Information processing device, information processing method, and program

The information processing device enhances classification accuracy and anomaly detection by using a supervised shapelet learning method trained on normal data to generate partial waveform patterns during operation, addressing the challenge of limited data availability and hyperparameter adjustment in time-series data classification.

JP2024167998AActive Publication Date: 2024-12-05KK TOSHIBA
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
JP2023084370
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-23
Publication Date
2024-12-05
Estimated Expiration
2043-05-23

AI Technical Summary

Technical Problem

Existing classification technologies for time-series data, such as sensor data, face challenges in clarifying the basis of classification and accurately detecting anomalies when only normal data is available for learning, particularly in infrastructure and manufacturing equipment, where anomaly cases are difficult to collect.

Method used

An information processing device that uses a supervised shapelet learning method to learn shapelets during operation, utilizing a detection model trained only on normal data to detect anomalies and generate partial waveform patterns as evidence, without requiring hyperparameter adjustment, and employs methods like Minirocket and ECOD for unsupervised feature extraction.

Benefits of technology

Enables accurate classification and anomaly detection with higher performance by generating shapelets during operation, improving the presentation of judgment grounds and reducing the need for hyperparameter tuning, even when trained on limited data states.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024167998000001_ABST
    Figure 2024167998000001_ABST
Patent Text Reader

Abstract

To provide an information processing device that performs classification using a model with higher accuracy even when the model is learned using learning data of only a part of a state.SOLUTION: An information processing device includes a processing unit. The processing unit detects whether input waveform data is in a first state using a detection model. The processing unit acquires, when the first state is detected, a plurality of pieces of second state waveform data, which is a second state that is detected in advance using the detection model. The processing unit uses first state waveform data detected as being in the first state and the plurality of pieces of second state waveform data as learning data to learn a classification model that classifies the waveform data as being in either the first state or the second state, and then generates one or more partial waveform patterns that are the basis for indicating that the first state waveform data is in the first state. The processing unit outputs the generated partial waveform patterns.SELECTED DRAWING: Figure 1
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] An embodiment of the present invention relates to an information processing device, an information processing method, and a program. [Background technology]

[0002] In class classification technology for time-series data (time-series waveform data) such as sensor data, it is desirable to clarify the basis for classification in addition to classification performance.As a technology for clarifying the basis for classification, a supervised shapelet learning method has been proposed, which learns shapelets, which are partial waveform patterns effective for classification, together with a classification model (classifier).

[0003] On the other hand, in anomaly detection for infrastructure facilities and manufacturing equipment, it may be difficult to collect anomaly cases (abnormal data) during learning. For this reason, in addition to improving the performance of presenting judgment grounds and anomaly detection performance, learning with only normal cases (normal data) is required. This example is an example of classification into two states (classes), abnormal and normal, but for other examples where classification is into multiple states, it may be necessary to learn with only learning data for some states. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 7166501 [Non-patent literature]

[0005] [Non-Patent Document 1] Isak Karlsson et al., “Generalized random shapelet forests.”, Data Mining and Knowledge Discovery 30, 5 (2016), 1053-1085 Summary of the Invention [Problem to be solved by the invention]

[0006] The present invention aims to provide an information processing device, an information processing method, and a program that can perform classification using a model with higher accuracy even when the model is trained using training data of only some states. [Means for solving the problem]

[0007] The information processing device includes a processing unit. The processing unit detects whether or not input waveform data is in a first state using a detection model. When it is detected that the input waveform data is in the first state, the processing unit acquires a plurality of second state waveform data that is a second state that has been detected in advance using the detection model. The processing unit uses the first state waveform data detected to be in the first state and the plurality of second state waveform data as learning data to learn a classification model that classifies the waveform data as being in either the first state or the second state, thereby generating one or more partial waveform patterns that are grounds for indicating that the first state waveform data is in the first state. The processing unit outputs the generated partial waveform patterns. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 is a block diagram of an information processing apparatus according to a first embodiment. [Diagram 2] 5 is a flowchart of a learning process according to the first embodiment. [Diagram 3] 4 is a flowchart of an estimation process according to the first embodiment. [Figure 4] 11 is a flowchart of a basis output process according to the first embodiment. [Diagram 5] FIG. 11 is a block diagram of an information processing apparatus according to a second embodiment. [Figure 6] 13 is a flowchart of a basis output process according to the second embodiment. [Figure 7] FIG. 13 is a diagram for explaining the necessity of the second embodiment. [Figure 8] FIG. 2 is a hardware configuration diagram of the information processing apparatus according to the embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of an information processing apparatus according to the present invention will be described in detail with reference to the accompanying drawings.

[0010] In the following, a case where the multiple states (classes) to be classified are abnormal and normal will be mainly described, but the states to be classified are not limited to these and may be any states. In addition, in the following, an example will be described in which time-series waveform data such as sensor data is input and the state of the input data is detected, but the data to be processed is not limited to time-series waveform data. In the following, time-series waveform data may be simply referred to as waveform data.

[0011] As described above, a supervised shapelet learning method that learns a classification model and shapelets together has been proposed and is attracting attention in the fields of data mining and machine learning. Such a technique is explainable because it identifies shapelets and presents them to the user, and has high performance in presenting judgment grounds because it can utilize class labels such as normal and abnormal during learning. In addition, the technique is still being actively updated, and techniques that eliminate the need for tuning hyperparameters and techniques that achieve high classification performance have been proposed.

[0012] Even if the same learning data is used, the shapelet used as the basis for judgment may be different each time learning is performed. As a result, the shapelet presented may be different each time even for the same anomaly, which may confuse the user. Therefore, it is desirable to present the same shapelet as much as possible for the same condition (such as an anomaly).

[0013] (First embodiment) The information processing device of the first embodiment is configured to satisfy the following requirements. - Learn without abnormal data (only normal data). -Present the basis for the decision. No need to tune hyperparameters. -High performance in detecting anomalies and in presenting the basis for judgments.

[0014] For example, the information processing device of the embodiment uses a supervised shapelet learning method when presenting a judgment basis using a detection model trained only with normal data. To achieve this, the information processing device learns shapelets not before operation but during operation when anomaly detection is performed. Since there is at least one abnormal data when anomaly detection is performed, the information processing device can present a judgment basis using the supervised shapelet learning method.

[0015] The operational phase is a phase in which estimation (also called detection, inspection, inference, etc.) is performed using a detection model. The pre-operational phase is a phase in which a detection model to be used for estimation is prepared, for example, by executing learning of the detection model. The pre-operational and operational phases are sometimes called the learning phase and estimation (inference) phase, respectively.

[0016] Fig. 1 is a block diagram showing an example of a configuration of an information processing device 100 according to a first embodiment. As shown in Fig. 1, the information processing device 100 includes a receiving unit 101, a learning unit 110, an estimation unit 120, an output control unit 102, a storage unit 131, and a display unit 132.

[0017] The receiving unit 101 receives input of various information used in the information processing device 100. For example, the receiving unit 101 receives designation of learning data used for learning by the learning unit 110, time-series waveform data to be estimated (input waveform data), and various parameters to be used in learning or estimation.

[0018] The parameters are, for example, the following information: The maximum and minimum number of shapelets to present (output) - Maximum number of iterations for generating the shapelet to be presented The number of normal waveform data (an example of second state waveform data) acquired for comparison with input waveform data in which an abnormality was detected (abnormal waveform data, an example of first state waveform data)

[0019] The learning unit 110 learns the detection model. In this embodiment, the learning unit 110 does not apply a supervised shapelet learning method. That is, the learning unit 110 does not specify a shapelet during learning. The detection model may be learned by any method as long as the model can detect abnormalities and normalities in waveform data.

[0020] The learning unit 110 learns a detection model, for example, by using a plurality of normal time-series waveform data (normal time-series data set) collected in advance as learning data. For example, the learning unit 110 generates unsupervised waveform features using Minirocket, and learns a detection model realized by the following technology using the generated waveform features. ·DeepSVDD(Support Vector Data Discription) ·ECOD(Empirical-Cumulative-distribution-based Outlier Detection)

[0021] When the combination of Minirocket and ECOD is used, it is not necessary to adjust the hyperparameters related to the detection model. The learning unit 110 stores the detection model obtained by learning and the feature amount used during learning, for example, in the storage unit 131. The stored detection model and feature amount are used for estimation by the estimation unit 120.

[0022] The estimation unit 120 estimates the state of the time-series waveform data to be estimated by using the learned detection model. The estimation unit 120 includes a detection unit 121, an acquisition unit 122, and a generation unit 123.

[0023] The detection unit 121 uses the detection model to detect whether the input waveform data is abnormal (an example of a first state).

[0024] When an abnormality is detected, the acquiring unit 122 acquires a plurality of normal waveform data that are normal (an example of a second state). The normal waveform data is waveform data that is detected in advance as normal waveform data using a detection model. The acquiring unit 122 acquires the plurality of normal waveform data from, for example, a normal waveform data set collected in advance.

[0025] The normal waveform data set is, for example, a data set including a plurality of waveform data estimated to be normal by a detection model, and is stored in advance in the storage unit 131. The acquisition unit 122 acquires a plurality of normal waveform data from the normal waveform data set stored in the storage unit 131 by at least one of the following methods (M1) to (M4).

[0026] (M1) When the number of cases (number of data) in the normal waveform data set is small, the acquiring section 122 acquires all normal waveform data included in the normal waveform data set.

[0027] (M2) The acquiring unit 122 acquires normal waveform data similar to the abnormal waveform data. Whether or not the data are similar is determined, for example, by whether or not the distance between the data is equal to or smaller than a threshold (distance threshold). The distance is, for example, the Euclidean distance between the input waveform data and the normal waveform data, or the DTW (Dynamic Time Warping) distance.

[0028] (M3) When a detection model that generates features when detecting an abnormality is used, the acquisition unit 122 acquires normal waveform data in which the features used in detection by the detection model are similar to the features of the abnormal waveform data.

[0029] (M4) When a detection model is used that includes multiple partial detection models according to differences in the state of the device, the season, etc., the acquisition unit 122 acquires normal waveform data used for detection by a partial detection model that detected an abnormality among the multiple partial detection models (normal waveform data related to the partial detection model that contributed to the detection of the abnormality).

[0030] An example of the above (M4) will be described. For example, the detection model includes a plurality of partial detection models according to the season. When such a detection model is used, the estimation unit 120 performs estimation on the input waveform data using, for example, a partial detection model according to the season at the time of detecting an abnormality. When the input waveform data is detected to be normal, the input waveform data is added to the normal waveform data set in the storage unit 131 as normal waveform data. At this time, the normal waveform data is stored in association with, for example, information indicating the season at the time of the detection or the partial detection model used for the detection. By referring to this information, the acquisition unit 122 can identify normal waveform data detected to be normal by the same partial detection model as the partial detection model used for detecting the abnormality.

[0031] When the receiving unit 101 receives a designation of the number of normal waveform data to be acquired, the acquiring unit 122 may acquire the designated number of second state waveform data.

[0032] When an abnormality in the input waveform data is detected by the detection unit 121, the generation unit 123 generates one or more shapelets using the input waveform data and a plurality of normal waveform data acquired by the acquisition unit 122. For example, the generation unit 123 generates shapelets by learning a classification model using the abnormal waveform data (first state waveform data) and the plurality of acquired normal waveform data as learning data.

[0033] The classification model is a model different from the detection model, and is a model that classifies input waveform data as either abnormal or normal. The one or more shapelets that are generated correspond to partial waveform patterns that are grounds for indicating that the input waveform data is abnormal, among a plurality of partial waveform patterns included in the input waveform data.

[0034] The shapelet generation process by the generation unit 123 can be realized by a procedure similar to that of the conventional supervised shapelet learning method. The process by the generation unit 123 will be described in detail later.

[0035] The output control unit 102 controls the output of various information used in the information processing device 100. For example, the output control unit 102 outputs (displays) on the display unit 132 information indicating one or more shapelets generated by the generation unit 123.

[0036] At least a part of each of the above units (reception unit 101, learning unit 110, estimation unit 120, and output control unit 102) may be realized by one processing unit. Each of the above units is realized by, for example, one or more processors. For example, each of the above units may be realized by having a processor such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit) execute a program, that is, by software. Each of the above units may be realized by a processor such as a dedicated IC (Integrated Circuit), that is, by hardware. Each of the above units may be realized by using a combination of software and hardware. When multiple processors are used, each processor may realize one of the units, or two or more of the units.

[0037] The storage unit 131 stores various information used in the information processing device. For example, the storage unit 131 stores a normal waveform data set, information indicating a learned detection model, and the like.

[0038] The storage unit 131 can be configured from any commonly used storage medium, such as a flash memory, a memory card, a RAM (Random Access Memory), a HDD (Hard Disk Drive), or an optical disk.

[0039] The display unit 132 is an example of a device that displays various information used in the information processing device 100. The display unit 132 is realized by a display device such as a liquid crystal display.

[0040] The information processing device 100 may be physically configured by one device, or may be physically configured by multiple devices. For example, the information processing device 100 may be constructed on a cloud environment. Furthermore, each unit in the information processing device 100 may be provided in a distributed manner in multiple devices. For example, the information processing device 100 (information processing system) may be configured to include a device (e.g., a learning device) having a function required for learning (e.g., the learning unit 110) and a device (e.g., an estimation device) having a function required for estimation (e.g., the estimation unit 120).

[0041] Next, a learning process by the information processing device 100 of the first embodiment will be described. The learning process is a process for learning a detection model before operation. Fig. 2 is a flowchart showing an example of the learning process in the first embodiment.

[0042] The learning unit 110 learns a detection model using a normal waveform data set collected in advance as learning data for learning the detection model (step S101). The learning unit 110 stores information indicating the learned detection model and feature amounts calculated during learning for the normal waveform data set in the storage unit 131 (step S102), and ends the learning process.

[0043] Next, an estimation process by the information processing device 100 of the first embodiment will be described. The estimation process is a process for estimating a state of waveform data (input waveform data) input as an estimation target by using a learned detection model. Fig. 3 is a flowchart showing an example of the estimation process in the first embodiment.

[0044] The detection unit 121 detects an abnormality in the input waveform data using the learned detection model (step S201). The estimation unit 120 determines whether or not an abnormality has been detected (step S202). If an abnormality has not been detected (step S202: No), the estimation process ends. The output control unit 102 may output information indicating that an abnormality has not been detected.

[0045] If an abnormality is detected (step S202: Yes), the estimation unit 120 judges whether or not to output the basis for the abnormality (step S203). For example, as a simple judgment method, the estimation unit 120 judges that the basis is to be output each time an abnormality is detected. As another judgment method, a method can be applied in which abnormal waveform data is accumulated for a certain period of time, and when the certain period has elapsed, the estimation unit 120 judges that the basis for the abnormality is to be output for one or more pieces of accumulated input waveform data. This makes it possible to collectively output the basis when, for example, abnormalities are detected consecutively within a short period of time.

[0046] If the basis is not to be output (step S203: No), the estimation process ends. If the basis is to be output (step S203: Yes), the estimation unit 120 executes a basis output process (step S204). After the basis output process, the estimation process ends.

[0047] As described above, in the evidence output process, a shapelet serving as evidence is generated and output using the supervised shapelet learning method. In this manner, whereas the supervised shapelet learning method is conventionally executed before operation (learning phase), in this embodiment, a shapelet is learned using the supervised shapelet learning method during operation (estimation phase) and output as evidence.

[0048] FIG. 4 is a flowchart showing an example of the basis output process in step S204.

[0049] The acquiring unit 122 acquires a plurality of normal waveform data from a normal waveform data set collected in advance (step S301). For example, the acquiring unit 122 acquires a plurality of normal waveform data similar to the current abnormal waveform data from the normal waveform data set.

[0050] The generating unit 123 sets the acquired multiple normal waveform data and abnormal waveform data, which is input waveform data in which an abnormality was detected, as learning data (step S302). If the reason for an abnormality is output every time an abnormality is detected, the number of abnormal waveform data is one. If the reason for an abnormality is output at regular intervals, the number of abnormal waveform data can be one or more. Steps S303 to S308 are repeated.

[0051] The generator 123 selects the length L of the shapelet, selects a partial waveform of length L from the abnormal waveform data, and adds it to the shapelet set as shapelet s1 (step S303).

[0052] The shapelet set is a set for storing candidates for shapelets to be output as evidence. At the end of the evidence output process, the shapelets stored in the shapelet set are output as shapelets that indicate evidence of anomaly.

[0053] In step S303, a shapelet s1 is selected from the abnormal waveform data. This makes it possible to obtain a shapelet specific to the abnormal waveform data. As a simple method, the generator 123 first randomly selects a shapelet with length L, and then randomly selects a partial waveform (segment) with length L from the abnormal waveform data to set it as shapelet s1.

[0054] When two or more pieces of abnormal waveform data are included in the training data, the generating unit 123 randomly selects one piece of abnormal waveform data from, for example, the two or more pieces of abnormal waveform data, and selects a shapelet s1 from the selected abnormal waveform data.

[0055] The generation unit 123 calculates an evaluation index o1 of the classification model using the shapelet s1 (step S304). For example, the generation unit 123 calculates features of the learning data using each shapelet included in the shapelet set to which the shapelet s1 has been added, and calculates an evaluation index o1 indicating classification performance when the calculated features are input to the classification model.

[0056] The feature amount can be calculated, for example, by the distance, similarity, or dissimilarity between the shapelet and each of the multiple waveform data included in the training data. As the distance, for example, the Euclidean distance at the point where the shapelet and the waveform data are most similar can be used.

[0057] The classification model is a model that inputs multiple feature amounts calculated for each shapelet included in the shapelet set and outputs whether each waveform data is abnormal or normal. Since the training data includes abnormal waveform data and normal waveform data, it corresponds to teacher data from which the correct answer of whether something is abnormal or normal is obtained. The classification model is trained so that the classification loss for such teacher data is small. The classification loss can be interpreted as the value of an error function or a loss function.

[0058] The classification model may be, for example, a logistic regression model, in which case cross-entropy loss may be used as the classification loss.

[0059] The generating unit 123 can use the following index as the evaluation index o1. Akaike information criterion (AIC), Bayesian information criterion (BIC) and Minimum Description Length (MDL) based on classification loss and classification model complexity. Classification loss on training data. Classification loss in validation data. For example, when there are multiple pieces of abnormal waveform data, validation data is selected from a portion of the training data set in step S302. The remaining training data is used in training in step S303 and thereafter.

[0060] The generation unit 123 may determine selection parameters used for selecting the shapelet s1 based on selection information obtained at the time of detection using the detection model, and select the shapelet s1 in accordance with the determined selection parameters.

[0061] The selection parameter is, for example, at least one of the length L of the shapelet to be selected and the range within which the shapelet is selected. The selection information is, for example, at least one of the length of the waveform data used when detecting that the input waveform data is abnormal (the waveform data that contributed to the abnormality detection) and the range including the partial waveform pattern detected to be abnormal among the partial waveform patterns included in the input waveform data.

[0062] An example of a detection model that outputs such selection information will be described. For example, the detection model may be an ensemble model that integrates multiple partial detection models that use different lengths of waveform data to detect anomalies. The ensemble model, for example, detects anomalies using each of the multiple partial detection models, integrates the multiple detection results, and outputs a final result indicating the presence or absence of anomalies. At this time, the ensemble model outputs, as selection information, information indicating the partial detection model that contributed to the detection of the anomaly, or the length of the waveform data used by the partial detection model that contributed to the detection of the anomaly.

[0063] When the detection model is a model that outputs selection information, the generation unit 123 can select a selection parameter according to the selection information output from the detection model and use it to select the shapelet s1. That is, while the detection model and the classification model have been used independently in the past, in this embodiment, the learning method of the classification model can be changed in consideration of the output by the detection model. This can further improve the presentation performance of the judgment basis, for example.

[0064] Next, the generator 123 calculates an evaluation index o2 of the classification model using shapelets included in the shapelet set assuming that the shapelet s2 that does not contribute to class classification is excluded from the current shapelet set (step S305). The evaluation index o2 can be calculated in the same manner as the above evaluation index o1.

[0065] For example, when a linear classification model such as logistic regression is used, the generation unit 123 selects a shapelet corresponding to a feature amount having a minimum absolute value of classification weight as a shapelet s2 that does not contribute to class classification. The generation unit 123 calculates an evaluation index o2 using a shapelet other than the selected shapelet s2 included in the shapelet set.

[0066] The generating unit 123 determines whether or not to exclude the shapelet s2 (step S306). As a simple method, the generating unit 123 compares the evaluation index o1 with the evaluation index o2, and determines to exclude the shapelet s2 if the evaluation index o2 (corresponding to the evaluation index when it is assumed that the shapelet s2 is excluded from the shapelet set) is a value indicating a better evaluation. If the maximum number of shapelets is specified as a parameter, the generating unit 123 may determine to exclude the shapelet s2 if the number of shapelets included in the shapelet set exceeds the maximum number.

[0067] In the first iteration of the repetitive process (steps S303 to S308), only one shapelet s1 is included in the shapelet set. Therefore, the process of excluding shapelet s2 from the shapelet set may be skipped. Even if it is not skipped, if it is assumed that shapelet s1 has been excluded, no shapelet is included in the shapelet set, so the value of evaluation index o2 does not indicate a better evaluation than evaluation index o1, and it is determined that the shapelet is not to be excluded.

[0068] If it is determined that the shapelet s2 is to be excluded (step S306: Yes), the generation unit 123 excludes the shapelet s2 from the shapelet set (step S307). When the shapelet s2 is excluded, the generation unit 123 updates the value of the evaluation index o1 to the value of the evaluation index o2.

[0069] After shapelet s2 is excluded, or if it is not determined that shapelet s2 is to be excluded (step S306: No), generating unit 123 determines whether or not to end the generation of shapelets (step S308). For example, generating unit 123 determines to end the generation of shapelets when the number of times the repetitive processing (steps S303 to S308) has been performed reaches an upper limit.

[0070] If it is determined that the generation of shapelets is not to be ended (step S308: No), the process returns to step S303, where the next shapelet is added and the process is repeated.

[0071] When it is determined that the generation of shapelets is to be ended (step S308: Yes), the output control unit 102 outputs a shapelet included in the shapelet set as a shapelet that is the basis for anomaly detection (step S309).

[0072] Through this processing, the generation unit 123 calculates features of the training data using the selected shapelets, calculates an evaluation index when the calculated features are input into the classification model, and generates one or more shapelets whose evaluation index is higher than other shapelets as shapelets that serve as the basis for anomaly detection.

[0073] As described above, the information processing device of the first embodiment can detect anomalies using a detection model trained only on normal data, and when an anomaly is detected (during operation), can learn and output a shapelet that is the basis for anomaly detection. This makes it possible to perform classification using the model with higher accuracy even when the model is trained only on training data in a partial state. Furthermore, in this embodiment, it can be configured to use a model that does not require adjustment of hyperparameters.

[0074] Second embodiment When anomaly detection and reason presentation are repeated for a plurality of input waveform data, the information processing device of the second embodiment outputs (presents) shapelets that have a small difference in determination performance and are similar to shapelets that have already been presented.

[0075] Fig. 5 is a block diagram showing an example of the configuration of an information processing device 100-2 according to the second embodiment. As shown in Fig. 5, the information processing device 100-2 includes a reception unit 101, a learning unit 110, an estimation unit 120-2, an output control unit 102, a storage unit 131, and a display unit 132.

[0076] In the second embodiment, the function of the generation unit 123-2 in the estimation unit 120-2 is different from that in the first embodiment. Other configurations and functions are the same as those in FIG. 1, which is a block diagram of the information processing device 100 in the first embodiment, so the same reference numerals are used and the description here is omitted.

[0077] The generating unit 123-2 is added with the following functions when generating shapelets for multiple input waveform data input as estimation targets. It is assumed that the generating unit 123-2 generates a shapelet sA (first partial waveform pattern) for input waveform data DA (first input waveform data). The generating unit 123-2 generates a shapelet sB (second partial waveform pattern) for input waveform data DB (second input waveform data) input thereafter, excluding shapelets that are not similar to the shapelet sA and whose difference in classification performance (evaluation index) with respect to the classification model learned with the shapelet sA is not equal to or greater than a threshold.

[0078] In this embodiment, the detection unit 121 detects whether or not the input waveform data is abnormal by using a detection model for each of a plurality of input waveform data. The reception unit 101 may be configured to receive a designation of a threshold value (an example of a parameter) that the generation unit 123-2 compares with the difference in classification performance.

[0079] The learning process of this embodiment is similar to that of the first embodiment shown in Fig. 2, and therefore will not be described. The overall flow of the estimation process of this embodiment is similar to that of the first embodiment shown in Fig. 3. In this embodiment, the details of the basis output process (step S204) included in the estimation process are different from those of the first embodiment.

[0080] The basis output process of this embodiment will be described below with reference to Fig. 6. Fig. 6 is a flowchart showing an example of the basis output process in the second embodiment. Note that Fig. 6 is executed every time one piece of input waveform data is input.

[0081] Steps S401 to S403 are similar to steps S301 to S303 of the evidence output process in the first embodiment, and therefore a description thereof will be omitted.

[0082] In this embodiment, the generating unit 123-2 determines whether or not the shapelet has been output (step S404). For example, when the processing is for input waveform data inputted for the second or subsequent time among a plurality of input waveform data, and the shapelet output processing (step S415, described later) has been executed for input waveform data inputted prior to the input waveform data, the generating unit 123-2 determines that the shapelet has been output.

[0083] If the shapelet has not been output (step S404: No), the same processing as in the first embodiment is executed. That is, the generation unit 123-2 executes steps S410 to S415, which correspond to steps S304 to S309 in the first embodiment. Steps S410 to S415 are the same processing as steps S304 to S309 in the first embodiment, and therefore the description thereof will be omitted.

[0084] If the shapelet has been output (step S404: Yes), the generation unit 123-2 acquires the output shapelet set and the evaluation index o1 calculated for the shapelet set (step S405). The generation unit 123-2 determines whether the shapelet included in the acquired shapelet set is similar to the currently selected shapelet s1 (step S406).

[0085] For example, the generation unit 123-2 calculates the Euclidean distance between each of one or more shapelets included in the shapelet set and the shapelet s1. If the calculated Euclidean distance includes a Euclidean distance equal to or smaller than a threshold, the generation unit 123-2 determines that the shapelet included in the shapelet set is similar to the shapelet s1.

[0086] If it is determined that they are similar (step S406: Yes), the generating unit 123- executes steps S410 to S415 which are the same processes as steps S304 to S309 in the first embodiment.

[0087] If it is determined that they are not similar (step S406: No), the generating unit 123-2 calculates an evaluation index o2 of the classification model using the added shapelet s1 (step S407). For example, the generating unit 123 calculates the feature amount of the training data using the shapelet s1, and calculates the evaluation index o2 indicating the classification performance when the calculated feature amount is input to the classification model.

[0088] The generating unit 123-2 compares the evaluation index o1 calculated for the output shapelet set and determines whether the evaluation index o2 calculated using the current shapelet s1 has been improved (step S408). For example, the generating unit 123-2 calculates the difference (difference) between the evaluation index o1 calculated for the output shapelet set (corresponding to shapelet sA) and the evaluation index o2 calculated using the shapelet s1, and determines that the evaluation index o2 has been improved if the difference is equal to or greater than a threshold value.

[0089] If the evaluation index o2 has been improved (step S408: Yes), the generating unit 123-2 executes steps S410 to S415, which are the same processes as steps S304 to S309 in the first embodiment.

[0090] If the evaluation index o2 is not improved (step S408: No), the generating unit 123-2 deletes the shapelet s1 from the shapelet set (step S409), and determines whether or not to end the generation of shapelets (step S414).

[0091] FIG. 7 is a diagram for explaining the necessity of this embodiment. The waveform data shown in the two graphs on the left side of FIG. 7 is an example of normal waveform data. The waveform data shown in the two graphs on the right side of FIG. 7 is an example of abnormal waveform data. The example shown in FIG. 7 is an example in which it is possible to determine whether normal waveform data and abnormal waveform data are abnormal or not using two shapelets 701 (a shapelet that points upward) and 702 (a shapelet that points downward). In other words, whichever of shapelets 701 and 702 is used, normal and abnormal are correctly classified.

[0092] If the processing of this embodiment is not performed, even if input waveform data of a similar shape is input, a different shapelet may be output as the basis for anomaly detection each time a judgment is made.

[0093] In this embodiment, in order to avoid the occurrence of such a situation, if a current abnormality can be detected by a shapelet similar to a shapelet previously presented as a basis for judgment, a shapelet having a shape similar to the previously presented shapelet is presented as the basis for judgment this time. In the example of Fig. 7, for example, if shapelet 701 was output in the past, shapelet 701 is also output as the basis for judgment for the current input waveform data.

[0094] In this way, in the second embodiment, it is possible to output a shapelet that has a small difference in judgment performance and is similar to a shapelet that has already been presented. This makes it possible to avoid, for example, a situation in which different shapelets are presented each time for the same anomaly, and to output the same shapelet as much as possible for the same anomaly.

[0095] As described above, according to the first and second embodiments, even when a model is trained using training data of only a portion of the states, classification using the model can be performed with higher accuracy.

[0096] Next, a hardware configuration of the information processing device according to the first or second embodiment will be described with reference to Fig. 8. Fig. 8 is an explanatory diagram showing an example of the hardware configuration of the information processing device according to the first or second embodiment.

[0097] The information processing device of the first or second embodiment includes a control device such as a CPU 51, a storage device such as a ROM (Read Only Memory) 52 or a RAM 53, a communication I / F 54 that connects to a network and communicates, and a bus 61 that connects each part.

[0098] The programs executed by the information processing device of the first or second embodiment are provided in a state that they are pre-installed in the ROM 52 or the like.

[0099] The program executed by the information processing device of the first or second embodiment may be configured to be provided as a computer program product by being recorded in an installable or executable format on a computer-readable recording medium such as a CD-ROM (Compact Disk Read Only Memory), a flexible disk (FD), a CD-R (Compact Disk Recordable), or a DVD (Digital Versatile Disk).

[0100] Furthermore, the program executed by the information processing device of the first or second embodiment may be configured to be stored on a computer connected to a network such as the Internet and provided by being downloaded via the network. Also, the program executed by the information processing device of the first or second embodiment may be configured to be provided or distributed via a network such as the Internet.

[0101] The program executed by the information processing device of the first or second embodiment can cause the computer to function as each part of the information processing device described above. In this computer, the CPU 51 can read the program from a computer-readable storage medium onto the main storage device and execute the program.

[0102] A configuration example of the embodiment will be described below. (Configuration example 1) Detecting whether the input waveform data is in a first state using the detection model; When the first state is detected, a plurality of second state waveform data are acquired, the second state being a second state different from the first state, which has been detected in advance using the detection model; using first state waveform data, which is the input waveform data detected to be in the first state, and a plurality of pieces of second state waveform data as learning data, to learn a classification model that classifies the waveform data as being in the first state or the second state, thereby generating one or more partial waveform patterns, among a plurality of partial waveform patterns included in the first state waveform data, that are grounds for indicating that the first state waveform data is in the first state; outputting the generated partial waveform pattern; Processing section An information processing device comprising: (Configuration example 2) the processing unit calculates a feature amount of the learning data by using the partial waveform patterns, calculates an evaluation index indicating a classification performance when the calculated feature amount is input to the classification model, and generates one or more of the partial waveform patterns having the evaluation index larger than other partial waveform patterns. The information processing device according to configuration example 1. (Configuration example 3) The processing unit includes: determining a selection parameter, which is at least one of a length of the partial waveform pattern to be selected and a range within which the partial waveform pattern is selected, based on selection information obtained at the time of detection by the detection model; selecting the partial waveform pattern from the first state waveform data in accordance with the determined selection parameter, and calculating the feature amount using the selected partial waveform pattern; The information processing device according to configuration example 2. (Configuration Example 4) The selection information is at least one of a length of the waveform data used when detecting that the input waveform data is in the first state and a range including the partial waveform pattern detected to be in the first state among the partial waveform patterns included in the input waveform data. The information processing device according to configuration example 3. (Configuration Example 5) The evaluation index is: A classification loss when classifying the training data using the classification model; Akaike Information Criterion, Bayes Information Criterion, and Minimum Description Length based on the classification loss and the complexity of the classification model; or The classification loss when classifying validation data using the classification model. The information processing device according to any one of configuration examples 2 to 4. (Configuration Example 6) the feature amount is a distance between the partial waveform pattern and the learning data, a similarity between the partial waveform pattern and the learning data, or a dissimilarity between the partial waveform pattern and the learning data; 6. The information processing device according to any one of configuration examples 2 to 5. (Configuration Example 7) The processing unit includes: Detecting whether the input waveform data is in the first state for a plurality of the input waveform data; generating a first partial waveform pattern for first input waveform data included in the plurality of input waveform data, and then generating a second partial waveform pattern as the partial waveform pattern for second input waveform data included in the plurality of input waveform data, excluding a partial waveform pattern that is not similar to the first partial waveform pattern and has a difference in classification performance with respect to the classification model learned with the first partial waveform pattern that is not equal to or greater than a threshold value; The information processing device according to any one of configuration examples 2 to 6. (Configuration Example 8) The processing unit includes: Accepting the designation of the threshold value; generating the second partial waveform pattern using the specified threshold value; The information processing device according to configuration example 7. (Configuration Example 9) The processing unit includes: Among the plurality of second state waveform data, The second state waveform data is similar to the input waveform data detected as being in the first state; and and the second state waveform data, the feature amount used in detection by the detection model being similar to the feature amount used in detection by the detection model for the input waveform data detected to be in the first state. The information processing device according to any one of configuration examples 1 to 8. (Configuration Example 10) The sensing model includes a plurality of partial sensing models; The processing unit acquires the second state waveform data used for detection by the partial detection model that detected the first state among the plurality of partial detection models. The information processing device according to any one of configuration examples 1 to 9. (Configuration Example 11) The processing unit includes: Accepting a designation of the number of the plurality of second state waveform data to be acquired; acquiring the number of second state waveform data; The information processing device according to any one of configuration examples 1 to 10. (Configuration Example 12) The processing unit includes: outputting the generated partial waveform pattern; The information processing device according to any one of configuration examples 1 to 11. (Configuration Example 13) The processing unit includes: a detection unit that detects whether the input waveform data is in the first state; An acquisition unit that acquires a plurality of second state waveform data; A generator for generating one or more of the partial waveform patterns; Equipped with 13. The information processing device according to any one of configuration examples 1 to 12. (Configuration Example 14) An information processing method executed by an information processing device, a detection step of detecting whether the input waveform data is in a first state using the detection model; an acquiring step of acquiring a plurality of second state waveform data, which is a second state different from the first state and which has been detected in advance using the detection model when the first state is detected; a generation step of generating one or more partial waveform patterns that are grounds for indicating that the first state waveform data is in the first state, among a plurality of partial waveform patterns included in the first state waveform data, by learning a classification model that classifies waveform data as being in either the first state or the second state using first state waveform data, which is the input waveform data detected to be in the first state, and a plurality of pieces of the second state waveform data, as learning data; an output control step of outputting the generated partial waveform pattern; An information processing method comprising: (Configuration Example 15) On the computer, a detection step of detecting whether the input waveform data is in a first state using the detection model; an acquiring step of acquiring a plurality of second state waveform data, which is a second state different from the first state and which has been detected in advance using the detection model when the first state is detected; a generation step of generating one or more partial waveform patterns that are grounds for indicating that the first state waveform data is in the first state, among a plurality of partial waveform patterns included in the first state waveform data, by learning a classification model that classifies waveform data as being in either the first state or the second state using first state waveform data, which is the input waveform data detected to be in the first state, and a plurality of pieces of the second state waveform data, as learning data; an output control step of outputting the generated partial waveform pattern; A program for executing.

[0103] Although some embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope and spirit of the invention, and are included in the scope of the invention and its equivalents described in the claims. [Explanation of symbols]

[0104] 100, 100-2 Information processing device 101 Reception 102 Output control section 110 Learning Department 120, 120-2 Estimation part 121 Detection unit 122 Acquisition Department 123, 123-2 generation part 131 Storage section 132 Display section

Claims

1. Detecting whether the input waveform data is in a first state using the detection model; When the first state is detected, a plurality of second state waveform data are acquired, the second state being a second state different from the first state, which has been detected in advance using the detection model; using, as learning data, first state waveform data, which is the input waveform data detected to be in the first state, and a plurality of pieces of the second state waveform data, to learn a classification model that classifies the waveform data as being in the first state or the second state, thereby generating one or more partial waveform patterns, among a plurality of partial waveform patterns included in the first state waveform data, that are grounds for indicating that the first state waveform data is in the first state; Processing section An information processing device comprising:

2. the processing unit calculates a feature amount of the learning data by using the partial waveform patterns, calculates an evaluation index indicating a classification performance when the calculated feature amount is input to the classification model, and generates one or more of the partial waveform patterns having the evaluation index larger than other partial waveform patterns. The information processing device according to claim 1 .

3. The processing unit includes: determining a selection parameter, which is at least one of a length of the partial waveform pattern to be selected and a range within which the partial waveform pattern is selected, based on selection information obtained at the time of detection by the detection model; selecting the partial waveform pattern from the first state waveform data in accordance with the determined selection parameter, and calculating the feature quantity using the selected partial waveform pattern; The information processing device according to claim 2 .

4. The selection information is at least one of a length of the waveform data used when it is detected that the input waveform data is in the first state, and a range including the partial waveform pattern detected to be in the first state among the partial waveform patterns included in the input waveform data. The information processing device according to claim 3 .

5. The evaluation index is: A classification loss when classifying the training data using the classification model; Akaike Information Criterion, Bayes Information Criterion, and Minimum Description Length based on the classification loss and the complexity of the classification model; or The classification loss when classifying validation data using the classification model. The information processing device according to claim 2 .

6. the feature amount is a distance between the partial waveform pattern and the learning data, a similarity between the partial waveform pattern and the learning data, or a dissimilarity between the partial waveform pattern and the learning data; The information processing device according to claim 2 .

7. The processing unit includes: Detecting whether the input waveform data is in the first state for a plurality of the input waveform data; a first partial waveform pattern is generated for a first input waveform data included in the plurality of input waveform data, and then a second partial waveform pattern is generated as the partial waveform pattern for a second input waveform data included in the plurality of input waveform data, excluding a partial waveform pattern that is not similar to the first partial waveform pattern and has a difference in classification performance with respect to the classification model learned with the first partial waveform pattern that is not equal to or greater than a threshold value; The information processing device according to claim 2 .

8. The processing unit includes: Accepting the designation of the threshold value; generating the second partial waveform pattern using the specified threshold value; The information processing device according to claim 7.

9. The processing unit includes: Among the plurality of second state waveform data, the second state waveform data being similar to the input waveform data detected as being in the first state; and and the second state waveform data, the feature amount used in detection by the detection model being similar to the feature amount used in detection by the detection model for the input waveform data detected to be in the first state. The information processing device according to claim 1 .

10. The sensing model includes a plurality of partial sensing models; The processing unit acquires the second state waveform data used for detection by the partial detection model that detected the first state among the plurality of partial detection models. The information processing device according to claim 1 .

11. The processing unit includes: Accepting a designation of the number of the plurality of second state waveform data to be acquired; acquiring the number of second state waveform data; The information processing device according to claim 1 .

12. The processing unit includes: outputting the generated partial waveform pattern; The information processing device according to claim 1 .

13. The processing unit includes: a detection unit that detects whether the input waveform data is in the first state; An acquisition unit that acquires a plurality of the second state waveform data; A generator for generating one or more of the partial waveform patterns; Equipped with The information processing device according to claim 1 .

14. An information processing method executed by an information processing device, a detection step of detecting whether the input waveform data is in a first state using the detection model; an acquiring step of acquiring a plurality of second state waveform data, which is a second state different from the first state and which has been detected in advance using the detection model when the first state is detected; a generation step of generating one or more partial waveform patterns that are grounds for indicating that the first state waveform data is in the first state, among a plurality of partial waveform patterns included in the first state waveform data, by learning a classification model that classifies waveform data as being in either the first state or the second state using first state waveform data, which is the input waveform data detected to be in the first state, and a plurality of pieces of the second state waveform data, as learning data; an output control step of outputting the generated partial waveform pattern; An information processing method comprising:

15. On the computer, a detection step of detecting whether the input waveform data is in a first state using the detection model; an acquiring step of acquiring a plurality of second state waveform data, which is a second state different from the first state and which has been detected in advance using the detection model when the first state is detected; a generation step of generating one or more partial waveform patterns that are grounds for indicating that the first state waveform data is in the first state, among a plurality of partial waveform patterns included in the first state waveform data, by learning a classification model that classifies waveform data as being in either the first state or the second state using first state waveform data, which is the input waveform data detected to be in the first state, and a plurality of pieces of the second state waveform data, as learning data; an output control step of outputting the generated partial waveform pattern; A program for executing the above.

Citation Information

Patent Citations

  • Abnormality detection system, and abnormality detection method

    JP2017009548A

  • Diagnosis device, diagnosis method, and computer program

    JP2019016209A

  • Time series data analysis method, time-series data analyzer and computer program

    JP2020170214A

  • Information processing apparatus, information processing method and program

    JP2021149652A

  • Machine learning program, machine learning method, and machine learning device

    WO2023053216A1