comprising: A recording unit, recording a parameter θ of a second probability distribution q2(x;θ); an abnormality degree estimation unit that estimates an abnormality degree indicating a degree of abnormality of the device based on the second probability distribution q2(x;θ) and from a sound emitted from the abnormality detection target device, i.e., the abnormality detection target sound; an abnormality degree determination unit, which generates a determination result indicating whether the abnormality detection target machine is abnormal based on the abnormality degree, The second probability distribution q2(x; θ) is a probability distribution representing the distribution of normal sounds emitted from the abnormality detection target machine, obtained by updating the first probability distribution using normal sounds emitted from the abnormality detection target machine, i.e., adaptive learning using normal sounds, wherein the first probability distribution represents the distribution of normal sounds emitted from one or more machines of the same type as the abnormality detection target machine and different from the abnormality detection target machine. The variable x of the second probability distribution q2(x;θ) is a variable representing input data generated based on the normal sound emitted from the abnormality detection target device. The variable x is transformed by using f i and the latent variable z0 is represented as x=f K (f K-1 (…(f1(z0))…)), where i=1,…,K, K is an integer greater than 1, relative to the transformation f i There exists an inverse transformation f i -1 , Let q0(z0) be the probability distribution of the latent variable z0, The probability distribution q2(x;θ) of the input data x is obtained by using the latent variable z0=f1 corresponding to the input data x. -1 (f2 -1 (…(f K -1 The probability distribution q0(z0) of (x))…)) is calculated.
2. The abnormality detection device according to claim 1, characterized in that: For the transformation f i At least one of the transforms whose inverse transform is adaptive batch normalization, where i=1,…,K.
3. The abnormality detection device according to claim 1, characterized in that: The second probability distribution q2(x;θ) is represented by the following formula: [Number 39] Where z1 = f1(z0; θ1), z2 = f2(z1; θ2), ..., x = f K (z K-1 θ K ), θ i is the same as the transformation f i The corresponding parameters, θ = [θ1 T ,θ2 T ,…,θ K T ] T , where i=1,…,K.
4. The abnormality detection device according to claim 2 or 3, characterized in that: For the transformation f i At least one transformation among them, whose inverse transformation is a linear transformation, where i = 1, ..., K, The matrix corresponding to the linear transformation is expressed as the product of a lower triangular matrix and an upper triangular matrix, or the product of a lower triangular matrix, a diagonal matrix, and an upper triangular matrix.
5. An abnormality detection device, comprising: A recording unit, recording the parameter θ of the second autoencoder; an abnormality degree estimation unit that estimates an abnormality degree indicating a degree of abnormality of the device based on a sound emitted from the abnormality detection target device, i.e., an abnormality detection target sound, based on the second autoencoder; an abnormality degree determination unit, which generates a determination result indicating whether the abnormality detection target machine is abnormal based on the abnormality degree, The second autoencoder is an autoencoder that restores the normal sound emitted by the abnormality detection target machine by updating the first autoencoder using the normal sound emitted by the abnormality detection target machine, i.e., the normal sound for adaptive learning. The first autoencoder restores the normal sound emitted by one or more machines of the same type as the abnormality detection target machine that are different from the abnormality detection target machine.
6. The abnormality detection device according to claim 5, characterized in that: The first autoencoder is a neural network including an AdaBN layer for performing the adaptive batch normalization calculation. Constructed through the network.
7. A computer program product, comprising a computer program for causing a computer to function as the abnormality detection device according to any one of claims 1 to 6.
