Learning data selection device, learning data selection method, and abnormality detection device

By identifying and screening false negative and false positive detection data and selecting appropriate sensor data as re-learning data for the learning model, the error detection problem in the prior art is solved and the accuracy and accuracy of abnormal detection is improved.

CN118900992APending Publication Date: 2024-11-05MITSUBISHI ELECTRIC CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202280093744.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2022-03-30
Publication Date
2024-11-05

Smart Images

  • Figure CN118900992A_ABST
    Figure CN118900992A_ABST
Patent Text Reader

Abstract

A learning data selection device (2) is configured so as to comprise: a sensor data acquisition unit (11) that acquires, from a sensor (1) that observes a failure detection object, a plurality of pieces of sensor data that indicate the observation results of the failure detection object, and a detection data acquisition unit (12) that acquires, from the sensor (1) that observes the failure detection object, a plurality of pieces of sensor data that indicate the observation results of the failure detection object; and a detection data acquisition unit (12) that assigns each piece of sensor data acquired by the sensor data acquisition unit (11) to a learning model (13) that learns the distribution of sensor data when the failure detection object is normal, and that acquires, from the learning model (13), detection data indicating whether the failure detection object is normal or abnormal. Furthermore, the learning data selection device (2) is provided with a learning data selection unit (14) that acquires a learning data selection unit (14) for selecting learning data from among the plurality of pieces of sensor data acquired by the sensor data acquisition unit (11). The identification information identifies which sensor data is the sensor data related to the false negative detection data indicating the normal object although the fault detection object is abnormal. From sensor data other than sensor data related to false-negative detection data among a plurality of pieces of sensor data acquired by a sensor data acquisition unit (11), learning data to be used for re-learning of a learning model (13) is used as learning data to be used for re-learning of the learning model (13). Sensor data related to detection data indicating that a failure detection object is normal is selected.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a learning data selection device, a learning data selection method and an abnormality detection device. Background Art

[0002] There is an abnormality detection device that obtains sensor data from a sensor that observes a fault detection object, assigns the sensor data to a learning model, and obtains detection data indicating whether the fault detection object is normal or abnormal from the learning model (for example, refer to Patent Document 1). The learning model is based on the premise that sensor data when the fault detection object is normal is assigned as learning data, and the distribution of sensor data when the fault detection object is normal is learned. The sensor data when the fault detection object is normal is, for example, sensor data output from a sensor during the initial stage of operation of the fault detection object.

[0003] In this abnormality detection device, in order to improve the detection accuracy of the learning model, the learning model may be relearned. The learning data used for relearning the learning model include sensor data for additional learning and sensor data in a normal period of the fault detection target.

[0004] The sensor data for additional learning is sensor data when there is a high possibility of false detection that the fault detection target is normal but detected as abnormal. The sensor data within the normal period of the fault detection target is sensor data determined to be within the normal period in any evaluation.

[0005] Patent Document 1: Japanese Patent Application Publication No. 2019-28565 Summary of the invention

[0006] In the abnormality detection device disclosed in Patent Document 1, sensor data within a normal period that is judged as a fault detection object among sensor data at any evaluation is used as learning data used for relearning the learning model. However, in the initial stage of operation, when an initial fault occurs in the fault detection object, or when the fault detection object deteriorates over time, the sensor data within the normal period may include sensor data when the fault detection object is abnormal. Therefore, sensor data within a normal period that is judged as a fault detection object among sensor data at any evaluation may also include sensor data that is erroneously detected as normal although the fault detection object is abnormal. In the case where the learning data used for relearning the learning model includes sensor data when the fault detection object is abnormal, there is the following problem, that is, even if the learning model is relearned, the abnormality detection device sometimes makes a false detection that the fault detection object is abnormal but still detected as normal.

[0007] The present invention is proposed to solve the above-mentioned problems, and its purpose is to obtain a learning data selection device and a learning data selection method that can select learning data that can reduce false detections of the subject that is detected as normal although the fault detection object is abnormal.

[0008] The learning data selection device of the present invention comprises: a sensor data acquisition unit, which acquires a plurality of sensor data indicating the observation result of the fault detection object from a sensor that observes the fault detection object; and a detection data acquisition unit, which assigns each sensor data acquired by the sensor data acquisition unit to a learning model that has learned the distribution of sensor data when the fault detection object is normal, and acquires detection data indicating whether the fault detection object is normal or abnormal from the learning model. In addition, the learning data selection device comprises a learning data selection unit, which acquires identification information for identifying which sensor data among the plurality of sensor data acquired by the sensor data acquisition unit is the sensor data related to the false negative detection data indicating that the fault detection object is normal although the fault detection object is abnormal, and selects the sensor data related to the detection data indicating that the fault detection object is normal from the sensor data other than the sensor data related to the false negative detection data among the plurality of sensor data acquired by the sensor data acquisition unit as the learning data used for re-learning the learning model.

[0009] Effects of the Invention

[0010] According to the present invention, it is possible to select learning data that can reduce false detections that detect a fault detection target as normal even though the target is abnormal. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a diagram showing the structure of the learning data selection device 2 involved in the first embodiment.

[0012] Figure 2 This is a hardware configuration diagram showing the hardware of the learning data selection device 2 according to the first embodiment.

[0013] Figure 3 This is a hardware configuration diagram of a computer when the learning data selection device 2 is implemented by software, firmware, or the like.

[0014] Figure 4 It is a flowchart showing the processing procedure of the learning data selection device 2, that is, the learning data selection method.

[0015] Figure 5 1 is an explanatory diagram showing the distribution of sensor data after learning by the learning model 13 when the fault detection target is normal.

[0016] Figure 6 This is an explanatory diagram showing an example of sensor data displayed on the display.

[0017] Figure 7 1 is an explanatory diagram showing the distribution of sensor data after relearning by the learning model 13 .

[0018] Figure 8 It is a diagram showing the structure of the learning data selection device 2 involved in the second embodiment.

[0019] Fig. 9 This is a hardware configuration diagram showing the hardware of the learning data selection device 2 according to the second embodiment.

[0020] Fig.10 It is an explanatory diagram showing two false-negative sensor data, two false-positive sensor data, and four normal sensor data.

[0021] Fig.11 Yes means Fig.10 Graphs for explaining the first evaluation value and the second evaluation value in four pieces of normal sensor data shown.

[0022] Fig.12 It is a diagram showing the structure of the learning data selection device 2 involved in the third embodiment.

[0023] Fig.13 Yes means Fig.10 Graphs for explaining the first evaluation value, the second evaluation value, and the third evaluation value in four pieces of normal sensor data are shown.

[0024] Fig.14 It is a diagram showing the structure of the learning data selection device 2 involved in the fourth embodiment.

[0025] Fig.15 It is a hardware configuration diagram showing the hardware of the learning data selection device 2 according to the fourth embodiment.

[0026] Fig.16 It is an explanatory diagram showing the ratio of the learning data output from the learning data selection unit 14 included in the learning data used for the previous learning of the learning model 13 .

[0027] Fig.17 This is a diagram showing the structure of an abnormality detection device according to the fifth embodiment.

[0028] Fig.18 This is a hardware configuration diagram showing the hardware of the abnormality detection device according to the fifth embodiment. DETAILED DESCRIPTION

[0029] Hereinafter, in order to explain the present invention in more detail, a mode for carrying out the present invention will be described with reference to the accompanying drawings.

[0030] Implementation method 1.

[0031] Figure 1 It is a diagram showing the structure of the learning data selection device 2 involved in the first embodiment.

[0032] Figure 2 This is a hardware configuration diagram showing the hardware of the learning data selection device 2 according to the first embodiment.

[0033] exist Figure 1 In the process, sensor 1 repeatedly observes the fault detection object.

[0034] The sensor 1 outputs a plurality of sensor data indicating observation results of a fault detection target to the learning data selection device 2 .

[0035] The learning data selection device 2 includes a sensor data acquisition unit 11 , a detection data acquisition unit 12 , and a learning data selection unit 14 .

[0036] The human-machine interface unit (hereinafter referred to as "human-machine IF unit") 3 has an input device and an output device. The input device is a device that receives user operations, such as a mouse or a keyboard. The output device is implemented by a display device that displays sensor data output from the learning data selection device 2.

[0037] The sensor data acquisition unit 11 uses, for example, Figure 2 The sensor data acquisition circuit 21 shown is implemented.

[0038] The sensor data acquisition unit 11 acquires a plurality of sensor data indicating observation results of a fault detection target from the sensor 1 .

[0039] The sensor data acquisition unit 11 outputs each sensor data to the detection data acquisition unit 12 and the learning data selection unit 14 .

[0040] The detection data acquisition unit 12 performs, for example, Figure 2 The detection data acquisition circuit 22 shown is implemented.

[0041] The detection data acquisition unit 12 includes a learning model 13 .

[0042] The detection data acquisition unit 12 assigns each sensor data acquired by the sensor data acquisition unit 11 to the learning model 13 , and acquires detection data indicating whether the fault detection target is normal or abnormal from the learning model 13 .

[0043] The detection data acquisition unit 12 outputs each detection data to the learning data selection unit 14 .

[0044] exist Figure 1 In the illustrated learning data selection device 2 , the detection data acquisition unit 12 includes the learning model 13 . However, this is just an example, and the learning model 13 may be provided outside the detection data acquisition unit 12 .

[0045] The learning model 13 is implemented by, for example, a neural network.

[0046] The learning model 13 is, for example, an unsupervised learning model.

[0047] During learning, the learning model 13 is given sensor data when the fault detection target is normal, and learns the distribution of the sensor data.

[0048] When inferring, the learning model 13 outputs detection data indicating that the fault detection target is normal if given sensor data when the fault detection target is normal. When inferring, the learning model 13 outputs detection data indicating that the fault detection target is abnormal if given sensor data when the fault detection target is abnormal.

[0049] However, when the detection accuracy of the learning model 13 is low, the learning model 13 may output false positive detection data indicating an abnormality even though the fault detection object is normal, or may output false negative detection data indicating a normality even though the fault detection object is abnormal, resulting in erroneous detection.

[0050] The learning data selection unit 14 performs, for example, Figure 2 The learning data selection circuit 24 shown is implemented.

[0051] The learning data selection unit 14 includes an identification information acquisition unit 15 , a sensor data selection unit 16 , and a learning data output unit 17 .

[0052] The learning data selection unit 14 acquires identification information for identifying which sensor data among the plurality of sensor data acquired by the sensor data acquisition unit 11 is sensor data related to false negative detection data indicating that the fault detection target is normal although it is abnormal.

[0053] The learning data selection unit 14 selects, based on the identification information, sensor data related to detection data indicating that the fault detection object is normal from sensor data other than sensor data related to false negative detection data among the multiple sensor data acquired by the sensor data acquisition unit 11 as learning data to be used for re-learning the learning model 13, and selects sensor data related to detection data indicating that the fault detection object is normal.

[0054] Furthermore, the learning data selection unit 14 acquires identification information for identifying which sensor data among the plurality of sensor data acquired by the sensor data acquisition unit 11 is sensor data related to false positive detection data indicating abnormality although the fault detection target is normal.

[0055] The learning data selection unit 14 selects sensor data related to false positive detection data from among the plurality of sensor data acquired by the sensor data acquisition unit 11 as learning data to be used for re-learning the learning model 13 based on the identification information.

[0056] The learning data selection unit 14 outputs the selected sensor data to the outside as learning data used for re-learning the learning model 13 .

[0057] The identification information acquisition unit 15 acquires each sensor data from the sensor data acquisition unit 11 , and acquires each detection data from the detection data acquisition unit 12 .

[0058] The identification information acquisition unit 15 outputs each sensor data and each detection data to the human-machine IF unit 3. The human-machine IF unit 3 displays each sensor data together with the detection results shown by each detection data, for example, on a display. If the sensor data displayed on the display is sensor data related to false-positive detection data, the user operates the input device of the human-machine IF unit 3 to set the identification information, and the identification information indicates that the sensor data is sensor data related to false-positive detection data. The human-machine IF unit 3 outputs the set identification information to the identification information acquisition unit 15. In addition, if the sensor data displayed on the display is sensor data related to false-negative detection data, the user operates the input device of the human-machine IF unit 3 to set the identification information, and the identification information indicates that the sensor data is sensor data related to false-negative detection data. The human-machine IF unit 3 outputs the set identification information to the identification information acquisition unit 15. Furthermore, if the sensor data displayed on the display is sensor data related to normal detection data, which indicates that the fault detection object is normal, the user operates the input device of the human-machine IF unit 3 to set identification information indicating that the sensor data is sensor data related to normal detection data. The human-machine IF unit 3 outputs the set identification information to the identification information acquisition unit 15.

[0059] The identification information acquisition unit 15 acquires the identification information output from the human-machine interface unit 3 , and outputs the identification information to the sensor data selection unit 16 .

[0060] Here, if the sensor data displayed on the display is sensor data related to normal detection data, which indicates the main idea of ​​normal when the fault detection object is normal, the user operates the input device of the human-machine IF unit 3 to set identification information, which indicates the main idea that the sensor data is sensor data related to normal detection data. Moreover, the human-machine IF unit 3 outputs the set identification information to the identification information acquisition unit 15. However, this is just an example, and the user can also set the identification information indicating the main idea of ​​sensor data related to normal detection data without operating the input device of the human-machine IF unit 3. In this case, the human-machine IF unit 3 does not output the identification information indicating the main idea of ​​sensor data related to normal detection data to the identification information acquisition unit 15.

[0061] The sensor data selection unit 16 acquires each sensor data from the sensor data acquisition unit 11 , acquires each detection data from the detection data acquisition unit 12 , and acquires identification information from the identification information acquisition unit 15 .

[0062] The sensor data selection unit 16 selects sensor data related to normal detection data from among the plurality of sensor data acquired by the sensor data acquisition unit 11 based on the identification information.

[0063] Furthermore, the sensor data selection unit 16 selects sensor data related to false positive detection data from among the plurality of sensor data acquired by the sensor data acquisition unit 11 based on the identification information.

[0064] The sensor data selection unit 16 outputs the sensor data related to the normal detection data and the sensor data related to the false positive detection data to the learning data output unit 17 .

[0065] The learning data output unit 17 acquires sensor data related to normal detection data and sensor data related to false positive detection data from the sensor data selection unit 16 .

[0066] If the similarity between the sensor data related to the normal detection data and the sensor data related to the false positive detection data is greater than or equal to the threshold, the learning data output unit 17 outputs the sensor data related to the normal detection data to the outside as the learning data used for re-learning the learning model 13.

[0067] If the similarity is smaller than the threshold value, the learning data output unit 17 does not output the sensor data related to the normal detection data to the outside as learning data used for re-learning the learning model 13 but discards it.

[0068] The learning data output unit 17 outputs sensor data related to false positive detection data to the outside as learning data used for relearning the learning model 13. The threshold value may be stored in the internal memory of the learning data output unit 17 or may be given from outside the learning data selection device 2.

[0069] Imagine Figure 1 In the example, the sensor data acquisition unit 11, the detection data acquisition unit 12, and the learning data selection unit 14, which are components of the learning data selection device 2, are each configured by Figure 2 Specifically, it is assumed that the learning data selection device 2 is implemented by the sensor data acquisition circuit 21 , the detection data acquisition circuit 22 , and the learning data selection circuit 24 .

[0070] The sensor data acquisition circuit 21 , the detection data acquisition circuit 22 , and the learning data selection circuit 24 are each, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.

[0071] The structural elements of the learning data selection device 2 are not limited to those implemented by dedicated hardware. The learning data selection device 2 can also be implemented by software, firmware, or a combination of software and firmware.

[0072] Software or firmware is stored as a program in the computer's memory. A computer refers to hardware that executes a program, such as a CPU (Central Processing Unit), a central processing unit, a processing unit, an arithmetic unit, a microprocessor, a microcomputer, a processor, or a DSP (Digital Signal Processor).

[0073] Figure 3 This is a hardware configuration diagram of a computer when the learning data selection device 2 is implemented by software, firmware, or the like.

[0074] When the learning data selection device 2 is implemented by software or firmware, a program for causing a computer to execute the processing sequence of the sensor data acquisition unit 11, the detection data acquisition unit 12, and the learning data selection unit 14 is stored in the memory 31. Then, the processor 32 of the computer executes the program stored in the memory 31.

[0075] In addition, Figure 2 In FIG. 1 , an example is shown in which each of the structural elements of the learning data selection device 2 is implemented by dedicated hardware. Figure 3In the example in which the learning data selection device 2 is implemented by software or firmware, etc., is shown. However, this is only an example, and some of the structural elements in the learning data selection device 2 may be implemented by dedicated hardware, and the remaining structural elements may be implemented by software or firmware, etc.

[0076] Next, Figure 1 The operation of the learning data selection device 2 shown will be described.

[0077] Figure 4 It is a flowchart showing the processing procedure of the learning data selection device 2, that is, the learning data selection method.

[0078] exist Figure 1 In the learning data selection device 2 shown, the fault detection object is, for example, a shaft connected to the engine of a car, and the sensor data output from the sensor 1 is, for example, the vibration frequency of the shaft. However, the fault detection object is not limited to the shaft, and the sensor data is not limited to the vibration frequency of the shaft. For example, the fault detection object may also be an air conditioner, and the sensor data may also be power consumption.

[0079] Figure 5 1 is an explanatory diagram showing the distribution of sensor data after learning by the learning model 13 when the fault detection target is normal.

[0080] exist Figure 5 In the figure, ○, △, □, and × are sensor data related to the detection data output from the learning model 13, respectively.

[0081] In particular, ○ is sensor data related to normal detection data indicating that the fault detection target is normal, and △ is sensor data related to false positive detection data indicating that the fault detection target is normal but abnormal.

[0082] □ is sensor data related to false negative detection data indicating that the fault detection target is normal although it is abnormal, and × is sensor data related to abnormal detection data indicating that it is abnormal when the fault detection target is abnormal.

[0083] Each of the detection results of the false-positive detection data and the false-negative detection data is an erroneous detection.

[0084] In order to reduce false detection, it is necessary to relearn the learning model 13. If the sensor data involved in the false positive detection data is used as the learning data used for relearning, the sensor data involved in the false positive detection data is included in the distribution of the sensor data when the fault detection object is normal. The sensor data involved in the false positive detection data is the sensor data when the fault detection object is normal, so the distribution of the sensor data when the fault detection object is normal becomes an appropriate distribution, and as a result, the false positive false detection is reduced.

[0085] If the sensor data involved in the false negative detection data is used as the learning data used for relearning, the sensor data involved in the false negative detection data is included in the distribution of the sensor data when the fault detection object is normal. The sensor data involved in the false negative detection data is the sensor data when the fault detection object is abnormal, so the distribution of the sensor data when the fault detection object is normal becomes an inappropriate distribution, and as a result, false negative misdetection may occur. If the sensor data involved in the false negative detection data is not used as the learning data used for relearning, the sensor data involved in the false negative detection data is not included in the distribution of the sensor data when the fault detection object is normal. As a result, false negative misdetection is reduced.

[0086] If sensor data when the fault detection target is normal is used as learning data used for re-learning, a distribution of sensor data when the fault detection target is normal is formed.

[0087] Therefore, it is effective to use sensor data related to false positive detection data and sensor data related to normal detection data as learning data used for re-learning the learning model 13 .

[0088] On the other hand, sensor data related to false negative detection data and sensor data related to abnormal detection data should not be used as learning data used for re-learning of the learning model 13 .

[0089] Furthermore, among the plurality of sensor data involved in the normal detection data, the sensor data close to the center of the distribution of the sensor data has less effect in clarifying the distribution of the sensor data when the fault detection object is normal than the sensor data close to the sensor data involved in the false positive detection data. Therefore, from the perspective of reducing the amount of learning data and performing efficient learning, the necessity of using the sensor data close to the center of the distribution of the sensor data as learning data is small.

[0090] Sensor 1 repeatedly observes the fault detection object.

[0091] The sensor 1 outputs a plurality of sensor data indicating observation results of the fault detection target to the sensor data acquisition unit 11 of the learning data selection device 2 .

[0092] The sensor data acquisition unit 11 acquires a plurality of sensor data ( Figure 4 Step ST1).

[0093] The sensor data acquisition unit 11 outputs each sensor data to the detection data acquisition unit 12 and the learning data selection unit 14 .

[0094] The detection data acquisition unit 12 acquires each sensor data from the sensor data acquisition unit 11 .

[0095] The detection data acquisition unit 12 assigns each sensor data to the learning model 13, and acquires detection data indicating whether the fault detection object is normal or abnormal from the learning model 13. Figure 4 Step ST2).

[0096] The detection data acquisition unit 12 outputs each detection data to the learning data selection unit 14 .

[0097] Depending on the detection accuracy of the learning model 13 , the plurality of detection data output from the learning model 13 may include false positive detection data or false negative detection data in addition to normal detection data and abnormal detection data.

[0098] If the plurality of detection data output from the learning model 13 include false positive detection data or false negative detection data, and the number of false positive detection data is large, the learning model 13 should be relearned.

[0099] The identification information acquisition unit 15 acquires each sensor data from the sensor data acquisition unit 11 , and acquires each detection data from the detection data acquisition unit 12 .

[0100] The identification information acquisition unit 15 outputs each sensor data and each detection data to the human-machine interface unit 3 .

[0101] The human-machine IF unit 3 displays each sensor data together with the detection result indicated by each detection data on, for example, a display.

[0102] Figure 6 This is an explanatory diagram showing an example of sensor data displayed on the display.

[0103] exist Figure 6In , ○ is sensor data indicating that the fault detection target is normal. However, in ○, in addition to sensor data related to normal detection data, sensor data related to false negative detection data may also be included.

[0104] The "×" indicates sensor data indicating that the fault detection target is abnormal. However, in addition to sensor data related to abnormal detection data, "×" may also include sensor data related to false positive detection data.

[0105] If the user viewing the display determines that sensor data associated with false negative detection data is included in the sensor data indicated by "○", the user operates the input device of the human-machine IF unit 3 to specify sensor data considered to be sensor data associated with false negative detection data.

[0106] The human-machine IF unit 3 outputs identification information indicating that the designated sensor data is sensor data related to false negative detection data (hereinafter referred to as “false negative label”) to the identification information acquisition unit 15 .

[0107] If the user viewing the display determines that sensor data related to false positive detection data is included in the xed sensor data, the user operates the input device of the human-machine IF unit 3 to designate sensor data related to the false positive detection data.

[0108] The human-machine IF unit 3 outputs identification information indicating that the designated sensor data is sensor data related to false positive detection data (hereinafter referred to as “false positive label”) to the identification information acquisition unit 15 .

[0109] If the user viewing the display determines that sensor data indicated by “○” includes sensor data related to normal detection data, the user operates the input device of the human-machine IF unit 3 to designate sensor data considered to be sensor data related to normal detection data.

[0110] The human-machine IF unit 3 outputs identification information indicating that the designated sensor data is sensor data related to detection data in a normal state (hereinafter referred to as “normal state label”) to the identification information acquisition unit 15 .

[0111] Here, the user operates the input device of the human-machine IF unit 3 to designate the sensor data of the sensor data involved in the detection data considered to be normal, and the human-machine IF unit 3 outputs the normal label to the identification information acquisition unit 15. However, this is just an example, and the user may designate the sensor data of the sensor data involved in the detection data considered to be normal without operating the input device of the human-machine IF unit 3. In this case, the human-machine IF unit 3 does not output the normal label to the identification information acquisition unit 15.

[0112] The identification information acquisition unit 15 acquires the false negative label, the false positive label and the normal label from the human-machine IF unit 3. Figure 4 Step ST3).

[0113] The identification information acquisition unit 15 outputs the false negative label, the false positive label, and the normal time label to the sensor data selection unit 16 .

[0114] If the label is not normally output from the human-machine IF unit 3 to the identification information acquisition unit 15 , the identification information acquisition unit 15 obtains the false negative label and the false positive label from the human-machine IF unit 3 , and outputs the false negative label and the false positive label to the sensor data selection unit 16 .

[0115] The sensor data selection unit 16 acquires each sensor data from the sensor data acquisition unit 11 , and acquires each detection data from the detection data acquisition unit 12 .

[0116] Furthermore, the sensor data selection unit 16 acquires the false negative label, the false positive label, and the normal time label from the identification information acquisition unit 15 .

[0117] The sensor data selection unit 16 selects sensor data related to false positive detection data from the plurality of sensor data acquired by the sensor data acquisition unit 11 based on the false positive label ( Figure 4 Step ST4).

[0118] The sensor data selection unit 16 outputs the sensor data related to the false positive detection data to the learning data output unit 17 .

[0119] The sensor data selection unit 16 selects sensor data related to one or more normal detection data from the plurality of sensor data acquired by the sensor data acquisition unit 11 based on the normal label ( Figure 4 Step ST5).

[0120] The sensor data selection unit 16 outputs the sensor data related to each normal detection data to the learning data output unit 17 .

[0121] If the normal-time label is not output from the identification information acquisition unit 15, the sensor data selection unit 16 selects sensor data other than sensor data involving false-negative detection data as sensor data involving normal-time detection data from the sensor data involving detection data indicating that the fault detection object is normal among the multiple sensor data acquired by the sensor data acquisition unit 11 based on the false-negative label.

[0122] The learning data output unit 17 acquires the sensor data related to each of the normal detection data and the sensor data related to the false positive detection data from the sensor data selection unit 16 .

[0123] The learning data output unit 17 calculates the similarity between the sensor data involved in each normal detection data and the sensor data involved in the false positive detection data. As a physical quantity representing the similarity, there is, for example, the Euclidean distance between the sensor data involved in each normal detection data and the sensor data involved in the false positive detection data. The learning data output unit 17 determines the position of the sensor data involved in the normal detection data in the Euclidean space and the position of the sensor data involved in the false positive detection data in the Euclidean space, and can calculate the straight-line distance between the determined two positions as the Euclidean distance.

[0124] The learning data output unit 17 selects sensor data having a similarity greater than or equal to a threshold value from among the sensor data related to the plurality of normal detection data output by the sensor data selection unit 16 .

[0125] The learning data output unit 17 outputs sensor data having a similarity greater than or equal to a threshold value to the outside as learning data used for relearning the learning model 13 ( Figure 4 Step ST6).

[0126] The learning data output unit 17 does not output to the outside the sensor data whose similarity is smaller than the threshold value, but discards it.

[0127] In addition, the learning data output unit 17 outputs the sensor data related to the false positive detection data to the outside as the learning data used for re-learning the learning model 13 ( Figure 4 Step ST6).

[0128] The learning model 13 relearns the distribution of the sensor data when the fault detection target is normal, using the sensor data output from the learning data output unit 17 .

[0129] Figure 7 1 is an explanatory diagram showing the distribution of sensor data after relearning by the learning model 13 .

[0130] like Figure 7 As shown, the distribution of sensor data after relearning only includes sensor data when the fault detection target is normal, and does not include sensor data related to false negative detection data. As a result, the distribution of sensor data when the fault detection target is normal becomes an appropriate distribution.

[0131] exist Figure 7 In the example of Figure 5 The sensor data involved in the 4 false positive detection data shown, and Figure 5 The sensor data shown are related to the four normal detection data.

[0132] In the above embodiment 1, the learning data selection device 2 is constructed in a manner including a sensor data acquisition unit 11 and a detection data acquisition unit 12. The sensor data acquisition unit 11 acquires a plurality of sensor data representing the observation results of the fault detection object from a sensor 1 that observes the fault detection object. The detection data acquisition unit 12 assigns a learning model 13 after learning the distribution of the sensor data when the fault detection object is normal to each sensor data acquired by the sensor data acquisition unit 11, and acquires detection data indicating whether the fault detection object is normal or abnormal from the learning model 13. In addition, the learning data selection device 2 includes a learning data selection unit 14, which obtains identification information for identifying which sensor data is the sensor data related to the false negative detection data indicating that the fault detection object is normal, among the plurality of sensor data obtained by the sensor data acquisition unit 11, and based on the identification information, selects the sensor data related to the detection data indicating that the fault detection object is normal from the sensor data other than the sensor data related to the false negative detection data among the plurality of sensor data obtained by the sensor data acquisition unit 11 as the learning data used for re-learning the learning model 13. Therefore, the learning data selection device 2 can select the learning data that can reduce the false detection of the fault detection object being detected as normal, although the fault detection object is abnormal.

[0133] Implementation method 2.

[0134] In the second embodiment, the learning data selection device 2 is described in which the learning data selection unit 41 includes the identification information acquisition unit 15 , the data classification unit 42 , the evaluation value calculation unit 43 , the priority calculation unit 44 , and the learning data output unit 45 .

[0135] Figure 82 is a diagram showing the structure of a learning data selection device 2 according to Embodiment 2. Figure 8 In, with Figure 1 The same reference numerals denote the same or corresponding parts, and thus their descriptions are omitted.

[0136] Fig. 9 2 is a hardware configuration diagram showing the hardware of the learning data selection device 2 according to the second embodiment. Fig. 9 In, with Figure 2 The same reference numerals denote the same or corresponding parts, and thus their descriptions are omitted.

[0137] Figure 8 The illustrated learning data selection device 2 includes a sensor data acquisition unit 11 , a detection data acquisition unit 12 , and a learning data selection unit 41 .

[0138] The learning data selection unit 41 performs, for example, Fig. 9 The learning data selection circuit 25 shown is implemented.

[0139] The learning data selection unit 41 includes the identification information acquisition unit 15 , a data classification unit 42 , an evaluation value calculation unit 43 , a priority calculation unit 44 , and a learning data output unit 45 .

[0140] The learning data selection unit 41 acquires identification information for identifying which sensor data among the plurality of sensor data acquired by the sensor data acquisition unit 11 is sensor data related to false negative detection data indicating that the fault detection target is abnormal but normal.

[0141] The learning data selection unit 41 selects, based on the identification information, sensor data related to detection data indicating that the fault detection object is normal from among the multiple sensor data acquired by the sensor data acquisition unit 11, other than the sensor data related to the false negative detection data, as learning data used for re-learning the learning model 13.

[0142] Furthermore, the learning data selection unit 41 acquires identification information for identifying which sensor data among the plurality of sensor data acquired by the sensor data acquisition unit 11 is sensor data related to false positive detection data indicating abnormality although the fault detection target is normal.

[0143] The learning data selection unit 41 selects sensor data related to false positive detection data from among the plurality of sensor data acquired by the sensor data acquisition unit 11 as learning data to be used for re-learning the learning model 13 based on the identification information.

[0144] The learning data selection unit 41 outputs the selected sensor data to the outside as learning data used for re-learning the learning model 13 .

[0145] The data classification unit 42 acquires each sensor data from the sensor data acquisition unit 11 , and acquires each detection data from the detection data acquisition unit 12 .

[0146] Furthermore, the data classification unit 42 acquires the false negative label, the false positive label, and the normal-time label as identification information from the identification information acquisition unit 15 .

[0147] The data classification unit 42 classifies each sensor data acquired by the sensor data acquisition unit 11 into sensor data involving false negative detection data (hereinafter referred to as "false negative sensor data"), sensor data involving false positive detection data (hereinafter referred to as "false positive sensor data"), or sensor data involving normal detection data (hereinafter referred to as "normal sensor data") based on false negative labels, false positive labels and normal time labels.

[0148] Among the plurality of sensor data acquired by the sensor data acquisition unit 11 , sensor data related to abnormal detection data is not classified by the data classification unit 42 .

[0149] The data classification unit 42 outputs the false negative sensor data, the false positive sensor data, and the normal sensor data to the evaluation value calculation unit 43 .

[0150] In addition, the data classification unit 42 outputs the false positive sensor data and the normal sensor data to the learning data output unit 45 .

[0151] If the normal label is not output from the identification information acquisition unit 15, the data classification unit 42 selects sensor data other than sensor data related to false negative detection data as sensor data related to normal detection data from sensor data related to detection data indicating that the fault detection object is normal among the plurality of sensor data acquired by the sensor data acquisition unit 11, based on the false negative label. Then, the data classification unit 42 classifies the selected sensor data as normal sensor data.

[0152] The evaluation value calculation unit 43 obtains the false negative sensor data, the false positive sensor data, and the normal sensor data from the data classification unit 42. Here, for convenience of explanation, it is assumed that the number of normal sensor data output from the data classification unit 42 is plural.

[0153] The evaluation value calculation unit 43 calculates, as the first evaluation value of each normal sensor data, an evaluation value whose absolute value becomes larger and whose sign becomes positive as the similarity between each normal sensor data and the false positive sensor data becomes higher.

[0154] The evaluation value calculation unit 43 calculates, as the second evaluation value of each normal sensor data, an evaluation value whose absolute value becomes larger and whose sign becomes negative as the similarity between each normal sensor data and the false-negative sensor data becomes higher.

[0155] The evaluation value calculation unit 43 outputs the first evaluation value of each normal sensor data and the second evaluation value of each normal sensor data to the priority calculation unit 44 .

[0156] The priority calculation unit 44 acquires the first evaluation value of each normal sensor data and the second evaluation value of each normal sensor data from the evaluation value calculation unit 43 .

[0157] The priority calculation unit 44 calculates the priority of each normal sensor data based on the first evaluation value of each normal sensor data and the second evaluation value of each normal sensor data.

[0158] The priority calculation unit 44 outputs the priority of each normal sensor data to the learning data output unit 45 .

[0159] The learning data output unit 45 selects one or more pieces of normal sensor data from among the plurality of normal sensor data output by the data classification unit 42 based on the priority calculated by the priority calculation unit 44 .

[0160] The learning data output unit 45 outputs the selected normal sensor data and false positive sensor data to the outside as learning data used for re-learning of the learning model.

[0161] Imagine Figure 8 In the example, the sensor data acquisition unit 11, the detection data acquisition unit 12, and the learning data selection unit 41, which are components of the learning data selection device 2, are each configured by Fig. 9 Specifically, it is assumed that the learning data selection device 2 is implemented by the sensor data acquisition circuit 21 , the detection data acquisition circuit 22 , and the learning data selection circuit 25 .

[0162] Each of the sensor data acquisition circuit 21 , the detection data acquisition circuit 22 , and the learning data selection circuit 25 is, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA, or a combination thereof.

[0163] The structural elements of the learning data selection device 2 are not limited to those implemented by dedicated hardware. The learning data selection device 2 can also be implemented by software, firmware, or a combination of software and firmware.

[0164] When the learning data selection device 2 is implemented by software or firmware, a program for causing a computer to execute the processing sequence of the sensor data acquisition unit 11, the detection data acquisition unit 12, and the learning data selection unit 41 is stored in Figure 3 The memory 31 shown. Moreover, Figure 3 The processor 32 shown executes the program stored in the memory 31 .

[0165] In addition, Fig. 9 In FIG. 1 , an example is shown in which each of the structural elements of the learning data selection device 2 is implemented by dedicated hardware. Figure 3 In the example in which the learning data selection device 2 is implemented by software or firmware, etc., is shown. However, this is only an example, and some of the structural elements in the learning data selection device 2 may be implemented by dedicated hardware, and the remaining structural elements may be implemented by software or firmware, etc.

[0166] Next, Figure 8 The operation of the learning data selection device 2 shown in FIG. 4 is described below. However, except for the learning data selection unit 41, Figure 1 Therefore, only the operation of the learning data selection unit 41 will be described here.

[0167] The data classification unit 42 acquires each sensor data from the sensor data acquisition unit 11 , and acquires each detection data from the detection data acquisition unit 12 .

[0168] Furthermore, the data classification unit 42 acquires the false negative label, the false positive label, and the normal-time label as identification information from the identification information acquisition unit 15 .

[0169] The data classification unit 42 classifies each sensor data into false-negative sensor data, false-positive sensor data, or normal sensor data based on the false-negative label, the false-positive label, and the normal-time label, respectively.

[0170] If the normal-time label cannot be obtained from the identification information acquisition unit 15, the data classification unit 42 selects sensor data other than sensor data related to false-negative detection data as sensor data related to normal-time detection data from sensor data related to detection data indicating that the fault detection object is normal among the plurality of sensor data acquired by the sensor data acquisition unit 11, based on the false-negative label. The data classification unit 42 classifies the selected sensor data as normal sensor data.

[0171] Among the plurality of sensor data acquired by the sensor data acquisition unit 11 , sensor data related to abnormal detection data is not classified by the data classification unit 42 .

[0172] The data classification unit 42 outputs the false negative sensor data, the false positive sensor data, and the normal sensor data to the evaluation value calculation unit 43 .

[0173] In addition, the data classification unit 42 outputs the false positive sensor data and the normal sensor data to the learning data output unit 45 .

[0174] The evaluation value calculation unit 43 obtains the false negative sensor data, the false positive sensor data, and the normal sensor data from the data classification unit 42 .

[0175] exist Figure 8 In the illustrated learning data selection device 2 , for convenience of explanation, it is assumed that the evaluation value calculation unit 43 obtains two false negative sensor data, two false positive sensor data, and four normal sensor data from the data classification unit 42 .

[0176] The evaluation value calculation unit 43 calculates the first evaluation value EV of each normal sensor data. Dm-FP1 ,EV Dm-FP2 , the higher the similarity between each normal sensor data and the false positive sensor data, the larger the absolute value and the positive sign of the evaluation value are calculated. m=1, 2, 3, 4.

[0177] As the first evaluation value EV Dm-FP1 ,EV Dm-FP2 For example, there is the Euclidean distance. Dm-FP1 ,EV Dm-FP2 The absolute value of the Euclidean distance between the normal sensor data and the false positive sensor data is greater as the similarity between the normal sensor data and the false positive sensor data increases. Dm-FP1 ,EV Dm-FP2 The sign of the Euclidean distance is positive. The evaluation value calculation unit 43 specifies the position of the normal sensor data in the Euclidean space and the position of the false positive sensor data in the Euclidean space, and can calculate the straight-line distance between the position of the normal sensor data and the position of the false positive sensor data as the Euclidean distance.

[0178] The evaluation value calculation unit 43 calculates the second evaluation value EV for each normal sensor data. Dm-FN1 ,EV Dm-FN2 , the higher the similarity between each normal sensor data and the false negative sensor data, the larger the absolute value and the negative sign of the evaluation value are calculated.

[0179] As the second evaluation value EV Dm-FN1 ,EV Dm-FN2 For example, there is the Euclidean distance. Dm-FN1 ,EV Dm-FN2 The absolute value of the Euclidean distance between the normal sensor data and the false negative sensor data is greater as the similarity between the normal sensor data and the false negative sensor data increases. Dm-FN1 ,EV Dm-FN2 The sign of the Euclidean distance is negative.

[0180] The evaluation value calculation unit 43 calculates the first evaluation value EV of each normal sensor data. Dm-FP1 ,EV Dm-FP2 and the second evaluation value EV of each normal sensor data Dm-FN1 ,EV Dm-FN2 The result is output to the priority calculation unit 44 .

[0181] Fig.10 It is an explanatory diagram showing two false-negative sensor data, two false-positive sensor data, and four normal sensor data.

[0182] exist Fig.10 , D1, D2, D3, and D4 are each normal sensor data, FP1 and FP2 are each false positive sensor data, and FN1 and FN2 are each false negative sensor data.

[0183] Fig.11 Yes means Fig.10 Graphs for explaining the first evaluation value and the second evaluation value in four pieces of normal sensor data shown.

[0184] exist Fig.11 In FIG. 1 , the numbers show the Euclidean distance as the first evaluation value or the Euclidean distance as the second evaluation value. However, Fig.11 The numbers shown do not represent exact Euclidean distances but are approximate values.

[0185] The priority calculation unit 44 obtains the first evaluation value EV of each normal sensor data from the evaluation value calculation unit 43. Dm-FP1 ,EV Dm-FP2 and the second evaluation value EV of each normal sensor data Dm-FN1 ,EV Dm-FN2 .

[0186] The priority calculation unit 44 calculates the first evaluation value EV based on each normal sensor data as shown in the following equations (1) to (4). Dm-FP1 ,EV Dm-FP2 and the second evaluation value EV of each normal sensor data Dm-FN1 ,EV Dm-FN2, the score SC of each normal sensor data m Perform calculations.

[0187] SC1

[0188] =EV D1-FP1 +EV D1-FP2 +EV D1-FN1 +EV D1-FN2

[0189] =110+100-80-60=70 (1)

[0190] SC2

[0191] =EV D2-FP1 +EV D2-FP2 +EV D2-FN1 +EV D2-FN2

[0192] =180+200-40-40=300 (2)

[0193] SC3

[0194] =EV D3-FP1 +EV D3-FP2 +EV D3-FN1 +EV D3-FN2

[0195] =50+70-160-180=-220 (3)

[0196] SC4

[0197] =EV D4-FP1 +EV D4-FP2 +EV D4-FN1 +EV D4-FN2

[0198] =50+100-40-60=50 (4)

[0199] The priority calculation unit 44 compares the four scores SC1 to SC4 with each other, and calculates the priority of the four normal sensor data D1, D2, D3, and D4 based on the comparison results of the scores SC1 to SC4. m The larger the value, the normal sensor data D m The higher the priority becomes.

[0200] exist Fig.11 In the example, the normal sensor data D2 has the first priority, the normal sensor data D1 has the second priority, the normal sensor data D4 has the third priority, and the normal sensor data D3 has the fourth priority.

[0201] The priority calculation unit 44 outputs the priority levels of the four normal sensor data D1 , D2 , D3 , and D4 to the learning data output unit 45 .

[0202] The learning data output unit 45 selects one or more pieces of normal sensor data from among the four pieces of normal sensor data D1 , D2 , D3 , and D4 output by the data classification unit 42 based on the priority calculated by the priority calculation unit 44 .

[0203] Specifically, the learning data output unit 45 selects N (N is an integer greater than or equal to 1) normal sensor data having the highest priority calculated by the priority calculation unit 44 from among the four normal sensor data D1 , D2 , D3 , and D4 .

[0204] exist Fig.11 In the example of , it is assumed that if N=2, normal sensor data D2 and normal sensor data D1 are selected from four normal sensor data D1, D2, D3, and D4.

[0205] Assuming that N=3, normal sensor data D2 , normal sensor data D1 , and normal sensor data D4 are selected from the four normal sensor data D1 , D2 , D3 , and D4 .

[0206] The learning data output unit 45 outputs the selected normal sensor data to the outside as learning data used for re-learning the learning model.

[0207] Furthermore, the learning data output unit 45 outputs the false positive sensor data FP1 and FP2 to the outside as learning data used for re-learning the learning model.

[0208] In the above-mentioned second embodiment, the learning data selection unit 41 is configured to include the identification information acquisition unit 15, the data classification unit 42, the evaluation value calculation unit 43, the priority calculation unit 44, and the learning data output unit 45. Figure 8 The learning data selection device 2 shown. Therefore, Figure 8 The learning data selection device 2 shown is Figure 1 The learning data selection device 2 shown in the figure can also select learning data that can reduce false detection of the subject that is detected as normal although the fault detection object is abnormal. Figure 8 The learning data selection device 2 shown can select learning data that can reduce false detections that detect an abnormality even though the fault detection target is normal.

[0209] exist Figure 8 In the learning data selection device 2 shown in the figure, the priority calculation unit 44 calculates the first evaluation value EV of each normal sensor data byDm-FP1 ,EV Dm-FP2 and the second evaluation value EV of each normal sensor data Dm-FN1 ,EV Dm-FN2 The scores SC1, SC2, SC3, and SC4 of each normal sensor data are calculated by adding them. However, this is only an example, and the priority calculation unit 44 may also calculate the scores SC1, SC2, SC3, and SC4 of each normal sensor data by, for example, calculating the first evaluation value EV Dm-FP1 ,EV Dm-FP2 and the second evaluation value EV of each normal sensor data Dm-FN1 ,EV Dm-FN2 The weighted average of the scores is calculated to obtain the scores SC1, SC2, SC3, and SC4 of each normal sensor data.

[0210] exist Figure 8 In the learning data selection device 2 shown, the learning data output unit 45 selects N normal sensor data with the highest priority calculated by the priority calculation unit 44 from the four normal sensor data D1, D2, D3, and D4. However, this is just an example, and the learning data output unit 45 may select one normal sensor data from a plurality of normal sensor data in descending order of the priority calculated by the priority calculation unit 44. Then, the selected one normal sensor data is assigned to the learning model 13, and the learning model 13 is relearned until the detection accuracy of the relearned learning model 13 is greater than or equal to the threshold value, and the learning data output unit 45 repeats the selection of one normal sensor data.

[0211] In this case, the learning data selection device 2 can improve the detection accuracy of the re-learned learning model 13 to be greater than or equal to the threshold.

[0212] Implementation method 3.

[0213] In the third embodiment, the learning data selection device 2 will be described in which the priority calculation unit 46 calculates the third evaluation value of each normal sensor data based on the acquisition time of each normal sensor data by the sensor data acquisition unit 11 .

[0214] Fig.12 2 is a diagram showing the structure of the learning data selection device 2 according to the third embodiment. Fig.12 In, with Figure 1 and Figure 8 The same reference numerals denote the same or corresponding parts, and thus their descriptions are omitted.

[0215] The priority calculation unit 46 obtains the first evaluation value of each normal sensor data and the second evaluation value of each normal sensor data from the evaluation value calculation unit 43 .

[0216] The priority calculation unit 46 obtains, from the sensor data acquisition unit 11 , the acquisition time of each normal sensor data by the sensor data acquisition unit 11 .

[0217] The priority calculation unit 46 calculates the third evaluation value of each normal sensor data based on the acquisition time of each normal sensor data. The third evaluation value is a value that becomes smaller as the acquisition time of the normal sensor data becomes earlier.

[0218] The priority calculation unit 46 calculates the priority of each normal sensor data based on the first evaluation value of each normal sensor data, the second evaluation value of each normal sensor data, and the third evaluation value of each normal sensor data.

[0219] The priority calculation unit 46 outputs the priority of each normal sensor data to the learning data output unit 45 .

[0220] Next, Fig.12 The operation of the learning data selection device 2 shown in FIG. 4 is described below. However, except for the priority calculation unit 46, Figure 8 The above-mentioned learning data selection device 2 is the same as that shown in the figure. Therefore, only the operation of the priority calculation unit 46 will be described here.

[0221] exist Fig.12 In the illustrated learning data selection device 2 , for convenience of explanation, an example is described in which normal sensor data are D1 , D2 , D3 , and D4 , false positive sensor data are FP1 , FP2 , and false negative sensor data are FN1 , FN2 .

[0222] The priority calculation unit 46 obtains the normal sensor data D from the evaluation value calculation unit 43. m The first evaluation value EV Dm-FP1 ,EV Dm-FP2 and normal sensor data D m The second evaluation value EV Dm-FN1 ,EV Dm-FN2 . m=1, 2, 3, 4.

[0223] In addition, the priority calculation unit 46 obtains the normal sensor data D related to the sensor data acquisition unit 11 from the sensor data acquisition unit 11. m The acquisition time t m .

[0224] The priority calculation unit 46 calculates the normal sensor data D m The acquisition time t m , for normal sensor data D mThe third evaluation value is calculated based on the normal sensor data D m The acquisition time t m The earlier the time, the smaller the value. Get the time t m The earlier the normal sensor data D m , the distribution of sensor data at the current time, that is, the distribution of sensor data when the fault detection target is normal, is more likely to deviate, so the time t is obtained m Early normal sensor data D m The third evaluation value of is calculated as a small value.

[0225] For example, if the acquisition time t2 of the normal sensor data D2 is the latest, the acquisition time t4 of the normal sensor data D4 is the second latest, the acquisition time t1 of the normal sensor data D1 is the third latest, and the acquisition time t3 of the normal sensor data D3 is the earliest, then the normal sensor data D m The third evaluation value EV 3,m For example, as follows.

[0226] EV 3 3,1 =60

[0227] EV 3 3,2 =100

[0228] EV 3 3,3 =20

[0229] EV 3 3,4 =90

[0230] In this example, in the normal sensor data D m The acquisition time t m In the example, the third evaluation value EV of the normal sensor data D2 is set based on the acquisition time t2 of the latest normal sensor data D2. 3,2 Set the value of to 100.

[0231] The priority calculation unit 46 compares the acquisition time t2 of the normal sensor data D2 with the other normal sensor data D m The acquisition time t m The time difference Δt m Calculate proportional values.

[0232] Then, the priority calculation unit 46 calculates the third evaluation value EV from the normal sensor data D2. 3,2 The value is 100 minus the time difference Δt m The value obtained by the proportional value is set as other normal sensor data D m The third evaluation value EV 3,m The value of .

[0233] Fig.13 It means respectively Fig.10 Graphs for explaining the first evaluation value, the second evaluation value, and the third evaluation value in four pieces of normal sensor data shown.

[0234] The priority calculation unit 46 calculates the priority level based on the normal sensor data D as shown in the following equations (5) to (8). m The first evaluation value EV Dm-FP1 ,EV Dm-FP2 , Normal sensor data D m The second evaluation value EV Dm-FN1 ,EV Dm-FN2 and normal sensor data D m The third evaluation value EV 3,m , for normal sensor data D m Score of SC m Perform calculations.

[0235] SC1

[0236] =EV D1-FP1 +EV D1-FP2 +EV D1-FN1 +EV D1-FN2+EV3,1

[0237] =110+100-80-60+60=130 (5)

[0238] SC2

[0239] =EV D2-FP1 +EV D2-FP2 +EV D2-FN1 +EV D2-FN2+EV3,2

[0240] =180+200-40-40+100=400 (6)

[0241] SC3

[0242] =EV D3-FP1 +EV D3-FP2 +EV D3-FN1 +EV D3-FN2+EV3,3

[0243] =50+70-160-180+20=-200 (7)

[0244] SC4

[0245] =EV D4-FP1 +EV D4-FP2 +EV D4-FN1 +EV D4-FN2+EV3,4

[0246] =50+100-40-60+90=140 (8)

[0247] The priority calculation unit 46 compares the four scores SC1 to SC4 with each other, and calculates the priority of the four normal sensor data D1, D2, D3, and D4 based on the comparison results of the scores SC1 to SC4. m The larger the value, the normal sensor data D m The higher the priority becomes.

[0248] exist Fig.13 In the example, the normal sensor data D2 has the first priority, the normal sensor data D4 has the second priority, the normal sensor data D1 has the third priority, and the normal sensor data D3 has the fourth priority.

[0249] exist Fig.11 In the example, the normal sensor data D2 has the first priority, the normal sensor data D1 has the second priority, the normal sensor data D4 has the third priority, and the normal sensor data D3 has the fourth priority.

[0250] Therefore, the priority level of the normal sensor data D1 and the priority level of the normal sensor data D4 are reversed.

[0251] The priority calculation unit 46 outputs the priority levels of the four normal sensor data D1 , D2 , D3 , and D4 to the learning data output unit 45 .

[0252] In the above-mentioned third embodiment, the priority calculation unit 46 calculates the third evaluation value of each normal sensor data based on the acquisition time of each normal sensor data involved in the sensor data acquisition unit 11, and calculates the priority of each normal sensor data based on the first evaluation value, the second evaluation value and the third evaluation value calculated by the evaluation value calculation unit 43, respectively. Fig.12 The learning data selection device 2 shown. Therefore, Fig.12 The learning data selection device 2 shown is Figure 1 The learning data selection device 2 shown in the figure can also select learning data that can reduce false detection of the subject that is detected as normal although the fault detection object is abnormal. Fig.12 The learning data selection device 2 shown is Figure 8 Similarly, the learning data selection device 2 shown in the figure can select learning data that can reduce the false detection of the main point that the fault detection object is detected as abnormal even though it is normal. Fig.12 The learning data selection device 2 shown is capable of lowering the priority level of each normal sensor data.

[0253] Implementation method 4.

[0254] In the fourth embodiment, a description will be given of a learning data selection device 2 including a relearning unit 50 for relearning a learning model 13 using learning data selected by a learning data selection unit 14 .

[0255] Fig.14 2 is a diagram showing the structure of a learning data selection device 2 according to Embodiment 4. Fig.14 In, with Figure 1 , Figure 8 and Fig.12 The same reference numerals denote the same or corresponding parts, and thus their descriptions are omitted.

[0256] Fig.15 2 is a hardware configuration diagram showing the hardware of the learning data selection device 2 according to the fourth embodiment. Fig.15 In, with Figure 2 and Fig. 9 The same reference numerals denote the same or corresponding parts, and thus their descriptions are omitted.

[0257] Fig.14 The illustrated learning data selection device 2 includes a sensor data acquisition unit 11 , a detection data acquisition unit 12 , a learning data selection unit 14 , and a re-learning unit 50 .

[0258] The relearning unit 50 performs, for example, Fig.15 The relearning circuit 26 shown is implemented.

[0259] The re-learning unit 50 re-learns the learning model 13 using the learning data selected by the learning data selection unit 14 .

[0260] exist Fig.14 In the learning data selection device 2 shown in FIG. 1 , the re-learning unit 50 is applied to Figure 1 However, this is only an example, and the re-learning unit 50 can also be applied to Figure 8 The learning data selection device 2 shown or Fig.12 The learning data selection device 2 is shown.

[0261] Imagine Fig.14 In the learning data selection device 2, the sensor data acquisition unit 11, the detection data acquisition unit 12, the learning data selection unit 14 and the re-learning unit 50 are each configured by Fig.15 Specifically, it is assumed that the learning data selection device 2 is implemented by the sensor data acquisition circuit 21 , the detection data acquisition circuit 22 , the learning data selection circuit 24 , and the re-learning circuit 26 .

[0262] The sensor data acquisition circuit 21 , the detection data acquisition circuit 22 , the learning data selection circuit 24 and the relearning circuit 26 are each, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA or a combination thereof.

[0263] The structural elements of the learning data selection device 2 are not limited to those implemented by dedicated hardware. The learning data selection device 2 can also be implemented by software, firmware, or a combination of software and firmware.

[0264] When the learning data selection device 2 is implemented by software or firmware, a program for causing a computer to execute the processing sequence of the sensor data acquisition unit 11, the detection data acquisition unit 12, the learning data selection unit 14, and the re-learning unit 50 is stored in Figure 3 The memory 31 shown. Moreover, Figure 3 The processor 32 shown executes the program stored in the memory 31 .

[0265] In addition, Fig.15 In FIG. 1 , an example is shown in which each of the structural elements of the learning data selection device 2 is implemented by dedicated hardware. Figure 3 In the example in which the learning data selection device 2 is implemented by software or firmware, etc., is shown. However, this is only an example, and some of the structural elements in the learning data selection device 2 may be implemented by dedicated hardware, and the remaining structural elements may be implemented by software or firmware, etc.

[0266] Next, Fig.14 The operation of the learning data selection device 2 shown in FIG. 1 is described below. However, except for the relearning unit 50, Figure 1 The same is true for the learning data selection device 2 shown. Therefore, only the operation of the re-learning unit 50 will be described here.

[0267] The re-learning unit 50 acquires normal sensor data and false positive sensor data from the learning data selection unit 14 as learning data used for re-learning the learning model 13 .

[0268] The relearning unit 50 uses the normal sensor data and the false positive sensor data to relearn the learning model 13. Thus, the learning model 13 uses the normal sensor data and the false positive sensor data to relearn the distribution of the sensor data when the fault detection object is normal. The distribution of the sensor data after relearning does not include the sensor data related to the false negative detection data. As a result, the distribution of the sensor data when the fault detection object is normal becomes an appropriate distribution.

[0269] When the re-learning unit 50 re-learns the learning model 13, Fig.16As shown, if the proportion of the learning data selected by the learning data selection unit 14 in the learning data used in the previous learning of the learning model 13 is less than or equal to the threshold, the hyperparameters of the relearned learning model 13 can be adjusted. The threshold can be stored in the internal memory of the relearning unit 50 or given from the outside of the learning data selection device 2.

[0270] Fig.16 It is an explanatory diagram showing the ratio of the learning data output from the learning data selection unit 14 included in the learning data used for the previous learning of the learning model 13 .

[0271] exist Fig.16 In FIG. 1 , ○ is normal sensor data, and ○ with diagonal lines is normal sensor data outputted from the learning data selection unit 14 as learning data among the learning data used in the previous learning of the learning model 13 . △ is false positive sensor data outputted from the learning data selection unit 14 as learning data.

[0272] The hyperparameters of the learning model 13 are, for example, values ​​for controlling the behavior of the algorithm in the learning model 13. By adjusting the hyperparameters of the learning model 13, for example, it is expected that the performance of the learning model 13 will be improved, overlearning will be suppressed, or learning efficiency will be improved.

[0273] In the above-mentioned fourth embodiment, the relearning unit 50 is provided to relearn the learning model 13 using the learning data selected by the learning data selection unit 14. Fig.14 The learning data selection device 2 shown. Therefore, Fig.14 The learning data selection device 2 shown is Figure 1 The learning data selection device 2 shown can similarly select learning data that can reduce false detections of fault detection targets that are detected as normal despite being abnormal, and can also relearn the distribution of sensor data for the learning model 13 .

[0274] Implementation method 5.

[0275] In the fifth embodiment, an abnormality detection device including the abnormality detection unit 60 will be described.

[0276] Fig.17 : is a structural diagram showing an abnormality detection device involved in Embodiment 5. Fig.17 In, with Fig.14 The same reference numerals denote the same or corresponding parts, and thus their descriptions are omitted.

[0277] Fig.18 FIG. 4 is a hardware configuration diagram showing the hardware of the abnormality detection device according to Embodiment 5. Fig.18In, with Fig.15 The same reference numerals denote the same or corresponding parts, and thus their descriptions are omitted.

[0278] Fig.17 The abnormality detection device shown includes a learning data selection device 2 and an abnormality detection unit 60 .

[0279] exist Fig.17 In the abnormality detection device shown, the learning model 13 is provided outside the detection data acquisition unit 12. However, this is just an example, and the detection data acquisition unit 12 may also include the learning model 13. The learning model 13 is a learning model relearned by the relearning unit 50.

[0280] The abnormality detection unit 60 performs, for example, Fig.18 The abnormality detection circuit 27 shown is implemented.

[0281] After the relearning unit 50 relearns the learning model 13 , the abnormality detection unit 60 assigns the sensor data obtained by the sensor data acquisition unit 11 to the relearned learning model 13 , and obtains detection data indicating whether the fault detection object is normal or abnormal from the relearned learning model 13 .

[0282] The abnormality detection unit 60 outputs the detection data to the outside.

[0283] Imagine Fig.17 In the embodiment, the sensor data acquisition unit 11, the detection data acquisition unit 12, the learning data selection unit 14, the relearning unit 50 and the abnormality detection unit 60, which are the structural elements of the abnormality detection device, are each Fig.18 Specifically, it is assumed that the abnormality detection device is realized by the sensor data acquisition circuit 21, the detection data acquisition circuit 22, the learning data selection circuit 24, the relearning circuit 26, and the abnormality detection circuit 27.

[0284] The sensor data acquisition circuit 21, the detection data acquisition circuit 22, the learning data selection circuit 24, the relearning circuit 26 and the abnormality detection circuit 27 are each, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA or a combination thereof.

[0285] The structural elements of the abnormality detection device are not limited to those implemented by dedicated hardware, and the learning data selection device 2 can also be implemented by software, firmware, or a combination of software and firmware.

[0286] When the abnormality detection device is implemented by software or firmware, a program for causing a computer to execute the processing sequence of the sensor data acquisition unit 11, the detection data acquisition unit 12, the learning data selection unit 14, the relearning unit 50, and the abnormality detection unit 60 is stored in Figure 3 The memory 31 shown. Moreover, Figure 3 The processor 32 shown executes the program stored in the memory 31 .

[0287] In addition, Fig.18 In FIG. 1 , an example is shown in which each of the structural elements of the abnormality detection device is implemented by dedicated hardware. Figure 3 In the embodiment, an example is shown in which the abnormality detection device is implemented by software or firmware. However, this is only an example, and some structural elements in the abnormality detection device may be implemented by dedicated hardware, and the remaining structural elements may be implemented by software or firmware.

[0288] Next, Fig.17 The operation of the abnormality detection device shown in FIG. Fig.14 The above-mentioned learning data selection device 2 is the same as that shown in FIG. Therefore, only the operation of the abnormality detection unit 60 will be described here.

[0289] The abnormality detection unit 60 acquires the sensor data from the sensor data acquisition unit 11 after the relearning unit 50 relearns the learning model 13 .

[0290] The abnormality detection unit 60 gives the sensor data to the relearned learning model 13 , and acquires detection data indicating whether the fault detection target is normal or abnormal from the relearned learning model 13 .

[0291] The abnormality detection unit 60 outputs the detection data to the outside.

[0292] In the above embodiment 5, Fig.14 The illustrated learning data selection device 2 and the abnormality detection unit 60 constitute an abnormality detection device, and after the learning model 13 is relearned by the relearning unit 50, the abnormality detection unit 60 assigns the sensor data obtained by the sensor data acquisition unit 11 to the relearned learning model 13, and obtains detection data indicating whether the fault detection object is normal or abnormal from the relearned learning model 13. Therefore, the abnormality detection device can reduce the false detection of the main point that the fault detection object is detected as normal although it is abnormal.

[0293] Furthermore, the present invention can be a free combination of the various embodiments, or arbitrary components of the various embodiments can be modified, or arbitrary components can be omitted in the various embodiments.

[0294] Industrial Applicability

[0295] The present invention is applied to a learning data selection device, a learning data selection method and an anomaly detection device.

[0296] Description of the label

[0297] 1 sensor, 2 learning data selection device, 3 human-machine IF unit, 11 sensor data acquisition unit, 12 detection data acquisition unit, 13 learning model, 14 learning data selection unit, 15 identification information acquisition unit, 16 sensor data selection unit, 17 learning data output unit, 21 sensor data acquisition circuit, 22 detection data acquisition circuit, 24, 25 learning data selection circuit, 26 re-learning circuit, 27 abnormality detection circuit, 31 memory, 32 processor, 41 learning data selection unit, 42 data classification unit, 43 evaluation value calculation unit, 44 priority calculation unit, 45 learning data output unit, 46 priority calculation unit, 50 re-learning unit, 60 abnormality detection unit.

Claims

1. A learning data selection device, comprising: a sensor data acquisition unit that acquires a plurality of sensor data indicating an observation result of the fault detection target from a sensor that observes the fault detection target; a detection data acquisition unit that assigns each sensor data acquired by the sensor data acquisition unit to a learning model that has learned the distribution of sensor data when the fault detection object is normal, and acquires detection data indicating whether the fault detection object is normal or abnormal from the learning model; and A learning data selection unit, which obtains identification information for identifying which sensor data among a plurality of sensor data obtained by the sensor data acquisition unit is sensor data related to false negative detection data indicating that the fault detection object is abnormal but normal, and selects, based on the identification information, sensor data related to detection data indicating that the fault detection object is normal from sensor data other than sensor data related to the false negative detection data among the plurality of sensor data obtained by the sensor data acquisition unit as learning data used for re-learning the learning model.

2. The learning data selection device according to claim 1, characterized in that: The learning data selection unit obtains identification information for identifying which sensor data among the multiple sensor data obtained by the sensor data acquisition unit is sensor data related to false positive detection data indicating an abnormality although the fault detection object is normal, and based on the identification information, selects the sensor data related to the false positive detection data as learning data used for re-learning the learning model from the multiple sensor data obtained by the sensor data acquisition unit.

3. The learning data selection device according to claim 2, characterized in that: The learning data selection unit has: an identification information acquisition unit that acquires identification information indicating which sensor data, among the plurality of sensor data acquired by the sensor data acquisition unit, is the sensor data related to the false-negative detection data, which sensor data is the sensor data related to the false-positive detection data, and which sensor data is the sensor data related to the normal detection data indicating that the fault detection object is normal; a sensor data selection unit that selects, from among the plurality of sensor data acquired by the sensor data acquisition unit, sensor data related to the normal detection data and sensor data related to the false positive detection data based on the identification information acquired by the identification information acquisition unit; as well as The learning data output unit outputs the sensor data selected by the sensor data selection unit as learning data used for re-learning the learning model.

4. The learning data selection device according to claim 3, characterized in that: If the identification information acquisition unit acquires identification information indicating which sensor data is the sensor data related to the false-negative detection data and which sensor data is the sensor data related to the false-positive detection data among the plurality of sensor data acquired by the sensor data acquisition unit, and does not acquire identification information indicating which sensor data is the sensor data related to the normal detection data, The sensor data selection unit then selects sensor data other than the sensor data involving the false-negative detection data from the sensor data involving the detection data indicating that the fault detection object is normal among the multiple sensor data acquired by the sensor data acquisition unit, as the sensor data involving the detection data during normal times.

5. The learning data selection device according to claim 3, characterized in that: If the similarity between the sensor data involved in the normal detection data and the sensor data involved in the false positive detection data selected by the sensor data selection unit is greater than or equal to a threshold value, the learning data output unit outputs the sensor data involved in the normal detection data as learning data used for re-learning the learning model; if the similarity is less than the threshold value, the sensor data involved in the normal detection data is discarded.

6. The learning data selection device according to claim 2, characterized in that: The learning data selection unit has: an identification information acquisition unit that acquires identification information indicating which sensor data, among the plurality of sensor data acquired by the sensor data acquisition unit, is the sensor data related to the false-negative detection data, which sensor data is the sensor data related to the false-positive detection data, and which sensor data is the sensor data related to the normal detection data indicating that the fault detection object is normal; a data classification unit that classifies each sensor data acquired by the sensor data acquisition unit into false-negative sensor data related to the false-negative detection data, false-positive sensor data related to the false-positive detection data, or normal sensor data related to the normal detection data based on the identification information acquired by the identification information acquisition unit; an evaluation value calculation unit, if there are multiple normal sensor data classified by the data classification unit, then, as a first evaluation value of each normal sensor data, an evaluation value having a larger absolute value and a positive sign is calculated as the higher the similarity between each normal sensor data and the false positive sensor data is, and, as a second evaluation value of each normal sensor data, an evaluation value having a larger absolute value and a negative sign is calculated as the higher the similarity between each normal sensor data and the false negative sensor data is; a priority calculation unit that calculates the priority of each normal sensor data based on the first evaluation value and the second evaluation value calculated by the evaluation value calculation unit; as well as A learning data output unit selects one or more normal sensor data from the plurality of normal sensor data based on the priority calculated by the priority calculation unit, and outputs the selected normal sensor data and the false positive sensor data as learning data used for re-learning the learning model.

7. The learning data selection device according to claim 6, characterized in that: If the identification information acquisition unit acquires identification information indicating which sensor data is the sensor data related to the false-negative detection data and which sensor data is the sensor data related to the false-positive detection data among the plurality of sensor data acquired by the sensor data acquisition unit, and does not acquire identification information indicating which sensor data is the sensor data related to the normal detection data, The data classification unit selects sensor data other than the sensor data involved in the false-negative detection data as the sensor data involved in the normal detection data from the sensor data involved in the detection data indicating that the fault detection object is normal among the multiple sensor data acquired by the sensor data acquisition unit based on the identification information indicating which sensor data is the sensor data involved in the false-negative detection data, and classifies the sensor data involved in the normal detection data as the normal sensor data.

8. The learning data selection device according to claim 6, characterized in that: The learning data output unit selects N normal sensor data having the highest priority calculated by the priority calculation unit from among the plurality of normal sensor data, where N is an integer greater than or equal to 1.

9. The learning data selection device according to claim 6, characterized in that: The learning data output unit selects one normal sensor data from the multiple normal sensor data in descending order of priority calculated by the priority calculation unit, assigns the selected one normal sensor data to the learning model, and re-learns the learning model until the detection accuracy of the re-learned learning model is greater than or equal to a threshold, and repeats the selection of one normal sensor data.

10. The learning data selection device according to claim 6, characterized in that: The evaluation value calculation unit, As the first evaluation value of each normal sensor data, a Euclidean distance whose absolute value is larger and whose sign is positive as the similarity between each normal sensor data and the false positive sensor data is higher is calculated. As the second evaluation value of each normal sensor data, a Euclidean distance whose absolute value becomes larger and whose sign becomes negative as the similarity between each normal sensor data and the false-negative sensor data increases is calculated.

11. The learning data selection device according to claim 6, characterized in that: The priority calculation unit calculates a third evaluation value for each normal sensor data based on the acquisition time of each normal sensor data by the sensor data acquisition unit. The priority level of each normal sensor data is calculated based on the first evaluation value and the second evaluation value calculated by the evaluation value calculation unit and the third evaluation value.

12. The learning data selection device according to claim 1, characterized in that: A re-learning unit is provided for re-learning the learning model using the learning data selected by the learning data selection unit.

13. The learning data selection device according to claim 12, characterized in that: The re-learning unit adjusts the hyper parameters of the re-learned learning model if the proportion of the learning data selected by the learning data selection unit in the learning data used in the previous learning of the learning model is less than or equal to a threshold.

14. A learning data selection method, The sensor data acquisition unit acquires a plurality of sensor data indicating observation results of the fault detection target from a sensor that observes the fault detection target. The detection data acquisition unit assigns each sensor data acquired by the sensor data acquisition unit to a learning model that has learned the distribution of sensor data when the fault detection object is normal, and acquires detection data indicating whether the fault detection object is normal or abnormal from the learning model, respectively. The learning data selection unit obtains identification information for identifying which sensor data among the multiple sensor data obtained by the sensor data acquisition unit is sensor data related to false negative detection data indicating that the fault detection object is abnormal but normal, and based on the identification information, selects sensor data related to detection data indicating that the fault detection object is normal from sensor data other than sensor data related to the false negative detection data among the multiple sensor data obtained by the sensor data acquisition unit as learning data used for re-learning the learning model.

15. An abnormality detection device, characterized in that: have: The learning data selection device as claimed in claim 12; and The abnormality detection unit, after relearning the learning model through the relearning unit, assigns the sensor data obtained by the sensor data acquisition unit to the relearned learning model, and obtains detection data indicating whether the fault detection object is normal or abnormal from the relearned learning model.