Time series data processing device

Abstract

To provide an information processing apparatus for time-series data, configured to support a user in determining a window width in a learning model for use in anomaly detection.SOLUTION: An information processing apparatus for time-series data is configured to support determining a window width that defines a range of training data for use in learning of a model, for time-series training data for training the model for use in anomaly detection for monitoring an anomaly of a target facility, with unsupervised learning, the apparatus comprising an arithmetic unit and a memory. The arithmetic unit includes: an analysis unit which calculates a plurality of candidate values of the window width and a plurality of feature quantities each representing a temporal feature of the training data with respect to the candidate values; and a display processing unit which displays, to a user, associations between the plurality of feature quantities and the plurality of candidate values calculated by the analysis unit.SELECTED DRAWING: Figure 1A

Description

[Technical field]

[0001] The present invention relates to an information processing apparatus for time series data. [Background technology]

[0002] The industrial sector is facing the problem of rising maintenance costs due to the aging of infrastructure and equipment. To solve this problem, instead of the conventional periodic inspections, systems and services that perform predictive maintenance (CBM: Condition Based Maintenance) are being introduced. This allows for the prediction of failures and deterioration due to aging (hereinafter referred to as anomalies) from measurement data of infrastructure and equipment, and maintenance is carried out based on the results.

[0003] Anomalies that appear in maintenance are often time-series data that changes, such as motor torque and current, plant pressure and temperature conditions, etc. Because the dynamic characteristics (time-series behavior) of such data change, an anomaly detection system that can handle aging changes is required to determine whether an anomaly exists by looking at the overall transition of the time-series data.

[0004] In such predictive maintenance systems, general machine learning algorithms are used to detect anomalies that can be observed from such time series data, but in addition to this, it is important to construct features from the time series data (preprocessing).In recent years, deep learning architectures that directly consider time series and incorporate preprocessing mechanisms have been developed, and examples of such architectures include CNN (Convolutional neural network) and LSTM (Long short-term memory).

[0005] It is necessary to detect abnormal phenomena by generating features for the time series data through such preprocessing of the time series data. To do this, it is necessary to set a common parameter called "window width" that greatly affects the performance of detecting time series abnormalities. The definition of this window width differs depending on the method. For example, in the case of the sliding window method (a method of converting time series data into a vector of a dimension corresponding to the window width), which is a representative preprocessing method, the number of dimensions when vectorizing the time series data is the window width. In addition, in the case of the short-time Fourier transform (a method of extracting peak frequencies, which are representative frequency components of time series data, and the corresponding amplitudes and phases), the window width is the length of the domain of the window function. In addition, in the case of a convolutional neural network (CNN), the kernel size of the convolutional layer is the window width.

[0006] In anomaly detection, it is important that the length of the window width that is set matches the phenomenon of the target equipment. For example, if the window width is set too short, time changes cannot be captured, and long-period vibrations and trends of the target equipment cannot be learned. Conversely, if the window width is set too long, patterns that have different characteristics and should be distinguished in the time series cannot be distinguished. In other words, the need to set an appropriate window width length to match the abnormal phenomenon is a challenge in general anomaly detection targeting time series.

[0007] The method of determining the window width, which selects a length that matches the target phenomenon, differs depending on the problem setting of anomaly detection. For example, in the case of "supervised learning" where labeled data is available, the window width can be easily determined by minimizing the label discrimination performance as an objective function (solving an optimization problem). However, as mentioned above, in many industrial sites, labeled data on anomalies is not available, or even if it is available, the amount of data is far less than the amount of normal data. In this case, "unsupervised learning" is used, and since a self-evident objective function like that used in supervised learning cannot be obtained, the window width must be determined on a case-by-case basis. If an appropriate window width cannot be determined, it is not possible to determine control of the target equipment to respond to anomalies, so it is necessary to consider a learning method that suits the characteristics of the time series data and the problem setting.

[0008] As an example of anomaly detection using unsupervised learning, for example, Patent Document 1 proposes an anomaly detection system that targets motor equipment and, when the operating frequency of the motor is known, determines an abnormal state of the motor equipment by using a window width determined based on the frequency width. [Prior art documents] [Patent documents]

[0009] [Patent Document 1] JP 2016-156643 A Summary of the Invention [Problem to be solved by the invention]

[0010] In Patent Document 1, the operating frequency is recognized as a feature value that represents the operation of the target equipment, and it is assumed that prior knowledge can be obtained, thereby determining the window width. However, in many industrial sites, such characteristics are not necessarily easily obtained, and there is no physical quantity that can be used as a reference for determining the window width. Therefore, the only way to determine the window width is to input statistics obtained from data (such as an anomaly score in anomaly detection) into an algorithm, which creates a problem of undermining the reliability of the judgment results.

[0011] In view of this, an object of the present invention is to provide an information processing device for time-series data that assists a user in determining a window width in a learning model used for anomaly detection. [Means for solving the problem]

[0012] The information processing device for time series data of the present invention assists in determining a window width that defines a range of time series training data used to learn a model used in anomaly detection that monitors abnormalities in target equipment by unsupervised learning, and is equipped with a calculation unit and a memory, wherein the calculation unit executes an analysis unit that calculates a plurality of candidate values ​​for the window width and a plurality of feature amounts that respectively represent temporal characteristics of the training data for the plurality of candidate values, and a display processing unit that displays to the user the correspondence between the plurality of candidate values ​​and the plurality of feature amounts calculated by the analysis unit. Effect of the Invention

[0013] It is possible to provide a user with an information processing device for time-series data that assists in determining the window width in a learning model used for anomaly detection. [Brief description of the drawings]

[0014] [Figure 1A] FIG. 1 is a block diagram showing the overall configuration of an unsupervised learning system (anomaly detection system) according to an embodiment of the present invention. [Figure 1B] 1 is an example of a hardware configuration of an information processing device for time-series data according to an embodiment of the present invention. [Diagram 2] 2 is a diagram showing an example of the configuration of a time-series data management table in FIG. 1; [Diagram 3] 13 is a flow chart showing a window width determination process according to an embodiment of the present invention. [Figure 4] 13 is an example of a selection screen for a reference value of a window width according to an embodiment of the present invention. [Diagram 5] 13 is an example of a display screen of a regression analysis result based on a selected window width reference value according to an embodiment of the present invention. [Figure 6] 11 is a flow chart showing a display flow of an inference result according to an embodiment of the present invention. [Figure 7] 11 is a display screen of an inference result according to an embodiment of the present invention.

[0015] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The following description and drawings are examples for explaining the present invention, and appropriate omissions and simplifications have been made for clarity of explanation. The present invention can be implemented in various other forms. Unless otherwise specified, each component may be singular or plural.

[0016] In order to facilitate understanding of the invention, the position, size, shape, range, etc. of each component shown in the drawings may not represent the actual position, size, shape, range, etc. Therefore, the present invention is not necessarily limited to the position, size, shape, range, etc. disclosed in the drawings.

[0017] (One embodiment of the present invention and overall configuration) (Figure 1A) The unsupervised learning system 1 is one of the anomaly detection systems that includes the time-series data information processing device of the present invention (FIG. 1B) and detects anomalies in target equipment 7. Hereinafter, the unsupervised learning system 1 will be described as anomaly detection system 1. The anomaly detection system 1 includes a time-series data management table 2, a display processing unit 3, a display unit 4, an analysis unit 5, and a learning and inference unit 6. The time-series data management table 2 acquires and records time-series data when the anomaly detection system 1 measures the target equipment 7. An example of the time-series data recorded in the time-series data management table 2 will be described later with reference to FIG. 2.

[0018] The time-series data management table 2 transmits the time-series data acquired from the target equipment 7 to the display processing unit 3, the analysis unit 5, and the learning and inference unit 6. The display unit 4 displays to the user the data processed by the display processing unit 3. The display processing unit 3, the analysis unit 5, and the learning and inference unit 6 will be described in detail in the description of Figs. 3 to 7.

[0019] (Figure 1B) An embodiment of a hardware configuration of an information processing device for time series data of the present invention will be described. An information processing device 10 equipped with an unsupervised learning system 1 (anomaly detection system 1) has an input device 11, an output display device 12, an input / output interface 13, a communication interface 14, a RAM (Random Access Memory) 15, a CPU (Central Processing Unit) 16, and a ROM (Read Only Memory) 17.

[0020] The information processing device 10 acquires time series data from the target equipment 7 via the communication interface 14, and the acquired time series data is stored in a database (not shown) having a storage function. The ROM 17, which is a memory, stores and holds pre-written data (programs, etc.). When the CPU 16 starts processing (executing as a calculation unit), the analysis unit 5, learning and inference unit 6, and display processing unit 3 function to acquire the time series data stored in the database. The analysis unit 5, learning and inference unit 6, and display processing unit 3 each perform data processing using the working area of ​​the RAM 15, which is a memory.

[0021] The data processed by the display processing unit 3, the analysis unit 5, and the learning and inference unit 6 is displayed to the user on the output display device 12 (display unit 4) via the input / output interface 13. The user judges the displayed data and gives instructions to the information processing device 10 by performing an input operation using the input device 11. Note that the input device 11 may be the display unit 4 if it is a touch panel type.

[0022] (Figure 2) The time-series data management table 2 shown in Fig. 2 is an example assuming that the target equipment 7 from which data is acquired is a mechanical device (particularly a motor), and the current and acceleration values ​​measured by multiple current sensors and acceleration sensors equipped in the target equipment 7 are recorded in the table as acquired data together with a timestamp. Note that #1 and #2 written in columns 201 and 202, respectively, represent sensor numbers. Also, the time-series data management table 2 may be defined by a database schema, or may be input each time using an Excel file, and there is no clear definition.

[0023] (Fig. 3, Fig. 4) The anomaly detection system 1 has a function of presenting to the user a recommended window width that determines the range of time-series training data (learning data) for learning a model used for anomaly detection. The flow of determining the window width presented to the user will be explained using the data interval selection screen in Fig. 4 along with the flowchart in Fig. 3. Note that the screen in Fig. 4 is assumed to be capable of allowing the user to input data.

[0024] First, the user selects time-series training data for learning a model used in anomaly detection by unsupervised learning from the data stored in the time-series data management table 2 (S1). Specifically, the user inputs a file path of the target time-series data in the data selection field 401 in Fig. 4. Note that the training data may be read not only in a file format, but also by specifying a column corresponding to the target time interval for the time-series data recorded in the database.

[0025] Next, the display processing unit 3 displays the time waveform for the training data selected in the data selection field 401 in the data display unit 402 (S2). The user inputs a reference value for the window width in the input unit 403 with reference to the time waveform displayed in the data display unit 402 (S3). In the input unit 403, the user can determine the reference value for the window width by inputting a value in the value input unit 4031 and selecting the unit of the value in the pull-down menu 4032.

[0026] The reference value of the window width set by the user in the input unit 403 is a value that is sufficiently larger than the length of the time fluctuation inherent in the training data. The meaning of taking a sufficiently large value will be explained. For example, the waveform (current, acceleration) of the selected training data shown in the data display unit 402 initially fluctuates at a constant value, but then the waveform oscillates and rises. The user visually confirms the time during which this waveform oscillates and rises, that is, the length of time when the waveform changes, and sets the reference value of the window width in the input unit 403 so that the length of that time falls within the range.

[0027] Specifically, in the data display section 402, the four time series graphs show significant changes in the graphs over a 20-second period from approximately 10:24:20 to 10:24:40, but to ensure that these 20 seconds are within the range, the user inputs a value in the input section 403 using 30 seconds as the reference value for the window width, being generous.

[0028] When the user sets a reference value for the window width in input unit 403 on the screen of display unit 4, analysis unit 5 acquires the reference value for the window width input in input unit 403 from display processing unit 3, and generates a plurality of candidate window width values. Analysis unit 5 performs a regression analysis on the generated plurality of candidate window width values, thereby making it possible to display a recommended window width value to the user, as shown in Fig. 5 described below. The implementation of this regression analysis corresponds to the pre-processing process described above.

[0029] The regression analysis performed by the analysis unit 5 to find the recommended window width will now be described. The analysis unit 5 uses a series of K window width candidate values ​​τ∈{τ1, τ2, ..., τ K A regression analysis is performed for each candidate value of τ ∈ {τ 1 , τ 2 , ..., τ} (S4). Note that the reference value of the window width input by the user via the input unit 403 is τ0, and the candidate values ​​of the window width are τ ∈ {τ 1 , τ 2 , ..., τ K}. The candidate values ​​τ (in terms of magnitude relationship) are τ0>τ1>τ2>…>τ K The value is set to satisfy the following. Also, K=10, τ K = τ0 / 100, τ0, τ1, τ2, ..., τ KA sufficiently large number of candidate window widths are chosen so that the logarithmic intervals are uniform.

[0030] Regression analysis begins with vectorization (calculation of equation (1) described below). Given a value yt taken at a discrete point in time t corresponding to a time among some time series data, and a window width τ, a vector is generated from time t by τ (including data delayed by τ-1). In other words, this operation has the effect of generating a vector that arranges past data from time t. In this way, the analysis unit 5 constructs a linear regression model (calculation of equation (2) described below) of the training data that predicts data from time t to the next time t+1 based on the generated vector for multiple candidate window width values ​​including the reference value of the window width.

[0031] A more detailed explanation will be given below. First, the analysis unit 5 converts the time-series training data into vectors using the following formula (1) with a window width τ. Note that this vector conversion is a common vector conversion called a sliding window in the fields of statistical analysis and machine learning, and a delayed coordinate system in the field of dynamical systems. In addition, in the time-series data, the value of the training data at time t is expressed as y t ∈R M (t=0, . . . N-1), where M is the dimension of the training data, and in this embodiment, M=4.

[0032]

number

[0033] Next, the analysis unit 5 analyzes the data X t (τ) Using the above, a linear regression model shown in the following equation (2) is trained (constructed) by the least squares method or the maximum likelihood method.

[0034]

number

[0035] Here, A τ ∈RMτ×Mτ is the matrix for one-time-ahead prediction. τ If it is possible to decompose A into eigenvalues, τ Eigenvalue λ1 (τ) , λ2 (τ) , … λ τ (t) The time scale (time constant and period of fluctuation) inherent in the training data can be expressed by using ∈C (where overlap is allowed). This time scale is used for displaying to the user as multiple feature quantities each representing the temporal characteristics of the training data for multiple window width candidate values ​​generated by the analysis unit 5.

[0036] Specifically, for each order p = 1, 2, ... τ, Re[lnλ p (τ) / Δt](eigenvalue λ at order p p (τ) The real part of the natural logarithm of λ divided by the time interval Δt of the training data represents the degree of divergence or decay per unit time, Im[lnλ p (τ) / Δt](eigenvalue λ at order p p (τ) The imaginary part of the natural logarithm of λ divided by the time interval Δt of the training data represents the speed (frequency) of oscillation per unit time. Therefore, the eigenvalue λ for the window width τ p (τ) is close to the value under the window width reference value τ0, it means that the long-period oscillations and trend components that can be learned under the window width τ0 can also be learned under the window width τ.

[0037] Also, the A of the linear regression model τ is a matrix with square elements, with a square matrix of dimension x dimension, the dimension of which is the window width. This allows equation (2) to function as an equation that predicts the characteristics of a time series, for example, the window width, and the data of the window width one time ahead. In other words, by using equation (2), it is possible to determine how the time series itself is deviating, and equation (2) functions as a regression model suitable for determining time series anomalies.

[0038] Although the linear regression model of formula (2) has been described above, formula (2) may be a nonlinear regression model. t (τ) This vector can be converted into another vector using a neural network, for example, and then a linear regression model can be constructed between the converted vectors. This means that the same processing can be achieved by performing one stage of neural network processing.

[0039] Based on the above, the analysis unit 5 calculates λ p (τ) and λ p (τ0) The window width τ=τ when the objective function consisting of the difference between * is calculated as the recommended window width to be shown to the user. This is the process of solving the optimization problem when showing the recommended window width. The objective function f(τ) is expressed, for example, by the following formula (3).

[0040]

number

[0041] The reason for introducing formula (3) will be explained. The prediction model (linear regression model) depends on the window width τ, but apart from the judgment of anomaly detection, the longer the window width, the more accurate the prediction. Therefore, under the longest standard window width τ0, the eigenvalue λ p (τ) However, if the eigenvalues ​​(feature quantities) for each window width can be reproduced with a window width shorter than the reference window width τ0, the long-period oscillations and trend components contained in the training data can be correctly learned. Therefore, if the difference between the eigenvalue (first feature quantity) under the reference value τ0 and the eigenvalue (second feature quantity) for other window widths τ is taken, and learning is performed so that this difference is reduced using equation (3), learning can be performed with a window width of the correct and appropriate length.

[0042] In other words, by using the linear regression model equation (2) for the window width determination problem and using equation (3) to solve the optimization problem based on the difference between each eigenvalue, a recommended window width can be extracted and displayed to the user.

[0043] Formula (3) will be specifically explained. d (τ)={Σ p∈p |λ p (τ) -λ p (τ0) | 2} 1 / 2 is the distance between the eigenvalue under window width τ and the value under window width τ0. The set p ⊆ {1, 2, ..., τ} of indexes p is determined by sorting the eigenvalues ​​by absolute value or the like. In this embodiment, in order to extract the vibration component, three indexes with the largest absolute value among those with positive imaginary parts of the eigenvalues ​​are extracted and sorted, and the set P is obtained.

[0044] Also, f r (τ)=τ0 / τ is a term to prevent the window width τ from becoming longer than necessary (i.e., approaching τ0). In the above equation, to ensure that the window width τ does not become too long, f r (τ) is multiplied by the parameter ε as a coefficient.

[0045] This parameter ε is determined based on the variance of eigenvalues ​​obtained using the bootstrap method, a type of method for sampling data using machine learning. The bootstrap method has the role of estimating parameters by sampling (appropriately extracting samples) and changing the data to be learned, and then averaging, for example, when you want to estimate the variance of a population of data but do not know the distribution function and must estimate it from the data.

[0046] The determination of ε is, for example, the data set {(x t (τ0) , x t+1 (τ0) )|t=τ-1, τ, . . . N-2}, we create multiple data sets by sampling with replacement. Then, we use each of these data sets to find the eigenvalue λ p (τ0)(p∈P) is trained. ε is determined by taking the average of the variance of the eigenvalues ​​for each calculated data set, that is, their variances for p∈P.

[0047] In the above regression analysis, the analysis unit 5 is assumed to learn a linear regression model. However, even in the case of learning a nonlinear regression model using a neural network or a kernel method, τ If there exists a matrix equivalent to

[0048] The results of the regression analysis by the analysis unit 5 are processed by the display processing unit 3, and the display unit 4 displays the processed results as a window width calculation result display area 501 shown in FIG. 5 (to be described later) (S5).

[0049] (Figure 5) FIG. 5 shows the results of pre-processing the data. In the window width calculation result display section 501, window width candidate values ​​501a are shown on a two-dimensional graph. Specifically, for each window width τ, the horizontal axis shows f d (τ) / √(ε), and the vertical axis represents the relative amount of data f r A number of window width candidate values ​​501a are plotted as (τ). Among the window width candidate values ​​501a, the recommended window width is the recommended window width value 501b indicated by a star.

[0050] The smaller the above-mentioned formula (3) is, the more appropriate the recommended window width value is (the smaller the value of the objective function f(τ) of the corresponding window width τ is). On this graph, the more appropriate the window width is, as the point of the window width candidate value 501a moves to the lower left. Furthermore, the recommended value τ calculated by formula (3) is * Based on this, a trade-off curve 501c, which is a curve showing the trade-off between the dynamic characteristic reproduction error and the relative data amount, is drawn on the graph of the window width calculation result display unit 501 using the following formula (4) obtained by transforming formula (3). Note that in the following formula (4), X and Y respectively represent the coordinates of the horizontal and vertical axes of the graph.

[0051]

number

[0052] Since trade-off curve 501c is a minimized evaluation function, all candidate window width values ​​displayed on trade-off curve 501c are recommended values ​​(in the present invention, only one recommended window width value 501b is displayed). In this way, by referring to the two-dimensional graph displayed in window width calculation result display section 501, the user can understand what trade-off was used to calculate the recommended window width value.

[0053] Furthermore, the display processing unit 3 has a function of displaying to the user the time scale (amount of time) of components that can be learned with each window width. This will be described. First, by clicking on each point on the two-dimensional graph in the window width calculation result display unit 501, a regression analysis result display unit 5011 is displayed as a property, and the details of the regression analysis with the corresponding window width can be confirmed. In addition to the dynamic characteristic reproduction error and the relative data amount, the regression analysis result display unit 5011 displays the natural period τ that can be learned with the selected window width. p (τ) =Im[lnλ p (τ) / Δt] (5s, 14.5s, 19.5s). The natural period is the eigenvalue under the window width, and the eigenvalue is a feature that indicates the time scale inherent in the training data.

[0054] The window width dependency display section 502 of the natural period is a screen for showing what the window width calculation result display section 501 physically means. In the window width dependency display section 502 of the natural period, the vertical axis is the natural period τ p (τ) Here, the top three eigenvalues ​​(natural periods) (5s, 14.5s, 19.5s) are extracted, and the changes in the top three natural periods depending on the window width are shown.

[0055] This makes it possible to visualize the effect on the natural period when the window width value is changed, and provides a basis for the user to make a final decision on the window width to be used for learning. In other words, the validity of the recommended window width value can be confirmed in the window width dependency display section 502 of the natural period by checking against the results of other window widths shown in the window width calculation result display section 501.

[0056] (Fig. 6, Fig. 7) The learning and inference unit 6 is a functional unit that performs sliding window or short-time Fourier transform based on the window width set in Fig. 4, generates features from time-series data, and uses these features to learn a machine learning model (performs unsupervised learning). The functions of the learning and inference unit 6 will be explained using the flowchart in Fig. 6 and the display screen of the inference results in Fig. 7.

[0057] The machine learning model may be any known machine learning model (algorithm), and can be realized by using known techniques such as k-nearest neighbors, clustering, linear regression, and autoencoder.

[0058] The learning and inference unit 6 uses training data (hereinafter, reference data) according to the window width determined by the user, inputs time series data different from the training data (hereinafter, test data) to the machine learning model, and performs unsupervised learning, and performs inference using the trained model (S11). The display processing unit 3 processes the results, and the display unit 4 displays them as shown in FIG. 7 (S12). In this way, the inference results are converted into the natural period τ p (τ) It gives the user an interpretation from the perspective of

[0059] 7 displays a time waveform 7011 of the anomaly score obtained by inference and its threshold value 7012. Here, when the time waveform 7011 of the anomaly score exceeds the threshold value 7012, it indicates that an anomaly has occurred in the target equipment 7.

[0060] In addition, in the inference result display unit 701, in order to show details of the abnormality of the target equipment 7, time waveforms 7013 of the test data of current #1, current #2, acceleration #1, and acceleration #2, and a waveform 7014 of reference data serving as a standard for the machine learning model are displayed alongside a time waveform 7011 of the abnormality score. Note that the reference data is data that is used as a standard when the machine learning model calculates the distance from the test data, and corresponds to, for example, the neighborhood of the test data in the case of the k-nearest neighbor method, or the prediction result in the case of a linear regression model.

[0061] In this way, time waveforms 7013, 7014 of the test data and reference data for multiple types of variables such as current #1, current #2, acceleration #1, and acceleration #2 are displayed side by side in the inference result display unit 701, allowing the user to visually confirm the deviation between the waveform 7013 of the test data and the waveform 7014 of the reference data when an abnormality occurs.

[0062] Furthermore, the learning and inference unit 6 receives the test data, the reference data, the window width τ used in learning and inference, and the corresponding eigenvalue λ p (τ) to the analysis unit 5, and requests the analysis unit 5 to interpret the inference result (S13).

[0063] The specific processing of the interpretation that the learning and inference unit 6 requests the analysis unit 5 to perform will be described. At time t, the test data and the reference data are transformed into delayed coordinates with a window width τ, and are denoted by x t (τ) , x ~ t (τ) In this case, the analysis unit 5 calculates each eigenvalue λ p (τ) Perform z-transform on (p∈P). The formula for z-transform is as shown in the following formula (5).

[0064]

number

[0065] Equation (5) expresses the eigenvalue λ in the test data and the reference data.p (τ) In this way, the analysis unit 5 calculates the contribution of each of the multiple feature amounts for the reference data and the test data. Based on the contribution of each natural period calculated by the formula (5), the display processing unit 3 calculates the contribution of the natural period T p (τ) For each test data, z-transform X p (τ) and (contribution to the test data), z-transformation of the reference data X ~ p (τ) and (degree of contribution to the training data) are displayed side by side in the inference result interpretation display section 702 (S14).

[0066] The natural periods of 5 s, 14.5 s, and 19.5 s displayed side by side in the inference result interpretation display section 702 are periods obtained when the recommended window width described above was determined, and represent the characteristic time scale for that window width. The inference result interpretation display section 702 breaks down the time series data into three representative natural periods and displays them.

[0067] The inference result interpretation display section 702 displays two bar graphs, one for the test data, which is the result at the time of inference, and the other for the reference data, which is the result at the time of model construction, side-by-side for each natural period. Looking at this, it can be seen that the contributions of the two bar graphs are roughly the same for a natural period of 5 s, and also for a natural period of 19.5 s, but only for a natural period of 14.5 s, the contributions of the two bar graphs differ, and are highlighted. This shows that the waveforms of the test data and the reference data are out of sync by the range of change of 14.5 s.

[0068] The user looks at the result of the inference result interpretation display section 702 and selects p (τ) and X ~ p (τ) When the difference between these is large, the length of time it takes for the abnormal phenomenon to occur (i.e., the time scale of the abnormality) can be ascertained as the natural period.

[0069] In other words, by characterizing the deviation between the waveform 7013 of the test data and the waveform 7014 of the reference data shown in the inference result display section 701, it is possible to visually confirm the duration over which the abnormal phenomenon of the target equipment 7 will occur.

[0070] In the above embodiment, the learning and inference unit 6 has been described as performing unsupervised anomaly detection. However, the present invention is not limited to this, and various unsupervised learning methods such as clustering and factor analysis may be performed.

[0071] As described above, the information processing device 10 for time series data provided in the anomaly detection system 1 (unsupervised learning system) of the present invention can use multivariate time series data consisting of measured values ​​of the target equipment 7, and does not require testing of an unsupervised learning model or statistical processing for determining the window width. In addition, while reproducing and clarifying the dynamic characteristics inherent in the time series data, it is possible to determine a window width value according to the purpose without using prior information on the target equipment 7, and detect changes over time in the dynamic characteristics of the target equipment 7. In addition, it is possible to reduce the process of determining the window width (learning time and execution time) while maintaining the performance of the machine learning model related to the accuracy of anomaly detection, and to visualize the anomaly detection results for the user.

[0072] In addition, by analyzing the effect that the window width value has on the unsupervised learning model and visualizing it as a physical quantity, an interpretation can be provided that enables the validity of the window width value to be confirmed before constructing a machine learning model, thereby assisting users in determining the window width.

[0073] In addition, by using such a predictive diagnosis system for abnormalities of the countermeasure equipment 7, the contents of the display screen can be used, for example, when presenting an abnormality report to a customer. Also, from the customer's point of view, by knowing the length of the abnormality, the equipment staff can determine the abnormal part of the machine equipment, which can encourage support for the abnormality detection part.

[0074] According to the embodiment of the present invention described above, the following advantageous effects are obtained.

[0075] (1) The information processing device 10 for time series data assists in the determination of a window width that defines a range of training data used for learning a model used in anomaly detection that monitors anomalies in a target facility 7, for time series training data for learning the model by unsupervised learning. The information processing device 10 for time series data includes a calculation unit and a memory. The calculation unit executes an analysis unit 5 that calculates multiple window width candidate values ​​and multiple feature quantities that respectively represent the temporal characteristics of the training data for the multiple candidate values, and a display processing unit 3 that displays to a user the correspondence between the multiple candidate values ​​and the multiple feature quantities calculated by the analysis unit 5. In this way, it is possible to provide the user with the information processing device 10 for time series data that assists in the determination of a window width in a learning model used for anomaly detection.

[0076] (2) The display processing unit 3 displays a display screen that allows the user to select a reference value for the window width, and the analysis unit 5 calculates multiple candidate values ​​based on the reference value selected by the user on the display screen. In this manner, a regression analysis can be performed and a recommended window width value can be proposed to the user.

[0077] (3) The analysis unit 5 constructs a regression model of the training data for each of a plurality of candidate values ​​including the reference value, calculates an objective function consisting of the difference between a first feature value generated from the regression model constructed for the reference value and a second feature value generated from each of the regression models constructed for the plurality of candidate values ​​excluding the reference value, and extracts the candidate value that minimizes the objective function as a recommended window width from among the plurality of candidate values, thereby calculating a recommended window width. In this way, a recommended window width can be proposed to the user.

[0078] (4) The feature amount is an eigenvalue of the regression model. The analysis unit 5 minimizes an objective function consisting of the difference and the ratio of the reference value to the window width candidate value for each of the multiple candidate values. The display processing unit 3 displays the difference and the ratio of the difference to the reference value for each of the multiple candidate values ​​on a two-dimensional graph. In this way, the recommended window width value can be displayed and visually confirmed by the user.

[0079] (5) The display processing unit 3 displays the change in the feature quantity with respect to the change in the window width. In this way, it is possible to show the user the physical meaning of the recommended window width value.

[0080] (6) The analysis unit 5 calculates the contribution degree of each of the multiple feature amounts for the training data and the test data, which is time-series data different from the training data. The display processing unit 3 displays the contribution degree for the training data and the contribution degree for the test data side by side, and when the difference between the two contribution degrees is large, the corresponding feature amount is highlighted. In this way, it is possible to visually confirm how long the abnormal phenomenon of the target equipment 7 will last.

[0081] (7) The information processing device 10 for time-series data includes a learning and inference unit 6 that performs unsupervised learning of a model using training data according to a window width determined by a user, and performs inference using a learned model. In this way, an interpretation of an abnormality in the target equipment 7 can be presented to the user, and support can be provided in making a judgment.

[0082] (8) The information processing device 10 for time-series data of the present invention performs anomaly detection using a trained model. In this manner, the information processing device 10 for time-series data of the present invention can be applied to anomaly detection in the target equipment 7.

[0083] The present invention is not limited to the above-described embodiment, and various modifications and other configurations can be combined without departing from the spirit of the present invention. Furthermore, the present invention is not limited to those having all of the configurations described in the above-described embodiment, and includes those in which some of the configurations are omitted. [Explanation of symbols]

[0084] 1...Unsupervised learning system (anomaly detection system), 2...Time series data management table, 3...Display processing unit, 4...Display unit, 5...Analysis unit, 6...Learning and inference unit, 7...Target equipment, information processing device...10

Claims

1. An information processing device for time series data, The method assists in determining a window width that defines a range of time-series training data used to learn a model used in anomaly detection for monitoring anomalies in a target facility by unsupervised learning, the window width defining the range of the training data used to learn the model, A calculation unit and a memory, The calculation unit, an analysis unit that calculates a plurality of candidate values ​​for the window width and a plurality of feature amounts that respectively represent temporal features of the training data for the plurality of candidate values; a display processing unit that displays to a user the correspondence relationships between the plurality of candidate values ​​and the plurality of feature quantities calculated by the analysis unit; 2. An information processing apparatus for time series data, comprising:

2. 2. The information processing device for time series data according to claim 1, the display processing unit displays a display screen that allows the user to select a reference value for the window width; The analysis unit calculates a plurality of the candidate values ​​based on the reference value selected by the user on the display screen.

23. An information processing apparatus for time series data comprising:

3. 3. The information processing device for time series data according to claim 2, The analysis unit includes: constructing a regression model of the training data for each of the plurality of candidate values ​​including the reference value; An objective function consisting of a difference between a first feature amount generated from the regression model constructed for the reference value and a second feature amount generated from each of the regression models constructed for the plurality of candidate values ​​excluding the reference value is calculated, and the recommended window width is calculated by extracting, from among the plurality of candidate values, a value that minimizes the objective function as the recommended window width.

23. An information processing apparatus for time series data comprising:

4. 4. The information processing device for time series data according to claim 3, the feature value is an eigenvalue of the regression model, The analysis unit minimizes, for each of the plurality of candidate values, an objective function consisting of the difference and a ratio of the reference value to the candidate value of the window width; The display processing unit displays, for each of the plurality of candidate values, the difference and a ratio of the difference to the reference value on a two-dimensional graph.

23. An information processing apparatus for time series data comprising:

5. 2. The information processing device for time series data according to claim 1, The display processing unit displays a change in the feature amount with respect to a change in the window width.

23. An information processing apparatus for time series data comprising:

6. 2. The information processing device for time series data according to claim 1, The analysis unit calculates a contribution degree of each of a plurality of feature amounts for the training data and test data, which is time-series data different from the training data; The display processing unit displays the contribution degree to the training data and the contribution degree to the test data side by side, and when a difference between the two contribution degrees is large, displays the corresponding feature amount in an emphasized manner.

23. An information processing apparatus for time series data comprising:

7. An information processing apparatus for time series data according to any one of claims 1 to 6, a learning and inference unit that performs unsupervised learning of the model using the training data corresponding to the window width determined by the user, and performs inference using the trained model.

23. An information processing apparatus for time series data comprising:

8. The information processing device for time series data according to claim 7, The anomaly detection is performed using the trained model.

23. An information processing apparatus for time series data comprising:

Images