Machine learning device and machine learning method

Abstract

To provide a machine learning device in which the accuracy of detecting an abnormality in an object to be detected can be enhanced.SOLUTION: A machine learning device according to one embodiment of the present disclosure includes: a smoothing circuit that can generate second flag data by performing a smoothing process on first flag data which is time-series data indicating whether or not an abnormality has occurred in an object to be detected and corresponding to a plurality of measurement data sets which are time-series data about a plurality of parameters of the object to be detected; and a machine learning circuit that can perform a machine learning process by performing a regression analysis in such a way that when the plurality of measurement data sets are inputted to a machine learning model which is configured to output two or more measurement data sets of the plurality of measurement data sets in response to input of the plurality of measurement data sets, a prediction value corresponding to two or more measurement data outputted from the machine learning model becomes equal to the second flag data.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to a machine learning device and a machine learning method for performing machine learning processing.

Background Art

[0002] In recent years, technologies for detecting abnormalities in detection targets using machine learning technologies have been developed. For example, Patent Document 1 discloses a technology for detecting abnormalities in a transmission in a vehicle using a machine learning technology.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In such an apparatus for detecting an abnormality in a detection target, it is desired that the accuracy of detecting the abnormality is high, and a further improvement in the accuracy of detecting the abnormality is expected.

[0005] It is desirable to provide a machine learning device and a machine learning method capable of enhancing the accuracy of detecting an abnormality.

Means for Solving the Problems

[0006] A machine learning device according to an embodiment of the present disclosure includes a smoothing circuit and a machine learning circuit. The smoothing circuit can generate second flag data by performing a smoothing process on first flag data, which is time-series data indicating a period during which an abnormality has occurred in a detection target object, in response to a plurality of measurement data that is time-series data of a plurality of parameters in the detection target object. The machine learning circuit can perform machine learning processing by inputting the plurality of measurement data into a machine learning model in which the plurality of measurement data is input and two or more measurement data that is a part of the plurality of measurement data is output, and performing regression analysis so that a predicted value corresponding to the two or more measurement data output from the machine learning model is the same as the second flag data.

[0007] A machine learning method according to an embodiment of the present disclosure includes generating second flag data by performing a smoothing process on first flag data, which is time-series data indicating a period during which an abnormality has occurred in a detection target object, in response to a plurality of measurement data that is time-series data of a plurality of parameters in the detection target object; inputting the plurality of measurement data into a machine learning model in which the plurality of measurement data is input and two or more measurement data that is a part of the plurality of measurement data is output; and performing machine learning processing by performing regression analysis so that a predicted value corresponding to the two or more measurement data output from the machine learning model is the same as the second flag data.

Advantages of the Invention

[0008] According to the machine learning device and the machine learning method according to an embodiment of the present disclosure, the accuracy of detecting an abnormality in a detection target object can be improved.

Brief Description of the Drawings

[0009]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings.

[0011] <Embodiment> [Configuration Example] FIG. 1 shows a configuration example of a machine learning device (machine learning device 1) according to an embodiment. In this example, an anomaly detection device 100 described later performs anomaly detection on a detection target using a machine learning model generated by the machine learning device 1. In this example, the detection target is a vehicle, but it is not limited to this, and various objects can be the detection target. The machine learning device 1 is, for example, a personal computer, and includes a processor, a memory, a storage, etc., and performs processing by executing a program. The machine learning device 1 includes a storage 10, a processing unit 20, and a user interface 30.

[0012] The storage 10 is configured using a non-volatile storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), and is configured to store a plurality of data sets DS1. The plurality of data sets DS1 are sets of data used in machine learning processing. Each of the plurality of data sets DS1 includes input data DT1 and flag data DF1.

[0013] The input data DT1 includes a plurality of measurement data DD which are time-series data of a plurality of parameters of the detection target. When the detection target is a vehicle, the plurality of parameters are, for example, hydraulic pressure, engine speed, drive torque, etc.

[0014] FIG. 2 shows an example of a plurality of measurement data DD in the input data DT1. The horizontal axis represents time, and the vertical axis represents parameter values. The input data DT1 may include, for example, about 10 to 100 measurement data DD. In FIG. 2, five of the measurement data DD are shown. The five measurement data DD are time-series data of five different parameters, and in this example, they have completely different waveforms from each other. When an anomaly occurs in the detection target, it is expected that a waveform corresponding to the anomaly will appear in at least a part of the about 10 to 100 measurement data DD.

[0015] The flag data DF1 is time-series data indicating whether an abnormality has occurred in the object to be detected. This flag data DF1 is created, for example, when an engineer identifies a period during which an abnormality has occurred in the object to be detected based on a plurality of measurement data DD included in the input data DT1. This flag data DF1 can take the value of "1" or "0". In this example, the flag data DF1 indicates "1" during a period when an abnormality has occurred and "0" during a period when no abnormality has occurred.

[0016] The storage 10 stores a plurality of data sets DS1, each of which includes such input data DT1 and flag data DF1. The machine learning device 1 is configured to perform machine learning processing based on this plurality of data sets DS1.

[0017] The processing unit 20 (FIG. 1) includes a smoothing processing unit 21, a machine learning processing unit 22, a statistical processing unit 23, a display processing unit 24, and a threshold setting unit 25.

[0018] The smoothing processing unit 21 is configured to generate flag data DF2 by performing smoothing processing on the flag data DF1. Specifically, the smoothing processing unit 21 can perform smoothing processing using, for example, a Gaussian distribution.

[0019] FIG. 3 shows an operation example of the smoothing processing unit 21. (A) shows the waveform of the flag data DF1, and (B) shows the waveform of the flag data DF2. As shown in FIG. 3(A), the flag data DF1 changes from "0" to "1" at timing t1 and from "1" to "0" at timing t2. Since this flag data DF1 indicates "1" during a period when an abnormality has occurred and "0" during a period when no abnormality has occurred, the period from timing t1 to t2 is an abnormal period P during which an abnormality has occurred in the object to be detected.

[0020] The smoothing processing unit 21 (Fig. 1) generates the flag data DF2 shown in Fig. 3(B) by performing smoothing processing on such flag data DF1 (Fig. 3(A)). The flag data DF2 continuously changes from "0" to "1" before the timing t1 when the flag data DF1 changes from "0" to "1", and continuously changes from "1" to "0" after the timing t2 when the flag data DF1 changes from "1" to "0". In this way, the smoothing processing unit 21 performs smoothing processing so that the value of the flag data DF2 continuously changes between "0" and "1" during the period when the flag data DF1 is "0".

[0021] The machine learning processing unit 22 is configured to generate a machine learning model M by performing machine learning processing based on the input data DT1 including a plurality of measurement data DD and the flag data DF2. In this machine learning model M, the input data DT1 including a plurality of measurement data DD is input, and the extraction data DT2 including two or more measurement data DD related to anomalies among these plurality of measurement data DD is output. The machine learning processing unit 22 performs machine learning processing by performing regression analysis so that the predicted value corresponding to two or more measurement data DD output from this machine learning model M is the same as the flag data DF2. Here, the predicted value is the regression line in regression analysis. Specifically, for example, the machine learning processing unit 22 sequentially selects two or more measurement data DD from the plurality of measurement data DD and calculates a predicted value based on the selected two or more measurement data DD. Then, the machine learning processing unit 22 performs regression analysis so as to select two or more measurement data DD such that this predicted value matches the flag data DF2 supplied from the smoothing processing unit 21. In other words, the machine learning processing unit 22 performs machine learning processing so as to extract two or more measurement data DD that can reproduce the flag data DF2 from the plurality of measurement data DD included in the input data DT1. The regression analysis is performed using, for example, methods such as random forest and XGBoost.

[0022] The machine learning processing unit 22 generates a machine learning model M by performing this machine learning process for each of a plurality of data sets DS1. Through this machine learning process, the machine learning model M can output two or more measurement data DD related to anomalies among these plurality of measurement data DD when a plurality of measurement data DD are input. Then, this machine learning model M is used in an anomaly detection device 100 (Fig. 4) described later. This anomaly detection device 100 is configured to perform anomaly detection of a detection target using this machine learning model M.

[0023] Also, the machine learning processing unit 22 is configured to generate extraction data DT2 including two or more measurement data DD among the plurality of measurement data DD included in the input data DT1 using the generated machine learning model M.

[0024] The statistical processing unit 23 is configured to generate statistical data DT3 by performing statistical processing based on two or more measurement data DD included in the extraction data DT2. In this statistical processing, the statistical processing unit 23 is configured to generate statistical data DT3 indicating time-series data of Mahalanobis distances by calculating Mahalanobis distances based on two or more measurement data DD.

[0025] The display processing unit 24 is configured to generate data to be displayed on the display device in the user interface 30 based on the extraction data DT2, the statistical data DT3, and the flag data DF2.

[0026] The threshold setting unit 25 is configured to set a threshold TH used when the anomaly detection device 100 described later detects an anomaly based on a user operation on the user interface 30.

[0027] The user interface 30 is configured to present information to the user and receive the user's operations, and includes, for example, a display device, a keyboard, a mouse, etc. The engineer determines how to set the threshold TH, for example, by viewing the information related to the extracted data DT2, the statistical data DT3, and the flag data DF2 displayed on the display device, and inputs the threshold TH into the keyboard.

[0028] Next, an anomaly detection device 100 that detects anomalies in the object to be detected using the machine learning model M generated by the machine learning device 1 will be described in detail.

[0029] FIG. 4 shows a configuration example of the anomaly detection device 100. The anomaly detection device 100 is, for example, a personal computer, and includes a processor, a memory, a storage, etc., and performs processing by executing a program. The anomaly detection device 100 includes a storage 110, a processing unit 120, and a user interface 130.

[0030] The storage 110 is configured using a non-volatile storage device such as an HDD or an SSD, similar to the storage 10 in the machine learning device 1, and is configured to store a plurality of data sets DS101 in this example. The plurality of data sets DS101 are, for example, sets of data related to a plurality of objects to be detected for anomaly detection. Each of the plurality of data sets DS101 includes input data DT101.

[0031] The input data DT101 includes a plurality of measurement data DD that are time-series data for a plurality of parameters in the object to be detected, similar to the input data DT1 (FIG. 1). The input data DT101 can include, for example, about 10 to 100 pieces of measurement data DD.

[0032] The processing unit 120 has an extraction processing unit 122, a statistical processing unit 123, and a detection processing unit 124.

[0033] The extraction processing unit 122 is configured to output extraction data DT102 including two or more measurement data DD related to anomalies among these plurality of measurement data DD by using the machine learning model M generated by the machine learning device 1 based on the input data DT101 including the plurality of measurement data DD.

[0034] Similar to the statistical processing unit 23 in the machine learning device 1, the statistical processing unit 123 is configured to generate statistical data DT103 by performing statistical processing based on two or more measurement data DD included in the extraction data DT102. In this statistical processing, similar to the statistical processing unit 23 in the machine learning device 1, the statistical processing unit 123 is configured to generate statistical data DT103 indicating time-series data of Mahalanobis distances by calculating Mahalanobis distances based on two or more measurement data DD.

[0035] The detection processing unit 124 is configured to generate data to be displayed on the display device in the user interface 130 based on the extraction data DT102 and the statistical data DT103, and to detect anomalies of the object to be detected based on the threshold TH generated by the machine learning device 1.

[0036] The user interface 130 is configured to present information to the user and receive the user's operations, and includes, for example, a display device, a keyboard, a mouse, and the like.

[0037] Here, the smoothing processing unit 21 corresponds to a specific example of the "smoothing circuit" in an embodiment of the present disclosure. The machine learning processing unit 22 corresponds to a specific example of the "machine learning circuit" in an embodiment of the present disclosure. The machine learning model M corresponds to a specific example of the "machine learning model" in an embodiment of the present disclosure. The plurality of measurement data DD included in the input data DT1 corresponds to a specific example of the "plurality of measurement data" in an embodiment of the present disclosure. The two or more measurement data DD included in the extraction data DT2 corresponds to a specific example of the "two or more measurement data" in an embodiment of the present disclosure. The flag data DF1 corresponds to a specific example of the "first flag data" in an embodiment of the present disclosure. The flag data DF2 corresponds to a specific example of the "second flag data" in an embodiment of the present disclosure.

[0038] [Operations and functions] Subsequently, the operations and functions of the machine learning device 1 and the abnormality detection device 100 of the present embodiment will be described.

[0039] (Overall operation overview) First, with reference to FIGS. 1 and 3, the overall operation overview of the machine learning device 1 and the abnormality detection device 100 will be described.

[0040] In the machine learning device 1 (Fig. 1), the smoothing processing unit 21 of the processing unit 20 generates flag data DF2 by performing smoothing processing on the flag data DF1. The machine learning processing unit 22 generates a machine learning model M by performing machine learning processing based on a plurality of measurement data DD included in the input data DT1 and the flag data DF2. Further, the machine learning processing unit 22 generates extraction data DT2 including two or more pieces of measurement data DD among the plurality of measurement data DD included in the input data DT1 by using the generated machine learning model M. The statistical processing unit 23 generates statistical data DT3 by performing statistical processing based on two or more pieces of measurement data DD included in the extraction data DT2. The display processing unit 24 generates data to be displayed on the display device in the user interface 30 based on the extraction data DT2, the statistical data DT3, and the flag data DF2. The threshold setting unit 25 sets a threshold TH used when the abnormality detection device 100 detects an abnormality based on a user operation on the user interface 30. This threshold TH is set in the detection processing unit 124 of the abnormality detection device 100 (Fig. 4).

[0041] In the abnormality detection device 100 (Fig. 4), the extraction processing unit 122 generates extraction data DT102 including two or more pieces of measurement data DD related to an abnormality based on the input data DT101 including a plurality of measurement data DD by using the machine learning model M generated by the machine learning device 1. The statistical processing unit 123 generates statistical data DT103 by performing statistical processing based on two or more pieces of measurement data DD included in the extraction data DT102. The detection processing unit 124 generates data to be displayed on the display device in the user interface 130 based on the extraction data DT102 and the statistical data DT103, and detects an abnormality of the detection target object based on the threshold TH generated by the machine learning device 1.

[0042] (Detailed operation) Next, the operations of the machine learning device 1 and the abnormality detection device 100 will be described in detail. The abnormality detection device 100 detects an abnormality of the detection target object by using the machine learning model M generated by the machine learning device 1.

[0043] FIG. 5 shows an operation example of the machine learning device 1 and the abnormality detection device 100. (A) shows the waveforms of five measurement data DD included in the extraction data DT102, (B) shows the waveform of the statistical data DT103, and (C) shows the waveforms of the flag data DF1 and DF2 when the machine learning device 1 generates the machine learning model M.

[0044] When the machine learning device 1 performs machine learning processing, the engineer creates the flag data DF1 by specifying the period during which an abnormality has occurred in the detection target based on a plurality of measurement data DD that are time-series data of a plurality of parameters in the detection target (FIG. 5(C)). In this example, the engineer determines that an abnormality has occurred in the detection target during the period from timing t11 to t12 based on the plurality of measurement data DD. This period from timing t11 to t12 is the abnormal period P. The smoothing processing unit 21 generates the flag data DF2 by performing smoothing processing on this flag data DF1 (FIG. 5(C)). The machine learning processing unit 22 generates the machine learning model M by performing machine learning processing using this flag data DF2. Further, the machine learning processing unit 22 generates extraction data DT2 including two or more of the plurality of measurement data DD included in the input data DT1 using the generated machine learning model M. The statistical processing unit 23 generates statistical data DT3 by performing statistical processing based on two or more of the measurement data DD included in the extraction data DT2. The display processing unit 24 generates data to be displayed on the display device in the user interface 30 based on the extraction data DT2, the statistical data DT3, and the flag data DF2. The engineer determines how to set the threshold TH based on the data displayed on the user interface 30 and inputs the threshold TH to the keyboard. The threshold setting unit 25 sets the threshold TH based on the user's operation on the user interface 30. The threshold TH is set to a value of, for example, "10" or more.

[0045] The anomaly detection device 100 performs anomaly detection on the object to be detected using the machine learning model M generated in this way. Specifically, first, the extraction processing unit 122 uses the machine learning model M generated by the machine learning device 1 to generate extraction data DT102 (Fig. 5(A)) including two or more measurement data DD related to anomalies based on the input data DT101 including a plurality of measurement data DD. In this example, the extraction data DT102 includes five measurement data DD. In Fig. 5, the parameter values of the measurement data DD are normalized in the range of 0 or more and 1 or less.

[0046] Then, the statistical processing unit 123 performs statistical processing based on two or more measurement data DD included in the extraction data DT102 to generate statistical data DT103 (Fig. 5(B)). In this example, the statistical processing unit 123 calculates the Mahalanobis distance based on five measurement data DD to generate statistical data DT103 indicating the time series data of the Mahalanobis distance between these five measurement data DD. A small Mahalanobis distance indicates a small variation between the five measurement data DD, and a large Mahalanobis distance indicates a large variation between the five measurement data DD. When an anomaly occurs in the object to be detected, the Mahalanobis distance increases.

[0047] The detection processing unit 124 compares this Mahalanobis distance with the threshold TH. Then, the detection processing unit 124 determines that the period during which the Mahalanobis distance exceeds the threshold TH is the period during which an anomaly has occurred in the object to be detected. In Fig. 5, the period during which the Mahalanobis distance exceeds the threshold TH is shown shaded.

[0048] In this example, the period during which the Mahalanobis distance exceeds the threshold TH is within the anomaly period P set by the engineer. That is, the machine learning device 1 can correctly perform machine learning processing based on the flag data DF1, and the accuracy of the machine learning model M generated by the machine learning device 1 is high.

[0049] In this way, in the machine learning device 1, the flag data DF2 is generated by performing smoothing processing on the flag data DF1, and the machine learning model M is generated using this flag data DF2. As a result, the machine learning device 1 can correctly perform machine learning processing and improve the accuracy of the machine learning model M. As a result, the anomaly detection device 100 can improve the accuracy of detecting anomalies in the object to be detected.

[0050] (Reference Example) Next, the operation of the present embodiment will be described in comparison with the machine learning device 1R according to the reference example. In this reference example, in the machine learning device 1 (FIG. 1) according to the present embodiment, the smoothing processing unit 21 is omitted, and the machine learning processing unit 22 performs machine learning processing based on the input data DT1 and the flag data DF1.

[0051] FIG. 6 shows a configuration example of the machine learning device 1R according to this reference example. The machine learning device 1R includes a processing unit 20R. The processing unit 20R includes a machine learning processing unit 22, a statistical processing unit 23, a display processing unit 24, and a threshold setting unit 25. That is, the processing unit 20R is obtained by omitting the smoothing processing unit 21 from the machine learning device 1 (FIG. 1) according to the present embodiment.

[0052] The machine learning processing unit 22 is configured to generate a machine learning model MR by performing machine learning processing based on the input data DT1 including a plurality of measurement data DD and the flag data DF1. The machine learning processing unit 22 performs machine learning processing by performing regression analysis so that the predicted value corresponding to two or more pieces of measurement data DD output from this machine learning model MR becomes the same as the flag data DF1. That is, the machine learning processing unit 22 according to the present embodiment performs regression analysis so that the predicted value corresponding to two or more pieces of measurement data DD becomes the same as the flag data DF2 supplied from the smoothing processing unit 21, but the machine learning processing unit 22 according to this reference example performs regression analysis so that the predicted value corresponding to two or more pieces of measurement data DD becomes the same as the flag data DF1 supplied from the storage 10.

[0053] FIG. 7 shows an operation example of the anomaly detection device 100 when using the machine learning model MR generated by the machine learning device 1R. (A) shows the waveforms of five measurement data DD included in the extraction data DT102, (B) shows the waveform of the statistical data DT103, and (C) shows the waveform of the flag data DF1 when the machine learning model MR was generated. The machine learning device 1R generates the machine learning model MR using the flag data DF1 (FIG. 7(C)) created by an engineer.

[0054] The extraction processing unit 122 of the anomaly detection device 100 uses this machine learning model MR to generate extraction data DT102 (FIG. 7(A)) including two or more measurement data DD related to anomalies based on input data DT101 including a plurality of measurement data DD in the object to be detected. In this example, the extraction data DT102 includes five measurement data DD. These five measurement data DD are different from the five measurement data DD (FIG. 5(A)) in the present embodiment. That is, since the machine learning model MR generated by the machine learning device 1R according to the reference example is different from the machine learning model M generated by the machine learning device 1 according to the present embodiment, the extraction processing unit 122 can extract five measurement data DD different from those in the case of the present embodiment (FIG. 5(A)).

[0055] Then, the statistical processing unit 123 of the abnormality detection device 100 performs statistical processing based on two or more measurement data DD included in the extraction data DT102 to generate statistical data DT103 (Fig. 7(B)). In this example, outside the period of the abnormal period P set by the engineer (part W1), the Mahalanobis distance is large, and within the period of the abnormal period P, there is no behavior such that the Mahalanobis distance becomes large. Therefore, unlike the case of this embodiment (Fig. 5(B)), it is not possible to determine whether an abnormality has occurred in the object to be detected by providing a threshold value TH and based on the Mahalanobis distance. That is, in this example, when the extraction processing unit 122 extracts two or more measurement data DD related to the abnormality among the plurality of measurement data DD, appropriate measurement data DD could not be extracted, so the Mahalanobis distance does not become large within the period of the abnormal period P. Therefore, the abnormality detection device 100 cannot determine that an abnormality has occurred in the object to be detected based on the Mahalanobis distance.

[0056] As described above, in the machine learning device 1R, since the machine learning model MR is generated using the flag data DF1, the accuracy of the machine learning model MR becomes low. That is, when an abnormality occurs in the object to be detected, it is considered that the abnormality does not suddenly start at a certain timing (timing t11 in Fig. 7), but rather starts to occur gradually before that timing. Therefore, when the machine learning model MR is generated using the flag data DF1 indicating that the abnormality suddenly starts at the timing t11, the accuracy of the machine learning model MR becomes low. As a result, in the abnormality detection device 100, when detecting an abnormality of the object to be detected using this machine learning model MR, the detection accuracy becomes low.

[0057] On the other hand, in the machine learning device 1 according to the present embodiment, the flag data DF2 is generated by performing smoothing processing on the flag data DF1, and the machine learning model M is generated using this flag data DF2. Therefore, the accuracy of the machine learning model M can be improved. That is, when an abnormality occurs in the detection target object, the abnormality gradually starts to occur before a certain timing (timing t11 in FIG. 5). Therefore, by using the flag data DF2 in which the flag value gradually changes from "0" to "1" from before this timing t11, the accuracy of the machine learning model M can be improved. As a result, in the abnormality detection device 100, the detection accuracy can be improved by detecting the abnormality of the detection target object using this machine learning model M.

[0058] (Regarding the smoothing process) In this way, in the machine learning device 1, the flag data DF2 is generated by performing smoothing processing on the flag data DF1, and the machine learning model M is generated using this flag data DF2. Hereinafter, this smoothing process will be described in detail.

[0059] First, the timing during which the flag data DF2 changes from "0" to "1" and the timing during which it changes from "1" to "0" will be described.

[0060] In the example of FIG. 5(C), the period during which the flag data DF2 changes from "0" to "1" is before the timing t11, and the period during which the flag data DF2 changes from "1" to "0" is after the timing t12. In this case, as shown in FIG. 5(B), the period during which the Mahalanobis distance exceeds the threshold TH is within the abnormal period P set by the engineer. That is, the accuracy of the machine learning model M can be increased.

[0061] On the other hand, for example, as shown in FIG. 8, when the period during which the flag data DF2 changes from "0" to "1" straddles the timing t11 and the period during which the flag data DF2 changes from "1" to "0" straddles the timing t12 (FIG. 8(C)), as shown in FIG. 8(A), the extraction processing unit 122 of the abnormality detection device 100 can extract five measurement data DD different from those in the case of the present embodiment (FIG. 5(A)). As a result, as shown in FIG. 8(B), outside the period of the abnormal period P set by the engineer (portion W2), the Mahalanobis distance increases, and within the period of the abnormal period P, there is no behavior such that the Mahalanobis distance increases. Therefore, the abnormality detection device 100 cannot determine that an abnormality has occurred in the detection target based on the Mahalanobis distance.

[0062] Also, for example, as shown in FIG. 9, when the period during which the flag data DF2 changes from "0" to "1" is after the timing t11 and the period during which the flag data DF2 changes from "1" to "0" is before the timing t12 (FIG. 9(C)), as shown in FIG. 9(A), the extraction processing unit 122 of the abnormality detection device 100 can extract five measurement data DD different from those in the case of the present embodiment (FIG. 5(A)). As a result, as shown in FIG. 9(B), outside the period of the abnormal period P set by the engineer (portion W3), the Mahalanobis distance increases, and within the period of the abnormal period P, there is no behavior such that the Mahalanobis distance increases. Therefore, the abnormality detection device 100 cannot determine that an abnormality has occurred in the detection target based on the Mahalanobis distance.

[0063] The smoothing processing unit 21 of the machine learning device 1 sets the period during which the flag data DF2 changes from "0" to "1" to be before the timing (timing t11) at which the flag data DF1 changes from "0" to "1". This timing t11 is the timing at which an engineer visually recognizes the start of an abnormality. Also, the period during which the flag data DF2 changes from "1" to "0" is set to be after the timing (timing t12) at which the flag data DF1 changes from "1" to "0". This timing t12 is the timing at which an engineer visually recognizes the end of an abnormality. As a result, the flag data DF2 has a waveform corresponding to the behavior that an abnormality gradually starts to occur in the detection object before the abnormal period P visually recognized by the engineer, and the abnormality gradually ends after the abnormal period P visually recognized by the engineer. By performing machine learning processing using such flag data DF2, the machine learning device 1 can improve the accuracy of the machine learning model M.

[0064] Next, the time widths of the periods during which the flag data DF2 changes from "0" to "1" and from "1" to "0" will be described. In this example, the case where the abnormality is a vibration phenomenon will be described as an example. For example, when the detection object is a vehicle, this vibration phenomenon can be, for example, abnormal vibration of a valve mechanism used in the vehicle's engine.

[0065] In the example of FIG. 5(A), a vibration phenomenon is observed in one of the five measured data DD extracted. This measured data DD vibrates with a vibration period T during the abnormal period P. In this example, the time widths of the periods during which the flag data DF2 changes from "0" to "1" and from "1" to "0" are approximately the same as this vibration period T. In this case, as shown in FIG. 5(B), the period during which the Mahalanobis distance exceeds the threshold TH is within the period of the abnormal period P set by the engineer. That is, the accuracy of the machine learning model M can be improved.

[0066] On the one hand, for example, as shown in FIG. 10, when the time width of the period during which the flag data DF2 changes from "0" to "1" is shorter than the oscillation period T and the time width of the period during which the flag data DF2 changes from "1" to "0" is shorter than the oscillation period T (FIG. 10(C)), as shown in FIG. 10(A), the extraction processing unit 122 of the abnormality detection device 100 can extract five measurement data DD different from those in the case of the present embodiment (FIG. 5(A)). As a result, as shown in FIG. 10(B), outside the period of the abnormal period P set by the engineer (portion W4), the Mahalanobis distance increases, and within the period of the abnormal period P, there is no behavior such that the Mahalanobis distance increases. Therefore, the abnormality detection device 100 cannot determine that an abnormality has occurred in the object to be detected based on the Mahalanobis distance.

[0067] Also, for example, as shown in FIG. 11, when the period during which the flag data DF2 changes from "0" to "1" is longer than the oscillation period T and the period during which the flag data DF2 changes from "1" to "0" is longer than the oscillation period T (FIG. 11(C)), as shown in FIG. 11(A), the extraction processing unit 122 of the abnormality detection device 100 can extract five measurement data DD different from those in the case of the present embodiment (FIG. 5(A)). As a result, as shown in FIG. 11(B), outside the period of the abnormal period P set by the engineer (portion W5), the Mahalanobis distance increases, and within the period of the abnormal period P, there is no behavior such that the Mahalanobis distance increases. Therefore, the abnormality detection device 100 cannot determine that an abnormality has occurred in the object to be detected based on the Mahalanobis distance.

[0068] For example, when the abnormality is a vibration phenomenon, the smoothing processing unit 21 of the machine learning device 1 makes the time width of the period during which the flag data DF2 changes from "0" to "1" approximately the same as the vibration period, and makes the time length of the period during which the flag data DF2 changes from "1" to "0" approximately the same as the vibration period. As a result, the flag data DF2 has a waveform corresponding to the behavior in which the vibration phenomenon gradually starts and gradually ends in the detection target object. By performing machine learning processing using such flag data DF2, the accuracy of the machine learning model M can be increased.

[0069] In this way, in the machine learning device 1, a smoothing processing unit 21 capable of generating second flag data (flag data DF2) by performing smoothing processing on first flag data (flag data DF1), which is time-series data indicating whether an abnormality has occurred in the detection target object, according to a plurality of measurement data DD, which is time-series data about a plurality of parameters in the detection target object, and a machine learning processing unit 22 capable of performing machine learning processing by performing regression analysis so that the predicted value according to two or more measurement data output from the machine learning model M becomes the same as the second flag data (flag data DF2) when the plurality of measurement data DD are input to the machine learning model M that inputs the plurality of measurement data DD and outputs two or more measurement data that are part of the plurality of measurement data DD are provided. As a result, the flag data DF2 has a waveform corresponding to the behavior in which an abnormality gradually begins to occur and gradually ends in the detection target object. By performing machine learning processing using such flag data DF2, the accuracy of the machine learning model M can be increased. In the abnormality detection device 100, by using this machine learning model M to detect an abnormality in the detection target object, the accuracy of detecting an abnormality in the detection target object can be improved.

[0070] In the machine learning device 1, the first flag data (flag data DF1) includes a first value (“0”) indicating that the detection target is normal and a second value (“1”) indicating that the detection target is abnormal. The smoothing processing unit 21 is capable of performing smoothing processing so that the value of the second flag data (flag data DF2) continuously changes between the first value and the second value during the period when the first flag data (flag data DF1) is the first value. As a result, the flag data DF2 has a waveform corresponding to the behavior in which an abnormality gradually starts to occur in the detection target before the abnormal period P visually recognized by the engineer and the abnormality gradually ends after the abnormal period P visually recognized by the engineer. The machine learning device 1 can improve the accuracy of the machine learning model M by performing machine learning processing using such flag data DF2. In the abnormality detection device 100, the accuracy of detecting an abnormality in the detection target can be improved by detecting the abnormality in the detection target using this machine learning model M.

[0071] In the machine learning device 1, when the abnormality in the detection target is a vibration phenomenon, the time width of the period during which the value of the second flag data (flag data DF2) continuously changes between the first value and the second value is set to be a time width corresponding to the period of the vibration phenomenon. As a result, the flag data DF2 has a waveform corresponding to the behavior in which the vibration phenomenon gradually starts and gradually ends in the detection target. The machine learning device 1 can improve the accuracy of the machine learning model M by performing machine learning processing using such flag data DF2. In the abnormality detection device 100, the accuracy of detecting an abnormality in the detection target can be improved by detecting the abnormality in the detection target using this machine learning model M.

[0072] [Effect] As described above, in the present embodiment, for the first flag data which is time series data indicating whether an abnormality has occurred in the detection target object according to a plurality of measurement data which are time series data of a plurality of parameters in the detection target object, a smoothing process is performed to generate the second flag data. A smoothing process unit capable of this, and a machine learning model into which a plurality of measurement data are input and two or more measurement data which are a part of the plurality of measurement data are output, the plurality of measurement data are input, and a machine learning process is performed by performing a regression analysis so that a predicted value according to the two or more measurement data output from the machine learning model becomes the same as the second flag data. A machine learning process unit capable of this is provided. Thereby, the accuracy of detecting an abnormality in the detection target object can be improved.

[0073] In the present embodiment, the first flag data includes a first value indicating that the detection target object is normal and a second value indicating that the detection target object is abnormal. The smoothing process unit can perform a smoothing process so that the value of the second flag data continuously changes between the first value and the second value during a period when the first flag data is the first value. Thereby, the accuracy of detecting an abnormality in the detection target object can be improved.

[0074] In the present embodiment, when the abnormality in the detection target object is a vibration phenomenon, the time width of the period during which the value of the second flag data continuously changes between the first value and the second value is set to be a time width corresponding to the period of the vibration phenomenon. Thereby, the accuracy of detecting an abnormality in the detection target object can be improved.

[0075] [Modification Example] In the above-described embodiment, the machine learning device 1 and the abnormality detection device 100 are configured as separate devices, respectively. However, the present invention is not limited to this, and instead, the machine learning device 1 and the abnormality detection device 100 may be configured as one device.

[0076] As described above, the present technology has been described with reference to the embodiment and several modification examples. However, the present technology is not limited to these embodiments and the like, and various modifications are possible.

[0077] For example, in the above embodiment, the present technology is applied to a vehicle, but it is not limited thereto, and the present technology can be applied to various devices and various equipment.

[0078] Note that the effects described in this specification are merely examples and are not limiting, and there may be other effects.

[0079] Furthermore, the present disclosure can take the following aspects.

[0080] (1) A smoothing circuit capable of generating second flag data by performing a smoothing process on first flag data, which is time-series data indicating whether an abnormality has occurred in the detection target object, according to a plurality of measurement data that is time-series data of a plurality of parameters in the detection target object, and A machine learning circuit capable of performing machine learning processing by inputting the plurality of measurement data into a machine learning model that inputs the plurality of measurement data and outputs two or more pieces of measurement data that are part of the plurality of measurement data, and performing regression analysis so that a predicted value corresponding to the two or more pieces of measurement data output from the machine learning model becomes the same as the second flag data A machine learning device comprising the above. (2) The first flag data includes a first value indicating that the detection target object is normal and a second value indicating that the detection target object is abnormal, The smoothing circuit can perform the smoothing process so that the value of the second flag data continuously changes between the first value and the second value during a period when the first flag data is the first value The machine learning device according to (1) above. (3) The first flag data includes a first value indicating that the detection target object is normal and a second value indicating that the detection target object is abnormal, When the abnormality in the object to be detected is a vibration phenomenon, the time width of the period during which the value of the second flag data continuously changes between the first value and the second value is a time width corresponding to the period of the vibration phenomenon The machine learning device according to (1) or (2) above (4) Generating second flag data by performing smoothing processing on first flag data, which is time-series data indicating whether an abnormality has occurred in the object to be detected, according to a plurality of measurement data, which is time-series data for a plurality of parameters in the object to be detected Inputting the plurality of measurement data into a machine learning model that receives the plurality of measurement data and outputs two or more measurement data that are part of the plurality of measurement data Performing machine learning processing by performing regression analysis so that a predicted value corresponding to the two or more measurement data output from the machine learning model becomes the same as the second flag data A machine learning method including the above

Explanation of symbols

[0081] 1... Machine learning device, 10... Storage, 20... Processing unit, 21... Smoothing processing unit, 22... Machine learning processing unit, 23... Statistical processing unit, 24... Display processing unit, 25... Threshold setting unit, 30... User interface, 100... Abnormality detection device, 110... Storage, 120... Processing unit, 122... Extraction processing unit, 123... Statistical processing unit, 124... Detection processing unit, 130... User interface, DD... Measurement data, DF1, DF2... Flag data, DS1, DS101... Data set, DT1... Input data, DT2, DT102... Extracted data, DT3, DT103... Statistical data, M... Machine learning model, P... Abnormal period, T... Vibration period, TH... Threshold

Claims

1. A smoothing circuit capable of generating second flag data by performing a smoothing process on first flag data, which is time-series data indicating whether an abnormality has occurred in the detection target object according to a plurality of measurement data that are time-series data of a plurality of parameters in the detection target object, and a machine learning circuit capable of performing machine learning processing by inputting the plurality of measurement data into a machine learning model in which the plurality of measurement data are input and two or more measurement data that are part of the plurality of measurement data are output, and performing regression analysis so that a predicted value according to the two or more measurement data output from the machine learning model becomes the same as the second flag data A machine learning device comprising the above.

2. The first flag data includes a first value indicating that the detection target object is normal and a second value indicating that the detection target object is abnormal, and the smoothing circuit is capable of performing the smoothing process so that the value of the second flag data continuously changes between the first value and the second value during a period in which the first flag data is the first value The machine learning device according to claim 1.

3. The first flag data includes a first value indicating that the detection target object is normal and a second value indicating that the detection target object is abnormal, and when the abnormality in the detection target object is a vibration phenomenon, the time width of the period in which the value of the second flag data continuously changes between the first value and the second value is a time width corresponding to the period of the vibration phenomenon The machine learning device according to claim 1.

4. Generating second flag data by performing a smoothing process on first flag data, which is time-series data indicating a period in which an abnormality has occurred in the detection target object according to a plurality of measurement data that are time-series data of a plurality of parameters in the detection target object; Inputting the plurality of measurement data into a machine learning model in which the plurality of measurement data are input and two or more measurement data that are part of the plurality of measurement data are output; and Performing machine learning processing by performing regression analysis so that a predicted value according to the two or more measurement data output from the machine learning model becomes the same as the second flag data A machine learning method including the above.

Images