JP2022025863A - Apparatus and method for detecting abnormality and program - Google Patents

Abstract

To provide an apparatus for detecting abnormality having a high accuracy of abnormality detection.SOLUTION: An apparatus for detecting abnormality is for detecting abnormality by utilizing a fact that the chronological waveforms occurring are different between at normality and at abnormality. The apparatus for detecting abnormality comprises a converting section 1, a machine learning section 2A, 2B and a determining section 3. The converting section 1 converts a waveform into a graphic multi-dimensional image having a time component and a frequency component. The machine learning section 2A learns in a manner to return an image X1 in which at least one area of a conversion image Y1 obtained by converting a waveform at normality in the converting section 1 is converted into a different image, into the conversion image Y1. The determining section 3 determines abnormality by comparing between an evaluation image Y2 and a prediction image Z, to be outputted by inputting an image X2 in which at least one area of an evaluation image Y2 obtained by converting the waveform to evaluate in the converting section 1 is converted into a different image, into a learnt machine learning section 2B.SELECTED DRAWING: Figure 1

Description

Application for application of Article 30, Paragraph 2 of the Patent Act ・ February 20, 2020, https: // ncsp. jp / registation / proc. Research on anomaly detection devices, anomaly detection methods, and programs was published on html. -From February 28th to March 2nd, 2020, research on anomaly detection devices, anomaly detection methods and programs was disclosed at RISP International Works on Nonliner Circuits, Communications and Signal Processing 2020.

The present invention relates to an abnormality detection device, an abnormality detection method and a program.

Patent Document 1 discloses a conventional abnormality detection device. This abnormality detection device determines an abnormality of the determination object displayed in the input image data. This abnormality detection device generates reconstructed image data from the feature amount extracted from the normal image data group, and performs abnormality determination based on the difference information between the generated reconstructed image data and the input image data. The feature amount can be acquired by using an autoencoder which is one of the neural network models.

Japanese Unexamined Patent Publication No. 2018-5773

However, in general, the autoencoder trains the input image and the output image to be the same, so if most of the input normal image and the abnormal image are similar, it was acquired using the autoencoder. It is difficult to make a difference between the reconstructed image data generated from the feature amount and the input image data. Therefore, when an autoencoder of a general learning model is used, it is difficult for the abnormality detection device of Patent Document 1 to obtain the accuracy of abnormality detection.

The present invention has been made in view of the above-mentioned conventional circumstances, and it is an object to be solved to provide an abnormality detection device, an abnormality detection method and a program having high accuracy of abnormality detection.

The abnormality detection device of the present invention detects an abnormality by utilizing the fact that the generated time-series waveform is different between the normal time and the abnormal time. This abnormality detection device includes a conversion unit, a machine learning unit, and a determination unit. The conversion unit converts the waveform into a graphed multidimensional image having a time component and a frequency component. The machine learning unit learns to return an image in which at least one region of the converted image obtained by converting the waveform in the normal state to a different image is returned to the converted image in the conversion unit. The judgment unit is a prediction image that is output by inputting an image in which at least one area of the evaluation image obtained by converting the waveform to be evaluated in the conversion unit is changed to a different image into the trained machine learning unit. , Judge the abnormality by comparing with the evaluation image.

The abnormality detection method of the present invention detects an abnormality by utilizing the fact that the generated time-series waveform is different between the normal time and the abnormal time. This abnormality detection method includes a conversion step, a learning step, and a determination step. The conversion step converts the waveform into a graphed multidimensional image with time and frequency components. In the learning step, the conversion step is executed, and the machine learning unit is made to learn to return the image in which at least one area of the converted image obtained by converting the waveform in the normal state is changed to a different image. In the judgment step, a conversion step is executed, and an image in which at least one area of the evaluation image obtained by converting the waveform to be evaluated is changed to a different image is input to the trained machine learning unit and output. The abnormality is judged by comparing the predicted image with the evaluation image.

In this abnormality detection device and abnormality detection method, in the conversion unit and the conversion process, at least one area of the conversion image obtained by converting the waveform at the normal time is changed to a different image, and the image is returned to the conversion image. The machine learning department learns. As described above, by making the problem learned by the machine learning unit difficult, the abnormality detection device and the abnormality detection method can improve the accuracy of the learning model. Thereby, the abnormality detection device and the abnormality detection method can clarify the difference between the normal state and the abnormal value in the comparison between the predicted image and the evaluation image in the determination unit. Therefore, this abnormality detection device and the abnormality detection method can simplify the evaluation criteria in the determination unit. Further, the abnormality detection device and the abnormality detection method can improve the accuracy of abnormality detection.

In the anomaly detection device and the anomaly detection method of the present invention, different images can be images of predetermined brightness. In this case, the problem to be learned in the machine learning unit can be easily made difficult.

The predetermined luminance of different images in the anomaly detection device and the anomaly detection method of the present invention may be the average value of the luminance in the region before changing to the predetermined luminance. In this case, the problem to be learned in the machine learning unit can be made moderately difficult.

The machine learning unit in the anomaly detection device and the anomaly detection method of the present invention can randomly change at least one of the position, size, shape, and number of one region to be changed to a different image. In this case, it is possible to eliminate the bias of the problem to be learned in the machine learning unit.

It is a block diagram which shows the learning method of the learning stage in the abnormality detection apparatus of Embodiment 1. FIG. It is a block diagram which shows the abnormality detection method of the stage which detects the abnormality in the abnormality detection apparatus of Embodiment 1. FIG. It is a conceptual diagram which shows the state of a chain. It is a graph which shows the acceleration data in a normal state. It is a graph which shows the acceleration data at the time of an abnormality. It is a graph which shows the degree of abnormality of the comparative example 1. FIG. It is a graph which shows the degree of abnormality of Example 1. FIG. It is a graph which shows the degree of abnormality of Example 2.

The first embodiment which embodies the abnormality detection device and the abnormality detection method of the present invention will be described with reference to the drawings.

<Embodiment 1>
As shown in FIGS. 1 and 2, the abnormality detection device of the first embodiment includes a conversion unit 1, machine learning units 2A and 2B, and a determination unit 3. The conversion unit 1 converts the acceleration data into spectrograms Y1 and Y2. The spectrograms Y1 and Y2 have a time component in the horizontal axis direction and a frequency component in the vertical axis direction, and are graphed multidimensional images representing the power of each frequency band using colors. In the stage where the machine learning unit 2A learns (hereinafter, referred to as "learning stage"), the spectrogram Y1 converted by the conversion unit 1 corresponds to a converted image. At the stage where the abnormality detection device detects an abnormality (hereinafter referred to as “abnormality detection stage”), the spectrogram Y2 converted by the conversion unit 1 corresponds to an evaluation image. Further, the conversion unit 1 converts the acceleration data into masked spectrograms X1 and X2. The "masked spectrograms X1 and X2" are spectrograms in which at least one region of the spectrogram (converted image Y1 and evaluation image Y2) converted by the conversion unit 1 is changed to a different image. That is, "masking" means changing at least one region of the spectrograms Y1 and Y2 to a different image. The "at least one place" is not limited to one place but may be a plurality of places.

As shown in FIG. 1, the abnormality detection device in the learning stage learns to return the spectrogram X1 masked by the machine learning unit 2A to the original spectrogram Y1. The "masked spectrogram" input to the machine learning unit 2A is obtained by converting the normal acceleration data into the masked spectrogram X1 in the conversion unit 1. The "original spectrogram" input to the machine learning unit 2A is the one in which the acceleration data at the normal time is converted into the spectrogram Y1 by the conversion unit 1, and corresponds to the converted image. In this way, the machine learning unit 2A learns to return the spectrogram X1 masked by the normal acceleration data to the spectrogram Y1 (converted image) of the normal acceleration data. The machine learning unit 2A is a kind of autoencoder (hereinafter referred to as "AE") and has a deep convolution encoder / decoder architecture.

As shown in FIG. 2, the abnormality detection device in the abnormality detection stage converts the acceleration data to be evaluated into the spectrogram Y2 in the conversion unit 1. This spectrogram Y2 corresponds to an evaluation image. Further, the abnormality detection device in the abnormality detection stage converts the acceleration data to be evaluated into the masked spectrogram X2 in the conversion unit 1. The masked spectrogram X2 is a spectrogram in which at least one region of the evaluation image Y2 is changed to a different image. The abnormality detection device inputs the masked spectrogram X2 to the trained machine learning unit 2B.

The determination unit 3 compares the predicted image Z output from the trained machine learning unit 2B with the evaluation image Y2. The determination unit 3 determines that the difference between the predicted image Z and the evaluation image Y2 is smaller than a predetermined threshold value, and determines that the difference is large. In making this determination, the weighted movement variance obtained by the peak signal-to-noise ratio (hereinafter referred to as “PSNR”) is used as the degree of anomaly. The PSNR is calculated from the predicted image Z and the evaluation image Y2. The weighted movement variance is calculated by rearranging the PSNRs in order. The weighted window size is 161 frames and the shift is 1 frame. Remove the top and bottom 3% of each frame from the frame to exclude outliers. The variance is obtained from the PSNR of the frame.

PSNR is an index value that looks at the square of the difference between the brightness values of the same pixel in two images. The higher the PSNR value, the higher the similarity between the two images. In the image I and the image K of m × n, the equation of PSNR is shown in the equation 1, and the equation of the mean square error is shown in the equation 2. MAX is the highest value of the luminance area (here, 255).

In the abnormality detection method using such an abnormality detection device, the machine learning unit 2A is learned by using the acceleration data at the normal time (learning stage). After that, the normality and abnormality of the acceleration data to be evaluated are determined by using the learned machine learning unit 2B (abnormality detection stage).

<Learning stage>
As shown in FIG. 1, the conversion unit 1 executes a conversion step of converting a plurality of normal acceleration data into masked spectrogram X1 and spectrogram Y1. Next, a learning step is executed in which the machine learning unit 2A learns so that the masked spectrogram X1 converted from the normal acceleration data is returned to the spectrogram Y1 which is an evaluation image converted from the same normal acceleration data. In the learning step, a predetermined number of epochs is repeatedly executed for the masked spectrogram X1 and the spectrogram Y1 converted from the plurality of normal acceleration data.

<Anomaly detection stage>
After the machine learning unit 2A finishes learning, as shown in FIG. 2, the conversion unit 1 executes a conversion step of converting the acceleration data to be evaluated into a masked spectrogram X2 and a spectrogram Y2. Next, the determination step shown below is executed to determine whether the acceleration data is normal or abnormal. In the determination step, first, the masked spectrogram X2 converted from the acceleration data to be evaluated is input to the trained machine learning unit 2B. Next, the determination unit 3 compares the predicted image Z output from the machine learning unit 2B with the spectrogram Y2, which is an evaluation image, to determine an abnormality.

The functions of each part that executes the abnormality detection method in the above-mentioned abnormality detection device are realized by a computer. The abnormality detection device includes a drive device, a memory device, a CPU, an interface device, and the like. The program to function as each part in the abnormality detection device is provided by a recording medium such as a CD-ROM. When the recording medium in which the program is stored is set in the drive device, the program is installed in the memory device from the recording medium via the drive device. However, the program does not necessarily have to be installed from the recording medium, and may be downloaded from another computer via the network. The CPU executes the functions of each part in the abnormality detection device according to the program stored in the memory device. The interface device is used as an interface for connecting to a network. The anomaly detection device acquires acceleration data via the interface device.

The abnormality detection device described above can detect abnormal behavior of vibration data of a mechanical device and can be used as a countermeasure against a failure of the mechanical device. When anomaly detection of a mechanical device equipped with a chain conveyor is performed using this anomaly detection device (Example 1 and Example 2), the machine learning unit 2B learned by a normal AE learning model is used. The degree of abnormality was compared with that when abnormality was detected (Comparative Example 1). The normal AE learning model is a model for learning so that the input image and the output image are the same. Further, in Comparative Example 1, the image input to the machine learning unit 2A in the learning stage and the image input to the learned machine learning unit 2B in the abnormality detection stage are spectrograms without a mask. That is, in Comparative Example 1, the original spectrogram is used as both the input image and the output image of the machine learning units 2A and 2B. Hereinafter, Example 1, Example 2, and Comparative Example 1 will be described.

In Example 1, Example 2, and Comparative Example 1, vibration data was collected by an accelerometer attached to a mechanical device equipped with a chain conveyor. The accelerometer data collected by the accelerometer is shown as a time series waveform.

Acceleration data was collected when the chain conveyor failed, the chain was replaced, and it was operating normally. That is, the chain conveyor broke down once during data acquisition, took first aid, and then replaced the chain. After replacing the chain, the condition of the chain becomes unstable and the total length tends to be long. For this reason, the chain conveyor cut the chain and adjusted the length shortly after the replacement, and then started to operate normally. As for the acceleration data, as shown in FIG. 3, the data after cutting the chain and adjusting the length was regarded as normal, and the data before the failure was regarded as abnormal. A part of the adjusted normal data was used for learning of the machine learning unit 2A. Table 1 shows the conditions for collecting acceleration data. The normal acceleration data is shown in FIG. The acceleration data at the time of abnormality is shown in FIG.

In the first and second embodiments, the masked spectrograms X1 and X2 input to the machine learning units 2A and 2B are different images of one part of the spectrogram (converted image Y1 and evaluation image Y2) converted by the conversion unit 1. It was changed to. That is, the masked spectrograms X1 and X2 are obtained by masking one part of the converted image Y1 and the evaluation image Y2 with different images. Different images have a given brightness. The predetermined brightness was the average value of the brightness within the range to be changed.

The masked spectrograms X1 and X2 in the first embodiment have a width of 1/32 with respect to the total vertical width of the spectrograms Y1 and Y2, and the entire horizontal width of the spectrograms Y1 and Y2 parallel to the time axis. It is masked by the stretched image. Each time the masked spectrograms X1 and X2 are input to the machine learning units 2A and 2B, the masked position is randomly moved in the vertical direction.

The masked spectrograms X1 and X2 in the second embodiment have a width of 1/8 with respect to the entire vertical width of the spectrograms Y1 and Y2, and are horizontally extended by an image extending over the entire horizontal width of the spectrograms Y1 and Y2. It is masked. Each time the masked spectrograms X1 and X2 are input to the machine learning units 2A and 2B, the masked position is randomly moved in the vertical direction.

The parameters of the spectrograms X1, X2, Y1 and Y2 are shown in Table 2, and the learning parameters are shown in Table 3. In Example 1 and Comparative Example 1, the number of learning epochs in the machine learning unit 2A is 301. In the second embodiment, the number of learning epochs in the machine learning unit 2A is 1001.

In Comparative Example, Example 1, and Example 2, as shown in FIGS. 6 to 8, the degree of abnormality exceeding the threshold value is shown before the failure. The threshold was 90% of the abnormality. From these, in Comparative Example 1, Example 1 and Example 2, it can be predicted that the failure will occur after the threshold value is exceeded.

As shown in FIG. 6, the degree of abnormality in Comparative Example 1 gradually increases before the failure. This indicates that the condition of the chain conveyor gradually deteriorated before the failure. The degree of abnormality in Comparative Example 1 is higher than the degree of abnormality in the normal state from failure to chain replacement. This indicates that the condition of the chain conveyor cannot be completely repaired by a rough repair.

As shown in FIGS. 7 and 8, the degree of abnormality in Examples 1 and 2 temporarily increases after the chain is replaced and after the length of the chain is adjusted. This indicates that the chain does not mesh well for a short time after replacing the chain and after adjusting the length of the chain.

Comparing the graphs of the degree of abnormality of Example 1 and Example 2, the degree of abnormality before failure is higher in Example 2. As a result, the increase in the degree of abnormality due to an unknown cause appears to be relatively small in Example 2 as compared with Example 1. It is considered that the larger the mask width, the larger the degree of abnormality before the failure, and the smaller the change in the value due to the unknown error. It is also believed that masking the spectrogram makes it difficult to reconstruct the spectrogram and makes it more susceptible to major incidents than minor ones.

The abnormality detection device of the first embodiment and the second embodiment detects an abnormality by utilizing the fact that the waveform of the time series generated is different between the normal time and the abnormal time. This abnormality detection device includes a conversion unit 1, machine learning units 2A and 2B, and a determination unit 3. The conversion unit 1 converts the waveform into spectrograms X1, X2, Y1, Y2 having a time component and a frequency component. The machine learning unit 2A extends a region extending over the entire left and right width in the horizontal direction, which is a region of the spectrogram Y1 (converted image) obtained by converting the waveform at the normal time in the conversion unit 1, to the brightness in the region. It is learned to return the masked spectrogram X1 changed to the image of the brightness of the average value of the original spectrogram Y1 (converted image). The determination unit 3 sets a region extending over the entire left-right width in the horizontal direction, which is a region of the spectrogram Y2 (evaluation image) obtained by converting the waveform to be evaluated in the conversion unit 1, into the region of brightness. An abnormality is determined by comparing the predicted image Z output by inputting the masked spectrogram X2 changed to the average brightness image into the trained machine learning unit 2B and the original spectrogram Y2 (evaluation image). ..

The abnormality detection method of the first embodiment and the second embodiment detects an abnormality by utilizing the fact that the generated time-series waveforms are different between the normal time and the abnormal time. This abnormality detection method includes a conversion step, a learning step, and a determination step. The conversion step converts the waveform into spectrograms X1, X2, Y1, Y2 having a time component and a frequency component. In the learning process, the conversion step is executed, and the region extending horizontally over the entire left-right width, which is one region of the spectrogram Y1 (converted image) obtained by converting the waveform at the normal time, is the brightness in the region. The machine learning unit 2A is trained to return the masked spectrogram X1 changed to the image of the brightness of the average value of the above to the original spectrogram Y1 (converted image). In the judgment step, the conversion step is executed, and the region extending horizontally over the entire left-right width, which is one region of the spectrogram Y1 (evaluation image) obtained by converting the waveform to be evaluated, is the brightness in the region. The predicted image Z output by inputting the masked spectrogram X2 changed to the image of the brightness of the average value of the above into the trained machine learning unit 2B is compared with the original spectrogram Y2 (evaluation image) to determine the abnormality. do.

The abnormality detection device and the abnormality detection method are learned by the machine learning unit 2A so as to return the masked spectrogram X1 to the original spectrogram Y1 (converted image) in the conversion unit 1 and the conversion step. As described above, the abnormality detection device and the abnormality detection method can improve the accuracy of the learning model by making the problem learned by the machine learning unit 2A difficult. Thereby, the abnormality detecting device and the abnormality detecting method can clarify the difference between the normal state and the abnormal value in the comparison between the predicted image Z and the evaluation image Y2 in the determination unit 3. Therefore, the abnormality detection device and the abnormality detection method can simplify the evaluation criteria in the determination unit 3 and the determination process. As described above, the abnormality detection device and the abnormality detection method can improve the accuracy of abnormality detection.

In the anomaly detection device and the anomaly detection method of the first and second embodiments, the different images are images having a luminance which is an average value of the luminance in the region before the change. In this case, the problem to be learned in the machine learning unit 2A can be made moderately difficult.

The present invention is not limited to the first embodiment described by the above description and the drawings, and for example, the following embodiments are also included in the technical scope of the present invention.
(1) In Examples 1 and 2, the image is masked with an image having brightness which is the average value of the brightness of the region before the change, but the image to be masked may be an image having arbitrary brightness and is changed. The image may be different from the image in the previous area.
(2) In Examples 1 and 2, one area of the converted image and the evaluation image is changed to a different image, but a plurality of areas may be changed to different images.
(3) In Examples 1 and 2, the region extending in the horizontal direction parallel to the time axis of the converted image and the evaluation image is masked, but the region extending in the vertical direction parallel to the frequency axis may be masked. good.
(4) In Examples 1 and 2, the size of the area to be masked by the machine learning unit was constant in the learning stage, but the size of the area to be masked may be changed according to the learning stage. .. In this case, the size of the masked area is reduced in the early stage of the learning stage to make it a simple problem, and as the learning stage progresses, the size of the masked area is increased to make it difficult to return to the original image. You may let them learn.
(5) In Examples 1 and 2, the abnormality detection device was used to detect an abnormality in a mechanical device equipped with a chain conveyor. In this abnormality detection device, when the generated time-series waveforms are normal and abnormal. If it is different from the above, it is possible to detect an abnormality not only in a mechanical device equipped with a chain conveyor.

1 ... Conversion unit, 2A, 2B ... Machine learning unit, 3 ... Judgment unit, X1, X2 ... Masked spectrogram, Y1 ... Spectrogram (converted image), Y2 ... Spectrogram (evaluation image), Z ... Predicted image

Claims (6)

It is an anomaly detection device that detects anomalies by utilizing the fact that the time-series waveforms that occur are different between normal and abnormal times.
A conversion unit that converts the waveform into a graphed multidimensional image having a time component and a frequency component,
A machine learning unit that learns to return an image in which at least one region of the converted image obtained by converting the waveform in the normal state to a different image in the conversion unit is returned to the converted image.
A predicted image output by inputting an image in which at least one region of the evaluation image obtained by converting the waveform to be evaluated in the conversion unit to a different image is input to the trained machine learning unit. A judgment unit that judges an abnormality by comparing with the evaluation image,
Anomaly detection device equipped with. The abnormality detection device according to claim 1, wherein the different images are images having a predetermined brightness. The abnormality detection device according to claim 2, wherein the predetermined brightness is an average value of the brightness in the region before the change to the predetermined brightness. The machine learning unit randomly changes at least one of the positions, sizes, shapes, and numbers of one region to be changed to the different image according to any one of claims 1 to 3. The described anomaly detector. It is an anomaly detection method that detects anomalies by utilizing the fact that the time-series waveforms that occur are different between normal and abnormal times.
A conversion step of converting the waveform into a graphed multidimensional image having a time component and a frequency component, and
A learning step of executing the conversion step and causing a machine learning unit to learn an image in which at least one region of the converted image obtained by converting the waveform in a normal state is changed to a different image and returned to the converted image. When,
Prediction that the conversion step is executed and an image in which at least one region of the evaluation image obtained by converting the waveform to be evaluated is changed to a different image is input to the trained machine learning unit and output. A judgment step of comparing an image with the evaluation image to determine an abnormality, and
Anomaly detection method equipped with. A program for making a computer function as each part in the abnormality detection device according to any one of claims 1 to 4.

Images