Sensor module and information processing device

Abstract

A sensor module according to one embodiment of the present invention is provided with: a sensor capable of detecting a physical quantity corresponding to an operation state of a device; and a processing circuit capable of performing, on the basis of a detection result from the sensor and by using a first learning model, monitoring processing for monitoring the operation state of the device. The processing circuit can perform relearning processing for the first learning model on the basis of the detection result from the sensor, perform, in the relearning processing, verification processing for verifying data to be input to the first learning model, and stop the relearning processing on the basis of the result of the verification processing.

Description

Sensor Module and Information Processing Device

[0001] The present invention relates to a sensor module and an information processing device for monitoring the operating state of a device.

[0002] In industrial machines, it may be desirable not only to detect that a failure has actually occurred, but also to detect in advance that a failure is likely to occur before it occurs. For example, Patent Document 1 discloses a failure prediction system that uses machine learning technology to predict failures in industrial machines.

[0003] Japanese Patent Application Laid-Open No. 2022-125288

[0004] Thus, in a system for predicting failures in industrial machines, it is desirable that the burden on the operator is small, and further reduction of the burden on the operator is expected.

[0005] It is desirable to provide a sensor module and an information processing device that can reduce the burden on the operator.

[0006] A sensor module according to an embodiment of the present invention includes a sensor and a processing circuit. The sensor is capable of detecting a physical quantity corresponding to the operating state of the device. The processing circuit is capable of performing a monitoring process for monitoring the operating state of the device using a first learning model based on the detection result of the sensor. The processing circuit is capable of performing a relearning process for the first learning model based on the detection result of the sensor, and in the relearning process, is capable of performing a verification process for verifying the data input to the first learning model, and is capable of aborting the relearning process based on the result of the verification process.

[0007] An information processing device according to one embodiment of the present invention includes a processing circuit. The processing circuit is capable of performing monitoring processing to monitor the operating state of a device using a first learning model, based on the detection results of a sensor capable of detecting a physical quantity corresponding to the operating state of the device. The processing circuit is capable of performing a retraining process of the first learning model based on the detection results of the sensor, and in the retraining process, it is capable of performing a verification process to verify the data to be input to the first learning model, and based on the results of the verification process, it is possible to cancel the retraining process.

[0008] According to one embodiment of the present invention, the sensor module and information processing device can reduce the burden on the operator.

[0009] Figure 1 is a block diagram showing an example configuration of a monitoring system according to one embodiment of the present invention. Figure 2 is a block diagram showing an example configuration of the sensor module shown in Figure 1. Figure 3 is an explanatory diagram showing an example operation of the data generation unit shown in Figure 2. Figure 4 is an explanatory diagram showing an example operation of the monitoring processing unit shown in Figure 2. Figure 5A is a flowchart showing an example operation of the sensor module shown in Figure 2. Figure 5B is another flowchart showing an example operation of the sensor module shown in Figure 2. Figure 6 is a flowchart showing an example operation of the initial startup verification unit shown in Figure 2. Figure 7 is an explanatory diagram showing an example operation of the initial startup verification unit shown in Figure 2. Figure 8 is a flowchart showing an example operation of the environmental change detection unit shown in Figure 2. Figure 9 is an explanatory diagram showing an example operation of the environmental change detection unit shown in Figure 2. Figure 10 is another explanatory diagram showing an example operation of the environmental change detection unit shown in Figure 2. Figure 11 is a flowchart showing an example operation of the data verification unit shown in Figure 2. Figure 12 is an explanatory diagram showing an example operation of the data verification unit shown in Figure 2. Figure 13 is a block diagram showing an example configuration of a monitoring system according to a modified example. Figure 14 is a flowchart illustrating one example of operation of the environmental change detection unit shown in Figure 13. Figure 15 is an explanatory diagram illustrating one example of operation of the environmental change detection unit shown in Figure 14. Figure 16 is another explanatory diagram illustrating one example of operation of the environmental change detection unit shown in Figure 14. Figure 17 is another explanatory diagram illustrating one example of operation of the environmental change detection unit shown in Figure 14.

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

[0011] <Embodiment> [Configuration Example] Figure 1 shows an example configuration of a monitoring system 1 equipped with a sensor module according to an embodiment of the present invention. The monitoring system 1 comprises a plurality of mechanical devices 100 (in this example, three mechanical devices 100A to 100C), a plurality of sensor modules 10 (in this example, three sensor modules 10A to 10C), an access point 91, and a personal computer 92.

[0012] Each of the multiple mechanical devices 100 is an industrial machine with movable mechanical parts, installed, for example, in a factory engaged in manufacturing. Such industrial machines may be small machines such as motors and pumps, or large machines such as industrial robots. In this example, the multiple mechanical devices 100 are of the same type and of the same model number.

[0013] Each of the multiple sensor modules 10 is configured to detect vibrations of the mechanical device 100. The multiple sensor modules 10 are each attached to one of the multiple mechanical devices 100. Based on the vibrations of the mechanical device 100, the sensor modules 10 monitor the operating state of the mechanical device 100 using a learning model M. The sensor modules 10 then transmit the results of monitoring the operating state of the mechanical device 100 to the personal computer 92 via the access point 91 by performing wireless communication W with the access point 91.

[0014] The access point 91 is, for example, a wireless LAN (Local Area Network) access point and is configured to communicate wirelessly W with each of the multiple sensor modules 10. The access point 91 also has the function of communicating via wired connection with the personal computer 92.

[0015] The personal computer 92 is a computer operated by the operator of the monitoring system 1. The personal computer 92 receives monitoring results of the operating status of the mechanical devices 100 transmitted from multiple sensor modules 10. The personal computer 92 then displays the monitoring results of the operating status of the multiple mechanical devices 100 on its display. Based on the information displayed on the personal computer 92's display, the operator can understand the operating status of the multiple mechanical devices 100.

[0016] In this monitoring system 1, first, an initial learning model M is set in the sensor modules 10A to 10C. This initial learning model M is generated by a personal computer (not shown) performing a learning process based on data showing vibrations of mechanical devices of the same model number as mechanical devices 100A to 100C. Then, each of the sensor modules 10A to 10C modifies this learning model M by performing a retraining process as needed during the initial startup. That is, because there are individual differences in mechanical devices, this learning model M may not be usable for monitoring a mechanical device 100 that is different from the mechanical device 100 used in the learning process of this learning model M. Therefore, each of the sensor modules 10A to 10C performs a retraining process as needed during the initial startup to absorb the individual differences of the mechanical device 100.

[0017] Furthermore, each of the sensor modules 10A to 10C modifies the learned model M by performing a retraining process when environmental changes occur. That is, for example, if the temperature or humidity changes, the characteristics of the mechanical device 100 may change. Also, for example, if another mechanical device is installed near the mechanical device 100, the vibrations of that mechanical device may be slightly superimposed on the vibrations of the mechanical device 100. Therefore, each of the sensor modules 10A to 10C is designed to perform a retraining process when environmental changes occur in order to minimize the impact of those environmental changes on the monitoring results of the operating state of the mechanical device 100.

[0018] (Sensor Module 10) Figure 2 shows an example configuration of the sensor module 10. The sensor module 10 includes an acceleration sensor 11, a processing circuit 20, and a wireless communication circuit 12.

[0019] The acceleration sensor 11 is configured to detect acceleration. The sensor module 10 is attached to the mechanical device 100. Therefore, the acceleration detected by the acceleration sensor 11 may change in accordance with the vibration of the mechanical device 100. The acceleration sensor 11 is configured to supply acceleration data indicating the detected acceleration to the processing circuit 20.

[0020] The processing circuit 20 is configured to monitor the operating state of the mechanical device 100 based on acceleration data supplied from the acceleration sensor 11. The processing circuit 20 is configured to include, for example, a microcontroller. By executing software, the processing circuit 20 can operate as a data generation unit 21, a monitoring processing unit 22, a learning processing unit 23, a data verification unit 24, a relearning determination unit 25, and a control processing unit 29.

[0021] The data generation unit 21 is configured to generate standardized spectral data D3 for acceleration by performing a Fourier transform based on the time-series data of acceleration data supplied from the acceleration sensor 11.

[0022] Figure 3 shows an example of the operation of the data generation unit 21. In Figure 3, the vertical axis of the three waveforms is an arbitrary unit. The data generation unit 21 generates spectral data D2 by performing a Fourier transform on time-series data (sensor data D1) for a period of approximately 2 seconds, which is included in the time-series data of acceleration data (Figure 3(A), (B)). Then, the data generation unit 21 generates spectral data D3 by performing a standardization process based on this spectral data D2 (Figure 3(C)). Specifically, the data generation unit 21 calculates the average value VAV of the amplitude values ​​at each frequency in spectral data D2, and generates spectral data D3 by subtracting this average value VAV from the amplitude values ​​in spectral data D2. In this example, the data generation unit 21 intermittently generates spectral data D3 at a rate of approximately once per minute and supplies the generated spectral data D3 to the monitoring processing unit 22.

[0023] Furthermore, when the learning processing unit 23 performs relearning processing, the data generation unit 21 continuously calculates the average value VAV and generates spectral data D3 based on the sensor data D1 in approximately 2-second intervals during the relearning processing period (relearning period). The data generation unit 21 also calculates the peak-to-peak value (PtoP value) VPP of the acceleration value in the sensor data D1. The PtoP value VPP is the value obtained by subtracting the minimum value from the maximum value of the acceleration value. In this example, the data generation unit 21 continuously supplies spectral data D3 to the learning processing unit 23 at a rate of approximately once every 2 seconds, and also supplies the average value VAV and PtoP value VPP to the data verification unit 24.

[0024] The monitoring processing unit 22 is configured to generate an abnormal value VRES corresponding to the operating state of the mechanical device 100, using a learning model M based on the spectral data D3. In this example, the spectral data D3 is supplied intermittently at a rate of approximately once every minute. The monitoring processing unit 22 generates the abnormal value VRES based on this spectral data D3.

[0025] Figure 4 shows an example of the operation of the monitoring processing unit 22. The learning model M is a machine learning model in which spectral data D3 is input and spectral data D4 is output.

[0026] As described above, the initial learning model M is generated by a personal computer (not shown) performing a learning process based on data showing vibrations of mechanical devices of the same model number as mechanical devices 100A to 100C. This initial learning model M is then supplied to the sensor module 10 via wireless communication W from the personal computer 92 and set in the monitoring processing unit 22. This personal computer for learning processing inputs spectral data similar to spectral data D3 into the learning model M and performs learning processing so that the learning model M outputs spectral data identical to the input data. In this learning process, spectral data showing vibrations of mechanical devices operating normally is used. Therefore, the spectral data output from the learning model M is similar to the spectral data related to mechanical devices operating normally.

[0027] The monitoring processing unit 22 operates using the learning model M generated in this manner. Therefore, for example, if the machine 100 is operating normally, when spectral data D3 related to the machine 100 is input to the learning model M, spectral data D4 similar to spectral data D3 is output from the learning model M. On the other hand, if the machine 100 is not operating normally, when spectral data D3 related to the machine 100 is input to the learning model M, spectral data D4 that is not similar to spectral data D3 is output from the learning model M. In other words, since the spectral data output from the learning model M is similar to the spectral data related to a machine that is operating normally, spectral data D4 is data that is not similar to spectral data D3.

[0028] The monitoring processing unit 22 generates an anomaly value VRES by performing a subtraction operation based on the spectral data D3 input to the learning model M and the spectral data D4 output from the learning model M. Specifically, it calculates the absolute value of the difference in amplitude at the same frequency for spectral data D3 and spectral data D4, and generates the anomaly value VRES by summing these absolute values ​​over all frequencies. The value indicated by the anomaly value VRES is small when spectral data D3 and spectral data D4 are similar to each other, and large when spectral data D3 and spectral data D4 are different to each other.

[0029] Therefore, when the machine 100 is operating normally, spectral data D3 and spectral data D4 are similar to each other, so the abnormal value VRES is small. Conversely, when the machine 100 is not operating normally, spectral data D3 and spectral data D4 are not similar to each other, so the abnormal value VRES is large.

[0030] In this way, the monitoring processing unit 22 generates an abnormal value VRES corresponding to the operating state of the machine device 100 based on the spectral data D3. The monitoring processing unit 22 is configured to process and generate the abnormal value VRES each time spectral data D3 is supplied.

[0031] In this example, the monitoring processing unit 22 generates the abnormal value VRES by using a learning model M that takes spectral data D3 as input and outputs spectral data D4, and performing a subtraction operation based on spectral data D3 and spectral data D4. However, it is not limited to this. Alternatively, for example, the monitoring processing unit 22 may generate the abnormal value VRES by using a learning model that takes spectral data D3 as input and outputs the abnormal value VRES.

[0032] The learning processing unit 23 (Figure 2) is configured to perform retraining processing based on spectral data D3 during the retraining period. Spectral data D3 is continuously supplied during the retraining period at a rate of approximately once every two seconds in this example. The learning processing unit 23 inputs spectral data D3 to the learning model M and performs retraining processing so that the learning model M outputs spectral data identical to the input data. The learning processing unit 23 modifies the learning model M by performing retraining processing based on this spectral data D3 during the retraining period.

[0033] The data verification unit 24 is configured to perform verification processing during the relearning period to verify whether the spectral data D3 supplied to the learning processing unit 23 is appropriate data for the relearning process, based on the average value VAV and PtoP value VPP. The average value VAV and PtoP value VPP are continuously supplied during the relearning period, at a rate of approximately once every two seconds in this example. The data verification unit 24 performs verification processing during this relearning period based on this average value VAV and PtoP value VPP.

[0034] The relearning determination unit 25 is configured to determine whether the learning processing unit 23 should perform relearning based on the abnormal value VRES. In this example, the abnormal value VRES is supplied intermittently at a rate of approximately once every minute. The relearning determination unit 25 performs processing based on this abnormal value VRES. The relearning determination unit 25 includes an initial startup verification unit 26 and environmental change detection units 27S and 27L.

[0035] The initial startup verification unit 26 is configured to determine whether the learning processing unit 23 should perform retraining based on the abnormal value VRES when the sensor module 10 is started for the first time. That is, as described above, the initial learning model M is generated by, for example, a personal computer (not shown) performing learning processing based on data showing vibrations of mechanical devices of the same model number as mechanical devices 100A to 100C. Since there are individual differences in mechanical devices, this initial learning model M may not be usable for monitoring a mechanical device 100 that is different from the mechanical device used in the learning process of this learning model M. Therefore, the initial startup verification unit 26 is configured to determine whether the learning processing unit 23 should perform retraining based on the abnormal value VRES when the system is first started.

[0036] Each of the environmental change detection units 27S and 27L is configured to detect environmental changes based on abnormal values ​​VRES, and to determine whether the learning processing unit 23 should perform retraining processing when an environmental change is detected. In this example, the environmental change detection unit 27S can detect that the environment has changed in a short period of time based on abnormal values ​​VRES supplied at a rate of approximately once every minute. As a result, the environmental change detection unit 27S can detect, for example, that another machine has been installed near the machine 100. In this example, the environmental change detection unit 27L can detect that the environment has changed slowly over a long period of time based on abnormal values ​​VRES supplied at a rate of approximately once every hour. As a result, the environmental change detection unit 27L can detect, for example, changes in temperature or humidity.

[0037] The control processing unit 29 is configured to control the operation of the data generation unit 21, the monitoring processing unit 22, the learning processing unit 23, the data verification unit 24, and the relearning determination unit 25 based on instructions supplied from the personal computer 92 via the wireless communication circuit 12. The control processing unit 29 is also configured to control the operation of the wireless communication circuit 12 so that it transmits the abnormal value VRES generated by the monitoring processing unit 22 to the personal computer 92.

[0038] The wireless communication circuit 12 is configured to perform wireless communication W with the access point 91.

[0039] With this configuration, the monitoring system 1 allows the sensor module 10 to detect if the mechanical device 100 is malfunctioning or is likely to malfunction by detecting vibrations of the mechanical device 100. In other words, if the mechanical device 100 is malfunctioning or is likely to malfunction, the vibrations of the mechanical device 100 may change. Therefore, the sensor module 10 monitors the operating state of the mechanical device 100 by detecting vibrations of the mechanical device 100. The sensor module 10 then transmits the monitoring results of the operating state of the mechanical device 100 to the personal computer 92. The personal computer 92 displays the monitoring results of the operating states of multiple mechanical devices 100 on its display. Based on the information displayed on the personal computer 92's display, the operator can understand the operating state of multiple mechanical devices 100. This allows the operator to understand if the mechanical device 100 is malfunctioning or is likely to malfunction.

[0040] Furthermore, the sensor module 10 can modify the learned model M by performing a retraining process as needed during its initial startup. The sensor module 10 can also modify the learned model M by performing a retraining process when environmental changes occur. When performing this retraining process, the sensor module 10 verifies whether the data used for the retraining process is appropriate for the process. This allows the sensor module 10 to perform the retraining process appropriately and maintain the learned model M as a model suitable for monitoring the mechanical device 100.

[0041] Here, the sensor module 10 corresponds to a specific example of the "sensor module" in one embodiment of the present disclosure. The acceleration sensor 11 corresponds to a specific example of the "sensor" in one embodiment of the present disclosure. The learning model M corresponds to a specific example of the "first learning model" in one embodiment of the present disclosure. The processing circuit 20 corresponds to a specific example of the "processing circuit" in one embodiment of the present disclosure. The sensor data D1 corresponds to a specific example of the "sensor data" in one embodiment of the present disclosure. The spectral data D2 and D3 correspond to specific examples of the "spectral data" in one embodiment of the present disclosure. The average value VAV corresponds to a specific example of the "average value" in one embodiment of the present disclosure. The PtoP value VPP corresponds to a specific example of the "peak-to-peak value" in one embodiment of the present disclosure. The abnormal value VRES corresponds to a specific example of the "monitoring result value" in one embodiment of the present disclosure.

[0042] [Operation and Function] Next, the operation and function of the sensor module 10 of this embodiment will be described.

[0043] (Overall Operation Summary) First, referring to FIG. 2, the operation of the sensor module 10 will be described. The acceleration sensor 11 detects acceleration and supplies acceleration data indicating the detected acceleration to the processing circuit 20. The data generation unit 21 of the processing circuit 20 generates spectrum data D3, average value VAV, and PtoP value based on the acceleration data supplied from the acceleration sensor 11. The monitoring processing unit 22 generates an abnormal value VRES corresponding to the operating state of the mechanical device 100 using the learning model M based on the spectrum data D3. The learning processing unit 23 performs a re-learning process based on the spectrum data D3 during the re-learning period. The data verification unit 24 performs a verification process to verify whether the spectrum data D3 supplied to the learning processing unit 23 is appropriate for the re-learning process based on the average value VAV and the PtoP value VPP during the re-learning period. The re-learning determination unit 25 determines whether the learning processing unit 23 should perform the re-learning process based on the abnormal value VRES. The control processing unit 29 controls the operations of the data generation unit 21, the monitoring processing unit 22, the learning processing unit 23, the data verification unit 24, and the re-learning determination unit 25 based on an instruction supplied from the personal computer 92 via the wireless communication circuit 12. In addition, the control processing unit 29 controls the operation of the wireless communication circuit 12 so as to transmit the abnormal value VRES generated by the monitoring processing unit 22 to the personal computer 92. The wireless communication circuit 12 performs wireless communication W with the access point 91.

[0044] (Detailed Operation) FIGS. 5A and 5B illustrate an operation example of the sensor module 10. When the sensor module 10 is activated, the sensor module 10 performs the following processes.

[0045] First, the control processing unit 29 checks whether the sensor module 10 has been started for the first time (step S101). The initial learning model M is supplied to the sensor module 10 from, for example, a personal computer 92 via wireless communication W and set in the monitoring processing unit 22. The control processing unit 29 checks whether the sensor module 10 has been started for the first time after the initial learning model M has been set in the monitoring processing unit 22. If the sensor module 10 has not been started for the first time (in step S101, “N”), the process proceeds to step S110.

[0046] In step S101, if the sensor module 10 has been started for the first time (in step S101, “Y”), the first startup verification unit 26 of the relearning determination unit 25 determines whether the learning processing unit 23 should perform a relearning process (step S102). That is, as described above, the initial learning model M is generated by, for example, a personal computer (not shown) performing a learning process based on data indicating the vibration of a machine device of the same model number as the machine devices 100A to 100C. Since there are individual differences in machine devices, this initial learning model M may not be usable for monitoring a machine device 100 different from the machine device used in the learning process of this learning model M. Therefore, in the sensor module 10, the monitoring processing unit 22 uses the initial learning model M based on the spectrum data D3 to generate an abnormal value VRES corresponding to the operating state of the machine device 100. Then, the first startup verification unit 26 determines whether the learning processing unit 23 should perform a relearning process based on this abnormal value VRES at the first startup. This process will be described in detail later. If it is determined not to perform the relearning process (in step S103, “N”), the process proceeds to step S110.

[0047] In step S103, if it is determined that the relearning process should be performed (in step S103, “Y”), the data verification unit 24 checks whether the spectrum data D3 supplied to the learning processing unit 23 is appropriate data for the relearning process (step S104). This process will be described in detail later.

[0048] In step S104, if the spectral data D3 is suitable for retraining (Y in step S104), the learning processing unit 23 performs retraining based on this spectral data D3 using the learning model M (step S105). Specifically, the learning processing unit 23 inputs this spectral data D3 to the learning model M and performs retraining so that the learning model M outputs spectral data identical to the input data.

[0049] Next, the learning processing unit 23 checks whether the relearning process in step S105 has been performed a predetermined number of times (step S106). If the relearning process has not yet been performed a predetermined number of times ("N" in step S106), the process returns to step S104, and the learning processing unit 23 performs the relearning process based on the next spectral data D3. The learning processing unit 23 repeats the processes in steps S104 to S106 while changing the spectral data D3 until the relearning process has been performed a predetermined number of times.

[0050] In step S106, if the retraining process is performed a predetermined number of times ("Y" in step S106), the learning processing unit 23 updates the learning model M of the monitoring processing unit 22 by setting the learning model M corrected by the retraining process to the monitoring processing unit 22 (step S107). Then, the process proceeds to step S110.

[0051] In step S104, if the spectral data D3 is not suitable for the retraining process ("N" in step S104), the learning processing unit 23 terminates the retraining process (step S108).

[0052] Then, based on instructions from the control processing unit 29, the wireless communication circuit 12 sends a notification to the personal computer 92 to cancel the relearning process (step S109). This process then ends.

[0053] If it is determined in step S103 that no retraining process is to be performed ("N" in step S103), or if the learning model M is updated in step S107, the sensor module 10 starts monitoring (step S110). In this monitoring process, the monitoring processing unit 22 generates an abnormal value VRES corresponding to the operating state of the mechanical device 100 using the learning model M based on the spectral data D3. The control processing unit 29 then controls the operation of the wireless communication circuit 12 to transmit this abnormal value VRES to the personal computer 92. The personal computer 92 displays the monitoring results of the operating states of the multiple mechanical devices 100 on its display. This allows the operator to understand the operating states of the multiple mechanical devices 100 based on the information displayed on the personal computer 92's display.

[0054] Next, the environmental change detection units 27S and 27L determine whether the learning processing unit 23 should perform relearning based on the environmental changes (step S111). This process will be described in detail later. If it is determined that relearning is not to be performed ("N" in step S112), the process proceeds to step S119.

[0055] If it is determined in step S112 that retraining should be performed ("Y" in step S112), the data verification unit 24, in the same manner as in step S104, checks whether the spectral data D3 supplied to the learning processing unit 23 is appropriate data for retraining (step S113).

[0056] In step S113, if the spectral data D3 is suitable for retraining (Y in step S113), the learning processing unit 23 performs retraining using the learning model M based on this spectral data D3 (step S114).

[0057] Next, the learning processing unit 23 checks whether the relearning process in step S114 has been performed a predetermined number of times (step S115). If the relearning process has not yet been performed a predetermined number of times ("N" in step S115), the process returns to step S113, and the learning processing unit 23 performs the relearning process based on the next spectral data D3. The learning processing unit 23 repeats the processes in steps S113 to S115 while changing the spectral data D3 until the relearning process has been performed a predetermined number of times.

[0058] In step S115, if the retraining process is performed a predetermined number of times ("Y" in step S115), the learning processing unit 23 updates the learning model M of the monitoring processing unit 22 by setting the learning model M corrected by the retraining process to the monitoring processing unit 22 (step S116). Then the process proceeds to step S119.

[0059] In step S113, if the spectral data D3 is not suitable for the retraining process ("N" in step S113), the learning processing unit 23 terminates the retraining process (step S117).

[0060] Then, based on instructions from the control processing unit 29, the wireless communication circuit 12 notifies the personal computer 92 to cancel the relearning process (step S118). The process then proceeds to step S119.

[0061] If it is determined in step S112 that the relearning process will not be performed ("N" in step S112), or if the learning model M is updated in step S116, or if a notification to cancel relearning is issued in step S118, the control processing unit 29 checks whether it has received an operation termination instruction from the personal computer 99 (step S119). If it has not received an operation termination instruction ("N" in step S119), the process returns to step S111.

[0062] Then, in step S119, if an operation termination instruction is received ("Y" in step S119), this process is terminated.

[0063] (Operation of the Initial Startup Verification Unit 26) As shown in step S102 of Figure 5A, the Initial Startup Verification Unit 26 determines whether the learning processing unit 23 should perform a retraining process when the sensor module 10 is started up for the first time. This operation will be explained in detail below.

[0064] Figure 6 shows an example of the operation of the sensor module 10. The initial startup verification unit 26 is intermittently supplied with an abnormal value VRES at a rate of approximately once every minute in this example. The initial startup verification unit 26 performs the following processing each time an abnormal value VRES is supplied.

[0065] First, the initial startup verification unit 26 checks whether a predetermined amount of time has elapsed since startup (step S201). The predetermined amount of time can be, for example, several tens of minutes to several hours. If the predetermined amount of time has elapsed since startup ("Y" in step S201), this process ends.

[0066] In step S201, if a predetermined amount of time has not yet elapsed since startup ("N" in step S201), the monitoring processing unit 22 generates an abnormal value VRES corresponding to the operating state of the machine device 100 using the initial learning model M based on the spectral data D3 (step S202).

[0067] Next, the initial startup verification unit 26 subtracts a predetermined value V1 from the abnormal value VRES and calculates the absolute value of the subtraction result to obtain a difference value (step S203). The predetermined value V1 is set in advance in the sensor module 10.

[0068] Next, the initial startup verification unit 26 checks whether the difference value obtained in step S203 is smaller than the threshold value TH1 (step S204). This threshold value TH1 is set in the sensor module 10 in advance. If the difference value is smaller than the threshold value TH1 ("Y" in step S204), this process ends. In other words, in this case, the initial startup verification unit 26 determines that the learning processing unit 23 does not perform the relearning process.

[0069] In step S204, if the difference value is greater than or equal to the threshold TH1 ("N" in step S204), the initial startup verification unit 26 determines that the learning processing unit 23 should perform retraining (step S205).

[0070] This completes the process.

[0071] The predetermined value V1 and threshold TH1 are set in advance, for example, based on the data used when the initial learning model M was generated. Specifically, as described above, the initial learning model M is generated by a personal computer (not shown) performing a learning process based on data showing the vibration of mechanical devices of the same model number as mechanical devices 100A to 100C. This personal computer for the learning process inputs spectral data similar to spectral data D3 into the learning model M and performs a learning process so that the learning model M outputs spectral data identical to the input data. Multiple anomaly values ​​VRES can be obtained using the multiple spectral data used at this time and the initial learning model M generated by the learning process. The average value of these multiple anomaly values ​​VRES can be calculated and this average value can be used as the predetermined value V1. In addition, the standard deviation of these multiple anomaly values ​​VRES can be calculated and three times this standard deviation can be used as the threshold TH1.

[0072] As a result, the initial startup verification unit 26 can determine whether the learning processing unit 23 should perform retraining based on whether the abnormal value VRES obtained in step S202 is approximately the same as the abnormal value VRES obtained using the data when the initial learning model M was generated.

[0073] For example, if the abnormal value VRES obtained in step S202 is about the same as the abnormal value VRES obtained using the data when the initial learning model M was generated, the initial startup verification unit 26 determines that the learning processing unit 23 does not need to perform retraining. In other words, in this case, the abnormal value VRES obtained in step S202 is close to the value of the abnormal value VRES obtained using the data when the initial learning model M was generated, so the individual difference between the machine device 100 and the machine device evaluated when the initial learning model M was generated is small. In this case, the sensor module 10 can use the initial learning model M as is. Therefore, the initial startup verification unit 26 determines that the learning processing unit 23 does not need to perform retraining.

[0074] Furthermore, if the abnormal value VRES obtained in step S202 is not the same as the abnormal value VRES obtained using the data when the initial learning model M was generated, the initial startup verification unit 26 determines that the learning processing unit 23 should perform retraining. In other words, in this case, the abnormal value VRES obtained in step S202 is far from the value of the abnormal value VRES obtained using the data when the initial learning model M was generated, so there is a large individual difference between the machine device 100 and the machine device evaluated when the initial learning model M was generated. In this case, the sensor module 10 should modify the learning model M based on the initial learning model M. Therefore, the initial startup verification unit 26 determines that the learning processing unit 23 should perform retraining.

[0075] Figure 7 shows a specific example of the operation of the initial startup verification unit 26.

[0076] In case C1, the abnormal value VRES is 0.02. In this example, the predetermined value V1 is 0.01, so the difference value is 0.01. Since this difference value is sufficiently small, the initial startup verification unit 26 determines that the learning processing unit 23 does not need to perform relearning.

[0077] In case C2, the abnormal value VRES is 1.9. In this example, the predetermined value V1 is 0.01, so the difference value is 1.89. Since this difference value is large, the initial startup verification unit 26 determines that the learning processing unit 23 should perform retraining.

[0078] (Operation of Environmental Change Detection Units 27S and 27L) As shown in step S111 of Figure 5B, the environmental change detection units 27S and 27L determine whether the learning processing unit 23 should perform retraining based on the environmental changes. The operation of the environmental change detection unit 27S will be explained in detail below, using it as an example. The operation of the environmental change detection unit 27L is similar.

[0079] Figure 8 shows an example of operation of the environmental change detection unit 27S. In this example, the environmental change detection unit 27S is intermittently supplied with an abnormal value VRES at a rate of approximately once every minute. Each time the environmental change detection unit 27S is supplied with an abnormal value VRES, it performs the following processing.

[0080] First, the environmental change detection unit 27S stores the supplied abnormal value VRES (step S301).

[0081] Next, the environmental change detection unit 27S checks whether the number of accumulated abnormal values ​​VRES is greater than or equal to a predetermined number (for example, 100) (step S302). If the number of accumulated abnormal values ​​VRES is less than the predetermined number ("N" in step S302), this process ends.

[0082] In step S302, if the number of accumulated abnormal values ​​VRES is greater than or equal to a predetermined number ("Y" in step S302), the environmental change detection unit 27S sets two groups G1 and G2 by setting a classification point P on the time series of the latest predetermined number (for example, 100) of abnormal values ​​VRES (step S303).

[0083] Then, the environmental change detection unit 27S calculates the mean value μ and variance value A of the abnormal value VRES for each of the two groups G1 and G2, and calculates the ratio of the two mean values ​​μ and the ratio of the two variance values ​​A (step S304).

[0084] Figure 9 shows an example of the calculation process for the ratio of mean values ​​μ and variance values ​​A in two groups G1 and G2. The environmental change detection unit 27S sets a division point P on the time series of abnormal values ​​VRES. The environmental change detection unit 27S then sets groups G1 and G2 such that all abnormal values ​​VRES to the left of division point P belong to group G1, and all abnormal values ​​VRES to the right of division point P belong to group G2. The environmental change detection unit 27S calculates the mean value μ (mean value μ1) and variance value A (variance value A1) for the abnormal values ​​VRES belonging to group G1. Similarly, the environmental change detection unit 27S calculates the mean value μ (mean value μ2) and variance value A (variance value A2) for the abnormal values ​​VRES belonging to group G2. The environmental change detection unit 27S then calculates the ratio of two average values ​​μ1 and μ2 (μ1 / μ2) and the ratio of two variance values ​​A1 and A2 (A1 / A2).

[0085] Next, the environmental change detection unit 27S checks whether the ratio of the average values ​​μ is within a predetermined range (step S305). Specifically, the environmental change detection unit 27S checks whether the ratio of the average values ​​μ is greater than threshold TH2 and less than threshold TH3. These thresholds TH2 and TH3 are set in advance in the sensor module 10. For example, threshold TH2 is set to a value less than 1, and threshold TH3 is set to a value greater than 1. If the average values ​​μ1 and μ2 in group G1 are approximately the same, the ratio of the average values ​​μ is within the predetermined range. If the average values ​​μ1 and μ2 in group G1 are significantly different, the ratio of the average values ​​μ falls outside the predetermined range. If the ratio of the average values ​​μ is not within the predetermined range ("N" in step S305), the process proceeds to step S307.

[0086] In step S305, if the ratio of the mean values ​​μ is within a predetermined range ("Y" in step S305), the environmental change detection unit 27S checks whether the ratio of the variance values ​​A is within a predetermined range (step S306). Specifically, the environmental change detection unit 27S checks whether the ratio of the variance values ​​A is greater than threshold TH4 and less than threshold TH5. These thresholds TH4 and TH5 are set in advance in the sensor module 10. For example, threshold TH4 is set to a value less than 1, and threshold TH5 is set to a value greater than 1. If the variance value A1 in group G1 and the variance value A2 in group G2 are approximately the same, the ratio of the variance values ​​A is within a predetermined range. If the variance value A1 in group G1 and the variance value A2 in group G2 are significantly different, the ratio of the variance values ​​A falls outside the predetermined range. If the ratio of the variance values ​​A is not within a predetermined range ("N" in step S306), the process proceeds to step S307.

[0087] If, in step S305, the ratio of the mean values ​​μ is not within a predetermined range ("N" in step S305), and if, in step S306, the ratio of the variance values ​​A is not within a predetermined range ("N" in step S306), the environmental change detection unit 27S determines that the learning processing unit 23 should perform a retraining process (step S307). In other words, in this case, there is a large difference between the abnormal value VRES in group G1 and the abnormal value VRES in group G2, so the environmental change detection unit 27S determines that an environmental change has occurred. Therefore, the environmental change detection unit 27S determines that the learning processing unit 23 should perform a retraining process. Then, this process ends.

[0088] In step S306, if the ratio of the variance values ​​A is within a predetermined range ("Y" in step S306), the environmental change detection unit 27S checks whether all the classification points P have been set (step S308). For example, if the number of abnormal values ​​VRES is 100, the environmental change detection unit 27S can set 99 classification points P.

[0089] In step S308, if not all classification points P have been set yet ("N" in step S308), the environmental change detection unit 27S sets two groups G1 and G2 by setting the unset classification points P (step S309). Then the process proceeds to step S304. In this way, the environmental change detection unit 27S repeats the operations of steps S304 to S309 until all classification points P have been set.

[0090] Then, if all classification points P are set in step S308 ("Y" in step S308), this process ends.

[0091] Figure 10 shows an example of operation of the environmental change detection unit 27S.

[0092] In case C11, the abnormal value VRES gradually increases over time. However, there is no significant change in the abnormal value VRES, nor is there a significant change in the variability of the abnormal value VRES. Therefore, regardless of which classification point P is set, the ratio of the mean value μ in groups G1 and G2 is within a predetermined range, and the ratio of the variance A in groups G1 and G2 is within a predetermined range. Therefore, the environmental change detection unit 27S determines that the learning processing unit 23 does not need to perform retraining.

[0093] In case C12, the abnormal value VRES becomes large after timing t1. Therefore, for example, if a classification point P is set at this timing t1, the ratio of the mean values ​​μ in groups G1 and G2 is not within the predetermined range. On the other hand, in this example, regardless of which classification point P is set, the ratio of the variance values ​​A in groups G1 and G2 is within the predetermined range. Therefore, the environmental change detection unit 27S determines that the learning processing unit 23 should perform retraining.

[0094] In case C13, the variability of the abnormal value VRES increases after timing t2. Therefore, for example, if a division point P is set at timing t2, the ratio of the variance values ​​A in groups G1 and G2 is not within the predetermined range. On the other hand, in this example, regardless of which division point P is set, the ratio of the mean values ​​μ in groups G1 and G2 is within the predetermined range. Therefore, the environmental change detection unit 27S determines that the learning processing unit 23 should perform retraining.

[0095] The operation of the environmental change detection unit 27S has been described above, but the same applies to the environmental change detection unit 27L. The environmental change detection unit 27S is supplied with abnormal values ​​VRES intermittently at a rate of approximately once every minute in this example. On the other hand, the environmental change detection unit 27L is supplied with abnormal values ​​VRES intermittently at a rate of approximately once every hour in this example. As a result, the environmental change detection unit 27S can detect when the environment has changed in a short period of time, and the environmental change detection unit 27L can detect when the environment has changed slowly over a long period of time.

[0096] (Operation of Data Verification Unit 24) The data verification unit 24 determines whether the spectral data D3 is suitable for the retraining process, as shown in step S104 in Figure 5A and step S113 in Figure 5B. This operation will be described in detail below.

[0097] Figure 11 shows an example of the operation of the data verification unit 24. During the retraining period, the average value VAV and PtoP value VPP are continuously supplied to the data verification unit 24 at a rate of approximately once every two seconds in this example. Each time the average value VAV and PtoP value VPP are supplied, the data verification unit 24 performs the following processing.

[0098] The data verification unit 24 checks whether the average value VAV is within a predetermined range (step S401). Specifically, the data verification unit 24 checks whether the average value VAV is greater than threshold TH6 and less than threshold TH7. These thresholds TH6 and TH7 are set in the sensor module 10 in advance. If the average value VAV is not within the predetermined range ("N" in step S401), the process proceeds to step S403.

[0099] If the average value VAV is within a predetermined range in step S401 ("Y" in step S401), the data verification unit 24 checks whether the PtoP value VPP is within a predetermined range (step S402). Specifically, the data verification unit 24 checks whether the PtoP value VPP is greater than the threshold TH8. This threshold TH8 is set in the sensor module 10 in advance. If the PtoP value VPP is not within a predetermined range ("N" in step S402), the process proceeds to step S403.

[0100] If the average value VAV is not within a predetermined range in step S401 ("N" in step S401), and if the PtoP value VPP is not within a predetermined range in step S402 ("N" in step S402), the data verification unit 24 determines that the learning processing unit 23 should stop the retraining process (step S403). Then, this process ends.

[0101] Thresholds TH6 to TH8 are pre-set based on data used to generate the initial learning model M, for example. Specifically, as described above, the initial learning model M is generated by a personal computer (not shown) performing a learning process based on data showing vibrations of mechanical devices of the same model number as mechanical devices 100A to 100C. This personal computer for learning processes inputs spectral data similar to spectral data D3 into the learning model M and performs a learning process so that the learning model M outputs spectral data identical to the input data. When generating this spectral data, a standardization process is performed to calculate the average value VAV. Based on the multiple average value VAVs related to the multiple spectral data used in this learning process, the average value of these multiple average value VAVs and the standard deviation of the multiple average value VAVs can be calculated. Then, the value obtained by subtracting three times the standard deviation from the average value of the multiple average value VAVs can be used as threshold TH6, and the value obtained by adding three times the standard deviation to the average value of the multiple average value VAVs can be used as threshold TH7. Furthermore, the point-to-point (PtoP) value can be calculated based on the sensor data, which is the source data for this spectral data. The minimum value among the multiple PtoP values ​​related to multiple sense data can then be used as the threshold value TH8.

[0102] As a result, the data verification unit 24 can determine whether the learning processing unit 23 should terminate the retraining process based on whether the average VAV is approximately the same as the average VAV obtained using the data when the initial learning model M was generated, and whether the PtoP value is approximately the same as the PtoP value obtained using the data when the initial learning model M was generated.

[0103] For example, if the average VAV is approximately the same as the average VAV obtained using the data generated when the initial learning model M was created, and the PtoP value is approximately the same as the PtoP value obtained using the data generated when the initial learning model M was created, the data verification unit 24 determines that the learning processing unit 23 should perform retraining based on this spectral data D3. In other words, in this case, the spectral data D3 generated by the data generation unit 21 is presumed to be data from a situation where no significant malfunction has occurred in the mechanical device 100. In this case, the data verification unit 24 can determine that this spectral data D3 is appropriate data for retraining. Therefore, the data verification unit 24 determines that the learning processing unit 23 should perform retraining based on this spectral data D3.

[0104] Furthermore, if, for example, the average VAV is not at a similar value to the average VAV obtained using the data from when the initial learning model M was generated, the data verification unit 24 determines that the learning processing unit 23 should not perform retraining based on this spectral data D3. Also, if, for example, the PtoP value is not at a similar value to the PtoP value obtained using the data from when the initial learning model M was generated, the data verification unit 24 determines that the learning processing unit 23 should not perform retraining based on this spectral data D3. In other words, in this case, the spectral data D3 generated by the data generation unit 21 is presumed to be data from a situation where there is a significant malfunction in the mechanical device 100. In this case, the data verification unit 24 can determine that this spectral data D3 is not suitable data for retraining. Therefore, the data verification unit 24 determines that the learning processing unit 23 should discontinue the retraining process.

[0105] Figure 12 shows a specific example of the operation of the data verification unit 24.

[0106] In case C21, the average VAV is 1.2 × 10⁻⁶ －４Therefore, the PtoP value VPP is 0.49. In this example, the average value VAV is within a predetermined range, and the PtoP value VPP is within a predetermined range. Thus, the data verification unit 24 determines that the spectral data D3 is suitable for the retraining process and decides to perform the retraining process based on this spectral data D3.

[0107] In case C22, the average VAV is 3.2 × 10⁻⁶ －４ Therefore, the PtoP value VPP is 0.75. In this example, the average value VAV is higher than the predetermined range, and the PtoP value VPP is within the predetermined range. Thus, the data verification unit 24 determines that the spectral data D3 is not suitable for the retraining process and decides that the retraining process should be terminated.

[0108] In case C23, the average VAV is 3.1 × 10⁻⁶ －６ Therefore, the PtoP value VPP is 0.1. In this example, the average value VAV is within a predetermined range, and the PtoP value VPP is below the predetermined range. Thus, the data verification unit 24 determines that the spectral data D3 is not suitable for the retraining process and decides to terminate the retraining process.

[0109] As described above, the sensor module 10 includes a sensor (accelerometer 11) capable of detecting a physical quantity corresponding to the operating state of the device (mechanical device 100), and a processing circuit 20 capable of performing monitoring processing to monitor the operating state of the device (mechanical device 100) using a first learning model (learning model M) based on the detection results of the sensor (accelerometer 11). The processing circuit 20 is capable of performing relearning processing on the first learning model (learning model M) based on the detection results of the sensor (accelerometer 11), and in the relearning processing, it is possible to perform verification processing to verify the data to be input to the first learning model (learning model M), and based on the results of the verification processing, it is possible to cancel the relearning processing. As a result, the sensor module 10 can determine whether the spectral data D3 generated based on the detection results of the acceleration sensor 11 is appropriate for the relearning processing, and if the spectral data D3 is not appropriate for the relearning processing, it can cancel the relearning processing without performing relearning processing based on this spectral data D3. The sensor module 10 performs a verification process based on the detection results of the acceleration sensor 11, and then cancels the retraining process based on the results of that verification process. As a result, the sensor module 10 can reduce the burden on the operator.

[0110] In other words, for example, if the operator has to check whether each of the multiple spectral data D3 is suitable for the retraining process when performing the retraining process, the burden on the operator becomes significant. On the other hand, the sensor module 10 performs a verification process based on the detection result of the acceleration sensor 11 and stops the retraining process based on the result of the verification process. As a result, the operator does not need to check whether each of the multiple spectral data D3 is suitable for the retraining process. Consequently, the sensor module 10 can reduce the burden on the operator.

[0111] Furthermore, in the sensor module 10, the processing circuit 20 can sequentially generate first spectral data (spectral data D2) by sequentially performing a Fourier transform on sensor data D1, which is data at a predetermined time included in the detection result of the sensor (accelerometer 11). It can generate second spectral data (spectral data D3) by standardizing the first spectral data (spectral data D2) using the average value VAV of each frequency component in the first spectral data (spectral data D2). Monitoring processing can be performed by sequentially inputting the second spectral data (spectral data D3) into the first learning model (learning model M). The processing circuit 20 can also confirm several conditions in the verification process, including that the peak-to-peak value (PtoP value VPP) in sensor data D1 is within a predetermined range, and that the average value VAV in the first spectral data (spectral data D2) generated based on sensor data D1 is within a predetermined range. The processing circuit 20 can also cancel the retraining process if at least one of the multiple conditions is not met. As a result, the processing circuit 20 can determine, for example, whether the spectral data D3 is suitable for retraining based on the PtoP value VPP in the sensor data D1 and the average value VAV in the spectral data D2. Consequently, the sensor module 10 can reduce the burden on the operator.

[0112] Furthermore, in the sensor module 10, the processing circuit 20 is capable of performing a retraining process based on the sensor data D1 and the second spectral data (spectral data D3) corresponding to the first spectral data (spectral data D2) used to check whether multiple conditions are met, when multiple conditions are met. As a result, the processing circuit 20 can individually determine whether the spectral data D3 related to the sensor data D1 and spectral data D2 is suitable for the retraining process, based on the PtoP value VPP in the sensor data D1 and the average value VAV in the spectral data D2. Consequently, the sensor module 10 can reduce the burden on the operator.

[0113] Furthermore, in the sensor module 10, the processing circuit 20 is configured to sequentially generate monitoring result values ​​(abnormal value VRES) according to the operating state of the device (mechanical device 100) by using a first learning model (learning model M). The processing circuit 20 is configured to confirm that when the sensor module 10 is started for the first time, it performs a first confirmation process to check the monitoring result value (abnormal value VRES). The processing circuit 20 is configured to perform a retraining process based on the result of the first confirmation process. In the first confirmation process, the processing circuit 20 is configured to confirm whether the monitoring result value (abnormal value VRES) is within a predetermined range. The processing circuit 20 is configured to perform a retraining process if the monitoring result value (abnormal value VRES) is not within the predetermined range. As a result, for example, if the abnormal value VRES is approximately the same as the abnormal value VRES obtained using the data when the initial learning model M was generated, the initial startup verification unit 26 can determine that the learning processing unit 23 does not perform a retraining process. Furthermore, if the abnormal value VRES is not at the same level as the abnormal value VRES obtained using the data generated when the initial learning model M was created, the initial startup verification unit 26 can determine that the learning processing unit 23 should perform retraining. As a result, the sensor module 10 can reduce the burden on the operator.

[0114] Furthermore, in the sensor module 10, the processing circuit 20 is configured to sequentially generate monitoring result values ​​(abnormal value VRES) according to the operating state of the device (mechanical device 100) by using a first learning model (learning model M). The processing circuit 20 is configured to perform a second verification process to check for changes in the monitoring result values ​​(abnormal value VRES) based on a plurality of monitoring result values ​​(abnormal value VRES) obtained over a predetermined period. The processing circuit 20 is configured to perform a relearning process based on the results of the second verification process. The processing circuit 20 is capable of sequentially setting classification points P over a predetermined period in the second verification process, calculating a first ratio (μ1 / μ2) which is the ratio of the average values ​​μ1 and μ2 of the monitored result values ​​(abnormal value VRES) in each of the two intervals divided by the classification points P, and calculating a second ratio (A1 / A2) which is the ratio of the variance values ​​(variance values ​​A1 and A2) of the monitored result values ​​(abnormal value VRES) in each of the two intervals, and is capable of verifying multiple conditions, including that the first ratio is within a predetermined range and that the second ratio is within a predetermined range. The processing circuit 20 is capable of performing a relearning process if at least one of the multiple conditions is not met. As a result, the environmental change detection units 27S and 27L can detect whether there is a large difference between the abnormal value VRES related to group G1 and the abnormal value VRES related to group G2 due to environmental changes. Furthermore, the environmental change detection units 27S and 27L can determine that the learning processing unit 23 should perform retraining if there is a large difference between these abnormal values ​​VRES. As a result, the sensor module 10 can reduce the burden on the operator.

[0115] [Effects] As described above, this embodiment includes a sensor capable of detecting a physical quantity corresponding to the operating state of the device, and a processing circuit capable of performing monitoring processing to monitor the operating state of the device using a first learning model based on the sensor's detection results. The processing circuit is capable of performing relearning processing for the first learning model based on the sensor's detection results, and in the relearning processing, it is capable of performing verification processing to verify the data input to the first learning model, and it is possible to cancel the relearning processing based on the results of the verification processing. This reduces the burden on the operator.

[0116] In this embodiment, the processing circuit can sequentially generate first spectral data by sequentially performing Fourier transforms on sensor data, which is data at a predetermined time included in the sensor's detection results. It can generate second spectral data by standardizing the first spectral data using the average value of each frequency component in the first spectral data. Monitoring processing can be performed by sequentially inputting the second spectral data into the first learning model. Furthermore, the processing circuit can confirm multiple conditions in the verification process, including that the peak-to-peak value in the sensor data is within a predetermined range, and that the average value in the first spectral data generated based on the sensor data is within a predetermined range. Furthermore, the processing circuit can stop the retraining process if at least one of the multiple conditions is not met. This reduces the burden on the operator.

[0117] In this embodiment, the processing circuit is configured to perform a retraining process based on the sensor data used to check whether multiple conditions are met and the second spectral data corresponding to the first spectral data, thereby reducing the burden on the operator.

[0118] In this embodiment, the processing circuit is configured to sequentially generate monitoring result values ​​according to the operating state of the device by using a first learning model. The processing circuit is configured to perform a first confirmation process to check the monitoring result values ​​when the sensor module is started for the first time. The processing circuit is configured to perform a relearning process based on the results of the first confirmation process. The processing circuit is configured to check whether the monitoring result values ​​are within a predetermined range during the first confirmation process. The processing circuit is configured to perform a relearning process if the monitoring result values ​​are not within the predetermined range. This reduces the burden on the operator.

[0119] In this embodiment, the processing circuit is capable of sequentially generating monitoring result values ​​according to the operating state of the device by using a first learning model. The processing circuit is capable of performing a second verification process to check for changes in the monitoring result values ​​based on a plurality of monitoring result values ​​obtained over a predetermined period. The processing circuit is capable of performing a relearning process based on the results of the second verification process. In the second verification process, the processing circuit is capable of sequentially setting division points over a predetermined period, calculating a first ratio which is the ratio of the average values ​​of the monitoring result values ​​in each of the two intervals divided by the division points, calculating a second ratio which is the ratio of the variation values ​​of the monitoring result values ​​in each of the two intervals, and confirming a plurality of conditions including that the first ratio is within a predetermined range and that the second ratio is within a predetermined range. The processing circuit is capable of performing a relearning process if at least one of the plurality of conditions is not met. This reduces the burden on the operator.

[0120] [Modification] In the above embodiment, the environmental change detection units 27S and 27L detected environmental changes based on the abnormal value VRES, but are not limited to this. Alternatively, for example, environmental changes may be detected based on spectral data D3. The sensor module 30 according to this modification will be described in detail below.

[0121] Figure 13 shows an example configuration of the sensor module 30. The sensor module 30 has a processing circuit 40. By executing software, the processing circuit 40 can operate as a data generation unit 41, a monitoring processing unit 22, a learning processing unit 23, a data verification unit 24, a relearning determination unit 45, and a control processing unit 49.

[0122] The data generation unit 41 is configured to generate standardized spectral data D3 for acceleration by performing a Fourier transform based on time-series data of acceleration data supplied from the acceleration sensor 11, similar to the data generation unit 21 in the above embodiment. In this example, the data generation unit 41 intermittently generates spectral data D3 at a rate of approximately once per minute and supplies the generated spectral data D3 to the environmental change detection unit 47S. In addition, in this example, the data generation unit 41 intermittently generates spectral data D3 at a rate of approximately once per hour and supplies the generated spectral data D3 to the environmental change detection unit 47L.

[0123] The retraining determination unit 45 includes environmental change detection units 47S and 47L. Each of the environmental change detection units 47S and 47L is configured to detect environmental changes based on spectral data D3 and to determine whether the learning processing unit 23 should perform retraining processing when an environmental change is detected. In this example, the environmental change detection unit 47S can detect that the environment has changed in a short period of time based on spectral data D3 supplied at a rate of approximately once every minute. In this example, the environmental change detection unit 47L can detect that the environment has changed slowly over a long period of time based on spectral data D3 supplied at a rate of approximately once every hour.

[0124] As shown in step S111 of Figure 5B, the environmental change detection units 47S and 47L determine whether the learning processing unit 23 should perform retraining based on the environmental change. The operation of the environmental change detection unit 47S will be explained in detail below, using it as an example. The operation of the environmental change detection unit 47L is similar.

[0125] Figure 14 shows an example of the operation of the environmental change detection unit 47S. Spectrum data D3 is supplied to the environmental change detection unit 47S intermittently, at a rate of approximately once every minute in this example. The environmental change detection unit 47S performs the following processing each time spectrum data D3 is supplied.

[0126] First, the environmental change detection unit 47S stores the supplied spectral data D3 (step S501).

[0127] Next, the environmental change detection unit 47S checks whether the number of accumulated spectral data D3 is greater than or equal to a predetermined number (for example, 100) (step S502). If the number of accumulated spectral data D3 is less than the predetermined number ("N" in step S502), this process ends.

[0128] In step S502, if the number of accumulated spectral data D3 is greater than or equal to a predetermined number (Y in step S502), the environmental change detection unit 47S performs a learning process based on the latest predetermined number (e.g., 100) of spectral data D3 to generate a learning model M2 in which spectral data D3 is input and label value VL is output (step S503). Here, the learning model M2 corresponds to a specific example of the "second learning model" in one embodiment of the present disclosure.

[0129] Figure 15 shows an example of a learning model M2. Spectral data D3 is input to the learning model M2. Then, a label value VL is output from the learning model M2. The label value VL is a value between 0 and 1, inclusive.

[0130] Figure 16 shows an example of the learning process. In this example, the learning process uses 100 spectral data points D3 and 100 corresponding label values ​​VL. The label values ​​VL are the training data. For the sake of explanation, Figure 16 shows 10 spectral data points D3 and 10 label values ​​VL. The horizontal axis represents time. That is, the leftmost spectral data point D3 represents the first spectral data point D3 out of 100 spectral data points D3, and the rightmost spectral data point D3 represents the last spectral data point D3 out of 100 spectral data points D3. Of the 100 spectral data points D3, for example, the 80 label values ​​VL corresponding to the first 80 spectral data points D3 are "0", and the 20 label values ​​VL corresponding to the remaining 20 spectral data points D3 are "1".

[0131] The environmental change detection unit 47S generates a learning model M2 by performing a learning process based on the spectral data D3 and label value VL.

[0132] Next, the environmental change detection unit 47S generates a label value VL by inputting each of the multiple spectral data D3 used in the learning process into the learning model M2 (step S504). Then, the environmental change detection unit 47S calculates the accuracy ACC of the label value VL generated in step S504 (step S505). In steps S504 and S505, the environmental change detection unit 47S calculates the accuracy ACC using the following formula. Here, CD_AI indicates processing using the learning model M2, and the arguments indicate the data input to the learning model M2. Data represents the spectral data D3. VLB is the binarized value of the label value VL. VLT is the training data for the label value VL. The subscript i indicates the index number. N indicates the number of spectral data D3.

[0133] First, the environmental change detection unit 47S obtains a label value VL by inputting spectral data D3 into the learning model M2, as shown in equation F1. Then, as shown in equation F2, the environmental change detection unit 47S obtains a label value VLB by binarizing the label value VL using 0.5 as the reference. Then, as shown in equation F3, the environmental change detection unit 47S calculates the average of the absolute values ​​of the difference between the binarized label value VLB and the training data label value VL, and calculates the accuracy score ACC by subtracting this average value from 1. For example, if all of the binarized label values ​​VLB are the same as the training data, the second term on the right side becomes 0 (zero), so the accuracy score ACC is 1. Also, if all of the binarized label values ​​VLB are different from the training data, the second term on the right side becomes 1, so the accuracy score ACC is 0 (zero).

[0134] Thus, the environmental change detection unit 47S generates a learning model M2 by performing a learning process in step S504, and verifies the learning model M2 in steps S505 and S506.

[0135] Next, the environmental change detection unit 47S checks whether the accuracy ACC is higher than the threshold TH9 (step S506). The threshold TH9 is set in advance. The threshold TH9 is, for example, 0.7. If the accuracy ACC is less than or equal to the threshold TH9 ("N" in step S506), this process ends.

[0136] In step S506, if the accuracy ACC is higher than the threshold TH9 ("Y" in step S506), the environmental change detection unit 47S determines that the learning processing unit 23 should perform retraining (step S507).

[0137] This completes the process.

[0138] Figure 17 schematically represents the determination method shown in Figure 16, where (A) shows the case where there is no environmental change, and (B) shows the case where there is an environmental change. In Figure 17, the rectangles represent spectral data D3. In Figure 17, the spectral data D3 are drawn at different positions to correspond to the fact that the spectral data D3 are different from each other. Shaded rectangles represent spectral data D3 for which the training data of the label value VL is 1, and unshaded rectangles represent spectral data D3 for which the training data of the label value VL is 0 (zero).

[0139] For example, in the absence of environmental fluctuations, multiple spectral data points D3 are similar to each other, as shown in Figure 17(A). When the environmental fluctuation detection unit 47S performs learning processing based on such spectral data points D3, it is difficult to distinguish between spectral data points D3 with a training data of 0 (zero) and spectral data points D3 with a training data of 1 using the learned model M2. Therefore, in this case, the accuracy ACC is low.

[0140] On the other hand, when there are environmental changes, as shown in Figure 17(B), the spectral data D3 before the environmental change occurs and the spectral data D3 after the environmental change occur may be different from each other. Therefore, when the environmental change detection unit 47S performs learning processing based on such spectral data D3, it can use the learned learning model M2 to distinguish between spectral data D3 where the training data is 0 (zero) and spectral data D3 where the training data is 1. In this case, the accuracy ACC is high.

[0141] The operation of the environmental change detection unit 47S has been described above, but the same applies to the environmental change detection unit 47L. Spectrum data D3 is supplied to the environmental change detection unit 47L intermittently, at a rate of approximately once every hour in this example. This allows the environmental change detection unit 47L to detect that the environment has changed slowly over a long period of time.

[0142] Thus, in the sensor module 30, the processing circuit 40 is capable of performing a learning process for a second learning model (learning model M2) so as to classify two or more second spectral data (spectral data D3) into two classes, performing a third verification process to confirm whether two or more second spectral data (spectral data D3) have been classified into two classes, and performing a relearning process based on the result of the third verification process. As a result, the environmental change detection units 47S and 47L can detect whether an environmental change has occurred based on the spectral data D3, and thereby determine whether the learning processing unit 23 should perform a relearning process. Consequently, the sensor module 30 can reduce the burden on the operator.

[0143] Although the present invention has been described above with reference to embodiments and modifications, the present invention is not limited to these embodiments and various modifications are possible.

[0144] For example, while an acceleration sensor 11 is provided in each of the embodiments described above, the invention is not limited to this. Various sensors can be provided to detect various physical quantities that can be used to monitor the operating state of the device, such as a temperature sensor for detecting temperature and a microphone for detecting sound.

[0145] For example, in the embodiments described above, the processing circuit 20 of the sensor module 10 performs the processing, but this is not the only option. Alternatively, a personal computer may perform the processing.

[0146] The effects described herein are illustrative only, and the effects of this disclosure are not limited to those described herein. Therefore, other effects may be obtained with respect to this disclosure.

[0147] Furthermore, this disclosure may take the following forms:

[0148] (1) A sensor module comprising: a sensor capable of detecting a physical quantity corresponding to the operating state of a device; and a processing circuit capable of performing monitoring processing to monitor the operating state of the device using a first learning model based on the detection result of the sensor, wherein the processing circuit is capable of performing retraining processing on the first learning model based on the detection result of the sensor; performing verification processing to verify the data to be input to the first learning model in the retraining processing; and being able to cancel the retraining processing based on the result of the verification processing. (2) The sensor module according to (1), wherein the processing circuit is capable of sequentially generating first spectral data by sequentially performing Fourier transforms on sensor data which is data at a predetermined time included in the detection result of the sensor; generating second spectral data by standardizing the first spectral data using the average value of each frequency component in the first spectral data; and performing the monitoring processing by sequentially inputting the second spectral data to the first learning model. (3) The sensor module according to (2), wherein the processing circuit can confirm a plurality of conditions in the verification process, including that the peak-to-peak value in the sensor data is within a predetermined range and that the average value in the first spectral data generated based on the sensor data is within a predetermined range, and the relearning process can be terminated if at least one of the plurality of conditions is not met. (4) The sensor module according to (3), wherein the processing circuit can perform the relearning process based on the sensor data and the second spectral data corresponding to the first spectral data used to confirm whether the plurality of conditions are met, when the plurality of conditions are met.(5) The sensor module according to any one of (1) to (4), wherein the processing circuit is capable of sequentially generating monitoring result values ​​according to the operating state of the device by using the first learning model, and is capable of performing a first confirmation process to confirm the monitoring result values ​​when the sensor module is started for the first time, and is capable of performing the relearning process based on the result of the first confirmation process. (6) The sensor module according to (5), wherein the processing circuit is capable of confirming in the first confirmation process whether the monitoring result values ​​are within a predetermined range, and is capable of performing the relearning process if the monitoring result values ​​are not within the predetermined range. (7) The sensor module according to any one of (1) to (6), wherein the processing circuit is capable of sequentially generating monitoring result values ​​according to the operating state of the device by using the first learning model, and is capable of performing a second confirmation process to confirm changes in the monitoring result values ​​based on a plurality of monitoring result values ​​obtained over a predetermined period, and is capable of performing the relearning process based on the result of the second confirmation process. (8) The sensor module according to (7), wherein the processing circuit is capable of sequentially setting division points in the second verification process over a predetermined period, calculating a first ratio which is the ratio of the average values ​​of the monitoring result values ​​in each of the two intervals divided by the division points, calculating a second ratio which is the ratio of the variation values ​​of the monitoring result values ​​in each of the two intervals, and confirming a plurality of conditions including that the first ratio is within a predetermined range and that the second ratio is within a predetermined range, and the processing circuit is capable of performing the relearning process if at least one of the plurality of conditions is not met.(9) The sensor module according to any one of (2) to (4), wherein the processing circuit is capable of performing a learning process of a second learning model so as to classify two or more of the second spectral data into two classes, is capable of performing a third verification process to confirm whether the two or more of the second spectral data have been classified into two classes, and is capable of performing the retraining process based on the result of the third verification process. (10) The sensor module according to any one of (1) to (8), wherein the device is a mechanical device having a movable member, and the sensor is an acceleration sensor. (11) The information processing device comprising a processing circuit capable of performing a monitoring process to monitor the operating state of the device using a first learning model based on the detection result of a sensor capable of detecting a physical quantity corresponding to the operating state of the device, wherein the processing circuit is capable of performing a retraining process of the first learning model based on the detection result of the sensor, is capable of performing a verification process to verify the detection result of the sensor in the retraining process, and is capable of canceling the retraining process based on the result of the verification process.

Claims

1. A sensor module comprising: a sensor capable of detecting a physical quantity corresponding to the operating state of a device; and a processing circuit capable of performing monitoring processing to monitor the operating state of the device using a first learning model based on the detection result of the sensor, wherein the processing circuit is capable of performing relearning processing for the first learning model based on the detection result of the sensor; performing verification processing to verify the data to be input to the first learning model in the relearning processing; and being able to cancel the relearning processing based on the result of the verification processing.

2. The sensor module according to claim 1, wherein the processing circuit can sequentially generate first spectral data by sequentially performing a Fourier transform on sensor data, which is data at a predetermined time included in the detection result of the sensor; can generate second spectral data by standardizing the first spectral data using the average value of each frequency component in the first spectral data; and can perform the monitoring process by sequentially inputting the second spectral data to the first learning model.

3. The sensor module according to claim 2, wherein the processing circuit can confirm, in the verification process, a plurality of conditions including that the peak-to-peak value in the sensor data is within a predetermined range and that the average value in the first spectral data generated based on the sensor data is within a predetermined range, and the retraining process can be terminated if at least one of the plurality of conditions is not met.

4. The sensor module according to claim 3, wherein the processing circuit is capable of performing the retraining process based on the sensor data used to check whether the multiple conditions are met and the second spectral data corresponding to the first spectral data, when the multiple conditions are met.

5. The sensor module according to claim 1, wherein the processing circuit is capable of sequentially generating monitoring result values ​​according to the operating state of the device by using the first learning model, can be confirmed to perform a first confirmation process to confirm the monitoring result values ​​when the sensor module is started for the first time, and can perform the relearning process based on the result of the first confirmation process.

6. The sensor module according to claim 5, wherein the processing circuit can confirm in the first confirmation process whether the monitoring result value is within a predetermined range, and can perform the relearning process if the monitoring result value is not within the predetermined range.

7. The sensor module according to claim 1, wherein the processing circuit is capable of sequentially generating monitoring result values ​​corresponding to the operating state of the device by using the first learning model, performing a second verification process to confirm changes in the monitoring result values ​​based on a plurality of the monitoring result values ​​obtained over a predetermined period, and performing the relearning process based on the result of the second verification process.

8. The sensor module according to claim 7, wherein the processing circuit is capable of sequentially setting division points during the predetermined period in the second verification process, calculating a first ratio which is the ratio of the average values ​​of the monitoring result values ​​in each of the two intervals divided by the division points, calculating a second ratio which is the ratio of the variation values ​​of the monitoring result values ​​in each of the two intervals, confirming a plurality of conditions including that the first ratio is within a predetermined range and that the second ratio is within a predetermined range, and the processing circuit is capable of performing the relearning process if at least one of the plurality of conditions is not met.

9. The sensor module according to claim 2, wherein the processing circuit is capable of performing a learning process of a second learning model so as to classify two or more of the second spectral data into two classes, is capable of performing a third verification process to confirm whether the two or more of the second spectral data have been classified into two classes, and is capable of performing the retraining process based on the result of the third verification process.

10. The sensor module according to claim 1, wherein the device is a mechanical device having a movable member, and the sensor is an acceleration sensor.

11. An information processing device comprising: a processing circuit capable of performing monitoring processing to monitor the operating state of a device using a first learning model based on the detection result of a sensor capable of detecting a physical quantity corresponding to the operating state of the device, wherein the processing circuit is capable of performing a retraining process of the first learning model based on the detection result of the sensor, is capable of performing a verification process to verify the data to be input to the first learning model in the retraining process, and is capable of canceling the retraining process based on the result of the verification process.

Images