WO2022234635A1

WO2022234635A1 - Data analysis device, data analysis method, and recording medium

Info

Publication number: WO2022234635A1
Application number: PCT/JP2021/017458
Authority: WO
Inventors: 裕子太田; 咲子美島
Original assignee: 日本電気株式会社
Priority date: 2021-05-07
Filing date: 2021-05-07
Publication date: 2022-11-10
Also published as: JPWO2022234635A1

Abstract

This invention accurately identifies abnormal sound from time-series data having various characteristics. A determination unit (11) determines a characteristic of time-series data. A selection unit (12) selects a method for analyzing the time-series data on the basis of the characteristic of the time-series data. An identification unit (13) identifies an abnormal sound included in the time-series data by using the selected method to analyze the time-series data.

Description

Data analysis device, data analysis method, and recording medium

The present invention relates to a data analysis device, data analysis method, and recording medium, and more particularly to a data analysis device, data analysis method, and recording medium for analyzing time-series data.

Research is being conducted on technologies for monitoring the operating status of equipment and parts in railway vehicles, automobile engine compartments, and factories. For example, in the related technology described in Patent Literature 1, a learning model generated by machine learning is used to diagnose the operating state of a machine in real time based on physical quantities detected by sensors.

Furthermore, in another example of related technology, non-negative matrix factorization (NMF) is used to analyze time series data. Specifically, related techniques transform time series data into an amplitude spectrogram and decompose the spectrogram into a basis matrix and an activation matrix. Then, by using the activation matrix as an acoustic feature quantity, abnormal sounds contained in the time-series data are identified.

Japanese Patent Application Laid-Open No. 2020-204937

NMF approximates a non-negative matrix, which is a spectrogram expression, by the product of lower-dimensional non-negative matrices. Therefore, when the period of peaks in the time-series data is not stable, or in a noisy environment, the accuracy of identifying abnormal sounds in the related technique using NMF decreases.

The present invention has been made in view of the above problems, and its purpose is to accurately identify abnormal sounds from time-series data of various properties.

A data analysis apparatus according to an aspect of the present invention includes determination means for determining properties of time-series data, and selection means for selecting a method for analyzing the time-series data based on the properties of the time-series data. and identification means for identifying abnormal sounds contained in the time-series data by analyzing the time-series data using the selected technique.

In a data analysis method according to one aspect of the present invention, properties of time-series data are determined, a method for analyzing the time-series data is selected based on the properties of the time-series data, and the selected method is selected. to identify abnormal sounds contained in the time-series data by analyzing the time-series data.

A recording medium according to an aspect of the present invention comprises determining properties of time-series data, selecting a technique for analyzing the time-series data based on the properties of the time-series data, and selecting A program is stored for causing a computer to analyze the time-series data using the method to identify abnormal sounds contained in the time-series data.

According to one aspect of the present invention, abnormal sounds can be accurately identified from time-series data of various properties.

1 is a block diagram showing the configuration of a data analysis device according to Embodiment 1; FIG. 4 is a flowchart for explaining the operation of the data analysis device according to Embodiment 1; 2 is a block diagram showing the configuration of a data analysis device according to Embodiment 2; FIG. FIG. 10 is a diagram showing an example of time-series data to be analyzed by the data analysis device according to the second embodiment; 9 is a flowchart for explaining the operation of the data analysis device according to Embodiment 2; FIG. 11 is a block diagram showing the configuration of a data analysis device according to Embodiment 3; It is a figure which shows an example of the spectrogram converted from time-series data. FIG. 4 is a graph showing an example of a frequency spectrum transformed from a time-width segment of time-series data; FIG. 10 is a flowchart for explaining the operation of the data analysis device according to Embodiment 3; FIG. 10 is an example of a graph showing the distribution of scores used to determine the threshold of peak intensity of frequency spectrum; FIG. 3 is a diagram showing an example of a hardware configuration of a data analysis device according to any one of Embodiments 1-3; FIG.

Several forms for carrying out the present invention will be described below.

[Embodiment 1]
Embodiment 1 will be described with reference to FIGS. 1 and 2. FIG.

(Abnormal sound identification device 10)
FIG. 1 is a block diagram showing the configuration of an abnormal noise identification device 10 according to the first embodiment. As shown in FIG. 1 , the abnormal noise identification device 10 includes a determination section 11 , a selection section 12 , an identification section 13 and a provision section 14 .

The determination unit 11 determines the nature of time-series data. The determination unit 11 is an example of determination means.

In the first example, the determination unit 11 tracks peaks in time-series data. Here, the determination unit 11 can use well-known peak tracking technology. The determination unit 11 measures the time from the first peak of the time-series data to the next peak. Subsequently, the determination unit 11 calculates the time from the second peak to the third peak. The determination unit 11 repeatedly calculates the time width (called period) between adjacent peaks in the time-series data. After that, the determination unit 11 calculates fluctuations in the period of peaks in the time-series data. For example, the determining unit 11 calculates the variance or standard deviation of the peak period in the time-series data as an index indicating the magnitude of fluctuation in the peak period in the time-series data. Then, when the magnitude of the period fluctuation of the peak in the time-series data is equal to or less than the threshold, the determination unit 11 determines that the time-series data has the property a. On the other hand, when the magnitude of the period fluctuation of the peak in the time series data exceeds the threshold value, the determination unit 11 determines that the time series data has the property b (Embodiment 1).

In the second example, the determination unit 11 Fourier-transforms the time-series data into a spectrum. The determination unit 11 calculates the peak intensity of the spectrum. Then, when all the peak intensities of the spectrum are equal to or greater than the threshold, the determination unit 11 determines that the time-series data has the property a. On the other hand, when one or more peak intensities of the time-series data are below the threshold, the determination unit 11 determines that the time-series data has property b (second embodiment). Note that the method by which the determination unit 11 determines the nature of the time-series data is not limited to the first and second examples described here.

The determination unit 11 outputs the determination result of the nature of the time-series data to the selection unit 12 . The determination unit 11 also outputs the time-series data to the identification unit 13 .

The selection unit 12 selects a method for analyzing time series data based on the properties of the time series data. The selection unit 12 is an example of selection means.

In one example, the selection unit 12 receives the determination result of the nature of the time-series data from the determination unit 11 . The selection unit 12 selects a technique for analyzing the time series data based on the determination result of the properties of the time series data. For example, when the time-series data has property a, the selection unit 12 selects the first method using nonnegative matrix factorization (NMF). In a technique using non-negative matrix factorization (hereinafter referred to as NMF), a spectrogram obtained by arranging spectra of time-series data in time order is decomposed into a base matrix and an activation matrix. The activation matrix thus obtained is the feature quantity in the first method.

On the other hand, when the time-series data has property b, the selection unit 12 selects the second method using Mel-Frequency Cepstrum Coefficients (MFCC). Mel-frequency cepstrum coefficients (hereinafter MFCC) are weighted cepstrum low-order components obtained by cepstrum analysis of time series data. The MFCC thus obtained is the feature quantity in the second method. The selection unit 12 notifies the identification unit 13 of the method (first method or second method) for analyzing the time-series data.

The identification unit 13 uses the selected method to analyze the time-series data to identify abnormal sounds contained in the time-series data. The identification unit 13 is an example of identification means.

In one example, the identification unit 13 receives time-series data from the determination unit 11 . Further, the identifying unit 13 is notified of the method (either the first method or the second method) for analyzing the time-series data from the selecting unit 12 . The identification unit 13 uses the method selected by the selection unit 12 to analyze the time-series data. For example, when the first technique is selected, the identification unit 13 first converts the time-series data into a spectrogram. Then, the identification unit 13 obtains an activation matrix by decomposing the spectrogram using NMF. The identifying unit 13 inputs the obtained activation matrix as a feature quantity to a classifier (hereinafter referred to as a classifier A) that performs machine learning using the activation matrix as a feature quantity. The discriminator A discriminates time-series data based on the input feature amount of the activation matrix, and outputs the discrimination result.

On the other hand, when the second method is selected, the identification unit 13 first obtains MFCC by performing cepstrum analysis on the time-series data. The identifying unit 13 inputs the MFCC obtained by the cepstrum analysis as a feature amount to a classifier (hereinafter referred to as a classifier B) that performs machine learning using the MFCC as a feature amount. The discriminator B discriminates the time-series data based on the input MFCC feature amount, and outputs the discrimination result. In this manner, the identification unit 13 identifies time-series data using the identification device A or the identification device B according to the method selected by the selection unit 12 . The identification unit 13 may output the time-series data identification result to a subsequent processing unit (not shown), or may provide it to a recording medium or an external device.

(Operation of Abnormal Sound Identification Device 10)
The operation of the abnormal noise identification device 10 according to the first embodiment will be described with reference to FIG. FIG. 2 is a flow chart showing the flow of processing executed by each part of the abnormal noise identification device 10. As shown in FIG.

First, time-series data is input to the abnormal sound identification device 10 . The time-series data is, for example, acoustic signals generated by collecting sounds emitted by equipment or parts with a microphone in train cars in operation, factories, engine rooms of automobiles, or the like. The abnormal sound identification device 10 receives time-series data such as acoustic signals via any wireless or wired network. After that, the abnormal noise identification device 10 starts the following operations.

As shown in FIG. 2, the determination unit 11 determines the nature of time-series data (S1). In one example, the determination unit 11 measures the time width (period) from the peak of the time-series data to the next peak. Then, when the magnitude of period fluctuation is equal to or less than the threshold, the determination unit 11 determines that the time-series data has property a. On the other hand, when the magnitude of period fluctuation exceeds the threshold, the determination unit 11 determines that the time-series data has property b. The determination unit 11 outputs the determination result of the nature of the time-series data to the selection unit 12 . The determination unit 11 also outputs the time-series data to the identification unit 13 .

Next, the selection unit 12 selects a technique for analyzing the time series data based on the properties of the time series data (S2). For example, when the time-series data has property a, the selection unit 12 selects the first technique using NMF. After that, when the time-series data has property b, the selection unit 12 selects the second method using MFCC. The selection unit 12 notifies the identification unit 13 of the method for analyzing the time-series data.

The identification unit 13 uses the method selected by the selection unit 12 to identify abnormal sounds contained in the time series data by analyzing the time series data (S3).

With this, the operation of the abnormal noise identification device 10 according to the first embodiment is completed.

(Effect of this embodiment)
According to the configuration of this embodiment, the determination unit 11 determines the nature of time-series data. The selection unit 12 selects a technique for analyzing time series data based on the properties of the time series data. The identification unit 13 identifies abnormal sounds contained in the time-series data by analyzing the time-series data using the selected technique. Time-series data contain various types of sounds (including allophones) and noise, and the nature of time-series data also varies. For example, the time-series data may include an abnormal sound with large period fluctuations, or the noise may be large (the target sound is small).

The abnormal sound identification device 10 first determines the nature of the time-series data, and selects a method for analyzing the time-series data based on the determination result. This makes it possible to accurately identify abnormal sounds from time-series data of various properties.

[Embodiment 2]
Embodiment 2 will be described with reference to FIGS. 3 to 5. FIG. In the second embodiment, an example of a method for determining properties of time-series data will be described. In Embodiment 2, the description of Embodiment 1 is cited with respect to the configuration described in Embodiment 1, and the description thereof is omitted.

(Abnormal sound identification device 20)
FIG. 3 is a block diagram showing the configuration of the abnormal noise identification device 20 according to the second embodiment. As shown in FIG. 3 , the abnormal noise identification device 20 includes a determination section 21 , a selection section 12 and an identification section 13 . Also, the determination unit 21 of the abnormal noise identification device 20 includes a peak detection unit 24 . The peak detector 24 detects peaks in the time-series data.

An example of a method for determining the properties of time-series data will be specifically described with reference to FIG. FIG. 4 illustrates time-series data with property a and time-series data with property b, respectively. In FIG. 4, the peaks detected by the peak detector 24 are indicated by dots (black circles).

In the second embodiment, the determination unit 21 determines the nature of the time series data based on the time width (called period) from the detection of the peak of the time series data to the detection of the next peak. .

In FIG. 4, the period of peaks in the time series data is represented by the distance between the points indicating the peaks of the time series data (that is, the length of the double arrow). In the time-series data on the upper side, the period of peaks in the time-series data is almost constant. In other words, the time-series data on the upper side has a small period difference (period fluctuation). On the other hand, in the time-series data on the lower side, there are variations in the period of peaks in the time-series data. In other words, the time-series data on the lower side has a large period difference (period fluctuation).

The determination unit 21 compares the magnitude of the difference in peak periods (period fluctuations) in the time-series data with a predetermined threshold. For example, the threshold X is 0.5 when the magnitude of period fluctuation of peaks in time-series data is represented by the deviation of the difference. In this example, when the period fluctuation of the peak in the time-series data is X=0.5 or less, the determination unit 21 determines that the time-series data has the property a. On the other hand, when the period fluctuation of peaks in the time-series data exceeds X=0.5, the determination unit 21 determines that the time-series data has property b.

(Operation of Abnormal Sound Identification Device 20: S1)
The operation of the abnormal noise identifying device 20 according to the second embodiment will be described with reference to FIG. Here, only the details of the flow of processing executed by the determination unit 21, that is, the contents of step S1 shown in FIG. 2 will be described.

The abnormal noise identification device 20 receives time-series data in the same manner as in the first embodiment. After that, the determination unit 21 of the abnormal noise identification device 20 determines the properties of the time-series data as described below.

As shown in FIG. 5, the peak detection unit 24 of the determination unit 21 detects peaks in the time-series data (S21).

The determination unit 21 calculates fluctuations in the period of peaks in the time series data based on the time width between peaks in the time series data (S22).

The determination unit 21 determines whether or not the magnitude of fluctuation in the period of peaks in the time-series data is equal to or less than a threshold (S23).

When the magnitude of the period fluctuation of the peak in the time series data is equal to or less than the threshold (Yes in S23), the determination unit 21 determines that the time series data has property a (S24A). On the other hand, when the magnitude of the period fluctuation of the peak in the time series data exceeds the threshold (No in S23), the determination unit 21 determines that the time series data has property b (S24B).

With this, the processing of the determination unit 21 ends. After that, the process proceeds to the processing (step S2) of the selection unit 12 described in the first embodiment. In the second embodiment, the description of the process after the process (step S2) of the selection unit 12 is omitted.

(Effect of this embodiment)
According to the configuration of this embodiment, the determination unit 21 determines the nature of time-series data. The selection unit 12 selects a technique for analyzing time series data based on the properties of the time series data. The identification unit 13 identifies abnormal sounds contained in the time-series data by analyzing the time-series data using the selected technique. Time-series data contain various types of sounds (including allophones) and noise, and the nature of time-series data also varies. For example, the time-series data may include an abnormal sound with large period fluctuations, or the noise may be large (the target sound is small).

Furthermore, according to the configuration of Embodiment 2, the determination unit 21 includes a peak detection unit 24 that detects peaks in the time-series data. The determination unit 21 determines the nature of the time-series data based on the time width from the detection of the peak of the time-series data to the detection of the next peak. From the time width from the detection of the peak of the time-series data to the detection of the next peak, the magnitude of the period fluctuation of the peak in the time-series data can be calculated. Then, by comparing the magnitude of the period fluctuation of the peaks in the time series data with the threshold value, it is found that the period fluctuations of the peaks in the time series data are relatively small and the period fluctuations of the peaks in the time series data are relatively small. Large properties can be determined.

[Embodiment 3]
Embodiment 3 will be described with reference to FIGS. 6 to 9. FIG. In the third embodiment, another example of the method for determining the properties of time-series data will be described. In Embodiment 3, the description of Embodiment 1 is cited with respect to the configuration described in Embodiment 1, and the description thereof is omitted.

(Abnormal sound identification device 30)
FIG. 6 is a block diagram showing the configuration of the abnormal noise identification device 30 according to the third embodiment. As shown in FIG. 6 , the abnormal noise identification device 30 includes a determination section 31 , a selection section 12 and an identification section 13 . Also, the determination unit 31 of the abnormal noise identification device 30 includes a data conversion unit 34 . The data conversion unit 34 converts time-domain signals such as time-series data and waveforms into frequency-domain signals such as spectra and spectrograms. An example of converting time-series data into a spectrogram will be described below.

Fig. 7 shows an example of a spectrogram converted from time-series data. In the spectrogram shown in FIG. 7, the frequency spectrum intensity is represented by shading. Also, the peak of the frequency spectrum is indicated by a thick line (bar). In the example shown in FIG. 7, part of the peak of the frequency spectrum is tilted with respect to the vertical and horizontal axes. This indicates that the peak frequency is transitioning with time. In other words, the period of peaks (=1/peak frequency) in the original time-series data fluctuates.

FIG. 8 is a graph showing an example of frequency spectrum converted from time-series data. A frequency spectrum corresponds to a given time span in the spectrogram. In FIG. 8, the peaks of the frequency spectrum are indicated by dots (black circles) on the graph.

A sharp and high peak in the frequency spectrum corresponds to the fact that the period of the peak in the original time-series data is almost constant (that is, the fluctuation of the period is small) in the predetermined time width. On the other hand, the fact that the peak of the frequency spectrum is dull and low corresponds to the fact that the period of the peak in the original time-series data varies (that is, the fluctuation of the period is large) in the predetermined time width.

The determination unit 31 determines the nature of the time-series data based on the peak intensity of the frequency spectrum cut out from the spectrogram for each predetermined time width. For example, the determination unit 31 calculates the difference between the peak intensity in the frequency spectrum and the average intensity in a predetermined band around the peak frequency. The determination unit 31 compares the obtained difference and the threshold value Y with each other. In this example, when the difference between the peak intensity in the frequency spectrum and the average intensity in a predetermined band centered on the peak frequency is equal to or greater than the threshold value Y, the determination unit 31 determines that the time-series data has property a. do. On the other hand, when the difference between the peak intensity in the frequency spectrum and the average intensity in the predetermined time span is less than the threshold value Y, the determination unit 31 determines that the time-series data has property b.

By feeding back information about the reliability of the abnormal noise identification result obtained by the identification unit 13 in the subsequent stage to the determination unit 31, the determination unit 31 increases the reliability of the abnormal noise identification result obtained by the identification unit 13. , the threshold Y may be updated.

(Operation of Abnormal Sound Identification Device 30: S1)
The operation of the abnormal noise identification device 30 according to the third embodiment will be described with reference to FIG. Here, only the details of the flow of processing executed by the determination unit 31, that is, the contents of step S1 shown in FIG. 2 will be described.

The abnormal noise identification device 30 receives time-series data in the same manner as in the first embodiment. After that, the determination unit 31 of the abnormal noise identification device 30 determines the properties of the time-series data as described below.

As shown in FIG. 9, the data conversion unit 34 of the determination unit 31 converts the time series data (time domain signal) into a spectrogram (FIG. 7) (frequency domain signal) (S31).

The determination unit 31 generates a frequency spectrum by cutting out a segment of a predetermined time width from the spectrogram. The determination unit 31 calculates peak intensity in the frequency spectrum (S32).

The determination unit 31 determines whether or not the peak intensity in the frequency spectrum is greater than or equal to the threshold (S33). For example, the threshold is the average intensity in a given band centered around the peak frequency.

When the peak intensity in the frequency spectrum is equal to or greater than the threshold (Yes in S33), the determination unit 31 determines that the time-series data has property a (S34A). On the other hand, when the peak intensity in the frequency spectrum is below the threshold (No in S33), the determination unit 31 determines that the time-series data has property b (S34B).

With this, the processing of the determination unit 31 ends. After that, the process proceeds to the processing (step S2) of the selection unit 12 described in the first embodiment. In the third embodiment, the description of the process after the process (step S2) of the selection unit 12 is omitted.

(Method for determining threshold)
Here, a configuration has been described in which the determination unit 31 determines the properties of time-series data by comparing the peak intensity of the frequency spectrum and the threshold. Here, an example of a method for determining the threshold of peak intensity of the frequency spectrum will be described.

FIG. 10 is an example of a graph showing the distribution of scores used to determine the peak intensity threshold of the frequency spectrum.

The determination unit 31 calculates scores for a large number of learning data that contain the same or approximately the same number of time-series data that have been determined to have property a and time-series data that have been determined to have property b. The score here is the difference between the average intensity in a predetermined band around the peak frequency and the peak intensity. From the score calculation results, a score distribution as shown in FIG. 10 is obtained. Based on the score distribution, the determination unit 31 then determines a threshold so that time-series data having property a can be distinguished from time-series data having property b. For example, the determination unit 31 determines twice the maximum score of the time-series data having property b as the threshold. When the score of certain time-series data is equal to or higher than the threshold, the time-series data has a high probability of having property a. The determination unit 31 can determine the properties of the time-series data as described above by using the thresholds determined in this manner.

(Effect of this embodiment)
According to the configuration of this embodiment, the determination unit 31 determines the nature of time-series data. The selection unit 12 selects a technique for analyzing time series data based on the properties of the time series data. The identification unit 13 identifies abnormal sounds contained in the time-series data by analyzing the time-series data using the selected technique. Time-series data contain various types of sounds (including allophones) and noise, and the nature of time-series data also varies. For example, the time-series data may include an abnormal sound with large period fluctuations, or the noise may be large (the target sound is small).

Furthermore, according to the configuration of Embodiment 3, the determination unit 31 includes a data conversion unit 34 that converts time-series data into a spectrogram. The determination unit 31 determines the properties of the time-series data based on the peak intensity of the frequency spectrum cut out from the spectrogram for each predetermined time width. A sharp and strong peak in the frequency spectrum corresponds to a small period fluctuation of the peak in the original time-series data. Then, by comparing the peak intensity of the frequency spectrum with a threshold value, it is determined whether the period fluctuation of the peaks in the time series data is relatively small or the period fluctuations of the peaks in the time series data are relatively large. be able to.

(Modification)
In one modified example of any one of the first to third embodiments, the identification unit 13 identifies abnormal sounds included in the time-series data using three or more identifiers.

For example, the identification unit 13 according to this modification includes a classifier B that performs machine learning using MFCC as a feature quantity, and a classifier that uses DCTC (Discrete Cosine Transform Coefficients) as a feature quantity (hereinafter referred to as a classifier C). used together. The identification unit 13 identifies abnormal sounds contained in the time-series data by using the two identifiers, respectively, and compares the reliability of the identification results. When the reliability of the discrimination result by the discriminator B is higher, the discriminating unit 13 outputs the discrimination result by the discriminator B. FIG. On the other hand, when the reliability of the discrimination result by the discriminator C is higher, the discriminating unit 13 outputs the discrimination result by the discriminator C. FIG.

In another modification, the identification unit 13 may use multiple identifiers according to the location where the acoustic signal that is the source of the time-series data was acquired. In this modified example, each discriminator is associated in advance with information indicating different locations. The identification unit 13 according to the present modification receives the time-series data as well as information indicating the location associated with the time-series data from the determination unit 11 . The identifying unit 13 selects a corresponding classifier from a plurality of classifiers based on the information indicating the location. Then, the identification unit 13 uses the selected identifier to identify abnormal sounds included in the time-series data.

According to the configuration of this modified example, one of the two discriminators with the higher reliability of the discrimination result is selected, so that the reliability of the discrimination result output by the discrimination unit 13 can be improved.

[Hardware configuration]
Each component of the abnormal

noise identifying devices

10, 20, and 30 described in the first to third embodiments represents a functional unit block. Some or all of these components are realized by an information processing device 900 as shown in FIG. 11, for example. FIG. 11 is a block diagram showing an example of the hardware configuration of the information processing device 900. As shown in FIG.

As shown in FIG. 11, the information processing device 900 includes the following configuration as an example.

- CPU (Central Processing Unit) 901
・ROM (Read Only Memory) 902
・RAM (Random Access Memory) 903
・Program 904 loaded into RAM 903
- Storage device 905 for storing program 904
A drive device 907 that reads and writes the recording medium 906
- A communication interface 908 that connects to the communication network 909
- An input/output interface 910 for inputting/outputting data
A bus 911 connecting each component
Each component of the abnormal

noise identifying devices

10, 20, and 30 described in the first to third embodiments is implemented by the CPU 901 reading and executing a program 904 that implements these functions. A program 904 that implements the function of each component is stored in advance in, for example, the storage device 905 or the ROM 902, and is loaded into the RAM 903 and executed by the CPU 901 as necessary. The program 904 may be supplied to the CPU 901 via the communication network 909 or may be stored in the recording medium 906 in advance, and the drive device 907 may read the program and supply it to the CPU 901 .

According to the above configuration, the abnormal

noise identifying devices

10, 20, and 30 described in the first to third embodiments are implemented as hardware. Therefore, the same effects as those described in the first to third embodiments can be obtained.

[Appendix]
One aspect of the present invention is also described in the following appendices, but is not limited to the following.

(Appendix 1)
Determination means for determining properties of time-series data;
selection means for selecting a technique for analyzing the time series data based on the properties of the time series data;
A data analysis device comprising identification means for identifying abnormal sounds contained in the time-series data by analyzing the time-series data using the selected technique.

(Appendix 2)
The determination means comprises peak detection means for detecting a peak of the time-series data,
Supplementary Note 1, wherein the determination means determines the property of the time-series data based on the time width from the detection of the peak of the time-series data to the detection of the next peak. The data analysis device described in .

(Appendix 3)
The determination means comprises data conversion means for converting the time-series data into a spectrogram,
The data analysis device according to Supplementary Note 1, wherein the determining means determines the property of the time-series data based on the peak intensity of a frequency spectrum extracted from the spectrogram for each predetermined time width.

(Appendix 4)
The determination means determines the magnitude of variation in the period of the periodic component included in the time-series data,
4. The method according to any one of appendices 1 to 3, wherein the selection means selects a method according to the magnitude of the variation in the period from among a plurality of methods for analyzing the time-series data. Data analysis device as described.

(Appendix 5)
5. The data analysis apparatus according to appendix 4, wherein the selection means selects NMF (Nonnegative Matrix Factorization) when the magnitude of the period variation is below a threshold.

(Appendix 6)
determine the nature of time-series data,
selecting a technique for analyzing the time series data based on the properties of the time series data;
A data analysis method comprising identifying abnormal sounds contained in the time-series data by analyzing the time-series data using the selected technique.

(Appendix 7)
determining the nature of the time series data;
selecting a technique for analyzing the time series data based on the properties of the time series data;
A non-temporary recording medium storing a program for causing a computer to identify abnormal sounds contained in the time-series data by analyzing the time-series data using the selected method.

The present invention can be used, for example, in an abnormal noise identification device that identifies abnormal sounds emitted by railways, automobile engine rooms, factories, and other equipment or parts.

REFERENCE SIGNS LIST 10 abnormal noise identifying device 11 determining unit 12 selecting unit 13 identifying unit 20 abnormal noise identifying device 24 peak detecting unit 30 abnormal noise identifying device 34 data converting unit

Claims

Determination means for determining properties of time-series data;
selection means for selecting a technique for analyzing the time series data based on the properties of the time series data;
A data analysis device comprising identification means for identifying abnormal sounds contained in the time-series data by analyzing the time-series data using the selected technique.
The determination means comprises peak detection means for detecting a peak of the time-series data,
3. The determining means determines the property of the time-series data based on the time width from the detection of the peak of the time-series data to the detection of the next peak. 2. The data analysis device according to 1.
The determination means comprises data conversion means for converting the time-series data into a spectrogram,
2. The data analysis apparatus according to claim 1, wherein said determination means determines said property of said time-series data based on peak intensity of a frequency spectrum extracted from said spectrogram for each predetermined time width. .
The determination means determines the magnitude of variation in the period of the periodic component included in the time-series data,
4. The method according to any one of claims 1 to 3, wherein the selecting means selects a method from a plurality of methods for analyzing the time-series data according to the magnitude of variation in the period. The data analysis device described in .
5. The data analysis apparatus according to claim 4, wherein the selection means selects NMF (Nonnegative Matrix Factorization) when the magnitude of the period variation is below a threshold.
Determine the nature of time series data,
selecting a technique for analyzing the time series data based on the properties of the time series data;
A data analysis method comprising identifying abnormal sounds contained in the time-series data by analyzing the time-series data using the selected technique.
determining the nature of the time series data;
selecting a technique for analyzing the time series data based on the properties of the time series data;
A non-temporary recording medium storing a program for causing a computer to identify abnormal sounds contained in the time-series data by analyzing the time-series data using the selected method.