Unsupervised industrial abnormal sound detection method and system based on time-frequency feature selective learning

Abstract

The invention discloses an unsupervised industrial abnormal sound detection method and system based on time-frequency feature selective learning, and the method comprises the following steps: 1, extracting log-Mel spectrogram features from a training set sound signal, executing a spectrogram frame selection strategy, and obtaining a spectrogram frame of each sound sample; step 2, inputting the spectrogram frame in the step 1 into a frequency domain feature selection auto-encoder model for training, wherein a loss function is LAEFS; step 3, extracting log-Mel spectrogram features from the collected unknown sound samples, and selecting a plurality of frames from a log-Mel spectrogram according to a spectrogram frame selection strategy; step 4, inputting the spectrogram frame obtained in the step 3 into the trained frequency domain feature selection auto-encoder model in the step 2, and taking the reconstruction loss LRec as an abnormal score S; and step 5, comparing the abnormal score S obtained in the step 4 with a threshold value A theta, if the abnormal score S is greater than the threshold value A theta, indicating that the input sound signal is abnormal, otherwise, indicating that the sound signal is normal.

Description

Technical Field

[0001] The present invention belongs to the technical field of abnormal sound detection, and in particular relates to an unsupervised industrial abnormal sound detection method and system based on selective learning of time-frequency features. Background Art

[0002] In industrial manufacturing, it is crucial to monitor the operating status of machines and equipment through the sounds they make and detect abnormalities in a timely manner to ensure the safety and stable operation of the equipment. With the rapid advancement of artificial intelligence and industrial big data technologies, more and more researchers are using deep learning technology to detect industrial abnormal sound. Considering that the equipment is usually in normal operation and there is relatively little abnormal data, most industrial abnormal sound detection tends to use self-supervised or unsupervised learning methods. Self-supervised learning relies on the metadata of the sound to train the model, but the metadata does not exist in all cases, which limits its application scenarios. In contrast, unsupervised deep learning methods are not subject to such restrictions. Specifically, unsupervised deep learning methods usually extract specific features from the sound of machine operation, use generative models to learn the characteristic patterns of normal sounds, and detect anomalies by comparing whether the samples conform to these patterns. In this process, the model's capture of sound features greatly affects the detection effect.

[0003] Affected by noise and other interference, the characteristics of equipment sound signals in different time periods and frequencies have different effects on anomaly detection results. The key to anomaly detection is to capture discriminative features. Compared with learning features in the entire time and frequency domain dimensions, focusing on learning features with significant discriminativeness is more helpful in improving detection performance. Based on this, the present invention discloses an unsupervised industrial abnormal sound detection method and system based on selective learning of time-frequency features. Summary of the invention

[0004] In view of the above-mentioned status of the prior art, the present invention proposes an unsupervised industrial abnormal sound detection method and system based on selective learning of time-frequency features. The present invention first extracts the log-Mel spectrogram features of the sound; selects several spectrogram frames containing the target sound signal and less affected by noise through the spectrogram frame selection (SFS) strategy; inputs the obtained spectrogram frames into the frequency domain feature selection autoencoder (AEFS) model to obtain the abnormality score; and determines whether the sound sample is abnormal based on the abnormality score.

[0005] The following is an explanation of the professional terms in the present invention:

[0006] Spectrogram frame: The log-Mel spectrogram converts the sound signal from the time domain to the frequency domain while taking into account the perceptual characteristics of the human auditory system. Usually, the log-Mel spectrogram is obtained by calculating several consecutive time frames of the original sound signal, which provides information about the frequency and time characteristics of the sound signal. In the log-Mel spectrogram, each frame along the dimension of the time axis is called a spectrogram frame. Each spectrogram frame represents the intensity or energy distribution of different frequency components in the current time domain, which is calculated from the corresponding time frame of the original audio signal. For audio of a certain length, its log-Mel spectrogram contains several spectrogram frames, which respectively represent the intensity or energy distribution of different frequency components in each time frame.

[0007] Stationary and non-stationary sound signals: Not all mechanical equipment emits sound continuously during the operation period. Generally, its sound signals can be roughly divided into two cases: stationary and non-stationary, namely: when the target sound event exists continuously during the working period of the mechanical equipment, all time periods contain the information of the target sound, which is called a stationary sound signal; when the target sound event rarely occurs during the working period of the mechanical equipment, the signal is sparse. At this time, only a small part of the time period has the target sound, and most of the time period is non-target sound (environmental noise or other interference), which is called a non-stationary sound signal. For stationary sound signals, its main information is distributed in most of the time periods. Therefore, most of the spectrogram frames in the log-Mel spectrogram can show the frequency domain characteristics of the target sound signal. For non-stationary sound signals, its main information may be scattered in very few time periods, so only very few spectrogram frames in the log-Mel spectrogram contain the frequency domain characteristics of the target signal, and most of the spectrogram frames are noise or interference. See the spectrogram examples of stationary and non-stationary sound signals. Figure 2 .

[0008] The technical solution adopted by the present invention to solve the technical problem is a method of using a plurality of spectrogram frames obtained by executing the SFS strategy on the sound log-Mel spectrogram as input and combining it with the AEFS neural network model, which is specifically as follows:

[0009] The unsupervised industrial abnormal sound detection method based on selective learning of time-frequency features includes the following steps:

[0010] Step 1: extract log-Mel spectrogram features from the training set sound signals, execute the spectrogram frame selection strategy, and obtain the spectrogram frame of each sound sample;

[0011] Step 2: Input the spectrogram frame in step 1 into the frequency domain feature selection autoencoder model for training, and the loss function is L AEFS ;

[0012] Step 3: extract log-Mel spectrogram features from the collected sound samples, and select several frames from the log-Mel spectrogram according to the spectrogram frame selection strategy;

[0013] Step 4: Input the frames selected in step 3 into the frequency domain feature selection autoencoder model trained in step 2, and reconstruct the loss L Rec as the abnormality score S;

[0014] Step 5: Compare the anomaly score S obtained in step 4 with the threshold A θ Compared with, greater than threshold A θ Indicates that the input sound signal is abnormal, otherwise it is normal, expressed as:

[0015]

[0016] Preferably, the SFS strategy is as follows:

[0017] The log-Mel spectrogram matrix extracted from the sound signal is expressed as:

[0018] X∈R k×d =[x1,x2,...x k ] T

[0019] Among them, x i ∈R 1×d is the i-th spectrogram frame, k is the total number of spectrogram frames, and d is the dimension of the spectrogram frame, i.e., the number of Mel filters;

[0020] Use r i Indicates the correlation between the i-th spectrum frame and the entire spectrum, defined as the spectrum frame x i The sum of the covariances with all other spectrogram frames:

[0021]

[0022] Among them, cov(xi,xj) is the vector x i The covariance between and xj, and then the set of correlation sizes can be obtained:

[0023] R=[r1,r2,...r k ]

[0024] According to the different characteristics of the running sound of industrial machines, different spectrogram frame selection strategies are designed. Specifically, if the input sound signal is a stationary signal, the n-frame features that are most correlated with other spectrogram frames (i.e., the spectrogram frames with the largest correlation) are selected from the log-Mel spectrum matrix X. The process can be described as:

[0025] I max=IndexMaxN(R,n)

[0026] X n =Select(X,I max )

[0027] Among them, IndexMaxN(R,n) represents the index of the first n maximum values ​​from the set R, n refers to the number of spectrogram frames to be selected, Select(X,I max ) indicates that according to index I max Select the corresponding vector x from the matrix X i , while X n is a log-Mel spectrogram matrix containing only the selected n spectrogram frames.

[0028] If the input sound signal is a non-stationary signal, the n-frame features that are least correlated with other spectrogram frames (i.e., the spectrogram frames with the smallest correlation) are selected from the log-Mel spectrogram matrix X. The process can be described as:

[0029] I min =IndexMinN(R,n)

[0030] X n =Select(X,I min )

[0031] Among them, IndexMinN(R,n) represents the index of the first n minimum values ​​taken from the set R.

[0032] After the spectrum frame selection, the log-Mel spectrum matrix composed of the selected spectrum frames can be obtained:

[0033]

[0034] in, Yes X n The i-th eigenvector in .

[0035] Preferably, the AEFS model is as follows:

[0036] The AEFS model is constructed by introducing a scaling gate (SG) after the autoencoder (AE), which selects frequency domain features through the scaling function of SG. Let S∈R 1×d Represents the weight parameter of SG. The weighted calculation of SG on all frequency dimension features can be expressed as:

[0037]

[0038] Where x and x′ represent the input and output of the scaling gate, respectively. Represents element-by-element multiplication between vectors or matrices, where e is the natural base, which is used here to ensure the non-negativity of the x weight.

[0039] The frequency domain feature selection process minimizes the reconstruction loss L of the frequency domain features. Rec and SG loss L SG Suppose for The output obtained directly through SG, for The output obtained by AEFS. Then the reconstruction loss of AEFS is L Rec and SG loss L SG Respectively expressed as:

[0040]

[0041] in, represents the square of the L2 norm, and They are:

[0042]

[0043] The total loss of AEFS includes the reconstruction loss L Rec and SG loss L SG , calculated as:

[0044] L AEFS =βL Rec +L SG

[0045] Among them, β is the loss L Rec and L SG A smaller β value (0<β<1) means that the model pays more attention to frequency domain feature selection; on the contrary, a larger β value (β>1) will cause the model to pay more attention to feature reconstruction. In fact, β can be selected based on experience. Reconstruction loss L Rec It encourages the model to learn the distribution pattern of the selected frequency domain features. The loss function L of SG SG , so that the sum of all dimensional parameters of SG has an increasing trend. However, since the parameters of SG are used as weight factors in the reconstruction error calculation, they are affected by L Rec The parameters of each dimension may change differently due to the constraints of the model. The parameters of some dimensions may increase, while the parameters of other dimensions may decrease.

[0046] The present invention also discloses an unsupervised industrial abnormal sound detection system based on selective learning of time-frequency features, which specifically includes the following modules:

[0047] Model training module: extract log-Mel spectrogram features from the training set sound signals, execute the spectrogram frame selection strategy, and obtain the spectrogram frame of each sound sample; input the spectrogram frame into the frequency domain feature selection autoencoder model for training, and the loss function is L AEFS ;

[0048] Anomaly detection module: extract log-Mel spectrogram features from the collected sound samples, select several frames from the log-Mel spectrogram according to the spectrogram frame selection strategy; input the selected frames into the trained frequency domain feature selection autoencoder model in the model training module, and reconstruct the loss L Rec As the anomaly score S; the obtained anomaly score S and the threshold A θ Compared with, greater than threshold A θ Indicates that the input sound signal is abnormal, otherwise it is normal, expressed as:

[0049]

[0050] The beneficial effects of the present invention are as follows:

[0051] In the technical solution of the present invention, the SFS strategy can remove the spectrogram frames in the log-Mel spectrogram that are interfered by noise or do not contain the target sound signal, providing a relatively pure spectrogram feature for subsequent related operations of anomaly detection, which is beneficial to improving the detection performance.

[0052] In the training process of AEFS of the present invention, the corresponding weights of feature dimensions with higher reconstruction errors will be reduced, because reducing these weights can minimize the overall loss. On the contrary, for feature dimensions with lower reconstruction errors, their corresponding weights will increase or slightly decrease. Ultimately, feature dimensions with lower reconstruction errors are selected by the model and their distribution patterns are learned. Related studies have shown that for generative models, features with lower reconstruction errors are relatively constant or representative features. Therefore, the essence of the AEFS model is to learn relatively constant or representative frequency domain features of normal sounds, which helps to improve anomaly detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a framework diagram of an unsupervised industrial abnormal sound detection method based on selective learning of time-frequency features according to a preferred embodiment of the present invention;

[0054] Figure 2 Examples of log-Mel spectrograms for stationary and non-stationary sound signals;

[0055] Figure 3 It is the structure and parameter diagram of AEFS model;

[0056] Figure 4 Diagram of the training and testing process of the AEFS model;

[0057] Figure 5 This is a flow chart of unsupervised industrial abnormal sound detection based on selective learning of time-frequency features in a preferred embodiment of the present invention;

[0058] Figure 6 This is a block diagram of an unsupervised industrial abnormal sound detection system based on selective learning of time-frequency features according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0059] The preferred embodiments of the present invention are described in detail below with reference to the drawings. The following description is only for illustration and does not limit the present invention in any form.

[0060] like Figure 1-5 As shown, the unsupervised industrial abnormal sound detection method based on selective learning of time-frequency features in this embodiment specifically includes the following steps:

[0061] Step 1: extract log-Mel spectrogram features from the training set sound signals, execute the spectrogram frame selection strategy, and obtain the spectrogram frame of each sound sample;

[0062] Step 2: Input the spectrogram frame in step 1 into the frequency domain feature selection autoencoder model for training, and the loss function is L AEFS ;

[0063] Step 3: extract log-Mel spectrogram features from the collected sound samples, and select several frames from the log-Mel spectrogram according to the spectrogram frame selection strategy;

[0064] Step 4: Input the frames selected in step 3 into the frequency domain feature selection autoencoder model trained in step 2, and reconstruct the loss L Rec as the abnormality score S;

[0065] Step 5: Compare the anomaly score S obtained in step 4 with the threshold A θ Compared with, greater than threshold A θ Indicates that the input sound signal is abnormal, otherwise it is normal, expressed as:

[0066]

[0067] In this embodiment, the SFS strategy is as follows:

[0068] The log-Mel spectrogram matrix extracted from the sound signal is expressed as:

[0069] X∈R k×d =[x1,x2,...x k ] T

[0070] Among them, xi ∈R 1×d is the i-th spectrogram frame, k is the total number of spectrogram frames, and d is the dimension of the spectrogram frame, i.e., the number of Mel filters;

[0071] Use r i Indicates the correlation between the i-th spectrum frame and the entire spectrum, defined as the spectrum frame x i The sum of the covariances with all other spectrogram frames:

[0072]

[0073] Among them, cov(xi,xj) is the vector x i The covariance between and xj, and then the set of correlation sizes can be obtained:

[0074] R=[r1,r2,...r k ]

[0075] According to the different characteristics of the running sound of industrial machines, different spectrogram frame selection strategies are designed. Specifically, if the input sound signal is a stationary signal, the n-frame features that are most correlated with other spectrogram frames (i.e., the spectrogram frames with the largest correlation) are selected from the log-Mel spectrum matrix X. The process can be described as:

[0076] I max =IndexMaxN(R,n)

[0077] X n =Select(X,I max )

[0078] Among them, IndexMaxN(R,n) represents the index of the first n maximum values ​​from the set R, n refers to the number of spectrogram frames to be selected, Select(X,I max ) indicates that according to index I max Select the corresponding vector x from the matrix X i , while X n is a log-Mel spectrogram matrix containing only the selected n spectrogram frames.

[0079] If the input sound signal is a non-stationary signal, the n-frame features that are least correlated with other spectrogram frames (i.e., the spectrogram frames with the smallest correlation) are selected from the log-Mel spectrogram matrix X. The process can be described as:

[0080] I min =IndexMinN(R,n)

[0081] X n =Select(X,I min )

[0082] Among them, IndexMinN(R,n) represents the index of the first n minimum values ​​taken from the set R.

[0083] After the spectrum frame selection, the log-Mel spectrum matrix composed of the selected spectrum frames can be obtained:

[0084]

[0085] in, Yes X n The i-th eigenvector in .

[0086] In this embodiment, the AEFS model is as follows:

[0087] The AEFS model is constructed by introducing a scaling gate (SG) after the autoencoder (AE), which selects frequency domain features through the scaling function of SG. Let S∈R 1×d Represents the weight parameter of SG. The weighted calculation of SG on all frequency dimension features can be expressed as:

[0088]

[0089] Where x and x′ represent the input and output of the scaling gate, respectively. Represents element-by-element multiplication between vectors or matrices, where e is the natural base, which is used here to ensure the non-negativity of the x weight.

[0090] The frequency domain feature selection process minimizes the reconstruction loss L of the frequency domain features. Rec and SG loss L SG Suppose for The output obtained directly through SG, for The output obtained by AEFS. Then the reconstruction loss of AEFS is L Rec and SG loss L SG Respectively expressed as:

[0091]

[0092] in, represents the square of the L2 norm, and They are:

[0093]

[0094] The total loss of AEFS includes the reconstruction loss L Rec and SG loss L SG , calculated as:

[0095] L AEFS =βL Rec +L SG

[0096] Among them, β is the loss L Rec and L SG A smaller β value (0<β<1) means that the model pays more attention to frequency domain feature selection; on the contrary, a larger β value (β>1) will cause the model to pay more attention to feature reconstruction. In fact, β can be selected based on experience. Reconstruction loss L Rec It encourages the model to learn the distribution pattern of the selected frequency domain features. The loss function L of SG SG , so that the sum of all dimensional parameters of SG has an increasing trend. However, since the parameters of SG are used as weight factors in the reconstruction error calculation, they are affected by L Rec The parameters of each dimension may change differently due to the constraints of the model. The parameters of some dimensions may increase, while the parameters of other dimensions may decrease.

[0097] Taking the sound of industrial equipment as an example, the unsupervised industrial abnormal sound detection method based on selective learning of time-frequency features is further explained as follows:

[0098] In this embodiment, the public dataset of machine abnormal sound detection of DCASE2020 TASK2 is used, which includes six machine types: Fan, Pump, Slider, Valve, ToyCar and Toyconveyor. Each machine type has 3 or 4 machine IDs. The dataset of each machine ID consists of approximately 1,000 normal sound samples for training and several normal and abnormal sound signal samples for testing. The sampling rate is 16 kHz and the duration is 10 s.

[0099] Model training:

[0100] Step 1: Extract log-Mel spectrogram features from the training set sound, using a Mel filter bank of size 128, with a frame length and frame shift of 1024 and 512 respectively. According to the SFS strategy, if it is a stationary sound signal, select the n frames with the largest sum of covariances with other spectrogram frames; otherwise, select the n frames with the smallest sum of covariances with other spectrogram frames. Here, Slider and Valve are non-stationary sound signals, and the rest are stationary sound signals; the number of spectrogram frames selected for each machine type is Fan (270 frames), Pump (250 frames), Slider (7 frames), Valve (7 frames), ToyCar (230 frames) and ToyConveyor (130 frames).

[0101] Step 2: Input the spectrogram frame obtained in step 1 into the AEFS model for training. The structure and parameters of the AEFS model are shown in Figure 3 , the loss function is L AEFS , using the Adam optimizer, the learning rate is set to 0.002, the BatchSize is 128, and the training is 6000 Epochs. Among them, L in the loss function AEFS The hyperparameter β is set to Fan(0.4), Pump(1.0), Slider(9.0), Valve(3.0), ToyCar(0.2) and ToyConveyor(7.0) according to different machine types.

[0102] Anomaly Detection:

[0103] For an unknown sound sample collected from a machine, perform the following processing to determine whether it is abnormal:

[0104] Step 1: Extract log-Mel spectrogram features. According to the SFS strategy and the corresponding machine type, select a number of frames from the log-Mel spectrogram, which should be consistent with the number of frames selected for model training.

[0105] Step 2: Input the selected frames into the trained AEFS model and reconstruct the loss L Rec As the anomaly score S.

[0106] Step 3: Compare the obtained anomaly score S with the threshold A θ Compared with, greater than threshold A θ Indicates that the input sound signal is abnormal, otherwise it is normal, expressed as:

[0107]

[0108] like Figure 6 As shown, this embodiment discloses an unsupervised industrial abnormal sound detection system based on selective learning of time-frequency features, which specifically includes the following modules:

[0109] Model training module: extract log-Mel spectrogram features from the training set sound signals, execute the spectrogram frame selection strategy, and obtain the spectrogram frame of each sound sample; input the spectrogram frame into the frequency domain feature selection autoencoder model for training, and the loss function is L AEFS ;

[0110] Anomaly detection module: extract log-Mel spectrogram features from the collected sound samples, select several frames from the log-Mel spectrogram according to the spectrogram frame selection strategy; input the selected frames into the trained frequency domain feature selection autoencoder model in the model training module, and reconstruct the loss L RecAs the anomaly score S; the obtained anomaly score S and the threshold A θ Compared with, greater than threshold A θ Indicates that the input sound signal is abnormal, otherwise it is normal, expressed as:

[0111]

[0112] For other contents of this embodiment, please refer to the above method embodiment.

[0113] The present invention discloses an unsupervised industrial abnormal sound detection method and system based on selective learning of time-frequency features. The method comprises: extracting log-Mel spectrogram features from original sound signals; selecting a number of spectrogram frames containing target information from the log-Mel spectrogram features through a covariance-based spectrogram frame selection (SFS) strategy, and removing the remaining spectrogram frames containing noise or irrelevant information; processing the sounds of a training set and a test set according to the aforementioned steps; inputting the processed training set data into an Auto-Encoder for Frequency-Feature Selection (AEFS) model for training to optimize model parameters, wherein the AEFS is constructed based on the auto-encoder and can realize selective learning of frequency-domain features; in a testing phase, the reconstruction error of the AEFS model for input samples is used as an abnormality score, the processed test set sound data is input into the trained AEFS model, and whether it is abnormal or not is determined by comparing the abnormality score with a threshold.

[0114] Any matters not described in the present invention are applicable to the prior art.

[0115] The specific implementation modes of the present invention are described above to facilitate the understanding of the present invention by those skilled in the art. However, the present invention is not limited to the scope of the specific implementation modes. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

Claims

1. Unsupervised industrial abnormal sound detection method based on selective learning of time-frequency features, characterized by The specific steps include: Step 1: extract log-Mel spectrogram features from the training set sound signals, execute the spectrogram frame selection strategy, and obtain the spectrogram frame of each sound sample; Step 2: Input the spectrogram frame in step 1 into the frequency domain feature selection autoencoder model for training, and the loss function is L AEFS ; Step 3: extract log-Mel spectrogram features from the collected sound samples, and select several frames from the log-Mel spectrogram according to the spectrogram frame selection strategy; Step 4: Input the frames selected in step 3 into the frequency domain feature selection autoencoder model trained in step 2, and reconstruct the loss L Rec as the abnormality score S; Step 5: Compare the anomaly score S obtained in step 4 with the threshold A θ Compared with, greater than threshold A θ Indicates that the input sound signal is abnormal, otherwise it is normal, expressed as:

2. The unsupervised industrial abnormal sound detection method based on selective learning of time-frequency features as claimed in claim 1 is characterized in that: The spectrogram frame selection strategy is specifically as follows: The log-Mel spectrogram matrix extracted from the sound signal is expressed as: X∈R k×d =[x1,x2,...x k ] T Among them, x i ∈R 1×d is the i-th spectrogram frame, k is the total number of spectrogram frames, and d is the dimension of the spectrogram frame; r i Indicates the correlation between the i-th spectrum frame and the entire spectrum, defined as the spectrum frame x i The sum of the covariances with all other spectrogram frames: Among them, cov(xi,xj) is the vector x i and the covariance between xj, and then obtain a set of correlation sizes: R＝[r1,r2,...r k ] If the input sound signal is a stationary signal, the n-frame features that are most relevant to other spectrogram frames are selected from the log-Mel spectrum matrix X, which can be described as: I max =IndexMaxN(R,n) X n =Select(X,I max ) Among them, IndexMaxN(R,n) represents the index of the first n maximum values ​​from the set R, n refers to the number of spectrogram frames to be selected, Select(X,I max ) indicates that according to index I max Select the corresponding vector x from the matrix X i , X n is the log-Mel spectrogram matrix containing only the selected n spectrogram frames; If the input sound signal is a non-stationary signal, the n-frame features that are least correlated with other spectrogram frames are selected from the log-Mel spectrum matrix X, which can be described as: I min =IndexMinN(R,n) X n =Select(X,I min ) Among them, IndexMinN(R,n) represents the index of the first n minimum values ​​from the set R; After the spectrum frame selection, the log-Mel spectrum matrix consisting of the selected spectrum frames is obtained in, Yes X n The i-th eigenvector in .

3. The unsupervised industrial abnormal sound detection method based on selective learning of time-frequency features according to claim 2 is characterized in that: The frequency domain feature selection autoencoder model is specifically: Let S∈R 1×d Represents the weight parameter of the scaling gate. The weighted calculation of the scaling gate on all frequency dimension features is expressed as: Where x and x′ represent the input and output of the scaling gate, respectively. Represents element-by-element multiplication between vectors or matrices, where e is the natural base; set up for The output directly passed through the scaling gate, for The output obtained by the frequency domain feature selection autoencoder is then the reconstruction loss L of the frequency domain feature selection autoencoder Rec It is expressed as: in, represents the square of the L2 norm, and They are: The total loss of the frequency domain feature selection autoencoder model includes the reconstruction loss L Rec and SG loss L SG , calculated as: L AEFS =βL Rec +L SG in: Among them, β is the loss L Rec and L SG The balance factor.

4. Unsupervised industrial abnormal sound detection system based on selective learning of time-frequency features, characterized by The modules include: Model training module: extract log-Mel spectrogram features from the training set sound signals, execute the spectrogram frame selection strategy, and obtain the spectrogram frame of each sound sample; input the spectrogram frame into the frequency domain feature selection autoencoder model for training, and the loss function is L AEFS ; Anomaly detection module: extract log-Mel spectrogram features from the collected sound samples, select several frames from the log-Mel spectrogram according to the spectrogram frame selection strategy; input the selected frames into the trained frequency domain feature selection autoencoder model in the model training module, and reconstruct the loss L Rec As the anomaly score S; the obtained anomaly score S and the threshold A θ Compared with, greater than threshold A θ Indicates that the input sound signal is abnormal, otherwise it is normal, expressed as:

5. The unsupervised industrial abnormal sound detection system based on selective learning of time-frequency features as claimed in claim 4, characterized in that: The spectrogram frame selection strategy is specifically as follows: The log-Mel spectrogram matrix extracted from the sound signal is expressed as: X∈R k×d =[x1,x2,...x k ] T Among them, x i ∈R 1×d is the i-th spectrogram frame, k is the total number of spectrogram frames, and d is the dimension of the spectrogram frame; r i Indicates the correlation between the i-th spectrum frame and the entire spectrum, defined as the spectrum frame x i The sum of the covariances with all other spectrogram frames: Among them, cov(xi,xj) is the vector x i and the covariance between xj, and then obtain a set of correlation sizes: R＝[r1,r2,...r k ] If the input sound signal is a stationary signal, the n-frame features that are most relevant to other spectrogram frames are selected from the log-Mel spectrum matrix X, which can be described as: I max =IndexMaxN(R,n) X n =Select(X,I max ) Among them, IndexMaxN(R,n) represents the index of the first n maximum values ​​from the set R, n refers to the number of spectrogram frames to be selected, Select(X,I max ) indicates that according to index I max Select the corresponding eigenvector x from the matrix X i , X n is the log-Mel spectrogram matrix containing only the selected n spectrogram frames; If the input sound signal is a non-stationary signal, the n-frame features that are least correlated with other spectrogram frames are selected from the log-Mel spectrum matrix X, which can be described as: I min =IndexMinN(R,n) X n =Select(X,I min ) Among them, IndexMinN(R,n) represents the index of the first n minimum values ​​from the set R; After the spectrum frame selection, the log-Mel spectrum matrix consisting of the selected spectrum frames is obtained in, Yes X n The i-th eigenvector in .

6. The unsupervised industrial abnormal sound detection system based on selective learning of time-frequency features as claimed in claim 5, characterized in that: The frequency domain feature selection autoencoder model is specifically: Let S∈R 1×d Represents the weight parameter of the scaling gate. The weighted calculation of the scaling gate on all frequency dimension features is expressed as: Where x and x′ represent the input and output of the scaling gate, respectively. Represents element-by-element multiplication between vectors or matrices, where e is the natural base; set up for The output directly passed through the scaling gate, for The output obtained by the frequency domain feature selection autoencoder is then the reconstruction loss L of the frequency domain feature selection autoencoder Rec It is expressed as: in, represents the square of the L2 norm, and They are: The total loss of the frequency domain feature selection autoencoder model includes the reconstruction loss L Rec and SG loss L SG , calculated as: L AEFS =βL Rec +L SG in: Among them, β is the loss L Rec and L SG The balance factor.

Images