Anomaly detection system

Abstract

The objective is to provide an anomaly detection system using a learning model with improved predictive performance compared to a system that uses false positive data only for retraining the learning model. [Solution] The anomaly detection system includes a processor, which acquires multiple operational data from the anomaly detection target, and uses these operational data, including false positive data indicating that the anomaly detection target is abnormal when it is actually normal, to train a learning model as initial training data.

Description

Technical Field

[0001] The present invention relates to an abnormality detection system.

Background Art

[0002] Patent Document 1 aims to provide a learning data selection device capable of selecting learning data that can reduce false detection of a failure detection target. Patent Document 1 discloses that, in order to solve this problem, sensor data related to false positive detection data is selected as learning data used for re-learning of a learning model.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] For example, when using a learning model for data acquired from an abnormality detection target such as production equipment, data immediately after maintenance is often simple in structure and easy for the learning model to detect abnormalities. On the other hand, data after a long time has passed since the previous maintenance may have a complicated structure as the equipment state changes, and may be difficult for the learning model to detect abnormalities. Therefore, there is a method of improving the prediction performance by training false positive data indicating that the abnormality detection target is abnormal although it is normal as learning data used for re-learning of the learning model. However, even if re-learning can be performed by this method, if the quality of the original initial model is low, it may have an adverse effect on the prediction performance of the model after re-learning. An object of the present invention is to provide an abnormality detection system using a learning model with higher prediction performance as compared with the case of using only false positive data for re-learning of the learning model.

Means for Solving the Problems

[0005] The invention described in claim 1 is an anomaly detection system comprising a processor, wherein the processor acquires a plurality of operating data from an anomaly detection target, and uses the plurality of operating data, which includes false positive data indicating that the anomaly detection target is abnormal even though it is normal, to train a learning model as initial training data. The invention described in claim 2 is an anomaly detection system according to claim 1, characterized in that the plurality of operating data includes the false positive data and data other than the false positive data, and the false positive data is given priority among the plurality of operating data to be used as initial training data to train the learning model. The invention described in claim 3 is an anomaly detection system according to claim 2, characterized in that only the false positive data is used to train the learning model as the initial training data. The invention described in claim 4 is an anomaly detection system according to claim 2, characterized in that the data other than the false positive data includes data similar to the false positive data, and the learning model is trained preferentially using the false positive data and / or data similar to the false positive data from among the plurality of operating data as the initial training data. The invention described in claim 5 is an anomaly detection system comprising a processor, wherein the processor reads the driving data to be inferred, acquires a learning model which has been trained on a plurality of past driving data, including false positive data indicating that the anomaly detection target is abnormal even though it is normal, as initial learning data, and uses the acquired learning model to detect whether or not the driving data to be inferred indicates an anomaly in the anomaly detection target. The invention described in claim 6 is an anomaly detection system according to claim 4, characterized in that the learning model can be adapted for an anomaly detection target different from the anomaly detection target described above. The invention described in claim 7 is an anomaly detection system according to claim 1 or 5, characterized in that the plurality of operating data are vibration data obtained from the vibration of the object to be detected as an anomaly, and the learning model learns the characteristics of the vibration data. [Effects of the Invention]

[0006] This invention provides an anomaly detection system using a learning model that has improved predictive performance compared to the case where false positive data is used only for retraining the learning model. [Brief explanation of the drawing]

[0007] [Figure 1] (a) and (b) are diagrams showing the overall configuration of the anomaly detection system. [Figure 2] (a) and (b) are diagrams showing the functional configuration of a learning device or inference device. [Figure 3] This describes the hardware configuration of the data storage device, learning device, and inference device. [Figure 4] Figures (a) to (c) show the characteristics of vibration data according to the time series. [Figure 5] This is a flowchart showing the processing flow of the learning device. [Figure 6] This is a flowchart showing the processing flow of the inference device. [Figure 7] This diagram explains the training data used to train the learning model. [Figure 8] This is a diagram illustrating false positive data. [Figure 9] (a) and (b) are diagrams showing the initial training or retraining of the learning model. [Figure 10] This graph shows the anomaly score output by a trained model over time. [Figure 11] This diagram shows the reuse of a trained model. [Modes for carrying out the invention]

[0008] Embodiments of the present invention will be described in detail below with reference to the attached drawings. Figures 1(a) and 1(b) show the overall configuration of the anomaly detection system. Figure 1(a) shows the configuration of the anomaly detection system 1 during the learning phase, and Figure 1(b) shows the configuration of the anomaly detection system 1 during the inference phase. The learning phase is the stage in which the learning model 150 of the learning device 100 performs learning, and the inference phase is the stage in which the inference device 200 uses the learned learning model 150 to detect anomalies in the pump 10. The learned learning model 150 will be described later in Figure 2.

[0009] First, as shown in Figure 1(a), in the learning phase, the anomaly detection system 1 has a pump 10, which is an example of an anomaly detection target and is a device whose presence or absence of an anomaly is monitored 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. Furthermore, the anomaly detection system 1 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 uses operating data as learning data to train the learning model 150. Furthermore, the anomaly detection system 1 has an inference device 200 that predicts whether the pump 10 is abnormal or not based on the learning model 150 trained by the learning device 100. The learning model 150 trained by the learning device 100 is sometimes referred to as the trained learning model 150.

[0010] Pump 10 is a pump used to transfer LNG (liquefied natural gas) at an LNG receiving terminal. Pump 10 deteriorates in condition due to human use and the passage of time. As deterioration progresses, maintenance becomes necessary. Humans check the vibration data of pump 10 from the data storage device 30 and monitor for any abnormalities in pump 10 based on the vibration data and other information. Other information includes, for example, the operating sound of pump 10 and data on the flow rate of LNG. If humans determine that there is an abnormality in pump 10, they disassemble pump 10. If there is actually an abnormality in pump 10 at this time, pump 10 is serviced by humans.

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

[0012] The data storage device 30 assigns identification information to the vibration data acquired from the sensor 20 and pre-stores it as past vibration data. Further, the data storage device 30 assigns identification information to the vibration data acquired from the inference device 200 and pre-stores it as past vibration data.

[0013] The identification information is information for identifying whether the vibration data is normal data of the pump 10 or abnormal data of the pump. Further, 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 the normal state of the pump 10. For example, the normal data is vibration data among the vibration data acquired by the data storage device 30 from the sensor 20 when the pump 10 was actually normal as a result of disassembling the pump 10. Also, for example, the normal data is vibration data among the vibration data acquired by the data storage device 30 from the inference device 200 when the pump 10 was actually normal as a result of disassembling the pump 10. Further, for example, the normal data is vibration data acquired by the data storage device 30 from the sensor 20 or the inference device 200 during a period before the data storage device 30 acquired vibration data when the pump 10 was actually normal. The period before the data storage device 30 acquired vibration data when the pump 10 was actually normal is, for example, the period from when the pump 10 was disassembled and maintained once until it was disassembled again.

[0015] Also, the abnormal data is vibration data indicating an abnormality of the pump 10. For example, the abnormal data is vibration data acquired from the sensor 20 when the pump 10 was actually abnormal as a result of disassembling the pump 10.

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

[0017] Next, as shown in Figure 1(b), in the inference stage, the anomaly detection system 1 includes a pump 10 and a sensor 20. The anomaly detection system 1 also includes an inference device 200. The inference device 200 acquires data to be inferred from the sensor 20. The data to be inferred is vibration data that does not have identification information attached and is in a state where it is unknown whether or not it indicates a pump abnormality. The data to be inferred is an example of the operating data to be inferred. Based on the data to be inferred, the inference device 200 predicts whether or not the pump 10 is abnormal using a trained learning model 150. The predicted result is then output to a human terminal or the like. If, based on this output result, the human determines that maintenance of the pump 10 is necessary, the pump 10 is maintained by the human.

[0018] Figures 2(a) and 2(b) show the functional configuration of the learning device 100 or the inference device 200. Figure 2(a) shows the functional configuration of the learning device 100. Figure 2(b) shows the functional configuration of the inference device 200.

[0019] As shown in Figure 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 sorting unit 120 that sorts the vibration data into normal data or abnormal data based on the identification information contained in the vibration data. Furthermore, the learning device 100 has a preprocessing unit 130 that processes the vibration data into a format that can be input into the learning model 150 and generates learning data. Furthermore, the learning device 100 has an input unit 140 that inputs the learning data into the learning model 150. The learning device 100 also has a learning model 150 that is implemented using a neural network or the like. Furthermore, the learning device 100 has a setting unit 160 that sets a threshold for the degree of abnormality.

[0020] The sorting unit 120 sorts vibration data as normal data if the identification information indicates normal data. It also sorts vibration data as abnormal data if the identification information indicates abnormal data.

[0021] The input unit 140 inputs multiple vibration data, which have been processed by the preprocessing unit 130 and acquired by the data storage device 30 from the sensor 20, into the learning model 150 as initial learning data. Initial learning data is data used by the learning model 150 for initial learning. If there are multiple pumps 10 and inference devices 200, vibration data acquired from the inference devices of other pumps 10 may be used as initial learning data in addition to the vibration data acquired from the inference device of one pump 10. Furthermore, vibration data acquired from other pumps 10 may be used as initial learning data in addition to the vibration data acquired from one pump 10.

[0022] Furthermore, the input unit 140 inputs multiple vibration data points acquired by the data storage device 30 from the inference device 200 as training data for relearning, from among the vibration data acquired from the data storage device 30 and processed by the preprocessing unit 130. The training data for relearning is data used by the learning model 150 to perform relearning. Alternatively, multiple vibration data points acquired by the data storage device 30 from the sensor 20 may be input to the learning model 150 as training data for relearning.

[0023] The learning model 150 performs initial training by inputting initial training data from the input unit 140. Initial training is the first training performed by the learning model 150, which has not yet undergone any training. Once initial training is complete, the learning model 150 becomes a trained learning model 150.

[0024] The trained model 150 undergoes relearning by receiving relearning data from the input unit 140. Relearning is the process in which the trained model 150 performs training again after a predetermined time has elapsed since the initial training.

[0025] The trained model 150 outputs an anomaly score for the training data. The anomaly score is a quantitative indicator representing the degree of failure or deterioration of the pump 10.

[0026] The configuration unit 160 grasps the degree of anomaly in the training data output by the trained training model 150 and sets an anomaly threshold. The anomaly threshold is a value used to classify whether the data to be inferred is normal or abnormal. The configuration unit 160 also sets hyperparameters such as the number of training iterations to the training model 150 based on human instructions or the like.

[0027] Furthermore, as shown in Figure 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 a trained learning model 150 from the learning device 100. In addition, the inference device 200 has a pre-processing unit 230 that processes the inference target data so that it can be input into the learning model. The inference device 200 also has an input unit 240 that inputs the processed inference target data into the trained learning model 150. It has a determination unit 250 that determines whether the pump 10 is abnormal or not based on the abnormality score output from the trained learning model 150.

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

[0029] When the input unit 240 receives the data to be inferred, which has been processed by the preprocessing unit 230, for the trained model 150, the trained model 150 outputs an anomaly score based on the similarity between the data to be inferred and the training data.

[0030] The determination unit 250 determines the degree of abnormality output by the trained learning model 150. If the degree of abnormality output by the trained learning model 150 exceeds a threshold, that is, if the learning model 150 predicts that there is an abnormality in the pump 10, the determination unit 250 determines that there is an abnormality in the pump 10. Conversely, if the degree of abnormality output by the trained learning model 150 does not exceed a threshold, that is, if the learning model 150 predicts that there is no abnormality in the pump 10, the determination unit 250 determines that the pump 10 is normal. The determination unit 250 outputs the result of its determination to a terminal such as a human. Note that the determination of the degree of abnormality by the determination unit 250 may be performed by the trained learning model 150, or the result determined by the trained learning model 150 may be output to a terminal such as a human.

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

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

[0033] The program executed by the CPU 211 can be provided to the learning device 100 while stored on a computer-readable recording medium such as a magnetic recording medium (magnetic tape, magnetic disk, etc.), an optical recording medium (optical disk, etc.), a magneto-optical recording medium, or semiconductor memory. Alternatively, the program executed by the CPU 211 may be provided to the learning device 100 using communication means such as the Internet.

[0034] In this embodiment, each process is executed on any computer. This computer may be implemented as a processor (hardware), a program (software), or a combination thereof. The computer may be a general-purpose computer, a computer designed for a specific purpose, a workstation, or any other system capable of executing each process.

[0035] The processor is configured to perform various processes in cooperation with the program. The processor can function as each unit or each means in this embodiment. The execution order of the processes performed by the processor is not limited to the order described in this embodiment and can be changed as needed.

[0036] A processor can be configured with one or more hardware components. The types of hardware that make up a processor are not limited to any particular type. For example, a processor may be a CPU211, an MPU (=Micro Processing Unit), a programmable logic device such as an FPGA (=Field Programmable Gate Array), a dedicated circuit for performing 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 can be configured not only with a combination of multiple hardware of the same type, but also with a combination of multiple hardware of different types. When multiple hardware is configured to execute one or more processes of a processor, the multiple hardware may reside in physically separate devices or in the same device. Hardware is composed of electrical circuits, etc., which are combinations of circuit elements such as semiconductor elements. In any embodiment, the execution order 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 firmware or software such as microcode. The program may also 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. In each embodiment, the program may be program code or multiple code segments stored in one or more non-temporary computer-readable media (e.g., semiconductor memory, magnetic or optical storage media, or other storage).

[0039] A program may be divided and stored on multiple non-temporary computer-readable media located on devices that are physically separated from each other. Program code and multiple code segments may be represented by any combination of procedures, functions, subprograms, routines, subroutines, modules, software packages, classes, instructions, data structures, and program statements. Program code and multiple code segments may be connected to other code segments or hardware circuits by sending and receiving information, data, arguments, parameters, or memory contents.

[0040] Figures 4(a) to 4(c) show the characteristics of vibration data over time. Figure 4(a) shows vibration data acquired from sensor 20 immediately after maintenance of pump 10, Figure 4(b) shows vibration data acquired during the intermediate period, and Figure 4(c) shows vibration data acquired immediately before maintenance of pump 10.

[0041] As shown in Figure 4(a), immediately after maintenance, when the equipment has not yet been serviced and there is little wear or deterioration, the vibration data output from the sensor 20 shows fewer peaks and a simpler structure compared to vibration data from other periods.

[0042] As shown in Figure 4(b), during the intermediate period, when the equipment has worn down and deteriorated due to human use and the passage of time compared to immediately after maintenance, the vibration data output from sensor 20 shows an increase in peaks and a greater complexity of structure compared to the vibration data immediately after maintenance.

[0043] Figure 4(c) shows the period immediately before maintenance, when the equipment has worn down and deteriorated further due to human use and the passage of time compared to the intermediate period, and when pump 10 shows abnormalities. In this period immediately before maintenance, the vibration data output from sensor 20 has increased peaks and increased in intensity and a more complex structure compared to the vibration data of the intermediate period.

[0044] Figure 5 is a flowchart showing the processing flow 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 sorting unit 120 sorts the acquired vibration data into normal data and abnormal data (step 602). Then, the preprocessing unit 130 preprocesses the vibration data and converts it into a format that can be input into the learning model 150 (step 603). After that, the input unit 140 inputs the vibration data processed by the preprocessing unit 130 as learning data into the learning model 150 (step 604). Then, the learning model 150 performs learning based on the learning data (step 605). Finally, the setting unit 160 sets an abnormality threshold based on the abnormality level output by the learning model 150 based on the input of the learning data (step 606).

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

[0046] Figure 7 illustrates the training data used to train the learning model 150 in this embodiment. Equipment used in factories, such as the pump 10, rarely experiences malfunctions, and there is less abnormal data indicating a malfunction of the pump 10 than normal data indicating a normal pump 10. Furthermore, the abnormal data is diverse and cannot be comprehensively covered. Therefore, in this embodiment, only normal data of the pump 10 is used as training data to train the learning model 150 and model whether it is normal or abnormal. In this embodiment, only normal data is used as training data, but abnormal data may also be used as training data.

[0047] Figure 8 illustrates false positive data. As shown in Figure 8, the vibration data of pump 10 includes normal data, which is the vibration data when pump 10 is functioning normally, and abnormal data, which is the vibration data when pump 10 is malfunctioning.

[0048] Furthermore, among the normal data, true negative data is data that indicates that pump 10 is actually normal and that the human or learning model 150 is normal. For example, true negative data is the vibration data from a past disassembly and maintenance case of pump 10 where the learning model 150 predicted that pump 10 was normal based on one vibration data, a human disassembled pump 10, and pump 10 was actually normal. Alternatively, true negative data is the vibration data from a first vibration data or other information where a human predicted that pump 10 was normal, a human disassembled pump 10, and pump 10 was actually normal.

[0049] Furthermore, among the normal data, false positive data indicates that the pump 10 is actually normal, but the human or the learning model 150 is abnormal. For example, a false positive data point is a single vibration data point from a past disassembly and maintenance case of the pump 10, where the learning model 150 predicted that the pump 10 was abnormal based on a single vibration data point, and a human disassembled the pump 10, but it was actually normal. Alternatively, a false positive data point is a single vibration data point from a single vibration data point or other information, where a human predicted that the pump 10 was abnormal, and a human disassembled the pump 10, but the pump 10 was actually normal.

[0050] Furthermore, among the abnormal data, false negative data is data that indicates that pump 10 is normal, even though it is actually abnormal, according to the human or the learning model 150. For example, false negative data is data from a past disassembly and maintenance case of pump 10 where the learning model 150 predicted that pump 10 was normal based on one piece of data, and a human disassembled pump 10, but in reality pump 10 was abnormal. Also, for example, among the abnormal data, false negative data is data where a human predicted that pump 10 was normal based on one piece of vibration data or information other than one piece of vibration data, and a human disassembled pump 10, but in reality pump 10 was abnormal.

[0051] Furthermore, among the abnormal data, true positive data indicates that pump 10 is actually abnormal, and that either the human or the learning model 150 is abnormal. This data represents a past disassembly and maintenance case of pump 10 where the learning model 150 predicted that pump 10 was abnormal based on one data point, a human disassembled pump 10, and it was indeed abnormal. In addition, true positive data represents a past disassembly and maintenance case of pump 10 where a human predicted that pump 10 was abnormal based on one data point or information other than one vibration data point, a human disassembled pump 10, and it was indeed abnormal.

[0052] Furthermore, data where a human predicts that pump 10 is abnormal and disassembles it, but pump 10 is actually normal, is sometimes referred to as false positive data from humans. Also, data where the learning model 150 predicts that pump 10 is abnormal and disassembles it, but pump 10 is actually normal, is sometimes referred to as false positive data from the learning model 150.

[0053] Here, vibration data like the vibration data in Figure 4(c), which is collected long after the last maintenance, becomes more complex in structure due to changes in the equipment condition, making it difficult for the learning model 150 to classify whether it is abnormal or not. In such cases, even if the pump 10 is actually functioning normally, the learned learning model 150 may mistakenly identify it as abnormal. Then, the data storage device 30 (see Figure 1) assigns identification information to the misidentified vibration data as false positive data for the learning model 150.

[0054] Furthermore, vibration data like that shown in Figure 4(c), collected long after the last maintenance, becomes more complex due to changes in the equipment's condition, making it difficult for humans to classify whether the data is abnormal or not. Such vibration data may lead to a misinterpretation by humans as abnormal, even when pump 10 is actually functioning normally.

[0055] However, if the sensor 20 acquires vibration data such as the vibration data in Figure 4(c), the human observer will also check other information besides the vibration data, such as the operating sound of the pump 10 and the flow rate of the pump 10. The human observer will then check both the vibration data and the other information to make a comprehensive prediction as to whether the pump 10 is malfunctioning or not. If, as a result, the human observer incorrectly determines that the pump 10 is malfunctioning when it is actually functioning normally, the vibration data used in the prediction will be identified as false positive data by the data storage device 30 (see Figure 1).

[0056] Furthermore, consider the case where the sensor 20 acquires vibration data during a period of low equipment wear, as shown in Figures 4(a) and (b). In this case, even if a human predicts that the vibration data indicates that the pump 10 is functioning normally, if other information also predicts that the pump 10 is malfunctioning, the human may disassemble the pump 10. If, after disassembling the pump 10, it is found that the pump 10 is actually functioning normally, the vibration data used for the prediction will be identified by the data storage device 30 (see Figure 1) as false positive data. Note that although the data used for this prediction is the only data that a human would check and predict that the pump 10 is functioning normally, it may be treated as data similar to the false positive data described later, as it was data that could have led to a misjudgment.

[0057] Thus, humans predict whether pump 10 is malfunctioning or not based on vibration data and other information. Therefore, there may be differences between a human's prediction of whether pump 10 is malfunctioning and the prediction made by the trained model 150 based on vibration data. Furthermore, the prediction may differ depending on whether a human checks only the vibration data or checks both the vibration data and other information.

[0058] For example, a trained model 150, which has been trained on a single vibration data, might predict that pump 10 is abnormal, while a human, based on the single vibration data and information other than the single vibration data, might predict that pump 10 is normal.

[0059] Furthermore, for example, if a human observes only one vibration data point, they may predict that pump 10 is malfunctioning, while if they observe both one vibration data point and other information, they may predict that pump 10 is functioning normally.

[0060] Furthermore, in human false positive data, characteristics of information other than vibration data may appear as vibration characteristics. For example, characteristics such as the operating sound of pump 10, which a human might mistakenly judge as abnormal, may appear as vibration characteristics in the vibration data.

[0061] Figures 9(a) and 9(b) show the initial training or retraining of the learning model 150. Figure 9(a) shows the initial training of the learning model 150. Figure 9(b) shows the retraining of the trained learning model 150.

[0062] In Figure 9(a), the group of normal data refers to multiple normal data points that have been separated from the vibration data acquired from the data storage device by the sorting unit 120 and processed by the preprocessing unit 130. Here, the normal data includes not only data from when the pump 10 was actually disassembled, but also data that is highly likely to be normal even though the pump 10 was not actually disassembled. For example, it includes data acquired from the sensor 20 from after the maintenance of the pump 10 until before the reference date on which false positive data from a human was acquired from the sensor 20. It also includes data acquired from the sensor 20 before the reference time described later.

[0063] Furthermore, among the group of normal performance data, data similar to human false positives refers to vibration data similar to human false positive data. For example, data similar to human false positives refers to vibration data similar to the frequency intensity and other features of false positive data. Alternatively, for example, data similar to human false positives refers to vibration data acquired from sensor 20 around the reference date, with the reference date being the day on which human false positive data was acquired from sensor 20. The period around the reference date is, for example, one week before the reference date. Alternatively, the operating time of pump 10 when human false positive data was acquired from sensor 20 may be used as the reference time, and vibration data acquired from sensor 20 around the reference time may be considered data similar to human false positive data. For example, if the operating time of pump 10 when human false positive data was acquired from sensor 20 is 200 hours, then vibration data acquired from sensor 20 when pump 10 has been operating for 175 hours may be used as data similar to human false positive data. Note that this operating time of pump 10 is the time measured after the first maintenance.

[0064] Furthermore, among the data set of normal performance data, data that does not resemble human false positives is vibration data that does not resemble human false positive data. For example, this refers to data where the features such as frequency intensity of the false positive data are not similar to the features of the vibration data. Also, for example, this refers to vibration data acquired outside the vicinity of the reference date, using the date on which human false positive data was acquired from sensor 20 as the reference date.

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

[0066] Furthermore, the input unit 140 may prioritize inputting human false positive data and data similar to human false positives to the learning model 150 as initial training data. For example, the amount of human false positive data and data similar to human false positives input may be greater than the amount of other data. In addition, the input ratio of false positive data and data similar to false positives in the input training data may be changed according to the time series. For example, using the day on which false positive data was acquired from the sensor 20 as the reference date, the input ratio of false positive data and data similar to false positives in the input training data may be reduced as the reference date is moved back in time. Alternatively, only data similar to human false positives may be input to the learning model 150.

[0067] The input unit 140 prioritizes human false positive data, and the learning model 150 performs initial training focusing on the characteristics of human false positive data. The characteristics of false positive data for the learning model 150 are oscillation characteristics that do not appear in data other than human false positive data. Examples of oscillation characteristics include oscillation frequency and intensity, and the learning model 150 learns where the oscillation frequency occurs, the degree of oscillation intensity, etc. Then, the learning model 150 becomes a trained learning model 150 that has been trained with a focus on the characteristics of human false positive data. This trained learning model 150 can output an anomaly score based on the similarity between the characteristics of the data to be inferred and the characteristics of normal data other than human false positive data, as well as the similarity between the characteristics of the data to be inferred and the characteristics of human false positive data, in response to input data to be inferred.

[0068] In Figure 9(b), the additional normal data set consists of the false positive data from the learning model 150 and data similar to the false positive data from the learning model 150. As mentioned earlier, the false positive data from the learning model 150 is the data used for inference when the trained learning model 150 determined that the pump 10 was abnormal based on the inference data, but the pump was actually functioning normally. The data similar to the false positives from the learning model 150 consists of additional data acquired around the day on which the false positive data from the learning model 150 was acquired from the sensor 20, with the reference date being used as the reference date. For example, the data similar to the false positives from the learning model 150 is additional data acquired from the sensor 20 up to two or three days before the reference date. As mentioned earlier, the operating time of the pump 10 may also be used as the reference time. In addition to the false positive data from the learning model 150 and data similar to the false positive data from the learning model 150, the actual normal data set may also be added as part of the additional normal data set.

[0069] The input unit 140 prioritizes inputting false positive data from the trained model 150 as training data for retraining into the trained model 150. For example, the amount of false positive data from the trained model 150 input is set to be greater than the amount of data other than the false positive data from the trained model 150. Alternatively, only the false positive data from the trained model 150 may be input into the trained model 150.

[0070] Furthermore, the input unit 140 may prioritize inputting false positive data from the learning model 150 and data similar to the false positive data from the learning model 150 as training data for retraining to the already trained learning model 150. For example, the amount of false positive data from the learning model 150 and data similar to the false positive data from the learning model 150 may be increased compared to data other than the false positive data from the learning model 150. In addition, the input ratio of false positive data from the learning model 150 and data similar to the false positives from the learning model 150 in the input training data may be changed according to the time series. For example, using the day on which the false positive data from the learning model 150 was acquired from the sensor 20 as the reference date, the input ratio of false positive data from the learning model 150 and data similar to the false positives from the learning model 150 may be decreased as the reference date is moved back. Alternatively, only data similar to the false positive data from the learning model 150 may be input to the learning model 150.

[0071] The input unit 140 prioritizes input of false positive data from the learning model 150, causing the trained learning model 150 to retrain with an emphasis on the features of the false positive data. The features of the false positive data from the learning model 150 are oscillation features that do not appear in data other than the false positive data from the learning model 150. Examples of oscillation features include oscillation frequency and intensity, and the trained learning model 150 learns where the oscillation frequency occurs, the degree of the oscillation intensity, etc. Then, the trained learning model 150 becomes a trained learning model 150 that has been retrained with an emphasis on the false positive data from the learning model 150. This trained learning model 150 can output an anomaly score based on the input of the data to be inferred, not only based on the similarity between the features of the data to be inferred and the features of normal data other than the false positive data from the learning model 150, but also based on the similarity between the features of the data to be inferred and the features of the false positive data from the learning model 150.

[0072] Figure 10 is a graph showing the anomaly score output by the trained model 150 over time. In the graph in Figure 10, the vertical axis represents the anomaly score, and the horizontal axis represents the operating time. The operating time represents the operating time of pump 10 after it has been serviced once. IVA represents the period immediately after service, IVB represents the intermediate period, and IVC represents the period immediately before service. It is assumed that the anomaly score threshold is 30.

[0073] In the IVA of Figure 10, for example, if the inference device 200 acquires vibration data as the data to be inferred, the trained learning model 150 determines that the characteristics of the data to be inferred are similar to the characteristics of the vibration data of the normal data including false positive data, and determines that the abnormality level is 1, and the determination unit 250 determines that the pump 10 is normal.

[0074] In IVB in Figure 10, for example, let's assume that the inference device 200 acquires vibration data as the data to be inferred, as shown in Figure 4(b). In this case, the trained model 150 determines that the characteristics of the data to be inferred are somewhat similar to the characteristics of the vibration data of the normal data including the false positive data, and determines that the abnormality level is 17.5, and the determination unit 250 determines that the pump 10 is normal. Somewhat similar means that although it is not as similar as the stage in IVA, the pump 10 is within the normal range.

[0075] In the IVC shown in Figure 10, let's assume, for example, that the inference device 200 acquires vibration data as the data to be inferred, as shown in Figure 4(c). In this case, the trained learning model 150 determines that the characteristics of the data to be inferred are not similar to the characteristics of the vibration data of normal data, including false positive data, and assigns an anomaly score of 40, and the judgment unit 250 determines that the pump 10 is abnormal. Then, a human confirms the detection result by the judgment unit 250, disassembles the pump 10, and repairs the pump 10 if an abnormality has actually occurred. On the other hand, if the pump 10 is disassembled but no abnormality has actually occurred, the vibration data as shown in Figure 4(c) is assigned identification information as false positive data by the learning model 150.

[0076] Figure 11 illustrates the repurposing of a trained model 150. In Figure 11, devices A11, B12, and C13 each represent different pumps 10. The trained model 150 acquires vibration data from devices A11 and B12 and uses this vibration data as training data for learning. After completing the learning process, it becomes the trained model 150. This trained model 150 can acquire inference target data from devices A11 and B12 and detect abnormalities in the pumps 10. Furthermore, this trained model 150 can also acquire inference target data from device C13 and detect abnormalities there as well. Device C13 is a device from which the trained model 150 did not acquire training data in the first place. Thus, the trained model 150 can be applied to the device from which it was trained, and can also be repurposed for devices from which it was not trained. [Explanation of Symbols]

[0077] 10...Pump, 20...Sensor, 30...Data storage device, 100...Learning device, 200...Inference device

Claims

1. Equipped with a processor, The aforementioned processor, Multiple operational data are acquired from the object where an anomaly was detected. The learning model is trained using the aforementioned multiple operational data, including false positive data indicating that the anomaly detection target is abnormal when it is actually normal, as initial training data. An anomaly detection system characterized by the following features.

2. The aforementioned plurality of operational data include the false positive data and data other than the false positive data, The anomaly detection system according to claim 1, characterized in that the false positive data from among the plurality of operating data is given priority and used as the initial training data to train the learning model.

3. The anomaly detection system according to claim 2, characterized in that only the false positive data is used as the initial training data to train the learning model.

4. The data other than the false positive data includes data similar to the false positive data. The anomaly detection system according to claim 2, characterized in that, from among the plurality of operating data, the false positive data and / or data similar to the false positive data are given priority and used to train the learning model as the initial training data.

5. Equipped with a processor, The aforementioned processor, Load the driving data to be inferred, A learning model is obtained by training on multiple past operating data, including false positive data indicating an anomaly when the target is actually normal, using these as initial training data. Using the acquired learning model, the system detects whether the driving data subject to inference indicates an anomaly targeted for anomaly detection. An anomaly detection system characterized by the following features.

6. The anomaly detection system according to claim 5, characterized in that the learning model can be adapted for an anomaly detection target different from the anomaly detection target.

7. The aforementioned multiple operational data are vibration data obtained from the vibrations targeted for abnormality detection, The anomaly detection system according to claim 1 or 5, characterized in that the learning model learns the characteristics of the vibration data.

Images