Industrial equipment unsupervised abnormal sound signal detection method based on similarity feature selection decoupling

By using the combination of similarity feature selection and β-variant autoencoder model in industrial equipment, the noise interference problem of abnormal acoustic signal detection in industrial equipment is solved, and fast and accurate abnormal detection is achieved, with strong adaptability and application for complex audio data.

CN119091908APending Publication Date: 2024-12-06ZHEJIANG SCI-TECH UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410907146.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-08
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

In industrial scenarios, long-term operation of industrial equipment leads to inability to detect machine failures in real time, and due to the scarcity and diversity of abnormal sounds, it is difficult to effectively detect through supervised learning methods, especially in the presence of noise and interference. It is difficult to effectively identify abnormal sound signals.

Method used

Using a method based on similarity feature selection and decoupling, features are extracted through logarithmic Meier spectrogram, and potential variable factors are separated using the β-variant autoencoder model, and abnormal detection is performed in combination with similarity measurement and reconstruction error to eliminate noise interference. , improve detection efficiency.

Benefits of technology

There is no need to rely on machine ID to assist training, the training time is short, the inference speed is fast, and the memory consumption is small. It can effectively identify unknown abnormal phenomena, improving the accuracy and efficiency of abnormal detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119091908A_ABST
    Figure CN119091908A_ABST
Patent Text Reader

Abstract

The invention discloses an industrial equipment unsupervised abnormal sound signal detection method based on similar feature selection decoupling, and the method comprises the steps: 1) carrying out the feature extraction of an original sound signal, and obtaining a logarithmic Mel spectrogram (feature matrix); 2) performing time domain feature selection on the logarithmic Mel spectrogram by using a similarity measurement-based method, and calculating the similarity between each frame in the spectrogram and the whole spectrogram; 3) according to the similarity in the step 2), designing different selection strategies by utilizing the characteristics of stable and non-stable sound signals; 4) processing sound signals of normal work of the industrial equipment according to the steps 1)-3) to obtain a data set after feature selection; 5) inputting the training set into a beta-variational auto-encoder, and training by taking minimization of a reconstruction error and KL divergence as optimization targets; 6) taking the mean square error as an abnormal score, and setting a threshold value according to the abnormal score of the training set; and 7) inputting the test set into the trained network to calculate an abnormal score, and judging whether the network is abnormal or not according to a threshold value.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of abnormal sound signal detection, and in particular relates to an unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling. Background Art

[0002] In industrial scenarios, most production equipment runs almost non-stop around the clock. The long-term operation of the equipment makes it impossible to simply use manual methods to detect machine anomalies in real time. Therefore, it is crucial to detect machine failures in advance, which can effectively prevent major damage and economic losses. The acoustic signals of industrial equipment reflect rich information about the operation of the equipment. The purpose of industrial abnormal sound detection is to identify whether the sound emitted by industrial machines is normal or abnormal. In actual applications, due to the scarcity and diversity of abnormal equipment sounds, it is difficult to collect enough abnormal sound data for supervised learning. Therefore, how to perform anomaly detection when there is only normal sound data becomes a key issue.

[0003] At present, industrial abnormal sound detection mainly includes self-supervised methods and unsupervised methods. The self-supervised learning method introduces additional information such as machine ID, etc., so that the model focuses on distinguishing the characteristics of abnormal sounds. However, in actual scenarios, it is still challenging to try to learn discriminative features without classification labels; the unsupervised learning method usually does not require additional information assistance, and only requires normal sound samples for data training to perform abnormal sound detection.

[0004] In actual scenarios, in addition to the target machine, there may be other machines running, which makes the detected sound contain noise or other interference factors, resulting in little difference between normal sound and abnormal sound, which greatly reduces the performance of abnormality detection. Therefore, reducing noise and non-target sound information from the detected sound and selecting target sound information is crucial for industrial abnormal sound signal detection. Summary of the invention

[0005] In view of the above problems, the present invention proposes an unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling. Usually, the sound signal of industrial equipment is affected by noise or other interference. In addition to the main information of the target equipment, it also contains certain redundant information. The time frames containing the main information in the logarithmic Mel spectrum of the equipment sound signal often have similarities. Based on this, the present invention first extracts the logarithmic Mel spectrum of the normal sound signal; then, a method based on similarity measurement is used to calculate the similarity between each frame in the spectrum and other frames, and the spectrogram frames containing the main sound information in the time domain are screened out; then, for the spectrogram after feature selection, the β-variational autoencoder model is used for decoupling to separate the potential change factors in the selected spectrogram frame; the model uses the mean square error and KL divergence of the input and reconstructed output as the loss function, and uses the mean square error as the abnormality score to judge whether the input sound is abnormal or not.

[0006] First, the professional terms in the present invention are explained:

[0007] Log-Mel spectrogram is a commonly used audio signal processing method, which is used to convert audio signals into frequency domain representation. The process of generating a log-Mel spectrogram includes the following steps: first, the audio signal is collected and preprocessed (pre-emphasis, framing, windowing); then, each frame is subjected to a fast Fourier transform (FFT) to convert the time domain signal into a frequency domain signal; next, the power spectrum is calculated and a Mel filter bank is applied to convert the power spectrum into a Mel spectrum; finally, the Mel spectrum is logarithmically compressed to obtain a log-Mel spectrogram. In a log-Mel spectrogram, one dimension on the time axis is called a "spectrogram frame", and each frame represents the intensity or energy distribution of different frequency components.

[0008] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0009] The unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling uses a similarity measurement method to select a certain number of spectrogram frames from the logarithmic Mel spectrum of the equipment sound signal. Then, the selected spectrogram frame features are used as the input of the β-variational autoencoder network for anomaly detection. The specific steps are as follows:

[0010] (I) Model training

[0011] Step 1: extract the logarithmic Mel spectrum feature of the original sound signal, including pre-emphasis, framing, windowing, fast Fourier transform and logarithmic Mel filtering, to obtain its logarithmic Mel spectrum feature X.

[0012] Step 2: Perform time domain feature selection on the logarithmic Mel spectrum obtained in step 1, and use a similarity measurement method to process the feature matrix X corresponding to the logarithmic Mel spectrum. i With other frames x j (x≠j), and then the similarity between each frame and the entire logarithmic Mel-spectrogram is obtained.

[0013] Step 3: For the similarity calculated in step 2, according to the characteristics of different equipment sound signals, select the spectrogram frame containing the main information of the equipment sound signal (information that can accurately capture the sound characteristics of the machine rather than the interference of noise or other factors) to reduce the interference of noise and other factors. According to the characteristics of the stationary sound signal, select the most similar (greater than the set threshold) spectrogram frame; according to the characteristics of the non-stationary sound signal, select the least similar (less than the set threshold) spectrogram frame; finally, combine all the selected frames into a new spectrogram feature.

[0014] Step 4: Process the collected acoustic signals of the normal operation of the industrial equipment according to steps 1-3 to obtain the training data set X train ={x1,x2,...,x n}, the collected sound signals of normal and abnormal operation of industrial equipment are used as test data sets and processed according to steps 1-3 to obtain in, Represents the data of normal sound signal after feature selection, Represents the data of abnormal sound signal after feature selection.

[0015] Step 5: The training set X in step 4 train Input the β-variational autoencoder for training. In this step, the β-variational autoencoder consists of 5 fully connected layers with a dimension of 128-64-32-32-32, and finally outputs a latent variable with a dimension of 16*2. The decoder is the opposite, with a dimension of 32-32-32-64-128. The training process uses the Adam optimizer for parameter optimization.

[0016] For the model training in this step, the loss function is to minimize the reconstruction error and KL divergence, and the loss is expressed as:

[0017]

[0018] Among them, x is the input feature after feature selection, x' is the output after reconstruction by β-variational autoencoder, β is the control factor, KL() is the KL divergence, q φ (z|x) is the encoded distribution, p(z) is the prior distribution, and z is the latent variable.

[0019] Step 6: Express the anomaly score using the reconstruction loss in step 5. train The anomaly score is calculated by threshold A θ , used for abnormal judgment during testing.

[0020] (II) Anomaly Detection

[0021] Step 7: Input the test data set into the trained network to calculate the anomaly score and determine whether it is abnormal based on the threshold.

[0022] Preferably, step 1 includes pre-emphasis, framing, windowing, fast Fourier transform and logarithmic Mel filtering, as follows:

[0023] Step 1.1, for pre-emphasis, the calculation formula is:

[0024] H(n)=1-αn -1

[0025] Where n represents the index of discrete time or sequence, H(n) is the output after pre-emphasis, and α is the pre-emphasis coefficient;

[0026] Step 1.2, for framing, the frame length is L and the frame shift is

[0027] Step 1.3, for windowing, use the Hamming window, and the calculation formula is:

[0028]

[0029] Among them, a is the sample point number in the window, and A is the length of the window, that is, the number of sample points;

[0030] Step 1.4, for fast Fourier transform, convert the signal obtained in step 1.1-step 1.3 into the frequency domain, and the calculation formula is:

[0031]

[0032] Where e is the base of the natural logarithm, j is the imaginary unit, and j satisfies 2 =-1, y(n) is the windowed acoustic signal, Y(k1) is the frequency domain signal corresponding to y(n), and N is the number of Fourier transform points;

[0033] Step 1.5: Process the spectrum Y(k1) obtained in step 1.4 through a log-Mel filter to obtain the final output log-Mel spectrum feature. The calculation formula is:

[0034]

[0035] Where r∈R 1×k, R is a set of real numbers, k is the number of frames, ω M ∈R d×B is a Mel filter bank, B is each discrete frequency component of the spectrum obtained by FFT, and d represents the dimension.

[0036] Preferably, step 2 is as follows:

[0037] Step 2.1: Generate a logarithmic Mel spectrum feature after processing in step 1 Defined as:

[0038]

[0039] Among them, x i ∈R d×1 ,i∈[1,2,...,k], d represents the dimension, k represents the number of frames;

[0040] Step 2.2, the similarity between each frame in the spectrogram and other frames is expressed as:

[0041] similarity(x i ,x j ),i≠j,(i,j=[1,2,...,k])

[0042] Step 2.3, x i The similarity S between the whole spectrum graph i It is expressed as:

[0043]

[0044] Preferably, in step 3, for a stationary sound signal, a spectrogram frame whose similarity is greater than a set value is selected; for a non-stationary sound signal, a spectrogram frame whose similarity is less than a set value is selected.

[0045] Preferably, in step 5, the β-variational autoencoder consists of 5 fully connected layers with a dimension of 128-64-32-32-32, and finally outputs a latent variable with a dimension of 16*2; the dimension of the decoder is 32-32-32-64-128, and the Adam optimizer is used.

[0046] Preferably, in step 5, the loss function is expressed as:

[0047]

[0048] Among them, x is the feature spectrum input after feature selection, x' is the output after reconstruction by β-variational autoencoder, β is the control factor, KL() is the KL divergence, q φ (z|x) is the encoded distribution, p(z) is the prior distribution, and z is the latent variable.

[0049] Preferably, in step 6, the threshold Among them, x is the feature after feature selection; k2 represents the number of time frames after feature selection; d represents the dimension of the data; ψ t is the representation of the original input feature at time frame t; Φ is the β-variational autoencoder network; Φ(ψ t ) is the latent representation reconstructed from the latent space; is l2 regularization.

[0050] Preferably, in step 7, for a collected unknown sound signal data, the following processing is performed to determine whether it is abnormal:

[0051] Step 7.1: Extract the logarithmic Mel spectrum feature matrix according to step 1 in (a)

[0052] Step 7.2, perform logarithmic Mel spectrum feature selection according to steps 2-3 in (a).

[0053] In step 7.3, the spectrogram frame feature x selected in step 7.2 is input into the β-variational autoencoder to obtain its reconstructed sample x'; then, the reconstruction loss is calculated as the anomaly score S.

[0054] Step 7.4: Compare the anomaly score obtained in step 7.3 with the threshold A θ Compared with, greater than threshold A θ is an abnormal signal, otherwise it is a normal signal, expressed as:

[0055]

[0056] The beneficial effects of the present invention are as follows:

[0057] There is no need to rely on machine ID for auxiliary training, the training time is short, the reasoning speed is fast, the memory usage is small, and unknown abnormal phenomena can be discovered. The present invention has strong adaptability, does not require manual intervention, and improves the efficiency of abnormality detection tasks for complex audio data.

[0058] Compared with the prior art method of using the entire logarithmic Mel spectrum for detection, the present invention can effectively screen out sound information that is effective for anomaly detection, while eliminating the interference of noise and other redundant factors, thereby enhancing the expressiveness of features and improving detection efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1This is a framework diagram for unsupervised abnormal sound signal detection of industrial equipment based on similarity feature selection decoupling; where Encoder is the encoder; Decoder is the decoder; μ is the mean; σ is the variance; z is the latent variable; β is a hyperparameter used to control the weight of the reconstruction loss and the KL divergence; MSE is the mean square error.

[0060] Figure 2 Schematic diagram of stationary signal and non-stationary signal.

[0061] Figure 3 It is the structure diagram of β-variational autoencoder; Linear is the linear layer, ReLU is the activation function, and BatchNormld is batch normalization.

[0062] Figure 4 is the specific implementation process diagram of β-variational autoencoder; where x tr and x te are the inputs for training and testing respectively, x t ' r and x t ' e are the reconstructed training input and the reconstructed test input, respectively, q φ (z tr |x tr ) and q φ (z te |x te ) are the encoding distributions for training and testing, respectively, p θ (x tr |z t ' r ) and p θ (x tr |z t ' e ) are the probability distributions of the observed data x for training and testing, respectively, and z tr and z te The input data x tr and x te The potential representation of z t ' r and z t ' e are new latent variables for training and testing, μ tr and μ te are the means of training and testing, σ tr and σ te are the standard deviations of training and testing, respectively, KL (N(μ tr ,σ tr )||N(0,I)) is the KL divergence.

[0063] Figure 5 This is a flowchart of unsupervised abnormal sound signal detection for industrial equipment based on similarity feature selection decoupling. Where log-Mel is the logarithmic Mel spectrum. DETAILED DESCRIPTION

[0064] The following takes a data set in industrial equipment as an example, and uses an unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling to further illustrate the present invention. The following description is only for illustration and does not impose any form of limitation on the present invention.

[0065] In the present invention, the public data sets MIMII and ToyADMOS for industrial machine sound anomaly detection are used. The MIMII data set contains four types of industrial equipment: valve, pump, slider, and fan. Each type of industrial equipment has four machine IDs. The data set for each machine ID consists of approximately 1,000 normal sound signal samples for training and 100-200 normal and abnormal sound signal samples for testing. The sampling rate is 16KHz and the duration is 10s. The ToyADMOS data set contains two types of industrial equipment: ToyCar and ToyConveyor, which have four and three machine IDs, respectively. The data set for each machine ID consists of approximately 1,000 normal sound signal samples for training and 200-400 normal and abnormal sound signal samples for testing. The sampling rate is 16KHz, and the duration is 11s and 10s, respectively.

[0066] like Figure 5 As shown, this embodiment is an unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling, and the specific steps are as follows:

[0067] Model training:

[0068] Step 1: extract the logarithmic Mel spectrum feature of the original sound signal, including pre-emphasis, framing, windowing, fast Fourier transform and logarithmic Mel filtering, to obtain its logarithmic Mel spectrum feature X. Specifically:

[0069] Step 1.1, for pre-emphasis, the calculation formula is:

[0070] H(n)=1-αn -1

[0071] Wherein, n represents the index of discrete time or sequence, H(n) is the output after pre-emphasis, and α is the pre-emphasis coefficient, which is 0.97.

[0072] Step 1.2, for framing, the frame length is L and the frame shift is L is set to 64 milliseconds.

[0073] Step 1.3, for windowing, use the Hamming window, and the calculation formula is:

[0074]

[0075] Among them, a is the sequence number of the sample point in the window, and A is the length of the window, that is, the number of sample points.

[0076] Step 1.4, for fast Fourier transform, convert the signal obtained in step 1.1-step 1.3 into the frequency domain, and the calculation formula is:

[0077]

[0078] Where e is the base of the natural logarithm, j is the imaginary unit, and j satisfies 2 =-1, y(n) is the windowed acoustic signal, Y(k1) is the frequency domain signal corresponding to y(n), and N is the number of Fourier transform points, which is 512 here.

[0079] Step 1.5: Process the spectrum Y(k1) obtained in step 1.4 through a logarithmic Mel filter to obtain the final output: The calculation formula is:

[0080]

[0081] Where r∈R 1×k , R is a set of real numbers, k is the number of frames, ω M ∈R d×B is the Mel filter bank, and B is each discrete frequency component of the spectrum obtained by FFT. Logarithmic Mel energy d represents the dimension, which is set to 128, the window size is set to 1024, and the frame shift is 512.

[0082] Step 2: Perform time domain feature selection on the logarithmic Mel spectrum obtained in step 1.5, process the logarithmic Mel spectrum feature matrix X, and measure each frame x in the spectrum. i With other frames x j (i≠j), and then the similarity between each frame and the entire logarithmic Mel-spectrogram is obtained.

[0083] Step 2.1, first, generate a logarithmic Mel spectrum matrix after processing in step 1 Defined as:

[0084]

[0085] where x i ∈R d×1 ,i∈[1,2,...,k], where d represents the dimension and k represents the number of frames;

[0086] Step 2.2, Secondly, the similarity between each frame in the spectrogram and other frames can be expressed as:

[0087] similarity(x i ,x j ),i≠j,(i,j=[1,2,...,k])

[0088] Step 2.3, therefore, x i The similarity S between the whole spectrum graph i It is expressed as:

[0089]

[0090] Step 3: Based on the similarity obtained in step 2, according to the characteristics of different equipment sound signals, select the spectrogram frame containing the main information of the equipment sound signal to reduce the interference of noise and other factors; according to the characteristics of the stationary sound signal, select the most similar spectrogram frame; according to the characteristics of the non-stationary sound signal, select the least similar spectrogram frame; finally, combine all the selected spectrogram frames into a new spectrogram feature;

[0091] Since the main information of the stationary sound signal is distributed throughout the time interval, almost all spectrogram frames in its logarithmic Mel-spectrogram show the frequency characteristics of the target sound signal, which leads to similarity between the spectrogram frames of the target sound. For non-stationary sound signals, the main information is relatively scattered, and only a few spectrogram frames show the frequency characteristics of the target signal, while most spectrogram frames are related to noise or other redundant factors, which leads to less similarity between the spectrogram frames of the target sound. For stationary sound signals, spectrogram frames with greater similarity are selected; and for non-stationary sound signals, spectrogram frames with less similarity are selected.

[0092] Step 4: Process the collected acoustic signals of the normal operation of the industrial equipment according to steps 1-3 to obtain the training data set X train ={x1,x2,...,x n}, the collected sound signals of normal and abnormal operation of industrial equipment are used as test data sets and processed according to steps 1-3 to obtain in, Represents the data of normal sound signal after feature extraction, Represents the data of abnormal sound signal after feature extraction.

[0093] Step 5: Transform the training set X trainAfter the selection process based on similarity metric, the β-variational autoencoder is input for training. The β-variational autoencoder consists of 5 fully connected layers with dimensions of 128-64-32-32-32, and finally outputs a latent variable with 16*2 dimensions. The decoder is the opposite, with dimensions of 32-32-32-64-128, and uses the Adam optimizer.

[0094] The training objective is to minimize the reconstruction error and KL divergence, and the mean square error and KL divergence are used as the loss function. The loss is expressed as:

[0095]

[0096] Among them, x is the input feature after feature selection, x' is the output after reconstruction by β-variational autoencoder, β is the control factor, KL() is the KL divergence, q φ (z|x) is the encoded distribution, p(z) is the prior distribution, and z is the latent variable.

[0097] Step 6: For the anomaly score and threshold, the loss in step 5 is used to represent the anomaly score, and the training data set X train The anomaly score is calculated by threshold A θ :

[0098]

[0099] Among them, x is the feature after feature selection; k2 represents the number of time frames after feature selection; d represents the dimension of the data; ψ t is the representation of the original input feature at time frame t; Φ is the β-variational autoencoder network; Φ(ψ t ) is the latent representation reconstructed from the latent space; is l2 regularization.

[0100] Step 7: Input the test data set into the trained network to calculate the anomaly score and determine whether it is abnormal based on the threshold. The following are the anomaly detection steps:

[0101] Step 7.1: Extract the logarithmic Mel spectrum feature matrix according to step 1

[0102] Step 7.2, perform logarithmic Mel-spectrogram feature selection according to the above steps 2-3.

[0103] In step 7.3, the spectral feature x selected in step 7.2 is input into the β-variational autoencoder to obtain its reconstructed sample x'; then, the reconstruction loss is calculated as the anomaly score S.

[0104] Step 7.4: Compare the anomaly score obtained in step 7.3 with the threshold A θCompared with, greater than threshold A θ is an abnormal signal, otherwise it is a normal signal, expressed as:

[0105]

[0106] Any matters not described in the present invention are applicable to the prior art.

[0107] 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 abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling, characterized by The specific steps include: Step 1: Extract the logarithmic Mel spectrum feature of the original sound signal to obtain the logarithmic Mel spectrum feature Step 2: Perform time domain feature selection on the obtained logarithmic Mel spectrum graph. Processing is performed by measuring each frame x in the spectrogram i With other frames x j (similarity between them, and then obtain the similarity between each frame and the entire logarithmic Mel-spectrogram, where i≠j); Step 3: Based on the similarity obtained in step 2, according to the characteristics of different equipment sound signals, select the spectrogram frame containing the main information of the equipment sound signal; according to the characteristics of the stationary sound signal, select the most similar spectrogram frame; according to the characteristics of the non-stationary sound signal, select the least similar spectrogram frame; finally, combine all the selected spectrogram frames into a new spectrogram feature; Step 4: Process the collected acoustic signals of the normal operation of the industrial equipment according to steps 1-3 to obtain the training data set X train ={x1,x2,...,x n }, the collected sound signals of normal and abnormal operation of industrial equipment are used as test data sets and processed according to steps 1-3 to obtain in, represents the data of normal sound signal after feature selection, x f i Represents the data of abnormal sound signal after feature selection; Step 5: The training set X in step 4 train After the selection processing based on similarity measurement, it is input into the β-variational autoencoder for training, with the loss function of minimizing the reconstruction error and KL divergence; Step 6: Express the anomaly score using the reconstruction loss in step 5; according to the training dataset X train The anomaly score is calculated by threshold A θ ; Step 7: Input the test data set into the trained network to calculate the anomaly score and determine whether it is abnormal based on the threshold.

2. The unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling as claimed in claim 1, characterized in that: Step 1 includes pre-emphasis, framing, windowing, fast Fourier transform and logarithmic Mel filtering, as follows: Step 1.1, for pre-emphasis, the calculation formula is H(n) = 1-αn -1 , where n represents the index of discrete time or sequence, H(n) is the output after pre-emphasis, and α is the pre-emphasis coefficient; Step 1.2, for framing, the frame length is L and the frame shift is Step 1.3, for windowing, use the Hamming window, and the calculation formula is Among them, a is the sample point number in the window, and A is the length of the window, that is, the number of sample points; Step 1.4, for fast Fourier transform, convert the signal obtained in step 1.1-step 1.3 into the frequency domain, and the calculation formula is: Where e is the base of the natural logarithm, j is the imaginary unit, and j satisfies 2 =-1, y(n) is the windowed acoustic signal, Y(k1) is the frequency domain signal corresponding to y(n), and N is the number of Fourier transform points; Step 1.5: Process the spectrum Y(k1) obtained in step 1.4 through a log-Mel filter to obtain the final output log-Mel spectrum feature. The calculation formula is: Where r∈R 1×k , R is a set of real numbers, k is the number of frames, ω M ∈R d×B is a Mel filter bank, B is each discrete frequency component of the spectrum obtained by FFT, and d represents the dimension.

3. The unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling as claimed in claim 1, characterized in that: Step 2 is as follows: Step 2.1: Generate a logarithmic Mel spectrum feature after processing in step 1 Defined as Among them, x i ∈R d×1 ,i∈[1,2,...,k], d represents the dimension, k represents the number of frames; Step 2.2, the similarity between each frame in the spectrogram and other frames is expressed as: similarity(x i ,x j ),i≠j,(i,j=[1,2,...,k]) Step 2.3, x i The similarity S between the whole spectrum graph i It is expressed as:

4. The unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling as claimed in claim 1, characterized in that: In step 3, for a stationary sound signal, a spectrogram frame whose similarity is greater than a set value is selected; for a non-stationary sound signal, a spectrogram frame whose similarity is less than a set value is selected.

5. The unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling as claimed in claim 1, characterized in that: In step 5, the β-variational autoencoder consists of 5 fully connected layers with a dimension of 128-64-32-32-32, and finally outputs a latent variable with a dimension of 16*2; the decoder has a dimension of 32-32-32-64-128 and uses the Adam optimizer.

6. The unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling as claimed in claim 1, characterized in that: In step 5, the loss function is expressed as: Among them, x is the feature spectrum input after feature selection, x' is the output after reconstruction by β-variational autoencoder, β is the control factor, KL() is the KL divergence, q φ (z|x) is the encoded distribution, p(z) is the prior distribution, and z is the latent variable.

7. The unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling according to any one of claims 1 to 6, characterized in that: In step 6, the threshold Among them, x is the feature after feature selection; k2 represents the number of time frames after feature selection; d represents the dimension of the data; ψ t is the representation of the original input feature at time frame t; Φ is the β-variational autoencoder network; Φ(ψ t ) is the latent representation reconstructed from the latent space; is l2 regularization.

8. The unsupervised abnormal sound signal detection method for industrial equipment based on similarity feature selection decoupling as claimed in claim 7, characterized in that: Step 7 is as follows: Step 7.1: Extract the logarithmic Mel spectrum features according to step 1 Step 7.2: Perform logarithmic Mel spectrum feature selection according to steps 2-3; Step 7.3: Input the spectrogram frame feature x selected in step 7.2 into the β-variational autoencoder to obtain its reconstructed sample x'; then, calculate the reconstruction loss as the anomaly score S; Step 7.4: Compare the anomaly score obtained in step 7.3 with the threshold A θ Compare, if it is greater than the threshold A θ It is an abnormal signal. If it is less than or equal to the threshold A θ , it is a normal signal.

Citation Information

Cited By

  • Voice recognition method and device, electronic equipment, storage medium and program product

    CN119580716A