Equipment malfunction detection system and equipment malfunction detection method

Abstract

We provide an equipment malfunction detection system that can detect equipment abnormalities more accurately than conventional systems. [Solution] The equipment anomaly detection system 10 includes a data acquisition unit 51 that repeatedly acquires equipment data indicating the operating status of the equipment 40, a selection setting unit 52 that receives a specification from the user 31 regarding the range of each of the multiple items indicating the features of the equipment data (i.e., waveform selection setting 57c), a learning unit 54 that trains a machine learning model 57d using equipment data acquired by the data acquisition unit 51 that has features that satisfy the waveform selection setting 57c received by the selection setting unit 52, an evaluation unit 55 that evaluates anomalies of the equipment 40 by inputting equipment data having features that satisfy the waveform selection setting 57c into the trained machine learning model 57d, and an output unit 56 that outputs data for displaying the results of the evaluation.

Description

Technical Field

[0001] The present disclosure relates to an equipment abnormality detection system and an equipment abnormality detection method, and particularly to an equipment abnormality detection system that detects equipment abnormalities using a machine learning model, etc.

Background Art

[0002] Conventionally, there is known an equipment abnormality detection system that detects equipment abnormalities using a machine learning model as AI (Artificial Intelligence) based on equipment data such as measurement data obtained from equipment (see, for example, Patent Document 1).

[0003] In a conventional equipment abnormality detection system such as the equipment abnormality detection system of Patent Document 1 (referred to as a "trend calculation system" in Patent Document 1), waveform data measured by an edge device is accumulated in a cloud server as equipment data, and AI is learned using the equipment data to monitor equipment abnormalities. The user can view the equipment data on the detailed display screen of the web application, specify a period for the data for AI learning, and perform learning. Thereby, during the operation of the equipment, the continuously sent equipment data is evaluated (i.e., inferred) by the learned AI to detect equipment abnormalities.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, conventional equipment anomaly detection systems often struggle to collect the equipment data necessary for AI training in the way users expect, resulting in a mix of data containing different behaviors and noise. As a result, the AI ​​may not be able to learn as intended, and the system may fail to achieve the performance required by the user.

[0006] Therefore, the purpose of this disclosure is to provide an equipment abnormality detection system and an equipment abnormality detection method that can detect equipment abnormalities more accurately than conventional methods. [Means for solving the problem]

[0007] To achieve the above objective, an equipment anomaly detection system according to one embodiment of the present disclosure is an equipment anomaly detection system that uses a machine learning model to detect equipment anomalies, comprising: a data acquisition unit that repeatedly acquires equipment data indicating the operating status of the equipment; a selection setting unit that receives from a user a specification of at least one item and the range of each feature of the at least one item from a plurality of items indicating the feature quantities of the equipment data acquired by the data acquisition unit; a feature extraction unit that extracts the feature quantities of the at least one item included in the specification received by the selection setting unit from the equipment data acquired by the data acquisition unit; and by referring to the feature quantities extracted by the feature extraction unit, the data acquisition The system includes: a learning unit that selects equipment data from the equipment data acquired by the data acquisition unit that has features satisfying the specifications received by the selection setting unit, and uses the selected equipment data to train the machine learning model to perform an evaluation of the equipment's abnormalities; an evaluation unit that, by referring to the features extracted by the feature extraction unit, selects equipment data from the equipment data acquired by the data acquisition unit that has features satisfying the specifications received by the selection setting unit, and inputs the selected equipment data into the machine learning model trained by the learning unit to perform an evaluation of the equipment's abnormalities; and an output unit that outputs data to display the results of the evaluation performed by the evaluation unit.

[0008] To achieve the above objective, an equipment anomaly detection method according to one embodiment of the present disclosure is an equipment anomaly detection method in which a computer detects an anomaly in equipment using a machine learning model, comprising: a data acquisition step of repeatedly acquiring equipment data indicating the operating status of the equipment; a selection setting step of receiving a specification from a user for at least one item and the range of each feature of the at least one item from a plurality of items indicating the feature quantities of the equipment data acquired in the data acquisition step; a feature quantity extraction step of extracting the feature quantities of the at least one item included in the specification received by the selection setting step from the equipment data acquired in the data acquisition step; and by referring to the feature quantities extracted in the feature quantity extraction step, the data The system includes: a learning step in which equipment data acquired in the data acquisition step is selected to have features that satisfy the specifications accepted in the selection setting step, and the machine learning model is trained using the selected equipment data to perform an evaluation of the equipment's abnormalities; an evaluation step in which equipment data acquired in the data acquisition step is selected to have features that satisfy the specifications accepted in the selection setting step by referring to the features extracted in the feature extraction step, and the machine learning model trained in the learning step is input with the selected equipment data to perform an evaluation of the equipment's abnormalities; and an output step in which the results of the evaluation performed in the evaluation step are displayed.

[0009] Furthermore, in order to achieve the above objectives, this disclosure may be implemented as a program that causes a computer to execute the above equipment abnormality detection method, as a non-temporary computer-readable recording medium such as a DVD on which the program is recorded, or as a program product. [Effects of the Invention]

[0010] This disclosure provides an equipment anomaly detection system and an equipment anomaly detection method that can detect equipment anomalies more accurately than conventional methods. [Brief explanation of the drawing]

[0011] [Figure 1] Figure 1 is a block diagram showing the configuration of the equipment abnormality detection system according to the embodiment. [Figure 2] Figure 2 is a functional block diagram showing the server configuration in Figure 1. [Figure 3] Figure 3 illustrates the phases from installation to operation of the equipment anomaly detection system according to the embodiment. [Figure 4] Figure 4 is a flowchart showing the operation of the equipment abnormality detection system according to the embodiment. [Figure 5] Figure 5 shows an example of a graphical user interface provided to the user via a terminal device by the sorting and setting unit of the equipment abnormality detection system according to the embodiment. [Figure 6] Figure 6 shows an example of a detailed display screen provided to the user via a terminal device by the output unit of the equipment abnormality detection system according to the embodiment. [Figure 7] Figure 7 is a diagram illustrating the significance of having a selection and setting unit in the equipment abnormality detection system according to the embodiment. [Figure 8] Figure 8 shows the relationship between the user's waveform selection settings workflow and the operation of the equipment anomaly detection system. [Modes for carrying out the invention]

[0012] (Insights gained by the inventors) Conventional equipment anomaly detection systems, such as those described in Patent Document 1, have difficulty collecting the equipment data necessary for AI training in the manner expected by the user, and equipment data containing different operations or noise may be mixed together. As a result, the AI ​​may not be able to learn as intended, and there are cases where the performance required by the user cannot be achieved.

[0013] The equipment data sent from the equipment includes measurement data showing waveforms corresponding to various equipment operations, each with different characteristics. Therefore, it is important for the user to select the equipment data they expect, but this is a complex task requiring specialized knowledge and tools. If such selection is not performed properly, there is a high risk that the AI ​​will learn from incorrect data and fail to achieve the expected performance. Therefore, there is a need for a means for users to intuitively reflect the selection settings.

[0014] Furthermore, during routine equipment diagnostics, measurement data showing waveforms corresponding to various equipment operations is continuously transmitted as equipment data. If equipment is evaluated indiscriminately using an ideally trained AI without filtering these waveforms, there is a risk that equipment data outside the training range will be input and incorrectly identified as a serious anomaly.

[0015] Therefore, an equipment abnormality detection system according to an aspect of the present disclosure is an equipment abnormality detection system that detects an abnormality of equipment using a machine learning model, and includes a data acquisition unit that repeatedly acquires equipment data indicating an operating state of the equipment, and a screening setting unit that receives, from a user, a designation of at least one item from a plurality of items indicating feature amounts of the equipment data acquired by the data acquisition unit and a range of each feature amount of the at least one item, a feature amount extraction unit that extracts a feature amount of the at least one item included in the designation received by the screening setting unit from the equipment data acquired by the data acquisition unit, a learning unit that selects equipment data having a feature amount that satisfies the designation received by the screening setting unit from the equipment data acquired by the data acquisition unit by referring to the feature amount extracted by the feature amount extraction unit, and performs learning to cause the machine learning model to perform an evaluation regarding the abnormality of the equipment using the selected equipment data, and an evaluation unit that selects equipment data having a feature amount that satisfies the designation received by the screening setting unit from the equipment data acquired by the data acquisition unit by referring to the feature amount extracted by the feature amount extraction unit, and inputs the selected equipment data into the machine learning model learned by the learning unit to perform an evaluation regarding the abnormality of the equipment, and an output unit that outputs data for displaying a result of the evaluation performed by the evaluation unit.

[0016] As a result, an equipment abnormality detection system is provided that can detect an abnormality of equipment more accurately than before, with the same condition settings being transparently and intuitively applied to the equipment data used during learning of the machine learning model and during evaluation using the learned machine learning model.

[0017] Specifically, by screening the equipment data, it becomes possible to narrow down the AI learning to a specific equipment operation pattern, improving the accuracy of the AI learning result and enabling accurate equipment diagnosis. Furthermore, when performing evaluation using the learned AI, the screening settings during learning are automatically reflected, making the operation simpler without the user having to remember the settings.

[0018] (Embodiment) Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. Note that all the embodiments described below are specific examples of the present disclosure. The sensors, signals, signal waveforms, numerical values, devices, components, arrangement positions and connection forms of the components, steps, step orders, etc. shown in the following embodiments are merely examples and are not intended to limit the present disclosure. Also, each figure is not necessarily drawn precisely. In each figure, substantially the same components are denoted by the same reference numerals, and overlapping descriptions are omitted or simplified.

[0019] FIG. 1 is a block diagram showing the configuration of an equipment abnormality detection system 10 according to an embodiment. The equipment abnormality detection system 10 is an equipment diagnosis system that detects abnormalities in equipment 40 installed in a factory 20 using a machine learning model, and includes a terminal device 30 and a server 50 that constitutes a cloud computer connected to the terminal device 30 via a communication network 60. In the present embodiment, the equipment abnormality detection system 10 is composed of the terminal device 30 and the server 50 shown in FIG. 1, but is not limited to such a configuration. For example, a system composed of only some of the components shown in FIG. 1, such as only the server 50, may be referred to as an equipment abnormality detection system. Furthermore, the equipment abnormality detection system 10 may include a measurement device such as a current sensor 43 described later.

[0020] The equipment 40 is a device including a motor 41 and a motor driver 42 that drives the motor 41, and is, for example, a production robot that manufactures products. In the present embodiment, a current sensor 43 that non - contact measures the current flowing through a drive cable connecting the motor 41 and the motor driver 42, an edge slave device 44 that generates measurement data indicating the current measured by the current sensor 43, an edge master device 45 that transmits the measurement data generated by the edge slave device 44 to the server 50 via the communication network 60, a terminal device 30 that communicates with the server 50 via the communication network 60, a router 61 that relays communication between the edge master device 45 and the terminal device 30, etc. are installed in the factory 20.

[0021] Furthermore, multiple current sensors 43 are provided to measure the current flowing through multiple locations, and the measurement data obtained from the multiple current sensors 43, which are distinguished by identification information, may be transmitted to the server 50 as equipment data. In addition, the router 61 may not only relay communication between the edge master unit 45 and the terminal device 30, but may also relay communication between the terminal device 30 and the edge master unit 45 and the server 50.

[0022] The terminal device 30 is connected to the server 50 via the communication network 60 and is a terminal device for users to interact with the server 50 via a web browser.

[0023] Server 50 is a cloud server that uses equipment data (in this case, measurement data) sent from the edge master unit 45 of the factory 20 via the communication network 60 to train a machine learning model, uses the trained machine learning model to evaluate anomalies in the equipment 40, and displays the obtained evaluation results on the terminal device 30 via the communication network 60.

[0024] Figure 2 is a functional block diagram showing the configuration of server 50 in Figure 1. This figure also shows the equipment 40 and the terminal device 30 used by user 31 in a simplified manner.

[0025] Server 50 is a remote computer that forms the main part of the equipment anomaly detection system 10, which uses a machine learning model to detect anomalies in the equipment 40, and includes a data acquisition unit 51, a selection and setting unit 52, a feature extraction unit 53, a learning unit 54, an evaluation unit 55, an output unit 56, and a storage unit 57.

[0026] The data acquisition unit 51 is implemented by a communication interface and / or a program executed by the processor, the selection and setting unit 52, feature extraction unit 53, learning unit 54, and evaluation unit 55 are implemented by a program executed by the processor, the output unit 56 is implemented by a communication interface, output devices such as a display and / or a program executed by the processor, and the storage unit 57 can be implemented by memory or a hard disk drive.

[0027] The data acquisition unit 51 repeatedly acquires equipment data indicating the operating status of the equipment 40 and stores it in the storage unit 57 as equipment data 57a. In this embodiment, the equipment data is measurement data indicating the operating status of the equipment 40 obtained using the current sensor 43, and is also waveform data indicated by the signal obtained from the equipment 40.

[0028] The sorting setting unit 52 receives a specification from the user 31 for at least one item and the range of each feature of at least one item from among several items representing the feature quantities of the equipment data acquired by the data acquisition unit 51 (this specification is also called the "waveform sorting setting"), and stores it in the storage unit 57 as the waveform sorting setting 57c. In this embodiment, the multiple items include the frequency, amplitude, and temporal length represented by the waveform data. Here, the frequency represented by the waveform data is, for example, the frequency of the maximum peak in the frequency spectrum of the waveform data; the amplitude of the waveform data is, for example, the peak-to-peak value in the time waveform of the waveform data; and the temporal length of the waveform data is the temporal length of the waveform data, which corresponds to, for example, the period during which the object driven by the motor 41 being measured is moving. The sorting setting unit 52 stores the input waveform sorting setting in the storage unit 57 in association with the identification information of the current sensor 43.

[0029] The feature extraction unit 53 extracts feature quantities for at least one item included in the waveform selection setting 57c received by the selection setting unit 52 from the equipment data acquired by the data acquisition unit 51, and stores them in the storage unit 57 as feature quantities 57b. The feature extraction unit 53 may also extract feature quantities for multiple fixed items from the equipment data acquired by the data acquisition unit 51, regardless of the waveform selection setting 57c received by the selection setting unit 52, and store them in the storage unit 57 as feature quantities 57b. The learning unit 54 and evaluation unit 55, which require feature quantities, may then select and read only the feature quantities corresponding to the waveform selection setting 57c received by the selection setting unit 52 from the feature quantities 57b stored in the storage unit 57.

[0030] The learning unit 54, by referring to the features 57b extracted by the feature extraction unit 53, selects (i.e., filters) equipment data from the equipment data acquired by the data acquisition unit 51 that have features satisfying the waveform selection setting 57c received by the selection setting unit 52, and creates a dataset. Using the equipment data included in the created dataset, the learning unit 54 trains the machine learning model 57d to evaluate anomalies in the equipment 40. At this time, the learning unit 54 may store the selected equipment data as a dataset in the storage unit 57 in association with the waveform selection setting 57c received by the selection setting unit 52, and perform training using a dataset selected by the user 31 from among the multiple saved datasets.

[0031] The machine learning model 57d is a one-class SVM (Support Vector Machine) that can recognize normal data as normal by being trained using a dataset created based only on equipment data obtained when equipment 40 is operating normally (i.e., using only normal data). The training dataset is, for example, the time evolution of the frequency spectrum obtained by Fourier analysis of waveform data.

[0032] The evaluation unit 55, by referring to the features 57b extracted by the feature extraction unit 53, selects (i.e., filters) equipment data from the equipment data 57a acquired by the data acquisition unit 51 that have features that satisfy the waveform selection setting 57c received by the selection setting unit 52. The selected equipment data is then input into the machine learning model 57d learned by the learning unit 54 to evaluate the abnormality of the equipment 40, and the evaluation result 57e is stored in the storage unit 57. The evaluation result 57e is, for example, a degree of change indicating the degree of deviation of the equipment 40 from a state in which it is operating normally.

[0033] Furthermore, even if the selection setting unit 52 receives a new waveform selection setting from the user 31 that is different from the waveform selection setting 57c used for learning by the learning unit 54, the evaluation unit 55 will, depending on the prior settings, either (1) perform the evaluation using equipment data that has features that satisfy the waveform selection setting 57c used at the time of learning by the learning unit 54 (i.e., by returning to the waveform selection setting used during learning), or (2) stop the evaluation.

[0034] The output unit 56 generates data for displaying the results of the evaluation performed by the evaluation unit 55 (i.e., the evaluation results 57e stored in the storage unit 57), and transmits it to the terminal device 30 or outputs it to a display or the like on the server 50 via the communication network 60.

[0035] More specifically, the output unit 56 has a detailed display screen mode that outputs data for displaying the equipment data acquired by the data acquisition unit 51 in a time-series graph for at least one feature quantity included in the waveform selection setting 57c received by the selection setting unit 52. In the detailed display screen mode, the output unit 56 outputs data for visually displaying the range of feature quantities included in the waveform selection setting 57c on a graph. At that time, the output unit 56 may also output data for visually displaying the range of feature quantities included in the waveform selection setting 57c after combining the ranges of feature quantities included in the waveform selection setting 57c in the detailed display screen mode.

[0036] Figure 3 is a diagram illustrating the phases (i.e., operating modes) from installation to operation of the equipment anomaly detection system 10 according to the embodiment. As shown in this figure, the equipment anomaly detection system 10 consists of three phases from installation to operation: installation phase 71, learning phase 72, and operation phase 73.

[0037] In the installation phase 71, the user 31 interacts with the data acquisition unit 51 and the sorting and setting unit 52 of the server 50 via a web application through the terminal device 30. Through the installation and setting screen 81 displayed on the terminal device 30, the user 31 sets information about the equipment 40 and the current sensor 43, and sets waveform sorting settings as filter conditions (installation and setting 71a). The user 31 also checks the measurement waveform obtained from the current sensor 43 through the measurement waveform display screen 82 displayed on the terminal device 30 (measurement waveform confirmation 71b).

[0038] In the learning phase 72, the user 31 interacts with the learning unit 54 of the server 50 via a web application through the terminal device 30. Through the dataset editing screen 83 displayed on the terminal device 30, the user 31 creates, edits, and verifies datasets by specifying the period of equipment data to be used for training the machine learning model 57d (i.e., the learning period) (dataset creation 72a). In dataset creation 72a, the user 31 can create one or more datasets while checking, in detailed display screen mode, whether the equipment data acquired by the data acquisition unit 51 satisfies the waveform selection settings 57c specified in interaction with the selection setting unit 52, using a time-series graph via the web application through the terminal device 30. Subsequently, the user 31 selects a dataset from among several to be used for training the machine learning model 57d, or executes the training, via the learning setting screen 84 displayed on the terminal device 30 (AI training preparation and execution 72b).

[0039] In operation phase 73, the user 31 interacts with the evaluation unit 55 of the server 50 via a web application through the terminal device 30, and selects a pre-trained machine learning model 57d via the AI ​​selection screen 85 displayed on the terminal device 30 to perform evaluation (i.e., inference) (setting the pre-trained AI as the basis 73a). In addition, it is possible to set not only newly acquired equipment data by the data acquisition unit 51 as the equipment data to be evaluated (i.e., set real-time evaluation), but also past equipment data that has already been acquired and stored by the data acquisition unit 51 (i.e., set past evaluation).

[0040] Furthermore, user 31 can select a measurement target they wish to check from the list of measurement targets screen 86a displayed on the terminal device 30, and then check the evaluation results for the selected measurement target via the evaluation results screen 86b (AI-based evaluation (inference) 73b). As an evaluation result, for example, the degree of change indicating the deviation of equipment 40 from a state in which it is operating normally is displayed, so user 31 can find out the degree of abnormality of equipment 40, or the signs of an abnormality.

[0041] Figure 4 is a flowchart illustrating the operation of the equipment anomaly detection system 10 according to the embodiment (i.e., the equipment anomaly detection method). Here, the operation of the server 50, in particular the operation of the sorting and setting unit 52 and its related components, is shown.

[0042] First, based on instructions from user 31, the server 50's data acquisition unit 51 repeatedly acquires equipment data indicating the operating status of equipment 40 and stores it in the storage unit 57 as equipment data 57a (S10). The feature extraction unit 53 extracts feature quantities for multiple items such as frequency, amplitude, and temporal length from the equipment data 57a stored in the storage unit 57 and stores them in the storage unit 57 as feature quantities 57b.

[0043] User 31 checks the accumulated equipment data 57a (S11) and repeats the accumulation of equipment data until sufficient equipment data is accumulated for training the machine learning model 57d (S10).

[0044] When sufficient equipment data for training the machine learning model 57d has been accumulated (Yes in S11), the learning unit 54 accepts the setting of the accumulation period for the equipment data 57a to create a training dataset (i.e., the learning period) based on the instructions from the user 31 (S12). Then, the selection setting unit 52 accepts the waveform selection setting based on the instructions from the user 31 and stores it in the storage unit 57 as the waveform selection setting 57c, corresponding to the identification information of the pre-specified current sensor 43 (S13).

[0045] In the detailed display screen mode, user 31 checks the equipment data 57a for the specified learning period in a graph, specifically the equipment data 57a whose feature quantities 57b satisfy the waveform selection setting 57c (i.e., valid data), to determine whether the amount of valid data is sufficient and the variation in feature quantities is suppressed (S14). If the result is negative (No in S14), the user readjusts the learning period as necessary (S12), and then readjusts the waveform selection setting 57c as necessary (S13), before repeating the judgment regarding the amount of valid data and the variation in feature quantities (S14).

[0046] On the other hand, if it is determined that the amount of valid data is sufficient and the variation in features is suppressed (Yes in S14), the user 31 uses the equipment data via the dataset editing screen 83 to create a dataset to be used for training the machine learning model 57d (S15). The dataset created here consists of equipment data having features that satisfy the waveform selection setting 57c received by the selection setting unit 52, and is saved in association with the waveform selection setting 57c.

[0047] Subsequently, based on instructions from user 31, the learning unit 54 uses the selected dataset to train the machine learning model 57d to perform an evaluation of anomalies in the equipment 40 (S16).

[0048] Then, based on instructions from user 31, evaluation unit 55 selects equipment data 57a from the equipment data 57a acquired by data acquisition unit 51 that have features that satisfy the waveform selection setting 57c received by selection setting unit 52, and inputs the selected equipment data 57a into machine learning model 57d learned by learning unit 54 to perform an evaluation of abnormalities in equipment 40, and saves the evaluation result 57e in storage unit 57 (S17). Output unit 56 presents the evaluation result 57e obtained by evaluation unit 55 to user 31 via a web application through terminal device 30.

[0049] User 31, after reviewing the evaluation result 57e, determines whether the evaluation result 57e is stable or not (S18). If the result is negative (No in S18), the user readjusts the learning period as necessary (S12), and then readjusts the waveform selection setting 57c as necessary (S13). After that, the user repeats the process of determining the variability of the effective data amount and features (S14), creating a dataset (S15), learning (S16), and evaluating (S17).

[0050] On the other hand, if the evaluation result 57e is deemed stable (Yes in S18), the user 31 continues the evaluation by the evaluation unit 55 (S19).

[0051] As described above, the equipment anomaly detection system 10 according to the embodiment is provided with a selection setting unit 52, and the user 31 can set waveform selection, which is a filter condition for equipment data 57a used for training the machine learning model 57d, and also a filter condition for equipment data 57a that is subject to evaluation by the trained machine learning model 57d (S13). Therefore, the same condition setting is transparently and intuitively applied to the equipment data used during training of the machine learning model 57d and during evaluation by the trained machine learning model 57d.

[0052] Figure 5 shows an example of a graphical user interface (i.e., an input screen for waveform selection settings) provided by the selection setting unit 52 of the equipment anomaly detection system 10 according to the embodiment to the user 31 via a web application through the terminal device 30. The selection setting unit 52 accepts specifications from the user 31 for the following items as waveform selection settings (i.e., feature items) that will be used as filter conditions for equipment data during training of the machine learning model 57d and evaluation by the trained machine learning model 57d: operation pattern 52a, cleansing 52b, frequency 52c, amplitude 52d, and time 52e.

[0053] The operation pattern 52a is an item that specifies the operation pattern of the object driven by the motor 41 being measured. Specifically, one of the following options is selected for the operation pattern: "Not specified", "Acceleration / Deceleration" which is the operation pattern from when the object starts moving until it stops, "Constant Velocity" which is when the object moves at a constant velocity, and "Constant Velocity (High Resolution)" which is a constant velocity that prioritizes frequency resolution over temporal resolution.

[0054] Cleansing 52b is an item for selecting whether to "specify a range" or "do not specify a range" for the following three features (frequency 52c, amplitude 52d, and time 52e).

[0055] Frequency 52c is an item that accepts the selection (check box) whether or not to adopt it as a waveform selection setting, and the specification of the frequency range if adopted (i.e., the frequency range of the maximum peak in the frequency spectrum of the waveform data shown by the equipment data).

[0056] The amplitude 52d is an item that accepts the selection (check box) whether or not to adopt it as a waveform selection setting, and the specification of the amplitude range if adopted (i.e., the range of peak-to-peak values ​​in the time waveform of the waveform data shown by the equipment data).

[0057] Time 52e is an item that accepts the selection (check box) of whether or not to adopt the waveform selection setting, and the specification of the time length range if adopted (i.e., the time length range of the waveform data indicated by the equipment data).

[0058] The waveform selection settings configured in the graphical user interface shown in Figure 5 are stored in the memory unit 57 as waveform selection settings 57c, and are used by the learning unit 54 as filter conditions for the equipment data to be used to create a dataset for training the machine learning model 57d, and are also used by the evaluation unit 55 as filter conditions for the equipment data to be evaluated by the machine learning model 57d.

[0059] Figure 6 shows an example of a detailed display screen 56a provided to a user 31 via a web application through a terminal device 30 by the output unit 56 of the equipment abnormality detection system 10 according to the embodiment. Here, the equipment data (i.e., the frequency of the equipment data) acquired by the data acquisition unit 51 for at least one feature quantity (in this case, "frequency") included in the waveform selection setting 57c received by the selection setting unit 52 is graphed (i.e., plotted and visualized) in time series.

[0060] More specifically, in the detailed display screen 56a of Figure 6, it is shown that the "frequency" is selected as the display data 56b from the feature quantities ("frequency", "amplitude", "time") included in the waveform selection setting 57c received by the selection setting unit 52, and that the display period 56c is set to "2024 / 04 / 24 10:12~2024 / 05 / 08 17:05". In graph 56d, the vertical axis is "frequency", which is the feature quantity selected in the display data 56b, and the horizontal axis is "2024 / 04 / 24 10:12~2024 / 05 / 08 17:05" set in the display period 56c, and the frequencies of the equipment data acquired by the data acquisition unit 51 during this period are plotted.

[0061] In the vertical axis direction of graph 56d, the central region 56d2, sandwiched between the two shaded regions 56d1 at the top and bottom, corresponds to the frequency range (the frequency range 52c in Figure 5) accepted by the selection setting unit 52 as the waveform selection setting 57c, while the two shaded regions 56d1 correspond to the range outside that frequency range.

[0062] Furthermore, the rectangular region 56d3 enclosed by a thick line in the center of graph 56d is the region containing equipment data that belongs to the training period specified by user 31 on the dataset editing screen 83 and has features (i.e., frequencies) that belong to the range of features (i.e., frequency range) specified by user 31 as the waveform selection setting 57c. By visually checking the equipment data contained in this rectangular region 56d3, user 31 can determine whether the amount of equipment data 57a (i.e., effective data) used for training the machine learning model 57d is sufficient and whether the variation in features is suppressed (S14 in Figure 4).

[0063] In Figure 6, the two shaded regions 56d1 and the central region 56d2 represent the ranges input in the waveform selection setting 57c for the feature quantity "frequency" selected in the display data 56b. However, the display is not limited to this configuration, and the ranges of multiple feature quantities (frequency 52c, amplitude 52d, and time 52e) input as the waveform selection setting 57c may be combined and displayed. For example, the vertical axis may be set to the normalized values ​​of multiple feature quantities (frequency 52c, amplitude 52d, and time 52e) between 0 and 100, the AND region (range belonging to all feature quantities) of the ranges of multiple feature quantities (frequency 52c, amplitude 52d, and time 52e) input as the waveform selection setting 57c may be displayed as the central region 56d2, the other regions may be displayed as the two shaded regions 56d1, and the feature quantity "frequency" selected in the display data 56b may be displayed in graph 56d. In other words, the two shaded regions 56d1 represent the OR region of the ranges of multiple features (frequency 52c, amplitude 52d, and time 52e). This allows user 31 to quickly check the amount of equipment data (effective data amount) that satisfies all ranges of the multiple features (frequency 52c, amplitude 52d, and time 52e) input as waveform selection settings 57c by looking at graph 56d.

[0064] Figure 7 is a diagram illustrating the significance of providing a selection and setting unit 52 in the equipment anomaly detection system 10 according to the embodiment. In the equipment anomaly detection system 10, the machine learning model 57d is trained using only normal data, as described above.

[0065] Figure 7(a1) shows an example of the frequency distribution of normal data (black circles) used for training the machine learning model 57d and data (white circles) used for evaluation, illustrating the case where the variability in the distribution of normal data is small. On the other hand, Figure 7(b1) shows an example of the frequency distribution of normal data (black circles) used for training the machine learning model 57d and data (white circles) used for evaluation, illustrating the case where the variability in the distribution of normal data is large.

[0066] Figure 7(a2) shows an example of the evaluation result obtained by evaluating the machine learning model 57d in the case shown in Figure 7(a1) (here, the degree of change indicating the deviation from the state in which the equipment 40 is operating normally). Here, it is shown that a "sensitive" machine learning model 57d is created that generates a predetermined degree of change (here, 0 to 60) for small changes (i.e., the length of the horizontally extending arrow). On the other hand, Figure 7(b2) shows an example of the evaluation result obtained by evaluating the machine learning model 57d in the case shown in Figure 7(b1) (here, the degree of change). Here, it is shown that a "blunt" machine learning model 57d is created that generates a predetermined degree of change (here, 0 to 60) for large changes (i.e., the length of the horizontally extending arrow).

[0067] The equipment anomaly detection system 10 according to this embodiment is provided with a selection setting unit 52. By setting the filter conditions for the equipment data 57a used to train the machine learning model 57d as waveform selection setting 57c, the user 31 can train the machine learning model 57d using only equipment data with specific features, thereby creating a "sensitive" machine learning model 57d that can detect anomalies in the equipment 40 with high sensitivity, as shown in Figures 7(a1) and (a2).

[0068] Figure 8 shows the workflow for waveform selection settings by user 31 and the relationship between this workflow and the operation of the equipment anomaly detection system 10. More specifically, Figure 8(a) shows the workflow for waveform selection settings by user 31 using the detailed display screen, and Figure 8(b) shows the workflow for training the machine learning model 57d by the learning unit 54 and the evaluation by the evaluation unit 55 using the trained machine learning model 57d.

[0069] Now, let's assume that user 31 has set waveform selection setting A through interaction with the selection setting unit 52, created a training dataset (S20), and started training (S21). Then, the learning unit 54 starts training the machine learning model 57d using equipment data that has features that satisfy waveform selection setting A (S30).

[0070] During the learning process, user 31 sets a new waveform selection setting B through interaction with the selection setting unit 52 (S22). After the learning is complete, user 31 starts the evaluation by the evaluation unit 55 based on instructions from the detailed display screen (S23).

[0071] Then, when learning is complete, the evaluation unit 55 detects that a new waveform selection setting B, different from the waveform selection setting A used during learning, has been set. At the start of the evaluation, in the settings on the detailed display screen, it overwrites waveform selection setting B with waveform selection setting A (S31), and performs the evaluation using equipment data that has features that satisfy waveform selection setting A (S32). At this time, regarding the display, if the selection setting unit 52 receives a new waveform selection setting 57c from the user that is different from the waveform selection setting 57c used during learning by the learning unit 54, the output unit 56 outputs data to display the waveform selection setting 57c used during learning by the learning unit 54, rather than the new waveform selection setting 57c, when learning by the learning unit 54 is completed and evaluation by the evaluation unit 55 begins.

[0072] Therefore, when user 31 starts the evaluation by the evaluation unit 55, he notices that the waveform selection setting B, which was newly set immediately before, has been automatically changed to waveform selection setting A (S23). In addition to the display in step S31 above, if the selection setting unit 52 receives a new waveform selection setting 57c from the user that is different from the waveform selection setting 57c used in learning, during learning by the learning unit 54 or evaluation by the evaluation unit 55, the output unit 56 may output data to display the waveform selection setting 57c used at the time of learning by the learning unit 54, rather than the new waveform selection setting 57c, triggered by a user action such as a display instruction.

[0073] As described above, when the selection setting unit 52 receives a new waveform selection setting from the user 31 that is different from the waveform selection setting used for learning by the learning unit 54, the evaluation unit 55 performs the evaluation using equipment data that has features that satisfy the waveform selection setting used at the time of learning by the learning unit 54, even if it is a new waveform selection setting (that is, when a new waveform selection setting is received, the waveform selection setting is reverted to the content at the time of learning). This ensures that the waveform selection setting used during the learning of the machine learning model 57d and the evaluation using the trained machine learning model 57d are the same.

[0074] In the example shown in Figure 8, when user 31 changed the waveform selection settings used during training, the evaluation unit 55 restored the waveform selection settings to those used during training and then started the evaluation. However, the process is not limited to this, and the evaluation may be stopped. This also prevents the waveform selection settings used during training of the machine learning model 57d from being different from those used during evaluation with the trained machine learning model 57d. Whether to overwrite the waveform selection settings or stop the evaluation may be switched in the prior settings for the evaluation unit 55.

[0075] As described above, the equipment anomaly detection system 10 according to the embodiment is a system that detects anomalies in equipment 40 using a machine learning model 57d, and comprises a data acquisition unit 51 that repeatedly acquires equipment data indicating the operating status of equipment 40, a selection setting unit 52 that receives from a user 31 a specification (i.e., waveform selection setting 57c) for at least one item and the range of each feature of at least one item from a plurality of items indicating the feature quantities of the equipment data acquired by the data acquisition unit 51, a feature quantity extraction unit 53 that extracts feature quantities for at least one item included in the waveform selection setting 57c received by the selection setting unit 52 from the equipment data acquired by the data acquisition unit 51, and data acquisition by referring to the feature quantities extracted by the feature quantity extraction unit 53. The system includes: a learning unit 54 that selects equipment data from the equipment data acquired by the data acquisition unit 51 that has features satisfying the waveform selection setting 57c accepted by the selection setting unit 52, and uses the selected equipment data to train a machine learning model 57d to perform an evaluation of abnormalities in the equipment 40; an evaluation unit 55 that, by referring to the features extracted by the feature extraction unit 53, selects equipment data from the equipment data acquired by the data acquisition unit 51 that has features satisfying the waveform selection setting 57c accepted by the selection setting unit 52, and inputs the selected equipment data into the machine learning model 57d trained by the learning unit 54 to perform an evaluation of abnormalities in the equipment 40; and an output unit 56 that outputs data to display the results of the evaluation performed by the evaluation unit 55.

[0076] As a result, the waveform selection settings 57c received by the selection setting unit 52 from the user 31 are used to filter the input equipment data not only when training the machine learning model 57d, but also when the trained machine learning model 57d evaluates abnormalities in the equipment 40. Therefore, the same condition settings are transparently and intuitively applied to the equipment data used during machine learning model training and evaluation by the trained machine learning model, resulting in an equipment abnormality detection system that can detect equipment abnormalities more accurately than before.

[0077] Here, the equipment data is measurement data indicating the operating status of the equipment 40 obtained using sensors, and the sorting setting unit 52 stores the waveform sorting setting 57c in association with the sensor identification information. This allows the waveform sorting setting 57c to be set for each sensor, and by setting a waveform sorting setting 57c suitable for the type of sensor and measurement location, the accuracy of the evaluation can be improved.

[0078] Furthermore, the equipment data is waveform data represented by the signal obtained from equipment 40, and multiple items include at least one of the frequency, amplitude, and temporal length represented by the waveform data. This makes it possible to set appropriate filter conditions (i.e., waveform selection settings) that match the operating state of equipment 40 by using the various waveform characteristics of the equipment data as filter conditions.

[0079] Furthermore, the output unit 56 has a detailed display screen mode that outputs data for displaying the equipment data acquired by the data acquisition unit 51 in a time-series graph for at least one feature quantity included in the waveform selection setting 57c received by the selection setting unit 52. More specifically, in the detailed display screen mode, the output unit 56 outputs data for visually displaying the range of feature quantities included in the waveform selection setting 57c on a graph. As a result, the user 31 can intuitively check the appropriateness of the waveform selection setting 57c using the graph on the detailed display screen, and can easily determine the waveform selection setting 57c that is suitable for anomaly detection.

[0080] Furthermore, if at least one item included in the waveform selection setting 57c is one of several items representing the features of the equipment data, the output unit 56 may output data for visualization and display on a graph after synthesizing the ranges of the features of the multiple items included in the waveform selection setting 57c in the detailed display screen mode. This allows the appropriateness of the training dataset to be checked for the synthesized ranges of multiple features when multiple feature ranges are specified as the waveform selection setting 57c, thus enabling quick verification of the appropriateness of the waveform selection setting 57c even when multiple feature ranges are specified.

[0081] Furthermore, even if the selection setting unit 52 receives a new waveform selection setting from the user 31 that is different from the waveform selection setting 57c used for learning by the learning unit 54, the evaluation unit 55 will perform the evaluation using equipment data that has features that satisfy the waveform selection setting 57c used at the time of learning by the learning unit 54, rather than the new waveform selection setting. This ensures that the waveform selection setting used during the learning of the machine learning model 57d and the evaluation using the trained machine learning model 57d are the same.

[0082] Furthermore, if the selection setting unit 52 receives a new waveform selection setting 57c from the user that is different from the waveform selection setting 57c used in learning during learning by the learning unit 54, the output unit 56 may output data to display the waveform selection setting 57c used at the time of learning by the learning unit 54, rather than the new waveform selection setting 57c, when learning by the learning unit 54 is completed and evaluation by the evaluation unit 55 begins. Alternatively, if the selection setting unit 52 receives a new waveform selection setting 57c from the user that is different from the waveform selection setting 57c used in learning during learning by the learning unit 54 or evaluation by the evaluation unit 55, the output unit 56 may, triggered by user action, output data to display the waveform selection setting 57c used at the time of learning by the learning unit 54, rather than the new waveform selection setting 57c.

[0083] This ensures that even if a new waveform selection setting 57c different from the one used for training is input by the user during training by the learning unit 54 or evaluation by the evaluation unit 55, the waveform selection setting used during training of the machine learning model 57d and during evaluation by the trained machine learning model 57d will be the same, and this sameness will also be ensured in the display to the user.

[0084] Furthermore, the evaluation unit 55 may stop the evaluation if the selection setting unit 52 receives a new waveform selection setting from the user 31 that is different from the waveform selection setting 57c used for training by the learning unit 54. This avoids the situation where the waveform selection setting used during training of the machine learning model 57d and the setting used during evaluation by the trained machine learning model 57d are different.

[0085] Furthermore, the learning unit 54 may store the selected equipment data in association with the waveform selection settings 57c received by the selection setting unit 52, and from the multiple saved datasets, it may perform training using the dataset selected by the user 31, and link the trained machine learning model 57d, the selected dataset, and the corresponding waveform selection settings 57c (i.e., store information indicating the correspondence). This allows the waveform selection settings 57c, dataset, and machine learning model 57d to be stored in a tripartite association depending on the type of equipment 40 being targeted, enabling operation tailored to the target of anomaly detection.

[0086] Furthermore, the equipment anomaly detection method according to the embodiment is a method in which a computer detects an anomaly in equipment 40 using a machine learning model 57d, and comprises: a data acquisition step (S10) in which equipment data indicating the operating status of equipment 40 is repeatedly acquired; a selection setting step (S13) in which a user 31 specifies at least one item and the range of each feature of at least one item from a plurality of items indicating the feature quantities of the equipment data acquired in the data acquisition step (i.e., waveform selection setting 57c); a feature quantity extraction step in which the feature quantities of at least one item included in the waveform selection setting 57c accepted by the selection setting step are extracted from the equipment data acquired in the data acquisition step; and data acquisition by referring to the feature quantities extracted in the feature quantity extraction step. The system includes a learning step (S14-S16) in which, from the equipment data acquired in the step, equipment data having features that satisfy the waveform selection setting 57c accepted in the selection setting step is selected, and the machine learning model 57d is trained using the selected equipment data to perform an evaluation of abnormalities in the equipment 40; an evaluation step (S17) in which, by referring to the features extracted in the feature extraction step, equipment data having features that satisfy the waveform selection setting 57c accepted in the selection setting step is selected from the equipment data acquired in the data acquisition step, and the selected equipment data is input into the machine learning model 57d trained in the learning step to perform an evaluation of abnormalities in the equipment 40; and an output step (S17) in which the results of the evaluation performed in the evaluation step are displayed.

[0087] As a result, the waveform selection settings 57c received by the selection setting unit 52 from the user 31 are used to filter the input equipment data not only when training the machine learning model 57d, but also when the trained machine learning model 57d evaluates abnormalities in the equipment 40. Therefore, the same condition settings are transparently and intuitively applied to the equipment data used during machine learning model training and evaluation by the trained machine learning model, resulting in an equipment abnormality detection system that can detect equipment abnormalities more accurately than before.

[0088] Furthermore, this disclosure may be implemented not only as an equipment anomaly detection system or equipment anomaly detection method, but also as a program that causes a computer to execute the steps included in the above-mentioned equipment anomaly detection method, as a non-temporary computer-readable recording medium such as a DVD on which the program is recorded, or as a program product.

[0089] The equipment anomaly detection system and equipment anomaly detection method relating to this disclosure have been described above based on embodiments, but this disclosure is not limited to these embodiments. As long as they do not depart from the spirit of this disclosure, various modifications to these embodiments that a person skilled in the art could conceive of, and other forms constructed by combining some of the components of the embodiments, are also included within the scope of this disclosure.

[0090] For example, in this embodiment, the equipment data transmitted from the factory 20 to the server 50 was waveform data measured by the current sensor 43, but it is not limited to this. Any data indicating the operating status of the equipment 40 is acceptable, such as information regarding the operation of the equipment 40, information regarding the power consumption of the equipment 40, or information regarding the temperature of the equipment 40.

[0091] Furthermore, the server 50 may detect anomalies in one piece of equipment 40, or it may detect anomalies in multiple pieces of equipment. In addition, the server 50 may be implemented not only by a single computer, but also by multiple computers that perform distributed processing.

[0092] Furthermore, although the equipment anomaly detection system 10 was a cloud system in this embodiment, it is not limited to this form. For example, the server 50 may be a local computer installed in the same factory 20 as the equipment 40, or the functions of the server 50 (i.e., programs and data) may be implemented within the terminal device 30 (i.e., a standalone type in which the functions of the terminal device and the functions of the server are integrated). [Industrial applicability]

[0093] The equipment anomaly detection system described herein is a system that uses a machine learning model to detect or predict equipment anomalies. In particular, it can be used as an equipment anomaly detection system that can detect equipment anomalies more accurately than conventional systems by transparently and intuitively applying the same conditions to the equipment data used during the training of the machine learning model and during evaluation by the trained machine learning model. [Explanation of Symbols]

[0094] 10. Equipment Anomaly Detection System 20 factories 30 Terminal devices 40 Equipment 41 Motor 42 Motor Drivers 43 Current Sensor 44 Edge Sub-unit 45 Edge Master Unit 50 servers 51 Data Acquisition Unit 52 Selection setting section 52a Operating Pattern 52b Cleansing 52c frequency 52d amplitude 52 hours 53 Feature Extraction Unit 54 Learning Department 55 Evaluation Department 56 Output section 56a Detail display screen 56b Display Data 56c Display period 56d Graph 56d1 Shaded area 56d2 central area 56d3 rectangular area 57 Memory section 57a Equipment Data 57b Features 57c Waveform Selection Settings 57d Machine Learning Model 57e Evaluation Results 60 Communication Networks 61 Routers 71 Installation Phase 71a Installation & Setup 71b Verification of measured waveform 72 Learning Phases 72a Creating the dataset 72b AI training preparation and execution 73 Operational Phase 73a Setting the pre-trained AI as the core. 73b AI-based evaluation (inference) 81 Installation & Settings Screen 82 Measured waveform display screen 83. Dataset Editing Screen 84 Learning Settings Screen 85. Screen for selecting the AI ​​to use 86a List screen 86b Evaluation results screen

Claims

1. An equipment anomaly detection system that uses a machine learning model to detect equipment abnormalities, A data acquisition unit that repeatedly acquires equipment data indicating the operating status of the equipment, A selection setting unit receives from the user the specification of at least one item and the range of each feature of the equipment data acquired by the data acquisition unit, from among a plurality of items representing the feature quantities of the equipment data acquired by the data acquisition unit, A feature extraction unit extracts feature quantities from the equipment data acquired by the data acquisition unit for at least one item included in the designation received by the selection setting unit, A learning unit selects equipment data from the equipment data acquired by the data acquisition unit that has features that satisfy the specifications received by the selection setting unit, by referring to the features extracted by the feature extraction unit, and uses the selected equipment data to train the machine learning model to perform an evaluation regarding abnormalities in the equipment. An evaluation unit, by referring to the features extracted by the feature extraction unit, selects equipment data from the equipment data acquired by the data acquisition unit that has features that satisfy the specifications received by the selection setting unit, and inputs the selected equipment data into the machine learning model learned by the learning unit, thereby performing an evaluation regarding abnormalities in the equipment. The system includes an output unit that outputs data for displaying the results of the evaluation performed by the evaluation unit. Equipment malfunction detection system.

2. The data of the aforementioned equipment is measurement data indicating the operating status of the equipment obtained using sensors. The sorting setting unit stores the designation in association with the identification information of the sensor. The equipment anomaly detection system according to claim 1.

3. The aforementioned equipment data is waveform data represented by the signal obtained from the equipment. The aforementioned multiple items include at least one of the frequency, amplitude, and temporal length represented by the waveform data. The equipment anomaly detection system according to claim 1.

4. The output unit has a detailed display screen mode that outputs data for displaying the equipment data acquired by the data acquisition unit in a time-series graph for the feature quantities of at least one item included in the specification received by the selection setting unit. The equipment anomaly detection system according to claim 1.

5. The output unit outputs data for visualizing and displaying the range of the feature quantities included in the specification on the graph in the detailed display screen mode. The equipment anomaly detection system according to claim 4.

6. The at least one item included in the designation is a plurality of items that represent the characteristic quantities of the equipment data, The output unit, in the detailed display screen mode, outputs data for visualizing and displaying on the graph after combining the ranges of the feature quantities of the multiple items included in the specification. The equipment malfunction detection system according to claim 5.

7. Even if the selection setting unit receives a new specification from the user that differs from the specification used for learning by the learning unit, the evaluation unit will perform the evaluation using equipment data having feature quantities that satisfy the specification used at the time of learning by the learning unit, rather than the new specification. The equipment anomaly detection system according to claim 1.

8. If, during learning by the learning unit, the selection setting unit receives a new specification from the user that differs from the specification used during learning, the output unit outputs data to display the specification used at the time of learning by the learning unit, rather than the new specification, when learning by the learning unit ends and evaluation by the evaluation unit begins. The equipment anomaly detection system according to claim 1.

9. If the selection setting unit receives a new specification from the user that differs from the specification used during learning by the learning unit or evaluation by the evaluation unit, the output unit, triggered by the user's action, outputs data to display the specification used at the time of learning by the learning unit, rather than the new specification. The equipment anomaly detection system according to claim 1.

10. The evaluation unit stops the evaluation when the selection setting unit receives a new specification from the user that is different from the specification used for learning by the learning unit. The equipment anomaly detection system according to claim 1.

11. The learning unit stores the selected equipment data in association with the specification received by the selection setting unit, performs the learning using the dataset selected by the user from the multiple stored datasets, and links the learned machine learning model, the selected dataset, and the corresponding specification. The equipment anomaly detection system according to claim 1.

12. A method for detecting equipment anomalies in which a computer detects equipment anomalies using a machine learning model, A data acquisition step that repeatedly acquires equipment data indicating the operating status of the equipment, A selection setting step in which the user specifies at least one item from among multiple items representing the characteristic quantities of the equipment data acquired in the data acquisition step, and the range of each characteristic quantity of the at least one item. A feature extraction step which extracts feature quantities for at least one item included in the specification received in the selection setting step from the equipment data acquired in the data acquisition step, A learning step in which, by referring to the features extracted in the feature extraction step, equipment data acquired in the data acquisition step that have features that satisfy the specifications accepted in the selection setting step are selected, and the machine learning model is trained to perform an evaluation of abnormalities in the equipment using the selected equipment data. An evaluation step which involves selecting equipment data from the equipment data acquired in the data acquisition step that has features that satisfy the specifications accepted in the selection setting step, by referring to the features extracted in the feature extraction step, and inputting the selected equipment data into the machine learning model learned in the learning step to evaluate the abnormalities of the equipment, The process includes an output step that displays the results of the evaluation performed in the aforementioned evaluation step, Equipment malfunction detection method.

13. (program) A program that causes a computer to execute the steps included in the equipment abnormality detection method described in claim 12.

Images