JP2024027788A - Current waveform analysis device and current waveform analysis method - Google Patents

Abstract

The present invention provides a current waveform analysis device and a current waveform analysis method that can realize highly accurate abnormality determination. [Solution] A current waveform analyzer 1 that analyzes a measured current waveform, which includes an image data creation section 2 that creates two-dimensional image data from the measured current waveform, and the two-dimensional image data trained. The analysis unit 3 calculates latent variable data by inputting the two-dimensional image data created by the image data creation unit 2 according to the current waveform to be analyzed into the trained neural network, and the analysis unit 3 calculates the latent variable data. The image data creation unit 2 includes a display unit 4 that displays latent variable data calculated when the current waveform is in a normal state, and latent variable data calculated when the current waveform is in an abnormal state. Provided is a current waveform analyzer 1 that creates the two-dimensional image data using a procedure that maximizes the difference from the latent variable data calculated in the current waveform analysis device 1. [Selection diagram] Figure 1

Description

The present invention relates to a technique for analyzing current waveforms.

In recent years, various techniques for detecting abnormalities using trained models obtained by machine learning have been devised, but paragraph [0015], paragraph [0017], FIGS. 7 and 10 of Patent Document 1, In their explanatory sections, techniques are disclosed for detecting the occurrence of abnormal sounds by evaluating test image data obtained by converting sound data caused by the motion of the test object.

JP 2021-189039 Publication

However, depending on the conversion method or the inspection image data, there is a problem that sufficient detection accuracy cannot be obtained.

The present invention was made to solve such problems, and an object of the present invention is to provide a current waveform analysis device and a current waveform analysis method that can realize highly accurate abnormality determination.

In order to solve the above problems, the present invention provides a current waveform analysis device that analyzes a measured current waveform, and includes an image data creation means that creates two-dimensional image data from the measured current waveform, and An analysis means for calculating latent variable data by inputting two-dimensional image data created by an image data creation means according to a current waveform to be analyzed into a trained neural network that has learned the data; The image data creation means displays the latent variable data calculated when the current waveform is in a normal state and the latent variable data calculated when the current waveform is in an abnormal state. Provided is a current waveform analysis device that creates the two-dimensional image data using a procedure that maximizes the difference from the latent variable data calculated in the current waveform analysis device.

In order to solve the above problems, the present invention provides a current waveform analysis method for analyzing a measured current waveform, which includes a first step of creating two-dimensional image data from the measured current waveform, and a first step of creating two-dimensional image data from the measured current waveform. a second step of calculating latent variable data by inputting the two-dimensional image data created in the first step according to the current waveform to be analyzed into the trained neural network that has trained the data; The third step displays the latent variable data calculated in the second step, and the first step displays the latent variable data calculated when the current waveform is in a normal state and the current waveform. Provided is a current waveform analysis method that creates the two-dimensional image data using a procedure that maximizes the difference from latent variable data calculated when the image data is in an abnormal state.

According to the present invention, it is possible to provide a current waveform analysis device and a current waveform analysis method that can realize highly accurate abnormality determination.

1 is a block diagram showing the configuration of a current waveform analyzer 1 according to an embodiment of the present invention. 2 is a diagram showing an example of the current waveform analyzer 1 shown in FIG. 1. FIG. 3 is a flowchart showing a current waveform analysis method according to an embodiment of the present invention. 4 is a diagram showing an example of the current waveform analysis method shown in FIG. 3. FIG. 5 is a diagram showing an example in which the two-dimensional image data shown in FIG. 5(b) is created from the current waveform shown in FIG. 5(a) in step S1 shown in FIG. 3. FIG. 6 is a diagram showing an example in which the two-dimensional image data shown in FIG. 6(b) is created from the current waveform shown in FIG. 6(a) in step S1 shown in FIG. 3. FIG. FIG. 7A is a first diagram for explaining the shaping process shown in step S13 of FIG. 4, and FIG. 7(b) is a diagram showing an example of padding the pixels. FIG. 8A is a second diagram for explaining the shaping process shown in step S13 of FIG. 4, and FIG. b) is a first example in which the pixels are arranged on a two-dimensional plane in the order shown by the arrows, and FIG. 8(c) is a second example in which the pixels are arranged in the order shown in the arrows on a two-dimensional plane. FIG. 8(d) shows a third example in which the pixels are arranged on a two-dimensional plane in the order indicated by arrows. This is a third diagram for explaining the shaping process shown in step S13 of FIG. 4, in which FIG. 9(a) is an example of arranging the pixels in the procedure indicated by the arrow, and FIG. 9(c) is the procedure shown in FIG. 9(a) rotated 90 degrees clockwise, and FIG. 9(c) is the procedure shown in FIG. 9(a) rotated 180 degrees clockwise. 9(d) shows the procedure shown in FIG. 9(a) rotated 270 degrees clockwise, and FIG. 9(e) shows the procedure shown in FIG. 9(a) reversed horizontally. . This is a first diagram showing the results of the visualization processing in steps S17 and S18 in FIG. 4, in which FIG. 10(a) is an example of current waveforms forming the reference group and the analysis target group, and FIG. 10(b) is the reference group. Examples of latent variable scatter plots generated from the and analysis target groups are shown. 11(b) is a second diagram showing the results of the visualization processing in steps S17 and S18 in FIG. 4, and FIG. An example of a latent variable scatter plot generated when the number of layers is increased is shown. 4 is a graph showing an example of the current waveform shown in FIG. 3. 13(a) shows a current waveform measured when the electronic board is in a normal state, and FIG. 13(b) shows an example of the current waveform shown in FIG. 3. Shows the current waveform measured when a temperature load is applied to create an abnormal state. This is a third diagram showing the results of the visualization processing in steps S17 and S18 in FIG. 4, and FIG. 14(a) shows a latent variable scatter diagram generated corresponding to the current waveform shown in FIG. 13(a). 14(b) shows a latent variable scatter diagram generated corresponding to the current waveform shown in FIG. 13(b).

Embodiments of the present invention will be described in detail below with reference to the drawings. Note that the same reference numerals in the figures indicate the same or equivalent parts.

FIG. 1 is a block diagram showing the configuration of a current waveform analyzer 1 according to an embodiment of the present invention. As shown in FIG. 1, the current waveform analysis device 1 according to the embodiment of the present invention includes a node N, a bus B connected to the node N, and an image data creation unit 2 connected to the bus B, respectively. It includes an analysis section 3, a display section 4, and a storage section 5.

Here, the image data creation unit 2 uses latent variable data that is calculated when the current waveform measured at the terminal or electronic board of the IoT device is in a normal state, and latent variable data that is calculated when the current waveform is in an abnormal state. Two-dimensional image data is created from the measured current waveform using a procedure that maximizes the difference in latent variable data. Note that the method for creating two-dimensional image data using the above procedure will be explained in detail later.

The analysis unit 3 sends the two-dimensional image created by the image data creation unit 2 from the current waveform to be analyzed to a trained neural network trained on the two-dimensional image data obtained corresponding to each of the above-mentioned states. By inputting data, latent variable data in a multidimensional space is calculated.

The display unit 4 compresses the high-dimensional latent variable data calculated by the analysis unit 3 into two-dimensional data and displays it on a monitor or the like, and the storage unit 5 stores various programs and data.

FIG. 2 is a diagram showing an embodiment of the current waveform analyzer 1 shown in FIG. As shown in FIG. 2, the image data creation section 2, analysis section 3, and display section 4 shown in FIG. 1 are realized by, for example, a central processing unit (CPU) 10. Note that the storage unit 5 is realized by a memory, a hard disk, or the like. An overview of the operation of the current waveform analyzer 1 according to this embodiment will be described below with reference to FIG. 2.

As shown in FIG. 2, the user connects to the CPU 10 via the network NW and uses the web user interface IF deployed on the browser to send a waveform data file representing a current waveform to the current waveform analyzer 1. The storage unit 5 stores this file as a temporarily saved waveform data file.

The CPU 10 creates two-dimensional image data corresponding to the temporarily stored waveform data file stored in the storage unit 5, causes the neural network to learn, and stores the resulting weight file in the storage unit 5. Here, the neural network includes a first convolutional neural network (CNN) that inputs the two-dimensional image data, and a first variational self-encoding system that is connected to the first CNN and performs encoding. A Variational Auto Encoder (VAE), a second VAE connected to the first VAE to perform decoding, and a second CNN connected to the second VAE to output latent variable data.

Then, the CPU 10 creates two-dimensional image data corresponding to the waveform data file indicating the current waveform to be analyzed, and inputs the created two-dimensional image data to the trained neural network defined by the weight file. A visualization data file is created to display the calculated latent variable data on a monitor or the like.

The visualization data file created in this manner is transmitted to a monitor, etc. owned by the user via the network NW and web user interface IF, and the latent variable data is visualized.

FIG. 3 is a flowchart showing a current waveform analysis method according to an embodiment of the present invention. As shown in FIG. 3, in the current waveform analysis method according to the embodiment of the present invention, in step S1, latent variable data calculated when the measured current waveform is in a normal state, and the current waveform Two-dimensional image data is created from the measured current waveform using a procedure that maximizes the difference from latent variable data calculated when the current waveform is in an abnormal state.

Next, in step S2, the two-dimensional image data created in step S1 according to the current waveform to be analyzed is input to the trained neural network that has been trained with the two-dimensional image data, so that the latent variable data is Calculate.

Next, in step S3, the latent variable data calculated in step S2 is displayed.

FIG. 4 is a diagram showing an example of the current waveform analysis method shown in FIG. 3. In the following, the current waveform analysis method shown in FIG. 4 will be explained using the current waveform analysis device 1 shown in FIGS. 1 and 2 as an example. Needless to say, this may be carried out by the following means.

The current waveform analysis method shown in FIG. 4 consists of a preprocessing stage consisting of steps S11 to S15, a main processing stage consisting of step S16, and a visualization stage consisting of steps S17 and S18. This is realized by the CPU 10 defining the procedure shown in FIG. 4 and executing the program stored in the storage unit 5. The preprocessing step corresponds to step S1 shown in FIG. 3, the main processing step corresponds to step S2, and the visualization step corresponds to step S3, respectively.

In the preprocessing stage, the target data, which is a set of one-dimensional arrays corresponding to the above-mentioned current waveforms, is converted into a two-dimensional image for each time-series system, and the reference data corresponding to the data obtained in the above-mentioned normal state and the data obtained in the above-mentioned abnormal state are converted into two-dimensional images. Divide and store comparison target data corresponding to the obtained data. Specifically, two-dimensional image data as shown in FIGS. 5(b) and 6(b) is generated from the current waveforms shown in FIGS. 5(a) and 6(a). , is stored in the storage unit 5. Note that in the current waveforms shown in FIGS. 5(a) and 6(a), the horizontal axis represents time and the vertical axis represents the magnitude of the current, respectively.

In step S13, shaping processing is performed on a plurality of pixels generated by converting each current value constituting the current waveform into a pixel having a gradation corresponding to its magnitude. Here, the gradation values in these plurality of pixels can be, for example, a value obtained by dividing each current value by the maximum value among the measured current values, multiplied by a constant.

Depending on the format of the target data, as shown in FIG. 4, various optional processes indicated by diagonal lines in step S11 and step S12 may occur before step S13, and various optional processes indicated by diagonal lines in step S14 and step S15 may occur after step S13. It is also possible to execute various optional processes shown in . The present shaping process will be explained below.

FIG. 7 is a first diagram for explaining the shaping process shown in step S13, and FIG. 7(a) is an example of trimming a pixel having a gradation corresponding to the magnitude of the current value. FIG. 7B is a diagram showing an example of padding the pixels. As shown in FIG. 7, in order to make the two-dimensional image data into a square image by setting the number of pixels to the square of a natural number, pixels are deleted (trimmed) as shown before and after the arrow in FIG. 7(a). , or adding (padding) pixels having a constant gradation such as a 0 value as shown by the arrow in FIG. 7(b).

Note that by arranging the same number of pixels vertically and horizontally, two-dimensional image data can be made into a square image as described above, making it easier to expand the data when the number of data used as a dataset is small. This is suitable because data processing is easy even when the number of layers is increased.

FIG. 8 is a second diagram for explaining the shaping process shown in step S13, and FIG. 8(a) is an example of arranging the pixels in a line in the direction indicated by the arrow. 8(b) is a first example in which the pixels are arranged on a two-dimensional plane in the order shown by the arrows, and FIG. 8(c) is a first example in which the pixels are arranged in the order shown in the arrows on a two-dimensional plane. In the second example, FIG. 8(d) shows a third example in which the pixels are arranged on a two-dimensional plane in the order indicated by arrows.

In this shaping process, as shown in FIG. 8(a), the pixels created as described above are arranged in a line as shown by the arrows. 8(d).

Here, since the three two-dimensional image data generated by arranging pixels in the steps shown in FIGS. 8(b) to 8(d) do not overlap each other due to rotation or inversion, they are When input to the network, the latent variable data output from the VAE will be different. Therefore, among these three two-dimensional image data, there are latent variable data calculated when the current waveform is in a normal state, and latent variable data calculated when the current waveform is in an abnormal state. Select the procedure with the largest difference. Here, the magnitude of the difference can be determined by, for example, the degree of separation between the cluster of latent variables corresponding to the normal state and the cluster of latent variables corresponding to the abnormal state displayed on the display unit 4. .

It should be noted that procedures other than those shown in FIGS. 8(b) to 8(d) may be adopted as long as the generated two-dimensional image data do not overlap with each other.

FIG. 9 is a third diagram for explaining the shaping process shown in step S13. FIG. 9(a) is an example of arranging the pixels in the procedure indicated by the arrow, ) is the procedure shown in FIG. 9(a) rotated 90 degrees clockwise, FIG. 9(c) is the procedure shown in FIG. 9(a) rotated 180 degrees clockwise, 9(d) shows the procedure shown in FIG. 9(a) rotated 270 degrees clockwise, and FIG. 9(e) shows the procedure shown in FIG. 9(a) reversed horizontally. show.

For example, if the two-dimensional image data shown in FIG. 9(a) is used as a reference, the two-dimensional image data shown in FIG. 9(b) to FIG. 9(e) overlap each other due to rotation or inversion. , the latent variable data output from the above VAE will be the same. Therefore, the procedures shown in FIGS. 9(b) to 9(e) cannot be adopted as different processing from the procedure shown in FIG. 9(a).

Note that in the above-mentioned shaping process, the case where the pixels are arranged linearly has been described, but the case where the pixels are arranged spirally can be similarly considered.

Next, the main processing stage in step S16 will be explained. In this main processing, the two-dimensional image data created in the preprocessing stage is input to the first CNN and encoded by the first VAE, and the data output from the first VAE is encoded by the second VAE. The latent variable data is output from the second CNN. In addition, in the learning phase of this neural network, various parameters in this neural network are optimized by this main processing.

Next, the visualization stage in steps S17 and S18 will be explained. In the visualization stage, if the latent variable data obtained in step S16 is high-dimensional data, visualization option processing is executed to dimensionally compress the latent variable data into two-dimensional (or three-dimensional) data.

For example, in step S18, a dimension reduction algorithm using t-distributed Stochastic Neighbor Embedding (t-SNE) is executed, but other optional processing is performed as indicated by diagonal lines. may be executed.

FIG. 10 is a first diagram showing the results of the visualization processing in steps S17 and S18, in which FIG. 10(a) is an example of the current waveform forming the reference group and the analysis target group, and FIG. 10(b) is the reference An example of a latent variable scatter plot generated from a group and an analysis target group is shown. Here, the two latent variable scatter diagrams shown in FIG. 10(b) compare the characteristics between the current waveforms forming the reference group and the current waveforms forming the analysis target group. These differences are visualized based on their distribution status.

According to the current waveform analysis device 1 and the current waveform analysis method according to the embodiment of the present invention as described above, by converting one-dimensional time series data into a two-dimensional image and then performing deep learning, the above two The original functions of CNN can be utilized for dimensional image data. As a result, by increasing the number of CNN layers before and after VAE, as illustrated in FIG. 11, it becomes possible to separate states (in this case, individual differences) that could not be separated in the past.

FIG. 11 is an example of a latent variable scatter diagram showing the results of the visualization processing in steps S17 and S18 in FIG. It was generated from. As shown in Figure 11(a), even if the latent variables obtained from the current waveforms measured on the two electronic boards are mixed and displayed, the number of CNN layers inserted before and after the VAE is doubled. By increasing the sensitivity of feature extraction, it is possible to display both latent variables separately at the point indicated by the arrow. Note that the difference in the shape of the latent variable distribution itself in FIG. 11 is due to the nature of the VAE used and is unrelated to abnormality determination.

Further, according to the current waveform analysis device 1 and the current waveform analysis method according to the embodiment of the present invention, in addition to the data group that follows the conventional Gaussian distribution, anomaly analysis can also be performed on the data group that converges on the Dirac measure space. Can be done.

Generally, an effective method for distinguishing between normal data and abnormal data is to identify abnormal data based on a reinforcement learning type knowledge base of normal data. This is based on the idea of the central limit theorem in probability theory, and assumes that randomly sampled and averaged data follows a Gaussian distribution. When time series data converges to a probability space formed by the Dirac measure (hereinafter referred to as the Dirac measure space), the mean and variance of randomly sampled and averaged data converge to a single point regardless of whether it is normal or abnormal. As the number of sample data increases, the characteristics attenuate and eventually disappear. In such cases, conventional data analysis methods cannot capture the characteristics with an insufficient amount of data, and the characteristics are lost with a sufficient amount of data.

Here, the three current waveforms shown in FIG. 12 are examples of current waveforms measured on a certain electronic board, and the current value of each waveform has an average value avr and a variance var shown on the right. As shown in FIG. 12, the dispersion of the current values shown by these waveforms is extremely small.

In the current waveform shown in FIG. 12, characteristics due to differences in the operating environment appear immediately after power is turned on, but with the conventional method, the false detection rate of abnormality determination is high. Further, when the state of the electronic board becomes stable, the characteristics of the current data are leveled out and the characteristics cannot be captured. According to the current waveform analysis device 1 and the current waveform analysis method according to the embodiment of the present invention, even in such a case, abnormality can be detected with high accuracy for time series data that converges to the Dirac measure space. It is possible to make a judgment.

Here, FIG. 13(a) shows a current waveform measured when the electronic board is in a normal state, and FIG. 13(b) shows a current waveform measured when a temperature load is applied to the electronic board to bring it into an abnormal state. The current waveform is shown below. Moreover, FIG. 14(a) shows a latent variable scatter diagram generated corresponding to the current waveform shown in FIG. 13(a), and FIG. 14(b) shows the current waveform shown in FIG. 13(b). Shows the latent variable scatter plot generated in response to .

As shown in Figure 13(b), there is a gap in the current waveform obtained when a temperature load is applied to the electronic board to bring it into an abnormal state, but the variance in the overall current value is small. It has become. Furthermore, the latent variables shown in Figure 14(a) are distributed in a circular shape in the center of the figure, but the latent variables shown in Figure 14(b) are distributed in the same way as in Figure 14(a). It can be seen that the distribution is not only circular in the center, but also spread to the outer edge.

Therefore, according to the current waveform analysis device 1 and the current waveform analysis method according to the embodiment of the present invention, it is possible to produce a clear difference in the latent variable scatter diagram even between such a normal state and an abnormal state. Therefore, highly accurate abnormality determination can be realized.

As described above, in the abnormality determination according to the embodiment of the present invention, regardless of whether the time series data converges or does not converge, the reference data corresponding to the data in the normal state is extracted based on the data obtained by extracting features using CNN. Then, phase conversion is performed by VAE. Then, dimension compression processing is performed on the input data mapped to the latent variable space, and normal and abnormal conditions are identified using a unique distance concept of similarity distance with the reference data. As a result, a current waveform analysis device and a current waveform analysis method are realized that cover the conventional techniques and are capable of determining even objects to which the conventional techniques are difficult to apply.

1 current waveform analyzer, 2 image data creation section, 3 analysis section, 4 display section.

Claims (10)

A current waveform analyzer that analyzes a measured current waveform,
image data creation means for creating two-dimensional image data from the measured current waveform;
An analysis of calculating latent variable data by inputting the two-dimensional image data created by the image data creation means according to the current waveform to be analyzed into a trained neural network that has trained the two-dimensional image data. means and
comprising display means for displaying the latent variable data calculated by the analysis means,
The image data creation means is arranged to determine the maximum difference between the latent variable data calculated when the current waveform is in a normal state and the latent variable data calculated when the current waveform is in an abnormal state. A current waveform analysis device that creates the two-dimensional image data according to the following procedure. The image data creation means arranges a plurality of pixels having a gradation corresponding to the magnitude of each current value constituting the current waveform on a two-dimensional plane in a predetermined procedure, so that they overlap each other by rotation or inversion. The current waveform analysis device according to claim 1, wherein the current waveform analysis device creates the two-dimensional image data. 3. The current waveform analysis device according to claim 2, wherein the image data creation means arranges the plurality of pixels in the same number vertically and horizontally. The neural network is
a first convolutional neural network inputting the two-dimensional image data;
a first variational autoencoder connected to the first convolutional neural network to perform encoding;
a second variational autoencoder connected to the first variational autoencoder to perform decoding;
The current waveform analysis device according to claim 1, further comprising a second convolutional neural network connected to the second variational autoencoder and outputting latent variable data. The current waveform analysis device according to claim 1, wherein the display means converts the latent variable data into two-dimensional data and displays the data. A current waveform analysis method for analyzing a measured current waveform, the method comprising:
A first step of creating two-dimensional image data from the measured current waveform;
A second step of calculating latent variable data by inputting the two-dimensional image data created in the first step according to the current waveform to be analyzed into the trained neural network that has trained the two-dimensional image data. The second step and
a third step of displaying the latent variable data calculated in the second step,
In the first step, the difference between the latent variable data calculated when the current waveform is in a normal state and the latent variable data calculated when the current waveform is in an abnormal state is maximum. A current waveform analysis method in which the two-dimensional image data is created by the following steps. In the first step, a plurality of pixels having gradations corresponding to the magnitude of each current value constituting the current waveform are arranged on a two-dimensional plane in a predetermined procedure, and are overlapped with each other by rotation or inversion. 7. The current waveform analysis method according to claim 6, wherein said two-dimensional image data that does not have to be generated is created. 8. The current waveform analysis method according to claim 7, wherein in the first step, the plurality of pixels are arranged in the same number vertically and horizontally. The neural network is
a first convolutional neural network inputting the two-dimensional image data;
a first variational autoencoder connected to the first convolutional neural network to perform encoding;
a second variational autoencoder connected to the first variational autoencoder to perform decoding;
The current waveform analysis method according to claim 6, comprising a second convolutional neural network connected to the second variational autoencoder and outputting latent variable data. 7. The current waveform analysis method according to claim 6, wherein in the third step, the latent variable data is converted into two-dimensional data and then displayed.

Images