Signal processing device, method and program

Abstract

To exactly extract a signal which should be extracted from time series signals without time deviation.SOLUTION: A signal processing device concerning an embodiment comprises an acquisition unit, a storage unit, a candidate selection unit, a position determination unit, and an extraction unit. The acquisition unit acquires time series signals. The storage unit preliminarily stores a template regarding a section signal which should be extracted among the time series signals. The candidate selection unit selects a plurality of candidates of a time position at which the section signal is extracted from the time series signals and excludes candidates within a predetermined time from prominent candidates selected on the basis of an order of similarity among the plurality of candidates on the basis of the similarity between the time series signals and the template. The position determination unit determines an extraction position on the basis of the prominent candidates remaining without being excluded. The extraction unit extracts the section signal according to the extraction position from the time series signals.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] FIELD Embodiments of the present invention relate to a signal processing device, method, and program. [Background technology]

[0002] Conventionally, a signal processing device is known that acquires a time-series signal output from a recorder that monitors the operating status of a monitored object and detects abnormalities in the operating status. For example, the signal processing device detects abnormalities in an air knocker connected to a dryer installed in a nuclear waste reprocessing plant. The air knocker strikes the dryer to prevent clogging of the dryer with hot powder. When striking the dryer, the air knocker generates mechanical knocking sounds at regular intervals (intermittently). The signal processing device acquires a time-series signal containing the mechanical knocking sounds, extracts the mechanical knocking sounds from the time-series signal, and detects abnormalities using an autoencoder.

[0003] Here, because multiple air knockers are connected to the dryer, mechanical knocking sounds similar to those of air knockers other than the air knocker being monitored are generated. In addition, artificial knocking sounds generated by factory maintenance personnel occur without temporal regularity. Therefore, the time-series signal includes signals representing multiple similar types of mechanical knocking sounds that occur intermittently and signals representing artificial knocking sounds that occur without temporal regularity. As a result, the signal processing device may not be able to accurately extract a signal representing the monitored object from the time-series signal. Furthermore, because mechanical knocking sounds may occur due to time interval fluctuations, the signal processing device may extract a time-series signal at a time that is different from the extraction time of the time-series signal that should be extracted. Therefore, it is necessary to accurately extract the signal that should be extracted from the time-series signal without any time lag. Note that if accurate extraction is not possible, the accuracy of the autoencoder's ability to detect abnormalities and estimate deterioration in the air knocker will decrease, and the capacity of the autoencoder stored in the signal processing device will increase, which is undesirable. Therefore, accurate extraction is necessary. Furthermore, if an attempt is made to extract a signal to be extracted based on a trigger signal that drives the air knocker, the signal processing device will need to acquire the trigger signal from the air knocker. Therefore, it is desirable to accurately extract a signal to be extracted from a time-series signal without relying on the trigger signal of the air knocker. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2010-266327 Summary of the Invention [Problem to be solved by the invention]

[0005] The problem to be solved by the present invention is to accurately extract a signal to be extracted from a time-series signal without time lag. [Means for solving the problem]

[0006] A signal processing device according to an embodiment includes an acquisition unit, a storage unit, a candidate selection unit, a position determination unit, and an extraction unit. The acquisition unit acquires a time series signal. The storage unit pre-stores a template for a section signal to be extracted from the time series signal. The candidate selection unit selects multiple candidates for a time position at which to extract the section signal from the time series signal based on the similarity between the time series signal and the template, and excludes, from the multiple candidates, candidates within a predetermined time from likely candidates selected based on the order of similarity. The position determination unit determines the extraction position based on the remaining likely candidates that have not been excluded. The extraction unit extracts the section signal from the time series signal according to the extraction position. [Brief explanation of the drawings]

[0007] [Figure 1] FIG. 1 is a diagram illustrating an example of a signal processing device according to the first embodiment and its peripheral configuration. [Figure 2] FIG. 2 is a diagram illustrating a trained model according to the first embodiment. [Figure 3] FIG. 3 is a flowchart illustrating an example of the operation in the first embodiment. [Figure 4] FIG. 4 is a flowchart illustrating an example of a part of the operation in the first embodiment. [Figure 5] FIG. 5 is a flowchart illustrating an example of a part of the operation in the first embodiment. [Figure 6] FIG. 6 is a flowchart illustrating an example of a part of the operation in the first embodiment. [Figure 7] FIG. 7 is a diagram for explaining extraction of a time-series signal in the first embodiment. [Figure 8] FIG. 8 is a scatter diagram showing an example of the calculation results of the abnormality degree according to the comparative example. [Figure 9] FIG. 9 is a scatter diagram showing an example of the determination result of the abnormality degree according to the first embodiment. [Figure 10]FIG. 10 is a diagram for comparing the performance of the signal processing device of the comparative example and the signal processing device of the first embodiment. [Figure 11] FIG. 11 is a diagram illustrating an example of a signal processing device according to the second embodiment and its peripheral configuration. [Figure 12] FIG. 12 is a flowchart illustrating an example of the operation in the second embodiment. [Figure 13] FIG. 13 is a scatter diagram showing an example of the calculation results of the abnormality degree according to the second embodiment. [Figure 14] FIG. 14 is a diagram illustrating a separation unit according to the second embodiment. [Figure 15] FIG. 15 is a diagram illustrating a hardware configuration of a signal processing device according to the third embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0008] Hereinafter, an embodiment of a signal processing device will be described with reference to the drawings. In the following description, the same reference numerals are used to designate substantially the same parts in different drawings, and redundant description will be omitted.

[0009] (First embodiment) 1 is a diagram showing an example of the configuration of a signal processing device 100 and its peripherals according to the first embodiment. The signal processing device 100 is connected to a recorder 20 that collects sound signals. The recorder 20 is provided in the vicinity of a monitored object 10 that generates mechanical hammering sounds that are part of the sound signals. The recorder 20 and the monitored object 10 are provided, for example, inside a room in a nuclear waste reprocessing plant (not shown).

[0010] The monitored object 10 is an object monitored by the signal processing device 100 based on the generated mechanical knocking sound. The monitored object 10 is, for example, an air knocker. The air knocker is connected to a dryer (not shown). For example, three air knockers, including the air knocker serving as the monitored object 10, are connected to the dryer. The signal processing device 100, for example, selects one of the three air knockers as the monitored object 10. The air knocker strikes the dryer to prevent hot powder from clogging the interior of the dryer. The air knocker has, for example, a solenoid valve that can be opened and closed by a trigger signal, and a piston with an embedded magnet slidably mounted within a cylinder. To strike the dryer, the piston is moved from its initial position toward the dryer by introducing compressed air from the solenoid valve, generating a loud mechanical knocking sound. After moving toward the dryer, the piston is returned to its initial position by the discharge of compressed air and the magnet, generating a further soft mechanical knocking sound. The piston periodically repeats the operation of moving from its initial position toward the dryer and then returning to its initial position. For example, the period of this operation is approximately 30 seconds. In this way, the air knocker periodically generates a loud mechanical knocking sound approximately once every 30 seconds. Because the mechanical knocking sound occurs periodically, it falls under the category of intermittent operation noise.

[0011] The recorder 20 is installed near the monitored object 10 and collects and records ambient sound signals from the vicinity. The recorder 20 includes three microphones, numbered first through third. Each of the three microphones is installed near, for example, one of the three air knockers. Of the three microphones, the first microphone, installed near the air knocker of the monitored object 10, primarily collects mechanical knocking sounds generated by the air knocker of the monitored object 10. However, the first microphone also collects similar mechanical knocking sounds generated by air knockers other than the air knocker of the monitored object 10. In addition, the first microphone also collects irregular artificial knocking sounds, such as those generated when a maintenance worker at a nuclear waste reprocessing plant hits a dryer with a hammer. The sound signals collected by the first microphone include, for example, the mechanical knocking sounds and artificial knocking sounds of each of the three air knockers. The sampling frequency of the A / D converter of the recorder 20 that digitizes the collected sound signal is, for example, 44.1 kHz. However, this is not limited to this and other frequencies may be used. The recorder 20 transmits a time series signal obtained by digitizing the sound signal to the signal processing device 100. Note that a time series signal related to the operating state of the monitored object 10 is called an operating time series signal. Furthermore, the signal processing device 100 is not limited to processing the operating time series signal, and may also process a time series signal that is not related to the operating state of the monitored object.

[0012] The signal processing device 100 is a device that processes a time-series signal acquired from the recorder 20. The signal processing device 100 includes, for example, an acquisition unit 110, a detection unit 120, a storage unit 130, a candidate selection unit 140, a position determination unit 150, a cutout unit 160, a model storage unit 170, and a determination unit 180. However, without being limited to this, the signal processing device 100 may be configured by appropriately omitting any additional components such as the detection unit 120, the model storage unit 170, and the determination unit 180. Furthermore, the signal processing device 100 may also be called a state monitoring device.

[0013] The acquisition unit 110 acquires the time-series signal transmitted from the recorder 20. The acquisition unit 110 divides the acquired time-series signal into multiple time-series signals at predetermined time intervals. For example, the acquisition unit 110 divides the acquired time-series signal into multiple time-series signals every 60 seconds. The acquisition unit 110 may also be called a time-series signal acquisition unit.

[0014] The detection unit 120 monitors multiple power components corresponding to each of multiple signals constituting the time-series signal and detects a section candidate signal composed of the multiple signals from the time-series signal based on the multiple power components. The section candidate signal includes multiple sample signals corresponding to multiple candidates for time positions at which section signals, which are signals to be extracted from the time-series signal, are extracted. Specifically, the detection unit 120 includes a power monitoring unit 121 and a candidate detection unit 122. Note that when the time-series signal is a motion time-series signal, the detection unit 120 monitors multiple power components corresponding to each of the multiple signals constituting the motion time-series signal and detects a motion section candidate signal composed of the multiple signals from the motion time-series signal based on the multiple power components. The detection unit 120 may also be referred to as a motion section candidate detection unit. To minimize the amount of processing, the detection unit 120 first widely detects section candidate signals using the power components of the time-series signal.

[0015] The power monitoring unit 121 further divides the time-series signal divided every 60 seconds into predetermined time intervals. Hereinafter, a time-series signal further divided into predetermined time intervals will be referred to as a frame. Furthermore, each of the multiple signals constituting the time-series signal will be referred to as a sample signal. A frame is made up of, for example, 100 sample signals. The power monitoring unit 121 extracts one frame from the multiple frames.

[0016] The power monitoring unit 121 monitors each of a plurality of power components corresponding to each of a plurality of sample signals constituting the extracted frame. For example, the power monitoring unit 121 calculates the average value and maximum value of the plurality of power components constituting the extracted frame based on each of the monitored power components. The power monitoring unit 121 executes the above-described process of extracting one frame from the plurality of frames, monitoring each of the plurality of power components, and calculating the average value and maximum value of the plurality of power components for all frames. Note that the power monitoring unit 121 may also monitor, for example, a power component of a specific frequency or a power component over the entire band. Note that a power component is a value equivalent to the square of the amplitude of a sample signal. The power component may also be called a power spectrum.

[0017] The candidate detection unit 122 detects section candidate signals composed of multiple signals from a time-series signal divided into 60-second intervals based on multiple power components. If the average and maximum values ​​of the multiple power components are each equal to or greater than a threshold, the candidate detection unit 122 selects multiple signals corresponding to the multiple power components. Specifically, for example, when a sampling frequency is 44.1 kHz and one frame is 100 samples, the candidate detection unit 122 selects 100 sample signals if the average value of the sum of squared amplitudes within one frame (short-term power) is 76 dB or greater and the maximum value of the squared amplitude within one frame is 84 dB or greater. In this embodiment, in most cases, 0 to 3 frames contain 100 samples with short-term power equal to or greater than 76 dB among the time-series signals divided into 60-second intervals. The candidate detection unit 122 detects time-series signals corresponding to the selected multiple signals and the multiple signals before and after them as section candidate signals. The candidate detection unit 122 detects, as section candidate signals, 3,000 sample signals, for example, from the sample signal 500 samples before the selected 100-sample signal to the sample signal 2,999 samples after the selected sample signal. The candidate detection unit 122 calculates the average and maximum values ​​of the multiple power components, selects multiple sample signals for which both values ​​are equal to or greater than a threshold, and detects them as section candidate signals. This process is performed for all frames constituting the time-series signal divided into 60-second intervals. By performing this process, the candidate detection unit 122 typically detects multiple section candidate signals from the time-series signal divided into 60-second intervals. By detecting multiple section candidate signals, the candidate detection unit 122 first broadly identifies multiple sample signals corresponding to multiple candidates. Note that the candidate detection unit 122 may detect section candidate signals based on, for example, power components of a specific frequency or power components across the entire frequency band. The candidate detection unit 122 may also be referred to as a motion section candidate extraction unit.

[0018] The storage unit 130 stores in advance a template relating to the section signal to be extracted. The storage unit 130 stores a plurality of templates in advance. For example, the storage unit 130 stores in advance a template relating to a loud mechanical hammering sound when the monitored object 10 is in a normal state. The storage unit 130 stores a template made up of 3000 sample signals including a plurality of sample signals corresponding to one mechanical hammering sound. The template may also be called a time waveform template. The storage unit 130 may also be called a template storage unit.

[0019] The candidate selection unit 140 acquires a template from the storage unit 130. The candidate selection unit 140 selects multiple candidates for the time position at which to extract a section signal from the time series signal based on the similarity between the time series signal and the template. The candidate selection unit 140 selects several likely candidates from the multiple candidates based on the order of similarity, and excludes candidates within a predetermined time from the likely candidates. Specifically, the candidate selection unit 140 includes a score calculation unit 141, a score sorting unit 142, and a search unit 143.

[0020] The score calculation unit 141 calculates the similarity using the section candidate signal as a time-series signal. The score calculation unit 141 calculates the similarity between the section candidate signal and a template. The score calculation unit 141 selects multiple signals obtained by shifting the section candidate signal by a predetermined number of signals. For example, to search for a time position 2000 samples later, the score calculation unit 141 calculates the similarity for each sample. The score calculation unit 141 calculates the similarity using the multiple signals as section candidate signals. The score calculation unit 141 selects multiple other signals obtained by delaying the section candidate signal, and sequentially performs the process of calculating the similarity using the multiple other signals as section candidate signals. If the maximum similarity among the calculated similarities is greater than a threshold, the score calculation unit 141 calculates the time position of the maximum section candidate signal, which is the multiple signals corresponding to the maximum similarity, based on the delay of the time position of the maximum section candidate signal, which is the multiple signals corresponding to the maximum similarity, relative to the time positions of multiple signals shifted by a predetermined number of signals. The score calculation unit 141 selects the calculated time position as a candidate. The score calculation unit 141 calculates, for example, a cross-correlation coefficient between the selected portion of signals and the template as the similarity. The cross-correlation coefficient may also be referred to as a score. Specifically, the score calculation unit 141 selects 3,000 sample signals obtained by shifting the section candidate signals 500 samples earlier. In other words, the score calculation unit 141 selects 3,000 sample signals, starting from a signal located 500 samples earlier than the time position of the sample signal that is the starting point of the section candidate signals, to a sample signal located 2,999 samples later than the sample signal. Hereinafter, these 3,000 sample signals are referred to as reference section candidate signals. The score calculation unit 141 selects 3,000 sample signals obtained by delaying the reference section candidate signals by one sample. In other words, the score calculation unit 141 selects 3,000 sample signals, starting from a sample signal located one sample later than the time position of the sample signal that is the starting point of the reference section candidate signals, to a sample signal located 2,999 samples later than the sample signal. The score calculation unit 141 calculates the cross-correlation coefficient between the 3000 sample signals delayed by one sample and the template.After the calculation, the score calculation unit 141 selects 3000 sample signals obtained by further delaying the 3000 sample signals delayed by one sample. In other words, the score calculation unit 141 selects 3000 sample signals, starting from a sample signal located two samples after the time position of the sample signal that is the starting point of the reference section candidate signal, to a sample signal located 2999 samples after the sample signal. The score calculation unit 141 calculates a cross-correlation coefficient between the 3000 sample signals delayed by one sample and the template. Note that the signal delayed by one sample and the 3000 sample signals delayed by one sample may also be referred to as section candidate signals. Furthermore, 3000 sample signals obtained by delaying the reference section candidate signal by τ (τ: positive integer) samples may also be referred to as section candidate signals. Furthermore, hereinafter, "delay" refers to the delay of a section candidate signal obtained by delaying a reference section candidate signal by τ samples relative to the reference section candidate signal. The score calculation unit 141 selects a section candidate signal obtained by delaying the reference section candidate signal by τ samples, calculates a cross-correlation coefficient between the selected signal and a template, and after the calculation, sequentially executes a process of delaying the selected signal by another sample. The score calculation unit 141 executes the above process 2000 times, for example. The score calculation unit 141 calculates 2000 cross-correlation coefficients by repeatedly executing the above process 2000 times.

[0021] If the maximum similarity among the calculated similarities is greater than a threshold, the score calculation unit 141 calculates the time position of the maximum section candidate signal, which is a plurality of signals corresponding to the maximum similarity, based on a delay in the time position of the maximum section candidate signal. The score calculation unit 141 selects the calculated time position as a candidate. Specifically, for example, if the maximum cross-correlation coefficient among the calculated 2000 cross-correlation coefficients is greater than a threshold of 0.2, the score calculation unit 141 calculates the time position of the sample signal that is the starting point of the maximum section candidate signal that corresponds to the maximum cross-correlation coefficient. The score calculation unit 141 calculates the time position of the sample signal that is the starting point of the maximum section candidate signal based on the time position of the reference section candidate signal and the delay in the time position of the maximum section candidate signal.

[0022] The score calculation unit 141 selects, as a candidate, the time position of the sample signal that is the starting point of the calculated maximum section candidate signal. The time position of the sample signal that corresponds to the maximum cross-correlation coefficient is called a time index. The score calculation unit 141 selects the time index as a candidate. The score calculation unit 141 repeatedly performs the above-mentioned process on all detected section candidate signals. In this way, the score calculation unit 141 selects multiple candidates. If the time series signal is a motion time series signal, the candidate selection unit 140 calculates the similarity related to the motion state by using the motion section candidate signal as the motion time series signal.

[0023] The score sorting unit 142 sorts the selected candidates in descending order of similarity.

[0024] The search unit 143 selects the candidate with the highest similarity from among the multiple candidates as the first likely candidate and excludes candidates within a predetermined time from the first likely candidate. If, after the exclusion, the remaining candidates that have not been excluded include a candidate other than the first likely candidate, the search unit 143 selects the candidate with the highest similarity from among the remaining candidates excluding the first likely candidate as the second likely candidate and excludes candidates within a predetermined time from the second likely candidate. Note that, if, after the exclusion, the remaining candidates that have not been excluded include a candidate other than the first likely candidate and the second likely candidate, the search unit 143 selects the candidate with the highest similarity from among the remaining candidates excluding the first likely candidate and the second likely candidate as the third likely candidate. The search unit 143 excludes candidates within a predetermined time from the third likely candidate. The search unit 143 repeatedly performs the process of selecting the candidate with the highest similarity as the likely candidate and excluding candidates within a predetermined time from the selected likely candidate until all remaining candidates are selected as likely candidates. When the candidates within a predetermined time from the third likely candidate are excluded, if the remaining candidates are the first to third likely candidates, the search unit 143 ends the above process at this stage. Note that the search unit 143 may also be called a time-constrained search unit.

[0025] The position determination unit 150 determines the cutout position based on the remaining likely candidates that were not excluded by the candidate selection unit 140. If the remaining likely candidates are the first likely candidate, the position determination unit 150 determines the first likely candidate as the cutout position. If the remaining likely candidates are the first likely candidate and the second likely candidate, the position determination unit 150 determines the first likely candidate and the second likely candidate as the cutout positions. Note that this is not limited to this, and the position determination unit 150 may determine all remaining likely candidates that were not excluded as cutout positions. In this example, since a loud mechanical tapping sound occurs approximately every 30 seconds, approximately 0 to 3 likely candidates are extracted from the time-series signal divided into 60-second segments, and typically 0 to 3 cutout positions are determined. If no tapping sound is heard in the time-series signal divided into 60 seconds, the number of cutout positions is 0. Furthermore, if a tapping sound is missing from the time-series signal divided into 60 seconds, the number of cutout positions is 1. In most cases, the number of cut-out positions is 2. In rare cases, the number of cut-out positions is 3. The position determination unit 150 may also be called a cut-out position determination unit.

[0026] The clipping unit 160 clips a section signal from the time-series signal according to the clipping position. The clipping unit 160 clips, as the section signal, 3000 sample signals including the sample signal corresponding to the clipping position. Specifically, the clipping unit 160 clips, for example, 3000 sample signals from the sample signal corresponding to the clipping position to the sample signal located 2999 samples later in time from the sample signal corresponding to the clipping position, as the section signal. Note that, when the time-series signal is a motion time-series signal, the clipping unit 160 clips, as the section signal, the motion section signal from the motion time-series signal according to the clipping position.

[0027] 2, the model storage unit 170 stores a trained model Md10 consisting of a trained machine learning model Md1. The model storage unit 170 stores an autoencoder (AE) including the machine learning model Md1 that has been trained in advance in an unsupervised manner. However, the model storage unit 170 is not limited to this, and may store a convolutional autoencoder (CAE) or a variational autoencoder (VAE) consisting of the machine learning model Md1 that has been trained in advance in an unsupervised manner, instead of an autoencoder.

[0028] FIG. 2(a) shows an example of learning the machine learning model Md1. The input data of the machine learning model Md1 is a normal operation interval signal, which is an operation interval signal when the monitored object 10 is in a normal state. Specifically, the normal operation interval signal is an operation interval signal extracted from the intermittent operation sound of the monitored object 10 during initial use (after component calibration, overhaul, or maintenance). The machine learning model Md1 outputs a reconstructed signal that is substantially identical to the normal operation interval signal as output data. The machine learning model Md1 is, for example, composed of an encoder and a decoder that constitute an autoencoder. The machine learning model Md1 learns to minimize a loss function. The loss function may be the reconstruction error or the sum of the reconstruction error and a regular term. For example, an autoencoder or convolutional autoencoder equipped with the machine learning model Md1 outputs the reconstruction error calculated based on the normal operation interval signal and the reconstructed signal as a loss function. The machine learning model Md1 learns network parameters to minimize the output reconstruction error. The reconstruction error may be, for example, the root mean squared error (RMSE). The variational autoencoder equipped with the machine learning model Md1 outputs the sum of the reconstruction error and a regular term calculated based on the normal operation section signal and the reconstructed signal as a loss function. The machine learning model Md1 learns network parameters so as to minimize the sum of the output reconstruction error and the regular term.

[0029] Furthermore, the trained model Md10 is an implemented trained machine learning model Md1, as shown in the example of an abnormality detection in Figure 2(b), and when an operation interval signal is input, it performs inference processing based on the operation interval signal and outputs a reconstruction signal.

[0030] The determination unit 180 determines the operating state based on the operation interval signal. The determination unit 180 inputs the operation interval signal to an autoencoder including the trained model Md10, outputs an abnormality degree, and determines the operating state based on the abnormality degree. Specifically, the determination unit 180 generates a reconstructed signal by inputting the operation interval signal to the trained model Md10. The autoencoder or convolutional autoencoder including the trained model Md10 calculates and outputs a reconstruction error as an abnormality degree based on the operation interval signal and the reconstructed signal. On the other hand, the variational autoencoder including the trained model Md10 calculates and outputs the negative log-likelihood of the reconstruction probability as an abnormality degree based on the operation interval signal and the reconstructed signal. If the abnormality degree is equal to or greater than a threshold, the determination unit 180 determines the operating state as abnormal. If the abnormality degree is not equal to or greater than the threshold, the determination unit 180 determines the operating state as normal. The abnormality degree may also be referred to as a deterioration degree. Examples of abnormalities include sudden failures and deterioration due to aging. The determining unit 180 outputs a determination result indicating normality (no abnormality) or abnormality (abnormality / deterioration). The determining unit 180 may also be called an operation state determining unit.

[0031] Next, an example of the operation of the signal processing device 100 configured as described above will be described with reference to FIGS. 3 to 7. FIG. 3 is a flowchart showing an example of the operation of the signal processing device 100. FIGS. 4 to 6 are flowcharts showing an example of a portion of the operation of the signal processing device 100. FIG. 7 is a diagram for explaining the extraction of a time-series signal in the signal processing device 100. FIG. 7 shows an example of the waveform of a time-series signal divided every 60 seconds. In FIG. 7, the horizontal axis represents time, and the vertical axis represents sound amplitude. Time positions P111 and P112 of the sample signals represent the time positions of the sample signals indicating a loud mechanical knocking sound generated when the piston of a first air knocker, one of the three air knockers, strikes the dryer. Note that this mechanical knocking sound occurs twice within 60 seconds. Time position P111 represents the time position of the sample signal indicating the earlier of the two loud mechanical knocking sounds that occurred. Furthermore, time position P112 represents the time position of the sample signal indicating the later of the two loud mechanical knocking sounds that occurred. Time position P121 represents the time position of a sample signal indicating a small mechanical knock generated by the piston of the first air knocker returning to its initial position. This small mechanical knock occurs once every 60 seconds. Time positions P211, P212, P311, and P312 represent large mechanical knocks generated by the pistons of the second and third air knockers striking the dryer. That is, time position PXYZ represents the time position of a sample signal indicating a mechanical knock generated by the first air knocker when the most significant digit X of the three-digit number XYZ following P is 1, the second air knocker when it is 2, or the third air knocker when it is 3. Time position PXYZ represents the time position of a sample signal indicating a large mechanical knock generated by the piston striking the dryer when the middle digit Y is 1, or a small mechanical knock generated by the piston returning to its initial position when it is 2. The time position PXYZ indicates the time position of the sample signal that indicates the first of the two mechanical tap sounds that occurred when the value of the least significant digit Z is 1, and the last of the two mechanical tap sounds that occurred when the value is 2.For example, when the second air knocker is the monitoring target 10, the monitoring target 10 is monitored based on the loud mechanical knocking sounds generated when the piston strikes the dryer, so the signals at time positions P211 and P212 are the section signals to be extracted. Also, in Fig. 7, the low amplitude portion s indicates a standing wave between the knocking sounds in the time series signal. The low amplitude portion will be described later.

[0032] First, the air knocker, which is the monitored object 10, generates a mechanical knocking sound. Of the microphones provided in the recorder 20, a microphone provided in a position near the monitored object 10 collects an operation time-series signal, which is a sound signal including the mechanical knocking sound. The recorder 20 transmits the collected operation time-series signal to the signal processing device 100. In this initial state, step ST10 starts.

[0033] In step ST10, the acquisition unit 110 acquires the action time-series signal transmitted from the recorder 20.

[0034] After step ST10, in step ST50, the acquiring unit 110 divides the acquired motion time-series signal into a plurality of motion time-series signals every 60 seconds. Specifically, the motion time-series signal is composed of 2646000 (=sampling frequency 44100 [Hz] × 60 [sec]) sample signals.

[0035] After step ST50, in step ST100, the power monitoring unit 121 divides the 2,646,000 sample signals that make up the operation time-series signal into frames, each frame being a set of 100 sample signals. In this case, the operation time-series signal is made up of 26,460 (=2,646,000 / 100) frames.

[0036] After step ST100, in step ST200, the power monitoring unit 121 detects a motion section candidate signal from the motion time-series signal divided into 60-second intervals.

[0037] Here, step ST200 includes steps ST210 to ST235, for example, as shown in FIG.

[0038] In step ST210, the power monitoring unit 121 extracts one frame from the 26460 frames.

[0039] After step ST210, in step ST215, each of 100 power components corresponding to each of 100 sample signals constituting the extracted frame is monitored. Based on each of the monitored 100 power components, the power monitoring unit 121 calculates the average value and maximum value of the 100 power components. Specifically, the power monitoring unit 121 calculates the average value of the sum of squares of amplitudes within one frame as the average value. Furthermore, the power monitoring unit 121 specifically calculates the maximum value of the squares of amplitudes within one frame as the maximum value.

[0040] After step ST215, in step ST220, the candidate detection unit 122 determines whether or not each of the average value and maximum value of the 100 power components is equal to or greater than a threshold. If the result of the determination is that they are equal to or greater than the threshold, the process proceeds to step ST225. Specifically, for example, if the average value of the sum of squared amplitudes in one frame is equal to or greater than 76 dB and the maximum value of squared amplitudes in one frame is equal to or greater than 84 dB, the process proceeds to step ST225. On the other hand, if the result of the determination is that they are not equal to or greater than the threshold, the process proceeds to step ST235.

[0041] After step ST220, in step ST225, the candidate detection unit 122 selects 100 sample signals that make up the extracted frame.

[0042] After step ST225, in step ST230, the candidate detection unit 122 detects 3000 sample signals, from the sample signal 500 samples before the sample signal that is the starting point of the selected 100 sample signals to the sample signal that is 2999 samples after the selected sample signal, as motion section candidate signals.

[0043] After step ST220 or ST230, in step ST235, the detection unit 120 calculates the average and maximum values ​​of 100 sample signals for all 26,460 frames and determines whether or not the process of detecting them as motion section candidate signals has been executed. If the result of the determination is that this process has not been executed for all 26,460 frames, the process proceeds to step ST210. If the result of the determination is that this process has been executed for all 26,460 frames, step ST200 ends upon completion of step ST235. In this case, the candidate detection unit 122 detects multiple motion section candidate signals from the motion time-series signals divided into 60-second intervals.

[0044] After step ST200, in step ST250, the score calculation section 141 acquires the template stored in the storage section .

[0045] After step ST250, in step ST300, the score calculation section 141 selects a plurality of candidates based on the cross-correlation coefficient (score), which is the degree of similarity between the motion section candidate signal and the template.

[0046] Here, step ST300 includes steps ST305 to ST345, for example, as shown in FIG.

[0047] In step ST305, the score calculation section 141 selects one movement section candidate signal from among the plurality of detected movement section candidate signals.

[0048] After step ST305, in step ST315, the score calculation section 141 selects a reference section candidate signal, which is a 3000-sample signal obtained by shifting the motion section candidate signal forward by 500 samples.

[0049] After step ST315, in step ST320, the score calculation unit 141 selects a motion section candidate signal obtained by delaying the selected motion section candidate signal. Specifically, since the reference section candidate signal is currently selected, the score calculation unit 141 selects a 3000 sample signal obtained by delaying the reference section candidate signal by one sample as the motion section candidate signal.

[0050] After step ST320, in step ST325, the score calculation section 141 calculates the cross-correlation coefficient between the selected motion section candidate signal and the template.

[0051] After step ST325, in step ST330, the score calculation unit 141 determines whether or not 2000 cross-correlation coefficients have been calculated. Currently, one cross-correlation coefficient is being calculated. Therefore, if the result of the determination shows that 2000 cross-correlation coefficients have not been calculated, the process proceeds to ST320.

[0052] After step ST330, in step ST320, the score calculation unit 141 selects a motion section candidate signal obtained by delaying the selected motion section candidate signal. Specifically, since 3000 sample signals obtained by delaying the reference section candidate signal by one sample are currently selected as the motion section candidate signal, the score calculation unit 141 selects 3000 sample signals obtained by delaying the signal by another sample as the motion section candidate signal.

[0053] After step ST320, in step ST325, the score calculation unit 141 calculates a cross-correlation coefficient between the motion section candidate signal and the template. The score calculation unit 141 sequentially executes the processes of steps ST320 and ST325. The score calculation unit 141 calculates 2000 cross-correlation coefficients by repeating the process 2000 times.

[0054] After step ST325, it is determined in step ST330 whether 2000 cross-correlation coefficients have been calculated. If it is determined that 2000 cross-correlation coefficients have been calculated, the process proceeds to ST335.

[0055] After step ST330, in step ST335, the score calculation unit 141 determines whether or not the maximum cross-correlation coefficient among the calculated 2000 cross-correlation coefficients is greater than a threshold value of 0.2. If the result of the determination is that the maximum cross-correlation coefficient is greater than 0.2, the process proceeds to step ST340. If the result of the determination is that the maximum cross-correlation coefficient is 0.2 or less, the process proceeds to step ST345.

[0056] After step ST335, in step ST340, the score calculation unit 141 selects the maximum section candidate signal corresponding to the maximum cross-correlation coefficient. The score calculation unit 141 calculates the time position (time index) of the maximum section candidate signal based on the time position of the reference section candidate signal and the delay of the time position of the maximum section candidate signal. The score calculation unit 141 selects the calculated time index as a candidate.

[0057] After step ST335 or ST340, in step ST345, the score calculation unit 141 determines whether or not the above process has been performed on all detected motion section candidate signals. If the result of the determination shows that the above process has not been performed on all motion section candidate signals, the process proceeds to step ST305. In this case, in step ST305, a motion section candidate signal for which the process of calculating a cross-correlation coefficient has not yet been performed is selected from the detected motion section candidate signals, and the subsequent processes are repeatedly performed until the above process has been performed on all motion section candidate signals. If the result of the determination shows that the process has been performed, step ST345 ends and step ST300 ends. As a result, the score calculation unit 141 selects multiple candidates. In FIG. 7, the multiple candidates are time positions P111, P112, P121, P211, P212, P221, P311, P312, P321, and P322.

[0058] After step ST300, in step ST350, the score sorting unit 142 sorts the selected candidates in descending order of the magnitude of the cross-correlation coefficient.

[0059] After step ST350, in step ST400, candidates that are within 20 seconds before and after the most likely candidate selected based on the order of the cross-correlation coefficient are excluded from the plurality of candidates.

[0060] Here, step ST400 includes steps ST410 to ST440, for example, as shown in FIG.

[0061] In step ST410, the search unit 143 selects the candidate with the largest cross-correlation coefficient as the first likely candidate from among the multiple candidates. As shown in Fig. 7, the search unit 143 selects time position P211 from among the multiple candidates as the first likely candidate C1.

[0062] After step ST410, in step ST420, the search unit 143 excludes candidates located within 20 seconds before and after the selected first likely candidate, taking into account that the air knocker periodically generates a loud mechanical knocking sound approximately once every 30 seconds. In FIG. 7, time t1 represents a time within 20 seconds before and after the first likely candidate C1, within the range of the motion time-series signal divided into 60 seconds. As shown in FIG. 7, the search unit 143 excludes time positions P311, P111, P321, P221, P312, P112, and P312, which are candidates located at time t1, from among the multiple candidates.

[0063] After step ST420, in step ST430, it is determined whether the remaining candidates that have not been excluded include any candidates other than the selected first likely candidate. In Fig. 7, after excluding candidates within 20 seconds before and after the first likely candidate C1, the remaining candidates that have not been excluded are at time positions P211, P121, and P212. As a result of the determination, the remaining candidates that have not been excluded include any candidates other than the selected first likely candidate C1, so the process proceeds to step ST440.

[0064] After step ST430, in step ST440, the search unit 143 selects the candidate with the largest cross-correlation coefficient from the remaining candidates not including the first likely candidate as the second likely candidate. As shown in Fig. 7, from the remaining candidates not including the first likely candidate, that is, time positions P121 and P212, the search unit 143 selects time position P212 of the signal as the second likely candidate C2.

[0065] After step ST440, in step ST420, the search unit 143 excludes candidates that are within 20 seconds before and after the selected second likely candidate. In Fig. 7, time t2 represents a time that is within 20 seconds before and after the second likely candidate C2 within the range of the action time-series signal divided into 60 seconds. As shown in Fig. 7, the search unit 143 excludes time position P121, which is the candidate that is at time t2, from the remaining candidates.

[0066] After step ST420, in step ST430, it is determined whether the remaining candidates that have not been excluded include any candidates other than the first and second likely candidates, which are the selected candidates. In Figure 7, after excluding candidates within 20 seconds before and after the first and second likely candidates, the remaining candidates that have not been excluded are the first and second likely candidates C1 and C2, which are the selected candidates, and do not include any candidates other than the selected candidates. As a result of the determination, no candidates other than the selected candidates are included, and therefore step ST400 ends upon completion of step ST430.

[0067] After step ST400, in step ST450, the position determining section 150 determines the first and second likely candidates, which are the remaining candidates that have not been excluded, as cut-out positions.

[0068] After step ST450, in step ST500, the extraction unit 160 extracts the movement section signal to be extracted from the movement time-series signal according to the extraction position. The extraction unit 160 extracts 3000 sample signals from the sample signals corresponding to time positions P211 and P212, which are the extraction positions, to the sample signal at the time position 2999 after the sample signals, as the movement section signal S, as shown in FIG.

[0069] After step ST500, in step ST550, the judgment unit 180 judges the operating state based on the operation interval signal. Based on the normal operation interval signal, which is the operation interval signal when the monitored object 10 is in a normal state, the judgment unit 180 generates a reconstructed signal by inputting the operation interval signal to a trained model that generates a reconstructed signal that is substantially identical to the normal operation interval signal. The judgment unit 180 calculates the degree of abnormality based on the operation interval signal and the reconstructed signal. If the degree of abnormality is equal to or greater than the threshold, the process proceeds to step ST600. If the degree of abnormality is not equal to or greater than the threshold, the process proceeds to step ST601.

[0070] After step ST550, in step ST600, the determination unit 180 determines that the operating state of the monitoring target 10 is abnormal. After step ST600, the process ends.

[0071] After step ST550, in step ST601, the judgment unit 180 judges that the operating state of the monitoring target 10 is normal. After step ST601, the process ends.

[0072] When processing a time-series signal that is not an action time-series signal, the signal processing device 100 executes processes ST10 to ST500. However, in the above description, "action time-series signal" shall be read as "time-series signal", "action section signal" as "section signal", and "action section candidate signal" as "section candidate signal".

[0073] As described above, according to the first embodiment, the signal processing device 100 includes an acquisition unit 110, a storage unit 130, a candidate selection unit 140, a position determination unit 150, and an extraction unit 160. The acquisition unit 110 acquires a time-series signal. The storage unit 130 pre-stores a template for a section signal to be extracted from the time-series signal. The candidate selection unit 140 selects multiple candidates for the time position at which to extract the section signal from the time-series signal based on the similarity between the time-series signal and the template, and excludes candidates within a predetermined time from the likely candidates selected based on the order of similarity. The position determination unit 150 determines the extraction position based on the remaining likely candidates. The extraction unit 160 extracts the section signal from the time-series signal according to the extraction position. In this way, the signal processing device 100 is configured to select likely candidates based on the similarity between the acquired time-series signal and the template for the section signal, determine the extraction position based on the likely candidates, and extract the section signal. Therefore, the signal to be extracted can be accurately extracted from the time-series signal without any time lag. Specifically, in a time-series signal containing multiple similar types of mechanical knocking sounds that occur intermittently and artificial knocking sounds that occur without any time regularity, it is possible to accurately extract a large mechanical knocking sound from the monitored object 10 without any time lag based on the characteristics of the acquired sound. Furthermore, since the signal to be extracted can be accurately extracted without any time lag, it is possible to improve the accuracy of anomaly detection and deterioration estimation of the air knocker by the autoencoder. Furthermore, it is possible to reduce the capacity of the autoencoder stored in the signal processing device 100.

[0074] Additionally, the room in which the air knocker and dryer are located in a nuclear waste reprocessing plant is contaminated by radiation. For this reason, the room is sealed, and it is undesirable for maintenance personnel at the nuclear waste reprocessing plant to repeatedly enter and exit the room to detect abnormalities in the air knocker. Here, the signal processing device 100 configured as described above can accurately extract the mechanical knocking sound of the air knocker without time lag. Therefore, unmanned and remote detection of abnormalities in the air knocker is possible, thereby improving the safety and work efficiency of plant maintenance personnel.

[0075] Furthermore, the signal processing device 100 configured as described above can accurately extract mechanical tapping sounds of the monitoring target 10 without time lag based on the sound characteristics. Therefore, it is possible to accurately extract mechanical tapping sounds of the monitoring target 10 without time lag, regardless of the trigger signal from the monitoring target 10.

[0076] Furthermore, even intermittent operation sounds that have a certain time regularity, such as the mechanical tapping sound of an air knocker, have fluctuations in the time at which the operation sound occurs. The signal processing device 100 configured as described above can accurately extract not only periodic steady sounds, but also intermittent operation sounds whose time intervals fluctuate, without any time lag. Furthermore, the signal processing device 100 can accurately extract non-stationary sounds without any time lag.

[0077] Furthermore, according to the first embodiment, the candidate selection unit 140 selects, from among multiple candidates, a first likely candidate that has the greatest similarity as the likely candidate, and excludes candidates that are within a predetermined time from the first likely candidate. Therefore, in addition to the effects described above, it is possible to exclude candidates that are within a predetermined time from the first likely candidate and have a lower similarity than the first likely candidate.

[0078] Furthermore, according to the first embodiment, if the remaining candidate that has not been excluded is the first likely candidate, the position determination unit 150 determines the first likely candidate as the extraction position. Therefore, in addition to the effects described above, the candidate with the greatest similarity can be determined as the extraction position, so that the mechanical tapping sound of the monitored object 10 can be extracted more accurately with less time lag.

[0079] Furthermore, according to the first embodiment, when the remaining candidates that have not been excluded include a candidate other than the first likely candidate, the candidate selection unit 140 selects, as the likely candidate, the second likely candidate, which is the candidate with the greatest similarity among the remaining candidates that do not include the first likely candidate. The candidate selection unit 140 excludes candidates that are within a predetermined time from the second likely candidate. Therefore, in addition to the effects described above, it is possible to exclude candidates that are within a predetermined time from the second likely candidate and have a lower similarity than the second likely candidate.

[0080] Furthermore, according to the first embodiment, if the remaining candidates that have not been excluded are the first and second likely candidates, the position determination unit 150 determines the first and second likely candidates as the extraction positions. Therefore, in addition to the effects described above, by utilizing the periodicity of the intermittent operation sound, it is possible to extract the mechanical tapping sound of the monitored object 10 more accurately and without time lag.

[0081] Furthermore, according to the first embodiment, the signal processing device 100 further includes a detection unit 120 that monitors a plurality of power components corresponding to each of a plurality of signals constituting a time-series signal and detects a motion section candidate signal composed of the plurality of signals from the time-series signal based on the plurality of power components. The candidate selection unit 140 calculates the similarity by treating the motion section candidate signal as a time-series signal. In this manner, the signal processing device 100 is configured to detect a motion section candidate signal, which is a part of the time-series signal, based on the plurality of power components from the time-series signal and calculate the similarity between the motion section candidate signal and a template. Therefore, in addition to the above-mentioned effects, the processing load of the signal processing device 100 can be reduced. Furthermore, the reduction in processing load allows the mechanical tapping sounds of the monitored object 10 to be accurately extracted without time lag even in an edge device with smaller computational resources than a server or the like.

[0082] Furthermore, according to the first embodiment, the detection unit 120 calculates the average value and maximum value of a plurality of power components, and detects the above-mentioned section candidate signal when each of the average value and maximum value is equal to or greater than a threshold. Therefore, in addition to the above-mentioned effects, the processing amount can be further reduced because the operation section candidate signal is detected only when each of the average value and maximum value of a plurality of power components is equal to or greater than a threshold.

[0083] Furthermore, according to the first embodiment, the candidate selection unit 140 calculates the similarity between the section candidate signal and a template. The candidate selection unit 140 selects multiple signals obtained by shifting the section candidate signal by a predetermined number of signals. The candidate selection unit 140 calculates the similarity using the multiple signals as the section candidate signal. The candidate selection unit 140 sequentially selects multiple other signals obtained by delaying the section candidate signal and calculates the similarity using the multiple other signals as the section candidate signals. If the largest similarity among the calculated similarities is greater than a threshold, the candidate selection unit 140 calculates the time position of the largest section candidate signal, which is the multiple signals corresponding to the largest similarity, based on the delay of the time position of the largest section candidate signal relative to the time positions of multiple signals shifted by a predetermined number of signals. The candidate selection unit 140 selects the calculated time position as a candidate. Therefore, in addition to the above-mentioned effects, by performing the process of calculating the similarity between the detected section candidate signal and the template while shifting the detected section candidate signal, it is possible to select, as a candidate, the time position of the section candidate signal shifted so as to achieve the largest similarity.

[0084] Furthermore, according to the first embodiment, the candidate selecting unit 140 calculates the cross-correlation coefficient between the section candidate signal and the template as the similarity. Therefore, in addition to the effects described above, the similarity can be calculated from the cross-correlation coefficient.

[0085] Furthermore, according to the first embodiment, the candidate selection unit 140 calculates the time position of the maximum duration candidate signal based on the time positions and delays of multiple signals shifted by a predetermined number of signals. Therefore, in addition to the above-mentioned effects, the time position of the maximum duration candidate signal can be calculated.

[0086] Furthermore, according to the first embodiment, the time-series signal is an operation time-series signal relating to the operation state of the monitoring target 10. The extraction unit 160 extracts an operation section signal, which is a section signal, from the operation time-series signal in accordance with the extraction position. The signal processing device 100 further includes a determination unit 180 that determines the operation state based on the operation section signal. Therefore, in addition to the effects described above, the operation state of the monitoring target 10 is determined based on the time-series signal that has been accurately extracted without any time lag, which makes it possible to suppress extraction of mechanical hammering sounds or artificial hammering sounds other than those of the monitoring target 10 and reduce erroneous determinations.

[0087] Furthermore, according to the first embodiment, the determination unit 180 generates a reconstructed signal by inputting the operation interval signal, which is an operation interval signal when the monitored object 10 is in a normal state, into a trained model that generates a reconstructed signal that is substantially identical to the normal operation interval signal. The determination unit 180 determines the operating state of the monitored object 10 based on the operation interval signal and the reconstructed signal. Therefore, in addition to the effects described above, the operating state of the monitored object 10 can be determined by inputting the operation interval signal into a trained model such as an autoencoder, for example.

[0088] Furthermore, according to the first embodiment, the determination unit 180 calculates the degree of abnormality based on the operation section signal and the reconstruction signal, and determines that the operation state is abnormal if the degree of abnormality is equal to or greater than a threshold. Therefore, in addition to the effects described above, when the operation state of the monitoring target 10 is abnormal, the user can be notified of the abnormality.

[0089] Here, a signal processing device (not shown) according to a comparative example will be described. FIG. 8 is a scatter diagram showing an example of the calculation results of the abnormality degree according to the comparative example. This scatter diagram is an example in which the horizontal axis represents the time at which the abnormality degree of the operation section signal is calculated, and the vertical axis represents the abnormality degree calculated at that time. In FIG. 8, the dotted line T represents the threshold value of the abnormality degree, and the arrow L represents the learning data section of the normal operation section signal. The standard deviation of the abnormality degree calculated based on multiple calculated abnormality degrees is denoted as σ. The threshold is set to, for example, 3σ. In FIG. 8, at a time period prior to the learning data section, a mixture of judgment results indicating an abnormality degree equal to or greater than the threshold and judgment results indicating an abnormality degree below the threshold is present. This mixture indicates that in addition to the mechanical hammering sound of the monitored object, sound signals other than the mechanical hammering sound are also extracted from the time-series signal. The signal processing device according to the comparative example does not have a configuration in which multiple candidates for the time position to be extracted are selected based on the similarity between the acquired time-series signal and a template related to the operation section signal, and the extraction position is determined based on the multiple candidates to extract the operation section signal. For this reason, the signal processing device according to the comparative example cannot accurately extract the mechanical tapping sounds of the monitored object 10 without time lag from the characteristics of the acquired sounds in a time series signal containing multiple types of mechanical tapping sounds and artificial tapping sounds.

[0090] Meanwhile, FIG. 9 is a scatter plot showing an example of the calculation results of the abnormality level according to the first embodiment. In FIG. 9, at a time period prior to the learning data interval, there is no mixture of judgment results indicating an abnormality level equal to or greater than the threshold and judgment results indicating an abnormality level below the threshold. This lack of mixture indicates that the mechanical hammering sounds of the monitored object 10 have been extracted. The signal processing device 100 according to the first embodiment is configured to select multiple candidates for the extraction time position based on the similarity between the acquired time-series signal and a template related to the operation interval signal, determine the extraction position based on the multiple candidates, and extract the operation interval signal. Therefore, the signal processing device 100 according to the first embodiment can accurately extract the mechanical hammering sounds of the monitored object 10 without time lag based on the characteristics of the acquired sounds from a time-series signal containing multiple types of mechanical hammering sounds and artificial hammering sounds.

[0091] FIG. 10 is a diagram for comparing the performance of the signal processing device of the comparative example and the first embodiment. FIG. 10 shows the number of air knocker hits extracted based on sound signals collected by three microphones (first to third microphones) provided near each of the three air knockers (first to third air knockers). The signal processing device 100 according to the first embodiment extracts more hit sounds from each of the three microphones than the signal processing device according to the comparative example. Because the signal processing device 100 according to the first embodiment extracts more hit sounds, it has a higher accuracy in extracting mechanical hit sounds of the monitoring target 10 than the signal processing device according to the comparative example. Therefore, the signal processing device 100 according to the first embodiment can improve the accuracy in extracting mechanical hit sounds of the monitoring target 10.

[0092] Furthermore, according to the first embodiment, the signal processing device 100 further includes a detection unit 120 that monitors a plurality of power components corresponding to each of a plurality of signals constituting the action time-series signal and detects a motion section candidate signal composed of the plurality of signals from the action time-series signal based on the plurality of power components. The candidate selection unit 140 calculates a similarity regarding the action state using the action section candidate signal as the action time-series signal. Therefore, in addition to the above-mentioned effects, the processing load of the signal processing device 100 can be reduced even when determining the action state of the monitoring target 10 based on the action time-series signal.

[0093] (Modification of the first embodiment) The first embodiment may be modified as follows: The modifications may be combined with each other, or may be combined with the following embodiments.

[0094] According to the first embodiment, the detection unit 120 is configured to detect section candidate signals from a time-series signal based on a plurality of power components corresponding to each of a plurality of signals constituting a time-series signal divided into 60 seconds, but this is not limiting. The detection unit 120 may be configured to determine whether the time-series signal is silent or not based on the plurality of power components before detecting a motion section candidate signal, and if the result of the determination indicates silence, omit the detection process and subsequent processes. This configuration can further reduce the processing load of the signal processing device 100.

[0095] Furthermore, the detection unit 120 may be configured to omit the detection process and subsequent processes based on the time at which the time-series signal is acquired. Specifically, for example, the detection process and subsequent processes may be omitted when a maintenance worker at a nuclear waste reprocessing plant is present in a room where an air knocker and a dryer are installed. With such a configuration, the processing load of the signal processing device 100 can be further reduced. Note that the function of omitting the process of detecting an operation interval candidate signal based on a plurality of power components and subsequent processes based on the time at which the time-series signal is acquired may be referred to as an imperfect data removal unit.

[0096] Furthermore, according to the first embodiment, the detecting unit 120 is configured to detect an operation section candidate signal based on the average value and maximum value of the calculated multiple power components, but the present invention is not limited to this. The detecting unit 120 may be configured to monitor power components of a specific frequency or power components of the entire band, and detect a section candidate signal based on these power components.

[0097] (Second embodiment) The second embodiment is a modified example of the first embodiment, and has a configuration in which a sudden wave signal is separated from an operation time series signal, and based on the separated sudden wave signal, a determination is made as to whether the monitored object 10 that generates the sudden wave is normal or abnormal.

[0098] Fig. 11 is a diagram showing an example of a signal processing device 100 according to the second embodiment and its peripheral configuration. As shown in Fig. 11, the signal processing device 100 further includes a separation unit 115, a removal unit 125, a template separation unit 129, a standing wave extraction unit 190, a standing wave model storage unit 195, and a standing wave determination unit 200 in addition to the configuration shown in Fig. 1.

[0099] More specifically, the monitoring object 10 includes a sudden wave monitoring object 11 that generates a sudden wave and a standing wave monitoring object 12 that generates a standing wave. The sudden wave monitoring object 11 is, for example, an air knocker. The air knocker generates a sudden wave signal. The standing wave monitoring object 12 is, for example, a dryer. The dryer generates a standing wave signal. The sudden wave may be called a sudden sound, a hammering sound, a P (Percussive) wave, or the like. The standing wave may be called an H (Harmonic) wave.

[0100] The separation unit 115 separates a sudden wave signal from the movement time-series signal. The separation unit 115 separates a standing wave signal from the movement time-series signal. The separation unit 115 is, for example, a harmonic percussive sound separation (HPSS) unit. The separation unit 115 is an example of a first separation unit.

[0101] Instead of the configuration described above, the detection unit 120 transmits the operation section candidate signal detected from the sudden wave signal to the removal unit 125.

[0102] The removal unit 125 removes a portion of signals corresponding to irregular artificial beat sounds from the motion section candidate signals transmitted from the detection unit based on the plurality of power components. The removal unit 125 selects a portion of signals (frames) from the section candidate signals, and if a predetermined number or more of the plurality of power components corresponding to the portion of signals are equal to or greater than a threshold, the removal unit 125 regards the portion of signals as a portion of signals corresponding to artificial beat sounds and removes the portion of signals from the time-series signal. As described above, in many cases, among 100 power components corresponding to 100 sample signals constituting one frame, there are approximately 0 to 2 power components equal to or greater than 84 dB. In contrast, there are cases where eight or more power components equal to or greater than 84 dB are present among the 100 power components. Artificial beat sounds are characterized by occurring frequently within a short period of time. Therefore, if eight or more power components equal to or greater than 84 dB are present, it can be estimated that artificial beat sounds are occurring. Therefore, if eight or more power components equal to or greater than 84 dB are present, the removal unit 125 regards the frame as a frame corresponding to artificial beat sounds and removes it from the section candidate signals. The removal unit 125 transmits the remaining signals that have not been removed as section candidate signals to the candidate selection unit 140. The removal unit 125 may also be called an artificial hit sound removal unit.

[0103] The template separating unit 129 separates a sudden wave template from the template. The template separating unit 129 separates a standing wave template from the template. The template separating unit 129 is an example of a second separating unit.

[0104] In addition to the above-mentioned configuration, the storage unit 130 stores the sudden wave templates and standing wave templates separated by the template separation unit 129.

[0105] Instead of selecting candidates based on an operation section candidate signal, the candidate selection unit 140 selects candidates based on an operation section candidate signal transmitted from the removal unit 125. Instead of the configuration described above, the candidate selection unit 140 acquires multiple sudden wave templates from the storage unit 130. In addition to the configuration described above, the candidate selection unit 140 uses a sudden wave template separated from a template as a template and calculates a similarity as a sudden wave similarity using the sudden wave signal as a time series signal. Based on the sudden wave similarity, the candidate selection unit 140 selects multiple candidates for the time position at which the sudden wave section signal, which is an operation section signal, is extracted from the sudden wave signal, and excludes, from the multiple candidates, candidates within a predetermined time from the candidates selected based on the order of sudden wave similarity.

[0106] In addition to the above-described configuration, the position determining unit 150 determines the extraction position of the sudden wave signal based on the remaining candidates that have been excluded from the candidates.

[0107] In addition to the above-described configuration, the cutout unit 160 cuts out the sudden wave section signal from the sudden wave signal according to the cutout position.

[0108] In addition to the configuration described above, the model storage unit 170 stores a trained model that generates a normal sudden wave section signal that is approximately identical to a normal sudden wave section signal based on a normal sudden wave section signal, which is a sudden wave section signal when the sudden wave monitoring object 11 is in a normal state.

[0109] In addition to the above-mentioned configuration, the determination unit 180 determines the operating state of the sudden wave monitoring target 11 based on the sudden wave section signal.

[0110] The standing wave extraction unit 190 extracts the sudden wave section signal from the standing wave signal according to the extraction position of the standing wave signal. The extraction position of the standing wave signal is determined, for example, by performing the same processing on the standing wave signal as the processing performed on the sudden wave signal by the detection unit 120, the candidate selection unit 140, and the position determination unit 150. Note that, without being limited to this, for example, the processing performed by the detection unit 120 on the sudden wave signal, in which the maximum value of the multiple power components constituting the extracted frame is calculated based on each of the monitored multiple power components, may be omitted as appropriate.

[0111] The standing wave model storage unit 195 stores a trained model that generates a normal standing wave section signal that is approximately identical to the normal standing wave section signal, based on the normal standing wave section signal, which is the standing wave section signal when the standing wave monitoring object 12 is in a normal state.

[0112] The standing wave determination unit 200 determines the operating state of the standing wave monitoring target 12 based on the standing wave section signal.

[0113] The other configurations are the same as those in the first embodiment.

[0114] Next, an example of the operation of the signal processing device 100 configured as above will be described with reference to the flowchart of FIG.

[0115] First, step ST10 is started in the initial state as described above. However, the storage unit 130 stores in advance the sudden wave templates separated by the template separation unit 129. Also, in the above explanation, "monitoring target 10" is read as "sudden wave monitoring target 11."

[0116] The signal processing device 100 executes the processes of steps ST10 to ST100 in the same manner as described above.

[0117] After step ST100, in step ST150, the separating section 115 separates the sudden wave signal from the motion time-series signal.

[0118] After step ST150, in step ST200A, the same process as in step ST200 described above is executed. The power monitoring unit 121 detects an operation section candidate signal from the sudden wave signal divided every 60 seconds.

[0119] After step ST200A, in step ST249A, if there are eight or more power components of 84 dB or more among the 100 power components corresponding to each of the 100 sample signals constituting one frame, the elimination unit 125 removes the frame from the movement section candidate signals as a frame corresponding to an artificial hitting sound. The elimination unit 125 transmits the remaining multiple signals that have not been removed as movement section candidate signals to the candidate selection unit 140.

[0120] After step ST249A, the signal processing device 100 executes the processes of steps ST250A to ST600A or ST601A in the same manner as the above-described steps ST250 to ST600 or ST601. However, in the description of the above-described steps ST250 to ST600 or ST601, the "action time-series signal" is replaced with the "spontaneous wave signal," the "action section signal" is replaced with the "spontaneous wave section signal," the "template" is replaced with the "spontaneous wave template," and the "similarity" is replaced with the "spontaneous wave similarity."

[0121] The above-described example of operation relates to the processing of the signal processing device 100 for determining the operating state of the sudden wave monitoring target 11. The signal processing device 100 may also determine the operating state of the standing wave monitoring target 12 by executing the processing of steps ST10 to ST600 or ST601 in the same manner as described above. However, in the description of steps ST10 to ST600 or ST601, the "sudden wave signal" should be read as the "standing wave signal," the "sudden wave section signal" should be read as the "standing wave section signal," the "sudden wave template" should be read as the "standing wave template," and the "sudden wave similarity" should be read as the "standing wave similarity." Similarly, the "cutout unit 160" should be read as the "standing wave cutout unit 190," the "model storage unit 170" should be read as the "standing wave model storage unit 195," and the "determination unit 180" should be read as the "standing wave determination unit 200."

[0122] As described above, according to the second embodiment, the monitoring target 10 includes the sudden wave monitoring target 11 that generates a sudden wave. The operation time series signal includes a sudden wave signal related to the operation state of the sudden wave monitoring target 11. The signal processing device 100 further includes a separation unit (first separation unit) that separates the sudden wave signal from the operation time series signal. The cutout unit 160 cuts out a sudden wave section signal, which is an operation section signal, from the separated sudden wave signal according to the cutout position. The determination unit 180 determines the operation state of the sudden wave monitoring target 11 based on the sudden wave section signal. In this way, the configuration is such that the operation state of the sudden wave monitoring target 11 is determined by separating the sudden wave signal from the operation time series signal and cutting out the sudden wave section signal from the sudden wave signal according to the cutout position. Therefore, in addition to the effects of the first embodiment, the operation state of the sudden wave monitoring target 11 can be determined more accurately. Specifically, for example, the operation state of an air knocker can be determined more accurately.

[0123] Supplementally, the separation unit 115 reduces standing waves from the operation time-series signal, thereby increasing the signal-to-noise ratio (SNR) of the sudden wave signal in the operation time-series signal. This makes it possible to extract the sudden wave signal more efficiently, improving the accuracy of extracting the sudden wave signal. Furthermore, since the operation state of the sudden wave monitoring target 11 is determined based on the sudden wave section signal with improved extraction accuracy, it is possible to further suppress the extraction of mechanical hammering sounds and artificial hammering sounds other than those of the sudden wave monitoring target 11, thereby further reducing erroneous determinations.

[0124] FIG. 13 is a scatter plot showing an example of the calculation results of the degree of anomaly according to the second embodiment. The AUC (Area Under the Curve) of the ROC (Receiver Operating Characteristic) curve corresponding to the calculation results in FIG. 9 is compared with the AUC of the ROC curve corresponding to the calculation results in FIG. 13. The closer the AUC of the ROC curve is to 1, the higher the anomaly detection performance. The AUC of the ROC curve (not shown) corresponding to the calculation results in FIG. 9 is 0.993. On the other hand, the AUC of the ROC curve (not shown) corresponding to the calculation results in FIG. 13 is 0.997. As such, the AUC of the ROC curve corresponding to the calculation results of the signal processing device 100 according to the second embodiment is closer to 1, and the anomaly detection performance of the signal processing device 100 according to the second embodiment is improved. Therefore, in addition to the effects of the first embodiment, the operating state of the sudden wave monitoring target 11 can be determined more accurately.

[0125] Moreover, the average precision indicating the accuracy rate of the normal / deteriorated determination corresponding to the determination result in Fig. 9 is 0.995. On the other hand, the average precision corresponding to the determination result in Fig. 9 is 0.999. As such, the average precision corresponding to the determination result of the signal processing device 100 according to the second embodiment is closer to 1, and the anomaly detection performance of the signal processing device 100 according to the second embodiment is improved. Therefore, in addition to the effect of the first embodiment, the operating state of the sudden wave monitoring target 11 can be determined more accurately.

[0126] FIG. 14 is a diagram illustrating a separation unit according to a second embodiment. FIG. 14(a) shows a waveform identical to the example of the waveform of the action time series signal shown in FIG. 7. FIG. 14(b) shows an example of the waveform of a signal obtained by separating a sudden wave signal from the action time series signal. Note that FIG. 14(c) will be described later. In FIGS. 14(a) and 14(b), dotted lines indicate that the time positions in FIG. 14(a) and FIG. 14(b) are the same time positions. In FIG. 14, at time positions P211 and P212, the sudden wave signal has reduced standing waves with time continuity compared to the action time series signal, making the attack (frequency continuity) of the hitting sound at the start of the sudden wave more noticeable. Specifically, as shown in FIGS. 7 and 14, the low-amplitude portions s indicating standing waves between hitting sounds in the sudden wave signal are almost zero, unlike the low-amplitude portions s between hitting sounds in the action time series signal. As a result, the time positions P211 and P212 of the hitting sounds are clearer in the sudden wave signal than in the operation time series signal. In this way, when the time positions of the hitting sounds of the sudden wave are clearer, the accuracy of calculating the cross-correlation coefficient is improved, making it easier to extract candidates. Therefore, in addition to the effect of the first embodiment, the operation state of the sudden wave monitoring target 11 can be determined more accurately.

[0127] Furthermore, according to the second embodiment, the template includes a sudden wave template related to the sudden wave section signal. The signal processing device 100 further includes a template separation unit (second separation unit) that separates the sudden wave template from the template. The candidate selection unit 140 uses the separated sudden wave template as a template and selects multiple candidates for the time position at which to extract the sudden wave section signal from the sudden wave signal based on the sudden wave similarity, which is the similarity calculated using the sudden wave signal as a time-series signal. The candidate selection unit 140 excludes, from the multiple candidates, candidates within a predetermined time from the candidates selected based on the order of sudden wave similarity. The position determination unit 150 determines the extraction position based on the remaining excluded candidates. Therefore, in addition to the above-mentioned effects, the extraction position can be determined based on the sudden wave similarity between the sudden wave signal and the sudden wave template separated from the template.

[0128] According to the second embodiment, the monitoring target 10 includes a standing wave monitoring target 12 that generates standing waves. The operation time-series signal includes a standing wave signal related to the operation state of the standing wave monitoring target 12. The signal processing device 100 further includes a separation unit (first separation unit) that separates the standing wave signal from the operation time-series signal. The cutout unit 160 cuts out a standing wave section signal, which is an operation section signal, from the separated standing wave signal according to the cutout position. The determination unit 180 determines the operation state of the standing wave monitoring target 12 based on the standing wave section signal. In this way, the operation state of the standing wave monitoring target 12 is determined by separating the standing wave signal from the operation time-series signal and cutting out the standing wave section signal from the standing wave signal according to the cutout position. Therefore, in addition to the above-described effects, the operation state of the standing wave monitoring target 12 can be determined in addition to the operation state of the sudden wave monitoring target 11. Specifically, for example, the operation state of a dryer can be determined in addition to the operation state of an air knocker.

[0129] FIG. 14(c) shows an example of the waveform of a signal obtained by separating a standing wave signal from an operation time series signal. In FIGS. 14(a) and 14(c), the dotted lines indicate that the time positions in FIG. 14(a) and FIG. 14(c) are the same. In FIG. 14, at time positions P211 and P212, the standing wave signal does not exhibit the attack characteristics of the impact sound at the beginning of the sudden wave compared to the operation time series signal, and the sudden wave signal is reduced. Therefore, the signal-to-noise ratio of the standing wave signal in the operation time series signal is increased, and in addition to the effects of the first embodiment, the operation state of the standing wave monitoring target 12 can be determined more accurately.

[0130] Moreover, according to the second embodiment, the device further includes a removal unit 125 that removes some signals corresponding to artificial hit sounds from the section candidate signals based on a plurality of power components, and transmits the remaining plurality of signals that have not been removed as section candidate signals to the candidate selection unit 140. Therefore, in addition to the effects described above, the state can be determined based on the section candidate signals from which frames corresponding to frequently occurring artificial hit sounds have been removed, and therefore the operating state of the monitoring target 10 can be determined more accurately.

[0131] (Modification of the second embodiment) The second embodiment may be modified as follows: The modifications may be combined with each other, or may be combined with the following embodiments.

[0132] According to the second embodiment, the operation time series signal is divided every 60 seconds, and a process for determining the cut-out position of the standing wave signal is performed for the time series signal. However, this is not limited to this. For example, the operation time series signal may be divided into minutes, which is longer than 60 seconds, and a process for determining the cut-out position of the standing wave signal for the time series signal may be performed. Therefore, according to this modification, a process for determining the cut-out position can be performed for the operation time series signal divided into lengths according to the characteristics of the standing wave, so that more efficient processing can be expected.

[0133] Furthermore, according to the second embodiment, the motion time-series signal divided every 60 seconds is divided into frames each consisting of 100 sample signals, and the average value and maximum value of the sum of squares of the amplitudes of the 100 sample signals constituting a frame are calculated, but this is not limiting. The number of sample signals constituting a frame may be less than or greater than 100. Therefore, according to this modification, the motion time-series signal is divided into frames each consisting of a number of sample signals according to the characteristics of the standing wave, and the average value and maximum value of the sum of squares of the amplitudes of the sample signals constituting the frame can be calculated, so that more efficient processing can be expected.

[0134] (Third embodiment) The third embodiment is a specific example of the first or second embodiment, and is a form in which the above-described signal processing device 100 is realized by a computer.

[0135] 15 is a block diagram illustrating a hardware configuration of a signal processing device 100 according to the third embodiment. The signal processing device 100 includes, as hardware, a CPU (Central Processing Unit) 31, a RAM (Random Access Memory) 32, a program memory 33, an auxiliary storage device 34, and an input / output interface 35. The CPU 31 communicates with the RAM 32, the program memory 33, the auxiliary storage device 34, and the input / output interface 35 via a bus.

[0136] The CPU 31 is an example of a general-purpose processor. The RAM 32 is used by the CPU 31 as a working memory. The RAM 32 includes a volatile memory such as a Synchronous Dynamic Random Access Memory (SDRAM). The program memory 33 stores a program for implementing each unit according to each embodiment. This program may be, for example, a program for causing a computer to implement each function of the signal processing device 100 described above. The program memory 33 may be, for example, a read-only memory (ROM), a part of the auxiliary storage device 34, or a combination thereof. The auxiliary storage device 34 stores data non-temporarily. The auxiliary storage device 34 includes a non-volatile memory such as a hard disc drive (HDD) or a solid state drive (SSD).

[0137] The input / output interface 35 is an interface for connecting to other devices, and is used to connect to the recorder 20, for example.

[0138] The program stored in the program memory 33 includes computer-executable instructions. When executed by the CPU 31, which is a processing circuit, the program (computer-executable instructions) causes the CPU 31 to perform a predetermined process. For example, when executed by the CPU 31, the program causes the CPU 31 to perform the series of processes described with reference to the components in FIGS. 1 and 11 . For example, when executed by the CPU 31, the computer-executable instructions included in the program cause the CPU 31 to perform a signal processing method. The signal processing method may include steps corresponding to the functions of the signal processing device 100 described above. For example, the signal processing method may include acquiring a time-series signal, storing in advance a template for a section signal to be extracted from the time-series signal, selecting multiple candidates for the time position at which to extract the section signal from the time-series signal based on the similarity between the time-series signal and the template, excluding, from the multiple candidates, a candidate within a predetermined time from the candidates selected based on the order of similarity, determining an extraction position based on the remaining candidates, and extracting the section signal from the time-series signal according to the extraction position. Furthermore, the signal processing method may include the steps shown in FIGS. 3, 4, 5, 6, and 12 as appropriate.

[0139] The program may be provided to the signal processing device 100 in a state where it is stored in a computer-readable storage medium. In this case, for example, the signal processing device 100 may further include a drive (not shown) for reading data from the storage medium and acquire the program from the storage medium. As the storage medium, for example, a magnetic disk, an optical disk (CD-ROM, CD-R, DVD-ROM, DVD-R, etc.), a magneto-optical disk (MO, etc.), a semiconductor memory, etc. may be used as appropriate. The storage medium may also be referred to as a non-transitory computer-readable storage medium. Alternatively, the program may be stored in a server on a communication network, and the signal processing device 100 may download the program from the server using the input / output interface 35.

[0140] The processing circuit that executes the program is not limited to a general-purpose hardware processor such as CPU 31, but may also be a dedicated hardware processor such as an ASIC (Application Specific Integrated Circuit). The term processing circuit (processing unit) includes at least one general-purpose hardware processor, at least one dedicated hardware processor, or a combination of at least one general-purpose hardware processor and at least one dedicated hardware processor. In the example shown in FIG. 15, CPU 31, RAM 32, and program memory 33 correspond to the processing circuit.

[0141] According to at least one of the embodiments described above, it is possible to accurately extract a signal to be extracted from a time-series signal without time lag. This also applies to at least one of the modified examples described above.

[0142] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, as well as within the scope of the invention described in the claims and their equivalents. [Explanation of symbols]

[0143] 10. Monitoring Targets 11. Targets for sudden wave monitoring 12 Standing wave monitoring target 20 Recorder 31 CPU 32 RAM 33 Program Memory 34 Auxiliary storage device 35 Input / Output Interface 100 Signal processing device 110 Acquisition Department 115 Separation section 120 Detector 121 Power Monitoring Unit 122 Candidate detection unit 125 Removal section 129 Template Separator 130 Storage area 140 Candidate Selection Section 141 Score Calculation Section 142 Score sorting section 143 Search Department 150 Positioning section 160 Cutout 170 Model Storage Unit 180 Judgment section 190 Standing wave cutout 195 Standing wave model storage area 200 Standing wave determination unit C1 1st likely candidate C2 2nd most likely candidate L arrow Md1 machine learning model Md10 pre-trained model P111~P322 Time position S Operation section signal s Low amplitude part T dotted line t1, t2 times

Claims

1. an acquisition unit that acquires a time series signal; a storage unit that stores in advance a template relating to a section signal to be extracted from the time series signal; a candidate selection unit that selects a plurality of candidates for a time position at which the section signal is extracted from the time series signal based on the similarity between the time series signal and the template, and excludes, from the plurality of candidates, candidates that are within a predetermined time from likely candidates selected based on the order of the similarity; a position determination unit that determines a cutout position based on the remaining likely candidates that have not been excluded; a cutout unit that cuts out the section signal from the time-series signal in accordance with the cutout position; A signal processing device comprising:

2. 2. The signal processing device according to claim 1, wherein the candidate selection unit selects a first likely candidate from the plurality of candidates that has the greatest similarity as the likely candidate, and excludes candidates that are within a predetermined time from the first likely candidate.

3. The signal processing device according to claim 2 , wherein, when the remaining candidate that has not been excluded is the first likely candidate, the position determining unit determines the first likely candidate as the extraction position.

4. 3. The signal processing device according to claim 2, wherein, when the remaining candidates that have not been excluded include a candidate other than the first likely candidate, the candidate selection unit selects a second likely candidate from the remaining candidates that do not include the first likely candidate, which is the candidate with the greatest similarity, as the likely candidate, and excludes candidates that are within a predetermined time period from the second likely candidate.

5. 5. The signal processing device according to claim 4, wherein, when the remaining candidates that have not been excluded are the first likely candidate and the second likely candidate, the position determination unit determines the first likely candidate and the second likely candidate as the extraction positions.

6. a detection unit that monitors a plurality of power components corresponding to each of a plurality of signals constituting the time-series signal, and detects, from the time-series signal, a section candidate signal constituted by the plurality of signals based on the plurality of power components; the candidate selection unit calculates the similarity using the section candidate signal as the time-series signal; The signal processing device according to any one of claims 1 to 5.

7. The signal processing device according to claim 6 , wherein the detection section calculates an average value and a maximum value of the plurality of power components, and detects the section candidate signal when each of the average value and the maximum value is equal to or greater than a threshold value.

8. 7. The signal processing device according to claim 6, wherein the candidate selection unit sequentially executes processes of selecting a plurality of signals obtained by shifting the section candidate signals by a predetermined number of signals, calculating the similarity using the plurality of signals as the section candidate signals, selecting another plurality of signals obtained by delaying the section candidate signals, and calculating the similarity using the other plurality of signals as the section candidate signals, and if a maximum similarity among the calculated similarities is greater than a threshold, calculating a time position of a maximum section candidate signal, which is a plurality of signals corresponding to the maximum similarity, based on a delay of the time position of the maximum section candidate signal relative to the time positions of the plurality of signals shifted by the predetermined number of signals, and selecting the calculated time position as the candidate.

9. The signal processing device according to claim 8 , wherein the candidate selection unit calculates a cross-correlation coefficient between the section candidate signal and the template as the degree of similarity.

10. The signal processing device according to claim 8 , wherein the candidate selection unit calculates the time position of the maximum section candidate signal based on the time positions of the plurality of signals shifted by the predetermined number of signals and the delay.

11. Further comprising a determination unit, the time-series signal is an operational time-series signal relating to the operational state of a monitoring target, the cutout unit cuts out a motion section signal, which is the section signal, from the motion time-series signal in accordance with the cutout position; The determination unit determines the motion state based on the motion section signal. The signal processing device according to claim 1 .

12. the determination unit generates a reconstructed signal by inputting a normal operation interval signal, which is an operation interval signal when the monitoring target is in a normal state, into a trained model that generates a reconstructed signal that is substantially identical to the normal operation interval signal, and makes the determination based on the operation interval signal and the reconstructed signal. The signal processing device according to claim 11 .

13. The signal processing device according to claim 12 , wherein the determination unit calculates a degree of abnormality based on the operation section signal and the reconstructed signal, and determines that the operation state is abnormal if the degree of abnormality is equal to or greater than a threshold value.

14. a detection unit that monitors a plurality of power components corresponding to each of a plurality of signals constituting the motion time-series signal, and detects a motion section candidate signal composed of the plurality of signals from the motion time-series signal based on the plurality of power components; The signal processing device according to claim 11 , wherein the candidate selection unit calculates a similarity regarding the motion state using the motion section candidate signal as the motion time-series signal.

15. Further comprising a first separation section, The monitoring target includes a sudden wave monitoring target that generates a sudden wave, The operation time series signal includes a sudden wave signal related to the operation state of the object to be monitored for sudden waves, The first separation unit separates the sudden wave signal from the motion time-series signal, The cutout unit cuts out a sudden wave section signal, which is the operation section signal, from the separated sudden wave signal according to the cutout position, The determination unit determines the operating state of the object to be monitored for sudden waves based on the sudden wave section signal. The signal processing device according to claim 11 .

16. Further provided with a second separation section, The template includes a burst wave template related to the burst wave section signal, The second separation unit separates the burst wave template from the template, The candidate selection unit uses the separated sudden wave template as the template, and selects a plurality of candidates for a time position at which the sudden wave section signal is extracted from the sudden wave signal based on the sudden wave similarity, which is the similarity calculated from the sudden wave signal as the time series signal, and excludes candidates within a predetermined time from the candidates selected based on the order of the sudden wave similarity from the plurality of candidates; the position determination unit determines the extraction position based on the remaining candidates that have been excluded. The signal processing device according to claim 15.

17. Further comprising a first separation section, The monitoring target includes a standing wave monitoring target that generates a standing wave, The operation time series signal includes a standing wave signal related to the operation state of the standing wave monitoring target, The first separation unit separates the standing wave signal from the operation time series signal, The cutout unit cuts out the standing wave section signal, which is the operation section signal, from the separated standing wave signal according to the cutout position, The determination unit determines the operating state of the standing wave monitoring target based on the standing wave section signal. The signal processing device according to claim 11 .

18. 7. The signal processing device according to claim 6, further comprising a removal unit that removes some signals corresponding to artificial hit sounds from the section candidate signals based on the plurality of power components, and transmits the remaining plurality of signals that have not been removed as the section candidate signals to the candidate selection unit.

19. Obtaining a time series signal; storing in advance a template relating to a section signal to be extracted from the time-series signal; selecting a plurality of candidates for a time position at which the section signal is extracted from the time series signal based on the similarity between the time series signal and the template, and excluding, from the plurality of candidates, a candidate within a predetermined time from a likely candidate selected based on the order of the similarity; determining an extraction position based on the remaining likely candidates that have not been excluded; extracting the section signal from the time-series signal according to the extraction position; A signal processing method comprising:

20. The function of acquiring time series signals; a function of storing in advance a template relating to a section signal to be extracted from the time-series signal; a function of selecting a plurality of candidates for a time position at which the section signal is extracted from the time series signal based on the similarity between the time series signal and the template, and excluding, from the plurality of candidates, candidates within a predetermined time from likely candidates selected based on the order of the similarity; a function of determining a cutout position based on the remaining promising candidates that have not been excluded; a function of extracting the section signal from the time-series signal according to the extraction position; A program that makes the computer realize the above.

Images