JP7433648B2 - Anomaly detection method - Google Patents

Description

Application of Article 30, Paragraph 2 of the Patent Act ▲1▼August 26, 2019 Published in the publication “2019 Autumn Lecture Conference Summary Collection: Materials and Processes” ▲2▼September 11, 2019 “178th Autumn Lecture ▲3▼February 1, 2020 Announced in the publication “Tetsu to Hagane”

Embodiments of the present invention relate to an anomaly detection method for detecting an unknown anomaly.

Conventionally, systems and methods for verification and anomaly detection using a mixture of hidden Markov models have been disclosed (see Patent Document 1).

JP2017-21790A

However, if the abnormal state of the monitored object is unknown, it is not easy to determine whether the data measured from the monitored object is normal or abnormal.

An object of the embodiments of the present invention is to provide an anomaly detection method for detecting an unknown anomaly in a monitored object.

An abnormality detection method according to an aspect of the present invention is an abnormality detection method that uses a computer to detect an abnormality, wherein the computer decomposes waveform data indicating vibrations of a monitored object by frequency analysis, and the computer A hidden Markov model is applied to the decomposed data decomposed by the frequency analysis to calculate a state transition probability of the anomaly detection target, and the computer calculates the state transition probability of the anomaly detection target based on an evaluation function that evaluates suitability as an anomaly detection. The computer evaluates the state transition probability of the abnormality detection target, and the computer calculates the evaluated state transition probability of the abnormality detection target from the frequency analysis and the hidden waveform data when the monitoring target shows normal vibration. The computer compares the state transition probability during normal conditions calculated based on the Markov model, and determines that, as a result of the comparison, there is a predetermined difference between the state transition probability of the abnormality detection target and the state transition probability during normal conditions. The evaluation function includes determining that the monitoring target is abnormal if there is a difference greater than or equal to a value, and the evaluation function is based on the probability that the probability of transitioning from one state to another state is the probability of transitioning from the one state to the same state. The smaller the value, the higher the evaluation .

According to embodiments of the present invention, it is possible to provide an anomaly detection method for detecting an unknown anomaly in a monitoring target.

FIG. 1 is a configuration diagram showing a general configuration of a hidden Markov model according to an embodiment of the present invention. FIG. 3 is a waveform diagram showing the analysis results of conveyor vibration data subjected to multi-resolution analysis according to the present embodiment. FIG. 3 is a state transition diagram showing an example when the number of latent states of the hidden Markov model according to the present embodiment is two. FIG. 3 is an image diagram showing a method for determining a normal state and an abnormal state by comparing time series models according to the present embodiment. FIG. 1 is a configuration diagram showing the configuration of an abnormality detection device according to the present embodiment. FIG. 3 is an image diagram showing a list showing ranking of models according to the present embodiment.

(Embodiment)
First, the theory of the abnormality detection method according to this embodiment will be explained.
Hidden Markov Model is one of the methods for modeling characteristic patterns of time series data. Hidden Markov models model that states do not need to be directly observed, only stochastic events are observed. The stochastic processes that represent this hidden state are called latent variables, and each is not independent, but is a mixture distribution model in which they work together through a Markov process. Each latent variable S={S ₁ ,. ．． ．． , S _N }, a hidden Markov model can be expressed (Equation (1)).

Here, the latent variable S is a state sequence with K states of a Markov process, and has a state transition probability matrix A. π is an initial probability parameter that determines the leading state S1. The general configuration of a Hidden Markov Model (HMM) for time-series data is shown in FIG. This figure shows a model called left-to-right HMM, with initial state probabilities p(S ₁ )=1, p(S _j )=0 (j≠1), and all columns have state j= It is limited to starting from 1.

On the other hand, in the case of a first-order Markov chain that depends only on one neighboring state, the state transition probability A is expressed by a K×K matrix, and the transition probability from state i to state j can be expressed as a _ij . can. This model also considers transitions in the opposite direction, such as returning from an abnormal state to a normal state.

By the way, in the case of vibration data for conveyor equipment, it is expected that the observed frequencies will vary depending on the source of vibration, such as the motor, bearing, or support, and the natural frequency. However, it is difficult to specify in advance which frequency range is suspicious, and it is also impossible to predict in what time period the system will transition to an abnormal state.

Therefore, we adopted a wavelet analysis method that allows continuous frequency analysis of time-series data that lasts for a long time. This method performs time-frequency analysis by translating and expanding/contracting a small waveform ψ(x) called a mother wavelet and multiplying it by the original waveform. Multiresolution analysis is one type of discrete wavelet transform that is applied to discrete problems to facilitate calculation. This is a method of continuously obtaining a resolution waveform at every doubling using a mother wavelet serving as a basis function. Haar's scaling function φ _H is defined in equation (2).

Using this, the function F _j can be decomposed as shown in equation (5).

Here, f _j (x) and g _j (x) are given by equations (6) and (7).

These equations computed for each k are called decomposition algorithms.

FIG. 2 shows the analysis results of conveyor vibration data obtained by performing multi-resolution analysis (also called discrete wavelet transform), which is one of the wavelet transform techniques. The bottom layer is conveyor vibration data measured using an acceleration sensor, and the higher layers are time-series waveforms in which the signal train is decomposed into half the resolution using a scaling function. Waveform characteristics that differ from the raw waveform appear in the decomposed waveform. Therefore, unknown abnormal data is detected by applying a hidden Markov model to each resolution waveform that has been subjected to wavelet transformation. This is an attempt to discover abnormal conditions from equipment vibration data by estimating the unknown steady state that is behind the observed frequency-based time series data and changes over time through a Markov process. .

Next, we will discuss a model selection method for different data in hidden Markov model analysis. Here, model selection using only the information criterion cannot be simply used for data that has been subjected to discrete wavelet transform. In addition, there are many hyperparameters, such as the resolution level of multiresolution analysis and the shape and number of distributions used in hidden Markov models, and it is necessary to find all their combinations through trial and error. For the purpose of reducing this effort and selecting model parameters suitable for anomaly detection, evaluation functions for variable selection and hyperparameter selection will be described.

From the perspective of abnormality diagnosis, too frequent state transitions are considered undesirable. Therefore, we believe that it is desirable to use a hidden Markov model in which the probability of transitioning from one state i to another state j is sufficiently smaller than the probability of transitioning from state i to the same state i, and we use the following equation (8) as the evaluation function. .

Here, S is a set of states, p(i,j) is a transition probability from state i to j, and the smaller the value of eval, the better.

A model with the smallest eval value is extracted using the eval function for the measurement data to be monitored.

The elements forming the model are as follows. For example, when the measurement data is vibration data, there are types of measurement data such as displacement, velocity, and acceleration. Furthermore, for each type of measurement data, there are combinations of wavelet decomposition levels and hidden Markov model hyperparameters (estimated distribution type, number of distributions, initial value of distribution, etc.). The estimated distribution type is a probability distribution such as a normal distribution or a mixed normal distribution. For example, the hyperparameters of a hidden Markov model have the following selection elements. The number of distributions (number of latent states) is selected from a predetermined range (for example, 2 to 4). When the probability distribution is a mixed normal distribution, the maximum number of normal distributions is selected from a predetermined range (for example, 2 to 4). When the probability distribution is a normal distribution, the initialization method is selected from random, k-means, and the like. FIG. 3 is an example of a hidden Markov model state transition diagram in which the number of latent states is two.

Using the eval function, it is possible to select a model in which the probability of transitioning from one state i to another state j is smaller than the probability of transitioning from state i to the same state i. This makes it possible to clearly separate a normal state and an abnormal state at a certain point in time.

Here, in reality, daily monitoring must be performed without knowing whether it is normal or abnormal. All of the measured data may be in a normal state, or all of it may be in an abnormal state. Or perhaps the data contains both normal and abnormal data. Therefore, in order to deal with these issues, we use the following analysis method.

Using the state transition probabilities (parameters) of the Hidden Markov Model and the sequential data output probabilities from each state as data, there are time series models created from known normal data, and cases created from data that is unknown whether it is normal or abnormal. Compare series models to determine normal and abnormal states. Figure 4 shows the concept.

A model created only from time series data that is known to be normal is Mk _i , i=1. ．． ．． Let it be m. This model analyzes all combinations (total m) of hyperparameters such as displacement, velocity, and acceleration as data, each level of wavelets, and the distribution of hidden Markov models, and stores the state transition probabilities for each. I'll keep it. Next, a model Mu that minimizes eval is created from time-series data whose normality or abnormality is unknown. Then, it is compared with Mk _i having the same hyperparameters as that model. If the unknown time series is normal, the state transition probabilities of the two models are expected to be similar, and if the unknown time series is abnormal, the two models are expected to be different. Therefore, the mean square error MSE of the state transition probability is defined as shown in equation (9) below.

Here, sk _ij and su _ij are the state transition probabilities of the known model and the unknown model, and k is the number of latent states. Using this MSE, the MSE between two time series data measured from a normal state and the MSE between two time series data measured from a normal state and an unknown state are compared.

If a significant difference is found as a result of the comparison, it means that an unknown abnormal state has been discovered from the state transition probabilities analyzed using this model. Whether or not there is a significant difference can be determined using a statistical method such as a t-test.

FIG. 5 is a configuration diagram showing the configuration of the abnormality detection device 10 according to this embodiment.
The abnormality detection device 10 is a device that detects an abnormality based on vibration data D1 acquired from a monitoring target. Here, abnormality means a state different from normal, and does not necessarily mean that there is a defect such as a failure, and may also include a state in which the device operates normally without any problems.

The object to be monitored by the abnormality detection device 10 may be any object as long as it generates vibrations. For example, monitoring targets are often related to civil engineering or machinery using motors, such as conveyor equipment, bridges, roads, high-rise buildings, or cranes. Furthermore, the monitoring target is not limited to something that physically vibrates, but may be anything related to finance, such as stock market prices or exchange rates, as long as vibrations can be shown in a waveform.

The abnormality detection device 10 includes an arithmetic processing section 1 and a display 2. The arithmetic processing unit 1 is mainly composed of a computer, and performs general arithmetic processing of the abnormality detection device 10 according to software such as a program. The number of computers is not limited to one, but may be composed of a plurality of computers, and may be configured in any computer system. The display 2 displays various information such as abnormality detection results. Further, the display 2 may display information for operating the abnormality detection device 10 or may have an input function for inputting information to the abnormality detection device 10.

The arithmetic processing unit 1 includes a waveform data extraction unit 11 , a frequency analysis unit 12 , a waveform data synthesis unit 13 , a hidden Markov model application unit 14 , a parameter evaluation unit 15 , a parameter comparison unit 16 , and an abnormality determination unit 17 . The arithmetic processing unit 1 accesses the vibration data D1, the normal waveform data D2, the hyperparameter data D3, and the normal parameter data D4 as necessary. These data D1 to D4 may be stored in one storage medium, or may be distributed and stored in two or more storage media.

The vibration data D1 is data that can be obtained from a monitoring target and extract vibrations as waveform data. For example, the vibration data D1 is a moving image captured with a high-speed camera using area sensing technology. Note that the vibration data D1 is not limited to a two-dimensional video, but may be a video in which vibrations are captured three-dimensionally by a plurality of cameras. Further, the vibration data D1 may be data obtained by measuring vibrations using a plurality of pressure sensors, or may be numerical data expressed discretely, such as market prices. Hereinafter, the vibration data D1 will be mainly explained as data that can extract physical vibrations in three directions of the X-axis, Y-axis, and Z-axis as waveform data.

The waveform data extraction unit 11 extracts the type of waveform data used for abnormality detection from the vibration data D1 input to the abnormality detection device 10. The waveform data extraction section 11 outputs the extracted waveform data to the frequency analysis section 12. Note that the number of waveform data to be extracted is not limited to one type, but may be two or more types. Further, if the vibration data D1 is already the type of waveform data used for abnormality detection, the processing by the waveform data extraction unit 11 may not be performed.

Displacement data in three directions, the X-axis, Y-axis, and Z-axis, can be extracted from the vibration data D1. Velocity data can be found by differentiating the displacement data. Acceleration data can be found by differentiating the velocity data. Therefore, nine types of waveform data can be extracted from the vibration data D1. The type of waveform data used for abnormality detection is selected from these nine types of waveform data, but all types of waveform data may be selected.

The frequency analysis section 12 performs frequency analysis on the waveform data input from the waveform data extraction section 11 and decomposes it into a plurality of waveform data. For example, frequency analysis is a multi-resolution analysis. The frequency analysis unit 12 outputs the type of waveform data used for abnormality detection to the waveform data synthesis unit 13 from among the plurality of decomposed waveform data. For example, in the case of multi-resolution analysis, the type of waveform data corresponds to the decomposition level of the wavelet, and each time the decomposition level increases by one, the resolution doubles. The type of waveform data to be subjected to frequency analysis is not limited to one type, but may be two or more types, or all types.

Note that the frequency analysis method used by the frequency analysis unit 12 is preferably a wavelet transform that allows continuous frequency analysis of time series data that continues for a long time, but a fast Fourier transform or a short time Fourier transform may be adopted.

The waveform data synthesis section 13 synthesizes the waveform data inputted from the frequency analysis section 12 with waveform data of the same type as this waveform data and acquired from the normal waveform data D2. The normal waveform data D2 includes various types of waveform data measured when the monitoring target is normal (normal time). The waveform data synthesis section 13 outputs the synthesized waveform data to the hidden Markov model application section 14.

For example, if the waveform data input from the frequency analysis unit 12 is the waveform data obtained by decomposing the X-axis velocity data at level 2 of multi-resolution analysis, the waveform data to be synthesized is based on the vibration data measured during normal conditions. This is waveform data obtained by decomposing generated X-axis velocity data at level 2 of multi-resolution analysis.

Note that the synthesis of waveform data by the waveform data synthesis section 13 may be performed before the frequency analysis by the frequency analysis section 12. In this case, normal waveform data of the same type as the waveform data extracted by the waveform data extraction unit 11 may be combined with the waveform data to be detected for abnormality, and the frequency analysis unit 12 may perform frequency analysis on the combined waveform data. .

The hidden Markov model application unit 14 applies a hidden Markov model to the waveform data input from the waveform data synthesis unit 13 based on the hyperparameter set included in the hyperparameter data D3. The hidden Markov model application unit 14 outputs a set of parameters (state transition probabilities) calculated by applying the hidden Markov model to the parameter evaluation unit 15. For example, when the state transition diagram of the hidden Markov model has the configuration shown in FIG. 3, the number of calculated parameters is eight.

Hyperparameter data D3 is data including a set of hyperparameters for forming a hidden Markov model for abnormality detection. For example, in the hyperparameter set, the number of latent states is "3", the number of normal states is "1", the number of abnormal states is "2", the probability distribution is "mixed normal distribution", and the maximum number of normal distributions is "2". ”, the initialization method for the normal distribution is “random”, etc., and includes information on each hyperparameter. Note that the hyperparameter data D3 may include two or more sets of hyperparameters, and the hidden Markov model application unit 14 may apply a plurality of hidden Markov models.

Based on the set of parameters input from the hidden Markov model application unit 14, the parameter evaluation unit 15 determines the combination of the applied hidden Markov model and the type of waveform data to which this hidden Markov model is applied (hereinafter referred to as “model '') is appropriate for abnormality detection. The parameter evaluation unit 15 uses the evaluation function shown in equation (8) to evaluate the model so that the probability of transitioning from one state i to another state j becomes higher as the probability of transitioning from state i to the same state i is smaller. rate highly. The parameter evaluation unit 15 determines the ranking of each model based on the evaluation results of each model. The parameter evaluation unit 15 selects a predetermined number (one or more) of models in descending order of evaluation. That is, the parameter evaluation unit 15 selects a model to be used for abnormality detection from a plurality of candidate models. The parameter evaluation unit 15 outputs the selected model and the set of parameters of the model to the parameter comparison unit 16. Note that the evaluation function is not limited to the one shown in equation (8), and any evaluation method may be used as long as the suitability of the model for detecting an abnormality can be evaluated.

For example, the parameter evaluation unit 15 ranks the models as shown in the list shown in FIG. Note that the list in FIG. 6 is an image diagram, and any ranking may be performed. Furthermore, the items (elements) listed at the top are merely examples, and some items may be omitted, or other items may be added.

Note that if the evaluation of the model by the parameter evaluation unit 15 is performed as an initial setting before the start of operation of the abnormality detection device 10, it does not need to be performed after the start of operation of the abnormality detection device 10. Furthermore, when there is one set of parameters input from the hidden Markov model application section 14, the parameter evaluation section 15 does not need to evaluate the model.

The parameter comparison unit 16 is the same model as the model input from the parameter evaluation unit 15, and extracts a set of parameters based on waveform data measured when the monitoring target is normal from the normal parameter data D4. The parameter comparison unit 16 compares, for each model, the set of parameters input from the parameter evaluation unit 15 and the set of normal parameters extracted from the normal parameter data D4, and calculates the comparison result as a numerical value. The parameter comparison unit 16 outputs the numerical comparison result to the abnormality determination unit 17.

For example, the parameter comparison unit 16 calculates the (MSE) mean square error shown in equation (9) for each parameter set. By performing a t-test on the two calculated MSEs, a p value is obtained as a comparison result. Note that the comparison of the parameter sets may be performed in any manner, and the comparison results may be quantified in any manner. Further, any statistical method may be used, or an arithmetic expression derived from empirical rules or the like may be used.

The abnormality determination unit 17 determines whether the monitoring target is in an abnormal state based on the comparison result input from the parameter comparison unit 16. For example, when the evaluation result is a p value determined by a t-test, if the p value is smaller than 0.05, the abnormality determining unit 17 determines that the monitoring target is in an abnormal state. When a plurality of evaluation results are input from the parameter comparison section 16, the abnormality determination section 17 determines that the monitoring target is in an abnormal state if it is determined to be in an abnormal state based on any one of the evaluation results. The abnormality determination unit 17 outputs the determination result to the display 2. If the determination result is an abnormal state, the abnormality determination unit 17 outputs information on the model corresponding to the comparison result determined to be an abnormal state to the display 2 together with the determination result.

Note that the abnormality determining unit 17 may determine the abnormal state in any manner. For example, when making a judgment based on a p value based on a t-test, the threshold value to be compared with the p value is not limited to 0.05, but may be any numerical value. Furthermore, if the monitoring target is not determined to be abnormal based on a plurality of evaluation results, it may be determined that the monitoring target is not in an abnormal state. Further, the abnormality determining unit 17 may determine that an abnormal state exists if the difference between two numerical values obtained from two sets of parameters using a predetermined arithmetic expression set in advance is greater than or equal to a threshold value.

The display 2 displays the diagnosis result of the monitoring device based on the judgment result input from the abnormality judgment section 17. If the diagnosis result is an abnormal state, the display 2 displays information about the model that is the basis for determining the abnormal state. As a result, if an abnormal condition is determined, the operator can determine whether the monitored object is malfunctioning or investigate the cause of the malfunction based on the model information displayed on the display 2. .

Note that the display 2 does not have to display all information regarding the model. For example, the type of waveform data may be displayed, and information regarding the applied Hidden Markov Model (eg, hyperparameters) may not be displayed.

Next, a method of generating the normal waveform data D2, hyperparameter data D3, and normal parameter data D4 will be described. These data D2 to D4 are generated at the time of initial setting of the abnormality detection device 10, or at the time of periodic inspection of the monitored object.

Vibration data D1 measured when the monitoring target is normal is input to the waveform data extraction unit 11. The normal vibration data D1 may be vibration data measured at any time as long as it is vibration data measured when the test object is in a normal state. For example, when the test target is a device, the normal vibration data D1 is measured when the device is shipped from the factory or during periodic inspection of the device.

The waveform data extraction unit 11 extracts all types of waveform data that can be extracted. For example, the waveform data extraction unit 11 extracts displacement data, velocity data, and acceleration data for each of the X-axis, Y-axis, and Z-axis directions. The waveform data extraction unit 11 outputs all types of extracted waveform data to the frequency analysis unit 12.

The frequency analysis section 12 performs frequency analysis on all the waveform data input from the waveform data extraction section 11 at the maximum decomposable level. For example, if the maximum decomposition level of multi-resolution analysis by the frequency analyzer 12 is 5, the frequency analyzer 12 generates waveform data with decomposition levels from 1 to 5. The frequency analysis unit 12 outputs all types of decomposed waveform data to the hidden Markov model application unit 14. At this time, it is not necessary to synthesize the waveform data by the waveform data synthesis section 13.

Note that some types of waveform data may be predetermined not to be processed (extracted or frequency analyzed) in the waveform data extraction unit 11 and the frequency analysis unit 12. For example, it is not necessary to process waveform data of a type in which it is known in advance that an abnormal state of the monitored object cannot be detected or is difficult to detect.

The hidden Markov model application unit 14 applies a hidden Markov model based on all hyperparameter sets included in the hyperparameter data D3 to all types of waveform data input from the frequency analysis unit 12. Here, the hyperparameter data D3 is initially set to include, as candidates, all sets of hyperparameters that are considered effective for detecting abnormalities in the monitored object. The hidden Markov model application unit 14 outputs a set of all parameters calculated by applying the hidden Markov model to the parameter evaluation unit 15.

The parameter evaluation unit 15 evaluates each model using an evaluation function based on the set of parameters input from the hidden Markov model application unit 14. As a result, a predetermined number (one or more) of models are determined in order of highest evaluation.

The normal waveform data D2 is updated to save the waveform data used in the determined model (waveform data after decomposition by the frequency analysis section 12 or waveform data extracted by the waveform data extraction section 11). Ru. The hyperparameter data D3 is updated to save the set of hidden Markov model hyperparameters used for the determined model. The normal parameter data D4 is updated to save the determined set of model parameters. Note that the hyperparameter data D3 does not need to be updated from the initial setting.

According to the present embodiment, in order to frequency-analyze waveform data indicating vibrations of a monitoring target and apply a hidden Markov model to detect abnormalities, a model based on a combination of waveform data types and hidden Markov models is evaluated. Select the appropriate function. By comparing the hidden Markov model parameters obtained from unknown vibration data using the selected model with the hidden Markov model parameters obtained from normal vibration data, unknown abnormalities can be detected.

Furthermore, if the various data D2 and D4 during normal operation are updated during periodic inspections of the monitored object, it is possible to suppress changes in the monitored object due to deterioration over time and detect unknown abnormalities.

Furthermore, the type of waveform data in which an abnormality was detected can be used as a clue for investigating the cause of the detected abnormality. For example, if an abnormality is detected from waveform data of a specific frequency, it is possible to infer that a component whose unique vibration frequency is this specific frequency is the cause of the abnormality.

Note that the present invention is not limited to the embodiments described above, and constituent elements may be deleted, added, or changed. Furthermore, a new embodiment may be created by combining or exchanging components of a plurality of embodiments. Even if such embodiments are directly different from the embodiments described above, those having the same meaning as the present invention are treated as embodiments of the present invention, and the explanation thereof will be omitted.

DESCRIPTION OF SYMBOLS 1... Arithmetic processing section, 2... Display unit, 11... Waveform data extraction section, 12... Frequency analysis section, 13... Waveform data synthesis section, 14... Hidden Markov model application section, 15... Parameter evaluation section, 16... Parameter comparison section , 17...Abnormality judgment unit, D1...Vibration data, D2...Normal waveform data, D3...Hyper parameter data, D4...Normal parameter data.

Claims (8)

An anomaly detection method for detecting an anomaly using a computer,
The computer decomposes waveform data indicating vibration of the monitored object by frequency analysis,
The computer applies a hidden Markov model to the decomposed data decomposed by the frequency analysis to calculate a state transition probability of the abnormality detection target,
the computer evaluates the calculated state transition probability of the abnormality detection target based on an evaluation function that evaluates suitability for abnormality detection;
The computer calculates the evaluated state transition probability of the abnormality detection target based on the frequency analysis and the hidden Markov model from the normal waveform data showing the vibration when the monitoring target is normal. compared to probability,
As a result of the comparison, the computer determines that the monitoring target is abnormal if there is a difference greater than a predetermined value between the state transition probability of the abnormality detection target and the normal state transition probability. including making judgments;
The evaluation function is such that the evaluation becomes higher as the probability of transitioning from one state to another state is smaller than the probability of transitioning from the one state to the same state.
An anomaly detection method characterized by : The abnormality detection method according to claim 1, wherein the hidden Markov model is applied to synthetic data in which the decomposed data of the abnormality detection target includes decomposed data based on the normal waveform data. . The abnormality detection method according to claim 1, wherein the frequency analysis uses wavelet transform. 2. The abnormality detection method according to claim 1, wherein the waveform data is calculated based on vibration data obtained by measuring vibrations of the monitoring target. 5. The abnormality detection method according to claim 4, wherein the vibration data is a moving image. The state transition probability of the abnormality detection target to be compared is selected from a plurality of state transition probabilities for a plurality of types of waveform data to which the hidden Markov model is applied, based on the evaluation function. The abnormality detection method according to claim 1 . The determination is characterized in that the monitoring target is determined to be abnormal if a difference in mean square error between the state transition probability of the abnormality detection target and the state transition probability during normal operation is equal to or greater than a threshold value. The abnormality detection method according to item 1. a frequency analysis means for decomposing waveform data indicating vibration of a monitored object by frequency analysis;
state transition probability calculation means for calculating a state transition probability of an abnormality detection target by applying a hidden Markov model to the decomposed data decomposed by the frequency analysis means;
evaluation means for evaluating the state transition probability of the abnormality detection target calculated by the state transition probability calculation means based on an evaluation function for evaluating suitability for abnormality detection;
The state transition probability of the abnormality detection target evaluated by the evaluation means is calculated based on the frequency analysis and the hidden Markov model from normal waveform data showing vibration when the monitoring target is normal. a comparison means for comparing the transition probabilities;
As a result of the comparison by the comparison means, if there is a difference of more than a predetermined value between the state transition probability of the abnormality detection target and the normal state transition probability, it is determined that the monitoring target is abnormal. and an abnormality judgment means to
The evaluation function is such that the evaluation becomes higher as the probability of transitioning from one state to another state is smaller than the probability of transitioning from the one state to the same state.
An anomaly detection device characterized by :

Images