Abnormality detection device, and abnormality detection method

The abnormality detection device uses a CNN to convert waveform signal data into spectrograms for precise identification of abnormality timing and frequency, addressing the limitations of existing devices in determining these parameters.

JP2024161935AActive Publication Date: 2024-11-21ONO SOKKI CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
JP2023077004
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-09
Publication Date
2024-11-21
Estimated Expiration
2043-05-09

AI Technical Summary

Technical Problem

Existing abnormality detection devices struggle to accurately determine the timing and frequency of abnormalities in detection targets based on waveform signal data.

Method used

An abnormality detection device utilizing a convolutional neural network (CNN) with a conversion map, anomaly map generation, and anomaly degree map generation to analyze waveform signal data, converting it into a spectrogram or mel spectrogram for visualization of abnormality timing and frequency.

Benefits of technology

Enables precise identification of abnormality timing and frequency in detection targets, reducing computational load through dimension reduction and efficient similarity calculations, facilitating easy visualization of abnormality areas.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024161935000001_ABST
    Figure 2024161935000001_ABST
Patent Text Reader

Abstract

To provide an abnormality detection device and abnormality detection method that can grasp a timing or frequency in which abnormality occurs in an object targeted for detection.SOLUTION: An abnormality detection device 20 detects abnormality of an object targeted for detection on the basis of waveform signal data 12. The abnormality detection device 20 comprises: a convolution neural network 25 to which a conversion map converting the waveform signal data 12, and having strength defined according to a time and a frequency is input; an abnormality map generation unit 40 that generates an abnormality map for each intermediate layer of the convolution neural network 25; a degree-of-abnormality map generation unit 45 that generates a degree of abnormality map on the basis of the abnormality map of each intermediate layer; and an output unit 50 that outputs the degree of abnormality map as visualized data.SELECTED DRAWING: Figure 3
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to an abnormality detection device and an abnormality detection method for detecting an abnormality based on waveform signal data. [Background technology]

[0002] Conventionally, there has been known anomaly detection devices that detect anomalies in a detection target by using waveform signal data obtained by measuring sounds and vibrations when the detection target device is driven. For example, Patent Document 1 discloses an anomaly detection device that detects anomalies in a detection target by comparing the results of frequency analysis of waveform signal data of a normal device with the results of frequency analysis of waveform signal data of the detection target. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2020-134479 A Summary of the Invention [Problem to be solved by the invention]

[0004] Incidentally, in order to identify the cause of an anomaly in a detection target, the timing and frequency at which the anomaly occurs are very important. In this regard, although the anomaly detection device described in Patent Document 1 is capable of detecting that the detection target is abnormal, it is difficult to grasp the timing and frequency at which the anomaly occurs. [Means for solving the problem]

[0005] The anomaly detection device that solves the above problem detects an anomaly of a detection target based on waveform signal data. The anomaly detection device includes a convolutional neural network to which a transformation map obtained by transforming the waveform signal data and having an intensity defined corresponding to time and frequency is input, an anomaly map generating unit that generates an anomaly map for each intermediate layer of the convolutional neural network, an anomaly degree map generating unit that generates an anomaly degree map based on the anomaly map of each intermediate layer, and an output unit that outputs the anomaly degree map. The anomaly map generating unit executes a process of acquiring a feature map output by the convolutional layer of the intermediate layer for each channel, a process of calculating a similarity of the feature map to a normal feature map output by the convolutional layer when a normal transformation map obtained by transforming normal waveform signal data of the detection target is input to the convolutional neural network for each channel, and a process of generating a map for each channel in which each component of the feature map is weighted by the similarity, and generating a map in which the generated maps are summed for each component as the anomaly map. The anomaly degree map generation unit executes a process of resizing the anomaly map of each intermediate layer to the same size as the transformation map, and a process of generating a map by summing the resized anomaly maps for each preset component as the anomaly degree map. Effect of the Invention

[0006] According to the present invention, it is possible to grasp the timing and frequency at which an abnormality occurs in a detection object. [Brief description of the drawings]

[0007] [Figure 1] 1 is a diagram illustrating a schematic configuration when an embodiment of an anomaly detection device is used to detect an anomaly in a detection target. FIG. [Diagram 2] FIG. 2 is a block diagram showing an example of a hardware configuration of an information processing device. [Diagram 3] 1 is a block diagram showing an example of a schematic configuration of an anomaly detection device; [Figure 4]FIG. 1 is a diagram illustrating an example of a schematic configuration of a convolutional neural network. [Diagram 5] 2 is a functional block diagram showing a configuration for detecting an abnormality in a detection target in the anomaly detection device. FIG. [Figure 6] 2 is a functional block diagram showing a configuration related to generation of an abnormality degree map in the anomaly detection device. FIG. [Figure 7] FIG. 1A is a diagram showing an example of an image represented by a transformation map, FIG. 1B is a diagram showing an example of an image represented by an abnormality degree map, and FIG. 1C is a diagram showing an example of a superimposed image. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] An embodiment of an anomaly detection device and an anomaly detection method will be described with reference to FIGS.

[0009] As shown in Fig. 1, the anomaly detection device 20 is a device that outputs an anomaly detection result for the detection target 10 based on waveform signal data 12 measured by a measurement sensor 11 for the detection target 10. The waveform signal data is acoustic data in which a sound pressure level is defined as intensity with respect to time, and vibration data in which a vibration level is defined as intensity with respect to time. In addition, the anomaly detection device 20 is a device that outputs visualization data, which is image data that visualizes the timing and frequency at which an anomaly occurs, when an anomaly occurs in the detection target 10. The anomaly detection device 20 is mainly configured with an information processing device.

[0010] (Hardware configuration) 2, a hardware configuration of an information processing device H10 constituting the anomaly detection device 20 will be described. The information processing device H10 includes a communication device H11, an input device H12, a display device H13, a storage device H14, and a processor H15. Note that this hardware configuration is an example, and it is also possible to realize the information processing device H10 using other hardware.

[0011] The communication device H11 is an interface that establishes a communication path with another device and transmits and receives data, and is, for example, a network interface or a wireless interface.

[0012] The input device H12 is a device that accepts input from a user, etc., and is, for example, a mouse, a keyboard, etc. The display device H13 is, for example, a display that displays various information. A touch panel display or the like functions as the input device H12 and the display device H13.

[0013] The storage device H14 is a storage device that stores data and various programs for executing various functions of the anomaly detection device 20. Examples of the storage device H14 include a ROM, a RAM, and a hard disk.

[0014] The processor H15 controls each process using programs and data stored in the storage device H14. Examples of the processor H15 include a CPU and an MPU. The processor H15 loads programs stored in a ROM or the like into a RAM and executes various processes for each service.

[0015] The processor H15 is not limited to a processor that performs software processing for all the processes it executes. For example, the processor H15 may include a dedicated hardware circuit (e.g., an application specific integrated circuit (ASIC)) that performs hardware processing for at least a part of the processes it executes. That is, the processor H15 may be configured as follows:

[0016] [1] One or more processors operating according to a computer program (software); [2] One or more dedicated hardware circuits that perform at least some of the various processes; or [3] Circuits that include combinations of these The processor includes a CPU and memory, such as RAM and ROM, that stores program codes or instructions configured to cause the CPU to execute processes. Memory, i.e., computer-readable media, includes any available media that can be accessed by a general-purpose or special-purpose computer.

[0017] (Anomaly detection device) As shown in FIG. 3, the anomaly detection device 20 includes, as functional units that function by executing a program, a conversion unit 21, a convolutional neural network 25 (hereinafter referred to as CNN 25), an anomaly detection unit 35, an anomaly map generation unit 40, an anomaly degree map generation unit 45, and an output unit 50.

[0018] (Conversion section) The conversion unit 21 converts the waveform signal data 12 into a conversion map 22 (see FIG. 4, etc.) and executes an input process of inputting the conversion map 22 to the CNN 25. The conversion map 22 is map data in which the intensity is defined in correspondence with time and frequency. Specifically, the conversion unit 21 converts the waveform signal data 12 into image data representing a spectrogram or a mel spectrogram. FIG. 7(a) shows, as an example, an image 55 represented by the conversion map 22 in which the waveform signal data 12 is converted into a mel spectrogram. In the following, the map data in which values ​​are arranged in a matrix form is referred to as a map.

[0019] (CNN) CNN25 is a pre-trained classification model such as PANNs (Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition). PANNs is a model trained for classifying environmental sounds using image data converted from waveform signal data of environmental sounds called AudioSet.

[0020] As shown in Fig. 4, the CNN 25 has an input layer 26, multiple intermediate layers 27, and an output layer 28. In the CNN 25, the input layer 26 divides the transformation map 22 into multiple channels based on the configuration of the transformation map 22 and inputs them to the uppermost intermediate layer 27. For example, when the transformation map 22 is RGB image data, the input layer 26 divides the transformation map 22 into R, G, and B channels and outputs them to the intermediate layer 27. The output layer 28 outputs the classification results for each class based on the output of the lowest intermediate layer 27.

[0021] As shown in FIG. 5, each hidden layer 27 of the CNN 25 includes a convolutional layer 30 and a GAP layer 32. The convolutional layer 30 generates a feature map x^ for each channel by performing convolution using a predetermined filter on the map data input from the upper layer. The convolutional layer 30 outputs the generated feature map x^ to the GAP layer 32 and the convolutional layer 30 of the lower intermediate layer 27. In addition, the lowest intermediate layer 27 outputs the generated feature map x^ to the output layer 28.

[0022] The GAP layer 32 is a global average pooling layer. The GAP layer 32 calculates the average value of all components of the feature map x^ for each channel, and then generates a feature vector x composed of these average values.

[0023] (Abnormality detection section) 5, the anomaly detection unit 35 executes a process of determining the presence or absence of an anomaly in the detection target 10. The anomaly detection unit 35 includes a similarity calculation unit 36 ​​and a determination unit 39.

[0024] The similarity calculation unit 36 ​​acquires the feature vector x for each intermediate layer 27, and then executes a process of calculating the similarity of the feature vector x to a normal feature vector. This similarity is used when detecting an abnormality in the detection target 10. Note that in Fig. 5, the similarity calculation unit 36 ​​corresponding to the uppermost intermediate layer 27 and the similarity calculation unit 36 ​​corresponding to the lowermost intermediate layer 27 are shown, and the similarity calculation units 36 corresponding to the other intermediate layers 27 are omitted.

[0025] Specifically, the similarity calculation unit 36 ​​calculates the Mahalanobis distance M of the feature vector x with respect to the normal feature vector as the similarity. The normal feature vector is a feature vector obtained when a normal transformation map obtained by transforming normal waveform signal data of a normal detection target 10 is input to the CNN 25.

[0026] The similarity calculation unit 36 ​​calculates the Mahalanobis distance M based on the normal data 37, which is information about the normal feature vector. The normal data 37 is data that the anomaly detection unit 35 and the anomaly map generation unit 40 hold in the storage device H14. The normal data 37 is data for each intermediate layer 27, and is composed of data based on a large number of normal feature vectors obtained when a large number of normal transformation maps are input to the CNN 25. The normal data 37 is composed of an average vector μ having the average value of each channel in the large number of normal feature vectors as its component, and a variance-covariance matrix Σ based on the normal feature vectors and the average vector μ. The similarity calculation unit 36 ​​calculates the Mahalanobis distance M of the feature vector x according to the following formula (1). As shown in FIG. 5, "Level" indicates the hierarchy of each intermediate layer 27, with the uppermost intermediate layer 27 being Level 1, and the lower the intermediate layer 27, the larger the value.

[0027]

number

[0028] Furthermore, the mean vector μ and the variance-covariance matrix Σ may be approximated as in the following equations (2) and (3): The feature vector x in equations (2) and (3) represents a normal feature vector.

[0029]

number

[0030] The determination unit 39 executes a process of determining whether the detection object 10 is normal or not based on the similarity of each intermediate layer 27, and outputting the determination result to the output unit 50 as an abnormality detection result. Specifically, the determination unit 39 determines that the detection object 10 is normal when a determination value obtained by summing the Mahalanobis distances M of each intermediate layer 27 is equal to or less than a predetermined threshold. The determination unit 39 determines that the detection object 10 is abnormal when the determination value exceeds the threshold. The determination unit 39 outputs the determination result to the output unit 50 as an abnormality detection result.

[0031] (Anomaly map generation section) As shown in FIG. 6, the anomaly map generating unit 40 executes an anomaly map generating process in which it generates an anomaly map for each intermediate layer 27 and outputs the generated anomaly map to an anomaly degree map generating unit 45.

[0032] The abnormality map generating unit 40 includes a similarity calculating unit 41 and a map generating unit 43 . The similarity calculation unit 41 calculates the similarity of the feature vector x to the normal feature vector for each channel of each hidden layer 27. This similarity is used to generate an abnormality map.

[0033] Specifically, the similarity calculation unit 41 calculates the Mahalanobis distance M* of the feature vector x to the normal feature vector for each channel as the similarity. The similarity calculation unit 41 calculates the Mahalanobis distance M* of the feature vector x based on the normal data 37 and the following formula (4). In formula (4), "○" indicates the Hadamard product. Furthermore, the sum of the Mahalanobis distances M* for each channel calculated by formula (4) is the same value as the Mahalanobis distance M calculated by the above formula (1).

[0034]

number

[0035] The map generating unit 43 generates an anomaly map for each intermediate layer 27 based on the Mahalanobis distance M* and the feature map x^. Specifically, as shown in the following formula (5), the map generating unit 43 generates a map in which each component of the feature map x^ is weighted by the Mahalanobis distance M* for each channel CH. Then, the map generating unit 43 generates an anomaly map X by adding up the weighted maps for each component. The size of the anomaly map X becomes smaller for the intermediate layer 27 at a lower hierarchical level. The anomaly map generating unit 40 outputs the anomaly map X for each intermediate layer 27 to the anomaly degree map generating unit 45.

[0036]

number

[0037] In addition, in FIG. 6, only the similarity calculation units 41 and the map generation units 43 corresponding to the topmost intermediate layer 27 and the bottommost intermediate layer 27 are shown, and the similarity calculation units 41 and the map generation units 43 corresponding to the other intermediate layers 27 are omitted.

[0038] (Abnormality degree map generation section) As shown in FIG. 6 , the anomaly degree map generating unit 45 executes an anomaly degree map generating process to acquire the anomaly map X of each intermediate layer 27 output by the anomaly map generating unit 40, and output the anomaly degree map generated based on the acquired anomaly map X to the output unit 50 as visualized data.

[0039] The abnormality degree map generating unit 45 includes a resizing unit 46 and a map generating unit 48 . The resizing unit 46 resizes the anomaly map X of each intermediate layer 27 into an anomaly map X′ of the same size as the transformation map 22 using various interpolation methods such as nearest neighbor interpolation, bilinear interpolation, and bicubic interpolation.

[0040] The map generating unit 48 generates a map by summing the resized abnormality map X' for each preset component as the abnormality degree map Z, as shown in the following formula (6). For example, the map generating unit 48 generates the abnormality degree map Z by summing the resized abnormality map X' for each component. In addition, for example, the map generating unit 48 generates the abnormality degree map Z by dividing each resized abnormality map X' into a plurality of regions and summing the representative values ​​of the divided regions for each region.

[0041]

number

[0042] The abnormality degree map Z generated in this manner has the same size as the conversion map 22, and is map data in which the degree of abnormality is defined in accordance with time and frequency. The map generation unit 48 outputs the generated abnormality degree map Z to the output unit 50.

[0043] (Output section) The output unit 50 executes an output process for outputting the abnormality detection result and the abnormality degree map Z, which is visualization data. The output unit 50 outputs the abnormality detection result and the abnormality degree map Z to the display device H13, thereby displaying the abnormality detection result and an image represented by the abnormality degree map Z on the display device H13. The output unit 50 may also display a superimposed image in which the image represented by the transformation map 22 and the image represented by the abnormality degree map Z are superimposed on the display device H13.

[0044] (action) FIG. 7(a) shows an image 55 represented by the conversion map 22. FIG. 7(b) shows an image 56 represented by the abnormality degree map Z. FIG. 7(c) shows a superimposed image 57 in which the images 55 and 56 are superimposed. In the images 55, 56, and 57, the horizontal axis indicates the time axis, and the vertical axis indicates the frequency axis. The time axis and the frequency axis have the same scale in the images 55, 56, and 57. In the images 55 and 56, the lighter the color, the higher the intensity. In the image 55, the part pointed to by the arrow 55a is the abnormal part. As shown in FIG. 7(b), in the image 56, the intensity (degree of abnormality) of the part corresponding to the abnormal part is high, that is, it is displayed in a light color.

[0045] The operation and effects of this embodiment will be described. (1) In the image 56, the intensity of the parts corresponding to the abnormal parts is displayed higher. As a result, the timing and frequency at which the abnormality occurs in the detection target 10 can be visualized.

[0046] (2) The CNN 25 has a GAP layer 32 for each channel in which the feature map x^ is input in each hidden layer 27. The anomaly map generator 40 calculates the similarity based on the output of the GAP layer 32. With this configuration, it is possible to calculate the similarity with a smaller dimension than the feature map x^. As a result, it is possible to reduce the load when calculating the similarity.

[0047] (3) The abnormality map generator 40 calculates the Mahalanobis distance M* as the similarity using the mean vector μ for each component obtained when the normal feature map is input to the GAP layer 32 and the variance-covariance matrix Σ based on the values ​​of the components. In addition, the mean vector μ and the variance-covariance matrix Σ are held as the normal data 37. This configuration can further reduce the load when calculating the similarity.

[0048] (4) The abnormality degree map generating unit 45 sums up the resized abnormality map X' for each of some preset components to generate the abnormality degree map Z. This configuration can reduce the load when generating the abnormality degree map Z compared to the case where all components are summed for each component.

[0049] (5) The output unit 50 displays on the display device H13 a superimposed image 57 obtained by superimposing an image 55 represented by the transformation map 22 and an image 56 represented by the abnormality degree map Z. According to this configuration, the abnormality degree map is superimposed on the transformation map, so that areas in the transformation map with a high degree of abnormality can be easily grasped.

[0050] (6) Anomaly detection device 20 includes a conversion unit 21 that executes a process of converting waveform signal data 12 into conversion map 22. According to this configuration, simply by inputting waveform signal data 12 to anomaly detection device 20, it is possible to detect an abnormality in detection target 10, and when an abnormality occurs in detection target 10, it is possible to grasp the timing of the abnormality occurrence and the abnormality frequency.

[0051] (7) The transformation map 22 is a spectrogram or a mel spectrogram. In this way, the waveform signal data 12 can be transformed into a spectrogram or a mel spectrogram. In addition, since the transformation map 22 is a mel spectrogram, the dimension of the frequency axis in the transformation map 22 can be reduced, and therefore the amount of information handled when generating the CNN 25 or various maps can be reduced.

[0052] (8) The waveform signal data 12 is acoustic data in which a sound pressure level is defined as intensity with respect to time, or vibration data in which a vibration level is defined as intensity with respect to time. With this configuration, the image 56 represented by the abnormality degree map Z makes it possible to grasp the timing at which abnormal sounds or abnormal vibrations occur, making it easier for humans to identify the abnormal sounds or abnormal vibrations.

[0053] (9) The resizing unit 46 resizes the abnormality map X to the same size as the conversion map 22 using nearest neighbor interpolation. In this case, it is possible to generate an abnormality map X' and an abnormality degree map Z in which the boundary of the change in the abnormality degree is clear. Furthermore, the resizing unit 46 resizes the abnormality map X to the same size as the conversion map 22 using an interpolation method such as bilinear interpolation or bicubic interpolation. In this case, it is possible to generate an abnormality map X' and an abnormality degree map Z in which the change in the abnormality degree is smooth.

[0054] This embodiment can be modified as follows: This embodiment and the following modifications can be combined with each other to the extent that there is no technical contradiction. The waveform signal data 12 is not limited to acoustic data or vibration data, as long as it can be measured by the measurement sensor 11 and has a specified intensity with respect to time.

[0055] The converter 21 converts the waveform signal data 12 using a spectrogram or a mel spectrogram as the conversion map 22. However, the converter 21 may convert the waveform signal data 12 into map data in which the intensity is defined corresponding to time and frequency. In other words, the converter 21 is not limited to a configuration in which the converter 21 converts the waveform signal data 12 using a spectrogram or a mel spectrogram as the conversion map 22.

[0056] The method for calculating the similarity is not limited to the method using the Mahalanobis distance, but may be a method using the Euclidean distance, clustering, or the like. The anomaly detection device 20 only needs to be able to detect an anomaly and output an anomaly degree map based on the conversion map 22. Therefore, the conversion unit 21 may be omitted in the anomaly detection device 20. In this case, map data obtained by converting waveform signal data in an external device is input to the anomaly detection device 20 as the conversion map 22.

[0057] The pooling layer is not limited to the GAP layer 32, but may be any layer that can reduce the dimension of the feature map x^ for each channel. Therefore, the pooling layer may be a normal pooling layer inserted between the convolution layers 30 so as to correspond to each channel. [Explanation of symbols]

[0058] H10...information processing device, H11...communication device, H12...input device, H13...display device, H14...storage device, H15...processor, 10...detection object, 11...measurement sensor, 12...waveform signal data, 20...anomaly detection device, 21...conversion unit, 22...conversion map, 25...convolutional neural network, 26...input layer, 27...intermediate layer, 28...output layer, 30...convolution layer, 32...GAP layer, 35...anomaly detection unit, 36...similarity calculation unit, 37...normal data, 39...judgment unit, 40...anomaly map generation unit, 41...similarity calculation unit, 43...map generation unit, 45...anomaly degree map generation unit, 46...resize unit, 48...map generation unit, 50...output unit, 55, 56...image, 57...superimposed image.

Claims

1. An anomaly detection device that detects an anomaly of a detection target based on waveform signal data, a convolutional neural network to which a transformation map obtained by transforming the waveform signal data, the transformation map having strengths defined in correspondence with time and frequency, is input; an anomaly map generator that generates an anomaly map for each intermediate layer of the convolutional neural network; an anomaly degree map generating unit that generates an anomaly degree map based on the anomaly maps of each intermediate layer; an output unit that outputs the abnormality degree map, The anomaly map generation unit is A process of acquiring a feature map output by the intermediate convolutional layer for each channel; a process of calculating a similarity of the feature map for each channel to a normal feature map output by the convolutional layer when a normal transformation map obtained by transforming the normal waveform signal data of the detection target is input to the convolutional neural network; generating a map for each of the channels by weighting each component of the feature map by the similarity, and generating a map for the anomaly map by summing the generated maps for each component; The abnormality degree map generating unit resizing each hidden layer anomaly map to the same size as the transformation map; and generating a map obtained by adding up the resized anomaly maps for each preset component as the anomaly degree map. Anomaly detection device.

2. The convolutional neural network includes, for each of the channels, a pooling layer to which the feature map is input, in each of the intermediate layers; The anomaly map generator calculates the similarity based on the output of the pooling layer. The anomaly detection device according to claim 1 .

3. The anomaly map generation unit is The average value of each component obtained when the normal feature map is input to the pooling layer and a variance-covariance matrix based on the values ​​of the respective components are stored, and a Mahalanobis distance is calculated as the similarity using the average value and the variance-covariance matrix. The anomaly detection device according to claim 2 .

4. The abnormality degree map generating unit The resized anomaly map is summed for each of a portion of components set in advance to generate the anomaly degree map. The anomaly detection device according to claim 1 .

5. The output unit displays an image represented by the abnormality degree map and an image represented by the transformation map in a superimposed manner. The anomaly detection device according to claim 1 .

6. The waveform signal data is converted into the conversion map. The anomaly detection device according to any one of claims 1 to 5.

7. The transform map is a spectrogram or a mel-spectrogram. The anomaly detection device according to claim 6.

8. The waveform signal data is sound data in which a sound pressure level is defined as intensity with respect to time, or vibration data in which a vibration level is defined as intensity with respect to time. The anomaly detection device according to claim 7.

9. 1. An anomaly detection method for detecting an anomaly in a detection target based on waveform signal data, comprising: An anomaly detection device that detects an anomaly in the detection target, an input process of inputting a transformation map obtained by transforming the waveform signal data, the transformation map having an intensity defined in correspondence with time and frequency, into a convolutional neural network; an anomaly map generation process for generating an anomaly map for each intermediate layer of the convolutional neural network; an anomaly degree map generation process for generating an anomaly degree map indicating an anomaly degree for each region of the transformation map based on the anomaly map of each intermediate layer; An output process for outputting the abnormality degree map is executed; In the anomaly map generation process, A process of acquiring a feature map output by the intermediate convolutional layer for each channel; a process of calculating a similarity of the feature map for each channel to a normal feature map output by the convolutional layer when a normal transformation map obtained by transforming the normal waveform signal data of the detection target is input to the convolutional neural network; generating a map for each of the channels by weighting each component of the feature map by the similarity, and generating a map for the anomaly map by summing the generated maps for each component; In the abnormality degree map generation process, resizing each hidden layer anomaly map to the same size as the transformation map; and generating a map obtained by adding up the resized anomaly maps for each preset component as the anomaly degree map. Anomaly detection methods.

Citation Information

Patent Citations

  • Abnormality detection device, abnormality detection method, and abnormality detection program

    JP2020042519A

  • Image processing device, image processing method, learning model generation device, learning model manufacturing method, learned model and program

    JP2022074861A

  • Abnormality detection program, abnormality detection apparatus, abnormality detection system, and abnormality detection method

    JP2023005822A

  • Abnormality detection method and abnormality detection device

    JP2020134479A