A feature extraction method for abnormal loudspeaker sound based on auditory masking combined with SVD-MRMR

Through the SVD-MRMR method based on auditory masking, the abnormal sound characteristics of the speaker are extracted, and the subjectivity and accuracy problems of the traditional detection methods are solved, and efficient and stable detection of abnormal sounds of the speaker is achieved.

CN115002642BActive Publication Date: 2024-11-15GUILIN UNIV OF ELECTRONIC TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210561080.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-23
Publication Date
2024-11-15
Estimated Expiration
2042-05-23

AI Technical Summary

Technical Problem

Traditional speaker abnormal sound detection methods rely on human ear listening, which has problems of subjective differences and low accuracy, and signal processing methods fail to effectively pay attention to the auditory perception characteristics of human ear, resulting in poor classification results.

Method used

Using a method based on auditory masking combined with SVD-MRMR, the auditory perception masking spectrum of the speaker response signal is calculated through the ITU-RBS.1387-1 psychoacoustic model, and singular value decomposition is performed. Combined with the MRMR algorithm, the optimal feature subset is selected to eliminate redundant Remaining information.

Benefits of technology

The stability and consistency detection of abnormal sound of speakers is achieved, the subjective influence of manual, the detection accuracy and production efficiency are improved, and the product quality is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115002642B_ABST
    Figure CN115002642B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of speaker quality detection, and specifically to a feature extraction method for abnormal speaker sound based on auditory masking combined with SVD-MRMR. The present invention first analyzes the speaker response signal of the collected sound by simulating artificial listening, and calculates the auditory perception masking spectrum of the speaker; then the auditory perception masking spectrum is decomposed by SVD to obtain a singular value feature sequence; then the MRMR algorithm is used to extract the optimal features in the singular value feature sequence, eliminate redundant information, obtain the features after dimensionality reduction, and eliminate redundant information, so as to avoid too many features in the classification process, resulting in too high a dimension of the classification model, affecting the efficiency of the model, and being more conducive to unified measurement and judgment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of speaker quality detection, and in particular to a method for extracting abnormal speaker sound features based on auditory masking combined with SVD-MRMR. Background Art

[0002] Speakers and their systems are key components of audio equipment, and their quality directly affects the sound quality of the entire sound system; however, during the production process, voice coil bottoming, diaphragm rupture, and excessive impedance can all lead to abnormal speaker sound. Therefore, it is of great practical significance to detect and classify abnormal sound from produced speakers.

[0003] The traditional method of detecting abnormal sound from loudspeakers is for a listener to listen to the sound with his or her ears. The listening results are affected by the complexity of the technology, the physical and mental state of the workers, and the degree of work fatigue. There are large subjective differences in the test results. It is difficult to further improve the accuracy and stability of loudspeaker detection, which can easily cause hearing damage, and it is difficult to further improve the company's production efficiency and product quality.

[0004] Signal processing methods can be used to detect abnormal speaker sounds, for example, by detecting the time-frequency diagram obtained by short-time Fourier transform or waveform transform, and by decomposing the speaker response signal and then detecting the time-frequency features, such as using LMD energy entropy (Chen Xiurong. Research on speaker abnormal sound fault diagnosis based on LMD and LSSVM [D]. Guilin University of Electronic Technology, 2020.) and VMD energy entropy (Zhou Jinglei, Yan Ting. Classification of abnormal speaker sounds using variational mode decomposition and energy entropy [J]. Journal of Acoustics, 2021, 46(02): 263-270.) and other feature extraction methods. Although the above methods can have a good effect on the identification and classification of abnormal speaker sounds, they do not pay attention to the auditory perception characteristics of the human ear. Some methods also extract the mean of psychoacoustic energy (Guo Qing, He Jiekai, Su Haitao, Wang Xingjian. Speaker abnormal sound detection algorithm based on psychoacoustics and support vector machine [J]. Journal of Donghua University (Natural Science Edition), 2020, 46(02): 275-281.) as a feature to classify abnormal speaker sounds, but the classification results are poor due to the high dimension. Summary of the invention

[0005] In order to solve the above problems, the present invention provides a feature extraction method for abnormal speaker sound based on auditory masking combined with SVD-MRMR. The specific technical solution is as follows:

[0006] A method for extracting abnormal speaker sound features based on auditory masking combined with SVD-MRMR, comprising the following steps:

[0007] A method for extracting abnormal speaker sound features based on auditory masking combined with SVD-MRMR, comprising the following steps:

[0008] Step S1: Use a microphone to collect the sound signal of the speaker. Suppose the sampling frequency of the microphone is f s , in Hz; select a logarithmic sweep signal of 20 to 20000 Hz as the excitation signal, the duration of the sound signal is T, in seconds; calculate the auditory perception masking spectrum E[k,n] of the loudspeaker response signal x(t) according to the ITU-RBS.1387-1 psychoacoustic model, k is the number of subbands, n is the number of frames, n refers to the number of frames in which the duration T of the sound signal is divided according to a frame length of 2048 and an overlap rate of 50%,

[0009] Step S2: construct the auditory perception masking spectrum E[k,n] of each sub-band of the loudspeaker response signal into a k×n matrix E with a rank of r (r=min(k,n)). Then perform SVD decomposition on the matrix E to obtain a singular value eigenvector s. The eigenvector s obtained for each sample constitutes a feature set X of the loudspeaker psychoacoustic energy spectrum. h×r , where h is the number of speaker samples and r is the feature dimension;

[0010] Step S3: Use the MRMR algorithm to extract the h×r Select the optimal feature subset Z from the , and eliminate redundant information.

[0011] Preferably, the step S1 specifically includes the following steps:

[0012] Step S11, the response signal x(t) of the loudspeaker is processed by framing, and each frame signal x n [t,n] is:

[0013]

[0014] Where t is the number of time domain sampling points per frame; N F is the number of samples per frame;

[0015] Step S12: Frame the loudspeaker response signal x n [t,n] uses Hanning window filtering and fast Fourier transform to convert the signal from time domain to frequency domain to obtain the spectrum F f [k f ,n]:

[0016] F f [k f ,n]=fft(h[t]·x n [t,n]);

[0017] Where: kf is the number of frequency domain points; h[t] is the Hanning window function;

[0018] Step S13, simulating the auditory characteristics of the human philtrum and outer ear, and performing the obtained frequency spectrum F f [k f ,n] to perform frequency domain weighting to obtain the output F after the middle and outer ear frequency weighting e [k f ,n]:

[0019]

[0020] Where W[f] is the frequency response function of the middle and outer ears;

[0021]

[0022] Step S14, a Bark scale is evenly divided into 4 frequency bands, the frequency band width is Δz = 1 / 4, and the Bark domain is evenly divided into 109 non-overlapping critical frequency sub-bands in the auditory range of 20Hz to 18kHz, k = 1, 2, ..., 109; the signal energy value in each sub-band is added to obtain the energy output spectrum P in a frequency sub-band e [k,n], where k represents the kth critical frequency subband;

[0023] The energy distribution of each subband is calculated by the following formula:

[0024]

[0025] where f u [k], f l [k] are the low-frequency and high-frequency of the k-th critical frequency subband respectively;

[0026] The energy output spectrum P in a frequency subband e [k,n] is calculated by the following formula:

[0027] P e [k,n]=max(∑U[k,k f ](F e [k f ,n]) 2 ,10 -12 );

[0028] Step S15, give the energy output spectrum P e [k,n] add a section of internal noise P noise [k], add them together to get P p [k,n]:

[0029]

[0030] P p [k,n]=P e [k,n]+P noise [k];

[0031] Where fc is the center frequency of the frequency subband;

[0032] Step S16: Apply the frequency extension function in the Terhardt psychoacoustic model to simulate the frequency domain masking effect; distribute the energy of each subband to the entire Bark domain space through the extension function, then the energy of the i-th subband is the weighted sum of the contributions of each subband energy in the subband, and the extension function is recorded as S dB (i,k,n) represents the contribution of the k-th subband energy to the i-th subband. The formula is as follows:

[0033]

[0034] Calculate the energy distribution E of each subband after frequency domain expansion s [k,n]:

[0035]

[0036] Where: B s (i) is the normalization factor;

[0037]

[0038] Step S117, perform time domain expansion, and use a first-order low-pass filter to simulate the time domain masking effect:

[0039] E f [k,n]=a·E f [k,n-1]+(1-a)·E s [k,n];

[0040] Among them, the parameter a is determined by the time coefficient and is related to the center frequency of each sub-band; thus, the auditory perception masking spectrum E[k,n] of each sub-band after time domain masking can be obtained:

[0041] E[k,n]=max(E f [k,n],E s [k,n]).

[0042] Preferably, the relationship between the parameter a and the time coefficient τ is as follows:

[0043]

[0044]

[0045] Where: τ 100 , τmin is a constant, generally τ 100 =0.03, τ min =0.008, f c [k] is the center frequency of each sub-band.

[0046] Preferably, the expression for performing SVD decomposition on the matrix E in step S2 is as follows:

[0047]

[0048] Where S is a k*k matrix, V is an n*n matrix, both S and V are unitary matrices, ∑=diag(σ1,σ2,...,σ n ); the singular values ​​of the matrix E are σ i , i=1,2,...,r and σ1≥σ2≥...≥σ r ≥0; then the speaker response signal eigenvector s=(σ1,σ2,...,σ r );The feature vector s obtained from each sample constitutes X h×r ={X1,X2,...X r} is the feature set of the loudspeaker psychoacoustic energy spectrum, where X i (i=1,2,...,r) are the singular values ​​σ of the auditory perception masking spectrum E of each loudspeaker response signal i (i=1,2,...,r), and then assign the labels corresponding to the speaker response signals of four different states, namely, normal, ring contact, air leakage, and low sound, to 1, 2, 3, and 4 respectively. Assume that Y h×1 ={Y1,Y2,...Y h} is the category set of speaker features, where h is the number of speaker samples and r is the feature dimension.

[0049] Preferably, the step S3 specifically includes the following steps:

[0050] Step S31: feature set X h×r Normalization, calculate the feature set X according to the MRMR method h×r Each feature X i The importance of R x , the formula is as follows:

[0051]

[0052] Among them, L x For each feature X i and the correlation between category Y, L x For each feature X i and its category Y mutual information, that is, L x =I(X i,Y);

[0053] N x For each feature X i and other features X in set Z l The redundancy between, where i≠l, N x Determined by the mutual information between each feature and other features:

[0054]

[0055] Step S32: According to the importance R of each feature x Sort in descending order, input the features into the classifier one by one for evaluation, and select the feature subset with the highest classification accuracy. If the input of the mth feature subset makes the classifier have the highest accuracy, it means that the first m features are selected, and the optimal features are selected from them to form feature subset Z through the MRMR criterion.

[0056] The beneficial effects of the present invention are as follows: the traditional method for detecting abnormal sound of a loudspeaker relies on completion, and it is difficult to use a stable and consistent standard to judge the type of abnormal sound of the loudspeaker. The present invention simulates artificial listening based on the ITU-RBS.1387-1 psychoacoustic model to analyze the loudspeaker response signal, eliminating the influence of the subjective feeling of the listening artifact. The present invention proposes that the loudspeaker response signal is subjected to singular value decomposition (SVD) after being processed by the psychoacoustic model, and the psychoacoustic energy singular values ​​are obtained as the feature set to be selected, which is convenient for feature extraction of fault information. The present invention selects the optimal features from the feature set by utilizing the maximum relevance minimum redundancy (MRMR) algorithm and eliminates redundant information, which is more conducive to unified measurement and judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0058] Figure 1 It is a flow chart of the detection implementation of the present invention;

[0059] Figure 2 It is the time domain diagram of the response signal of a normal speaker;

[0060] Figure 3 It is the time domain diagram of the response signal of the touch circle speaker;

[0061] Figure 4 is the time domain diagram of the response signal of the leaky speaker;

[0062] Figure 5It is the time domain diagram of the response signal of the small sound speaker;

[0063] Figure 6 is the psychoacoustic masking energy spectrum of the response signal of a normal loudspeaker;

[0064] Figure 7 is the psychoacoustic masking energy spectrum of the response signal of the touch circle speaker;

[0065] Figure 8 Psychoacoustic masking energy spectrum for the response signal of the leaky loudspeaker;

[0066] Fig. 9 is the psychoacoustic masking energy spectrum of the response signal of the small sound speaker;

[0067] Fig.10 The psychoacoustic energy singular values ​​of the loudspeaker signal in four states after SVD decomposition;

[0068] Fig.11 Classification accuracy curves of the test set with different numbers of features.

[0069] Fig.12 Visualization result of feature extraction by the method of the present invention

[0070] Fig.13 Visualization results of feature extraction using the LMD energy entropy method

[0071] Fig.14 Visualization results of feature extraction using VMD energy entropy method

[0072] Fig.15 Visualization results of features extracted by psychoacoustic energy mean method DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0075] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0076] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0077] like Figure 1 As shown, a specific embodiment of the present invention provides a feature extraction method for abnormal speaker sound based on auditory masking combined with SVD-MRMR, comprising the following steps:

[0078] Step S1: Use a microphone to collect the sound signal of the speaker. Suppose the sampling frequency of the microphone is f s , in Hz; a logarithmic sweep signal of 20 to 20000 Hz is selected as the excitation signal, and the duration of the sound signal is T, in seconds; the auditory perception masking spectrum E[k,n] of the loudspeaker response signal x(t) is calculated according to the ITU-RBS.1387-1 psychoacoustic model, where k is the number of subbands and n is the number of frames, wherein k is a Bark scale that is evenly divided into 4 frequency bands, and the frequency band width is Δz=1 / 4, and the Bark domain is evenly divided into 109 non-overlapping critical frequency subbands in the auditory range of 20 Hz to 18 kHz, k=1,2,...,109, and n refers to the number of frames for dividing the duration T of the sound signal according to a frame length of 2048 and an overlap rate of 50%,

[0079] The test object of this experiment is a speaker unit with a diameter of 4 cm, a power of 0.5 W, and an impedance of 8 Ω. In the embodiment of the present invention, a logarithmic sweep frequency signal of 20 to 20000 Hz is selected as the excitation signal, and the microphone used is the Danish Bruel & The company's Type 4966-H-041 microphone, the main performance parameters of the 4966 microphone are: dynamic range of 14.6 ~ 144dB, frequency range of 5Hz ~ 20000Hz, inherent noise of 144.6dB A, lower limit frequency -3dB: 3Hz, sensitivity of 50mV / Pa, sampling rate of 65536Hz. This experimental data acquisition system uses the Danish B&K acoustic and vibration acquisition and analysis system as the acquisition system, and saves the collected speaker response signal.

[0080] The main tests were conducted on speakers in four states: normal, low-sound, ring-touching, and air-leaking. There were 50 speakers in each state, 50 in each state, for a total of 200 speakers. Each speaker was tested 5 times, for a total of 1,000 sample sets. The time domain diagrams of the response signals of various speakers were collected as shown below: Figure 2-Figure 5 ,in Figure 2 is the response signal of a normal speaker, Figure 3 is the response signal of the touch circle speaker, Figure 4 is the response signal of the air leakage speaker, Figure 5 It is the response signal of the small speaker.

[0081] right Figure 2-Figure 5 The auditory perception masking spectrum E[k,n] of the loudspeaker response signal is calculated according to the ITU-R BS.1387-1 psychoacoustic model, as follows:

[0082] The specific steps include:

[0083] Step S11, the response signal x(t) of the loudspeaker is processed by framing, and each frame signal x n [t,n] is:

[0084] Where t is the number of time domain sampling points per frame; N F is the number of samples per frame;

[0085] Step S12: Frame the loudspeaker response signal x n [t,n] uses Hanning window filtering and Fast Fourier Transformation (FFT) to convert the signal from time domain to frequency domain to obtain the spectrum F f [k f ,n]:

[0086] F f [k f ,n]=fft(h[t]·x n [t,n]);

[0087] Where: k f is the number of frequency domain points; h[t] is the Hanning window function;

[0088] Step S13, simulating the auditory characteristics of the human philtrum and outer ear, and performing the obtained frequency spectrum F f [k f ,n] to perform frequency domain weighting to obtain the output F after the middle and outer ear frequency weighting e [k f ,n]:

[0089]

[0090] Where W[f] is the frequency response function of the middle and outer ears;

[0091]

[0092] Step S14, Bark scale critical subband mapping, scale transformation is a nonlinear transformation process of sound frequency that simulates the Bark resonance characteristics of the basilar membrane in the human cochlea, which can map the linear frequency domain to the psychoacoustic frequency domain (Bark domain). After the signal is transformed to the Bark domain, the response signal is given a high-resolution nonlinear stretching, which has an obvious amplification effect, thereby being able to more delicately and intuitively portray the high and low frequency components in the signal.

[0093] A Bark scale is evenly divided into 4 frequency bands, with a frequency band width of Δz = 1 / 4. The Bark domain is evenly divided into 109 non-overlapping critical frequency sub-bands in the auditory range of 20Hz to 18kHz. The signal energy values ​​in each sub-band are added together to obtain the energy output spectrum P in the sub-band at a frequency. e [k,n], where k represents the kth critical frequency subband;

[0094] The energy distribution of each subband is calculated by the following formula:

[0095]

[0096] where f u [k], f l [k] are the low-frequency and high-frequency of the k-th critical frequency subband respectively;

[0097] The energy output spectrum P in a frequency subband e [k,n] is calculated by the following formula:

[0098] P e [k,n]=max(∑U[k,k f ](F e [k f ,n]) 2 ,10 -12 );

[0099] Step S15: Since human blood and heartbeat may affect human hearing to a certain extent, the energy output spectrum P is given. e [k,n] add a section of internal noise P noise [k], add them together to get P p [k,n]:

[0100]

[0101] P p[k,n]=P e [k,n]+P noise [k];

[0102] Among them, f c is the center frequency of the frequency sub-band.

[0103] Step S16: Apply the frequency extension function in the Terhardt psychoacoustic model to simulate the frequency domain masking effect; distribute the energy of each subband to the entire Bark domain space through the extension function, then the energy of the i-th subband is the weighted sum of the contributions of each subband energy in the subband; the extension function is denoted as S dB (i,k,n) represents the contribution of the k-th subband energy to the i-th subband. The formula is as follows:

[0104]

[0105] Calculate the energy distribution E of each subband after frequency domain expansion s [k,n]:

[0106]

[0107] Where: B s (i) is the normalization factor;

[0108]

[0109] Step S117, perform time domain expansion, and use a first-order low-pass filter to simulate the time domain masking effect:

[0110] E f [k,n]=a·E f [k,n-1]+(1-a)·E s [k,n];

[0111] Among them, the parameter a is determined by the time coefficient and is related to the center frequency of each sub-band; thus, the auditory perception masking spectrum E[k,n] of each sub-band after time domain masking can be obtained:

[0112] E[k,n]=max(E f [k,n],E s [k,n]).

[0113] The relationship between the parameter a and the time coefficient τ is as follows:

[0114]

[0115]

[0116] Where: τ 100 , τ minis a constant, generally τ 100 =0.03, τ min =0.008, f c [k] is the center frequency of each sub-band.

[0117] The psychoacoustic masking energy spectrum is obtained as Figure 6-Figure 9 shown. Figure 6 is the psychoacoustic masking energy spectrum of a normal loudspeaker. Figure 7 is the psychoacoustic masking energy spectrum of the touch circle speaker, Figure 8 is the psychoacoustic masking energy spectrum of the leaky loudspeaker, Fig. 9 is the psychoacoustic masking energy spectrum of the small speaker. Figure 6-Figure 9 It can be analyzed that the loudspeaker response signals of different fault types have the largest energy proportion in the interval of critical frequency sub-bands 49 to 64, and the corresponding center frequency is 2026.266 to 3519.344 Hz, which is in line with the characteristic that the human ear is most sensitive to sounds of 1000 to 3000 Hz. In addition, the energy amplitude of the response signals of different fault types in this interval is also different. Therefore, the use of auditory perception masking spectrum can simulate the keen human auditory perception system, suppress the low and high frequency parts of the collected loudspeaker response signals, and enhance the mid-frequency part of the signals, so that the analysis results are more in line with the auditory perception of the human ear.

[0118] Step S2: construct the auditory perception masking spectrum E[k,n] of each sub-band of the loudspeaker response signal into a k×n matrix E with a rank of r (r=min(k,n)), and then perform SVD decomposition on the matrix E to obtain a singular value feature sequence; the expression for SVD decomposition of the matrix E is as follows:

[0119]

[0120] Where S is a k*k matrix, V is an n*n matrix, both S and V are unitary matrices, ∑=diag(σ1,σ2,...,σ r ); the singular values ​​of the matrix E are σ i , i=1,2,...,r and σ1≥σ2≥...≥σ r ≥0;. It contains all the information of the auditory perception masking spectrum E[k,f], so this information is used to analyze the types of abnormal sounds to achieve the purpose of classification. The speaker response signal feature vector s=(σ1,σ2,...,σ r ). Fig.10 is the singular value of the psychoacoustic energy of the loudspeaker signal under different states, given by Fig.10It can be seen that for speakers in different states, the arrangement of singular values ​​is in descending order, reflecting that the arrangement of feature numbers is independent of the state of the speaker, showing the stability advantage of SVD. Therefore, using singular values ​​as signal features can effectively characterize auditory masking and achieve the purpose of feature dimensionality reduction.

[0121] The feature vector s obtained from each sample constitutes X h×r ={X1,X2,...X r} is the feature set of the loudspeaker psychoacoustic energy spectrum, where X i (i=1,2,...,r) are the singular values ​​σ of the auditory perception masking spectrum E of each loudspeaker response signal i (i=1,2,...,r), and then assign the labels corresponding to the speaker response signals of four different states, namely, normal, ring contact, air leakage, and low sound, to 1, 2, 3, and 4 respectively. Assume that Y h×1 ={Y1,Y2,...Y h} is the category set of speaker features, where h is the number of speaker samples and r is the feature dimension.

[0122] Step S3: Use the MRMR algorithm to extract the h×r ={X1,X2,...X r} to select the optimal feature subset Z and eliminate redundant information. The specific steps include:

[0123] Step S31: feature set X h×r Normalization, calculate the feature set X according to the MRMR method h×r Each feature X i The importance of R x , the formula is as follows:

[0124]

[0125] Among them, L x For each feature X i and the correlation between category Y, L x For each feature X i And its category Y mutual information determines:

[0126] L x =I(X i ,Y);

[0127] N x For each feature X i and other features X in set Z l The redundancy between, where i≠l, N x Determined by the mutual information between each feature and other features:

[0128]

[0129] Assume that two continuous random variables are x and y, their probability density functions are p(x) and p(y) respectively, and their joint probability density function is p(x, y); then the mutual information between x and y is:

[0130]

[0131] Step S32: According to the importance R of each feature x Sort in descending order, input the features into the classifier one by one for evaluation, and select the feature subset with the highest classification accuracy. If the input of the mth feature subset makes the classifier have the highest accuracy, it means that the first m features are selected, and the optimal features are selected from them to form feature subset Z through the MRMR criterion.

[0132] Specifically: normalize the feature set and calculate the R of each feature according to the MRMR method. x , and then sort the features in descending order according to their importance, and input the features into the classifier one by one for classification. The changes in the classification accuracy are shown in the following figure. Fig.11 As shown. Fig.11 It can be seen that the accuracy of feature set classification gradually increases with the increase in the number of feature values, but when the accuracy reaches the maximum value, it will basically remain unchanged or even decrease. This shows that when the accuracy reaches the maximum value, the increase in the number of features not only has little effect on improving the classification accuracy, but also affects the classification effect. The results show the importance and necessity of feature selection in feature recognition and classification. Fig.11 It can be seen that when the number of features is 10, the accuracy reaches the highest and the best classification effect is achieved, so the first 10 features are selected as the optimal feature subset.

[0133] In order to illustrate the algorithm more vividly, the algorithm is visualized using the t-distributed Stochastic Neighbor Embedding algorithm. The t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm is a technology that combines dimensionality reduction and visualization. It converts the high-dimensional Euclidean distance between data points into conditional probabilities representing similarities. After this visualization algorithm maps high-dimensional data to low-dimensional data, points separated from each other in the high-dimensional space remain unchanged in the low-dimensional space. After learning converges, t-SNE can project the data set into two-dimensional or three-dimensional space. The feature subset Z selected by the MRMR algorithm is visualized using t-SNE, and the results are shown below. Fig.12 shown.

[0134] In order to better verify the accuracy of the feature extraction method in this paper for classifying abnormal speaker sounds, the LMD energy entropy of the speaker response signal (Chen Xiurong. Research on speaker abnormal sound fault diagnosis based on LMD and LSSVM [D]. Guilin University of Electronic Technology, 2020.), VMD energy entropy (Zhou Jinglei, Yan Ting. Classification of speaker abnormal sound using variational mode decomposition and energy entropy [J]. Journal of Acoustics, 2021, 46(02): 263-270.) and psychoacoustic energy mean (Guo Qing, He Jiekai, Su Haitao, Wang Xingjian. Speaker abnormal sound detection algorithm based on psychoacoustics and support vector machine [J]. Journal of Donghua University (Natural Science Edition), 2020, 46(02): 275-281.) are extracted respectively. Then the data set is normalized and visualized using t-SNE. The results are shown in Figures 13-15. Fig.13 This is the visualization result of feature extraction using the LMD energy entropy method. Fig.14 This is the visualization result of feature extraction using VMD energy entropy method. Fig.15 Visualization of the features extracted for the psychoacoustic energy mean.

[0135] Through Figure 12-Figure 15 From the analysis, it can be obtained that the speaker abnormal sound feature extraction method proposed in the present invention has a high recognition effect on normal, touch circle, leakage, and small sound, especially the feature distributions between normal, touch circle, and leakage are completely separated, and each is gathered in the corresponding area. For the features extracted by the LMD energy entropy method, only the small sound is completely separated from the features of normal, touch circle, and leakage. Although the feature distributions of normal, touch circle, and leakage have a tendency to gradually separate, they are still gathered together; for the features extracted by the VMD energy entropy method, the feature distributions of touch circle and small sound have a tendency to gradually separate from other features, but they are not gathered together, and the degree of dispersion is large, especially for the feature extraction of leakage and other faults. It can be seen from the figure that the leakage feature is mixed with the feature distributions of normal, touch circle, and small sound; for the features extracted by the psychoacoustic energy mean, only the feature distributions of touch circle are completely separated from the features of normal, leakage, and small sound. Although the feature distributions of normal, leakage, and small sound have a tendency to gradually separate, they are still mixed together. It can be concluded that the feature extraction method proposed in the present invention is superior to the current feature extraction method of speaker abnormal sound, especially for the classification of normal speakers.

[0136] The present invention first analyzes the speaker response signal of the collected sound by simulating artificial listening, and calculates the auditory perception masking spectrum of the speaker; then performs SVD decomposition on the auditory perception masking spectrum to obtain a singular value feature sequence; then uses the MRMR algorithm to extract the optimal features in the singular value feature sequence, eliminates redundant information, obtains the features after dimensionality reduction, and eliminates redundant information, so as to avoid too many features in the classification process leading to too high a dimension of the classification model, affecting the efficiency of the model, and being more conducive to unified measurement and judgment.

[0137] Those of ordinary skill in the art will appreciate that the units of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition of each example has been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0138] In the embodiments provided in the present application, it should be understood that the division of units is merely a logical function division, and there may be other division methods in actual implementation, for example, multiple units may be combined into one unit, one unit may be split into multiple units, or some features may be ignored.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

Claims

1. A feature extraction method for abnormal speaker sound based on auditory masking combined with SVD-MRMR, characterized in that: The following steps are involved: Step S1: Use a microphone to collect the sound signal of the speaker. Suppose the sampling frequency of the microphone is f s , in Hz; select a logarithmic sweep signal of 20 to 20000 Hz as the excitation signal, the duration of the sound signal is T, in seconds; calculate the auditory perception masking spectrum E[k,n] of the loudspeaker response signal x(t) according to the ITU-RBS.1387-1 psychoacoustic model, k is the number of subbands, n is the number of frames, n refers to the number of frames in which the duration T of the sound signal is divided according to a frame length of 2048 and an overlap rate of 50%, Step S2: construct the auditory perception masking spectrum E[k,n] of each sub-band of the loudspeaker response signal into a k×n matrix E with a rank of r, r=min(k,n), and then perform SVD decomposition on the matrix E to obtain a singular value eigenvector s. The eigenvector s obtained for each sample constitutes a feature set X of the loudspeaker psychoacoustic energy spectrum. h×r , where h is the number of speaker samples and r is the feature dimension; Step S3: Use the MRMR algorithm to extract the h×n Select the optimal feature subset Z from the , and eliminate redundant information.

2. The feature extraction method of abnormal speaker sound based on auditory masking combined with SVD-MRMR according to claim 1 is characterized in that: The step S1 specifically includes the following steps: Step S11, the response signal x(t) of the loudspeaker is processed by framing, and each frame signal x n [t,n] is: Where t is the number of time domain sampling points per frame; N F is the number of samples per frame; Step S12: Frame the loudspeaker response signal x n [t,n] uses Hanning window filtering and fast Fourier transform to convert the signal from time domain to frequency domain to obtain the spectrum F f [k f ,n]: F f [k f ,n]=fft(h[t]·x n [t,n]); Where: k f is the number of frequency domain points; h[t] is the Hanning window function; Step S13, simulating the auditory characteristics of the human philtrum and outer ear, and performing the obtained frequency spectrum F f [k f ,n] to perform frequency domain weighting to obtain the output F after the middle and outer ear frequency weighting e [k f ,n]: Where W[f] is the frequency response function of the middle and outer ears; Step S14, a Bark scale is evenly divided into 4 frequency bands, the frequency band width is Δz = 1 / 4, and the Bark domain is evenly divided into 109 non-overlapping critical frequency sub-bands in the auditory range of 20Hz to 18kHz, k = 1, 2, ..., 109; the signal energy value in each sub-band is added to obtain the energy output spectrum P in a frequency sub-band e [k,n], where k represents the kth critical frequency subband; The energy distribution of each subband is calculated by the following formula: where f u [k], f l [k] are the low-frequency and high-frequency of the k-th critical frequency subband respectively; The energy output spectrum P in a frequency subband e [k,n] is calculated by the following formula: P e [k,n]=max(∑U[k,k f ](F e [k f ,n]) 2 ,10 -12 ); Step S15, give the energy output spectrum P e [k,n] add a section of internal noise P noise [k], add them together to get P p [k,n]: P p [k,n]=P e [k,n]+P noise [k]; Where fc is the center frequency of the frequency subband; Step S16: Apply the frequency extension function in the Terhardt psychoacoustic model to simulate the frequency domain masking effect; distribute the energy of each subband to the entire Bark domain space through the extension function, then the energy of the i-th subband is expressed as the weighted sum of the contributions of each subband energy to the subband, and the extension function is recorded as S dB (i,k,n) represents the contribution of the k-th subband energy to the i-th subband. The formula is as follows: Calculate the energy distribution E of each subband after frequency domain expansion s [k,n]: Where: B s (i) is the normalization factor; Step S117, perform time domain expansion, and use a first-order low-pass filter to simulate the time domain masking effect: E f [k,n]=a·E f [k,n-1]+(1-a)·E s [k,n]; Among them, the parameter a is determined by the time coefficient and is related to the center frequency of each sub-band; thus, the auditory perception masking spectrum E[k,n] of each sub-band after time domain masking can be obtained: E[k,n]=max(E f [k,n],E s [k,n]).

3. The feature extraction method of abnormal speaker sound based on auditory masking combined with SVD-MRMR according to claim 2 is characterized in that: The relationship between the parameter a and the time coefficient τ is as follows: Where: τ 100 , τ min is a constant, f c [k] is the center frequency of each sub-band.

4. The feature extraction method of abnormal speaker sound based on auditory masking combined with SVD-MRMR according to claim 1, characterized in that: The expression for performing SVD decomposition on the matrix E in step S2 is as follows: Where S is a k*k matrix, V is an n*n matrix, both S and V are unitary matrices, ∑=diag(σ1,σ2,...,σ r ); the singular values ​​of the matrix E are σ i , i=1,2,...,r and σ1≥σ2≥...≥σ r ≥0; then the speaker response signal eigenvector s=(σ1,σ2,...,σ r );The feature vector s obtained from each sample constitutes X h×r ={X1,X2,...X r } is the feature set of the loudspeaker psychoacoustic energy spectrum, where X i is the singular value σ of the auditory perception masking spectrum E of each loudspeaker response signal i Then the labels corresponding to the speaker response signals in four different states, normal, ring contact, air leakage, and low sound, are assigned as 1, 2, 3, and 4 respectively. Assume that Y h×1 ={Y1,Y2,...Y h } is the category set of speaker features, where h is the number of speaker samples and r is the feature dimension.

5. The feature extraction method of abnormal speaker sound based on auditory masking combined with SVD-MRMR according to claim 4 is characterized in that: The step S3 specifically comprises the following steps: Step S31: feature set X h×r Normalization, calculate the feature set X according to the MRMR method h×r Each feature X i The importance of R x , the formula is as follows: Among them, L x For each feature X i and the correlation between category Y, L x For each feature X i and its category Y mutual information, that is, L x =I(X i ,Y); N x For each feature X i With the set X h×r Other features within X l The redundancy between, where i≠l, N x Determined by the mutual information between each feature and other features: Step S32: According to the importance R of each feature x Sort in descending order, input the features into the classifier one by one for evaluation, and select the feature subset with the highest classification accuracy. If the input of the mth feature subset makes the classifier have the highest accuracy, it means that the first m features are selected, and the optimal features are selected from them to form feature subset Z through the MRMR criterion.