Feature amount extraction device, recurrence plot generation device, analysis system, feature amount extraction method, recurrence plot generation method, analysis method, feature amount extraction program, recurrence plot generation program, and analysis program

Abstract

The present invention improves the versatility of a feature amount in time series data. A feature amount extraction device (20) uses an inference model to extract a feature amount from a non-binary recurrence plot. In the non-binary recurrence plot, for each of all combinations of two samples selected from a plurality of samples included in the time-series data, a distance scale between the two samples is set to two-dimensional image information at positions determined by the order of one of the two samples and the order of the other in the time-series data. The inference model is constructed by machine learning which uses, as learning data, a non-binary recurrence plot generated from time-series data for learning.

Description

Feature extraction device, recurrence plot generation device, analysis system, feature extraction method, recurrence plot generation method, analysis method, feature extraction program, recurrence plot generation program, and analysis program

[0001] The present disclosure relates to a feature extraction device, a recurrence plot generation device, an analysis system, a feature extraction method, a recurrence plot generation method, an analysis method, a feature extraction program, a recurrence plot generation program, and an analysis program.

[0002] In detecting anomalies in time series data, it may be necessary to track features that represent recurrence in the time series data. In this regard, in order to extract such features from time series data, a recurrence plot that represents the recurrence of one-dimensional time series data as a two-dimensional image may be used. For example, International Publication No. 2021 / 044626 (Patent Document 1) discloses a technique for converting one-dimensional time series data into a recurrence plot and extracting features from the recurrence plot using machine learning.

[0003] International Publication No. 2021 / 044626

[0004] Recurrence plots are generated as binary image information by converting two-dimensional image information into binary information based on a threshold. Therefore, the selection of a threshold can significantly affect conventional analysis processes for extracting features from time-series data using recurrence plots. However, selecting an appropriate threshold is often difficult, and the threshold setting often depends on the user's subjective judgment. As a result, the versatility of the configuration for realizing conventional analysis processes can be reduced.

[0005] The present disclosure has been made to solve the above-mentioned problems, and its purpose is to improve the versatility of feature quantities of time-series data.

[0006] A feature extraction device according to one aspect of the present disclosure includes an acquisition unit and an extraction unit. The acquisition unit acquires a non-binary recurrence plot generated from time-series data. The extraction unit extracts features from the non-binary recurrence plot using an inference model. In the non-binary recurrence plot, for each of all combinations of two samples selected from a plurality of samples included in the time-series data, a distance measure between the two samples is set in two-dimensional image information at a position identified by the order of one of the two samples and the order of the other of the two samples in the time-series data. The inference model is constructed by machine learning using the non-binary recurrence plot generated from the time-series data for training as training data.

[0007] A recurrence plot generation device according to another aspect of the present disclosure includes an acquisition unit and a generation unit. The acquisition unit acquires time-series data. The generation unit generates a non-binary recurrence plot from the time-series data. The generation unit generates the non-binary recurrence plot by setting, for each of all combinations of two samples selected from a plurality of samples included in the time-series data, a distance measure between the two samples in two-dimensional image information at a position identified by the order of one of the two samples and the order of the other of the two samples in the time-series data.

[0008] A feature extraction method according to another aspect of the present disclosure includes the steps of obtaining a non-binary recurrence plot generated from time-series data and extracting features from the non-binary recurrence plot using an inference model. In the non-binary recurrence plot, for each combination of two samples selected from a plurality of samples included in the time-series data, a distance measure between the two samples is set in two-dimensional image information at a position identified by the order of one of the two samples in the time-series data and the order of the other of the two samples in the time-series data. The inference model is constructed by machine learning using the non-binary recurrence plot generated from the training time-series data as training data.

[0009] A recurrence plot generating method according to another aspect of the present disclosure includes the steps of acquiring time-series data and generating a non-binary recurrence plot from the time-series data, wherein the generating step generates the non-binary recurrence plot by setting, for each of all combinations of two samples selected from a plurality of samples included in the time-series data, a distance measure between the two samples to two-dimensional image information at positions identified by the order of one and the order of the other of the two samples in the time-series data.

[0010] According to the feature extraction device, recurrence plot generation device, analysis system, feature extraction method, recurrence plot generation method, analysis method, feature extraction program, recurrence plot generation program, and analysis program disclosed herein, the versatility of features of time series data can be improved by extracting features of time series data from a non-binary recurrence plot generated from the time series data.

[0011] FIG. 1 is a diagram showing an example of the configuration of an analysis system according to a first embodiment. FIG. 2 is a diagram showing a non-binary recurrence plot according to the first embodiment and a conventional recurrence plot together. FIG. 3 is a diagram showing an example of the hardware configuration of the recurrence plot generation device of FIG. 1. FIG. 4 is a flowchart showing an example of the flow of processing performed by a processing circuit 61 that executes the non-binary recurrence plot generation program of FIG. 3. FIG. 5 is a diagram showing an example of the hardware configuration of the feature extraction device of FIG. 1. FIG. 6 is a flowchart showing an example of the flow of processing performed by a processing circuit that executes the feature extraction program of FIG. 5. FIG. 7 is a diagram showing an example of the configuration of an anomaly determination system according to a second embodiment. FIG. 8 is a diagram showing an example of the configuration of a clustering system according to a third embodiment. FIG. 9 is a diagram showing an example of the configuration of a correlation analysis system according to a fourth embodiment.

[0012] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. In the drawings, the same or corresponding parts are designated by the same reference numerals, and their description will not be repeated in principle.

[0013] Embodiment 1. Fig. 1 is a diagram showing an example of the configuration of a system 100 according to embodiment 1. As shown in Fig. 1, the system 100 analyzes input time series data (input time series data) and outputs feature quantities of the input time series data. The input time series data is arbitrary and may be two or more types of time series data. Note that time series data is data in which multiple samples are ordered in the order in which they were acquired. Each of the multiple samples may be a scalar or a vector.

[0014] The system 100 includes a recurrence plot generation device 10 and a feature extraction device 20. The recurrence plot generation device 10 includes an acquisition unit 11 and a generation unit 12. The acquisition unit 11 acquires input time series data and outputs it to the generation unit 12. The generation unit 12 creates a non-binary recurrence plot from the input time series data. The generation unit 12 may also create a non-binary recurrence plot from time series data from which unnecessary data has been removed by filtering. Filtering the time series data in this manner makes it possible to flexibly generate a non-binary recurrence plot and select a distance measure appropriate for the task.

[0015] The generation unit 12 generates a non-binary recurrence plot by setting, for each of all combinations (sample pairs) of two samples x(i) and x(j) (i and j are integers) selected from the multiple samples included in the input time series data, a distance measure d(x(i), x(j)) between the two samples x(i) and x(j) to two-dimensional image information R(i, j) (pixel values) at a position specified by the order i of one of the two samples and the order j of the other sample in the input time series data (see Equation (1) below). That is, the generation unit 12 creates the non-binary recurrence plot by plotting the distance measure d for each pair as a two-dimensional histogram without binarization processing based on a threshold judgment. The distance measure d may be any index indicating the proximity or similarity between two samples, and may include, for example, Euclidean distance, Euclidean norm, Manhattan distance, or Mahalanobis distance.

[0016]

[0017] A conventional recurrence plot is generated as binary image information by converting two-dimensional image information R(i, j) into binary information based on a threshold value r, as shown in the following equation (2).

[0018]

[0019] Therefore, the selection of the threshold r can have a significant impact on conventional analytical processes for extracting features from time-series data using recurrence plots. However, selecting an appropriate threshold r is often difficult, and the setting of the threshold r often depends on the user's subjective judgment. This can reduce the versatility of the system implementing the conventional process or the reproducibility of the processing results. Furthermore, conventional processes are computationally expensive and unsuitable for large-scale time-series data or real-time processing. Furthermore, the conventional process's reliance on the threshold r can prevent it from fully capturing the underlying dynamics of the time-series data, limiting the versatility and accuracy of processing according to the time-series data.

[0020] The recurrence plot generation device 10 achieves improved processing speed and prevents task fixation compared to when binarization processing based on threshold judgment is performed, thereby improving the versatility of processing for extracting features from time-series data. Furthermore, the recurrence plot generation device 10 does not discard information due to binarization processing, making it possible to represent subtle fluctuations in time-series data as two-dimensional histograms. The process of generating a non-binary recurrence plot by the recurrence plot generation device 10 corresponds to the process of generating a conventional recurrence plot, except that it does not perform binarization processing based on threshold judgment.

[0021] The feature extraction device 20 includes an acquisition unit 21, an extraction unit 22, and a learning unit 23. The learning unit 23 performs machine learning using the non-binary recurrence plot as training data to capture patterns in time-series data corresponding to the non-binary recurrence plot and construct an inference model (trained model) that extracts features from the non-binary recurrence plot. The machine learning performed by the learning unit 23 is arbitrary and includes, for example, a neural network, a decision tree, or a support vector machine.

[0022] The acquisition unit 21 acquires the non-binary recurrence plot and outputs it to the extraction unit 22. The extraction unit 22 extracts features from the non-binary recurrence plot using the inference model constructed by the learning unit 23 and outputs the features to an external device (for example, a user's terminal device or display device). By utilizing the trained inference model, the analytical performance of the system 100 can be improved. Note that the inference model may be a model trained by an external learning device instead of the learning unit 23.

[0023] In addition, when multiple non-binary recurrence plots are created for the same time-series data under multiple conditions with different contents, the feature extraction device 20 may extract features indicated by the time-series data through machine learning or image processing using these multiple non-binary recurrence plots. Using machine learning from different perspectives enables multifaceted analysis of the time-series data. Furthermore, the feature extraction device 20 may extract features after filtering the non-binary recurrence plot with an image filter. Applying an image filter can eliminate unnecessary information contained in the non-binary recurrence plot or emphasize important information.

[0024] The multiple conditions for the same time series data, which have different contents, include, for example, an embedding dimension or delay amount based on the Turnens embedding theorem. Turnens embedding theorem is a method for capturing the nonlinearity of nonlinear time series data by mapping the time series data into a high-dimensional space. In other words, Turnens embedding theorem allows for the acquisition of multiple dimensions that represent the target nonlinear time series data. By converting the changes in the multiple dimensions into multiple nonbinary recurrence plots, two steps are achieved: "capturing nonlinearity using Turnens embedding theorem" and "emphasizing recurrence using nonbinary recurrence plots." As a result, the recurrence of nonlinear time series data can be emphasized without artificial parameter adjustment and independently of the waveform shape of the nonlinear time series data.

[0025] FIG. 2 shows a non-binary recurrence plot Rnb according to the first embodiment and a conventional recurrence plot Rb. As shown in FIG. 2, the non-binary recurrence plot Rnb and the recurrence plot Rb are generated from time-series data Dts. The horizontal axis of the time-series data Dts represents the time at which multiple samples included in the time-series data Dts were acquired. The times correspond to the ordinal values ​​i and j of the two samples included in the sample pair for which the distance measure is calculated. The vertical axis of the time-series data Dts represents the value x of the sample. The vertical and horizontal axes of the non-binary recurrence plot Rnb and the recurrence plot Rb correspond to the ordinal values ​​i and j of the two samples included in the sample pair, respectively.

[0026] The two-dimensional image information of the recurrence plot Rb is composed of two values, one corresponding to white and one corresponding to black. On the other hand, the two-dimensional image information of the non-binary recurrence plot Rnb includes not only the binary values ​​corresponding to white and black, but also values ​​corresponding to multiple shades of gray between white and black. In this way, it is visually apparent that the amount of information contained in the non-binary recurrence plot Rnb is greater than the amount of information contained in the recurrence plot Rb.

[0027] Fig. 3 is a diagram showing an example of the hardware configuration of the recurrence plot generation device 10 of Fig. 1. As shown in Fig. 3, the recurrence plot generation device 10 includes a processing circuitry 61, a memory 62, an input unit 63, an output unit 64, and a bus 65. The processing circuitry 61, the memory 62, the input unit 63, and the output unit 64 are communicatively connected to one another via the bus 65. The recurrence plot generation device 10 includes, for example, a PC (Personal Computer) or a workstation.

[0028] The processing circuitry 61 includes at least one central processing unit (CPU) that executes programs stored in the memory 62. The processing circuitry 61 may also include at least one graphics processing unit (GPU). The functions of the recurrence plot generation device 10 are realized by software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in the memory 62. The processing circuitry 61 reads and executes the program stored in the memory 62. The CPU is also called a central processing unit, processing device, arithmetic unit, microprocessor, microcomputer, processor, or DSP (digital signal processor).

[0029] The memory 62 may include non-volatile or volatile semiconductor memory (e.g., RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), or EEPROM (Electrically Erasable Programmable Read Only Memory)), as well as a magnetic disk, a flexible disk, an optical disk, a compact disk, a minidisk, or a DVD (Digital Versatile Disc). The memory 62 stores a non-binary recurrence plot generation program Pg11. A processing circuit 61 that executes the non-binary recurrence plot generation program Pg11 corresponds to the acquisition unit 11 and the generation unit 12 in FIG. 1. Although not shown in FIG. 1, the memory 62 may also store an operating system (OS).

[0030] The input unit 63 receives operations from a user and includes a mouse, a keyboard, a microphone, and a touch panel. The output unit 64 outputs the processing results of the processing circuit 61 to the user. The output unit 64 includes, for example, a touch panel, a display, and a speaker.

[0031] 4 is a flowchart showing an example of the flow of processing performed by the processing circuitry 61 executing the non-binary recurrence plot generation program Pg11 of FIG. 3. Hereinafter, steps will be simply referred to as S. As shown in FIG. 4, the processing circuitry 61 acquires time-series data in S101 and proceeds to S102. In S102, the processing circuitry 61 generates a non-binary recurrence plot and proceeds to S103. In S103, the processing circuitry 61 outputs the non-binary recurrence plot and ends the processing.

[0032] FIG. 5 is a diagram illustrating an example of the hardware configuration of the feature extraction device 20 of FIG. 1 . As illustrated in FIG. 5 , the feature extraction device 20 includes a processing circuit 71, a memory 72, an input unit 73, an output unit 74, and a bus 75. The processing circuit 71, the memory 72, the input unit 73, and the output unit 74 are communicatively connected to one another via the bus 75. The feature extraction device 20 includes, for example, a personal computer (PC) or a workstation. The processing circuit 71, the memory 72, the input unit 73, the output unit 74, and the bus 75 have the same configurations as the processing circuit 61, the memory 62, the input unit 63, the output unit 64, and the bus 65 of FIG. 3 , respectively, and therefore description of the same configurations will not be repeated.

[0033] The memory 72 stores an inference model Me, a feature extraction program Pg21, a machine learning program Pg22, and multiple pieces of training data Ld. The processing circuit 71 that executes the feature extraction program Pg21 corresponds to the acquisition unit 21 and the extraction unit 22 in Figure 1. The processing circuit 71 that executes the feature extraction program Pg21 corresponds to the acquisition unit 21 and the extraction unit 22 in Figure 1. The processing circuit 71 that executes the machine learning program Pg22 corresponds to the training unit 23 in Figure 1. The training data Ld includes time-series data for training or non-binary recurrence plots generated from the time-series data for training.

[0034] 6 is a flowchart showing an example of the flow of processing performed by the processing circuitry 71 executing the feature extraction program Pg21 of FIG. 5. As shown in FIG. 6, in S201, the processing circuitry 71 acquires a non-binary recurrence plot generated from time-series data and proceeds to S202. In S202, the processing circuitry 71 extracts features from the non-binary recurrence plot using an inference model and proceeds to S103. In S203, the processing circuitry 71 outputs the features and ends the processing.

[0035] Modification of First Embodiment The acquisition unit 11 of the recurrence plot generation device 10 of FIG. 1 may receive input time series data and divide the time series data into multiple pieces of time series data based on a predetermined trigger. In this case, the acquisition unit 11 resamples each of the multiple pieces of time series data after division so that the multiple pieces of time series data after division have the same number of samples. The acquisition unit 11 outputs the multiple pieces of time series data after division to the generation unit 12. The generation unit 12 generates a non-binary recurrence plot from each of the multiple pieces of time series data.

[0036] The trigger does not need to be included in the input time-series data, and may be a signal separate from the time-series data to be divided. For example, the shaft voltage waveform may be divided using the applied voltage waveform of the motor as a trigger. Note that multiple waveforms may be used as triggers, and for example, the shaft voltage waveform may be divided using the feature quantities of the applied voltage waveform and the shaft current waveform of the motor as triggers. Furthermore, the shaft voltage waveform may be divided based on a predetermined feature quantity of the shaft voltage waveform.

[0037] The number of samples (resampling number) of each time series data after resampling by the acquiring unit 11 is arbitrary, but since the recurrence plot is a process based on determinism, it is adjusted so as to maintain the global structure of the time series data. The resampling by the acquiring unit 11 is performed using a method such as data thinning or averaging.

[0038] When the generation unit 12 creates a non-binary recurrence plot, using all of the time series data to be analyzed may slow down the processing speed for creating the recurrence plot. This is because, when converting time series data into a non-binary recurrence plot, the number of plots in the non-binary recurrence plot is equal to the square of the number of samples in the original time series data. For example, if the time series data consists of 100 samples, it is converted into a non-binary recurrence plot as a two-dimensional histogram with 10,000 (100 squared) plots. In other words, as the number of samples in the time series data increases, the number of processes required to generate the non-binary recurrence plot increases exponentially. Therefore, the acquisition unit 11 can perform resampling in advance to reduce the number of samples included in the divided time series data, thereby improving the processing speed of the generation unit 12. Note that the acquisition unit 11 may also increase the number of samples in the divided time series data to increase the amount of information in the divided time series data.

[0039] As described above, the feature extraction device, recurrence plot generation device, feature extraction method, recurrence plot generation method, feature extraction program, and recurrence plot generation program according to the first embodiment and its modifications can improve the versatility of features of time-series data.

[0040] Embodiment 2. Fig. 7 is a diagram showing an example of the configuration of an abnormality determination system 200, which is an example of an analysis system according to embodiment 2. As shown in Fig. 7, the abnormality determination system 200 outputs a determination result as to whether or not input time-series data is abnormal. The abnormality determination system 200 includes the system 100 and an abnormality determination device 30. The configuration of the system 100 is similar to the configuration shown in Fig. 1, and therefore description of the similar configuration will not be repeated.

[0041] The anomaly determination device 30 sets a threshold value based on the feature value output from the system 100. The threshold value is used as an anomaly determination criterion and may be arbitrary or quantitatively calculated. The number of threshold values ​​does not have to be one. The threshold value is selected according to the requirements of the anomaly determination system 200 or the characteristics of the time-series data. If the feature value output from the system 100 exceeds the threshold value, the anomaly determination device 30 determines that an anomaly exists in the input time-series data. This determination result is used to detect or predict an anomaly.

[0042] According to the anomaly determination system 200, since no binarization process is performed when creating a recurrence plot, it is possible to perform anomaly determination for complex data such as nonlinear time series data, based on the recurrence of the time series data in the same way as when using a conventional recurrence plot, but with greater speed and detail than when using a conventional recurrence plot.

[0043] As described above, the feature extraction device, recurrence plot generation device, analysis system, feature extraction method, recurrence plot generation method, analysis method, feature extraction program, recurrence plot generation program, and analysis program according to embodiment 2 can improve the versatility of features of time-series data.

[0044] Embodiment 3. Fig. 8 is a diagram showing an example of the configuration of a clustering system 300, which is an example of an analysis system according to embodiment 3. As shown in Fig. 8, the clustering system 300 classifies input time series data into corresponding groups (clusters) based on features extracted from the input time series data, and outputs a classification result (clustering result). The clustering system 300 includes the system 100 and a clustering device 40. The configuration of the system 100 is similar to the configuration shown in Fig. 1, and therefore description of the similar configuration will not be repeated.

[0045] The clustering device 40 performs clustering on the input time-series data into two or more categories based on the feature quantities output from the system 100. Clustering is a method for grouping data based on similarity. In the clustering method, an arbitrary threshold may be set, or clustering may be performed using image processing such as machine learning.

[0046] According to the clustering system 300, since no binarization process is performed when creating a recurrence plot, it is possible to perform faster and more detailed clustering than when using conventional recurrence plots for complex data such as nonlinear time series data, while still based on the recurrence of the time series data in the same way as when using conventional recurrence plots.

[0047] As described above, the feature extraction device, recurrence plot generation device, analysis system, feature extraction method, recurrence plot generation method, analysis method, feature extraction program, recurrence plot generation program, and analysis program according to embodiment 3 can improve the versatility of features of time-series data.

[0048] Fourth Embodiment. Fig. 9 is a diagram showing an example of the configuration of a correlation analysis system 400, which is an example of an analysis system according to a fourth embodiment. As shown in Fig. 9, the correlation analysis system 400 performs correlation analysis on input time series data and outputs the results of the correlation analysis (analysis results). The correlation analysis can calculate parameters used for anomaly detection, clustering, factor analysis, etc. The correlation analysis system 400 includes a system 100 and a correlation analysis device 50. The configuration of the system 100 is similar to the configuration shown in Fig. 1, and therefore, description of the similar configuration will not be repeated.

[0049] The correlation analysis device 50 analyzes the correlation between input time series data based on the feature quantities extracted from the system 100. Note that the correlation analysis device 50 does not need to use all of the feature quantities output from the system 100, and may perform correlation analysis using only some of the feature quantities. By limiting the feature quantities used in the correlation analysis, efficient correlation analysis becomes possible. Limiting the feature quantities used in the correlation analysis in this way is suitable, for example, when the correlation analysis is performed as real-time processing.

[0050] According to the correlation analysis system 400, since no binarization process is performed when creating a recurrence plot, it is possible to perform a correlation analysis on complex data such as nonlinear time series data, based on the recurrence of the time series data in the same way as when using a conventional recurrence plot, but at a higher speed and in more detail than when using a conventional recurrence plot.

[0051] The correlation analysis system 400 does not necessarily include the feature extraction device 20. That is, the correlation analysis device 50 may perform image correlation analysis on the non-binary recurrence plot created by the recurrence plot generation device 10. The image correlation analysis refers to anomaly detection, clustering, or factor analysis.

[0052] As described above, the feature extraction device, recurrence plot generation device, analysis system, feature extraction method, recurrence plot generation method, analysis method, feature extraction program, recurrence plot generation program, and analysis program according to embodiment 4 can improve the versatility of features of time-series data.

[0053] The recurrence plot generation device 10 and feature extraction device 20 of the first to fourth embodiments, the anomaly determination device 30 of the second embodiment, the clustering device 40 of the third embodiment, and the correlation analysis device 50 of the fourth embodiment may each be realized by dedicated hardware. Part or all of each device may be realized by a general-purpose or dedicated circuit, a processor, optical computing using a diffractive optical element, or a combination of these. These may be configured by a single chip or by multiple chips connected via a bus. Part or all of the components of each device may be realized by a combination of the above-mentioned circuits, etc. and a program.

[0054] The use of dedicated hardware enables high-speed processing, efficient resource utilization, and flexible scalability. In particular, optical computing enables high-speed analysis of non-binary recurrence plots while saving memory and power consumption.

[0055] When the recurrence plot generation device 10, the feature extraction device 20, the anomaly determination device 30, the clustering device 40, and the correlation analysis device 50 are each realized by multiple information processing devices or multiple circuits, the multiple information processing devices or multiple circuits may be centrally located in one location or distributed across various locations. For example, the multiple information processing devices or multiple circuits may be realized in a form connected to each other via a communication network, such as a client-server system or a cloud computing system. In this way, by centrally or distributedly arranging the multiple information processing devices or multiple circuits, it becomes possible to flexibly expand the processing power of the system formed by the multiple information processing devices or multiple circuits.

[0056] Furthermore, if the functions of the recurrence plot generation device 10, the feature extraction device 20, the anomaly determination device 30, the clustering device 40, and the correlation analysis device 50 are realized by software, the software may be installed in a device other than the recurrence plot generation device 10, the feature extraction device 20, the anomaly determination device 30, the clustering device 40, and the correlation analysis device 50 via a memory device such as a USB (Universal Serial Bus) memory. By storing the software that realizes the functions of each device in a memory device in this way, the portability and security of the functions can be improved. Furthermore, it becomes possible to selectively use the required functions in each of the recurrence plot generation device 10, the feature extraction device 20, the anomaly determination device 30, the clustering device 40, and the correlation analysis device 50.

[0057] Aspects of the present disclosure are summarized below as appendices. [Appendix 1] A feature extraction device includes: an acquisition unit that acquires a non-binary recurrence plot generated from time-series data; and an extraction unit that extracts features from the non-binary recurrence plot using an inference model, wherein, for each of all combinations of two samples selected from a plurality of samples included in the time-series data, a distance measure between the two samples is set in two-dimensional image information at a position specified by the order of one of the two samples in the time-series data and the order of the other of the two samples in the time-series data, and the inference model is constructed by machine learning using the non-binary recurrence plot generated from time-series data for training as training data.

[0058] [Supplementary Note 2] The feature extraction device according to Supplementary Note 1, wherein in the machine learning, a plurality of non-binary recurrence plots generated from the training time-series data based on a plurality of conditions in which at least one of an embedding dimension and a delay amount is changed for the training time-series data are used as training data.

[0059] [Supplementary Note 3] A recurrence plot generation device comprising: an acquisition unit that acquires time series data; and a generation unit that generates a non-binary recurrence plot from the time series data, wherein the generation unit generates the non-binary recurrence plot by setting, for each of all combinations of two samples selected from a plurality of samples included in the time series data, a distance measure between the two samples to two-dimensional image information at a position specified by the order of one of the two samples and the order of the other of the two samples in the time series data.

[0060] [Supplementary Note 4] The recurrence plot generating device according to Supplementary Note 3, wherein the generating unit generates the non-binary recurrence plot from the time series data from which unnecessary data has been removed by a filter.

[0061] [Supplementary Note 5] The recurrence plot generation device according to Supplementary Note 3 or 4, wherein the generation unit divides the time series data into a plurality of time series data based on a predetermined trigger, and generates the non-binary recurrence plot from each of the plurality of time series data.

[0062] [Supplementary Note 6] An analysis system comprising: the recurrence plot generation device according to any one of Supplementary Notes 3 to 5; the feature extraction device according to Supplementary Note 1 or 2 that extracts the feature from the non-binary recurrence plot; and an analysis device that performs a specific processing based on the feature, wherein the specific processing includes at least one of anomaly detection, clustering, and correlation analysis.

[0063] [Supplementary Note 7] A feature extraction method comprising: steps of acquiring a non-binary recurrence plot generated from time series data; and steps of extracting features from the non-binary recurrence plot using an inference model, wherein, for each of all combinations of two samples selected from a plurality of samples included in the time series data, a distance measure between the two samples is set in two-dimensional image information at a position specified by the order of one of the two samples in the time series data and the order of the other of the two samples in the time series data, and the inference model is constructed by machine learning using, as training data, the non-binary recurrence plot generated from time series data for training.

[0064] [Supplementary Note 8] The feature extraction method according to Supplementary Note 7, wherein in the machine learning, a plurality of non-binary recurrence plots generated from the training time-series data based on a plurality of conditions in which at least one of an embedding dimension and a delay amount is changed for the training time-series data are used as training data.

[0065] [Supplementary Note 9] A recurrence plot generating method comprising: acquiring time series data; and generating a non-binary recurrence plot from the time series data, wherein the generating step generates the non-binary recurrence plot by setting, for each of all combinations of two samples selected from a plurality of samples included in the time series data, a distance measure between the two samples to two-dimensional image information at a position specified by the order of one of the two samples and the order of the other of the two samples in the time series data.

[0066] [Supplementary Note 10] The recurrence plot generating method according to Supplementary Note 9, wherein the generating step generates the non-binary recurrence plot from the time series data from which unnecessary data has been removed by a filter.

[0067] [Supplementary Note 11] The recurrence plot generating method according to Supplementary Note 9 or 10, wherein the generating step divides the time series data into a plurality of time series data based on a specific trigger, and generates the non-binary recurrence plot from each of the plurality of time series data.

[0068] [Supplementary Note 12] An analysis method including: a step included in the recurrence plot generation method according to any one of Supplements 9 to 11; a step included in the feature extraction method according to Supplementary Note 7 or 8, which extracts the feature from the non-binary recurrence plot; and a step of performing a specific process based on the feature, wherein the specific process includes at least one of anomaly detection, clustering, and correlation analysis.

[0069] The embodiments disclosed herein are intended to be combined as appropriate within the scope of compatibility. The embodiments disclosed herein should be considered to be illustrative and not restrictive in all respects. The scope of the present disclosure is defined by the claims, not the above description, and is intended to include all modifications within the meaning and scope of the claims.

[0070] 10 Recurrence plot generation device, 11, 21 Acquisition unit, 12 Generation unit, 20 Feature extraction device, 22 Extraction unit, 23 Learning unit, 30 Abnormality determination device, 40 Clustering device, 50 Correlation analysis device, 61, 71 Processing circuit, 62, 72 Memory, 63, 73 Input unit, 64, 74 Output unit, 65, 75 Bus, 100 System, 200 Abnormality determination system, 300 Clustering system, 400 Correlation analysis system, Dts Time series data, Ld Learning data, Me Inference model, Pg11 Non-binary recurrence plot generation program, Pg21 Feature extraction program, Pg22 Machine learning program, Rb Recurrence plot, Rnb Non-binary recurrence plot.

Claims

1. A feature extraction device comprising: an acquisition unit that acquires a non-binary recurrence plot generated from time series data; and an extraction unit that extracts features from the non-binary recurrence plot using an inference model, wherein in the non-binary recurrence plot, for each of all combinations of two samples selected from a plurality of samples included in the time series data, a distance measure between the two samples is set in two-dimensional image information of a position specified by the order of one of the two samples in the time series data and the order of the other, and the inference model is constructed by machine learning using the non-binary recurrence plot generated from time series data for training as training data.

2. The feature extraction device of claim 1, wherein the machine learning uses as training data a plurality of non-binary recurrence plots generated from the training time series data based on a plurality of conditions in which at least one of the embedding dimension and the delay amount for the training time series data is varied.

3. A recurrence plot generation device comprising: an acquisition unit that acquires time series data; and a generation unit that generates a non-binary recurrence plot from the time series data, wherein the generation unit generates the non-binary recurrence plot by setting, for each of all combinations of two samples selected from a plurality of samples included in the time series data, a distance measure between the two samples to two-dimensional image information at a position specified by the order of one of the two samples and the order of the other of the two samples in the time series data.

4. The recurrence plot generating device according to claim 3, wherein the generating unit generates the non-binary recurrence plot from the time series data from which unnecessary data has been removed by a filter.

5. A recurrence plot generating device as described in claim 3 or 4, wherein the generating unit divides the time series data into multiple time series data based on a specific trigger, and generates the non-binary recurrence plot from each of the multiple time series data.

6. An analysis system comprising: a recurrence plot generation device according to any one of claims 3 to 5; a feature extraction device according to claim 1 or 2 that extracts the feature from the non-binary recurrence plot; and an analysis device that performs specific processing based on the feature, wherein the specific processing includes at least one of anomaly detection, clustering, and correlation analysis.

7. A feature extraction method comprising: steps of obtaining a non-binary recurrence plot generated from time series data; and steps of extracting features from the non-binary recurrence plot using an inference model, wherein in the non-binary recurrence plot, for each of all combinations of two samples selected from a plurality of samples included in the time series data, a distance measure between the two samples is set to two-dimensional image information of a position specified by the order of one of the two samples in the time series data and the order of the other of the two samples, and the inference model is constructed by machine learning using the non-binary recurrence plot generated from time series data for training as training data.

8. The feature extraction method according to claim 7, wherein in the machine learning, a plurality of non-binary recurrence plots generated from the training time series data based on a plurality of conditions in which at least one of the embedding dimension and the delay amount for the training time series data is changed are used as training data.

9. A recurrence plot generating method comprising: a step of acquiring time series data; and a step of generating a non-binary recurrence plot from the time series data, wherein the generating step generates the non-binary recurrence plot by setting, for each of all combinations of two samples selected from a plurality of samples included in the time series data, a distance measure between the two samples to two-dimensional image information at a position specified by the order of one of the two samples and the order of the other of the two samples in the time series data.

10. The recurrence plot generating method according to claim 9, wherein said generating step generates said non-binary recurrence plot from said time series data from which unnecessary data has been removed by a filter.

11. A recurrence plot generation method according to claim 9 or 10, wherein the generating step divides the time series data into a plurality of time series data based on a specific trigger, and generates the non-binary recurrence plot from each of the plurality of time series data.

12. An analysis method comprising: steps included in the recurrence plot generation method of any one of claims 9 to 11; steps included in the feature extraction method of claim 7 or 8, which extracts the feature from the non-binary recurrence plot; and a step of performing a specific processing based on the feature, wherein the specific processing includes at least one of anomaly detection, clustering, and correlation analysis.

13. A feature extraction program that, when executed by a processing circuit, causes the processing circuit to execute the feature extraction method according to claim 7 or 8.

14. A recurrence plot generating program which, when executed by a processing circuit, causes the processing circuit to perform the recurrence plot generating method according to any one of claims 9 to 11.

15. An analysis program which, when executed by a processing circuit, causes the processing circuit to perform the analysis method of claim 12.

Images