Anomaly Detection System

Abstract

To provide an anomaly detection system using a learning model with improved predictive performance compared to a learning model trained using only false positive data in which the learning model erroneously determines that an anomaly detection target is abnormal when in fact it is normal. [Solution] The anomaly detection system includes a processor that acquires multiple data from an anomaly detection target, and trains a learning model on the multiple data, including data in which a human has mistakenly determined that the anomaly detection target is abnormal when in fact the target is normal, as learning data.

Description

[Technical field]

[0001] The present invention relates to an anomaly detection system. [Background technology]

[0002] Patent Document 1 aims to provide a learning data selection device capable of selecting learning data capable of reducing false positives of a fault detection target. In order to solve this problem, Patent Document 1 discloses a method of providing sensor data to a learning model, acquiring detection data from the learning model, and selecting the detection data as learning data to be used for re-learning the learning model. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2023 / 188052 Summary of the Invention [Problem to be solved by the invention]

[0004] For example, a person may empirically judge the anomaly of an anomaly detection target from various data acquired from the anomaly detection target such as a production facility. In this case, the person may not be able to completely distinguish between a failure and a normal state, and may erroneously judge a normal state as an anomaly. This may result in unnecessary maintenance being performed, leading to cost loss. In order to detect anomalies in an anomaly detection target, there is a learning model that detects anomalies in equipment by using false positive data that indicates that the anomaly detection target is abnormal when it is normal as learning data for the learning model. However, in this learning model, only false positive data in which the learning model erroneously judges that the anomaly detection target is abnormal when it is normal is used. The present invention aims to provide an anomaly detection system that uses a learning model with improved predictive performance compared to a learning model trained using only false positive data in which the learning model erroneously determines that the target for anomaly detection is abnormal when in fact it is normal. [Means for solving the problem]

[0005] The invention described in claim 1 is an anomaly detection system comprising a processor, the processor acquiring a plurality of data from an object for anomaly detection, and training a learning model to learn the plurality of data, including data in which a human has mistakenly determined that the object for anomaly detection is abnormal when in fact the object is normal, as learning data. The invention described in claim 2 is the anomaly detection system described in claim 1, characterized in that the data in which the object of anomaly detection has been mistakenly judged to be abnormal by a human being when the object is normal is data in which the human has mistakenly judged to be abnormal including information other than the erroneously judged data. The invention described in claim 3 is the anomaly detection system described in claim 2, characterized in that the erroneously judged data is vibration data obtained from vibrations of the object of anomaly detection, and information other than the erroneously judged data is information other than the vibration data. The invention described in claim 4 is an anomaly detection system comprising a processor, which reads driving data of an inference target, obtains a learning model trained using multiple driving data including data in which a human has erroneously determined that an anomaly detection target is abnormal when in fact the target is normal, and uses the obtained learning model to detect whether the driving data of the inference target indicates an anomaly in the anomaly detection target. The invention described in claim 5 is the anomaly detection system described in claim 4, characterized in that the learning model can be diverted to an anomaly detection target different from the anomaly detection target. Effect of the Invention

[0006] The present invention can provide an anomaly detection system that uses a learning model with improved predictive performance compared to a learning model trained only using false positive data in which the learning model erroneously determines that the target for anomaly detection is abnormal when in fact it is normal. [Brief description of the drawings]

[0007] [Figure 1] 1A and 1B are diagrams showing the overall configuration of an anomaly detection system. [Diagram 2] 1A and 1B are diagrams showing the functional configuration of a learning device or an inference device. [Diagram 3] This shows the hardware configuration of a data storage device, a learning device, and an inference device. [Figure 4] 6(a) to 6(c) are diagrams showing characteristics of vibration data according to time series. [Diagram 5] 13 is a flowchart showing a process flow of the learning device. [Figure 6] 13 is a flowchart showing a process flow of the inference device. [Figure 7] FIG. 2 is a diagram illustrating learning data that is a learning target for a learning model. [Figure 8] FIG. 13 is a diagram illustrating false positive data. [Figure 9] 13(a) and 13(b) are diagrams showing initial learning or relearning of a learning model. [Figure 10] 13 is a graph showing the degree of anomaly output by a trained learning model over time. [Figure 11] FIG. 13 is a diagram illustrating the reuse of a trained learning model. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. 1(a) and (b) are diagrams showing the overall configuration of the anomaly detection system. FIG. 1(a) shows the configuration of the anomaly detection system 1 in the learning stage, and FIG. 1(b) shows the configuration of the anomaly detection system 1 in the inference stage. The learning stage is a stage in which a learning model 150 possessed by a learning device 100 performs learning, and the inference stage is a stage in which an inference device 200 performs anomaly detection of a pump 10 using the trained learning model 150. The trained learning model 150 will be described later with reference to FIG. 2.

[0009] First, as shown in FIG. 1(a), in the learning stage, the anomaly detection system 1 has a pump 10, which is an example of an anomaly detection target and is a device monitored for the presence or absence of an anomaly by a human or an inference device 200. The anomaly detection system 1 also has a sensor 20 that detects the vibration state of the pump 10. The anomaly detection system 1 also has a data storage device 30 that acquires vibration data, which is data indicating the vibration state of the pump 10, from the sensor 20 and stores it as past vibration data. The anomaly detection system 1 also has a learning device 100 that has a learning model 150 and causes the learning model 150 to learn operation data as learning data. The anomaly detection system 1 also has an inference device 200 that predicts whether the pump 10 is abnormal or not using the learning model 150 learned by the learning device 100. The learning model 150 learned by the learning device 100 may be referred to as a learned learning model 150.

[0010] The pump 10 is a pump for transferring LNG (liquefied natural gas) at an LNG receiving terminal. The condition of the pump 10 deteriorates due to human use and the passage of time. As the deterioration progresses, maintenance becomes necessary. A human checks the vibration data of the pump 10 from the data storage device 30, and monitors the presence or absence of an abnormality in the pump 10 based on the vibration data and information other than the vibration data. Information other than the vibration data is, for example, data on the operating sound of the pump 10 and the flow rate of LNG. Then, if a human determines that an abnormality has occurred in the pump 10, the pump 10 is disassembled by the human. At this time, if there is actually an abnormality in the pump 10, the pump 10 is maintained by the human.

[0011] The sensor 20 is installed inside the pump 10. The sensor 20 outputs information on vibrations caused by operation of the pump 10 as vibration data to the data storage device 30. The vibration data is an example of operation data. Note that the sensor 20 does not have to detect vibrations, and may be a sensor that detects other information of the pump 10, such as the flow rate.

[0012] The data storage device 30 assigns identification information to the vibration data acquired from the sensor 20 and stores the data in advance as past vibration data. The data storage device 30 also assigns identification information to the vibration data acquired from the inference device 200 and stores the data in advance as past vibration data.

[0013] The identification information is information for identifying whether the vibration data is normal data or abnormal data of the pump 10. Furthermore, the identification information includes information for identifying whether the vibration data is false positive data, which will be described later.

[0014] Here, the normal data is vibration data indicating that the pump 10 is normal. For example, the normal data is vibration data obtained by the data storage device 30 from the sensor 20 when the pump 10 is actually normal as a result of disassembling the pump 10, among the vibration data obtained by the data storage device 30 from the inference device 200. For another example, the normal data is vibration data obtained by the data storage device 30 from the sensor 20 or the inference device 200 during a period before the data storage device 30 obtains vibration data obtained when the pump 10 is actually normal. The period before the data storage device 30 obtains vibration data obtained when the pump 10 is actually normal is, for example, a period from when the pump 10 is disassembled and serviced once until it is disassembled again.

[0015] Furthermore, the abnormal data is vibration data that indicates an abnormality in the pump 10. For example, the abnormal data is vibration data acquired from the sensor 20 when the pump 10 is actually found to be abnormal as a result of disassembling the pump 10.

[0016] The learning device 100 is a learning device that trains the learning model 150 based on the features of vibration data. For example, the learning device 100 trains the learning model 150 to learn the features of vibration intensity and vibration frequency based on the features of the vibration data.

[0017] Next, as shown in FIG. 1(b), in the inference stage, the anomaly detection system 1 has a pump 10 and a sensor 20. The anomaly detection system 1 also has an inference device 200. The inference device 200 acquires inference target data from the sensor 20. The inference target data is vibration data to which no identification information is attached and in a state in which it is unknown whether or not the data indicates an abnormality in the pump. The inference target data is an example of operation data to be inferred. The inference device 200 predicts whether or not the pump 10 is abnormal using the trained learning model 150 based on the inference target data. Then, the inference device 200 outputs the prediction result to a human terminal or the like. If the human determines that maintenance of the pump 10 is necessary according to this output result, the pump 10 is maintained by the human.

[0018] 2(a) and (b) are diagrams showing the functional configuration of learning device 100 or inference device 200. Fig. 2(a) is a diagram showing the functional configuration of learning device 100. Fig. 2(b) is a diagram showing the functional configuration of inference device 200.

[0019] As shown in FIG. 2(a), the learning device 100 has an acquisition unit 110 that acquires vibration data from the data storage device 30. The learning device 100 also has a classification unit 120 that classifies the vibration data into normal data or abnormal data based on identification information included in the vibration data. The learning device 100 also has a pre-processing unit 130 that processes the vibration data into a format that can be input to the learning model 150 and generates learning data. The learning device 100 also has an input unit 140 that inputs the learning data to the learning model 150. The learning device 100 also has the learning model 150 that is realized by a neural network or the like. The learning device 100 also has a setting unit 160 that sets a threshold value for the degree of abnormality.

[0020] The classification unit 120 classifies the vibration data as normal data when the identification information indicates normal data, and classifies the vibration data as abnormal data when the identification information indicates abnormal data.

[0021] The input unit 140 inputs, as initial learning data, a plurality of pieces of vibration data acquired by the data storage device 30 from the sensor 20, among the vibration data processed by the preprocessing device 130, to the learning model 150. The initial learning data is data for the learning model 150 to perform initial learning. When there are a plurality of pumps 10 and inference devices 200, in addition to the vibration data acquired from the inference device of one pump 10, vibration data acquired from the inference device of another pump 10 may be used as the initial learning data. In addition to the vibration data acquired from one pump 10, vibration data acquired from another pump 10 may be used as the initial learning data.

[0022] Furthermore, the input unit 140 inputs, as re-learning learning data, a plurality of vibration data acquired by the data storage device 30 from the inference device 200, among the vibration data acquired from the data storage device 30 and processed by the pre-processing device 130, to the learning model 150. The re-learning learning data is data for re-learning by the learning model 150. Note that, as the re-learning learning data, a plurality of vibration data acquired by the data storage device 30 from the sensor 20 may be input to the learning model 150 as the re-learning learning data.

[0023] The learning model 150 performs initial learning by inputting initial learning data to the input unit 140. The initial learning is learning that is first performed by the learning model 150 that has not yet performed learning. When the initial learning is completed, the learning model 150 becomes the learned learning model 150.

[0024] The trained learning model 150 performs re-learning by inputting re-learning learning data to the input unit 140. Re-learning is learning that the trained learning model 150 performs again after a predetermined time has elapsed since the previous learning.

[0025] The trained learning model 150 outputs an anomaly degree for the training data. The anomaly degree is a quantitative index that represents the degree of failure or deterioration of the pump 10.

[0026] The setting unit 160 grasps the degree of anomaly for the learning data output by the trained learning model 150, and sets a threshold value for the degree of anomaly. The threshold value for the degree of anomaly is a value for classifying the data to be inferred as normal data or abnormal data. In addition, the setting unit 160 sets hyperparameters, etc., which determine the number of times of learning, etc., in the learning model 150 according to human instructions, etc.

[0027] 2(b), the inference device 200 has an inference target data acquisition unit 210 that acquires inference target data from the sensor 20. The inference device 200 also has a trained learning model acquisition unit 220 that acquires the trained learning model 150 from the learning device 100. The inference device 200 also has a pre-processing unit 230 that processes the inference target data so that it can be input to the learning model. The inference device 200 also has an input unit 240 that inputs the processed inference target data to the trained learning model 150. The inference device 200 also has a determination unit 250 that determines whether or not the pump 10 is abnormal based on the degree of abnormality output from the trained learning model 150.

[0028] The preprocessing unit 230 performs the same processing as the preprocessing unit 130 of the learning device 100 on the inference target data.

[0029] When the input unit 240 inputs the inference target data processed by the pre-processing unit 230 to the trained learning model 150, the trained learning model 150 outputs the degree of anomaly based on the similarity between the inference target data and the training data.

[0030] The judgment unit 250 judges the degree of abnormality output by the trained learning model 150. If the degree of abnormality output from the trained learning model 150 exceeds a threshold, that is, if the learning model 150 predicts that the pump 10 is abnormal, the judgment unit 250 judges that the pump 10 is abnormal. If the degree of abnormality output from the trained learning model 150 does not exceed a threshold, that is, if the learning model 150 predicts that the pump 10 is not abnormal, the judgment unit 250 judges that the pump 10 is normal. The judgment unit 250 outputs the judgment result to a terminal or the like of a person. Note that the judgment of the degree of abnormality by the judgment unit 250 may be performed by the trained learning model 150, or the judgment result of the trained learning model 150 may be output to a terminal or the like of a person.

[0031] Fig. 3 shows the hardware configuration of data storage device 30, learning device 100, and inference device 200. As shown in Fig. 3, the computer realizing data storage device 30, learning device 100, and inference device 200 has a processing unit 201 that executes digital arithmetic processing according to a program. The computer realizing data storage device 30, learning device 100, and inference device 200 also has an information storage device 202 that stores information, and a network interface 203 that realizes communication via a LAN (=Local Area Network) cable or the like.

[0032] The processing unit 201 is configured by a computer. The processing unit 201 has a CPU (=Central Processing Unit) 211 as an example of a processor that executes various processes. The processing unit 201 also has a ROM (Read Only Memory) 212 in which a program is stored, and a RAM (Random Access Memory) 213 used as a work area. The information storage device 202 is realized by an existing device such as a hard disk drive, a semiconductor memory, or a magnetic tape. The processing unit 201, the information storage device 202, and the network interface 203 are connected via a bus 206 or a signal line (not shown).

[0033] The program executed by CPU211 may be provided to study device 100 in a state where it is stored in a computer-readable recording medium such as a magnetic recording medium (such as a magnetic tape or a magnetic disk), an optical recording medium (such as an optical disk), a magneto-optical recording medium, or a semiconductor memory. In addition, the program executed by CPU211 may be provided to study device 100 using a communication means such as the Internet.

[0034] In this embodiment, each process is executed by an arbitrary computer. The arbitrary computer may be realized as a processor as hardware, a program as software, or a combination of both. The arbitrary computer may be a general-purpose computer, a computer for specific purposes, a workstation, or any other system capable of executing each process.

[0035] The processor is configured to execute various processes in cooperation with the program. The processor can function as each unit or each means in the present embodiment. The order in which the processor executes the processes is not limited to the order described in the present embodiment, and can be changed as necessary.

[0036] The processor can be configured with one or more pieces of hardware. The type of hardware that configures the processor is not limited to a specific type. For example, the processor may be the CPU 211, an MPU (Micro Processing Unit), a programmable logic device such as an FPGA (Field Programmable Gate Array), a dedicated circuit for executing specific processing such as an ASIC (Application Specific Integrated Circuit), a GPU (Graphic Processing Unit), or hardware such as an NPU (Neural Processing Unit).

[0037] A processor is not limited to a combination of multiple pieces of the same type of hardware, but may also be configured by a combination of multiple pieces of different types of hardware. When multiple pieces of hardware are configured to execute one or more processes of a certain processor, the multiple pieces of hardware may exist in devices physically separated from each other, or may exist in the same device. The hardware is configured by an electric circuit (circuitry) combining circuit elements such as semiconductor elements. In any of the embodiments, the order of execution of each process by the processor is not limited to the order described in each embodiment, and can be changed as necessary.

[0038] The program may be software such as firmware or microcode. The program may be, for example, a group of program modules. Each function constituting the group of program modules may be implemented by a processor configured to execute each function. The program in each embodiment may be a program code or a plurality of code segments stored in one or more non-transitory computer-readable media (e.g., semiconductor memory, magnetic or optical storage medium, or other storage).

[0039] The program may be stored in multiple non-transitory computer-readable media that are physically separate from each other. The program code or multiple code segments may be represented as any combination of a procedure, a function, a subprogram, a routine, a subroutine, a module, a software package, a class, an instruction, a data structure, or a program statement. The program code or multiple code segments may be connected to other code segments or hardware circuits by transmitting or receiving information, data, arguments, parameters, or memory contents.

[0040] Figures 4(a) to (c) are diagrams showing the characteristics of vibration data according to time series. Figure 4(a) is vibration data acquired from the sensor 20 immediately after maintenance of the pump 10, Figure 4(b) is vibration data acquired in an intermediate period, and Figure 4(c) is vibration data acquired immediately before maintenance of the pump 10.

[0041] As shown in FIG. 4(a), immediately after maintenance by a human, when there is little wear or deterioration of the equipment, sensor 20 outputs vibration data that has fewer peaks and shows a simpler structure than vibration data at other times.

[0042] As shown in FIG. 4(b), in an intermediate period when wear and deterioration of the equipment has progressed due to human use and the passage of time compared to immediately after maintenance, sensor 20 outputs vibration data with increased and expanded peaks and more complex structure than the vibration data immediately after maintenance.

[0043] 4(c) shows a period immediately prior to maintenance when the equipment has become more worn and deteriorated due to human use and the passage of time compared to the intermediate period, and when the pump 10 shows an abnormality. During this period immediately prior to maintenance, the sensor 20 outputs vibration data with a more complicated structure and with an increased peak compared to the vibration data during the intermediate period.

[0044] FIG. 5 is a flowchart showing the flow of processing of the learning device 100. First, the acquisition unit 110 of the learning device 100 acquires past vibration data from the data storage device 30 (step 601). Next, the classification unit 120 classifies normal data and abnormal data from the acquired vibration data (step 602). Then, the preprocessing unit 130 performs preprocessing of the vibration data and converts the vibration data into a format that can be input to the learning model 150 (step 603). After that, the input unit 140 inputs the vibration data processed by the preprocessing unit 130 to the learning model 150 as learning data (step 604). Then, the learning model 150 performs learning based on the learning data (step 605). Then, the setting unit 160 sets a threshold value of the degree of abnormality based on the degree of abnormality output by the learning model 150 in response to the input of the learning data (step 606).

[0045] FIG. 6 is a flowchart showing the flow of processing of the inference device 200. First, the inference target data acquisition unit 210 of the inference device 200 acquires vibration data as inference target data from the sensor 20 (step 701). Then, the preprocessing unit 230 performs preprocessing of the vibration data and converts the vibration data into a format that can be input to the learning model 150 (step 702). After that, the learning model acquisition unit 220 acquires the learned learning model 150 from the learning device 100 (step 703). At this time, the learning model acquisition unit 220 acquires information such as a set threshold value of the abnormality degree and hyperparameters. Then, the input unit 240 inputs the vibration data processed by the preprocessing unit 230 to the learning model 150 as inference target data (step 704). The learned learning model 150 predicts the abnormality degree of the inference target data (step 705). The judgment unit 250 performs an abnormality judgment based on the set threshold value of the abnormality degree and outputs the abnormality judgment (step 706).

[0046] FIG. 7 is a diagram for explaining the learning data to be learned by the learning model 150 in this embodiment. In equipment used in a factory, such as the pump 10, abnormalities occur rarely, and there is less abnormal data indicating an abnormality in the pump 10 than normal data indicating the normality of the pump 10. In addition, the abnormal data is diverse and cannot be covered in its entirety. Therefore, in this embodiment, only the normal data of the pump 10 is used as learning data to learn the learning model 150 and model normal / abnormal. Note that, although only normal data is used as learning data in this embodiment, abnormal data may also be used as learning data.

[0047] Fig. 8 is a diagram illustrating false positive data. As shown in Fig. 8, the vibration data of the pump 10 includes normal data, which is vibration data when the pump 10 is normal, and abnormal data, which is vibration data when the pump 10 is abnormal.

[0048] Among the normal data, the true negative data is data that indicates that the pump 10 is actually normal and that a human or the learning model 150 indicates that the pump 10 is normal. For example, the true negative data is a piece of vibration data in a past overhaul and maintenance case of the pump 10, in which the learning model 150 predicts that the pump 10 is normal from a piece of vibration data, a human disassembles the pump 10, and the pump 10 is actually normal. Also, for example, the true negative data is a piece of vibration data in which a human predicts that the pump 10 is normal from a piece of vibration data or information other than the piece of vibration data, a human disassembles the pump 10, and the pump 10 is actually normal.

[0049] Among the normal data, false positive data is data in which a human or the learning model 150 indicates that the pump 10 is abnormal, even though the pump 10 is actually normal. For example, false positive data is a piece of data in a past overhaul and maintenance case of the pump 10, in which the learning model 150 predicts that the pump 10 is abnormal based on a piece of vibration data, and the pump 10 is disassembled by a human, but the pump 10 is actually normal. For example, false positive data is a piece of vibration data in which a human predicts that the pump 10 is abnormal based on a piece of vibration data or information other than the piece of vibration data, and the pump 10 is disassembled by a human, but the pump 10 is actually normal.

[0050] Among the abnormal data, false negative data is data in which a human or the learning model 150 indicates that the pump 10 is normal, when in fact the pump 10 is abnormal. For example, false negative data is data in a past overhaul and maintenance case of the pump 10, in which the learning model 150 predicts that the pump 10 is normal from one piece of data, and a human disassembles the pump 10, but in fact the pump 10 is abnormal. Among the abnormal data, false negative data is data in which a human predicts that the pump 10 is normal from one piece of vibration data or information other than the one piece of vibration data, and a human disassembles the pump 10, but in fact the pump 10 is abnormal.

[0051] Among the abnormal data, true positive data is data in which the pump 10 is actually abnormal and a human or the learning model 150 indicates that the pump 10 is abnormal. This is data in which, in a past overhaul and maintenance case of the pump 10, the learning model 150 predicts that the pump 10 is abnormal from one piece of data, a human disassembles the pump 10, and the pump 10 is actually abnormal. Moreover, true positive data is data in which, in a past overhaul and maintenance case of the pump 10, a human predicts that the pump 10 is abnormal from information other than the one piece of data or one piece of vibration data, a human disassembles the pump 10, and the pump 10 is actually abnormal.

[0052] Note that data obtained when a human predicts that the pump 10 is abnormal and disassembles the pump 10, but the pump 10 is actually normal, may be referred to as human false positive data. Also, data obtained when the learning model 150 predicts that the pump 10 is abnormal and disassembles the pump 10, but the pump 10 is actually normal, may be referred to as learning model 150 false positive data.

[0053] Here, vibration data obtained long after the last maintenance, such as the vibration data in FIG. 4(c), becomes complicated in structure as the equipment condition changes, making it difficult for the learning model 150 to classify whether the data is abnormal or not. Such vibration data may be erroneously determined by the trained learning model 150 to be abnormal, even though the pump 10 is actually normal. Then, the data storage device 30 (see FIG. 1) assigns identification information to the erroneously determined vibration data as false positive data of the learning model 150.

[0054] In addition, vibration data obtained long after the last maintenance, such as the vibration data shown in Fig. 4(c), becomes complicated as the equipment condition changes, making it difficult for humans to classify whether the data is abnormal or not. Such vibration data may be erroneously determined by humans to be abnormal, even though the pump 10 is actually normal.

[0055] However, when the sensor 20 acquires vibration data such as that in Fig. 4(c), a human checks the vibration data as well as other information besides the vibration data, such as the operating sound of the pump 10 and data on the flow rate of the pump 10. The human then checks the vibration data together with other information to comprehensively predict whether the pump 10 is abnormal or not. As a result, if the human erroneously determines that the pump 10 is abnormal when in fact the pump 10 is normal, the vibration data used for the prediction is given identification information by the data storage device 30 (see Fig. 1) as false positive data of the human.

[0056] Also, assume that the sensor 20 acquires vibration data during a period when the equipment is less worn, such as the vibration data in FIGS. 4(a) and (b). In this case, even if a human predicts that the vibration data indicates that the pump 10 is normal, if the human predicts that information other than the vibration data indicates that the pump 10 is abnormal, the human may disassemble the pump 10. If the pump 10 is actually normal as a result of disassembling the pump 10, the data storage device 30 (see FIG. 1) assigns identification information to the vibration data used for the prediction as human false positive data. Note that the data used for this prediction may be treated as data similar to false positive data, which will be described later, since it is data that a human would check to predict that the pump 10 is normal based on the data alone, but which may lead to an erroneous judgment.

[0057] In this way, a human predicts whether or not the pump 10 is abnormal based on the vibration data and information other than the vibration data. Therefore, a human's prediction of whether or not the pump 10 is abnormal may differ from a prediction of whether or not the pump 10 is abnormal made by the trained learning model 150 based on the vibration data. In addition, a prediction may differ between a case where a human checks only the vibration data and a case where a human checks the vibration data and information other than the vibration data.

[0058] For example, the learning model 150 trained based on one vibration data may predict that the pump 10 is abnormal, and a human being may predict that the pump 10 is normal based on the one vibration data and information other than the one vibration data. Also, for example, the learning model 150 trained based on one vibration data may predict that the pump 10 is normal, and a human being may predict that the pump 10 is abnormal based on the one vibration data and information other than the one vibration data.

[0059] Also, for example, when a human checks only one piece of vibration data, the human may predict that the pump 10 is abnormal, whereas when the human checks the one piece of vibration data and information other than the one vibration data, the human may predict that the pump 10 is normal. Also, for example, when a human checks only one piece of vibration data, the human may predict that the pump 10 is normal, whereas when the human checks the one piece of vibration data and information other than the one vibration data, the human may predict that the pump 10 is abnormal.

[0060] In addition, in the false positive data of humans, characteristics of information other than the vibration data may appear as vibration characteristics. For example, characteristics of the operating sound of the pump 10 that may lead a human to erroneously determine that the pump 10 is abnormal may appear as vibration characteristics in the vibration data.

[0061] 9(a) and (b) are diagrams showing initial learning or re-learning of the learning model 150. Fig. 9(a) is a diagram showing initial learning of the learning model 150. Fig. 9(b) is a diagram showing re-learning of the learned learning model 150.

[0062] In FIG. 9(a), the actual normal data group refers to a plurality of normal data that are classified as normal data by the classification unit 120 from among the vibration data group acquired from the data storage device and processed by the preprocessing unit 130. Here, the normal data is not limited to data obtained when the pump 10 is actually disassembled, but also includes data that is highly likely to be normal data even though the pump 10 is not actually disassembled. For example, the normal data includes data obtained from the sensor 20 from after the maintenance of the pump 10 until the reference date, where the date on which false positive data of a human being is acquired from the sensor 20 is set as the reference date. In addition, the normal data also includes data obtained from the sensor 20 before the reference time described later.

[0063] Furthermore, among the actual normal data group, the data similar to human false positives is vibration data similar to human false positive data. For example, the data similar to human false positives is vibration data similar to a feature such as the intensity of the frequency of the false positive data. For example, the data similar to human false positives is vibration data acquired from the sensor 20 around the reference date, with the date on which the human false positive data was acquired from the sensor 20 being set as the reference date. The vicinity of the reference date is, for example, one week before the reference date. Furthermore, the operation time of the pump 10 when the human false positive data was acquired from the sensor 20 may be set as the reference time, and the vibration data acquired from the sensor 20 around the reference time may be used as data similar to human false positive data. For example, if the operation time of the pump 10 when the human false positive data was acquired from the sensor 20 is 200 hours, the vibration data acquired from the sensor 20 when the pump 10 was operated for 175 hours may be used as data similar to human false positive data. Note that the operation time of the pump 10 is the time measured after one maintenance.

[0064] Furthermore, the data that is not similar to the human false positive data in the actual normal data group is vibration data that is not similar to the human false positive data. For example, it is data in which the feature amount of the false positive data, such as the frequency intensity, is not similar to the feature amount of the vibration data. Also, for example, it is vibration data acquired outside the vicinity of the reference date, where the date on which the human false positive data was acquired from the sensor 20 is set as the reference date.

[0065] The input unit 140 preferentially inputs human false positive data as initial learning data to the learning model 150. For example, the amount of human false positive data input is made greater than the amount of data other than human false positive data. Note that only human false positive data may be input to the learning model 150.

[0066] In addition, the input unit 140 may preferentially input human false positive data and data similar to human false positives to the learning model 150 as initial learning data. For example, the amount of human false positive data and data similar to human false positives input may be greater than data other than human false positive data and data similar to human false positives. Furthermore, the input ratio of false positive data and data similar to false positives in the input learning data may be changed according to time series. For example, the date on which the false positive data was acquired from the sensor 20 may be set as a reference date, and the input ratio of false positive data and data similar to false positives in the input learning data may be decreased as the reference date goes back. Note that only data similar to human false positives may be input to the learning model 150.

[0067] By inputting the human false positive data of the input unit 140 with priority, the learning model 150 performs initial learning with emphasis on the characteristics of the human false positive data. The characteristics of the false positive data of the learning model 150 are vibration characteristics that do not appear in data other than the human false positive data. Examples of vibration characteristics include the frequency and intensity of vibration, and the learning model 150 learns where the vibration frequency occurs, how strong the vibration is, and so on. The learning model 150 then becomes a trained learning model 150 that has been trained with emphasis on the characteristics of the human false positive data. This trained learning model 150 is able to output the degree of anomaly based on the similarity between the characteristics of the inference target data and the characteristics of normal data other than the human false positive data, as well as the similarity between the characteristics of the inference target data and the characteristics of the human false positive data, in response to the input of the inference target data.

[0068] In FIG. 9(b), the additional normal data group is the false positive data of the learning model 150 and data similar to the false positive data of the learning model 150. As described above, the false positive data of the learning model 150 is inference target data in the case where the learned learning model 150 judges the pump 10 to be abnormal from the inference target data, but the pump is actually normal. Furthermore, the data similar to the false positive of the learning model 150 is additional data acquired around the date on which the false positive data of the learning model 150 is acquired from the sensor 20 as a reference date. For example, the data similar to the false positive of the learning model 150 is additional data acquired from the sensor 20 up to two or three days before the reference date. Note that, as described above, the operation time of the pump 10 may be the reference time. Note that, as the additional normal data group, in addition to the false positive data of the learning model 150 and data similar to the false positive data of the learning model 150, an actual normal data group may be added.

[0069] The input unit 140 preferentially inputs the false positive data of the learning model 150 as re-learning learning data to the learned learning model 150. For example, the amount of false positive data of the learning model 150 to be input is made greater than the amount of data other than the false positive data of the learning model 150. Note that only the false positive data of the learning model 150 may be input to the learning model 150.

[0070] Furthermore, the input unit 140 may preferentially input the false positive data of the learning model 150 and data similar to the false positive data of the learning model 150 to the already-learned learning model 150 as the learning data for re-learning. For example, the amount of the false positive data of the learning model 150 and data similar to the false positive data of the learning model 150 input may be greater than the amount of data other than the false positive data of the learning model 150. Furthermore, the input ratio of the false positive data of the learning model 150 and data similar to the false positives of the learning model 150 in the input learning data may be changed according to the time series. For example, the date on which the false positive data of the learning model 150 was acquired from the sensor 20 may be set as the reference date, and the input ratio of the false positive data of the learning model 150 and data similar to the false positives of the learning model 150 may be decreased as the reference date goes back. Note that only data similar to the false positive data of the learning model 150 may be input to the learning model 150.

[0071] By inputting the false positive data of the learning model 150 of the input unit 140 with priority given to the false positive data of the learning model 150, the learned learning model 150 performs re-learning with emphasis on the features of the false positive data of the learning model 150. The features of the false positive data of the learning model 150 are vibration features that do not appear in data other than the false positive data of the learning model 150. Examples of vibration features include the frequency and intensity of vibration, and the learned learning model 150 learns where the vibration frequency occurs, how strong the vibration is, and the like. The learned learning model 150 becomes a learned learning model 150 that has been re-learned with emphasis on the false positive data of the learning model 150. This learned learning model 150 can output the degree of anomaly based on the similarity between the features of the inference target data and the features of the false positive data of the learning model 150 in response to the input of inference target data, in addition to the similarity between the features of the inference target data and the features of normal data other than the false positive data of the learning model 150.

[0072] FIG. 10 is a graph showing the degree of anomaly output by the trained learning model 150 according to a time series. In the graph of FIG. 10, the vertical axis shows the degree of anomaly, and the horizontal axis shows the operation time. The operation time shows the operation time of the pump 10 after the pump 10 has been maintained once. In addition, IVA shows the period immediately after maintenance, IVB shows an intermediate period, and IVC shows the period immediately before maintenance. It is assumed that the threshold value of the degree of anomaly is 30.

[0073] In the IVA of Figure 10, for example, when the inference device 200 acquires vibration data such as that shown in Figure 4(a) as inference target data, the trained learning model 150 judges the degree of abnormality to be 1 since the characteristics of the inference data are similar to the characteristics of the vibration data of normal data including false positive data, and the judgment unit 250 judges that the pump 10 is normal.

[0074] In IVB of Fig. 10, for example, it is assumed that the inference device 200 acquires vibration data as shown in Fig. 4(b) as inference target data. In this case, the trained learning model 150 judges the abnormality level to be 17.5 since the characteristics of the inference target data are somewhat similar to the characteristics of the vibration data of normal data including false positive data, and the judgment unit 250 judges that the pump 10 is normal. "Somewhat similar" indicates that the similarity is less than the stage in IVA, but that the pump 10 is within the normal range.

[0075] In the IVC of FIG. 10, for example, it is assumed that the inference device 200 acquires vibration data as shown in FIG. 4(c) as inference target data. In this case, the trained learning model 150 judges the abnormality level to be 40 because the characteristics of the inference target data are not similar to the characteristics of the vibration data of normal data including false positive data, and the judgment unit 250 judges that the pump 10 is abnormal. Then, a human checks the detection result by the judgment unit 250, disassembles the pump 10, and repairs the pump 10 if an abnormality actually occurs. On the other hand, if an abnormality does not actually occur despite disassembling the pump 10, the vibration data as shown in FIG. 4(c) is assigned identification information as false positive data of the learning model 150.

[0076] FIG. 11 is a diagram showing the diversion of the trained learning model 150. In FIG. 11, the equipment A11, the equipment B12, and the equipment C13 each show a different pump 10. The trained learning model 150 acquires vibration data from the equipment A11 and the equipment B12, and learns using the vibration data as learning data. Then, the learning is completed, and the trained learning model 150 is obtained. This trained learning model 150 acquires inference target data from the equipment A11 and the equipment B12, and becomes capable of detecting an abnormality in the pump 10. Furthermore, this trained learning model 150 acquires inference target data for the equipment C13, and becomes capable of detecting an abnormality. The equipment C13 is an equipment for which the learning model 150 has not acquired learning data in the first place. In this way, the trained learning model 150 can be applied to the equipment that has been trained, and can also be diverted to equipment that has not been trained in the first place. [Explanation of symbols]

[0077] 10... pump, 20... sensor, 30... data storage device, 100... learning device, 200... inference device

Claims

1. A processor is provided. The processor, Acquire multiple data from the anomaly detection target, The plurality of data, including data in which the object of anomaly detection is erroneously determined to be abnormal by a human being when the object is actually normal, are trained as training data in a learning model. The anomaly detection system according to the present invention is characterized in that

2. The anomaly detection system according to claim 1, characterized in that the data in which a human has mistakenly determined that the object of anomaly detection is normal but is in fact abnormal is data in which a human has mistakenly determined that the object of anomaly detection is abnormal, including information other than the mistakenly determined data.

3. The anomaly detection system according to claim 2, characterized in that the erroneously determined data is vibration data obtained from vibration of the object of anomaly detection, and the information other than the erroneously determined data is information other than the vibration data.

4. A processor is provided. The processor, Read the driving data to be inferred, A learning model is obtained by learning a plurality of driving data including data in which a human erroneously judges an abnormality to be present when the abnormality detection target is actually normal, and Using the acquired learning model, it is detected whether or not the driving data of the inference target indicates an abnormality of the abnormality detection target. The anomaly detection system according to the present invention is characterized in that

5. 5. The anomaly detection system according to claim 4, wherein the learning model can be diverted to an anomaly detection target different from the anomaly detection target.

Images