WO2022270124A1

WO2022270124A1 - Learning device, inference device, learning method, and inference method

Info

Publication number: WO2022270124A1
Application number: PCT/JP2022/017153
Authority: WO
Inventors: 友実堀; 孝吉岡; 隆彦増▲崎▼; 道康日下; 堅也杉原
Original assignee: 三菱電機株式会社
Priority date: 2021-06-22
Filing date: 2022-04-06
Publication date: 2022-12-29
Also published as: JPWO2022270124A1

Abstract

A learning device includes: a data acquisition unit that acquires time-series data of a normal time; a time frequency analysis unit that analyzes a chronical change of a frequency feature to the time-series data of the normal time acquired by the data acquisition unit; a feature difference emphasis range extraction unit that extracts, from the data that has been analyzed by the time frequency analysis unit, a signal intensity data in a range of a specified frequency band and a specified time period where a feature difference between the normal time and an abnormal time increases; and a model generation unit that uses the signal intensity data extracted by the feature difference emphasis range extraction unit in the range of the specified frequency band and the specified time period to generate a learned model for inferring similarity with the data of the normal time.

Description

Learning device, reasoning device, learning method and reasoning method

The present disclosure relates to a learning device, an inference device, a learning method, and an inference method.

Patent Document 1 discloses a method for judging the state of a prime mover. According to this method, the state of the prime mover can be determined by reducing the amount of data while retaining the characteristics of the original data.

JP 2020-165330 A

However, in the state determination method described in Patent Document 1, data may be acquired under poor conditions. In this case, anomalies cannot be detected accurately.

The present disclosure was made to solve the above problems. An object of the present disclosure is to provide a learning device, an inference device, a learning method, and an inference method that can accurately detect anomalies even from data acquired under poor conditions.

A learning device according to the present disclosure includes a data acquisition unit that acquires normal time-series data, and a time-frequency and an analysis unit, and feature difference enhancement for extracting signal strength data in a specific frequency band and a specific time range where the feature difference between normal and abnormal conditions is large from the data analyzed by the time-frequency analysis unit. Using the range extracting unit and the signal intensity data in the range of the specific frequency band and the specific time band extracted by the feature difference emphasizing range extracting unit, the learned for inferring the similarity with the normal data and a model generator for generating a model.

Another learning device according to the present disclosure includes a data acquisition unit that acquires normal time-series data, and analyzes temporal changes in frequency characteristics for the normal time-series data acquired by the data acquisition unit. A feature of extracting signal strength data in a specific frequency band and a specific time range where the feature difference between normal and abnormal conditions is large from the time-frequency analysis unit and the data analyzed by the time-frequency analysis unit. a difference emphasizing range extracting unit; a statistic calculating unit that calculates a statistic from signal strength data in a specific frequency band and a specific time band extracted by the feature difference emphasizing range extracting unit; a model generation unit that generates a trained model for inferring the degree of similarity with normal data from the statistics calculated by the model generation unit; a normalization parameter derivation unit for inputting all and deriving normalization parameters.

An inference device according to the present disclosure includes a data acquisition unit that acquires time-series data of an object under test, and analyzes temporal changes in frequency characteristics of the time-series data of the object under test acquired by the data acquisition unit. A feature of extracting signal strength data in a specific frequency band and a specific time range where the feature difference between normal and abnormal conditions is large from the time-frequency analysis unit and the data analyzed by the time-frequency analysis unit. Using a difference-enhancement range extraction unit and a trained model for inferring the degree of similarity with normal data, the range of a specific frequency band and a specific time zone extracted by the feature difference-enhancement range extraction unit and an inference unit for inferring the degree of similarity between the signal strength data and the normal data.

Another inference device according to the present disclosure includes a data acquisition unit that acquires time-series data of an object to be inspected; From the time-frequency analysis unit to be analyzed and the data analyzed by the time-frequency analysis unit, extract signal strength data in a specific frequency band and a specific time range where the feature difference between normal and abnormal conditions is large. a statistic calculation unit for calculating a statistic from the signal strength data in a specific frequency band and a specific time range extracted by the feature difference emphasis range extraction unit; an inference unit for inferring the degree of similarity between the statistic calculated by the statistic calculation unit and normal data using a trained model for inferring the degree of similarity with the data; a dissimilarity deriving unit that converts the obtained similarity into a dissimilarity using the normalization parameter stored in the storage device and normalizes the similarity.

The learning method according to the present disclosure includes a data acquisition step of acquiring normal time-series data using a learning device, and a normal time-series data acquired by the data acquisition step using the learning device. On the other hand, a time-frequency analysis step of analyzing temporal changes in frequency characteristics, and using the learning device, from the data analyzed by the time-frequency analysis step, a feature difference between normal and abnormal times is specified. A feature difference emphasis range extraction step for extracting signal strength data in a frequency band and a specific time zone range, and a specific frequency band and specific frequency band extracted by the feature difference emphasis range extraction step using the learning device and a model generation step of generating a trained model for inferring similarity with normal data, using the signal intensity data within the time range.

Another learning method according to the present disclosure includes a data acquisition step of acquiring normal time series data using a learning device, and a normal time series data acquired by the data acquisition step using the learning device. A time-frequency analysis step of analyzing temporal changes in frequency characteristics of data, and from the data analyzed by the time-frequency analysis step using the learning device, the feature difference between normal and abnormal is large. a feature difference emphasis range extraction step of extracting signal strength data in a specific frequency band and a specific time range, and a specific frequency band extracted in the feature difference emphasis range extraction step and A statistic calculation step of calculating a statistic of signal strength data in a range of a specific time period; a model generating step of generating a trained model for inferring the similarity; using the learning device, inputting part or all of learning data into the trained model generated by the model generating step; and a normalization parameter derivation step for deriving a normalization parameter.

An inference method according to the present disclosure includes a data acquisition step of acquiring time-series data of a subject using an inference device, and a time-series data of the subject acquired by the data acquisition step using the inference device. A time-frequency analysis step of analyzing temporal changes in frequency characteristics of the data, and the data analyzed by the time-frequency analysis step using the inference device, the feature difference between normal and abnormal is large. A feature difference enhancement range extraction step of extracting signal intensity data in a specific frequency band and a specific time range, and a trained model for inferring similarity to normal data using the inference device and an inference step of inferring the degree of similarity between the signal intensity data in the range of the specific frequency band and the specific time band extracted by the feature difference enhancement range extraction step and the normal data, using

Another inference method according to the present disclosure includes a data acquisition step of acquiring time-series data of an object to be inspected using an inference device; A time-frequency analysis step of analyzing temporal changes in frequency characteristics of time-series data; and using the inference device, from the data analyzed by the time-frequency analysis step, a feature difference between normal and abnormal conditions. a feature difference emphasis range extracting step of extracting signal strength data in a specific frequency band and a specific time band range where the A statistic calculation step of calculating the statistic of the signal strength data in the range of the time period and the specific time period; , an inference step of inferring the degree of similarity with normal data using the statistics calculated in the statistic calculation step; and a dissimilarity derivation step of converting to dissimilarity and normalizing using the normalization parameter stored in the storage device.

According to the present disclosure, data within a specific frequency band and a specific time range where the feature difference between normal and abnormal conditions is large is used. Therefore, anomalies can be accurately detected even from data acquired under poor conditions.

1 is a configuration diagram of a learning device according to Embodiment 1; FIG. 4 is a diagram showing time-series data acquired by the learning device according to Embodiment 1. FIG. 4 is a diagram showing a mother wavelet function used for wavelet transform by the learning device according to Embodiment 1; FIG. 4 is a diagram showing a mother wavelet function used for wavelet transform by the learning device according to Embodiment 1; FIG. 4 is a first example of a feature difference emphasis range of a rope tester using the learning device according to Embodiment 1. FIG. FIG. 10 is a second example of a feature difference emphasis range of a rope tester using the learning device according to Embodiment 1; FIG. It is an example of the feature difference emphasis range when there is a break in the rope in the rope tester using the learning device according to Embodiment 1. FIG. FIG. 10 is a third example of a feature difference emphasis range of a rope tester using the learning device according to Embodiment 1; FIG. 4 is a flowchart for explaining learning processing of the learning device according to Embodiment 1; 1 is a configuration diagram of an inference device according to Embodiment 1; FIG. 4 is a flowchart for explaining inference processing of the inference device according to Embodiment 1; 2 is a hardware configuration diagram of a learning device according to Embodiment 1. FIG. FIG. 10 is a configuration diagram of a learning device according to Embodiment 2; 10 is a flowchart for explaining learning processing of the learning device according to Embodiment 2; FIG. 11 is a configuration diagram of an inference device according to Embodiment 2; 10 is a flowchart for explaining inference processing of the inference device according to Embodiment 2; FIG. 11 is a configuration diagram of a learning device according to Embodiment 3; 14 is a flow chart for explaining learning processing of the learning device according to Embodiment 3. FIG. FIG. 11 is a configuration diagram of an inference device according to Embodiment 3; 14 is a flowchart for explaining inference processing of an inference device according to Embodiment 3;

An embodiment will be described according to the attached drawings. In addition, the same code|symbol is attached|subjected to the part which is the same or corresponds in each figure. Redundant description of the relevant part will be simplified or omitted as appropriate.

Embodiment 1.
FIG. 1 is a block diagram of a learning device according to Embodiment 1. FIG.

The learning device 1 in FIG. 1 is used to detect abnormalities in equipment and members. For example, the learning device 1 is used to detect rope abnormalities using a rope tester, which is a device for detecting wire rope breakage.

The learning device 1 includes a data acquisition unit 2, a time-frequency analysis unit 3, a feature difference emphasis range extraction unit 4, and a model generation unit 5.

The data acquisition unit 2 captures the time series of the voltage signal generated when the magnetic flux (leakage magnetic flux) leaking from the surface of the wire rope (leakage magnetic flux) passes through the coil when magnetizing a normal wire rope that is sent out in the longitudinal direction without disconnection. Get data.

The time-frequency analysis unit 3 performs time-frequency analysis for analyzing temporal changes in frequency characteristics for time-series data. For example, the time-frequency analysis unit 3 performs analysis using a short-time Fourier transform as the time-frequency analysis. For example, the time-frequency analysis unit 3 performs analysis using wavelet transform using a general mother wavelet function such as Biorthogonal 3.3 as the time-frequency analysis. For example, the time-frequency analysis unit 3 performs analysis using wavelet transform using a mother wavelet function similar to the voltage waveform at the time of wire rope disconnection.

The feature difference enhancement range extracting unit 4 extracts a specific frequency band and a specific frequency band in which the feature difference between normal and abnormal is large from the data after the time-frequency analysis for each time in the data of all the times to be inspected. Sequentially extract signal strength data for a range of time periods. In the rope tester, by limiting the data range in the frequency axis direction to the high frequency side, the feature difference between the normal state and the abnormal state becomes large. Furthermore, by limiting the data range to a bilaterally symmetrical data range a little away from the time of interest, the feature difference between the normal state and the abnormal state becomes large.

The model generation unit 5 uses learning data in the range extracted by the feature difference emphasis range extraction unit 4 to learn normal data. Specifically, the model generation unit 5 generates a trained model for inferring the degree of similarity between normal data and normal data when a normal wire rope without disconnection is measured by a rope tester. .

For example, the model generator 5 uses a known algorithm for unsupervised learning. Specifically, the model generation unit 5 learns the features in the learning data from the learning data that does not contain the result (label).

For example, the model generation unit 5 uses One Class Supported Vector Machine (One class SVM), which is unsupervised learning. For example, the model generation unit 5 learns normal data by unsupervised learning according to the grouping method by One Class SVM. Specifically, the model generation unit 5 is an algorithm applied to a one-class classification problem in One class SVM, and detects outliers by learning one class and determining a discrimination boundary.

More specifically, the model generation unit 5 uses kernel tricks to map the learning data xi to cluster "1" and the origin to cluster "-1" in the feature space of the high-dimensional space. After that, the model generator 5 determines a discriminant plane from the support vectors on the boundary between the origin and the learning data.

The storage device 6 stores the learned model output from the model generation unit 5.

Next, the time-series data will be explained using FIG. 2 is a diagram showing time-series data acquired by the learning device according to Embodiment 1. FIG.

For example, the time-series data is obtained by measuring the leakage magnetic flux from the wire rope with one coil. For example, the time-series data in the normal state is measured by a plurality of coils connected in series after the leakage magnetic flux from the wire rope is arranged with some distance in the longitudinal direction of the wire rope.

For example, when two coils Co1 and Co2 are used as shown in FIG. 2, a voltage waveform Vw with two spikes is obtained when the broken portion Bs of the wire rope Wr passes near the coils Co1 and Co2. As a result, the broken portion Bs of the wire rope Wr is emphasized to obtain time-series data.

Next, wavelet transform will be explained using FIGS. 3 and 4. FIG. 3 and 4 are diagrams showing mother wavelet functions used for wavelet transform by the learning device according to Embodiment 1. FIG.

For example, as shown in FIG. 3, in the learning device 1, the time-frequency analysis unit 3 uses wavelet transform using a general mother wavelet function such as Biorthogonal 3.3. For example, as shown in FIG. 4, in the learning device 1, the time-frequency analysis unit 3 uses wavelet transform using a mother wavelet function similar to the voltage waveform Vw when the wire rope Wr is disconnected.

Next, using FIGS. 5 to 7, a first example and a second example of the feature difference emphasis range will be described. FIG. 5 shows a first example of the feature difference enhancement range of the rope tester using the learning device according to the first embodiment. FIG. 6 shows a second example of the feature difference enhancement range of the rope tester using the learning device according to the first embodiment. FIG. 7 shows an example of the feature difference emphasis range when there is a break in the rope in the rope tester using the learning device according to the first embodiment.

As shown in FIG. 5 or FIG. 6, in the learning device 1, the feature difference enhancement range extracting unit 4 extracts a specific frequency band in which the feature difference between the normal state and the abnormal state becomes large and in the positive direction around the time of interest Tf. Signal strength data of specific time ranges Rp1, Rp2, Rm1, Rm2 symmetrical in the negative direction are extracted as feature difference emphasis ranges. As shown in FIG. 7, this range corresponds to ranges Rp2 and Rm1 where the signal strength is higher than the surrounding area and ranges Rp1 and Rm2 where the signal strength is lower than the surrounding area when the wire rope is disconnected.

Note that the feature difference emphasis range extraction unit 4 may process using the numerical values inside the ranges Rp1, Rp2, Rm1, and Rm2. For example, the feature difference emphasis range extracting unit 4 may take the difference between data symmetrical with respect to the time of interest Tf within the ranges Rp1, Rp2, Rm1 and Rm2 shown in FIG. 5 or 6 . For example, the difference between all data values within the ranges Rp1, Rp2, Rm1, and Rm2 may be taken. For example, within the ranges Rp1, Rp2, Rm1, and Rm2, differences in values may be obtained at regular intervals. For example, the difference between the average values of the data ranges Rp1, Rp2, Rm1, and Rm2 of data symmetrical with respect to the time of interest Tf may be taken.

Next, using FIG. 8, a third example of the feature difference emphasis range will be described. FIG. 8 shows a third example of the feature difference enhancement range of the rope tester using the learning device according to the first embodiment.

As shown in FIG. 8, the data after the time-frequency analysis shows the strength and weakness of the signal like a pattern Pt in which a number of mountains overlap. For this reason, in the feature difference emphasis range extracting unit 4, the signal strengths of the ranges Rf and Rb of a specific frequency band and a specific time band before and after the pattern Pt from the overlapping portion Op of the two peaks closest to the target time Tf are calculated. You may take the difference with the average value of data.

In addition, by comparing the data of the normal part of the wire rope and the data of the abnormal part and searching for the part where the feature difference is large, it is possible to extract the range where the feature difference between normal and abnormal is large. good. For example, a range in which the difference between the data of the normal part of the wire rope and the data of the broken part of the wire rope is large may be extracted. For example, the correlation coefficient between the data of the normal part of the wire rope and the data of the disconnected part may be obtained, and the range with low correlation may be extracted.

Next, the learning process of the learning device 1 will be explained using FIG. FIG. 9 is a flowchart for explaining learning processing of the learning device according to Embodiment 1. FIG.

In step S11, the data acquisition unit 2 acquires time-series data of the normal state when a normal wire rope without disconnection is measured with a rope tester.

After that, in step S12, the time-frequency analysis unit 3 performs time-frequency analysis on the normal time-series data.

After that, in step S13, the feature difference enhancement range extracting unit 4 extracts signal strength data in a range of a specific frequency band and a specific time zone in which the feature difference between normal and abnormal conditions is large from the data after the time-frequency analysis. Extract.

After that, in step S14, the model generation unit 5 uses the range data extracted by the feature difference emphasis range extraction unit 4 to generate a learned model.

In step S15, the model generation unit 5 stores the learned model in the storage device 6.

Next, the inference device 7 will be explained using FIG. 10 is a configuration diagram of an inference apparatus according to Embodiment 1. FIG.

The inference device 7 includes a data acquisition unit 8, a time-frequency analysis unit 9, a feature difference emphasis range extraction unit 10, and an inference unit 11.

The data acquisition unit 8 acquires a time series of voltage signals generated when the magnetic flux (leakage magnetic flux) leaking from the surface of the wire rope (leakage magnetic flux) passes through the coil when the wire rope, which is the object to be inspected, is magnetized and sent out in the longitudinal direction. Get data. However, the method of data acquisition must be the same as during learning.

The time-frequency analysis unit 9 performs time-frequency analysis for analyzing temporal changes in frequency characteristics for time-series data. However, the method of time-frequency analysis must be the same as during learning.

The feature difference enhancement range extracting unit 10 extracts signal strength data in a specific frequency band and a specific time range where the feature difference between normal and abnormal conditions is large from the data after the time-frequency analysis. However, it is necessary to match the range to be extracted with that at the time of learning.

The inference unit 11 uses the learned model stored in the storage device 6 to infer the degree of similarity with normal data. Specifically, the inference unit 11 inputs the data of the range extracted by the feature difference emphasis range extraction unit to the learned model, thereby determining whether the data of the object to be inspected belongs to the normal data. infer. The inference unit 11 outputs an inference result as to whether or not the data of the object to be inspected belongs to the normal data to the determination device 12 as a degree of similarity with the normal data.

It should be noted that the inference unit 11 may output the degree of similarity with normal data based on a learned model acquired from the outside such as another rope tester.

The determination device 12 determines normality or abnormality using the degree of similarity from the inference unit 11 as a score.

Next, the inference processing of the inference device 7 will be explained using FIG. 11 is a flowchart for explaining the inference processing of the inference device according to Embodiment 1. FIG.

In step S21, the data acquisition unit 8 acquires time-series data of the subject.

After that, in step S22, the time-frequency analysis unit 9 performs time-frequency analysis for analyzing temporal changes in the frequency characteristics of the time-series data during measurement of the object to be inspected.

After that, in step S23, the feature difference enhancement range extracting unit 10 extracts signal strength data in a range of a specific frequency band and a specific time range in which the feature difference between normal and abnormal conditions is large, from the data after the time-frequency analysis. Extract.

Thereafter, in step S24, the inference unit 11 inputs the data of the object to be inspected output from the feature difference emphasis range extraction unit 10 to the learned model stored in the storage device 6, and calculates the degree of similarity with the normal data. infer.

After that, in step S25, the inference unit 11 outputs to the determination device 12 the degree of similarity with the normal data obtained from the learned model.

After that, the determination device 12 uses the degree of similarity to identify the disconnection point in the wire rope to be inspected. When the disconnection point in the wire rope is identified, the determination device 12 notifies the operator by display and buzzer sound. At this time, the determination device 12 may store information that associates the wire rope feeding amount with the degree of similarity to the normal data. Based on the information, a disconnection point of the wire rope may be detected.

According to the first embodiment described above, the learning device 1 uses the signal strength data in the range of the specific frequency band and the specific time band in which the feature difference between normal and abnormal conditions is large, Generate a trained model for inferring the similarity with the data of For this reason, the characteristic difference between normal and abnormal conditions becomes large. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

In addition, the learning device 1 uses wavelet transform as a method of analyzing temporal changes in frequency characteristics. For this reason, the characteristic difference between normal and abnormal conditions becomes large. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

Also, in the wavelet transform, the learning device 1 uses a waveform similar to the waveform at the time of abnormality as a mother wavelet. For this reason, the characteristic difference between normal and abnormal conditions becomes large. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

In addition, the learning device 1 extracts signal intensity data in a range of a specific frequency band and a specific time band, which have a low correlation with the data of the abnormal part with respect to the data of the normal part, as the feature difference emphasis range. For this reason, the characteristic difference between normal and abnormal conditions becomes large. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

Also, the learning device 1 generates a trained model using the One class SVM. Therefore, even if the distribution of normal features is non-linear, it is possible to appropriately obtain the degree of similarity with the normal time. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

It should be noted that when unsupervised learning is realized in the learning device 1, other known methods capable of clustering may be applied. For example, a kernel density estimation method, an MT method, or the like may be applied. By applying an appropriate method according to the distribution of normal features, the degree of similarity or dissimilarity with the normal state can be obtained appropriately.

In addition, the inference device 7 uses a trained model for inferring the degree of similarity with the data in the normal state, and the degree of similarity between the data in the range where the feature difference between the normal state and the abnormal state becomes large and the data in the normal state. to infer For this reason, the difference in similarity between the normal state and the abnormal state increases. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

In addition, the inference device 7 uses the normalization parameters stored in the storage device 6 to convert into dissimilarity and normalize. This normalization enables uniform processing in the post-processing. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

In addition, the inference device 7 uses wavelet transform as a method of analyzing temporal changes in frequency characteristics. As a result, the amount of information about the features increases, and the feature differences between the normal state and the abnormal state partially increase. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

In wavelet transform, the inference device 7 may use a general mother wavelet function such as Biorthogonal 3.3, or may use a waveform similar to the waveform at the time of abnormality as the mother wavelet function. For this reason, the feature difference between the normal state and the abnormal state partially increases. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

In addition, the inference device 7 infers the degree of similarity with normal data using a trained model using the One class SVM. For this reason, the difference in similarity between the normal state and the abnormal state increases. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

In addition, in the inference device 7, the degree of similarity to the normal data may be inferred using a trained model by the kernel density estimation method, the MT method, or the like. Also in this case, the difference in the similarity between the normal state and the abnormal state becomes large. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

Also, as a learning algorithm, an algorithm such as reinforcement learning, supervised learning, or semi-supervised learning may be applied.

Also, as a learning algorithm, a deep learning algorithm that learns to extract the feature quantity itself may be applied.

Also, the learning device 1 and the reasoning device 7 may be connected to the rope tester via a network separately from the rope tester. The learning device 1 and the reasoning device 7 may be built in the rope tester. The learning device 1 and the reasoning device 7 may exist on a cloud server.

In addition, the model generation unit 5 may learn normal data according to learning data created for a plurality of rope testers. The model generator 5 may acquire learning data from a plurality of rope testers used in the same area. The model generator 5 may acquire learning data collected from a plurality of rope testers operating independently in different areas.

Also, a rope tester that collects learning data may be added or removed from the target on the way.

The learning device 1 that has learned the normal data for the rope tester may be applied to another rope tester, and the normal data for the other rope tester may be re-learned and updated.

Next, an example of the learning device 1 will be described using FIG. FIG. 12 is a hardware configuration diagram of the learning device according to the first embodiment.

Each function of the learning device 1 can be realized by a processing circuit. For example, the processing circuitry comprises at least one processor 100a and at least one memory 100b. For example, the processing circuitry comprises at least one piece of dedicated hardware 200 .

When the processing circuit includes at least one processor 100a and at least one memory 100b, each function of the learning device 1 is realized by software, firmware, or a combination of software and firmware. At least one of software and firmware is written as a program. At least one of software and firmware is stored in at least one memory 100b. At least one processor 100a implements each function of learning device 1 by reading and executing a program stored in at least one memory 100b. The at least one processor 100a is also referred to as a central processing unit, processing unit, arithmetic unit, microprocessor, microcomputer, DSP. For example, the at least one memory 100b is a nonvolatile or volatile semiconductor memory such as RAM, ROM, flash memory, EPROM, EEPROM, magnetic disk, flexible disk, optical disk, compact disk, mini disk, DVD, or the like.

Where the processing circuitry comprises at least one piece of dedicated hardware 200, the processing circuitry may be implemented, for example, in single circuits, multiple circuits, programmed processors, parallel programmed processors, ASICs, FPGAs, or combinations thereof. be. For example, each function of the learning device 1 is implemented by a processing circuit. For example, each function of the learning device 1 is collectively realized by a processing circuit.

A part of each function of the learning device 1 may be realized by dedicated hardware 200, and the other part may be realized by software or firmware. For example, the function of the model generation unit 5 is implemented by a processing circuit as dedicated hardware 200, and the functions other than the function of the model generation unit 5 are implemented by at least one processor 100a and a program stored in at least one memory 100b. may be implemented by reading and executing

Thus, the processing circuit implements each function of the learning device 1 with hardware 200, software, firmware, or a combination thereof.

Although not shown, each function of the inference device 7 can be implemented by a processing circuit equivalent to the processing circuit that implements each function of the learning device 1.

Embodiment 2.
FIG. 13 is a configuration diagram of a learning device according to the second embodiment. The same reference numerals are given to the same or corresponding parts as those of the first embodiment. Description of this part is omitted.

The learning device 1 of the second embodiment is a device in which a normalization parameter derivation unit 13 is added to the learning device 1 of the first embodiment.

The normalization parameter derivation unit 13 inputs part or all of the learning data to the trained model generated by the model generation unit 5, and derives normalization parameters.

For example, the normalization parameter is the maximum score derived using the One class SVM classifier.

The storage device 6 stores the normalization parameters output from the normalization parameter derivation unit 13 .

Next, the learning process of the learning device 1 will be described using FIG. FIG. 14 is a flow chart for explaining learning processing of the learning device according to the second embodiment.

Steps S31 to S34 are the same as steps S11 to S14 in FIG.

After step S34, in step S35, the normalization parameter derivation unit 13 inputs part or all of the learning data to the trained model generated by the model generation unit 5, and derives normalization parameters.

After that, in step S36, the model generation unit 5 stores the learned model in the storage device 6. The normalization parameter derivation unit 13 causes the storage device 6 to store the normalization parameter. At this time, the trained model and the normalized parameter are stored in association with each other.

Next, the inference device 7 will be explained using FIG. FIG. 15 is a configuration diagram of an inference device according to Embodiment 2. FIG.

The inference device 7 of the second embodiment is a device in which a dissimilarity derivation unit 14 is added to the inference device 7 of the first embodiment.

The dissimilarity deriving unit 14 converts the similarity between the normal data derived by the inference unit 11 into a dissimilarity using the normalization parameter stored in the storage device 6 and normalizes it. . The dissimilarity derivation unit 14 outputs the normalized dissimilarity to the determination device 12 .

The determination device 12 determines normality or abnormality using the normalized dissimilarity from the dissimilarity derivation unit 14 as a score.

Next, the inference processing of the inference device 7 will be explained using FIG. 16 is a flowchart for explaining the inference processing of the inference device according to Embodiment 2. FIG.

Steps S41 to S44 are the same as steps S21 to S24 in FIG.

After step S44, in step S45, the non-similarity derivation unit 14 uses the normalization parameter stored in the storage device 6 to calculate the similarity between the normal data derived by the inference unit 11 and the non-similarity derivation unit 14. Convert to similarity and normalize.

After that, in step S46, the dissimilarity derivation unit 14 outputs the normalized dissimilarity to the determination device 12.

According to the second embodiment described above, the learning device 1 inputs part or all of the learning data to the trained model and derives the normalization parameter.

In addition, the inference device 7 uses the normalization parameters stored in the storage device 6 to convert into dissimilarity and normalize. Therefore, the difference in dissimilarity between normal and abnormal cases increases. As a result, anomalies can be accurately detected even from data acquired under poor conditions.

Embodiment 3.
17 is a configuration diagram of a learning device according to Embodiment 3. FIG. The learning device 1 of FIG. 17 is used to detect an insulation abnormality of a circuit breaker. For example, the learning device 1 is used to detect insulation faults in circuit breakers using a partial discharge tester.

The learning device 1 includes a data acquisition unit 2, a time-frequency analysis unit 3, a feature difference enhancement range extraction unit 4, a statistic calculation unit 15, a model generation unit 5, a storage device 6, and a normalization parameter derivation unit. 13 and. The learning device 1 according to the third embodiment is a device obtained by adding a statistic calculation unit 15 to the learning device 1 according to the second embodiment. In this embodiment, as in

Embodiments

1 and 2, the feature difference enhancement range extraction unit 4 extracts a specific frequency band and a specific to extract signal strength data for a range of time periods.

The statistic calculation unit 15 calculates the minimum value, maximum value, median value, mode value, average value, Calculate statistics such as variance and standard deviation.

The model generation unit 5 generates a trained model from the statistics output from the statistics calculation unit 15 using unsupervised learning such as One Class SVM or kernel density estimation method.

Next, the learning process of the learning device 1 will be explained using FIG. FIG. 18 is a flow chart for explaining the learning process of the learning device 1 according to the third embodiment.

Steps S51 to S53 are the same as steps S31 to S33 in FIG. Steps S55 to S57 are the same as steps S34 to S36 in FIG.

In step S53, as in the first and second embodiments, the feature difference enhancement range extracting unit 4 extracts data from the data after the time-frequency analysis in a specific frequency band and at a specific time in which the feature difference between the normal state and the abnormal state increases. Extract signal strength data for a range of bands. After that, in step S54, the statistic calculation unit 15 calculates the minimum value, maximum value, median value, maximum Calculate statistics such as frequency, mean, variance and standard deviation.

After that, in step S55, the model generation unit 5 generates a trained model from the statistics output from the statistics calculation unit 15 using unsupervised learning such as One Class SVM or kernel density estimation method.

Next, the inference device 7 will be explained using FIG. FIG. 19 is a configuration diagram of an inference device according to Embodiment 3. FIG.

The inference device 7 of the third embodiment is a device in which a statistic calculation unit 16 is added to the inference device 7 of the second embodiment. In the present embodiment, as in

Embodiments

1 and 2, feature difference enhancement range extraction section 10 extracts a specific frequency band and Extract signal strength data for a specific time range.

The statistic calculation unit 16 calculates the minimum value, maximum value, median value, mode value, average value, Calculate statistics such as variance and standard deviation. The inference unit 11 uses the learned model stored in the storage device 6 to infer the degree of similarity between the statistic output from the statistic calculation unit 16 and the normal data.

Next, the inference processing of the inference device 7 will be explained using FIG. FIG. 20 is a flowchart for explaining the inference processing of the inference device 7 according to the third embodiment.

Steps S61 to S63 are the same as steps S41 to S43 in FIG. Steps S65 to S67 are the same as steps S44 to S46 in FIG.

After step S63, in step S64, the statistic calculation unit 16 calculates the minimum value, maximum value, median value, mode value, average value, variance, standard deviation, etc. of the data extracted by the feature difference emphasis range extraction unit 10. Calculate the statistic of

After that, in step S65, the inference unit 11 uses the learned model stored in the storage device 6 to infer the degree of similarity to normal data from the statistic output from the statistic calculation unit 16.

According to the third embodiment described above, the learning device 1 uses the statistics of the signal strength data in the range of the specific frequency band and the specific time band where the feature difference between the normal state and the abnormal state is large. to generate

In addition, the inference device 7 uses a learned model from statistics of signal strength data in a range of a specific frequency band and a specific time period where the feature difference between normal and abnormal conditions is large, and compares the data with the data during normal conditions. Infer similarity. As a result, the difference in dissimilarity between the normal state and the abnormal state becomes large, and the number of data to be input to the model generation unit and the inference unit is reduced. As a result, anomalies can be accurately detected in a short time even from data acquired under poor conditions.

In Embodiment 1, Embodiment 2, or Embodiment 3, an output signal from a sensor that senses sound, vibration, pressure, temperature, magnetic flux, light, etc. is sampled by an A/D converter and quantized. Then you can process the data.

1 learning device, 2 data acquisition unit, 3 time-frequency analysis unit, 4 feature difference enhancement range extraction unit, 5 model generation unit, 6 storage device, 7 reasoning device, 8 data acquisition unit, 9 time-frequency analysis unit, 10 feature difference Emphasis range extraction unit, 11 reasoning unit, 12 determination device, 13 normalization parameter derivation unit, 14 dissimilarity derivation unit, 15 statistics calculation unit, 16 statistics calculation unit

Claims

a data acquisition unit that acquires normal time-series data;
a time-frequency analysis unit that analyzes temporal changes in frequency characteristics of the normal time-series data acquired by the data acquisition unit;
a feature difference enhancement range extracting unit for extracting signal intensity data in a range of a specific frequency band and a specific time band in which the feature difference between normal and abnormal conditions is large, from the data analyzed by the time-frequency analysis unit; ,
A model for generating a trained model for inferring the degree of similarity to normal data using signal intensity data in a specific frequency band and a specific time range extracted by the feature difference emphasis range extracting unit. a generator;
A learning device with
a normalization parameter derivation unit that inputs a part or all of the learning data to the trained model generated by the model generation unit and derives a normalization parameter;
The learning device according to claim 1, comprising:
a data acquisition unit that acquires normal time-series data;
a time-frequency analysis unit that analyzes temporal changes in frequency characteristics of the normal time-series data acquired by the data acquisition unit;
a feature difference enhancement range extracting unit for extracting signal intensity data in a range of a specific frequency band and a specific time band in which the feature difference between normal and abnormal conditions is large, from the data analyzed by the time-frequency analysis unit; ,
a statistic calculation unit that calculates a statistic from the signal intensity data in the specific frequency band and the specific time range extracted by the feature difference emphasis range extraction unit;
a model generation unit that generates a trained model for inferring the degree of similarity between the statistics calculated by the statistics calculation unit and the normal data;
a normalization parameter derivation unit that inputs a part or all of the learning data to the trained model generated by the model generation unit and derives a normalization parameter;
A learning device with
The feature difference enhancement range extracting unit extracts the data analyzed by the time-frequency analysis unit in a positive direction and a negative direction centering on a specific frequency band where the feature difference between the normal state and the abnormal state increases and the time of interest. 4. The learning device according to any one of claims 1 to 3, which extracts signal strength data in a symmetrical specific time range.
The feature difference enhancement range extracting unit extracts the data analyzed by the time-frequency analysis unit in a positive direction and a negative direction centering on a specific frequency band where the feature difference between the normal state and the abnormal state increases and the time of interest. 4. The learning device according to any one of claims 1 to 3, which extracts a difference in signal strength data within a specific symmetric time zone.
The feature difference enhancement range extracting unit extracts signal strength data in a range of a specific frequency band and a specific time band in which the feature difference between normal and abnormal conditions is large, from the data analyzed by the time-frequency analysis unit. 4. The learning device according to any one of claims 1 to 3, wherein a difference from an average value of signal intensity data in a surrounding range is extracted.
The learning device according to any one of claims 1 to 6, wherein the time-frequency analysis unit uses wavelet transform as a method of analyzing temporal changes in frequency characteristics.
The learning device according to claim 7, wherein the time-frequency analysis unit uses a waveform similar to the waveform at the time of abnormality as a mother wavelet function in the wavelet transform.
The learning device according to any one of claims 1 to 8, wherein the feature difference emphasis range extracting unit extracts a range having a low correlation between the data of the normal section and the data of the abnormal section as the range of feature difference emphasis. .
The learning device according to any one of claims 1 to 9, wherein the model generation unit generates a trained model using One class SVM.
The learning device according to any one of claims 1 to 9, wherein the model generation unit generates a trained model using a kernel density estimation method.
a data acquisition unit that acquires time-series data of an object to be inspected;
a time-frequency analysis unit that analyzes temporal changes in frequency characteristics of the time-series data of the object to be inspected acquired by the data acquisition unit;
a feature difference enhancement range extracting unit for extracting signal intensity data in a range of a specific frequency band and a specific time band in which the feature difference between normal and abnormal conditions is large, from the data analyzed by the time-frequency analysis unit; ,
Signal strength data in a specific frequency band and a specific time range extracted by the feature difference emphasis range extracting unit using a trained model for inferring similarity with normal data and normal time an inference unit that infers similarity with data;
A reasoning device with
a dissimilarity deriving unit that converts the similarity inferred by the inference unit into a dissimilarity and normalizes it using a normalization parameter stored in a storage device;
13. A reasoning apparatus according to claim 12, comprising:
a data acquisition unit that acquires time-series data of an object to be inspected;
a time-frequency analysis unit that analyzes temporal changes in frequency characteristics of the time-series data of the object to be inspected acquired by the data acquisition unit;
a feature difference enhancement range extracting unit for extracting signal intensity data in a range of a specific frequency band and a specific time band in which the feature difference between normal and abnormal conditions is large, from the data analyzed by the time-frequency analysis unit; ,
a statistic calculation unit that calculates a statistic from the signal intensity data in the specific frequency band and the specific time range extracted by the feature difference emphasis range extraction unit;
an inference unit that infers the degree of similarity between the statistic calculated by the statistic calculation unit and the data in the normal state using a trained model for inferring the degree of similarity with the data in the normal state;
a dissimilarity derivation unit that converts the similarity inferred by the inference unit into a dissimilarity and normalizes it using a normalization parameter stored in a storage device;
A reasoning device with
The feature difference enhancement range extracting unit extracts the data analyzed by the time-frequency analysis unit in a positive direction and a negative direction centering on a specific frequency band where the feature difference between the normal state and the abnormal state increases and the time of interest. 15. A reasoning apparatus according to any one of claims 12 to 14 for extracting signal strength data for a symmetrical specific time range.
The feature difference enhancement range extracting unit extracts the data analyzed by the time-frequency analysis unit in a positive direction and a negative direction centering on a specific frequency band where the feature difference between the normal state and the abnormal state increases and the time of interest. 15. The reasoning apparatus according to any one of claims 12 to 14, wherein the difference of signal strength data in a specific symmetrical range of time is extracted.
As the feature difference enhancement range extraction unit, from the data analyzed by the time-frequency analysis unit, from the signal strength data in the range of the specific frequency band and the specific time band where the feature difference between normal and abnormal is large 15. The reasoning apparatus according to claim 12, wherein a difference from an average value of signal strength data in a surrounding range is extracted.
The inference device according to any one of claims 12 to 17, wherein the time-frequency analysis unit uses wavelet transform as a method of analyzing temporal changes in frequency characteristics.
19. The inference device according to claim 18, wherein the time-frequency analysis unit uses a waveform similar to an abnormal waveform as a mother wavelet function in the wavelet transform.
20. The inference device according to any one of claims 12 to 19, wherein the inference unit infers the degree of similarity with normal data using a trained model using One class SVM.
The inference device according to any one of claims 12 to 19, wherein the inference unit infers the degree of similarity with normal data using a trained model using a kernel density estimation method.
A data acquisition step of acquiring normal time-series data using a learning device;
a time-frequency analysis step of using the learning device to analyze temporal changes in frequency characteristics of the normal time-series data acquired in the data acquisition step;
Using the learning device, from the data analyzed by the time-frequency analysis step, signal strength data in a specific frequency band and a specific time range where the feature difference between normal and abnormal is large is extracted. a feature difference enhancement range extraction step;
Using the learning device, using the signal intensity data of the specific frequency band and the specific time range extracted by the feature difference emphasis range extraction step, for inferring the similarity with normal data a model generation step for generating a trained model;
A learning method with
A normalization parameter derivation step of inputting part or all of the learning data to the trained model generated in the model generation step using the learning device and deriving a normalization parameter;
23. The learning method of claim 22, comprising:
A data acquisition step of acquiring normal time-series data using a learning device;
a time-frequency analysis step of using the learning device to analyze temporal changes in frequency characteristics of the normal time-series data acquired in the data acquisition step;
Using the learning device, from the data analyzed by the time-frequency analysis step, signal strength data in a specific frequency band and a specific time range where the feature difference between normal and abnormal is large is extracted. a feature difference enhancement range extraction step;
a statistic calculation step of calculating, using the learning device, the statistic of the signal strength data in the specific frequency band and the specific time range extracted in the feature difference emphasis range extraction step;
A model generation step of generating a trained model for inferring the similarity to normal data using the statistic calculated by the statistic calculation step using the learning device;
a normalization parameter derivation step of inputting part or all of the learning data to the trained model generated in the model generation step using the learning device and deriving a normalization parameter;
A learning method with
In the feature difference enhancement range extraction step, from the data analyzed in the time-frequency analysis step, a specific frequency band in which the feature difference between normal and abnormal becomes large and in the positive and negative directions centering on the time of interest. 25. A learning method according to any one of claims 22 to 24, wherein the signal strength data for a range of symmetrical specific time periods is extracted.
In the feature difference enhancement range extraction step, from the data analyzed in the time-frequency analysis step, a specific frequency band in which the feature difference between normal and abnormal becomes large and in the positive and negative directions centering on the time of interest. 25. The learning method according to any one of claims 22 to 24, wherein a difference of signal strength data in a symmetrical specific time range is extracted.
As the feature difference enhancement range extraction step, from the data analyzed in the time-frequency analysis step, from signal strength data in a specific frequency band and a specific time range where the feature difference between normal and abnormal is large. 25. The learning method according to any one of claims 22 to 24, wherein a difference from an average value of signal intensity data in the surrounding range is extracted.
28. The learning method according to any one of claims 22 to 27, wherein the time-frequency analysis step uses wavelet transform as a method of analyzing temporal changes in frequency characteristics.
29. The learning method according to claim 28, wherein the time-frequency analysis step uses a waveform similar to a waveform at the time of abnormality as a mother wavelet function in the wavelet transform.
30. The learning method according to any one of claims 22 to 29, wherein the feature difference-enhanced range extracting step extracts, as a feature difference-enhanced range, a range having a low correlation between normal-area data and abnormal-area data. .
31. The learning method according to any one of claims 22 to 30, wherein the model generation step generates a trained model using One class SVM.
The learning method according to any one of claims 22 to 30, wherein the model generation step uses a kernel density estimation method to generate a learned model.
a data acquisition step of acquiring time-series data of an object to be inspected using an inference device;
a time-frequency analysis step of using the inference device to analyze temporal changes in frequency characteristics of the time-series data of the subject acquired in the data acquisition step;
Using the inference device, from the data analyzed by the time-frequency analysis step, signal strength data in a specific frequency band and a specific time range where the feature difference between normal and abnormal conditions is large is extracted. a feature difference enhancement range extraction step;
Using the inference device, using a trained model for inferring the degree of similarity with normal data, the range of the specific frequency band and the specific time zone extracted by the feature difference emphasis range extraction step an inference step of inferring similarities between the signal strength data and normal data;
An inference method with
A dissimilarity derivation step of converting the similarity inferred by the inference step into a dissimilarity and normalizing it using a normalization parameter stored in a storage device;
34. The reasoning method of claim 33, comprising:
a data acquisition step of acquiring time-series data of an object to be inspected using an inference device;
a time-frequency analysis step of using the inference device to analyze temporal changes in frequency characteristics of the time-series data of the subject acquired in the data acquisition step;
Using the inference device, from the data analyzed by the time-frequency analysis step, signal strength data in a specific frequency band and a specific time range where the feature difference between normal and abnormal conditions is large is extracted. a feature difference enhancement range extraction step;
a statistic calculation step of calculating a statistic of data in the range extracted in the feature difference emphasized range extraction step using the inference device;
Using the inference device, using a learned model for inferring the similarity with normal data, using the statistic calculated by the statistic calculation step, the similarity with normal data an inference step of inferring
a dissimilarity deriving step of converting the similarity inferred by the inference step using the inference device into a dissimilarity and normalizing it using a normalization parameter stored in a storage device;
An inference method with
The feature difference enhancement range extracting step extracts from the data analyzed in the time-frequency analysis step, a specific frequency band in which the feature difference between normal and abnormal is large, and in the positive and negative directions centering on the time of interest. 36. A reasoning method according to any one of claims 33 to 35, wherein signal strength data for a range of symmetrical specific time periods are extracted.
The feature difference enhancement range extracting step extracts from the data analyzed in the time-frequency analysis step, a specific frequency band in which the feature difference between normal and abnormal is large, and in the positive and negative directions centering on the time of interest. 36. A reasoning method according to any one of claims 33 to 35, wherein the difference of signal strength data for a range of symmetrical specific time periods is extracted.
The feature difference enhancement range extracting step extracts from the data analyzed in the time-frequency analysis step signal strength data in a range of a specific frequency band and a specific time range in which the feature difference between normal and abnormal conditions is large. 36. A reasoning method according to any one of claims 33 to 35, wherein a difference from an average value of signal strength data in a surrounding range is extracted.
39. The inference method according to any one of claims 33 to 38, wherein the time-frequency analysis step uses wavelet transform as a method of analyzing temporal changes in frequency characteristics.
40. The inference method according to claim 39, wherein the time-frequency analysis step uses a waveform similar to a waveform at the time of abnormality as a mother wavelet function in the wavelet transform.
41. The inference method according to any one of claims 33 to 40, wherein the inference step infers the degree of similarity with normal data using a trained model using One class SVM.
41. The inference method according to any one of claims 33 to 40, wherein the inference step infers the degree of similarity with normal data using a trained model using a kernel density estimation method.