Anomaly detection device, method and program
The abnormality detection device addresses the challenge of accurately identifying abnormal signals by calculating reconstruction errors and period information, using smoothing and machine learning to enhance the detection of periodic abnormalities in sensor signals.
Patent Information
- Application Number
- JP2021150373
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2021-09-15
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2041-09-15
AI Technical Summary
Existing anomaly detection methods using principal component analysis struggle to accurately distinguish abnormal signals from noise, particularly when accidental noise components cause locally large reconstruction errors, leading to incorrect detection of abnormalities.
An abnormality detection device that calculates reconstruction errors and period information from time-series sensor signals, using methods like principal component analysis, and employs smoothing, weighting, and machine learning to determine the presence of abnormal signals based on both error magnitude and periodicity.
Improves the accuracy of abnormality detection by considering both the magnitude and periodicity of reconstruction errors, effectively distinguishing noise from periodic abnormalities, thereby enhancing the detection of periodic abnormal signals.
Smart Images

Figure 0007581167000002 
Figure 0007581167000003 
Figure 0007581167000004
Abstract
Description
[Technical field]
[0001] An embodiment of the present invention relates to an anomaly detection device, method, and program. [Background technology]
[0002] In recent years, the spread of IoT (Internet of Things) has made it easier to use sensors to monitor the status of equipment in social infrastructure, manufacturing lines, etc. As a result, technology that uses sensor signals obtained from sensors to detect abnormalities and failures in equipment has been attracting attention.
[0003] One of the anomaly detection techniques using one-dimensional signals such as sensor signals is a method using principal component analysis. In a method using principal component analysis, reconstruction is performed after dimensional reduction is performed on the input signal. At this time, if the majority of the input signal is normal or noise, only the noise components are reconstructed and the abnormal signal components are difficult to reconstruct. For this reason, it is possible to detect abnormal signals by utilizing the magnitude of the reconstruction error, which is the error between the reconstructed signal and the input signal.
[0004] When an input signal contains accidental noise components, the reconstruction error may become large locally. In a method using principal component analysis, anomaly detection is performed by focusing only on the magnitude of the reconstruction error, so the accidental noise components may be erroneously detected as an abnormal signal. In addition, when the reconstruction error becomes locally large due to accidental noise, the reconstruction error at the time when an abnormal signal occurs becomes smaller than the reconstruction error at the time when a normal signal occurs, and the abnormal signal may not be detected normally. [Prior art documents] [Non-patent literature]
[0005] [Non-Patent Document 1] IT Jolliffe, "Principal Component Analysis" (USA), 2nd edition, Springer, 2002, P.37, 232-268 Summary of the Invention [Problem to be solved by the invention]
[0006] An object of the present invention is to provide an abnormality detection device, method, and program capable of determining with high accuracy whether or not an abnormal signal is present in an input signal. [Means for solving the problem]
[0007] In order to solve such problems, an anomaly detection device according to an embodiment includes a reconstruction error calculation unit, a periodic information calculation unit, and a determination unit. The reconstruction error calculation unit calculates a reconstruction error of an input signal that is a time-series signal. The periodic information calculation unit calculates periodic information that represents periodicity of the reconstruction error. The determination unit determines the presence or absence of an abnormal signal in the input signal based on the periodic information or based on the periodic information and the reconstruction error. [Brief description of the drawings]
[0008] [Figure 1] FIG. 1 is a diagram illustrating an example of the configuration of an anomaly detection device according to the first embodiment. [Diagram 2] FIG. 2 is a diagram illustrating an example of an input signal acquired by the anomaly detection device according to the first embodiment. [Diagram 3] FIG. 3 is a diagram illustrating an example of a reconstruction error calculated by the anomaly detection device according to the first embodiment. [Figure 4] FIG. 4 is a diagram illustrating an example of the reconstruction error before smoothing calculated by the anomaly detection device according to the first embodiment. [Diagram 5] FIG. 5 is a diagram showing an example of the reconstruction error after the reconstruction error in FIG. 4 is smoothed. [Figure 6]FIG. 6 is a diagram showing an example of a reconstruction error obtained by subtracting the reconstruction error after smoothing shown in FIG. 5 from the reconstruction error before smoothing shown in FIG. [Figure 7] FIG. 7 is a diagram showing a state in which reconstruction errors equal to or greater than a predetermined threshold are selected from the reconstruction errors shown in FIG. [Figure 8] FIG. 8 is a diagram showing a state in which reconstruction errors are selected in descending order from the reconstruction errors shown in FIG. [Figure 9] FIG. 9 is a diagram illustrating an example of a classifier used as a machine learning model in the anomaly detection device according to the first embodiment. [Figure 10] FIG. 10 is a diagram illustrating an example of a learning method for the classifier illustrated in FIG. [Figure 11] FIG. 11 is a diagram showing an example of a method for calculating a weighted reconstruction error using the reconstruction error shown in FIG. [Figure 12] FIG. 12 is a flowchart illustrating a processing procedure of the anomaly detection process performed by the anomaly detection device according to the first embodiment. [Figure 13] FIG. 13 is a diagram illustrating an example of the configuration of the anomaly detection device according to the second embodiment. As illustrated in FIG. [Figure 14] FIG. 14 is a diagram illustrating an example of an input signal acquired by the anomaly detection device according to the second embodiment. [Figure 15] FIG. 15 is a diagram illustrating an example of a short-time signal generated by the anomaly detection device according to the second embodiment. [Figure 16] FIG. 16 is a diagram showing a reconstruction error calculated for each short-time signal by the anomaly detection device according to the second embodiment. [Figure 17] FIG. 17 is a diagram showing representative values selected for each short-time signal by the anomaly detection device according to the second embodiment. [Figure 18] FIG. 18 is a flowchart illustrating a processing procedure of an anomaly detection process performed by the anomaly detection device according to the second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0009] Hereinafter, an embodiment of an anomaly detection device, a method, and a program will be described in detail with reference to the drawings. In the following description, components having substantially the same functions and configurations are denoted by the same reference numerals, and duplicated descriptions will be given only when necessary.
[0010] The anomaly detection device of the following embodiments is a diagnostic device that acquires a sensor signal obtained from an equipment, judges whether or not there is an abnormal signal in the acquired sensor signal, and diagnoses the equipment's abnormality, failure, state, etc. based on the judgment result. Examples of such diagnostic devices include a device that analyzes a sensor signal to diagnose a failure in high-voltage power supply equipment, and a device that analyzes the rotation sound of a rotor used in a coil, etc. to detect an abnormality. The anomaly detection device may also be called a signal processing device.
[0011] (First embodiment) 1 is a diagram showing the configuration of an anomaly detection device 100 according to a first embodiment. The anomaly detection device 100 is connected to a high-voltage power supply facility via a network or the like.
[0012] The anomaly detection device 100 detects a sensor signal for detecting anomalies in high-voltage power supply equipment or coils. For example, a general AE sensor attached to high-voltage power supply equipment is used as the sensor. The anomaly detection device 100 determines whether or not there is a signal indicating an anomaly in the sensor signal (hereinafter, referred to as an anomaly signal), and detects an anomaly based on the determination result. When the size of the sensor signal is large, the anomaly detection device 100 may perform constant monitoring and analyze only data determined to be abnormal.
[0013] The network is, for example, a local area network (LAN). Note that the connection to the network may be either a wired connection or a wireless connection. Furthermore, the network is not limited to a LAN, and may be the Internet, a public communication line, or the like.
[0014] The anomaly detection device 100 includes a processing circuit that controls the entire anomaly detection device 100, and a storage medium (memory). The processing circuit is a processor that executes the functions of a reconstruction error calculation unit 101, a period information calculation unit 102, and a determination unit 103 by calling and executing a program in the storage medium. The processing circuit is formed of an integrated circuit including a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or the like. The processor may be formed of one integrated circuit or multiple integrated circuits.
[0015] The storage medium stores a processing program used by the processor, and parameters and tables used in the calculations of the processor. The storage medium is a storage device such as an HDD (Hard Disk Drive), an SSD (Solid State Drive), or an integrated circuit that stores various information. In addition to an HDD or SSD, the storage device may be a portable storage medium such as a CD (Compact Disc), a DVD (Digital Versatile Disc), or a flash memory, or may be a drive that reads and writes various information between the storage device and a semiconductor memory element such as a flash memory or a RAM (Random Access Memory).
[0016] Each function of the reconstruction error calculation unit 101, the period information calculation unit 102, and the determination unit 103 may be realized by a single processing circuit, or a processing circuit may be configured by combining multiple independent processors, and each processor may execute a program to realize each function. Also, each function of the reconstruction error calculation unit 101, the period information calculation unit 102, and the determination unit 103 may be implemented as an individual hardware circuit.
[0017] The reconstruction error calculation unit 101 acquires an input signal from a sensor attached to a high-voltage power supply facility or a coil. The input signal is a voltage value acquired from the sensor. FIG. 2 is a diagram showing an example of the input signal. The horizontal axis of FIG. 2 indicates time. The vertical axis of FIG. 2 indicates voltage value.
[0018] The input signal is a time series signal that includes noise generated from peripheral devices, etc. The noise includes stationary noise and non-stationary noise. Of the input signals, a signal that is generated steadily due to stationary noise is called a normal signal, and a signal that is generated non-stationarily due to non-stationary noise is called noise. When an abnormality occurs in the high-voltage power supply equipment or the coil, the input signal becomes a signal in which an abnormal signal is mixed with a normal signal and noise, as shown in Figure 2. The occurrence timing of the abnormal signal is periodic. Area S1 in Figure 2 indicates the total value of the normal signal and the abnormal signal, and area S2 in Figure 2 indicates the value of the abnormal signal.
[0019] The reconstruction error calculation unit 101 calculates the reconstruction error based on the acquired input signal by a predetermined calculation method. Examples of the reconstruction error calculation method include principal component analysis, kernel principal component analysis, Auto Encoder, Variational Auto Encoder (VAE), Generative Adversarial Network (GAN), etc.
[0020] In this embodiment, a case where principal component analysis is used as a method for calculating a reconstruction error will be described. In principal component analysis, an input signal is divided into a plurality of short-time frames. Then, reconstruction is performed after performing dimensionality reduction on each divided frame. When performing reconstruction, abnormal signal components are difficult to reconstruct. Therefore, the reconstructed signal is a signal in which only normal signal and noise components of the input signal are normally reflected. Then, by calculating the difference between the reconstructed signal and the input signal for each divided frame, a reconstruction error reflecting the magnitude of the abnormal signal is calculated. FIG. 3 is a diagram showing an example of a reconstruction error. The horizontal axis of FIG. 3 indicates frames. The vertical axis of FIG. 3 indicates a value of a reconstruction error.
[0021] Moreover, the reconstruction error calculation unit 101 corrects the bias (waviness) of the reconstruction error after calculating the reconstruction error. At this time, the reconstruction error calculation unit 101 performs a smoothing process on the calculated reconstruction error, and corrects the fluctuation of the reconstruction error by subtracting the smoothed reconstruction error from the reconstruction error before smoothing, for example. FIG. 4 is a diagram showing an example of the reconstruction error before smoothing. FIG. 5 is a diagram showing an example of the reconstruction error after smoothing the reconstruction error shown in FIG. 4. FIG. 6 is a diagram showing an example of the reconstruction error after subtracting the smoothed reconstruction error shown in FIG. 5 from the reconstruction error before smoothing shown in FIG. 4. In FIGS. 4 to 6, the horizontal axis indicates frames, and the vertical axis indicates the value of the reconstruction error. Examples of the smoothing process include a convolution process using a moving average filter, an approximation process to an arbitrary function, and the like.
[0022] The period information calculation unit 102 calculates period information representing the periodicity of the reconstruction error. The period information is a feature value representing the periodicity of the reconstruction error. The period information is, for example, a coefficient, a spectrum, a spectrogram, or the like obtained by Fourier transforming the reconstruction error.
[0023] The periodic information may be an index showing the degree of variation between two or more reconstruction error times selected based on the magnitude of the value. For example, a parameter related to the distance between the times at which the reconstruction error is equal to or greater than a predetermined threshold is used as the periodic information. In this case, the periodic information calculation unit 102 selects the times at which the reconstruction error is equal to or greater than a predetermined threshold as t=(t1, t2, ... tN), and calculates the difference d between the selected times. i (=|t i -t j |, i ≠ j, 1 ≦ i, j ≦ N) is calculated as period information. For example, the average value or standard deviation of the reconstruction errors is used as the threshold. Fig. 7 is a diagram showing a state in which reconstruction errors equal to or greater than a predetermined threshold are selected from the reconstruction errors shown in Fig. 3.
[0024] Alternatively, the period information may be a parameter related to the distance between the times of two or more reconstruction errors selected in descending order. In this case, the period information calculation unit 102 may select a plurality of times t=(t1, t2, ... tN) in descending order of the reconstruction errors, and calculate the difference d between the selected times. i (=|t i -t j |, i ≠ j, 1≦i, j≦N) is calculated as the period information. FIG 8 is a diagram showing a state in which the five largest reconstruction errors are selected from the reconstruction errors shown in FIG 3 .
[0025] In addition, the difference d i Instead of , the difference d i The statistical quantity may be used as the period information. For example, the average, standard deviation, maximum value, minimum value, etc. may be used as the statistical quantity. In addition, the difference d i In addition, the difference d i The statistics may be used as the period information.
[0026] The period information may be a similarity between a reference signal synchronized with the period of the input signal and the reconstruction error. In this case, the period information calculation unit 102 acquires, as a reference signal, for example, an AC voltage signal or a trigger signal applied when acquiring the input signal, and calculates the inter-vector distance between the reconstruction error and the reference signal as the period information.
[0027] The determination unit 103 determines whether or not an abnormal signal exists in the input signal based on the period information or based on the period information and the reconstruction error, and outputs the determination result on the presence or absence of an abnormal signal to an external device such as a display device.
[0028] As a method for determining whether or not an abnormal signal exists in the input signal, for example, a method is available in which the weighted sum of the period information and the reconstruction error is calculated as the degree of abnormality, and whether or not the degree of abnormality is greater than a preset threshold is determined. In this method, for a time when the degree of abnormality is greater than the threshold, the determination unit 103 determines that an abnormal signal exists in the input signal. Also, for a time when the degree of abnormality is equal to or less than the threshold, the determination unit 103 determines that no abnormal signal exists in the input signal.
[0029] Alternatively, as a method for determining whether or not an abnormal signal is present in an input signal, a method may be used in which a vector generated from period information and a reconstruction error is acquired as a feature vector, and a discrimination method using a machine learning model is applied to the acquired feature vector. In this case, the determination unit 103 determines whether or not the probability that the reconstruction error belongs to an abnormal class is greater than a preset threshold, thereby determining whether or not an abnormal signal is present in the input signal. Examples of discrimination methods using a machine learning model include discrimination methods such as neural networks, LightGBM (Light Gradient Boosting Machine), SVM (Support Vector Machine), linear regression, and random forest, and outlier detection methods such as CSVM (One Class SVM), MT (Maharanobis-Taguchi) method, and LOF (Local Outlier Factor) method.
[0030] 9 is a diagram showing an example of a classifier used as a machine learning model in a classification method using a machine learning model. The classifier is trained to receive a reconstruction error and period information as input, and output a normal probability and an abnormal probability of the input reconstruction error. Specifically, the normal probability is the probability that an abnormal signal is not included in the input signal. Also, the abnormal probability is the probability that an abnormal signal is included in the input signal.
[0031] FIG. 10 is a diagram showing an example of a learning method of the classifier shown in FIG. 9. As shown in FIG. 10, when learning the classifier, learning data and a correct answer label associated with the learning data are used. The learning data includes normal data in which an input signal does not include an abnormal signal, and abnormal data in which an input signal includes an abnormal signal. These learning data are sequentially input to the classifier, and the classifier is caused to output normal probabilities and abnormal probabilities as estimation results. Thereafter, the estimation results are compared with the correct answer label, and it is determined whether the estimation results are correct. Then, the classifier is further trained so that the accuracy rate of the estimation results is increased.
[0032] As a method for determining whether or not there is an abnormal signal in the input signal, there is a method for weighting the reconstruction error based on periodic information, and determining whether or not there is an abnormal signal based on the weighted reconstruction error. Here, an example of a method for calculating the weighted reconstruction error will be described. FIG. 11 is a diagram showing an example of a method for calculating the weighted reconstruction error using the reconstruction error shown in FIG. 3. If the difference d between times is used as periodic information, i Hereinafter, an example of a method for calculating the weighted reconstruction error when using the following equation (1) will be described. i The difference in the neighborhood of d i and the difference d i The average value of d ave The reconstruction error r is calculated using the absolute value of the difference between i By weighting the weighted reconstruction error r iw Calculate.
[0033]
number
[0034] In the above method, by considering not only the magnitude of the reconstruction error but also the period of the input signal, the detection accuracy of an abnormal signal having periodicity can be improved. On the other hand, when the signal value of the noise is large, it is difficult to grasp the characteristics of the signal from the reconstruction error. For this reason, when the signal value of the noise is large, it is also preferable to determine the presence or absence of an abnormal signal by referring to only the period information. In this case, for example, the presence or absence of an abnormal signal in the input signal is determined by performing threshold processing on the value of the acquired period information. Alternatively, the presence or absence of an abnormal signal in the input signal may be determined using a machine learning model that inputs the period information of the input signal and outputs the abnormality probability of the input signal.
[0035] Next, the operation of the process executed by the anomaly detection device 100 will be described. FIG. 12 is a flow chart showing an example of the procedure of the anomaly detection process. The anomaly detection process is a process for detecting an anomaly in an equipment based on a signal obtained from an equipment such as a high-voltage power supply facility or a coil. Note that the process procedure in each process described below is merely an example, and each process can be appropriately changed as much as possible. Also, for the process procedure described below, steps can be omitted, replaced, or added as appropriate depending on the embodiment.
[0036] (Abnormality detection processing) (Step S101) The reconstruction error calculation unit 101 first acquires an input signal and calculates a reconstruction error based on the acquired input signal. Then, the reconstruction error calculation unit 101 outputs the calculated reconstruction error to the period information calculation unit 102 and the determination unit 103.
[0037] (Step S102) Next, the reconstruction error calculation unit 101 determines whether or not to correct the reconstruction error calculated in step S101. At this time, the reconstruction error calculation unit 101 determines whether or not to execute a reconstruction error correction process, for example, according to a pre-stored setting or a user input. If the reconstruction error is to be corrected (step S102-Yes), the process proceeds to step S103. If the reconstruction error is not to be corrected (step S102-No), the process proceeds to step S104.
[0038] (Step S103) The reconstruction error calculation unit 101 smoothes the reconstruction error by performing a predetermined smoothing process. After that, the reconstruction error calculation unit 101 corrects the reconstruction error by subtracting the reconstruction error after smoothing from the reconstruction error before smoothing. After that, the process proceeds to step S104.
[0039] (Step S104) The period information calculation unit 102 calculates period information based on the reconstruction error, and outputs the calculated period information to the determination unit 103 .
[0040] (Step S105) The determination unit 103 determines whether or not to weight the reconstruction error. At this time, the determination unit 103 determines whether or not to execute weighting processing of the reconstruction error, for example, according to a pre-stored setting or a user input. If the reconstruction error is to be weighted (step S105-Yes), the process proceeds to step S106. If the reconstruction error is not to be weighted (step S105-No), the process proceeds to step S107.
[0041] (Step S106) The determination unit 103 performs weighting processing on the reconstruction error based on the period information. At this time, the determination unit 103 calculates the weighted reconstruction error using, for example, the above-mentioned formula (1). After performing weighting processing on the reconstruction error, the process proceeds to step S107.
[0042] (Step S107) The determination unit 103 determines whether or not an abnormal signal exists in the input signal based on the period information and the reconstruction error, and outputs the determination result. For example, the determination unit 103 calculates a weighted sum of the period information and the reconstruction error as an abnormality degree, and determines whether or not the abnormality degree is greater than a preset threshold value, thereby determining whether or not an abnormal signal exists in the input signal.
[0043] The effects of the anomaly detection device 100 according to this embodiment will be described below.
[0044] When calculating the reconstruction error from an input signal using principal component analysis, if the input signal contains accidental noise components, the reconstruction error may become locally large. In a method using principal component analysis, anomaly detection is performed by focusing only on the magnitude of the reconstruction error, so the accidental noise components may be erroneously detected as an abnormal signal. In addition, if the reconstruction error becomes locally large due to accidental noise, the reconstruction error at the time when an abnormal signal occurs may be smaller than the reconstruction error at the time when a normal signal occurs, and the abnormal signal may not be detected normally.
[0045] The anomaly detection device 100 according to this embodiment calculates the reconstruction error of an input signal, which is a time-series signal, calculates periodic information that indicates the periodicity of the reconstruction error, and can determine the presence or absence of an abnormal signal in the input signal based on the periodic information or on the periodic information and the reconstruction error.
[0046] The periodic information is a feature that indicates the periodicity of the reconstruction error, and is, for example, a coefficient obtained by Fourier transforming the reconstruction error, a spectrum, a spectrogram, the degree of variation in time of a plurality of reconstruction errors selected based on the magnitude of the value, etc. Alternatively, the periodic information is the degree of variation in two or more times selected based on the magnitude of the reconstruction error.
[0047] When determining whether or not there is an abnormal signal, for example, a weighted sum of the period information and the reconstruction error is calculated as the degree of abnormality, and it is determined whether or not the degree of abnormality is greater than a preset threshold. Alternatively, a vector generated from the period information and the reconstruction error is obtained as a feature vector, and a predetermined identification method using machine learning is applied to the obtained feature vector to determine whether or not there is an abnormal signal.
[0048] When an input signal contains a periodic abnormal signal, the abnormal signal affects the periodicity of the input signal and the reconstruction error. Therefore, the anomaly detection device 100 according to the present embodiment generates periodic information indicating the periodicity of the reconstruction error, and uses the generated periodic information to determine the presence or absence of an abnormal signal in the input signal, thereby making it possible to determine whether the reconstruction error is caused by the periodic abnormal signal or whether the selected reconstruction error is caused by noise that does not have periodicity. This makes it possible to improve anomaly detection performance for periodic abnormal signals.
[0049] The period information may be a reference signal synchronized with the period of the input signal. In this case, the anomaly detection device 100 can obtain the reference signal synchronized with the period of the input signal and calculate the similarity between the reconstruction error and the reference signal as the period information. This configuration can further improve the anomaly detection performance for abnormal signals having periodicity.
[0050] Furthermore, the anomaly detection device 100 according to this embodiment can calculate a smoothed signal based on the reconstruction error and correct the reconstruction error by subtracting the smoothed signal from the reconstruction error. With this configuration, the anomaly detection device 100 according to this embodiment can further improve the anomaly detection performance for an abnormal signal by correcting the reconstruction error using a smoothing process and determining the presence or absence of an abnormal signal using the corrected reconstruction error.
[0051] Furthermore, the anomaly detection device 100 according to the present embodiment performs weighting processing on the reconstruction error based on the periodic information, and determines the presence or absence of an abnormal signal based on the periodic information and the weighted reconstruction error. With this configuration, the anomaly detection device 100 according to the present embodiment can further improve the anomaly detection performance for an abnormal signal by determining the presence or absence of an abnormal signal using the reconstruction error weighted based on the periodic information.
[0052] Second embodiment A second embodiment will be described. This embodiment is a modification of the configuration of the first embodiment as follows. Descriptions of the configuration, operation, and effects similar to those of the first embodiment will be omitted.
[0053] 13 is a diagram showing a configuration of an anomaly detection device 200 according to a second embodiment. The anomaly detection device 200 according to the second embodiment is used in a situation where reference information used to determine a division unit of an input signal can be acquired. The anomaly detection device 200 divides the input signal based on the input signal and the acquired reference information, and detects an anomaly signal based on the divided input signal.
[0054] In the anomaly detection device 200 according to this embodiment, the processing circuitry calls up and executes a program stored in a storage medium, thereby performing the functions of the reconstruction error calculation unit 101, the periodic information calculation unit 102, and the determination unit 103, as well as the function of the signal division unit 201.
[0055] The signal division unit 201 acquires an input signal and reference information. The reference information is information used to determine a division unit of the input signal. For example, the voltage value of the AC voltage applied to the high-voltage power supply equipment or the number of phases of the AC voltage is used as the reference information. Alternatively, the number of trigger signals used when measuring the input signal, or the number or number of rotations of a rotating body included in the device that is the source of the abnormal signal may be used as the reference information.
[0056] The signal division unit 201 divides an input signal into two or more short signals (hereinafter, referred to as short-time signals) using reference information. The signal division unit 201 outputs the generated multiple short-time signals to the reconstruction error calculation unit 101 as input signals.
[0057] The signal dividing unit 201 divides the input signal into a plurality of short-time signals by dividing one period of the voltage phase as a division unit, for example, using the voltage value of the AC voltage or the number of phases of the AC voltage in the input signal as reference information. When the voltage value of the AC voltage is acquired as reference information, the signal dividing unit 201 calculates the length of the phase of one period of the AC voltage using the voltage value of the AC voltage, and divides the input signal based on the calculation result. When the number of phases of the AC voltage included in the input signal is acquired as reference information, the signal dividing unit 201 divides the input signal into the same number of short-time signals as the number of phases. As a result, the input signal is divided into short-time signals including input signals with phases of one period of the AC voltage.
[0058] In addition, when the number of trigger signals used when measuring the input signal, or the number of rotations or number of rotating bodies included in the device that is the source of the abnormal signal is used as reference information, the signal dividing unit 201 divides the input signal into the same number as the number of reference information, and generates the same number of short-time signals as the number of reference information.
[0059] FIG. 14 is a diagram showing an example of an input signal. The horizontal axis of FIG. 14 indicates time. The vertical axis of FIG. 14 indicates the voltage value of the input signal. The input signal shown in FIG. 14 includes a signal equivalent to five periods of AC voltage. When dividing the input signal shown in FIG. 14, the signal dividing unit 201 obtains, for example, the number of periods "5" as reference information, and divides the input signal into five short-time signals. FIG. 15 is a diagram showing five short-time signals generated by dividing the input signal shown in FIG. 14. The horizontal axis of FIG. 15 indicates the value of a label (hereinafter referred to as an index) indicating the timing in the short-time signal. The vertical axis of FIG. 15 indicates the voltage value. The index corresponds to the phase in one period of AC voltage.
[0060] The reconstruction error calculation unit 101 acquires, as input signals, a plurality of short-time signals generated in the signal division unit 201. The reconstruction error calculation unit 101 calculates a reconstruction error for each short-time signal by applying a predetermined calculation method to the acquired short-time signal. The reconstruction error calculation method may be the same as the method described in the first embodiment.
[0061] Fig. 16 is a diagram showing an example of the reconstruction error calculated for each short-time signal for each of five short-time signals. Each of the five graphs shown in Fig. 16 includes the reconstruction error calculated based on a signal equivalent to one period of an AC voltage. The horizontal axis of Fig. 16 indicates the value of the index. The vertical axis of Fig. 16 indicates the value of the reconstruction error.
[0062] The period information calculation unit 102 calculates period information based on the magnitude of the reconstruction error calculated for each short-time signal. The period information is, for example, information on the degree of variation between short-time signals of the timing (index) corresponding to the reconstruction error selected in each short-time signal. For example, the time in the short-time signal or the phase value of the AC voltage is used as the timing in the short-time signal. In this case, the degree of phase variation related to the reconstruction error selected for each short-time signal is used as the period information.
[0063] Here, an example of a method for calculating period information will be described with reference to Fig. 16 and Fig. 17. When calculating period information using the five reconstruction errors shown in Fig. 16, the period information calculation unit 102 first selects the maximum reconstruction error as a representative value from among the reconstruction errors calculated for each short-time signal. Fig. 17 is a diagram showing only the selected representative value. The horizontal axis of Fig. 17 indicates the index. The vertical axis of Fig. 17 indicates the reconstruction error.
[0064] Next, the period information calculation unit 102 extracts the representative value with the largest reconstruction error from the selected five representative values as the maximum value. For each representative value other than the maximum value, the period information calculation unit 102 calculates a distance based on an index between the maximum value and each of the representative values other than the maximum value. At this time, the difference between the index for each of the four representative values other than the maximum value and the index for the maximum value is calculated as the index-based distance between the maximum value and each of the four representative values other than the maximum value. Then, the period information calculation unit 102 outputs the four calculated distances as period information.
[0065] As the period information, instead of or in addition to the above parameters, one or more of the following may be used: a distance based on an index between the top two representative values having the largest reconstruction error among the representative values; a distance based on the smallest index among the distances between the two representative values; a distance based on the largest index among the distances between the two representative values; a distance based on the smallest index among the distances between the top three representative values having the largest reconstruction error; and a standard deviation of the indexes of the representative values.
[0066] The determination unit 103 acquires the standard deviation of the index value of the reconstruction error selected for each short-time signal as period information, and performs threshold determination on the acquired standard deviation to determine whether the selected reconstruction error is caused by an abnormal signal. After that, the determination unit 103 outputs the determination result of the presence or absence of an abnormal signal to an external device such as a display device.
[0067] Next, the operation of the process executed by the anomaly detection device 200 will be described. FIG. 18 is a flowchart showing an example of the procedure of the anomaly detection process. Note that the process procedure in each process described below is merely an example, and each process can be appropriately changed as much as possible. In addition, steps can be omitted, replaced, or added to the process procedure described below as appropriate depending on the embodiment.
[0068] (Abnormality detection processing) (Step S201) The signal dividing unit 201 first acquires an input signal and reference information. The reference information includes the number of phases of the AC voltage of the input signal. Based on the number of phases of the AC voltage, the signal dividing unit 201 divides the input signal into short-time signals with one period of the voltage phase as a division unit, the number of which is the same as the number of phases. The signal dividing unit 201 outputs the short-time signals to the reconstruction error calculation unit 101 as input signals.
[0069] (Step S202) The reconstruction error calculation unit 101 calculates a reconstruction error for each short-time signal. The reconstruction error calculation unit 101 outputs the reconstruction error calculated for each short-time signal to the period information calculation unit 102 and the determination unit 103.
[0070] (Step S203) The period information calculation unit 102 calculates period information for each short-time signal based on the calculated reconstruction error, and outputs the calculated period information to the determination unit 103. Specifically, the period information calculation unit 102 calculates, as period information, the standard deviation of the phase related to the reconstruction error selected for each short-time signal.
[0071] (Step S204) The determination unit 103 determines whether or not an abnormal signal is present in the input signal based on the period information or the period information and the reconstruction error for each short-time signal. For example, the determination unit 103 acquires the standard deviation of the index value of the reconstruction error selected for each short-time signal as period information, and performs threshold judgment on the acquired standard deviation to determine whether or not the selected reconstruction error is caused by the abnormal signal. After that, the determination unit 103 outputs the judgment result to an external device.
[0072] The effects of the anomaly detection device 200 according to this embodiment will be described below.
[0073] The anomaly detection device 200 according to this embodiment can divide an input signal into two or more short-time signals based on reference information, calculate a reconstruction error for each short-time signal, and calculate, as periodic information, the degree of timing variation in the short-time signals with respect to the reconstruction error selected for each of the short-time signals.
[0074] The input signal is, for example, a voltage value of an AC voltage. The reference information is, for example, a voltage value of the AC voltage or a phase number of the AC voltage. The short-time signal includes, for example, an input signal having a phase of one cycle of the AC voltage. The timing in the short-time signal indicates, for example, the phase of the AC current.
[0075] With the above configuration, the anomaly detection device 200 according to this embodiment can, for example, divide an input signal into short-time signals of one phase period and use the degree of variation in timing of the reconstruction error in the short-time signal selected for each short-time signal, thereby making it possible to more accurately determine the periodicity of the reconstruction error. This can further improve the anomaly detection performance for periodic abnormal signals.
[0076] Thus, according to any of the above-described embodiments, it is possible to provide an anomaly detection device, method, and program that can provide an anomaly detection device that can determine with high accuracy the presence or absence of an abnormal signal in an input signal.
[0077] Although some embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope of the invention and its equivalents described in the claims, as well as in the scope and spirit of the invention. [Explanation of symbols]
[0078] 100, 200... anomaly detection device, 101... reconstruction error calculation unit, 102... period information calculation unit, 103... judgment unit, 201... signal division unit.
Claims
1. a reconstruction error calculation unit that calculates a reconstruction error of an input signal that is a time series signal; a periodic information calculation unit that calculates periodic information representing a periodicity of the reconstruction errors, the periodic information being a feature amount generated by performing a Fourier transform on a vector generated by arranging the reconstruction errors; a determination unit that determines whether or not an abnormal signal exists in the input signal based on the period information or based on the period information and the reconstruction error; An anomaly detection device comprising:
2. the period information calculation unit acquires a reference signal synchronized with a period of the input signal, and calculates a similarity between the reconstruction error and the reference signal as the period information; The anomaly detection device according to claim 1 .
3. The period information is a degree of variation between two or more times selected based on the magnitude of the reconstruction error. The anomaly detection device according to claim 1 or 2.
4. a signal division unit that divides the input signal into two or more short-time signals based on reference information; the reconstruction error calculation unit calculates the reconstruction error for each of the short-time signals; The period information calculation unit calculates, as the period information, a degree of timing variation in the short-time signals with respect to a reconstruction error selected in each of the short-time signals. The anomaly detection device according to claim 1 .
5. the input signal is a voltage value of an AC voltage, the reference information is a voltage value of the AC voltage or a phase number of the AC voltage, the short-time signal includes an input signal having a phase corresponding to one cycle of the AC voltage; the timing within the short time signal indicates the phase of the AC voltage; The anomaly detection device according to claim 4.
6. the reconstruction error calculation unit calculates a smoothed signal based on the reconstruction error, and corrects the reconstruction error by subtracting the smoothed signal from the reconstruction error. The anomaly detection device according to any one of claims 1 to 5.
7. the determination unit performs a weighting process on the reconstruction error based on the period information, and determines the presence or absence of the abnormal signal based on the period information and the weighted reconstruction error. The anomaly detection device according to any one of claims 1 to 6.
8. Calculating a reconstruction error of an input signal, which is a time series signal; calculating periodic information, which is a feature quantity generated by performing a Fourier transform on a vector generated by arranging the reconstruction errors, and which represents a periodicity of the reconstruction errors; determining whether or not an abnormal signal exists in the input signal based on the period information or based on the period information and the reconstruction error; A method for providing the above.
9. On the computer, A function to calculate the reconstruction error of the input signal, which is a time series signal; a function of calculating periodic information that is a feature quantity generated by performing a Fourier transform on a vector generated by arranging the reconstruction errors and that represents a periodicity of the reconstruction errors; a function of determining whether or not there is an abnormal signal in the input signal based on the period information or based on the period information and the reconstruction error; A program to achieve this.
Citation Information
Patent Citations
Control type vibration cotnrol supporter
JP1996270723A
System and method for production device control and program
JP2005251925A
Processing state monitoring device
JP2012138044A
Abnormality detector and abnormality detection method
JP2020191050A
Computer program, abnormality determination method and abnormality determination device
JP2022051074A
