JP2019036112A - Abnormal sound detector, abnormality detector, and program - Google Patents

Abstract

To provide an abnormality detection technology capable of suppressing a false detection using an abnormality detection model generated assuming a uniform distribution or a generalized normal distribution as a probability distribution of regression errors.SOLUTION: An abnormality detector is comprised of: a regression error calculation unit for calculating a regression error t- f(x) of input data x which is obtained from a pair (x, t) of correct data t for the input data x and the input data x while using a regression model f(.); and an abnormality detection unit that generates an abnormality detection result for the input data x from the regression error t- f(x) while using an abnormality detection model; wherein: the abnormality detection model is to output either a first value or a second value as an abnormality degree from the regression error of the input data; and the abnormality detection unit generates an abnormality detection result for indicating that the input data is normal when the abnormality degree is the first value, and generates an abnormality detection result for indicating that the input data is abnormal when the abnormality degree is the second value.SELECTED DRAWING: Figure 3

Description

  The present invention relates to an abnormality detection technique for detecting an abnormal state from an input, such as detecting a failure from an operation sound of a machine.

  For example, it is important from the viewpoint of business continuity to find a machine failure before the failure or to quickly find a failure after the failure. As a method for saving labor, there is a technical field called anomaly detection in which “abnormality”, which is a deviation from a normal state, is detected from data acquired using a sensor by an electric circuit or a program. In particular, an apparatus that uses a sensor that converts sound into an electrical signal, such as a microphone, is called abnormal sound detection.

  Conventionally, there is a method for detecting an abnormality using a regression model. The outline will be described below. First, a regression model is generated from sound data of machine operating sounds that are considered to be in a normal state with no failure. Next, the output value (output vector) by the above regression model is calculated for the sound data of the operation sound whose abnormality / normality is not known, and the difference between the output value and the actual value (correct value) (output vector and correct vector) The regression error is calculated. The “abnormality” is calculated on the basis of the likelihood in the probability distribution followed by the regression error determined in advance. Finally, it is determined whether or not there is an abnormality based on the threshold based on the degree of abnormality.

  As such a method, a Gaussian distribution is assumed as a probability distribution followed by a regression error, linear regression is used as a regression model, and a least square method is used for regression model parameter estimation (Non-patent Document 1), and a linear model is used as a regression model. Some use a neural network instead of regression (Non-Patent Document 2).

Takeshi Ide, “Introductory Anomaly Detection by Machine Learning-Practical Guide by R”, Corona, pp.161, 2015. Pankaj Malhotra, Lovekesh Vig, Gautam Shroff, Puneet Agarwal, “Long Short Term Memory Networks for Anomaly Detection in Time Series”, ESANN 2015 proceedings, pp.89-94, 2015.

  Even if the sound is classified as normal, it has a complicated regularity in the time direction, such as a sound that changes abruptly, such as a sound corresponding to the rotation or stoppage of a motor, or a sound corresponding to a machine operation sequence. There is something to have. That is, even if it is a normal sound, it is not necessarily a monotonous sound. In the conventional method, when the normal sound is a monotonous sound and the abnormal sound is a sound that changes suddenly, the abnormality can be accurately detected. Abnormalities cannot be detected well when there is a sudden change or when there is a difficulty in predicting, such as lack of regularity, and the normal sound's regression error tends to be large. Despite this, there was a phenomenon that the degree of abnormality increased instantaneously.

  In the conventional linear regression or neural network method, learning is performed by the least square method or the mean square error. This is equivalent to assuming that the regression error between the output vector of the regression model and the correct vector follows a Gaussian distribution. The probability density function of the Gaussian distribution assumes that the absolute value has a positive value even for very large positive and negative values, that is, it can be a probability, and when learning under such assumption, Although only normal data is included in the learning data, only a part of the patterns that are difficult to predict are learned so that the prediction may be omitted, that is, the likelihood is lowered. Usually, the degree of abnormality is designed to increase as the likelihood decreases, so the phenomenon that the degree of abnormality described above instantaneously increases occurs, and in abnormality detection that performs abnormality determination by dividing the degree of abnormality by a threshold value It will cause false detection.

  This problem of false detection is not limited to sound data, and can generally occur when a Gaussian distribution is assumed as a probability distribution of a regression error generated from a regression model.

  Therefore, the present invention aims to provide an anomaly detection technique that can suppress false detection by using an anomaly detection model that is generated assuming a uniform distribution or a generalized normal distribution as a probability distribution of regression errors. And

In one embodiment of the present invention, θ ^ is a set (x _n , t _n ) (x) of N (N is an integer of 1 or more) normal input data x _n and correct data t _n for the input data x _n . 1 ≦ n ≦ N) is used as a regression model parameter, f _{θ ^} (・) is a regression model defined by the regression model parameter θ ^, and using the regression model f _{θ ^} (・), Using a regression error calculation unit that calculates a regression error t- f _{θ ^} (x) of the input data x from a set (x, t) of the input data x and the correct answer data t with respect to the input data x, and an abnormality detection model An abnormal sound detection device including an abnormality detection unit that generates an abnormality detection result for the input data x from the regression error t-f _{θ ^} (x), wherein the abnormality detection model is based on a regression error of the input data. The first value or the second value is output as the degree of abnormality, and the abnormality detection unit has the degree of abnormality as the first value. If, it generates an abnormality detection result indicating that the input data is normal, if the error probability is the second value, to generate the abnormality detection result indicating that the input data is abnormal.

In one embodiment of the present invention, θ ^ is a set (x _n , t _n ) (x) of N (N is an integer of 1 or more) normal input data x _n and correct data t _n for the input data x _n . 1 ≦ n ≦ N) is used as a regression model parameter, f _{θ ^} (・) is a regression model defined by the regression model parameter θ ^, and using the regression model f _{θ ^} (・), Using a regression error calculation unit that calculates a regression error t- f _{θ ^} (x) of the input data x from a set (x, t) of the input data x and the correct answer data t with respect to the input data x, and an abnormality detection model An abnormal sound detection device including an abnormality detection unit that generates an abnormality detection result for the input data x from the regression error t-f _{θ ^} (x), wherein p is an integer of 3 or more and θ is the regression model a regression model parameters generated in the process of learning the parameters theta ^, n pieces of the regression error _{t n - f θ (x n} ) loss for (1 ≦ n ≦ n) is, p Nowaayama The abnormality detection model outputs a negative logarithmic likelihood corresponding to the p-th power as an abnormality degree from the regression error of the input data, and the abnormality detection unit sets a threshold value. And using the degree of abnormality to generate the abnormality detection result.

In one embodiment of the present invention, θ ^ is a set (x _n , t _n ) (x) of N (N is an integer of 1 or more) normal input data x _n and correct data t _n for the input data x _n . 1 ≦ n ≦ N) is used as a regression model parameter, f _{θ ^} (・) is a regression model defined by the regression model parameter θ ^, and using the regression model f _{θ ^} (・), Using a regression error calculation unit that calculates a regression error t- f _{θ ^} (x) of the input data x from a set (x, t) of the input data x and the correct answer data t with respect to the input data x, and an abnormality detection model An anomaly detection unit including an anomaly detection unit that generates an anomaly detection result for the input data x from the regression error t-f _{θ ^} (x), wherein the anomaly detection model is based on the regression error of the input data, Either the first value or the second value is output as the degree of abnormality, and the abnormality detection unit is configured to output the abnormality value when the degree of abnormality is the first value. Generates an abnormality detection result indicating that the input data is normal, if the error probability is the second value, to generate the abnormality detection result indicating that the input data is abnormal.

  According to the present invention, it is possible to suppress erroneous detection by using an abnormality detection model generated assuming a uniform distribution or a generalized normal distribution as a probability distribution of regression errors.

FIG. 3 is a block diagram showing an example of the configuration of the abnormality detection model generation apparatus 100. 5 is a flowchart showing an example of the operation of the abnormality detection model generation device 100. The block diagram which shows an example of a structure of the abnormal sound detection apparatus 200. FIG. The flowchart which shows an example of operation | movement of the abnormal sound detection apparatus 200. The block diagram which shows an example of a structure of the abnormality detection model production | generation apparatus 300. FIG. The flowchart which shows an example of operation | movement of the abnormality detection model production | generation apparatus 300. FIG. The block diagram which shows an example of a structure of the abnormality detection apparatus 400. FIG. 7 is a flowchart showing an example of the operation of the abnormality detection apparatus 400.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the structure part which has the same function, and duplication description is abbreviate | omitted.

<Technical background>
In anomaly detection using a regression model, a Gaussian distribution is assumed as a probability distribution followed by a regression error, and regression model learning by the least square method is often used. However, here, for the Gaussian distribution, a distribution that is far from the center with a certain value as a boundary, that is, a probability density value with a small absolute value is small is assumed. Examples of such distribution include a uniform distribution and a generalized normal distribution that is an approximation thereof. If a uniform distribution is assumed, the regression model is learned by the maximum absolute value method. When a generalized normal distribution is assumed, the regression model is learned by a method having a power that is larger than the square, such as a least-fourth method or a least-octet method. Furthermore, by deriving an expression of “abnormality” that expresses the anomaly of the input data as a numerical value, which is expressed in a normal information amount format corresponding thereto, and using the abnormality degree according to the expression, the above-described problem of false detection can be solved. Solve.

  The uniform distribution is a continuous probability distribution having the upper and lower limits of a random variable section as parameters. If it is one-dimensional, it has two parameters indicating an upper limit value and a lower limit value. If it is multidimensional, it has parameters indicating an upper limit value and a lower limit value for each dimension. Unlike the Gaussian distribution, the uniform distribution cannot be realized outside the interval specified by the upper limit value and the lower limit value, that is, the probability is zero.

  In general, in regression problems, it is often better to assume a distribution that allows outliers. However, assuming a uniform distribution, all of the training data can be treated as normal data, and the regression error In the field of anomaly detection where there is little need to lower it, it can be said that it is appropriate to have the opposite idea (assuming a uniform distribution). Assuming a uniform distribution as the distribution that the regression error follows, when learning the regression model, if the range where the regression error exists exceeds a predetermined bounded area, a very large penalty will be incurred. If it falls within, it will be the same penalty. Therefore, although the regression error in the entire learning data becomes large, even if it is a normal pattern that is difficult to predict, the prediction will be drastically lost and the likelihood will not decrease and the abnormality will not increase. .

  Hereinafter, an abnormality detection model generation procedure when it is assumed that the regression error follows a uniform distribution will be described. Here, the abnormality detection model is a model that outputs an abnormality degree with respect to input data. This generation procedure is roughly composed of two procedures: learning of a regression model and generation of an abnormality detection model.

(1: Learning of regression model)
Making the Gaussian distribution a regression error distribution means adopting the l-2 norm as the norm and the Euclidean distance as the distance. On the other hand, to make the uniform distribution a regression error distribution is to adopt l-∞ norm (uniform norm) and Chebyshev distance (chessboard distance) as the distance. Optimization methods for minimizing the l-∞ norm are known, and are described in, for example, Reference Non-Patent Document 1 and Reference Non-Patent Document 2.
(Reference Non-Patent Document 1: Yongduek Seo, Richard Hartley, "A fast method to minimize L ∞ error norm for geometric vision problems", IEEE 11th International Conference on Computer Vision (ICCV) 2007, Nov. 2007.)
(Reference Non-Patent Document 2: Christian Clason, “L ^∞ fitting for inverse problems with uniform noise”, Inverse Problems, Vol.28, No.10, pp.1-23, 2012.)
Regression models expressed using linear regression or neural networks can be learned using these optimization methods so that the regression error due to the l-∞ norm is reduced.

(2: Generation of anomaly detection model)
Consider defining the degree of abnormality based on the likelihood of regression error. If each dimension is independent, the range in which the likelihood is not zero in the uniform distribution is expressed as a section in one dimension, a rectangle in two dimensions, a rectangular parallelepiped in three dimensions, and generally a super rectangular parallelepiped. Further, when independence between dimensions is not required, a circle (in the case of two dimensions) or a sphere (in the case of three dimensions) can be set as a range in which the likelihood is not zero. If it is two-dimensional, the probability density function indicating the distribution of the regression error is represented by a cylindrical shape.

Let θ be a regression model parameter by linear regression or a neural network. In general, θ is composed of a plurality of parameters, and in the case of a neural network, it is a parameter expressing a weight or a bias. In addition, since the regression model can be regarded as a function that outputs output data from input data, a regression model whose regression model parameter is θ is represented as f _θ (·). Also, x _{n is} input data, and t _n is correct data for the input data x _n . For example, the input data x _n is a sound wave sample point, a power spectrum, and a filter bank output. In general, input data and output data are often expressed as vectors. In this case, the input data and output data are also referred to as an input vector and an output vector, respectively. The output data from the regression model can be expressed as f _θ (x _n ). The set of regression errors for N (x _n , t _n ) (1 ≦ n ≦ N) is {t _n −f _θ (x _n ) | n∈ {1,..., N}}. When a vector generated from sound is used as input data, an autoregressive model may be used as the regression model, for example, t _n = x _{n + 1} . In general, t _n can be a vector generated from x ₁ ,..., X _N.

The set of correct data t _n is the learning data and the normal input data x _n for the input data x _n (x _n, t _n) the set of regression errors from _{{t n -f θ (x n} ) | n∈ {1, ..., N}}, and the parameters are obtained by maximum likelihood estimation for a uniform distribution in which the range where the value does not become zero is a super rectangular parallelepiped. If there is a regression error t _n −f _θ (x _n ) that exceeds the range defined by the parameters for one dimension, the likelihood is zero. Further, the integral value of the probability density function of uniform distribution on the super rectangular parallelepiped is 1. Here, assuming that a hypercube contains all the regression errors t _n -f _θ (x _n ) (n∈ {1,…, N}), if the range of the hypercube is expanded even a little, 1 is larger Divided by the volume of the super rectangular parallelepiped, the likelihood decreases. Therefore, 2d values representing the range representing the minimum hypercube T including the set of regression errors {t _n −f _θ (x _n ) | n∈ {1,…, N}} are one in this case. This is the maximum likelihood estimation solution for the uniform distribution parameter. If the maximum likelihood estimation solution is not taken into account, the bounded region B including the smallest super rectangular parallelepiped can be set as a range in which the value does not become zero.

  The degree of abnormality can be defined as negative log likelihood (amount of information) based on the likelihood, for example. The likelihood for a regression error corresponding to a pair of correct and correct input vectors and correct vectors is constant if the regression error is included in the super-rectangular T, and 0 otherwise. Corresponding to the degree, constant and infinity can be used as the abnormal degree of the regression error. Therefore, when a Gaussian distribution is assumed as a distribution of regression errors, a threshold value is often used for calculating the degree of abnormality. However, when a uniform distribution is assumed, the degree of abnormality is binary, so it is necessary to use a threshold value. However, if the regression error is included in the super rectangular parallelepiped, the degree of abnormality is constant, so that it is normal, and if it is not included in the super rectangular parallelepiped, the degree of abnormality is infinite, so it can be handled as abnormal.

<First embodiment>
Here, an abnormality detection model generation apparatus and an abnormal sound detection apparatus based on an autoregressive model using vectors generated from sound as learning data and abnormality detection data will be described.

(Abnormality detection model generation device 100)
Hereinafter, the abnormality detection model generation device 100 will be described with reference to FIGS. FIG. 1 is a block diagram illustrating a configuration of the abnormality detection model generation device 100. FIG. 2 is a flowchart showing the operation of the abnormality detection model generation device 100. As shown in FIG. 1, the abnormality detection model generation device 100 includes a sound data generation unit 110, a regression model parameter learning unit 120, an abnormality detection model generation unit 130, and a recording unit 190. The recording unit 190 is a component that appropriately records information necessary for processing of the abnormality detection model generation device 100.

  The abnormality detection model generation apparatus 100 receives the sound waveform of the learning sound as an input. The learning sound is, for example, an operating sound of a machine that operates normally.

The operation of the abnormality detection model generation device 100 will be described with reference to FIG. The sound data generation unit 110 generates the sound data x _n (1 ≦ n ≦ N) of the learning sound from the sound waveform of the learning sound (S110). Specifically, the sound data x _n (1 ≦ n ≦ N) of the learning sound is generated as follows. The sound waveform of the learning sound is AD (analog-digital) converted at an appropriate sampling frequency to generate quantized waveform data. The waveform data quantized in this way may be used as the sound data x _n (1 ≦ n ≦ N) of the learning sound, and the data in which one-dimensional values are arranged in time series as it is, concatenation of multiple samples, discrete Fourier transform, The sound extraction data x _n (1 ≤ n ≤ N) that has been subjected to feature extraction processing that uses filter bank processing etc. to expand in multiple dimensions may be used. The sound data x _n (1 ≦ n ≦ N) of the learning sound may be preprocessed such as normalizing the width. In general, the sound data _xn of the learning sound is expressed as a vector.

The regression model parameter learning unit 120 learns the regression model parameter θ ^ using the sound data x _n (1 ≦ n ≦ N) of the learning sound (S120). For example, the correct answer data t _n of the sound data x _n can be learned regression model parameter theta ^ using an autoregressive model to x _{n + 1.}

For learning the regression model, linear regression, or a neural network such as a recurrent neural network or an LSTM network (Long Short Term Memory Networks) can be used. When linear regression is used, a regression model parameter θ that minimizes the uniform norm can be obtained using a numerical solution such as linear programming. When using a neural network, a loss function that returns a scalar value for the regression error set {t _n -f _θ (x _n ) | n∈ {1,…, N}} is expressed as Err [{t _n -f _θ (x _n ) | n∈ {1,..., N}}], and d = 1 (where d is the vector dimension of correct data t _n ), the maximum absolute value error expressed by the following equation is used.

This is used in place of the square sum error function of the following equation, which is widely used in regression model learning.

When the maximum absolute value error is viewed as a function, non-differentiable points are included. Therefore, when using a learning algorithm such as the error back propagation method, it is preferable to use a subgradient. However, since the learning in this way may not be stable, suitable smooth function (e.g., C ^k-class functions) may be approximated maximum absolute error in.

  The end of learning is determined when a predetermined number of repetitions has been executed.

  In the case of multi-dimensions (that is, d> 1), the maximum value of the value of Equation (1) obtained for each dimension is the maximum absolute value error.

  Alternatively, p may be an integer of 3 or more, and the p-th power sum error of the following equation may be used instead of the maximum absolute value error. That is, the maximum absolute value error is approximated by a p-th power sum error.

Further, p _1, ..., a p _m as an integer of 3 or more, respectively, p ₁ sum error, ..., may be used instead that the p _m square sum error is a linear combination.

Further, learning may be started from an appropriate value of p, and p may be increased stepwise. In other words, p ₁ , ..., p _m is an integer of 3 or more that satisfies p ₁ <... <p _m respectively, and after p ₁ sum-of-power errors, ..., p _m-th power sum errors and an appropriate number of iterations, It may be switched.

  For terms that change only monotonically with respect to the error, such as the power root and the average at the number N of training data sets, the calculation can be omitted because it ultimately results in an optimization problem. Good.

  In the case of multi-dimensions (that is, d> 1), the sum of errors for each dimension may be simply set as the p-th power sum error assuming that each dimension is independent. Also, for example, the error for each dimension may be weighted in consideration of the importance for each dimension, such as dividing the error for each dimension by a constant determined for each dimension.

  The learning of the regression model parameter θ described above can be formulated as the following optimization problem.

Note that the value of the loss function Err [{t _n −f _θ (x _n ) | n∈ {1,…, N}}] is expressed as N regression errors t _n −f _θ (x _n ) (1 ≦ This is a loss for n ≦ N).

The abnormality detection model generation unit 130 generates an abnormality detection model using the regression model parameter θ ^ learned in S120 (S130). Here, the abnormality detection model outputs either the first value or the second value indicating the degree of abnormality of the sound data of the input sound. Specifically, the abnormality detection model is generated as follows. First, the abnormality detection model generation unit 130 calculates a regression error t _n −f _{θ ^} (x _n ) from (x _n , t _n ) (1 ≦ n ≦ N) using the regression model parameter θ ^, Generate a regression error set {t _n -f _{θ ^} (x _n ) | n∈ {1, ..., N}}. In the case of t _n = x _{n + 1} , t _n −f _{θ ^} (x _n ) = x _{n + 1} −f _{θ ^} (x _n ).

Next, the anomaly detection model generation unit 130 generates a bounded region B of the d-dimensional real vector space including the regression error set {t _n −f _{θ ^} (x _n ) | n∈ {1,..., N}}. Generate. For example, the regression error set {t _n −f _{θ ^} (x _n ) | n∈ {1,…, N, which is a maximum likelihood estimation solution with uniform distribution, as described in <Technical Background>. }} Can be the smallest super rectangular parallelepiped. This super rectangular parallelepiped can represent the range using 2d parameter values. In this case, for example, by obtaining the maximum and minimum values of the elements of the regression error set {t _n -f _{θ ^} (x _n ) | n∈ {1,…, N}} for each dimension, It may be obtained, or the hypercube B may be obtained so as to be further reduced in consideration of the rotation of the coordinate axes, or the regression error set {t _n -f _{θ ^} (x _n ) | n∈ {1,… , N}} are averaged, and the coordinates are regarded as the center of the super rectangular parallelepiped, and the super rectangular parallelepiped B can be obtained by calculating only the width of d, the average, the super rectangular parallelepiped center, etc. The super rectangular parallelepiped B may be obtained by setting the parameter defining the center to 0 vector and obtaining only d widths. In addition, the region B may be a minimum sphere (high-dimensional ellipsoid) including a regression error set as a hypersphere or a high-dimensional ellipsoid instead of the super-cuboid. Further, the region B may be manually adjusted according to a request for a task in which actual abnormal sound detection is performed. Further, as described in <Technical Background>, a region larger than the minimum region including the regression error set may be set as the region B.

  In S120, when approximation by p-th power sum error is used, a parameter (for example, standard deviation) that defines the spread of the generalized normal distribution corresponding to the p-th power is calculated by maximum likelihood estimation, and region B is obtained. Also good. The generalized normal distribution corresponding to the p-th power has a domain such as a rectangular parallelepiped independent of each dimension, and a domain such as a spherical symmetry without independence of each dimension. Thus, the area B may be obtained.

  Finally, the abnormality detection model generation unit 130 generates the following abnormality detection model.

[Anomaly detection model]
A set of correct data s _m for sound data y _m of the sound data y _m and the input sound of the input sound (y _m, s _m), the abnormality detection model is the regression error s _m of the sound data y _m of the input sound - f respect _{θ ^} (y _m), the first value if _{s m -f θ ^ (y m} ) ∈B is established, otherwise outputting a second value as a degree of abnormality. However, the first value and the second value can be arbitrary values as long as they are different values.

For example, if the correct answer data s _m of the sound data y _m using an autoregressive model to y _{m + 1,} the regression error _{s m -f θ ^ (y m} ) = y m + 1 -f θ ^ (y m) It becomes. In general, when dealing with time series data, calculate the regression error using the necessary data in the past (y ₁ , ..., y _m-1 ) or the future (y _{m + 1} , ..., y _M ) be able to.

  For example, the negative logarithm of the value obtained by dividing 1 by the hypercubic volume (likelihood) may be the first value, infinity may be the second value, or simply 1 (value indicating normality) is the first value. , 0 (value indicating abnormality) may be set as the second value.

In S120, when approximation by p-th power sum error is used, the information amount (negative log likelihood) of the generalized normal distribution corresponding to p-th power can be set as the degree of abnormality. Specifically, if the regression error is one-dimensional, use the constant α and the natural logarithm ln to calculate the sum of | s _m -f _{θ ^} (y _m ) | / α raised to the p power and ln (α) However, this may be used as the degree of abnormality. In the case of multi-dimensions, different α may be used for each dimension, and after calculating the above sum, the sum may be taken as the degree of abnormality. α is a parameter that determines how the error spreads when normal in each dimension when the regression error is multi-dimensional, and may be determined by humans for each dimension, or the sound data used for model learning You may obtain | require using the maximum likelihood estimation etc. from another sound data for hyperparameter determination, and the regression error set produced | generated from the learned model.

(Abnormal sound detection device 200)
Hereinafter, the abnormal sound detection apparatus 200 will be described with reference to FIGS. FIG. 3 is a block diagram illustrating a configuration of the abnormal sound detection device 200. FIG. 4 is a flowchart showing the operation of the abnormal sound detection apparatus 200. As shown in FIG. 3, the abnormal sound detection device 200 includes a sound data generation unit 110, a regression error calculation unit 220, an abnormality detection unit 230, an abnormality detection result integration unit 240, and a recording unit 290. The recording unit 290 is a component that appropriately records information necessary for processing of the abnormal sound detection device 200. For example, the recording unit 290 records the regression model parameter θ ^ generated by the abnormality detection model generation apparatus 100 and the abnormality detection model.

  The abnormal sound detection apparatus 200 receives the sound waveform of the input sound as input. The input sound is, for example, an operation sound of a machine whose normality / abnormality is unknown. That is, the input sound is a sound that is an object of abnormality detection.

The operation of the abnormal sound detection apparatus 200 will be described with reference to FIG. The sound data generation unit 110 generates sound data y _m (1 ≦ m ≦ M) of the input sound from the sound waveform of the input sound (S110).

Regression error calculating unit 220, using the regression model _f θ _^ (·), the regression error s _m of the sound data y _m from the input sound sound data _{y m (1 ≦ m ≦ M} ) - f θ ^ (y m ) Is calculated (S220). In the previous example, s _m = y _{m + 1} .

The abnormality detection unit 230 generates an abnormality detection result for the sound data y _m from the regression error s _m −f _{θ ^} (y _m ) using the abnormality detection model (S230). Specifically, an abnormality detection result is generated as follows. First, the anomaly detection unit 230 uses an anomaly detection model to set the first value when s _m −f _{θ ^} (y _m ) ∈B holds, and the second value otherwise. Calculate as Next, the abnormality detecting unit 230, if the error probability is a first value, the abnormality detection result indicating that the sound data y _m is normal, if the error probability is the second value, the sound data y An abnormality detection result indicating that _m is abnormal is generated.

  When the abnormality detection model uses the information amount (negative log likelihood) of the generalized normal distribution corresponding to the p-th power as the abnormality degree, the abnormality detection unit 230 generates an abnormality detection result from the abnormality degree using the threshold value. To do.

The abnormality detection result integration unit 240 generates an abnormality detection result of the input sound from the abnormality detection result of the sound data y _m (1 ≦ m ≦ M) (S240). Out of the abnormal detection results of M sound data y _m (1 ≤ m ≤ M), if the detection results of a predetermined ratio or higher (greater than the predetermined ratio) are normal, the abnormal detection result of the input sound is normal In other cases, the input sound abnormality detection result may be made abnormal.

  According to the invention of this embodiment, erroneous detection can be suppressed by using an abnormality detection model generated assuming a uniform distribution or a generalized normal distribution as a probability distribution of regression errors. In the conventional method that assumed a Gaussian distribution as the probability distribution of the regression error, there was a case where the degree of abnormality was high for normal input data. However, the probability distribution of the regression error is a uniform distribution or a generalized normal distribution. As a result, occurrence of such a case is suppressed, and false detection can be reduced.

<Second embodiment>
In the first embodiment, the abnormality detection model generation device and the abnormal sound detection device have been described using the regression model for sound data. However, data that can be used for regression model learning is not limited to sound data. . Therefore, in the second embodiment, an example using a regression model for general data including sound data will be described.

(Abnormality detection model generation device 300)
Hereinafter, the abnormality detection model generation apparatus 300 will be described with reference to FIGS. FIG. 5 is a block diagram illustrating a configuration of the abnormality detection model generation device 300. FIG. 6 is a flowchart showing the operation of the abnormality detection model generation device 300. As shown in FIG. 5, the abnormality detection model generation device 300 includes a regression model parameter learning unit 120, an abnormality detection model generation unit 130, and a recording unit 190. The recording unit 190 is a component that appropriately records information necessary for processing of the abnormality detection model generation device 300.

Abnormality detection model generating device 300, the set of correct data t _n for the input data x _n and the normal input data _{x n (x n, t n} ) (1 ≦ n ≦ N, N is an integer of 1 or more) the training data As input.

The operation of the abnormality detection model generation device 300 will be described with reference to FIG. The regression model parameter learning unit 120 learns the regression model parameter θ ^ using N sets (x _n , t _n ) (1 ≦ n ≦ N) (S120). Specifically, the regression model parameter θ ^ is learned as follows. Let θ be the regression model parameter generated in the course of learning the regression model parameter θ ^. The loss for N regression errors t _n -f _θ (x _n ) (1 ≦ n ≦ N) is defined by the maximum absolute value error (Equation (1)), and the regression model parameter θ is set so as to minimize the loss. Learn ^. Or, let p be an integer greater than or _equal to 3, and define the loss for N regression errors t _n -f _θ (x _n ) (1 ≤ n ≤ N) by p-power sum error (Equation (2)). The regression model parameter θ ^ is learned so as to minimize. In this case, when learning the regression model parameter, the value of p may be increased stepwise.

The abnormality detection model generation unit 130 generates an abnormality detection model using the regression model parameter θ ^ learned in S120 (S130). Specifically, the abnormality detection model is generated as follows. Let f _{θ ^} (•) be the regression model defined by the regression model parameter θ ^. First, the abnormality detection model generation unit 130 calculates a regression error t _n −f _{θ ^} (x _n ) from (x _n , t _n ) (1 ≦ n ≦ N). Next, the abnormality detection model generation unit 130 generates a bounded region B including N regression errors t _n −f _{θ ^} (x _n ) (1 ≦ n ≦ N). Finally, the abnormality detection model generation unit 130 generates the following abnormality detection model.

[Anomaly detection model]
A set of correct data t (x, t) with respect to the input data x and the input data x, anomaly detection model, regression error _t- f θ of the input data x _^ with respect to _{(x), t- f θ ^} (x The first value is output as the degree of abnormality when) ∈B holds, and the second value is output otherwise. However, the first value and the second value can be arbitrary values as long as they are different values.

(Abnormality detection device 400)
Hereinafter, the abnormality detection device 400 will be described with reference to FIGS. FIG. 7 is a block diagram illustrating a configuration of the abnormality detection device 400. FIG. 8 is a flowchart showing the operation of the abnormality detection device 400. As shown in FIG. 7, the abnormality detection device 400 includes a regression error calculation unit 220, an abnormality detection unit 230, and a recording unit 290. The recording unit 290 is a component that appropriately records information necessary for processing of the abnormality detection device 400. For example, the recording unit 290 records the regression model parameter θ ^ generated by the abnormality detection model generation device 300 and the abnormality detection model.

  The abnormality detection device 400 receives as input a set (x, t) of input data x and correct answer data t for the input data x. That is, (x, t) is data that is the target of abnormality detection.

The operation of the abnormality detection device 400 will be described with reference to FIG. The regression error calculation unit 220 calculates the regression error t−f _{θ ^} (x) of the input data x from (x, t) using the regression model f _{θ ^} (•) (S220).

The abnormality detection unit 230 generates an abnormality detection result for the input data x from the regression error t−f _{θ ^} (x) using the abnormality detection model (S230). Specifically, an abnormality detection result is generated as follows. First, the anomaly detection unit 230 uses the anomaly detection model to calculate the first value when t−fθ _^ (x) ∈B holds, and the second value as the degree of abnormality otherwise. To do. Next, the abnormality detection unit 230 displays an abnormality detection result indicating that the input data is normal when the abnormality degree is the first value, and indicates that the input data is abnormal when the abnormality degree is the second value. An anomaly detection result indicating that it exists is generated.

(Modification)
The anomaly detection model generated by the anomaly detection model generation apparatus 300 outputs a binary value of a first value and a second value, but N regression errors t _n −f _θ (x _n ) (1 When the loss for ≦ n ≦ N) is defined by the p-th power sum error, an abnormality detection model that outputs a continuous value can be obtained. Hereinafter, the abnormality detection model and the abnormality detection unit 230 will be described.

The anomaly detection model is output from the regression error t- f _{θ ^} (x) of the input data as the degree of anomaly as a negative log likelihood corresponding to the pth power.

  The abnormality detection unit 230 generates an abnormality detection result from the degree of abnormality using a threshold value. Specifically, when α is a threshold value and the degree of abnormality is less than or equal to α (less than α), an abnormality detection result indicating that the input data is normal is indicated. In other cases, the input data is abnormal. An abnormality detection result indicating that is generated. Of course, the degree of abnormality may be defined so that the degree of abnormality increases as the data is normal. In this case, if the degree of abnormality is greater than or equal to α (greater than α), the input data is normal. Otherwise, an abnormality detection result indicating that the input data is abnormal is generated.

  According to the invention of this embodiment, erroneous detection can be suppressed as in the first embodiment. For example, the present invention can be applied to abnormality detection for general data such as sensor data such as temperature, pressure, and displacement, and traffic data such as network traffic.

<Supplementary note>
The apparatus of the present invention includes, for example, a single hardware entity as an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected, and a communication device (for example, a communication cable) capable of communicating outside the hardware entity Can be connected to a communication unit, a CPU (Central Processing Unit, may include a cache memory or a register), a RAM or ROM that is a memory, an external storage device that is a hard disk, and an input unit, an output unit, or a communication unit thereof , A CPU, a RAM, a ROM, and a bus connected so that data can be exchanged between the external storage devices. If necessary, the hardware entity may be provided with a device (drive) that can read and write a recording medium such as a CD-ROM. A physical entity having such hardware resources includes a general-purpose computer.

  The external storage device of the hardware entity stores a program necessary for realizing the above functions and data necessary for processing the program (not limited to the external storage device, for example, reading a program) It may be stored in a ROM that is a dedicated storage device). Data obtained by the processing of these programs is appropriately stored in a RAM or an external storage device.

  In the hardware entity, each program stored in an external storage device (or ROM or the like) and data necessary for processing each program are read into a memory as necessary, and are interpreted and executed by a CPU as appropriate. . As a result, the CPU realizes a predetermined function (respective component requirements expressed as the above-described unit, unit, etc.).

  The present invention is not limited to the above-described embodiment, and can be appropriately changed without departing from the spirit of the present invention. In addition, the processing described in the above embodiment may be executed not only in time series according to the order of description but also in parallel or individually as required by the processing capability of the apparatus that executes the processing. .

  As described above, when the processing functions in the hardware entity (the apparatus of the present invention) described in the above embodiments are realized by a computer, the processing contents of the functions that the hardware entity should have are described by a program. Then, by executing this program on a computer, the processing functions in the hardware entity are realized on the computer.

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

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, a hardware entity is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

Claims (8)

θ ^ is a set (x _n , t _n ) (1 ≦ n ≦ N) of normal input data x _n and correct data t _n for the input data x _{n, where} N is an integer greater than or _equal to 1 The regression model parameter, f _{θ ^} (・), learned by using the regression model parameter θ ^ as a regression model,
Using the regression model f _{θ ^} (·), the regression error t- f _{θ ^} (x) of the input data x is calculated from the set (x, t) of the input data x and the correct answer data t for the input data x A regression error calculation unit,
An abnormality detection unit including an abnormality detection unit that generates an abnormality detection result for the input data x from the regression error t-f _{θ ^} (x) using an abnormality detection model,
The abnormality detection model outputs either the first value or the second value as the degree of abnormality from the regression error of the input data,
The abnormality detection unit generates an abnormality detection result indicating that the input data is normal when the abnormality degree is the first value, and the input data is abnormal when the abnormality degree is the second value. An abnormal sound detection device that generates an abnormality detection result indicating that. The abnormal sound detection device according to claim 1,
The abnormality detection model outputs the degree of abnormality using a bounded region B including N regression errors t _n −f _{θ ^} (x _n ) (1 ≦ n ≦ N). Abnormal sound detection device. The abnormal sound detection device according to claim 2,
θ is a regression model parameter generated in the process of learning the regression model parameter θ ^,
The loss for N regression errors t _n -f _θ (x _n ) (1 ≤ n ≤ N) is defined using the maximum absolute value error,
The abnormal sound detection apparatus, wherein the regression model parameter θ ^ is learned so as to minimize the loss. The abnormal sound detection device according to claim 2,
p is an integer of 3 or more, θ is a regression model parameter generated in the course of learning the regression model parameter θ ^,
The loss for N regression errors t _n -f _θ (x _n ) (1 ≤ n ≤ N) is defined using the p-power sum error,
The abnormal sound detection apparatus, wherein the regression model parameter θ ^ is learned so as to minimize the loss. The abnormal sound detection device according to claim 2,
p is an integer of 3 or more, θ is a regression model parameter generated in the course of learning the regression model parameter θ ^,
The loss for N regression errors t _n -f _θ (x _n ) (1 ≤ n ≤ N) is defined using the p-power sum error,
The abnormal sound detection apparatus, wherein the regression model parameter θ ^ is learned so as to minimize the loss, and the value of p is increased step by step during the learning. θ ^ is a set (x _n , t _n ) (1 ≦ n ≦ N) of normal input data x _n and correct data t _n for the input data x _{n, where} N is an integer greater than or _equal to 1 The regression model parameter, f _{θ ^} (・), learned by using the regression model parameter θ ^ as a regression model,
Using the regression model f _{θ ^} (·), the regression error t- f _{θ ^} (x) of the input data x is calculated from the set (x, t) of the input data x and the correct answer data t for the input data x A regression error calculation unit,
An abnormality detection unit including an abnormality detection unit that generates an abnormality detection result for the input data x from the regression error t-f _{θ ^} (x) using an abnormality detection model,
p is an integer of 3 or more, θ is a regression model parameter generated in the process of learning the regression model parameter θ ^,
The loss for N regression errors t _n -f _θ (x _n ) (1 ≤ n ≤ N) is defined using the p-power sum error,
The abnormality detection model outputs a negative log likelihood corresponding to the p-th power as an abnormality degree from a regression error of input data,
The abnormality detection unit generates the abnormality detection result from the abnormality degree using a threshold value. θ ^ is a set (x _n , t _n ) (1 ≦ n ≦ N) of normal input data x _n and correct data t _n for the input data x _{n, where} N is an integer greater than or _equal to 1 The regression model parameter, f _{θ ^} (・), learned by using the regression model parameter θ ^ as a regression model,
Using the regression model f _{θ ^} (·), the regression error t- f _{θ ^} (x) of the input data x is calculated from the set (x, t) of the input data x and the correct answer data t for the input data x A regression error calculation unit,
An anomaly detection device including an anomaly detection unit that generates an anomaly detection result for the input data x from the regression error t-f _{θ ^} (x) using an anomaly detection model,
The abnormality detection model outputs either the first value or the second value as the degree of abnormality from the regression error of the input data,
The abnormality detection unit generates an abnormality detection result indicating that the input data is normal when the abnormality degree is the first value, and the input data is abnormal when the abnormality degree is the second value. An anomaly detection device that generates an anomaly detection result indicating that   A program for causing a computer to function as the abnormal sound detection device according to any one of claims 1 to 6 or the abnormality detection device according to claim 7.

Images