Cluster-based classification for time series data

Abstract

In order to analyze time series data monitored on a machine, the device (200) divides the time series data into a plurality of time periods (230). In addition, the device (200) determines clusters of time periods estimated to have the same time series data dynamics. The device (200) then classifies the clusters based on tag information associated with at least one of the time periods. If none of the time periods of the cluster has associated tag information, the device (200) presents at least a portion of the time series data of the cluster to the user (10). Based on user input received in response to presenting the time series data, the device (200) classifies the clusters and generates tag information associated with the time periods of the clusters. The generated tag information indicates a result of classifying the clusters.

Description

Technical Field

[0001] The present disclosure relates to apparatus and methods for analyzing time series data. Background Art

[0002] In various fields of technology, the operation of machines or machine systems needs to be monitored. For example, in the field of oil or gas production, a large number of pumps, usually electric submersible pumps (ESP), can be monitored in order to provide preventive maintenance and thereby ensure high availability.

[0003] In order to effectively monitor a large number of machines, it is necessary to utilize monitoring mechanisms that are at least partially based on artificial intelligence (AI). Such AI-based mechanisms can analyze time series data acquired by monitoring multiple machines, automatically identify abnormal behavior, and trigger measures to address the problem causing the abnormal behavior, for example, by issuing warnings to operators. When monitoring pumps, the AI-based mechanism can, for example, automatically classify detected anomalies as, for example, blockages, leaks, shaft damage, etc.

[0004] AI mechanisms can be based on predefined rules, for example, developed by domain experts, and / or based on machine learning. However, developing and maintaining appropriate rules is often a complex task. On the other hand, machine learning methods often correspond to black box methods. That is, the output of the AI ​​mechanism trained by the machine learning process may be difficult to understand and often does not provide further explanations for the reasons for its decisions or classifications. In addition, the machine learning process often requires a considerable amount of manually prepared training data, and preparing such training data can be a tedious process. Furthermore, when machine learning is utilized, the AI ​​mechanism can only be applied after the machine learning process has been completed at least to some extent, so it is often not possible to immediately address new types of behavior.

[0005] WO 2017 / 034512 A1 describes interactive analysis of time series data, involving automatic segmentation and clustering of time series data based on machine learning and user interaction.

[0006] WO 2018 / 140337 A1 describes a semi-supervised method for machine condition monitoring and fault diagnosis, involving partitioning time series data into segment clusters and clustering the segment clusters into cluster prototypes, which are used to learn a discriminant model for predicting class labels. Users can manually provide labels for segment clusters after they are created.

[0007] US2007 / 279490 A1 uses a two-stage training process to mine surveillance videos to detect abnormal events. In the first stage, segments of the video stream are clustered unsupervised, and a set of abnormal events is combined with user feedback to generate a clean training set. In the second group, a clean training set is used to train a more accurate model.

[0008] Therefore, there is a need for techniques that allow time series data acquired by monitoring machines to be efficiently and accurately classified. Summary of the invention

[0009] In view of the above needs, the present disclosure provides an apparatus as described in the present invention and a method as described in the present invention.

[0010] According to one embodiment, a device for analyzing time series data is provided. The device is configured to:

[0011] - Acquire time series data monitored on machines (e.g. pumps, mills, electric motors, internal combustion engines and / or turbines);

[0012] -Split time series data into multiple time periods;

[0013] - determining clusters of time periods estimated to have the same time series data dynamics;

[0014] - checking whether tag information associated with at least one of the time periods is available;

[0015] - if label information associated with at least one of the time periods is available, classifying the clusters based on the label information associated with at least one of the time periods;

[0016] - if none of the time periods of the cluster has associated label information, presenting at least a portion of the time series data of the cluster to the user; and

[0017] - Based on user input received in response to presenting at least a portion of the time series data, classifying the clusters and generating label information associated with time periods of the clusters, the generated label information indicating a result of classifying the clusters.

[0018] In this way, the device can achieve effective classification of time series data by leveraging existing label information for other time series data that show similar dynamics, and can handle new behavior types by invoking user input to classify the time series and generate new label information. Therefore, users, such as domain experts or machine learning experts, can manage the generation of new label information in an intuitive and effective manner, and can effectively learn new behavior types identified in time series data during routine monitoring operations.

[0019] According to one embodiment, the device is configured to determine the time periods of the time series data and the clusters based on a probabilistic model of the time series data dynamics. In this way, the similarity of time series data at different times can be evaluated in an accurate and reproducible manner.

[0020] According to one embodiment, the device is further configured to:

[0021] - If two or more time periods of a cluster have conflicting associated label information, presenting at least a portion of the time series data of the cluster to a user.

[0022] Therefore, conflicting classifications can be resolved in an efficient and intuitive way.

[0023] According to one embodiment, the device is further configured to:

[0024] -If two or more time periods of a cluster have conflicting associated label information, selecting a non-conflicting portion of the label information and classifying the cluster based on the selected non-conflicting portion of the label information. The device may be configured to select the non-conflicting portion of the label information based on a cumulative time length of the time periods associated with the label information. Alternatively or additionally, the device may be configured to select the non-conflicting portion of the label information based on a probabilistic model of the dynamics of the time series data.

[0025] In this way, conflicts in tag information can be automatically resolved, or the device can efficiently help the user resolve conflicts in tag information.

[0026] According to one embodiment, the device is configured to classify clusters by assigning them to one of a plurality of categories, including: normal operation of the machine, abnormal operation of the machine, unknown dynamics, and conflicting label information. However, it should be noted that other categories may also be used, such as different types of abnormal operation or different types of normal operation. These categories may enable effective monitoring of the machine during routine monitoring operations and effective management of learning new behavior types identified in time series data.

[0027] According to a further embodiment, a method for analyzing time series data is provided. The method comprises:

[0028] - Acquire time series data monitored on machines (e.g. pumps, mills, electric motors, internal combustion engines and / or turbines);

[0029] -Split time series data into multiple time periods;

[0030] - determining clusters of time periods estimated to have the same time series data dynamics;

[0031] - checking whether tag information associated with at least one of the time periods is available;

[0032] - if label information associated with at least one of the time periods is available, classifying the clusters based on the label information associated with at least one of the time periods;

[0033] - if none of the time periods of the cluster has associated label information, presenting at least a portion of the time series data of the cluster to the user; and

[0034] - Based on user input received in response to presenting at least a portion of the time series data, classifying the clusters and generating label information associated with time periods of the clusters, the generated label information indicating a result of classifying the clusters.

[0035] In this way, the method can achieve effective classification of time series data by leveraging existing label information for other time series data that show similar dynamics, while handling new behavior types by invoking user input to classify the time series and generate new label information. Therefore, users, such as domain experts or machine learning experts, can manage the generation of new label information in an intuitive and effective manner, and can effectively learn new behavior types identified in time series data during routine monitoring operations.

[0036] According to one embodiment, the method further comprises:

[0037] -Determine time periods and clusters of time series data based on probabilistic models of the time series data dynamics.

[0038] In this way, the method can enable the assessment of the similarity of time series data at different times in a precise and reproducible manner.

[0039] According to one embodiment, the method further comprises:

[0040] - If two or more time periods of a cluster have conflicting associated label information, presenting at least a portion of the time series data of the cluster to a user.

[0041] In this way, the method may enable resolving conflicting classifications in an efficient and intuitive manner.

[0042] According to one embodiment, the method further comprises:

[0043] -If two or more time periods of a cluster have conflicting associated label information, a non-conflicting portion of the label information is selected and the cluster is classified based on the selected non-conflicting portion of the label information. The selection of the non-conflicting portion of the label information can be based on the cumulative time length of the time periods associated with the label information and / or based on a probabilistic model of the time series data dynamics.

[0044] In this way, the conflict of label information can be automatically resolved, and the conflict of label information can also be efficiently helped to solve the problem for users.

[0045] According to one embodiment, the method further comprises:

[0046] -Classify clusters by assigning them to one of several categories: normal operation of the machine, abnormal operation of the machine, unknown dynamics, and conflicting label information.

[0047] However, it should be noted that other categories may also be used, such as different types of abnormal operation or different types of normal operation. These categories can enable effective monitoring of the machine and effective management of learning new behavior types identified in time series data during routine monitoring operations.

[0048] In the above-described embodiments of the device or method, the machine may include a pump. In this case, the time series data may include at least one of the following: motor temperature of the pump, inlet temperature of the pump, operating power of the pump, pressure inside the pump, and pressure outside the pump. Thus, the device or method may be used to implement enhanced monitoring of the pump, for example, through an AI-based mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Hereinafter, embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings.

[0050] Figure 1 Examples of time series data that can be analyzed according to embodiments of the present disclosure are schematically shown.

[0051] Figure 2 A monitoring system according to an embodiment of the present disclosure is schematically shown.

[0052] Figure 3 A flow chart for schematically illustrating a method according to an embodiment of the present disclosure is shown.

[0053] Figure 4 A block diagram for schematically illustrating the structure of a device according to the present disclosure is shown. DETAILED DESCRIPTION

[0054] It should be noted that in the detailed description of the following embodiments, the drawings are only schematic and the illustrated elements are not necessarily shown to scale. On the contrary, the drawings are intended to illustrate the functions and collaboration of the components. Here, it should be understood that any connection or coupling of a function block, device, component or other physical or functional element can also be realized by an indirect connection or coupling, for example, via one or more intermediate elements. The connection or coupling of an element or component can be realized, for example, by a combination based on a wired, wireless connection and / or based on a wired and wireless connection. The function block can be realized by dedicated hardware, by firmware and / or software installed on programmable hardware, and / or by a combination of dedicated hardware and firmware or software.

[0055] Figure 1Examples of time series data are schematically shown. The time series data may be provided, for example, by sensors on an electric submersible pump (ESP). The time series data may, for example, include pressures measured inside the ESP, such as the head pressure of the ESP. Furthermore, the time series data may include pressures measured outside the ESP, such as the annulus pressure or the pressure above ground. Further, the time series data may include the operating power of the ESP, for example, in terms of motor current.

[0056] Figure 1 Further illustrated are exemplary clusters 11, 12 of time series data. Each cluster 11, 12 corresponds to multiple time periods of time series data that are estimated to exhibit similar time series data dynamics. Cluster 11 may, for example, include time periods in which the time series data exhibits dynamics corresponding to normal operation of the ESP, while cluster 12 may include time periods in which the time series data exhibits dynamics corresponding to abnormal operation of the ESP. It should be noted that Figure 1 The clusters 11 and 12 shown are only exemplary, and in actual scenarios, more clusters can be identified in the time series data. Even if two clusters show different time series data dynamics, they may belong to the same category, for example, normal operation or abnormal operation. Typically, a cluster will contain multiple time periods of time series data, which occur at different times.

[0057] Figure 2 Schematically illustrated is a monitoring system equipped with an analysis tool 200 configured to perform classification of time series data according to the illustrated concept.

[0058] As shown, the analysis tool 200 receives time series data from the machine 100. As described above, the machine 100 may be a pump, in particular an ESP. However, it should be understood that the analysis tool 200 may also be used to analyze time series data monitored on other types of machines (e.g., industrial mills, turbines, or internal combustion engines).

[0059] In the example shown, the analysis tool 200 is provided with a memory 210 for storing time series data monitored on the machine. The time series data is then provided to a segmentation stage 220. The segmentation stage 220 segments the time series data into a plurality of time periods in which the time series data exhibits a certain type of dynamics or behavior. The segmentation may be based on fitting one or more probabilistic models to the time series data. The probabilistic model may be based, for example, on a Gaussian mixture pattern. As a result, the segmentation stage 220 outputs a plurality of time periods 230 of the time series data. In addition, the segmentation stage 220 also identifies clusters of the time periods 230. Clusters correspond to groups of time periods that exhibit similar time series data dynamics, such as Figure 1Clusters 11, 12 in the example. For example, clusters may be identified based on time periods matching the same probability model. As a result, the segmentation stage 220 outputs cluster information 240 describing the identified clusters.

[0060] The cluster information 240 is provided to the classification stage 250. The classification stage 250 classifies the clusters into different categories. In the example shown, these categories include "normal operation", "abnormal operation", "unknown behavior" and "conflict". As a result, the classification stage 250 outputs cluster classification information 270.

[0061] The classification is based on label information stored in the memory 260. For a certain time period, the label information may indicate the result of an earlier classification. The earlier classification may be a manual classification based on user input, such as a manual classification performed by a domain expert, or may be an automatic or semi-automatic classification performed by the analysis tool 200 itself or by some other automatic analysis system. However, for some time periods of the cluster, no label information may be available.

[0062] If for a time period of a cluster, label information is available and indicates a certain classification, the classification stage 250 can classify other time periods of the cluster according to the same classification. For example, if for one of the time periods of a cluster, label information indicates a classification in the "normal operation" category, then other time periods of the cluster and even the entire cluster can be classified into the "normal operation" category. Similarly, if for one of the time periods of a cluster, label information indicates a classification in the "abnormal operation" category, then other time periods of the cluster and even the entire cluster can be classified into the "abnormal operation" category.

[0063] On the other hand, if there is no label information available for the time period of the cluster, the classification stage 250 may first classify the cluster in the category “unknown behavior”. At least a portion of the time series data of the cluster may then be presented on the user interface 280, for example, over one or more time periods, to prompt input by the user 10, such as a domain expert. The presentation of the time series data on the user interface 280 may also be supplemented by explanatory information, such as when or how long the corresponding behavior was observed, or reliability information of the measured time series data, for example, in terms of mean and standard deviation.

[0064] Based on the user input, the cluster may then be reclassified into one of the categories “normal operation” and “abnormal operation” and the corresponding label information provided to the memory 260 .

[0065] In some cases, there may also be conflicting label information for the time periods of the cluster. For example, for one or more time periods of the cluster, the label information may indicate an earlier classification into the category "normal operation", while for one or more other time periods of the cluster, the label information may indicate an earlier classification into the category "abnormal operation". In this case, the classification level 250 may first classify the cluster in the category "conflict". At least a portion of the time series data of the cluster may then be presented on the user interface 280 to prompt the user 10 for input. The presentation of the time series data on the user interface 280 may also be supplemented by explanatory information, such as when or how long the corresponding behavior was observed, or reliability information of the measured time series data, for example, in terms of mean and standard deviation.

[0066] Based on the user input, the conflict of label information may be resolved, and then the cluster may be reclassified into one of the categories “normal operation” and “abnormal operation” and the corresponding label information may be provided to the memory 260 .

[0067] In some scenarios, user input provided via user interface 280 may also be used to adjust parameters used in segmentation stage 220, for example, according to the definition of relevant features to be considered in the segmentation process.

[0068] The analysis tool 200 can repeat these processes for newly arrived time series data, thereby continuously accumulating further label information and refining existing label information. The accumulated and refined label information further increases the possibility of being able to automatically classify the newly arrived time series data into the "normal operation" category or the "abnormal operation" category, and avoid prompting further user input.

[0069] Figure 3 A flow chart is shown for illustrating a method that can be used to implement the illustrated concept in a device, such as in the above-mentioned analysis tool 200. The method can be implemented, for example, by executing program code by one or more processors of the device. The program code can be stored in a memory of the device.

[0070] At block 310, the device acquires time series data monitored on a machine such as the above-described machine 100. The machine may correspond to or include at least one of the following: a pump such as an ESP, a mill, an electric motor, an internal combustion engine, or a turbine. If the machine includes a pump, the time series data may include at least one of the following: a motor temperature of the pump, an inlet temperature of the pump, an operating power of the pump, a pressure inside the pump, and a pressure outside the pump.

[0071] At block 320, the device segments the time series data into a plurality of time periods. The time periods of the time series data may be determined based on a probabilistic model of the dynamics of the time series data.

[0072] At block 330 , the device determines clusters of time periods estimated to have the same time series data dynamics. The clusters may be determined based on a probability model of the time series data dynamics, for example, based on the same probability model used to segment the time series data at block 320 .

[0073] The device then classifies the cluster based on the tag information associated with at least one time period of the cluster. This may involve assigning the cluster to one of a plurality of categories, including: normal operation of the machine, abnormal operation of the machine, unknown dynamics, and conflicting tag information.

[0074] At block 340 , the device checks whether tag information is available for at least one time period of the cluster.

[0075] If tag information is found to be available for at least one time period of the cluster at block 340, the device proceeds to block 350 and classifies the cluster based on the tag information associated with the at least one time period. This may involve classifying other time periods of the cluster for which tag information is not yet available into the category indicated by the available tag information.

[0076] If it is found at block 340 that no label information is available for the time period of the cluster, the device proceeds to block 360 and presents at least a portion of the time series data of the cluster to the user. At block 370, the device then classifies the cluster based on the user input received in response to presenting the time series data at block 360. In addition, the device generates label information associated with the time period of the cluster. The generated label information indicates the result of classifying the cluster.

[0077] In some cases, two or more time periods of a cluster may have conflicting associated label information. Also in this case, the device may present at least a portion of the time series data of the cluster to the user and classify the cluster based on user input received in response to presenting the time series data. Alternatively or in addition, the device may select non-conflicting portions of the label information and classify the cluster based on the selected non-conflicting portions of the label information. Selecting the non-conflicting portions of the label information may be based on the cumulative time length of the time periods associated with the label information and / or based on a probability model of the time series data dynamics, for example, the same probability model used to segment the time series data at box 320 or the same probability model used for clusters at box 330.

[0078] Can be repeated Figure 4 The method 400 is used to identify and classify multiple clusters in time series data and newly acquired time series data.

[0079] Figure 4A block diagram is shown for schematically illustrating the structure of the device 400. The device 400 may, for example, be used to implement the above-mentioned analysis tool 200. In the example shown, the device 400 includes at least one processor 410 and a memory 420. Further, the device 400 includes at least one data interface 430, which can be used to obtain the above-mentioned time series data and initial label information. In addition, the device 400 may also include a user interface 440, which is used to present the time series data to the user and receive user input, such as classifying clusters without existing label information, resolving conflicting classifications, or defining label information. The components of the device 400, namely the at least one processor 410, the memory 420, the at least one data interface 440 and the user interface 440, may be coupled to each other, for example, via a data bus 450 or a similar internal communication infrastructure of the device 400. By executing program code by at least one processor 410, the device 400 can be configured to operate according to the above-mentioned functions of the analysis tool 200 and / or implement such as in combination with Figure 3 The method 300 is explained.

[0080] It should be understood that the above description of examples is intended to be illustrative and that the concepts described are susceptible to various modifications. For example, the concepts described can be applied to monitoring various types of machines or machine systems and various types of data that can be monitored on the machines or machine systems, and the classification of time series data can be utilized in various ways.

Claims

1. A device (200, 400) for analyzing time series data, the device (200, 400) being configured to: - Get time series data monitored on the machine; - dividing the time series data into a plurality of time periods (230); - determining clusters of time periods estimated to have the same time series data dynamics (11, 12); providing the cluster information (240) to a classification stage (250) as a first input; - checking whether tag information indicating classification information associated with at least one of the time periods (230) is available; - if label information indicating classification information associated with at least one of the time periods (230) is available, providing said label information as a second input to said classification stage (250) for classifying said clusters (11, 12) based on said label information associated with at least one of said time periods (230); - if none of the time periods of the cluster (11, 12) has associated label information, presenting at least a portion of the time series data of the cluster to a user (10); as well as - Based on user input received in response to presenting at least a portion of the time series data, classifying the clusters (11, 12) and generating label information associated with the time periods of the clusters (11, 12), the generated label information indicating a result of classifying the clusters (11, 12).

2. The device (200, 400) according to claim 1, wherein: The device (200, 400) is configured to determine the time period (230) of the time series data and of the cluster (11, 12) based on a probabilistic model of the dynamics of the time series data.

3. The device (200, 400) according to claim 1 or 2, wherein: The device (200, 400) is further configured to: - if two or more of the time periods of the cluster have conflicting associated label information, presenting at least a portion of the time series data of the cluster to the user (10).

4. The device (200, 400) according to claim 1 or 2, wherein: The device (200, 400) is further configured to: - if two or more of the time segments of the cluster (11, 12) have conflicting associated label information, selecting a non-conflicting portion of the label information and classifying the cluster (11, 12) based on the selected non-conflicting portion of the label information.

5. The device (200, 400) according to claim 4, wherein: The device (200, 400) is configured to select the non-conflicting portion of the tag information based on a cumulative time length of the time periods associated with the tag information.

6. The device (200, 400) according to claim 4, wherein: The apparatus (200, 400) is configured to select the non-conflicting portion of the tag information based on a probabilistic model of the time series data dynamics.

7. The device (200, 400) according to claim 1 or 2, wherein: The device (200, 400) is configured to classify the cluster (11, 12) by assigning the cluster to one of a plurality of categories, the plurality of categories comprising: normal operation of the machine, abnormal operation of the machine, unknown dynamics, and conflicting label information.

8. The device (200, 400) according to claim 1 or 2, wherein: The machine includes at least one of the following: a pump, a mill, an electric motor, an internal combustion engine, and a turbine.

9. The device (200, 400) according to claim 1 or 2, wherein: The machine includes a pump, and the time series data includes at least one of the following: a motor temperature of the pump, an inlet temperature of the pump, an operating power of the pump, a pressure inside the pump, and a pressure outside the pump.

10. A method (300) for analyzing time series data, the method (300) comprising: - Get time series data monitored on the machine; - dividing the time series data into a plurality of time periods (230); - determining clusters of time periods estimated to have the same time series data dynamics (11, 12); - providing the cluster information (240) to the classification stage (250) as a first input; - checking whether tag information indicating classification information associated with at least one of the time periods (230) is available; - if label information indicating classification information associated with at least one of the time periods (230) is available, providing said label information as a second input to said classification stage (250) for classifying said clusters (11, 12) based on said label information associated with at least one of said time periods (230); - if none of the time periods of the cluster (11, 12) has associated label information, presenting at least a portion of the time series data of the cluster to a user (10); as well as - Based on user input received in response to presenting at least a portion of the time series data, classifying the clusters (11, 12) and generating label information associated with the time periods of the clusters (11, 12), the generated label information indicating a result of classifying the clusters (11, 12).

11. The method (300) according to claim 10, further comprising: - determining the time periods (230) of the time series data and of the clusters (11, 12) based on a probabilistic model of the dynamics of the time series data.

12. The method (300) according to claim 10 or 11, further comprising: - if two or more of the time periods of the cluster (11, 12) have conflicting associated label information, presenting at least a portion of the time series data of the cluster to the user (10) and / or selecting a non-conflicting portion of the label information and classifying the cluster (11, 12) based on the selected non-conflicting portion of the label information.

13. The method (300) according to claim 12, comprising: - selecting the non-conflicting portion of the tag information based on the cumulative time length of the time periods associated with the tag information and / or selecting the non-conflicting portion of the tag information based on a probabilistic model of the time series data dynamics.

14. The method (300) according to claim 10 or 11, further comprising: - Classifying the clusters (11, 12) by assigning them to one of a plurality of categories, the plurality of categories comprising: normal operation of the machine, abnormal operation of the machine, unknown dynamics, and conflicting label information.

15. A computer program product comprising a non-transitory computer readable medium storing computer executable code, which, when executed on a processor, performs the method according to any one of claims 10 to 14.

Images