JP6882397B2 - Abnormal noise detection system, equipment, method and program - Google Patents

Description

The present invention relates to a system or the like that detects an abnormal section from a predetermined time series signal.

Traditionally, industrial products such as automobiles and electric wires have been inspected for abnormalities before shipping or during periodic inspections. As one of the inspection methods, for example, an industrial product that is an inspection target is operated or an external force (vibration, etc.) is applied to put it in a specific state, and various time-series signals generated in that state are displayed. There is a method of estimating the presence or absence of an abnormality in the inspection object based on the properties.

Patent Document 1 discloses a method of estimating the presence or absence of an abnormality in an inspection object from the presence or absence of abnormal noise by focusing on a sound signal emitted from the inspection object as a time-series signal. More specifically, in this example, the presence or absence of abnormal noise is identified based on the abnormal noise sample sound including abnormal noise and the non-abnormal noise sample sound not containing abnormal noise, whereby the abnormality of the inspection object is identified. Is estimated.

That is, in the conventional method, when specifying an abnormal section in a predetermined time-series signal, first, a sample corresponding to the abnormality is specified as a pretreatment, and in the case of sound, an abnormal sound sample is specified, and then the abnormal sound sample is specified. The abnormal section was identified in the comparison. Conventionally, these operations have been performed manually by an operator.

Japanese Unexamined Patent Publication No. 2015-175770

However, it is a lot of work to identify a sample showing an abnormality from a large amount of data to be inspected, and it may be difficult to judge whether or not the sample shows an abnormality in the first place. As a result, the work of detecting the abnormal section from the time-series signal acquired from the inspection object has become an excessive burden on the operator.

The present invention has been made under the above-mentioned technical background, and an object of the present invention is to reduce the burden on the operator when detecting an abnormal section from a time-series signal acquired from an inspection object. To do.

Still other objects of the invention, as well as effects, will be readily understood by those skilled in the art by reference to the following description of the specification.

The above-mentioned technical problems can be solved by an abnormality detection system, a method, and a program having the following configurations.

That is, the abnormality detection system according to the present invention has a time-series signal acquisition unit that acquires a time-series signal generated by placing an inspection object in a specific state, and a predetermined feature based on the time-series signal. A feature amount generation unit that generates a quantity, a clustering processing unit that clusters the feature quantity based on a predetermined clustering algorithm, and a detection unit that detects an abnormal section in the time series signal based on the result of the clustering. , Is equipped.

According to such a configuration, since the presence or absence of an abnormality in the time-series signal related to the inspection object is detected based on the clustering result, the abnormality can be detected by unsupervised learning based on the time-series signal itself. As a result, it is not necessary to prepare a sample showing an abnormality and to compare it with the sample, so that the burden on the operator who performs the abnormality detection work is reduced. Further, since clustering is performed after generating a predetermined feature amount from the time series signal, the accuracy of clustering is improved. The specific state includes, for example, a state in which the inspection object is physically operated, a state in which the inspection object is electrically operated, a state in which an external force such as vibration is applied, and the like. Further, the time-series signals generated by placing the inspection object in a specific state include, for example, time-series sound signals and sensor signals acquired by a microphone or various sensors.

The time-series signal may include a time-series sound signal which is a time-series signal relating to a sound generated from the inspection object.

According to such a configuration, since the abnormality is detected based on the sound generated from the inspection object, the inspection can be performed non-destructively. Further, since it is not necessary to prepare a sound sample in the case of an abnormal state and to compare the sound sample with the sound sample, the burden on the operator who performs the abnormality detection work is reduced.

The time-series signal may further include a time-series sensor signal which is a time-series signal related to sensor information indicating an operating state of the inspection object.

According to such a configuration, not only the sound emitted from the object but also the sensor information (for example, acceleration) indicating the operating state of the object that may have a causal relationship with the abnormality is utilized to detect the abnormality accurately. Can be further improved.

The feature amount generation unit further includes a frequency region signal generation unit that converts the time-series sound signal and the time-series sensor signal into a frequency region to generate a frequency region sound signal and a frequency region sensor signal. Psychoacoustic evaluation features that generate loudness features related to sound loudness, sharpness features related to sound sharpness, fluctuation intensity features related to sound fluctuation intensity, and roughness features related to sound roughness based on frequency region sound signals. It may include a quantity generation unit and an MFCC feature quantity generation unit that generates an MFCC feature quantity including a mel frequency Keptram coefficient based on the frequency region sound signal and the frequency region sensor signal.

According to such a configuration, since the feature amount in consideration of the human hearing characteristic is used for the sound feature, it is possible to improve the detection accuracy of the abnormality accompanied by the abnormal sound which is particularly easy for the human to perceive.

The feature amount generation unit further includes an input neuron, an output neuron, and an intermediate neuron between them, and is based on a trained neural network trained to function as an autoencoder, and the frequency domain sound signal and the frequency domain. It may be provided with a coded feature amount generation unit that outputs the firing value of the intermediate neuron when the sensor signal is input as a coded feature amount.

According to such a configuration, the clustering accuracy can be improved by utilizing the property that the feature is compressed in the intermediate neuron (the neuron in the hidden layer), and the abnormality detection accuracy can be further improved.

The inspection object is a rotating body, the time-series sensor signal includes a time-series rotation speed signal of the rotating body, and the feature amount generator further comprises the time-series rotation speed signal and the frequency domain sound signal. Based on the above, a rotation order feature quantity generation processing unit that generates a predetermined rotation order as a feature quantity may be provided.

According to such a configuration, since the rotation speed indicating the rotation state is converted into the rotation order and used as the feature quantity, the abnormality detection accuracy in the rotating body can be improved.

The inspection object may be a rotating body, and the time-series sensor signal may be a time-series signal related to acceleration, torque, and rotation speed.

According to such a configuration, abnormality detection of the rotating body can be performed with high accuracy.

The clustering algorithm may include a first clustering algorithm that predetermines the number of clusters and a second clustering algorithm that does not predetermine the number of clusters.

According to such a configuration, it is possible to deal with both the case where the number of clusters can be predicted to some extent and the case where the number of clusters is not, and it is possible to perform anomaly detection more robustly.

The first clustering algorithm may be one or both of k-means clustering and GMM, and the second clustering algorithm may be DBSCAN.

According to such a configuration, the abnormality detection accuracy can be further improved by using an algorithm suitable for abnormal noise detection.

The abnormality detection system further includes an abnormality detection mode acquisition unit that acquires an abnormality detection mode, a detection basis determination unit that determines the feature amount and the clustering algorithm that are the basis of detection based on the abnormality detection mode. It may be provided with.

According to such a configuration, since the clustering algorithm which is the basis of the detection is changed according to the abnormality type, the target abnormality can be detected more accurately.

The abnormality detection system further includes a graph output data generation unit that generates display data for displaying a graph of the time-series sound signal having an abnormality section specified as a result of abnormality detection in the detection unit. There may be.

According to such a configuration, since the abnormal portion can be visually confirmed through the graph, the identification of the abnormal portion becomes more intuitive and easy.

The feature amount generation unit further converts the time-series sound signal into a frequency region to generate a frequency region sound signal, and a loudness regarding sound loudness based on the frequency region sound signal. It may be provided with a feature amount, a sharpness feature amount related to sound sharpness, a fluctuation intensity feature amount related to sound fluctuation intensity, and a psychoacoustic evaluation feature amount generation unit that generates a roughness feature amount related to sound roughness. ..

According to such a configuration, since the feature amount in consideration of the human hearing characteristic is used for the sound feature, it is possible to improve the detection accuracy of an abnormality accompanied by an abnormal sound that is easily perceived by a human.

The abnormality detection system further includes a clustering accuracy determination unit that determines the clustering accuracy based on the result of the clustering, and the detection unit is determined by the clustering accuracy determination unit to have an accuracy equal to or higher than a predetermined reference. Based on the result of the clustering, the abnormal section in the time series signal may be detected.

According to such a configuration, abnormality detection can be performed based on the clustering result in which the clustering accuracy is equal to or higher than a predetermined reference value, so that more appropriate abnormality detection can be performed. In addition, the processing time can be shortened because the number of data used for abnormality detection is reduced.

The clustering accuracy determination unit may determine the accuracy of the clustering result based on the silhouette coefficient calculated for the clustering result.

According to such a configuration, it is possible to determine whether or not each data is properly classified based on the degree of cluster cohesion and the degree of divergence.

The clustering accuracy determination unit may determine the accuracy of the clustering result based on the ratio of outliers of the clustering result.

According to such a configuration, it is possible to determine whether or not each data is properly classified according to the amount of outliers.

The present invention can also be conceived as an abnormality detection device. That is, the abnormality detection device according to the present invention has a time-series signal acquisition unit that acquires a time-series signal generated by placing an inspection object in a specific state, and a predetermined feature based on the time-series signal. A feature amount generation unit that generates a quantity, a clustering processing unit that clusters the feature quantity based on a predetermined clustering algorithm, and a detection unit that detects an abnormal section in the time series signal based on the result of the clustering. , Is equipped.

The present invention can also be considered as an abnormality detection method. That is, the abnormality detection method according to the present invention has a predetermined feature based on a time-series signal acquisition step of acquiring a time-series signal generated by placing an inspection object in a specific state and the time-series signal. A feature quantity generation step for generating a quantity, a clustering processing step for clustering the feature quantity based on a predetermined clustering algorithm, and a detection step for detecting an abnormal section in the time series signal based on the result of the clustering. , Is equipped.

The present invention can also be thought of as an abnormality detection program. That is, the abnormality detection program according to the present invention is based on a time-series signal acquisition unit that acquires a time-series signal generated by placing an inspection object in a specific state and the time-series signal. A feature amount generation unit that generates a predetermined feature amount, a clustering processing unit that clusters the feature amount based on a predetermined clustering algorithm, and an abnormal section in the time series signal are detected based on the result of the clustering. It functions as an abnormality detection device including a detection unit.

The present invention can also be thought of as an abnormal noise detection system. That is, the abnormality detection system according to the present invention transmits a time-series sound signal which is a time-series signal related to the sound generated from the drive train and a time-series sensor signal which is a time-series signal related to sensor information indicating the operating state of the drive train. A time-series signal acquisition unit to be acquired, a feature amount generation unit that generates a predetermined feature amount based on the time-series signal, a clustering processing unit that clusters the feature amount based on a predetermined clustering algorithm, and the clustering. Based on the result of the above, the detection unit for detecting an abnormal section in the time-series sound signal is provided.

According to the present invention, it is possible to reduce the burden on the operator when detecting an abnormal section from a time-series signal acquired from an inspection object.

FIG. 1 is an overall configuration diagram of a data measurement system. FIG. 2 is a configuration diagram relating to the abnormality detection device. FIG. 3 is a general flowchart (first embodiment) regarding the operation of the abnormality detection device. FIG. 4 is a functional block diagram (first embodiment). FIG. 5 is an explanatory diagram relating to the detection process of the abnormal section. FIG. 6 is a functional block diagram (second embodiment). FIG. 7 is a general flowchart (second embodiment) regarding the operation of the abnormality detection device. FIG. 8 is a detailed flowchart of the data selection process. FIG. 9 is an explanatory diagram showing the distribution of silhouette scores.

Hereinafter, an embodiment of the abnormality detection system, apparatus, method, and program according to the present invention will be described in detail with reference to the accompanying drawings.

<1. First Embodiment>
The first embodiment of the present invention will be described with reference to FIGS. 1 to 5. In the first embodiment, a drive train (power transmission mechanism) that transmits rotational power from an automobile engine is set as an inspection object, and time-series data of sounds generated from the inspection object and rotational acceleration of the inspection object are used. , Torque, rotation speed, and other time-series sensor data are measured, and an abnormality section is detected based on these data, thereby estimating an abnormality in the drive train.

<1.1 Configuration>
First, the configurations of the data measurement system 50 and the abnormality detection device 10 according to the present embodiment will be described with reference to FIGS. 1 and 2.

<1.1.1 Data measurement system configuration>
FIG. 1 shows the overall configuration of a data measurement system 50 for measuring time-series data related to the operating state of the drive train, which is the inspection object 51.

As is clear from the figure, the data measurement system 50 includes an inspection object 51, a front end 53, a measurement system 54, a control server 52, and a storage device 55. The control server 52 controls the engine connected to the drive train according to the automatic measurement program. Specifically, automatic measurement is performed by generating a predetermined test pattern, setting measurement conditions, and the like. Further, information on the measurement conditions and the like is transmitted to the measurement system 54 and stored in the storage device 55. The front end 53 reads out time-series signals related to sound, acceleration, rotation speed, and torque acquired from a microphone (not shown) installed on the inspection object 51 and various sensors as each channel signal and outputs the time-series signals to the measurement system 54. To do. The measurement system 54 performs predetermined preprocessing such as setting of a sampling frequency and a channel on the front end 53. In addition, the data acquired via the front end is stored in the storage device 55, and the predetermined data is post-processed.

That is, by operating the data measurement system 50, time-series signals relating to sound, acceleration, rotation speed, and torque acquired from a microphone (not shown) and various sensors installed on the drive train, which is the inspection object 51, are stored. It will be accumulated in the device 55.

<1.1.2 Configuration of anomaly detection device>
FIG. 2 is a configuration diagram of an abnormality detection device 10 for detecting an abnormality from the data measured by the data measurement system 50.

As is clear from the figure, the abnormality detection device 10 includes a control unit 1, a display output unit 2, a storage unit 3, an operation input unit 4, a voice output unit 5, a communication unit 6, and an I / O unit 7 inside. ing. The control unit 1 is composed of a CPU and the like, and executes execution processing of various programs. The display output unit 2 performs image output processing for displaying on a connected display or the like (not shown). The storage unit 3 is composed of a ROM, a RAM, or the like, and stores various programs and data. The operation input unit 4 processes the operation signal input via the keyboard or the like. The audio output unit 5 performs a process of outputting audio to an audio output device such as a speaker (not shown). The communication unit 6 is composed of a communication unit or the like, and performs communication processing with an external device. The I / O unit 7 performs input / output processing by connecting to an external device.

That is, the abnormality detection device 10 sequentially reads out various time-series data from the storage device 55 in which various time-series data measured by the data measurement system 50 is stored, or reads out various time-series data to the storage unit 3 in advance, and controls the control unit. In step 1, the detection process of whether or not an abnormal section is included in the simultaneous series data is performed.

In the present embodiment, the data measurement system 50 and the abnormality detection device 10 are configured as described above, but the present invention is not limited to such a configuration, and various known configurations can be adopted. Therefore, for example, instead of the storage device 55, it may be configured to store directly in the abnormality detection device 10.

<1.2 Operation of abnormality detection device>
Next, the operation of the abnormality detection device 10 will be described with reference to FIGS. 3 to 5.

<1.2.1 Acquisition of time series signals>
FIG. 3 is a general flowchart relating to the operation of the abnormality detection device 10. Further, FIG. 4 is a functional block diagram focusing on the function of the abnormality detection device 1.

As is clear from the figure, when the processing is started, the time-series signals related to sound, the acceleration, the torque, and the time-series signals related to the rotation speed are read out from the storage device 55 or the storage unit 3 to the control unit 1, respectively (S1). .. That is, the time-series signal related to sound is the sound signal time-series data acquisition unit 111, the time-series signal related to acceleration is the acceleration signal time-series data acquisition unit 112, the time-series signal related to torque is the torque signal time-series data acquisition unit 113, and the number of revolutions. The time-series signal relating to the above is read into the control unit 1 via the rotation number signal time-series data acquisition unit 114. At this time, the parameters that define various operations described later are also read via the parameter acquisition unit 115.

<1.2.2 Conversion process to spectrogram>
Each time-series signal read out to the control unit 1 is converted into a frequency domain (spectral conversion) in a predetermined unit, and a spectrogram is generated (S2). Specifically, the time-series signal related to sound is converted into a sound signal in the frequency region via the spectrogram conversion unit 131, and the time-series signal related to acceleration is converted into an acceleration signal in the frequency region via the spectrogram conversion unit 132. The time-series signal related to torque is converted into a torque signal in the frequency region via the spectrogram conversion unit 133, and the time-series signal related to the rotation speed is converted into a rotation speed signal in the frequency region via the spectrogram conversion unit 134.

<1.2.3 Feature generation process>
After performing spectrogram conversion processing (S2) for each time-series signal, processing for generating features is performed based on them (S3). In the present embodiment, four feature quantities, MFCC feature quantity, psychoacoustic evaluation feature quantity, rotation order feature quantity, and coded feature quantity, are adopted as the feature quantity.

<MFCC features>
The MFCC (Mel-Frequency Cepstrum Cooperatives) feature quantity is an MFCC feature quantity generation process in which the sound, acceleration, torque, and rotation frequency signals in the frequency domain after spectral gram conversion (S2) are input. Generated in part 135.

The MFCC feature quantity is a feature quantity related to cepstrum to which a scale obtained from the perceptual characteristics of human sound is applied. In the present embodiment, the MFCC feature quantity includes not only the mel frequency cepstrum coefficient but also its derivative value and its energy. Specifically, the MFCC feature amount includes not only the 26-dimensional Melkeptrum coefficient in the present embodiment, but also ΔMFCC (26 dimensions) obtained by differentiating it once and ΔΔMFCC (26 dimensions) obtained by differentiating it twice. Further, the root mean square energy (RMSE (Root Mean Squared Energy)) of the corresponding time-series signal, its first derivative (ΔRMSE), and its second derivative (ΔΔRMSE) are included one dimension each, for a total of 81 dimensions (ΔΔRMSE). It is a feature amount of 26 × 3 + 3).

According to such a configuration, since the feature amount in consideration of the human hearing characteristic is used, it is possible to improve the detection accuracy of the abnormal section that is easily perceived by the human.

<Psychoacoustic evaluation features>
The psychoacoustic evaluation feature amount is generated by the psychoacoustic evaluation feature amount generation unit 136 by inputting a sound signal in the frequency domain.

Psychoacoustic evaluation features include sound quality evaluation parameters that quantify the hearing of sound, such as loudness ([sone]) for loudness, sharpness ([acum]) for pitch, and slow sound. It contains four types of feature quantities corresponding to the fluctuation intensity ([vacil]) related to the fluctuation and the roughness ([asper]) related to the roughness of the sound. Specifically, the psychoacoustic evaluation feature values are the average value (AVE), the one-time differentiation of the average value (ΔAVE), and the average value twice for each of the four feature quantities of loudness, sharpness, fluctuation intensity, and roughness, respectively. Defined as a total of 24 dimensions (4 x 6) features including the calculated values of the differential (ΔΔAVE), standard deviation (STDEV), standard deviation single differential (ΔSTDEV), and standard deviation double differential (ΔΔSTDEV). To.

According to such a configuration, since the feature amount in consideration of the audibility of the sound is used, it is possible to improve the detection accuracy of an abnormality accompanied by an abnormal sound that is particularly noticeable to humans.

<Rotation order feature amount>
The rotation order feature amount is generated by the rotation order feature amount generation unit 137 by inputting the time series rotation speed signal and the sound signal in the frequency domain.

The rotation order feature amount is a one-dimensional feature amount extracted based on the rotation order tracking analysis. The rotation order tracking is to find the relationship between the rotation speed ([RPM]) of the inspection object and the sound signal ([Hz]) in the frequency domain at that time, and from there, a peculiar signal component based on the signal strength. Is detected, and the corresponding rotation order is extracted as a feature quantity, which is the rotation order feature quantity.

According to such a configuration, since the rotation speed indicating the rotation state is converted into the rotation order and used as the feature quantity, the abnormality detection accuracy in the rotating body can be improved.

<Coded features>
The coded feature amount is generated by the coded feature amount generation processing unit 138, respectively, with the sound, acceleration, torque, and rotation speed signals in the frequency domain as inputs.

The coded feature amount is a feature amount corresponding to the firing value of the neurons in the intermediate layer of the autoencoder using the artificial neural network composed of the input layer, the intermediate layer (hidden layer) and the output layer. It is known that an autoencoder is acquired by supervised learning of a neural network using the same supervised data as an input, and features corresponding to the input are expressed in neurons in the intermediate layer. That is, feature extraction or dimensional compression can be performed using the firing values of neurons in the middle layer.

In the present embodiment, the number of dimensions of the input / output layer is 1025 per frame, and the number of dimensions of the intermediate layer (output dimension) is 32. That is, 32 dimensional features are generated corresponding to each of the sound, acceleration, torque, and rotation speed signals in the frequency domain. Further, in the present embodiment, a feature amount (one-dimensional or multiple-dimensional) corresponding to the firing value of one or a plurality of intermediate layer neurons having a dynamic range equal to or higher than a predetermined threshold value is further added as the coded feature amount.

According to such a configuration, the clustering accuracy can be improved by utilizing the property that the feature is compressed in the intermediate layer, and the abnormality detection accuracy can be improved.

<12.4 Clustering process>
When each feature amount is generated (S3), the clustering processing unit 139 then performs a clustering process for each feature amount (S4). In the present embodiment, clustering processing of each feature amount is performed using three clustering methods of k-means method, GMM (Gaussian mixture model), and DBSCAN (Density-based spatial cruise of applications with noise).

The k-means clustering method is a clustering method in which each sample belongs to any one cluster by setting the number of clusters in advance, and parameters such as the number of clusters are set in advance by the parameter acquisition unit 115 or the like. GMM is a clustering method in which each sample stochastically belongs to any one cluster by setting the number of clusters in advance, and parameters such as the number of clusters are set in advance by the parameter acquisition unit 115 or the like. DBSCAN is a clustering method that automatically determines the number of clusters based on the density of sample points, and parameters such as radius (ε) are set in advance by the parameter acquisition unit 115 or the like.

According to such a configuration, the presence or absence of an abnormality in the time-series signal acquired from the inspection object is detected based on the clustering result, so that the abnormality can be detected by unsupervised learning based on the time-series signal itself. it can. As a result, it is not necessary to prepare a sample showing an abnormality and to compare it with the sample, so that the burden on the operator who performs the abnormality detection work is reduced. Further, since clustering is performed after generating a predetermined feature amount from the time series signal, the accuracy of clustering is improved.

<1.2.5 Abnormality detection mode acquisition process>
After the clustering process (S4) for each feature amount is completed, the abnormality detection mode acquisition process is performed (S5). The abnormality detection mode set by the user or the like and stored in the storage unit 3 is read out by the abnormality detection mode acquisition unit 116 and output to the abnormality detection mode setting unit 141. In response to this output, the abnormality detection mode setting unit 141 sets the abnormality detection mode for the abnormality detection unit 140.

The anomaly detection mode is set according to the characteristics of the anomaly to be detected, and as will be described later, the feature amount and the clustering algorithm that are the basis for performing anomaly detection according to the anomaly detection mode are determined. Will be done. In the present embodiment, the purpose of each anomaly detection mode is to detect a non-stationary anomaly signal, and the anomaly detection mode is an impact anomaly (separated impact noise) having an extremely short duration in the target signal. A first abnormality detection mode for detecting the presence or absence, a second abnormality detection mode for detecting the presence or absence of an abnormality (intermittent noise) related to the steadyness or intermittentness of the target signal, and an increase in rotation during acceleration / deceleration of the inspection target from the target signal. Alternatively, the presence or absence of an impact abnormality (quasi-steady impact noise) in the target signal, which has an extremely short duration in the target signal, and the third abnormality detection mode for detecting the presence or absence of an impact abnormality (fluctuation noise) due to the rotation descent It includes a fourth anomaly detection mode, which detects with high accuracy.

<12.6 Abnormality detection processing>
When the abnormality detection mode is set (S5), the abnormality detection unit 140 performs the abnormality detection process (S6). More specifically, the abnormality detection unit 140 specifies a feature amount and a clustering algorithm that are the basis of abnormality detection according to the abnormality detection mode, and performs abnormality detection based on the clustering result.

When the first abnormality detection mode is set, the abnormality detection unit 140 relates to the MFCC feature amount related to sound, acceleration, torque and rotation speed, the psychoacoustic evaluation feature amount, and the sound, acceleration, torque and rotation speed. The result of applying DBSCAN as a clustering algorithm to each coded feature is specified as the basis of abnormality detection. After that, when the feature amount indicating an outlier is specified in any of the classification results on which the abnormality detection is based, the abnormality detection unit 140 abnormalizes the feature amount and the time-series signal section corresponding to the feature amount. Detect as.

When the second abnormality detection mode is set, the abnormality detection unit 140 relates to the MFCC feature amount related to sound, acceleration, torque and rotation speed, the psychoacoustic evaluation feature amount, and the sound, acceleration, torque and rotation speed. The results of applying the k-means clustering method and GMM as clustering algorithms to the coded features are specified as the basis for abnormality detection. After that, when the feature amount indicating belonging to a predetermined class is specified in any of the classification results based on the abnormality detection, the abnormality detection unit 140 sets the feature amount and the time series corresponding to the feature amount. The signal section is detected as abnormal. In the present embodiment, the predetermined class is defined by the ratio occupied by each class in a predetermined time interval (for example, a time frame corresponding to 60 seconds). For example, a class having a small proportion in a predetermined time interval may be specified as a predetermined class showing an abnormality. The predetermined class may be defined by another method. For example, a class belonging to a steady state may be specified first, and other classes may be defined as a predetermined class indicating an abnormality.

When the third abnormality detection mode is set, the abnormality detection unit 140 relates to the MFCC feature amount related to sound, acceleration, torque and rotation speed, the psychoacoustic evaluation feature amount, and the sound, acceleration, torque and rotation speed. The results of applying the k-means clustering method and GMM as clustering algorithms to the coded features and the results of applying the k-means clustering method, GMM and DBSCAN as clustering algorithms to the rotation order features are detected as abnormalities. Identify as the basis. After that, in any of the classification results based on the abnormality detection, the abnormality detection unit 140 is out of the feature quantity indicating belonging to a predetermined class in the case of the k-means method and GMM, or in the case of DBSCAN. When a feature amount indicating a value is found, the feature amount and the time-series signal section corresponding to the feature amount are detected as anomalies. In the present embodiment, the predetermined class is defined by the ratio occupied by each class in a predetermined time interval (for example, a time frame corresponding to 60 seconds). For example, a class having a small proportion in a predetermined time interval may be specified as a predetermined class showing an abnormality. The predetermined class may be defined by another method. For example, a class belonging to a steady state may be specified first, and other classes may be defined as a predetermined class indicating an abnormality.

When the fourth abnormality detection mode is set, the abnormality detection unit 140 relates to the MFCC feature amount related to sound, acceleration, torque and rotation speed, the psychoacoustic evaluation feature amount, and the sound, acceleration, torque and rotation speed. The results of applying the k-means clustering method and GMM as clustering algorithms to the coded features, and the k-means clustering method and GMM as clustering algorithms for the coded features extracted assuming that the dynamic range is equal to or greater than a predetermined threshold. The result of applying GMM and the result of applying GMM are specified as the basis of abnormality detection. After that, in any of the classification results based on the abnormality detection, the abnormality detection unit 140 is out of the feature quantity indicating belonging to a predetermined class in the case of the k-means method and GMM, or in the case of DBSCAN. When the feature amount indicating the value is specified, the feature amount and the time-series signal section corresponding to the feature amount are detected as anomalies. In the present embodiment, the predetermined class is defined by the ratio occupied by each class in a predetermined time interval (for example, a time frame corresponding to 60 seconds). For example, a class having a small proportion in a predetermined time interval may be specified as a predetermined class showing an abnormality. The predetermined class may be defined by another method. For example, a class belonging to a steady state may be specified first, and other classes may be defined as a predetermined class indicating an abnormality.

FIG. 5 is an explanatory diagram for exemplifying the detection process of the abnormal section. The figure shows the case where MFCC features related to sound are clustered by DBSCAN in the first abnormality detection mode, and the frequency intensity (upper row) of the sound signal with respect to time ([min]) and the state of class transition at that time. (Lower) is represented by a graph.

As is clear from the figure, it is confirmed that the class has transitioned twice from 0 to -1 at a time slightly to the right of the center of the time-class graph. In the graph, -1 indicates an outlier. Therefore, the MFCC feature amount related to the sound and the time-series signal section corresponding to the feature amount are detected as abnormalities (that is, abnormal sounds in this case).

In the present embodiment, when an abnormality is detected in the clustering result of any of the features, the time interval is detected as an abnormality, but the present embodiment is not limited to such an embodiment. Therefore, for example, only when an abnormality is detected in a plurality of or all feature quantities in the same time interval, the time interval may be detected as an abnormality.

<12.7 Output processing>
When the abnormality detection process (S6) is completed, a process of outputting the detection result is performed (S7). As an output mode, there are an output via the abnormal output unit 151 and an output via the graph output unit 152. After these output processes, the process ends.

According to the output via the abnormal output unit 151, various data obtained so far can be output in an arbitrary data format such as CSV (Comma Separated Values). Further, according to the output via the graph output unit 152, the time series data and spectrogram of each signal (sound, acceleration, torque, rotation speed) that may include an abnormal section are displayed on the display or the like via the display output unit 2. be able to. At this time, the abnormal portion may be emphasized and displayed for each feature amount. In addition, only the portion recognized as abnormal in a plurality of feature quantities may be displayed as an abnormal portion. Further, the sound signal at a portion corresponding to the abnormal signal may be reproduced via the voice output unit 5 so that it can be confirmed whether or not the abnormal portion contains an abnormal sound or the like.

According to such a configuration, since the abnormal portion can be visually or audibly confirmed via a graph or voice, the identification of the abnormal portion becomes more intuitive and easy.

That is, according to the configuration according to the present embodiment, since the presence or absence of an abnormality in the time-series signal related to the inspection object is detected based on the clustering result, the abnormality is detected by unsupervised learning based on the time-series signal itself. be able to. As a result, it is not necessary to prepare a sample showing an abnormality and to compare it with the sample, so that the burden on the operator who performs the abnormality detection work is reduced. Further, since clustering is performed after generating a predetermined feature amount from the time series signal, the accuracy of clustering is improved.

Further, according to the configuration according to the present embodiment, since the abnormality detection is performed based on the sound generated from the inspection target object, the inspection can be performed non-destructively. Further, since it is not necessary to prepare a sound sample in the case of an abnormal state and to compare the sound sample with the sound sample, the burden on the operator who performs the abnormality detection work is reduced.

Further, according to the configuration according to the present embodiment, not only the sound emitted from the object but also the sensor information (for example, acceleration) indicating the operating state of the object that may have a causal relationship with the abnormality is utilized (for example). Multi-sensing), the accuracy of abnormality detection can be further improved.

<2. Second embodiment>
Next, a second embodiment of the present invention will be described with reference to FIGS. 6 to 9.

In the first embodiment, after performing the clustering process (S4), anomaly detection (S5) was performed based on the clustering result. However, depending on the nature of the data, clustering may not proceed properly even if clustering is performed. Based on such clustering results, it may not be possible to properly detect anomalies. In addition, when referring to all clustering results including the result of clustering not progressing properly, unnecessary processing time may occur. In the present embodiment, by selectively using the clustering result, the accuracy of abnormality detection is improved and the processing time is shortened.

In this embodiment, the hardware configuration is substantially the same as that in the first embodiment, and thus the description thereof will be omitted. It should be noted that the same reference numerals are given to the same configurations as those in the first embodiment.

FIG. 6 is a functional block diagram according to the present embodiment. As is clear from the figure, in this embodiment, in addition to the functional configuration of the first embodiment, a data selection unit 161 is provided between the clustering processing unit 139 and the abnormality detection unit 162. The data selection unit 161 performs a process of selecting an appropriate clustering result as described later.

FIG. 7 is a general flowchart relating to the operation of the abnormality detection determination device according to the present embodiment. When the processing is started, the time series signal acquisition processing (S1), the spectrogram conversion processing (S2), the feature amount generation processing (S3), and the clustering processing (S4) are performed as in the first embodiment. Will be done. In the present embodiment, after the clustering process (S4) is completed, the data selection process (S11) is performed by the data selection unit 161.

FIG. 8 is a detailed flowchart of the data selection process (S11). As is clear from the figure, when the process is started, the process of calculating the silhouette score (numerical value from -1 to 1) is performed based on the classification result.

The calculation method of the silhouette score will be described. First, for the target clustering result, a predetermined number of samples are silhouetted from the following equation based on the ^{value (a (i) ) representing the degree of cluster cohesion and the value (b (i)} ) representing the degree of cluster dissociation. Calculate the coefficient S ^(i).

^{The value (a (i)} ) representing the degree of cluster cohesion is the average distance from all other samples in the same cluster, and the degree of cluster divergence (b ⁽ⁱ⁾ ) is in the nearest cluster. Average distance to all samples. Further, max {a ⁽ⁱ⁾ , b ⁽ⁱ⁾ } represents the larger of ^{a (i)} and b ^(i). Further, i indicates a sample number.

After calculating the silhouette coefficient S ⁽ⁱ⁾ , the silhouette score is calculated. The silhouette score is calculated as the arithmetic mean value of ^{the silhouette coefficient S (i).}

Returning to FIG. 8, a process of specifying the reference data based on the calculated silhouette score is performed (S113). More specifically, among the plurality of silhouette scores calculated based on each feature amount, data having a silhouette score equal to or higher than a predetermined threshold value is specified as reference data.

FIG. 9 is an explanatory diagram showing the distribution when a plurality of silhouette scores are sorted in ascending order (in descending order of numerical values). As is clear from the figure, among the silhouette scores sorted in ascending order, the silhouette score drops sharply when it exceeds the 40th silhouette score (model number is 40). The data selection unit 161 sets the threshold value Th by automatically detecting the inflection point that suddenly drops by using the inclination or the like. Then, data having a silhouette score having a silhouette score of the threshold value Th or more (in the example of the figure, about 0.3 or more) is selected as reference data. The threshold value Th may be a fixed value or may be set by the user.

After the reference data is specified, a flag setting (flag ON) process (S115) is performed on the reference data, and the data selection process (S11) ends.

Returning to FIG. 7, when the data selection process (S11) is completed, the abnormality detection mode acquisition process is performed (S5) as in the first embodiment. That is, the abnormality detection mode set by the user or the like and stored in the storage unit 3 is read out by the abnormality detection mode acquisition unit 116 and output to the abnormality detection mode setting unit 141.

After that, the abnormality detection unit 140 performs the abnormality detection process based on the abnormality detection mode and the clustering result flagged by the data selection process (S11) (S16). That is, as in the first embodiment, the abnormality detection unit 140 identifies the feature amount and the clustering algorithm that are the basis of the abnormality detection according to the abnormality detection mode, and detects the abnormality based on the flagged clustering result. I do. After that, the output process is performed in the same manner as in the first embodiment (S7), and the process ends.

According to such a configuration, anomaly detection can be performed based only on the result of good clustering, so that anomaly detection can be performed more appropriately. Moreover, since the number of data used for abnormality detection is reduced, the processing time can be shortened.

<3. Modification example>
In the above-described embodiment, abnormality detection is performed using time-series signals relating to sound, acceleration, torque, and rotation speed. However, the application target of the present invention is not limited to the above-mentioned signals, and can be applied to various time series signals. Therefore, abnormality detection can be performed even with other time-series signals. Also, it does not necessarily have to include a signal related to sound.

Further, in the above-described embodiment, the drive train (power transmission mechanism) that transmits the rotational power from the engine of the automobile is the inspection target, but the inspection target of the present invention is not limited to this, and the rotating body is also used. Is not limited. Therefore, the present invention can be applied to, for example, abnormality detection of an object that performs other motions without rotation or an object that does not involve physical motions.

Further, in the above-described embodiment, the k-means method, GMM and DBSCAN are used as the clustering method. However, other clustering methods can be applied to the present invention, and for example, SOM (Self-Organizing Maps) and the like can also be used.

In addition, in the above-described embodiment, the MFCC feature amount, the psychoacoustic evaluation feature amount, the rotation order feature amount, and the coded feature amount are used as the feature amount to be clustered. However, the present invention is not limited to such a configuration. Therefore, for example, only a part of the above-mentioned feature amounts may be used, or another completely different feature amount may be adopted.

Further, in the second embodiment, the data selection process is performed using the silhouette score, but the present invention is not limited to such a configuration. Therefore, for example, data selection may be performed using another index indicating the accuracy of clustering. Alternatively, for example, DBSCAN or the like may be used to calculate the outlier ratio and select to exclude data having a high outlier ratio.

The present invention can be used in all industries that require inspection of industrial products and the like.

1 Control unit 2 Display output unit 3 Storage unit 4 Operation input unit 5 Voice output unit 6 Communication unit 7 I / O unit 10 Abnormality detection device 50 Data measurement system 51 Inspection target 52 Control server 53 Front end 54 Measurement system 55 Storage device

Claims (17)

A time-series signal acquisition unit that acquires a time-series signal generated by placing the inspection object in a specific state, and
A feature amount generation unit that generates a predetermined feature amount based on the time-series signal,
A clustering processing unit that clusters the features based on a predetermined clustering algorithm,
A detection unit that detects an abnormal section in the time-series signal based on the result of the clustering is provided .
The time-series signal includes a time-series sound signal which is a time-series signal related to a sound generated from the inspection object and a time-series sensor signal which is a time-series signal related to sensor information indicating an operating state of the inspection object. Including anomaly detection system. The feature amount generation unit further
A frequency domain signal generator that converts the time-series sound signal and the time-series sensor signal into a frequency domain to generate a frequency domain sound signal and a frequency domain sensor signal.
Psychoacoustic evaluation that generates loudness features related to sound loudness, sharpness features related to sound sharpness, fluctuation intensity features related to sound fluctuation intensity, and roughness features related to sound roughness based on the frequency region sound signal. Feature amount generator and
The abnormality detection system according to claim 1 , further comprising an MFCC feature amount generation unit that generates an MFCC feature amount including a mel frequency cepstrum coefficient based on the frequency domain sound signal and the frequency domain sensor signal. The feature amount generation unit further
When the frequency domain sound signal and the frequency domain sensor signal are input based on a trained neural network having an input neuron, an output neuron, and an intermediate neuron between them and trained to function as an autoencoder. The abnormality detection system according to claim 2 , further comprising a coded feature amount generator that outputs the firing value of the intermediate neuron as a coded feature amount. The object to be inspected is a rotating body.
The time-series sensor signal includes a time-series rotation speed signal of the rotating body.
The feature amount generation unit further
The abnormality detection system according to claim 3 , further comprising a rotation order feature quantity generation processing unit that generates a predetermined rotation order as a feature quantity based on the time-series rotation speed signal and the frequency domain sound signal. The object to be inspected is a rotating body.
The abnormality detection system according to claim 1 , wherein the time-series sensor signal is a time-series signal related to acceleration, torque, and rotation speed. The abnormality detection system according to claim 1, wherein the clustering algorithm includes a first clustering algorithm that determines the number of clusters in advance and a second clustering algorithm that does not determine the number of clusters in advance. The first clustering algorithm is one or both of k-means clustering and GMM.
The anomaly detection system according to claim 6 , wherein the second clustering algorithm is DBSCAN. The anomaly detection system further
Anomaly detection mode acquisition unit that acquires anomaly detection mode,
The abnormality detection system according to claim 1, further comprising the feature amount as a basis for detection and a detection basis determination unit for determining the clustering algorithm based on the abnormality detection mode. The anomaly detection system further
Anomaly detection system according to claim 1, further comprising a graphical output data generating unit, for generating display data for displaying the time chart series sound signal having the results identified the abnormal section of the abnormality detection in the detection unit. The feature amount generation unit further
A frequency domain signal generator that converts the time-series sound signal into a frequency domain and generates a frequency domain sound signal.
Psychoacoustic evaluation that generates loudness features related to sound loudness, sharpness features related to sound sharpness, fluctuation intensity features related to sound fluctuation intensity, and roughness features related to sound roughness based on the frequency region sound signal. It includes a feature amount generating unit, the abnormality detection system according to claim 1. The anomaly detection system further
A clustering accuracy determination unit for determining clustering accuracy based on the clustering result is provided.
The abnormality detection according to claim 1, wherein the detection unit detects an abnormality section in the time-series signal based on the result of the clustering determined by the clustering accuracy determination unit to have an accuracy equal to or higher than a predetermined reference. system. The abnormality detection system according to claim 11 , wherein the clustering accuracy determination unit determines the accuracy of the clustering result based on the silhouette coefficient calculated for the clustering result. The abnormality detection system according to claim 11 , wherein the clustering accuracy determination unit determines the accuracy of the clustering result based on the ratio of outliers of the clustering result. A time-series signal acquisition unit that acquires a time-series signal generated by placing the inspection object in a specific state, and
A feature amount generation unit that generates a predetermined feature amount based on the time-series signal,
A clustering processing unit that clusters the features based on a predetermined clustering algorithm,
A detection unit that detects an abnormal section in the time-series signal based on the result of the clustering is provided .
The time-series signal includes a time-series sound signal which is a time-series signal related to a sound generated from the inspection object and a time-series sensor signal which is a time-series signal related to sensor information indicating an operating state of the inspection object. Including anomaly detector. A time-series signal acquisition step for acquiring a time-series signal generated by placing an inspection object in a specific state, and
A feature amount generation step for generating a predetermined feature amount based on the time-series signal, and
A clustering process step for clustering the features based on a predetermined clustering algorithm, and
A detection step for detecting an abnormal section in the time-series signal based on the result of the clustering is provided .
The time-series signal includes a time-series sound signal which is a time-series signal related to a sound generated from the inspection object and a time-series sensor signal which is a time-series signal related to sensor information indicating an operating state of the inspection object. Abnormality detection method including. Computer,
A time-series signal acquisition unit that acquires a time-series signal generated by placing the inspection object in a specific state, and
A feature amount generation unit that generates a predetermined feature amount based on the time-series signal,
A clustering processing unit that clusters the features based on a predetermined clustering algorithm,
Based on the result of the clustering, a shall to function as the abnormality detection apparatus and a detection unit for detecting the abnormal section in the series signals,
The time-series signal includes a time-series sound signal which is a time-series signal related to a sound generated from the inspection object and a time-series sensor signal which is a time-series signal related to sensor information indicating an operating state of the inspection object. Including anomaly detection program. A time-series signal acquisition unit that acquires a time-series sound signal that is a time-series signal related to a sound generated from a drive train and a time-series sensor signal that is a time-series signal related to sensor information indicating an operating state of the drive train.
A feature amount generation unit that generates a predetermined feature amount based on the time-series signal,
A clustering processing unit that clusters the features based on a predetermined clustering algorithm,
An abnormal sound detection system including a detection unit that detects an abnormal section in the time-series sound signal based on the result of the clustering.

Images