Model generation support device, model generation support method, and program

Abstract

To provide a model generation support device.SOLUTION: A model generation support device comprises: a data acquisition unit which acquires time-series data detected by a sensor installed on an industrial machine; a condition acceptance unit which accepts extraction conditions for extracting section data from the time-series data from a user; a section extraction unit which extracts the section data from the time-series data on the basis of the extraction conditions; a model generation unit which generates an abnormality detection model by unsupervised learning on the basis of the section data; and a distribution display unit which displays a graph indicative of the distribution of abnormality degrees calculated by the abnormality detection model.SELECTED DRAWING: Figure 5

Description

[Technical Field]

[0001] The present invention relates to a model generation support device, a model generation support method, and a program. [Background technology]

[0002] Patent Document 1 discloses an anomaly detection method that focuses on the behavior of data over time, divides trajectories into clusters over time, models the divided clusters using the subspace method, and calculates outliers as anomaly candidates in order to detect anomalies in a plant or facility at an early stage. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 5048625 Summary of the Invention [Problem to be solved by the invention]

[0004] Anomaly detection using time-series data collected by sensors installed in industrial machinery is prone to overlooking abnormal signs when they are buried in complex operational patterns, or to overdetection when changes in operational conditions are mistakenly interpreted as abnormal signs. Therefore, it is necessary to apply appropriate preprocessing to the time-series data. However, it is difficult to extract interval data suitable for anomaly detection from time-series data and to obtain an accurate anomaly detection model.

[0005] The present invention has been made in consideration of the above-mentioned problems, and its main purpose is to provide a model generation support device, a model generation support method, and a program that make it easy for a user to evaluate an anomaly detection model generated from section data extracted using extraction conditions specified by the user. [Means for solving the problem]

[0006] In order to solve the above problem, one aspect of the present invention provides a model generation support device that includes: a data acquisition unit that acquires time-series data detected by a sensor installed in an industrial machine; a condition acceptance unit that accepts, from a user, extraction conditions for extracting section data from the time-series data; a section extraction unit that extracts the section data from the time-series data based on the extraction conditions; a model generation unit that generates an anomaly detection model through unsupervised learning based on the section data; and a distribution display unit that displays a graph showing the distribution of anomalies calculated by the anomaly detection model. This makes it easy for a user to evaluate the anomaly detection model generated from the section data extracted using the extraction conditions specified by the user.

[0007] In the above aspect, the apparatus may further include an extraction display unit that displays the section data extracted based on the extraction conditions, thereby making it easier to visually grasp the extracted section data.

[0008] In the above aspect, the extraction display unit may further display a data selection receiving area for receiving selection of time-series data from which the section data is to be extracted from the plurality of types of time-series data, thereby facilitating selection of time-series data.

[0009] In the above aspect, the extraction display unit may further display a condition input receiving area for receiving input of the extraction conditions, which makes it easier to input the extraction conditions.

[0010] In the above aspect, the condition input receiving area may further receive selection of auxiliary data related to control of the industrial machine, which is used to extract the section data, thereby facilitating selection of auxiliary data.

[0011] In the above aspect, the extraction display unit may further display an extraction target display area that displays the time-series data and an extraction result display area that displays the section data extracted from the time-series data, thereby making it easy to compare the time-series data to be extracted with the section data of the extraction result.

[0012] In the above aspect, the extraction result display area may further display auxiliary data related to control of the industrial machine that was used to extract the section data, making it easier to visually grasp the relationship between the section data and the auxiliary data.

[0013] In the above aspect, the distribution display unit may further display a data selection receiving area for receiving a switch of the display of the section data in the graph, which makes it easy to switch the display of the section data.

[0014] In the above aspect, the distribution display unit may further display a parameter setting reception area for receiving settings of parameters of the anomaly detection model, which makes it easier to set the parameters.

[0015] In the above aspect, the section extraction unit may extract the section data by detecting change points in the time-series data, identifying a code representing a trend at each of the change points, and clustering the change patterns of the codes. This makes it possible to extract the section data by clustering the change patterns of the codes.

[0016] In the above aspect, the distribution display unit may display, as the graph, a contour diagram in which the degree of abnormality is divided into a predetermined number of gradations, thereby making it easier to visually grasp the distribution of the degree of abnormality.

[0017] In the above aspect, the distribution display unit may display the section data in the graph and further include a correction receiving unit that receives a correction of the section data from a user, and the model generation unit may re-train the anomaly detection model based on the corrected section data. This makes it possible to re-train the anomaly detection model based on the section data corrected by the user.

[0018] In the above aspect, the model generation unit may calculate a decision boundary for anomaly detection, and the distribution display unit may display the decision boundary within the graph and further include a correction receiving unit that receives a correction of the decision boundary from a user. This makes it possible to receive a correction of the decision boundary from a user.

[0019] Another aspect of the model generation support method of the present invention acquires time series data detected by a sensor installed in an industrial machine, receives extraction conditions from a user for extracting section data from the time series data, extracts the section data from the time series data based on the extraction conditions, generates an anomaly detection model by unsupervised learning based on the section data, and displays a graph showing the distribution of anomalies calculated by the anomaly detection model.

[0020] Furthermore, another aspect of the program of the present invention causes a computer to acquire time series data detected by a sensor installed in industrial machinery, receive from a user extraction conditions for extracting section data from the time series data, extract the section data from the time series data based on the extraction conditions, generate an anomaly detection model by unsupervised learning based on the section data, and display a graph showing the distribution of anomaly degrees calculated by the anomaly detection model. [Effects of the Invention]

[0021] According to the present invention, it becomes easy for a user to evaluate an anomaly detection model generated from section data extracted using extraction conditions specified by the user. [Brief explanation of the drawings]

[0022] [Figure 1] FIG. 1 illustrates an example of the configuration of a model generation support device. [Figure 2] FIG. 4 is a diagram illustrating a display example of a display unit. [Figure 3] FIG. 4 is a diagram illustrating a display example of a display unit. [Figure 4] FIG. 4 is a diagram illustrating a display example of a display unit. [Figure 5] FIG. 2 is a diagram illustrating an example of the configuration of a control unit. [Figure 6] FIG. 10 is a diagram illustrating an example of a procedure of a model generation support method. [Figure 7] FIG. 1 illustrates an example of a segmentation technique. [Figure 8] FIG. 1 illustrates an example of a segmentation technique. [Figure 9] FIG. 1 illustrates an example of a segmentation technique. [Figure 10] FIG. 1 illustrates an example of a segmentation technique. [Figure 11] FIG. 1 illustrates an example of a segmentation technique. [Figure 12] FIG. 4 is a diagram illustrating a display example of a display unit. [Figure 13] FIG. 4 is a diagram illustrating a display example of a display unit. [Figure 14] FIG. 4 is a diagram illustrating a display example of a display unit. [Figure 15] FIG. 10 is a diagram illustrating an example of a procedure of a model generation support method. [Figure 16] FIG. 4 is a diagram illustrating a display example of a display unit. [Figure 17] FIG. 10 is a diagram illustrating an example of a procedure of a model generation support method. [Figure 18] FIG. 4 is a diagram illustrating a display example of a display unit. DETAILED DESCRIPTION OF THE INVENTION

[0023] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0024] 1 is a block diagram showing an example configuration of a model generation support device 1. The model generation support device 1 includes a control unit 10, a storage unit 2, an operation unit 3, and a display unit 4. The model generation support device 1 is a device that supports the generation of an anomaly detection model that detects anomalies in industrial machinery 100.

[0025] The control unit 10 is a computer including a CPU, RAM, ROM, nonvolatile memory, an input / output interface, etc. The CPU of the control unit 10 executes information processing in accordance with a program loaded from the ROM or nonvolatile memory to the RAM.

[0026] The program may be supplied via an information storage medium such as an optical disk or a memory card, or may be supplied via a communication network such as the Internet or a LAN. The storage unit 2 is, for example, a hard disk drive (HDD) or a solid state drive (SSD), etc. The operation unit 3 is, for example, a keyboard, a mouse, a touch panel, etc. The display unit 4 is, for example, a liquid crystal display device or an organic EL display device, etc.

[0027] The storage unit 2 stores sensor signals detected by sensors 102 installed in the industrial machine 100. The sensor signals are digital signals representing physical quantities detected by the sensors 102, and are an example of time-series data.

[0028] The industrial machine 100 is a machine used in, for example, a factory, etc. The industrial machine 100 is equipped with a plurality of types of sensors 102, and the storage unit 2 stores a plurality of types of sensor signals.

[0029] The sensor signal detected by the sensor 102 is supplied from the storage unit 2 to the control unit 10. However, the sensor signal may be supplied directly from the sensor 102 to the control unit 10, or may be supplied to the control unit 10 from a remote storage device via a communication network.

[0030] However, when detecting anomalies in industrial machinery that involves complex operating patterns, load fluctuations, and rotation control, it is difficult to obtain an accurate anomaly detection model simply by applying unsupervised anomaly detection methods using machine learning to sensor signals.

[0031] This is because unsupervised anomaly detection methods using machine learning usually assume the existence of relatively simple operational patterns, such as a single operational state, periodicity or stationarity of data, etc. If such methods are applied directly to sensor signals with complex operational patterns, they are likely to be overlooked because abnormal signs are buried in the complex operational patterns, or to be overdetected by mistaking simple changes in the operational status for abnormal signs.

[0032] For this reason, there are conventional examples in which sensor signals are divided into multiple data segments by using data segmentation using discrete auxiliary signals or partial time series clustering, and anomaly detection methods are applied to each cluster. However, even if the data segments show similar patterns, the actual operating conditions may be different, and there are limitations to methods that can extract specific operating modes and conditions without taking domain knowledge into account.

[0033] In particular, the current values ​​of drive motors used in industrial machinery are easily affected by subtle differences in operating conditions, and it is expected that abnormal signs will appear among small fluctuations, so utilizing domain knowledge is essential.

[0034] Additionally, there are challenges with model tuning. When adjusting an unsupervised anomaly detection model using machine learning, it is customary to tune it based on indicators such as the false negative rate, false positive rate, and F1 score, assuming that a small amount of anomalous data has been collected. Furthermore, when no anomalous data has been collected, pseudo-anomalous data, which is normal data to which stochastic noise has been added, is widely used.

[0035] However, when the operational situation is complex, past operational situations that constitute abnormal cases are unique, so it may not be appropriate to adjust the model based on the abnormal data collected, or the pseudo-abnormal data may be far removed from the actual situation at the time of the abnormality, making it difficult to use for model evaluation. Therefore, it is difficult to achieve high detection accuracy by simply adjusting the model based on indicators without using domain knowledge.

[0036] To deal with such situations, it is essential to utilize domain knowledge to perform hypothesis-driven data preprocessing based on anticipated failure scenarios (for example, extracting data during no-load operation, which is assumed to make it easier to identify abnormal characteristics), and to have operating staff with extensive knowledge of industrial machinery evaluate model behavior.

[0037] Therefore, in this embodiment, as will be described below, a mechanism is realized that allows even operation staff who are not familiar with machine learning or programming to easily perform data preprocessing and model generation / evaluation.

[0038] 2 and 3 are diagrams showing examples of displays on the display unit 4 of the model generation support device 1. Fig. 2 is a diagram showing an example of a data preprocessing screen 5 for performing preprocessing of sensor signals. Fig. 3 is a diagram showing an example of a model evaluation screen 6 for generating and evaluating an anomaly detection model.

[0039] 2, the data preprocessing screen 5 has a data selection reception area 51, an extraction target display area 52, a condition input area 53, and an extraction result display area 54. On the data preprocessing screen 5, it is possible to perform preprocessing and visualize the processing results by operating buttons.

[0040] The data selection receiving area 51 is an area for receiving the selection of a sensor signal to be segmented from multiple types of sensor signals. Segmentation is an example of extracting section data. The data selection receiving area 51 includes a file selection area 511, a process setting area 512, and a selection adjustment area 513.

[0041] The file selection area 511 is an area for selecting a file containing a sensor signal. The processing setting area 512 is an area for setting smoothing and downsampling for the sensor signal. The selection adjustment area 513 is an area for selecting the sensor signal to be used for anomaly detection and setting the time shift and difference calculation.

[0042] The extraction target display area 52 is an area for visualizing the sensor signal of the segmentation target, and includes extraction target graphs 521 and 522 that show the change over time in the sensor signal of the segmentation target.

[0043] The condition input area 53 is an area for receiving input of a segmentation method. The segmentation method is an example of an extraction condition. The condition input area 53 is also an area for receiving an auxiliary signal used for segmentation. The auxiliary signal is a digital signal related to the control of the industrial machine 100 and is an example of auxiliary data.

[0044] The condition input area 53 includes an extraction method selection area 531 and a method detail setting area 532. The extraction method selection area 531 is an area for selecting one of a plurality of segmentation methods and for selecting an auxiliary signal. The method detail setting area 532 is an area for setting the details of the segmentation method.

[0045] The extraction result display area 54 is an area for visualizing the segmentation results. The extraction result display area 54 includes extraction result graphs 541 and 542 that show segments extracted from the sensor data by segmentation. A segment is an example of section data.

[0046] The extraction result graphs 541 and 542 distinguishably display the segments by changing the color or thickness of the segments, etc. The extraction result graphs 541 and 542 may further display auxiliary signals used in the segmentation.

[0047] By displaying the extraction target display area 52 and the extraction result display area 54, it becomes easy to visually grasp the relationship between the sensor signal to be segmented and the segments resulting from the segmentation.

[0048] 3, the model evaluation screen 6 has a data selection reception area 61, a parameter setting reception area 62, and a graph display area 63. On the model evaluation screen 6, an anomaly detection model can be generated and evaluated by operating buttons.

[0049] The data selection receiving area 61 is an area for receiving the selection of the segments and the target period to be used for generating an anomaly detection model by unsupervised learning. The data selection receiving area 61 is also an area for receiving switching of the display of segments in the anomaly degree graph 631 displayed in the graph display area 63.

[0050] The parameter setting reception area 62 is an area for receiving settings of hyperparameters used to generate an anomaly detection model by unsupervised learning.

[0051] The graph display area 63 is an area for visualizing the anomaly detection model generated by unsupervised learning and the segments used to generate it. The graph display area 63 includes an anomaly degree graph 631 that shows the distribution of anomaly degrees calculated by the anomaly detection model.

[0052] 4, the abnormality degree graph 631 is a contour diagram in which the degree of abnormality is divided into a predetermined number of gradations. The abnormality degree graph 631 is also a heat map in which the gradations of the degree of abnormality are expressed by color or shade.

[0053] Training data points 632 representing the segments used to generate the anomaly detection model and a decision boundary 633 serving as a threshold for anomaly detection are displayed within the anomaly degree graph 631. The anomaly degree graph 631 can also be considered a scatter plot on which the training data points 632 are plotted.

[0054] 5 is a block diagram showing an example of the functional configuration of the control unit 10 of the model generation support device 1. The control unit 10 includes a data acquisition unit 11, a condition reception unit 12, a section extraction unit 13, an extraction display unit 14, a setting reception unit 15, a correction reception unit 16, a model generation unit 17, a distribution display unit 18, and a model storage unit 19.

[0055] These functional units are realized by the CPU of the control unit 10 executing information processing in accordance with a program loaded from the ROM or nonvolatile memory to the RAM.

[0056] The data acquisition unit 11 acquires from the storage unit 2 a sensor signal (an example of time-series data) detected by a sensor 102 installed in the industrial machine 100.

[0057] The condition receiving unit 12 receives input of a segmentation method (an example of an extraction condition). The segmentation method is input into a condition input area 53 of a data preprocessing screen 5 (see FIG. 2) displayed on the display unit 4 by a user operating the operation unit 3.

[0058] The condition receiving unit 12 also receives a selection of a sensor signal to be segmented. The selection of the sensor signal is input into the data selection receiving area 51 of the data preprocessing screen 5 by the user operating the operation unit 3.

[0059] The section extraction unit 13 performs segmentation based on the segmentation method accepted by the condition acceptance unit 12, and extracts segments (examples of section data) from the sensor signal.

[0060] The extraction display unit 14 displays the sensor signal to be segmented in the extraction target display area 52 of the data preprocessing screen 5. The extraction display unit 14 also displays the segments extracted from the sensor signal by segmentation in the extraction result display area 54 of the data preprocessing screen 5.

[0061] The setting reception unit 15 receives selection of segments to be used for generating an anomaly detection model and switching of display. The selection of segments and switching of display are input by the user operating the operation unit 3 into the data selection reception area 61 on the model evaluation screen 6 (see FIG. 3 ).

[0062] The setting receiving unit 15 also receives settings of hyperparameters used to generate an anomaly detection model. The hyperparameter settings are input into the parameter setting receiving area 62 on the model evaluation screen 6 by a user operating the operation unit 3.

[0063] The correction receiving unit 16 receives a correction of the learning data points 632 or the decision boundary 633 (see FIG. 4 ) displayed in the anomaly degree graph 631. The correction of the learning data points 632 or the decision boundary 633 is input into the graph display area 63 of the model evaluation screen 6 by a user operating the operation unit 3.

[0064] The model generation unit 17 generates an anomaly detection model by unsupervised learning based on the segments extracted from the sensor signal by the section extraction unit 13.

[0065] The distribution display unit 18 displays, in the graph display area 63 of the model evaluation screen 6, an anomaly degree graph 631 that indicates the distribution of the anomaly degrees calculated by the anomaly detection model.

[0066] The model saving unit 19 saves the anomaly detection model generated by the model generating unit 17 in the storage unit 2. The model saving unit 19 also saves in the storage unit 2 the preprocessing logic used for segmentation.

[0067] 6 is a flow diagram showing an example of the procedure of a model support method realized in the model generation support device 1. The control unit 10 of the model support device 1 executes the information processing shown in the diagram in accordance with a program.

[0068] In S11, the control unit 10 selects sensor signals to be used for generating an anomaly detection model. Specifically, the control unit 10 acquires a file containing sensor signals from the storage unit 2 based on the content of an input to the data selection reception area 51 of the data preprocessing screen 5 (see FIG. 2), and selects multiple (e.g., two) sensor signals to be used for anomaly detection. The control unit 10 also optionally performs smoothing processing or downsampling. The control unit 10 also displays the selected sensor signals or the smoothed sensor signals in the extraction target display area 52 of the data preprocessing screen 5. The control unit 10 may also perform time shift processing (creating a lag feature and considering a value from, for example, t seconds ago as the current value) or difference (differential) processing on the sensor signals.

[0069] In S12, the control unit 10 selects an auxiliary signal to be used for segmentation. Specifically, the control unit 10 selects an auxiliary signal to be used for segmentation based on the input contents in the condition input area 53 of the data preprocessing screen 5. The selection of the auxiliary signal is intended to extract data corresponding to a specific operation pattern. Note that, depending on the segmentation method selected, an auxiliary signal may not be selected.

[0070] In S13, the control unit 10 constructs segmentation logic. Specifically, the control unit 10 determines a segmentation method based on the input contents in the condition input area 53 of the data preprocessing screen 5. The control unit 10 also displays the segments extracted from the sensor data by the determined segmentation method in the extraction result display area 54 of the data preprocessing screen 5.

[0071] Segmentation methods include, for example, (a) discretization by threshold processing (see FIG. 7), (b) partial time series clustering based on the pattern of the sign of the differential value at the change point (see FIG. 8), (c) on / off of an auxiliary signal (see FIG. 9), and (d) extraction using the above method plus a timer (see FIG. 10). These methods may also be combined as appropriate.

[0072] 7 is a diagram showing an example of a technique for applying threshold processing to a sensor signal or an auxiliary signal. Segments 1 to 4 are extracted depending on the relationship between the signal value y of the sensor signal or auxiliary signal and thresholds A to C. For example, of the sensor signal, a section where y≧A is extracted as segment 1, a section where A≧y≧B is extracted as segment 2, a section where B≧y≧C is extracted as segment 3, and a section where C≧y is extracted as segment 4.

[0073] Fig. 8 shows an example of partial time series clustering based on the pattern of differential values ​​at change points of the sensor signal. Specifically, as shown in Fig. 11, (a) change points of the sensor signal are detected, (b) the sign of the differential value at each change point is identified, and (c) the change pattern of the sign is clustered to extract segments.

[0074] In (a), the change point may be detected by a known change point detection method such as a locally stationary AR model, or by threshold processing of the differential value (difference value) of the sensor signal. A change point means a time point at which the characteristics of the time series data change suddenly.

[0075] In (b), the sign of the differential value is specified as, for example, a binary value, positive or negative. The differential value represents the tendency of change at the change point. Alternatively, the sign of the differential value may be obtained by binarizing any summary feature quantity near each change point using a threshold value.

[0076] In (c), the change patterns of the signs of the differential values ​​at adjacent change points are clustered. The clusters of the segments between adjacent change points are determined by the change patterns of the signs of the differential values ​​at the adjacent change points (i.e., the change points before and after the segment).

[0077] For example, in the sensor signal, a section where the change pattern of the sign of the differential value at adjacent change points is "+→-" is extracted as segment 1, a section where it is "-→-" is extracted as segment 2, a section where it is "-→+" is extracted as segment 3, and a section where it is "+→+" is extracted as segment 4.

[0078] In this example, the change patterns of the sign of the differential value at adjacent change points are clustered, but this is not limiting, and the change patterns of the sign of the differential value at three or more change points may be clustered. As a method for allocating clusters, the change patterns of the sign of the differential value and clusters may be allocated one-to-one, or a general clustering method may be used.

[0079] Hierarchical clustering may also be performed based on the binary distance of binary numbers constructed as follows. Binary numbers are constructed for n points around the segment, with positive and negative (0, 1) being defined as the ones', two's and four's digits, in order from the change point closest to the segment. However, here, the p-adic distance between x and y (p is a prime number, x and y are integers) is generally expressed as "(xy)=p V ×m" (m and p are relatively prime), then p -V is the distance defined by

[0080] By using this distance measure, if the differential value of the nearest change point between the segment in question and the segment being compared does not match in sign, the distance becomes large, so it is possible to neglect the situation far from the segment in question and emphasize the match of the nearby situation. Furthermore, as long as the feature value extracted from the change point (and its vicinity) is a discrete value, this method can be applied to cases where the feature value has three or more values, and the difference value can be anything other than positive or negative.

[0081] 9 is a diagram showing an example of a technique that uses the on / off of an auxiliary signal. For example, a section of the sensor signal where the auxiliary signal is on is extracted as segment 1, and a section where the auxiliary signal is eff is extracted as segment 2.

[0082] 10 shows an example of a technique for combining timer-based extraction. For example, a predetermined start timer and end timer are started when the auxiliary signal is turned on, and the section from when the start timer stops to when the end timer stops is extracted as segment 1, and the other section is extracted as segment 2.

[0083] The example in FIG. 10 shows a combination of the technique of using on / off of the auxiliary signal (see FIG. 9) and extraction using a timer, but this is not limiting, and extraction using a timer may also be combined with the technique of applying threshold processing (see FIG. 7).

[0084] Returning to the explanation of Figure 6, in S14, the control unit 10 generates an anomaly detection model. Specifically, the control unit 10 selects a segment to be subjected to anomaly detection and a period to be used based on the content of an input in the data selection receiving area 61 of the model evaluation screen 6 (see Figure 3). The control unit 10 also sets hyperparameters of the anomaly detection model based on the content of an input in the parameter setting receiving area 62 of the model evaluation screen 6. Then, the control unit 10 generates an anomaly degree graph 631 based on the selected segment and the set anomaly detection model, and displays it in the graph display area 63 of the model evaluation screen 6 (see Figure 3).

[0085] This allows operational staff to check the model behavior and adjust the hyperparameters to achieve the desired model behavior, even if there are no abnormal data items.

[0086] The abnormality degree graph 631 is a graph in which the gradation region is divided into n regions (n ​​is a predetermined number) so that the areas of the gradation regions are equal to each other, and each gradation region is color-coded. In other words, the gradation boundaries are set so that the areas of the gradation regions are equal to each other. In the example of Fig. 4, the abnormality degree graph 631 is divided into five gradation regions.

[0087] Even if there are three or more sensor signals used for anomaly detection, similar graph drawing is possible by compressing the data to two dimensions using a dimension reduction method such as principal component analysis, or by selecting two variables to visualize. For example, when principal component analysis is used, a graph can be drawn using the anomaly levels of data points on the plane spanned by the first and second principal component vectors.

[0088] Furthermore, when two variables to be visualized are arbitrarily selected, for the variables not selected, the mode or predicted values ​​from a regression analysis using the two selected variables as explanatory variables is used to calculate the degree of anomaly for each area on a graph with the two selected variables on the vertical and horizontal axes, allowing for graph drawing. Furthermore, based on the linear correlation of each variable (alternatively, sparse linear correlation using the graphical lasso, Spearman's rank correlation, or nonlinear correlation coefficients such as MIC may also be used), it is also possible to present candidate combinations of two variables that select highly correlated pairs.

[0089] 12, the graph display area 63 of the model evaluation screen 6 may display multiple anomaly degree graphs 631. The multiple anomaly degree graphs 631 are graphs generated based on multiple anomaly detection models that use different conditions, such as the segments, hyperparameters, and thresholds used. This allows the operation staff to compare multiple anomaly detection models with different conditions and consider the optimal segmentation method, hyperparameters, thresholds, and the like.

[0090] When evaluating and adjusting an anomaly detection model, it is important to note that operational data from industrial machinery is generally "non-iid (non-independent identically distributed)" time series data. For example, when operational factors such as motor acceleration and deceleration cause areas where the motor current value fluctuates according to a certain pattern to alternate with areas where the current value is stable, if data with an equal number of areas with small fluctuations and large fluctuations is plotted on a graph, the areas with less data fluctuation on the graph will have a higher data density.

[0091] When an outlier detection method such as LOF (Local Outlier Factor), which tightly evaluates dense regions and loosely evaluates sparse regions, is applied to data with different data densities in each region and an anomaly graph is drawn, the decision boundary will be relatively far away from the vicinity of normal data points in regions with low data density due to factors such as fluctuations in motor current values ​​caused by acceleration and deceleration.On the other hand, in regions with little data fluctuation and high data density, such as constant load areas, the decision boundary will remain near the normal data points.

[0092] 13, in a region Ac where the data density is relatively low, the degree of anomaly of the LOF model changes gradually with increasing distance from the position of the training data point 632. On the other hand, in a region Ar where the data density is relatively high, the degree of anomaly of the LOF model changes rapidly with increasing distance from the position of the training data point 632.

[0093] However, even in areas where the data density is relatively low, if operations are performed according to a certain operational pattern, the data pattern on the graph itself may be stable, and there may be cases where it is not necessary to set the decision boundary loosely. In such cases, an operation staff member who understands the relationship between each area on the graph and the operating state of the industrial machinery can easily determine, based on their own knowledge, just by looking at the graph, that it is inappropriate to set the decision boundary in a sparse area loosely.

[0094] For this reason, in this embodiment, the segmentation method, the use of segments, the hyperparameter settings, etc. can be easily changed by the operation staff on the data preprocessing screen 5 (see FIG. 2) and the model evaluation screen 6 (see FIG. 3). Also, in this embodiment, as will be described later, the operation staff can modify the training data points and the decision boundary.

[0095] Since the operating staff can determine that there are unnecessary learning data points on the graph, if it can be determined from the graph that data that should be excluded (for example, motor current value during load fluctuation) has been mixed into a selected segment (for example, motor current value during constant load) due to incomplete segmentation, the relevant data can be deleted by operating the screen, and a function may be provided that makes it possible to check changes in the anomaly distribution and judgment boundary (details will be described later).

[0096] Additionally, based on the local density of training data points, unnecessary regions (e.g., regions with too high a density that destabilizes the behavior of density-based anomaly detection logic such as LOF) may be automatically removed from the training and evaluation targets.

[0097] Furthermore, a function may be provided that allows operation staff to freely correct the decision boundary without using machine learning, redraw it, and set up anomaly detection logic (details will be described later). This allows operation staff to check and improve model behavior through intuitive screen operations.

[0098] Returning to the explanation of Figure 6, in S15, the control unit 10 deploys the anomaly detection model. That is, it implements the anomaly detection model. Specifically, the control unit 10 saves or implements the anomaly detection model based on the input content on the model saving screen 7 (see Figure 14). The control unit 10 also saves the preprocessing logic used in segmentation.

[0099] This allows operational staff to use their own knowledge to confirm that there are no problems with both the segmentation results and model behavior before saving the preprocessing logic and anomaly detection model.

[0100] 14, the model saving screen 7 has a save destination setting area 71, a past data processing area 72, a setting saving area 73, and a file reading area 74. The model saving screen 7 allows the user to save preprocessing logic and anomaly detection models by operating buttons.

[0101] The save destination setting area 71 is an area for setting the save destination of the preprocessing logic and the anomaly detection model. The past data processing area 72 is an area for performing preprocessing of past data and anomaly detection processing.

[0102] The setting save area 73 is an area for saving or implementing the preprocessing logic and the anomaly detection model. By checking "Implementation," the preprocessing logic and the anomaly detection model are saved and the anomaly detection model is implemented online at the same time.

[0103] The file reading area 74 is an area for reading, editing, and deleting saved files. It is possible to read, modify, or delete saved preprocessing logic and anomaly detection models.

[0104] Fig. 15 is a flowchart showing an example of a procedure for processing related to the correction of learning data points, and Fig. 16 is a diagram for explaining the correction of learning data points.

[0105] First, the control unit 10 accepts segment selection and parameter setting based on the input contents in the data selection acceptance area 61 and the parameter setting acceptance area 62 of the model evaluation screen 6 (see FIG. 3) (S21, processing as the setting acceptance unit 15).

[0106] Next, the control unit 10 generates an anomaly detection model based on the received selection of the segment and the setting of the parameters, and calculates the degree of anomaly and the judgment boundary (S22, processing as the model generation unit 17).

[0107] Next, the control unit 10 creates an abnormality degree graph 631 based on the calculation results of the abnormality degree and the judgment boundary, and displays it in the graph display area 63 of the model evaluation screen 6 (S23, processing as the distribution display unit 18).

[0108] Next, the control unit 10 determines whether or not a segment or parameter modification has been accepted (S24, processing as the modification accepting unit 16). A segment modification is accepted based on the content entered into the data selection accepting area 61. A parameter modification is accepted based on the content entered into the parameter accepting area 62.

[0109] Without being limited to this, segment correction may be accepted, for example, by directly pointing to a learning data point 632 in an anomaly degree graph 631 displayed in the graph display area 63 with a pointer that is displayed on the screen of the display unit 4 and is linked to the operation of the operation unit 3.

[0110] If a modification of the segment or parameter is accepted (S24: YES), the control unit 10 re-learns the anomaly detection model based on the modified segment or parameter, and re-calculates the degree of anomaly and the judgment boundary (S25, processing as the model generation unit 17).

[0111] Next, the control unit 10 creates the anomaly degree graph 631 again based on the calculation results of the anomaly degree and the judgment boundary, and displays it in the graph display area 63 of the model evaluation screen 6 (S26, processing as the distribution display unit 18). The corrected anomaly degree graph 631 may be displayed alongside the anomaly degree graph 631 before correction.

[0112] Thereafter, the control unit 10 stores the anomaly detection model based on the input content on the model storage screen 7 (see FIG. 14) (S27, processing as the model storage unit 19). This completes the processing related to the correction of the learning data points.

[0113] As shown in FIG. 16, if there are learning data points 632e (a group of points arranged in a straight line in the figure) that the user determines to be unnecessary in the anomaly degree graph 631d before correction, the user can perform an operation to delete the unnecessary learning data points 632e, and a corrected anomaly degree graph 631n will be generated based on the learning data points 632 from which the unnecessary learning data points 632e have been removed.

[0114] Here, the deletion of the learning data point 632e has been given as an example of modifying a segment, but this is not limiting. For example, the learning data point 632 may be added to the anomaly degree graph 631, or the learning data point 632 may be moved with a pointer.

[0115] Fig. 17 is a flow diagram showing an example of the procedure for processing related to the correction of the decision boundary. Fig. 18 is a diagram for explaining the correction of the decision boundary. Steps that overlap with the above example are given the same numbers and detailed explanations are omitted.

[0116] When the control unit 10 displays the abnormality degree graph 631 (S23), it determines whether or not a correction to the determination boundary 633 has been accepted (S34, processing as the correction accepting unit 16).

[0117] Corrections to the judgment boundary 633 are accepted, for example, by directly pointing to the abnormality degree graph 631 displayed on the screen of the display unit 4 and linked to the operation of the operation unit 3 within the graph display area 63.

[0118] Next, the control unit 10 corrects the anomaly degree graph 631 so as to include the corrected decision boundary 633, and displays it in the graph display area 63 (S35, S36, processing as the distribution display unit 18). Thereafter, the anomaly detection model is saved (S27), and the processing related to correcting the decision boundary ends.

[0119] 18, for example, the new decision boundary 633n is applied and the pre-correction decision boundary 633d is deleted when the user performs an operation of drawing a new decision boundary 633n within the anomaly level graph 631 with the pointer 34. The new decision boundary 633n may be set, for example, between the learning data point 632 and the pre-correction decision boundary 633d, or may be set outside the pre-correction decision boundary 633d.

[0120] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and it goes without saying that various modifications can be made by those skilled in the art. [Explanation of symbols]

[0121] 1 Model generation support device, 2 Memory unit, 3 Operation unit, 4 Display unit, 10 Control unit, 11 Data acquisition unit, 12 Condition reception unit, 13 Section extraction unit, 14 Extraction display unit, 15 Setting reception unit, 16 Correction reception unit, 17 Model generation unit, 18 Distribution display unit, 19 Model storage unit, 5 Data preprocessing screen, 51 Data selection reception area, 52 Condition input area, 53 Extraction target display area, 54 Extraction result display area, 6 Model evaluation screen, 61 Data selection reception area, 62 Parameter setting reception area, 63 Graph display area, 631 Anomaly degree graph, 632 Learning data point, 633 Decision boundary, 100 Industrial machine, 102 Sensor

Claims

1. a data acquisition unit that acquires time-series data detected by a sensor installed in the industrial machine; a condition receiving unit that receives, from a user, extraction conditions for extracting section data from the time-series data; a section extraction unit that extracts the section data from the time-series data based on the extraction condition; a model generation unit that generates an anomaly detection model by unsupervised learning based on the section data; a distribution display unit that displays a graph showing a distribution of the degree of anomaly calculated by the anomaly detection model; Equipped with the section extraction unit detects change points in the time-series data, identifies a code representing a trend at each of the change points, and extracts the section data by clustering change patterns of the codes; Model generation support device.

2. further comprising an extraction display unit that displays the section data extracted based on the extraction conditions. The model generation support device according to claim 1 .

3. the extraction display unit further displays a data selection receiving area for receiving selection of time series data from which the section data is to be extracted from the plurality of types of time series data. The model generation support device according to claim 2 .

4. the extraction display unit further displays a condition input receiving area for receiving input of the extraction conditions. The model generation support device according to claim 2 .

5. the condition input receiving area further receives selection of auxiliary data related to control of the industrial machine, which is used for extracting the section data; 5. The model generation support device according to claim 4.

6. the extraction display unit further displays an extraction target display area that displays the time series data, and an extraction result display area that displays the section data extracted from the time series data. The model generation support device according to claim 2 .

7. the extraction result display area further displays auxiliary data related to control of the industrial machine, which was used to extract the section data. The model generation support device according to claim 6.

8. the distribution display unit further displays a data selection receiving area for receiving switching of the display of the section data in the graph. The model generation support device according to claim 1 .

9. the distribution display unit further displays a parameter setting reception area for receiving settings of parameters of the anomaly detection model. The model generation support device according to claim 1 .

10. the distribution display unit displays, as the graph, a contour diagram in which the degree of abnormality is divided into a predetermined number of gradations. The model generation support device according to claim 1 .

11. the distribution display unit displays the section data within the graph; further comprising a correction receiving unit that receives corrections to the section data from a user; the model generation unit re-learns the anomaly detection model based on the corrected section data. The model generation support device according to claim 1 .

12. the model generation unit calculates a decision boundary for anomaly detection; the distribution display unit displays the judgment boundary within the graph; further comprising a correction receiving unit that receives a correction of the decision boundary from a user. The model generation support device according to claim 1 .

13. Acquire time-series data detected by sensors installed on industrial machinery, receiving, from a user, extraction conditions for extracting section data from the time series data; extracting the section data from the time series data based on the extraction conditions; generating an anomaly detection model by unsupervised learning based on the section data; displaying a graph showing the distribution of the degree of anomaly calculated by the anomaly detection model; A model generation support method, The extraction of the section data includes detecting change points in the time-series data, identifying a code representing a trend at each of the change points, and clustering the change patterns of the codes to extract the section data. A method for assisting model generation.

14. Acquiring time series data detected by a sensor installed in the industrial machine; receiving, from a user, extraction conditions for extracting section data from the time series data; extracting the section data from the time series data based on the extraction condition; generating an anomaly detection model by unsupervised learning based on the section data; and displaying a graph showing the distribution of the degree of anomaly calculated by the anomaly detection model; on the computer, The extraction of the section data includes detecting change points in the time-series data, identifying a code representing a trend at each of the change points, and clustering the change patterns of the codes to extract the section data. program.

Images