A method for classifying abnormal loudspeaker sounds based on auditory perception combined with principal component analysis

By combining the time-varying loudness frequency band energy information entropy weighting based on auditory perception and two-dimensional principal component analysis with the SVM model, the problems of instability and low computational efficiency in loudspeaker abnormal sound classification are solved, and efficient and stable loudspeaker abnormal sound classification is achieved.

CN115022790BActive Publication Date: 2024-11-01GUILIN UNIV OF ELECTRONIC TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202210561514.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-23
Publication Date
2024-11-01
Estimated Expiration
2042-05-23

AI Technical Summary

Technical Problem

Existing methods for classifying abnormal loudspeaker sounds rely on human listening, which leads to unstable results and errors. Furthermore, existing signal processing methods fail to effectively simulate human hearing, resulting in low computational efficiency and high redundancy.

Method used

A method based on auditory perception combined with principal component analysis is adopted. By calculating the time-varying characteristic loudness frequency band energy information entropy weight coefficient, the time-varying loudness under information entropy weight is used to simulate human hearing. Combined with two-dimensional principal component analysis and support vector machine (SVM) for loudspeaker abnormal sound classification, redundant information is eliminated and computational efficiency is improved.

Benefits of technology

It achieves objectivity and stability in classifying abnormal loudspeaker sounds, reduces the workload of auditors, protects hearing health, and improves the consistency of classification results and productivity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115022790B_ABST
    Figure CN115022790B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of speaker quality detection, and specifically to a speaker abnormal sound classification method based on auditory perception combined with principal component analysis. The present invention first calculates the time-varying characteristic loudness of the collected sound according to the German DIN 45631 / A1 standard to obtain data on the characteristic loudness changing over time; then obtains the time-varying characteristic loudness band energy information entropy, and calculates the weight coefficient and the time-varying characteristic loudness graph to obtain the time-varying characteristic loudness spectrum under the weight; then, through a two-dimensional principal component analysis method, extracts the main information of the time-varying characteristic loudness spectrum, eliminates redundant information, and obtains the sample features after dimensionality reduction; and divides it into a training set and a test set, and uses the sample features as the input of a support vector machine, establishes a support vector machine model for training, and realizes the classification of speaker abnormal sounds. The present invention simulates the human ear listening process based on auditory perception, eliminating the influence of subjective feelings on the classification results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of loudspeaker quality detection, and in particular to a loudspeaker abnormal sound classification method based on auditory perception combined with principal component analysis. Background Art

[0002] Speakers and speaker systems are key components of audio equipment. Their quality directly affects the sound quality of the entire audio system. However, during the production process, due to voice coil bottoming, diaphragm rupture, excessive impedance, etc., the speakers produced produce abnormal sounds during the response process. Therefore, it is of great practical significance to classify the abnormal sounds of the speakers produced.

[0003] The traditional method of classifying abnormal loudspeaker sounds is for listeners to detect and classify them on the production line through manual listening. However, the results of manual listening are affected by factors such as the listener's technical maturity, physical and emotional condition, and work fatigue. The subjective feelings of the listening test vary greatly, making it difficult to further improve the accuracy and stability of loudspeaker classification. It is also easy to cause hearing damage, making it difficult to further improve the company's production efficiency and product quality.

[0004] At present, there are also some methods for classifying abnormal speaker sounds, such as the method based on variational mode decomposition and energy entropy (Zhou Jinglei, Yan Ting. Classification of abnormal speaker sounds based on variational mode decomposition and energy entropy [J]. Journal of Acoustics, 2021, 46(02): 263-270.), the method based on harmonic distortion (Hou Menghan. Method for identifying abnormal sound faults of micro speakers based on harmonic distortion [D]. Jilin University, 2018.). However, these methods fail to judge from the perspective of human auditory perception, but only classify from the perspective of traditional signal processing, resulting in errors between their objective test results and artificial listening results.

[0005] At present, many methods for detecting and classifying abnormal sounds of loudspeakers are analyzed through psychoacoustic models. Zhang Ping et al. applied the human ear hearing model to the abnormal sound detection of loudspeakers. After frame-by-frame processing, the signal perception difference curve was calculated to distinguish the abnormal sound samples (Zhang Ping, Feng Haihong, Hu Xiaocheng. Application of PEAQ in abnormal sound detection of loudspeakers [J]. Acoustic Technology, 2015, 34(06): 529-534.). Guo Qing et al. input the loudspeaker sound response signal into the psychoacoustic model and obtained the abnormal sound classification by taking the psychoacoustic energy mean as a feature (Guo Qing, He Jiekai, Su Haitao et al. Loudspeaker abnormal sound detection algorithm based on psychoacoustics and support vector machine [J]. Journal of Donghua University (Natural Science Edition), 2020, 46(02): 275-281.). However, these methods do not consider the problem of signal asynchrony. These feature extraction methods are performed frame by frame. If there is no alignment between frames, feature extraction is bound to produce large errors. In addition, the psychoacoustic model has problems such as high spatial dimension, large redundancy and low computational efficiency. Summary of the invention

[0006] In order to solve the above problems, the present invention provides a speaker abnormal sound classification method based on auditory perception combined with principal component analysis, and proposes a time-varying characteristic loudness band energy information entropy weight coefficient based on the nonlinear effect of the human ear. The time-varying loudness under the information entropy weight is used to simulate the human auditory perception to classify different types of abnormal sounds of the speaker response signal, thereby replacing manual listening, which can not only reduce the work intensity of the listener and protect hearing health, but also ensure the consistency of the objective classification results with the manual listening results, and improve the competitiveness of the production enterprise. The specific technical scheme is as follows:

[0007] The speaker abnormal sound classification method based on auditory perception combined with principal component analysis includes the following steps:

[0008] S1: Use a microphone to collect the sound signal of the speaker. Suppose the sampling frequency of the microphone is f s , in Hz; the duration of the sound signal is T, in seconds; according to the German DIN45631 / A1 standard, the time-varying characteristic loudness spectrum v(n,z i ), where n refers to the number of points at which the duration of the sound signal T is divided into 2ms increments; i means that the 24 critical frequency bands Bark are divided into 240 sub-bands according to 0.1 Bark, i is the corresponding frequency band number, i = 1, 2, ..., 240; z i Bark's frequency scale with 24 critical bands for psychoacoustics;

[0009] S2: Calculate the information entropy weight coefficient of each frequency band and obtain the time-varying characteristic loudness spectrum s(n,zi );

[0010] S3: Through the two-dimensional principal component analysis method, the main information of the time-varying characteristic loudness spectrum is extracted, and the redundant information is eliminated to obtain the feature matrix B after dimensionality reduction; the labels corresponding to the four different abnormal types of speaker response signals, namely normal, ring contact, air leakage, and low sound, are assigned to 1, 2, 3, and 4 respectively, and the features and labels are combined into a sample matrix, and the samples are divided into a training set and a test set;

[0011] S4: Select training samples corresponding to labels 1 and 2, 1 and 3, 1 and 4, 2 and 3, 2 and 4, and 3 and 4 as a training set, and obtain 6 training sets. For each training set, construct an unknown SVM binary classification model, and obtain the optimal hyperparameters by cross-validating the training samples, thereby obtaining 6 trained SVM binary classification models.

[0012] S5: Input the sample features extracted from the speaker response signal to be identified into the 6 SVM binary classification models in step S3, each model outputs a label, and the category corresponding to the label with the largest number of occurrences is determined as the abnormal type of the speaker response signal.

[0013] Preferably, the step S2 of calculating the information entropy weight coefficient of each frequency band specifically includes the following steps:

[0014] S21: Obtain the time-varying characteristic loudness v(n,z i )Energy E(z) in each frequency band i ):

[0015]

[0016] S22: Calculate the proportion of each frequency band to the total energy P(z i ):

[0017]

[0018] S23: Calculate the information entropy weight coefficient g(z i ):

[0019] g(z i )=-P(z i )logP(z i ).

[0020] Preferably, in step S2, the time-varying characteristic loudness spectrum s(n, z i ) Specifically, it uses the nonlinear effect of human hearing to obtain the time-varying characteristic loudness s(n,z i ), that is, the time-varying characteristic loudness spectrum v(n,z i) multiplied by the information entropy weight coefficient g(z i ), the specific calculation formula is as follows:

[0021] s(n,z i )=v(n,z i )·g(z i ).

[0022] Preferably, the step S3 specifically includes the following steps:

[0023] S31: Centralize the time-varying characteristic loudness spectra of all speaker response signals. There are M image samples in total, and the jth image sample is recorded as matrix A j , the average image of all samples is The covariance matrix G of the obtained image is t :

[0024]

[0025] S32: The covariance matrix G of the time-varying characteristic loudness spectrum image t Input into the eig function in MATLAB to get the image covariance matrix G t The eigenvalue λ k and the corresponding eigenvector s k ;

[0026] S33: The feature vector s k Sort in descending order according to the size of their corresponding eigenvalues;

[0027] S34: Select the main components of the time-varying characteristic loudness spectrum when the cumulative contribution rate is greater than 90%, and take the eigenvectors corresponding to the first d eigenvalues ​​as the characteristic subspace ω, s k is the feature vector, 1≤k≤d, d is the number of features corresponding to the cumulative contribution rate greater than 90%, ω=(s 1 ,s 2 ,...s d );

[0028] S35: According to the characteristic vector s k Construct an image feature matrix for feature extraction. For the time-varying feature loudness spectrum s(n,z i ) The corresponding image matrix A can obtain the projection feature vector Y k , Y k Represents the feature vector s k The corresponding projection feature vector, k is the feature number variable, d is the feature number corresponding to the cumulative contribution rate greater than 90%, Y k =As k (k=1,2,...,d);

[0029] S36: Construct a feature matrix B based on the projected feature vector, where B is a z×d-dimensional matrix, Y 1 ,Y 2 ,...,Y d are projection eigenvectors, B=(Y 1 ,Y 2 ,...,Y d );

[0030] S37: Assign labels corresponding to the speaker response signals in four different states as 1, 2, 3 and 4 respectively, form a sample matrix with features and labels, and divide it into a training set and a test set. The training set label is consistent with the training data set, and the test set label is consistent with the test data set.

[0031] Preferably, step S4 comprises the following steps:

[0032] S41: construct an unknown SVM binary classification model, wherein the SVM binary classification model is:

[0033] g(x)=w T Φ(x)+b;

[0034] Where x is the sample feature, x = B; Φ(x) is the mapping function that maps the sample feature x to a high dimension, and w, b are the parameters to be determined;

[0035] S42: Constructing a target optimization function according to the SVM binary classification model:

[0036]

[0037] c is a hyperparameter, y i is the label of sample i, x i To represent the sample characteristics of sample i, the number of samples is m; ξ i is the relaxation factor, ξ i =max(0,1-y i (w·x i +b), each sample has a corresponding slack variable, which represents the degree to which the sample does not satisfy the constraints;

[0038] S43: For the target optimization function, introduce Lagrange multipliers to solve and obtain the optimal classification hyperplane parameters where α=(α 1 ,α 2 ,...,α i ) is the Lagrange multiplier, x s is a component α in the Lagrange multiplier s The corresponding sample features, y s is the label corresponding to the sample feature;

[0039] S44: The calculated w 0 and b 0 Substituting into the nonlinear SVM classifier model expression:

[0040]

[0041] In the formula, K(x i ,x)=Φ T (x i )Φ(x) is the kernel function, and the Gaussian kernel function K(x i ,x)=exp(-γ||x i -x|| 2 ), γ is a hyperparameter;

[0042] S45: taking sample features and sample labels as sample input matrix, taking 70% to 80% of sample features as training samples for cross-validation, inputting the entire matrix into Classification Learner in MATLAB, selecting Gaussian nonlinear SVM classifier model, setting the values ​​of hyperparameters c and γ, selecting training samples corresponding to labels 1 and 2, 1 and 3, 1 and 4, 2 and 3, 2 and 4, 3 and 4 as a group of training sets, and performing machine learning on each group of training sets;

[0043] S46: Change the values ​​of hyperparameters c and γ, repeat step S45 until the preset number of times is reached, find the hyperparameters c and γ corresponding to the maximum accuracy as the optimal hyperparameters, and set the model with these parameters as the optimal SVM binary classification model;

[0044] S47: Input the remaining samples of each group of training samples into the optimal SVM binary classification model obtained in step S46 for classification, and compare the classification results with the corresponding labels. When the accuracy of the classifier is higher than the preset value, it is determined that the SVM binary classification model training is completed, thereby obtaining 6 trained SVM binary classification models.

[0045] Preferably, the step S5 includes the following steps:

[0046] S51: The sample features extracted from the speaker response signal to be identified are expressed as W;

[0047] S52: W is input into the obtained 6 SVM binary classification models. After the 6 SVM binary classification models have completed the classification, the category corresponding to the label with the largest number of occurrences is determined as the abnormal type of the speaker response signal.

[0048] The beneficial effects of the present invention are as follows: the traditional loudspeaker abnormal sound detection and classification method relies on audiometry workers, and it is difficult to use stable and consistent standards to judge the classification of loudspeaker abnormal sounds. In order to replace manual detection, the present invention analyzes loudspeaker abnormal sounds based on auditory perception simulation of artificial listening, obtains the time-varying characteristic loudness spectrum under the information entropy weight for the stable time-varying loudness data, and extracts the main information of the weighted time-varying characteristic loudness spectrum, eliminates redundant information, and eliminates the subjective influence of artificial audiometry. Compared with the existing methods, the present invention has the following outstanding advantages:

[0049] (1) Traditional classification methods rely on manual listening tests for judgment, while the present invention simulates the human ear listening process based on auditory perception, eliminating the influence of subjective feelings on the classification results;

[0050] (2) In this invention, we propose a weight coefficient under the time-varying characteristic loudness frequency band energy information entropy, which enhances the frequency band characteristic information and facilitates the feature extraction of fault information.

[0051] (3) Two-dimensional principal component analysis is used to process the weighted time-varying characteristic loudness spectrum, eliminating redundant information and retaining key information, thereby improving the computational efficiency for subsequent abnormal sound classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] 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.

[0053] Figure 1 This is the time domain diagram of the response signal of a normal speaker

[0054] Figure 2 This is the time domain diagram of the response signal of the touch circle speaker

[0055] Figure 3 The time domain diagram of the response signal of the leaky speaker

[0056] Figure 4 This is the time domain diagram of the response signal of the small speaker

[0057] Figure 5 It is the time-varying characteristic loudness diagram of the response signal of a normal speaker.

[0058] Figure 6 It is the time-varying characteristic loudness diagram of the response signal of the touch circle speaker.

[0059] Figure 7 It is the time-varying characteristic loudness diagram of the response signal of the leaky speaker.

[0060] Figure 8 It is the time-varying characteristic loudness diagram of the response signal of the small sound speaker.

[0061] Fig. 9 This is a graph showing the change in the proportion of the energy of each frequency band to the total energy of the entire frequency band of the time-varying characteristic loudness of the response signal of a normal loudspeaker.

[0062] Fig.10 This is a graph showing the change in the proportion of the energy of each frequency band to the total energy of the entire frequency band, showing the time-varying characteristic loudness of the response signal of the touch circle speaker.

[0063] Fig.11 This is a graph showing the change in the proportion of the energy of each frequency band of the time-varying characteristic loudness of the response signal of the leaky speaker to the total energy of the entire frequency band.

[0064] Fig.12 This is a graph showing the change in the proportion of the energy of each frequency band to the total energy of the entire frequency band of the time-varying characteristic loudness of the response signal of a small sound speaker.

[0065] Fig.13 Time-varying characteristic loudness frequency band energy information entropy weight map of the response signal of a normal loudspeaker.

[0066] Fig.14 Time-varying characteristic loudness frequency band energy information entropy weight map of the response signal of the touch circle speaker.

[0067] Fig.15 The time-varying characteristic loudness frequency band energy information entropy weight map of the response signal of a leaky speaker.

[0068] Fig.16 The time-varying characteristic loudness frequency band energy information entropy weight map of the response signal of a small sound speaker.

[0069] Fig.17 Time-varying characteristic loudness map of the response signal of a normal loudspeaker under information entropy weighting.

[0070] Fig.18 Time-varying characteristic loudness map of the response signal of the touch circle speaker under information entropy weighting.

[0071] Fig.19 Time-varying characteristic loudness map of the response signal of a leaky loudspeaker weighted by information entropy.

[0072] Fig. 20 The time-varying characteristic loudness graph of the response signal of a small speaker under information entropy weighting.

[0073] Fig.21 The following is a flow chart of the classification implementation of the present invention. DETAILED DESCRIPTION

[0074] 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.

[0075] 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.

[0076] 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.

[0077] 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.

[0078] like Fig.21 As shown, the specific implementation of the present invention provides a speaker abnormal sound classification method based on auditory perception combined with principal component analysis. The present invention first calculates the time-varying characteristic loudness of the collected sound according to the German DIN 45631 / A1 standard to obtain data on the characteristic loudness changing over time; then obtains the time-varying characteristic loudness band energy information entropy, and calculates the weight coefficient and the time-varying characteristic loudness map to obtain the time-varying characteristic loudness spectrum under the weight; then, through the two-dimensional principal component analysis method, extracts the main information of the time-varying characteristic loudness spectrum, eliminates redundant information, and obtains the sample features after dimensionality reduction; and divides it into a training set and a test set, and uses the sample features as the input of a support vector machine, establishes a support vector machine model for training, and realizes the speaker abnormal sound classification. Specifically, it includes the following steps:

[0079] S1: Use a microphone to collect the sound signal of the speaker. Suppose the sampling frequency of the microphone is f s , in Hz; the duration of the sound signal is T, in seconds; according to the German DIN45631 / A1 standard, the time-varying characteristic loudness spectrum v(n,z i ), where n refers to the number of points at which the duration of the sound signal T is divided into 2ms increments; i means that the 24 critical frequency bands Bark are divided into 240 sub-bands according to 0.1 Bark, i is the corresponding frequency band number, i = 1, 2, ..., 240; z i Bark's frequency scale with 24 critical bands for psychoacoustics.

[0080] The test object of the specific embodiment of the present invention 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 16.5 ~ 134dB, frequency range of 6.3Hz ~ 20000Hz, inherent noise of 16.5dB A, lower limit frequency -3dB: 3Hz, sensitivity of 51.4mV / 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.

[0081] 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 1-Figure 4 ,in Figure 1 is the response signal of a normal speaker, Figure 2 is the response signal of the touch circle speaker, Figure 3 is the response signal of the air leakage speaker, Figure 4 It is the response signal of the small speaker.

[0082] right Figure 1-Figure 4 The time-domain signal in the calculation of the time-varying characteristic loudness is based on the German DIN 45631 / A1 standard, and the result of the time-varying characteristic loudness changing with time is as follows: Figure 5-Figure 8 shown. Figure 5 It is the time-varying characteristic loudness diagram of the response signal of a normal speaker. Figure 6 is the time-varying characteristic loudness diagram of the response signal of the touch circle speaker, Figure 7 is the time-varying characteristic loudness diagram of the response signal of the leaky speaker, Figure 8 It is the time-varying characteristic loudness diagram of the response signal of the small sound speaker.

[0083] S2: Calculate the information entropy weight coefficient of each frequency band and obtain the time-varying characteristic loudness spectrum s(n,z i ).

[0084] Calculating the information entropy weight coefficient of each frequency band specifically includes the following steps:

[0085] S21: Obtain the time-varying characteristic loudness v(n,z i )Energy E(z) in each frequency band i ):

[0086]

[0087] S22: Calculate the proportion of each frequency band to the total energy P(z i ):

[0088]

[0089] S23: Calculate the information entropy weight coefficient g(z i ):

[0090] g(z i )=-P(z i )logP(z i ).

[0091] Obtain the time-varying characteristic loudness spectrum s(n,z i ) Specifically, it uses the nonlinear effect of human hearing to obtain the time-varying characteristic loudness s(n,z i ), that is, the time-varying characteristic loudness spectrum v(n,z i ) multiplied by the information entropy weight coefficient g(z i ), the specific calculation formula is as follows:

[0092] s(n,z i )=v(n,z i )g(z i ).

[0093] right Figure 5-Figure 8 The time-varying characteristic loudness v(n,z i ) calculates the proportion of the energy of each frequency band to the total energy of the entire frequency band, and obtains the change results of the energy proportion of each frequency band as follows: Figure 9-12 shown. Fig. 9 This is a graph showing the change in the proportion of the energy of each frequency band to the total energy of the entire frequency band of the time-varying characteristic loudness of the response signal of a normal speaker. Fig.10 This is a graph showing the change in the proportion of the energy of each frequency band to the total energy of the entire frequency band of the time-varying characteristic loudness of the response signal of the touch circle speaker. Fig.11 The figure is a graph showing the change in the proportion of the energy of each frequency band to the total energy of the entire frequency band of the time-varying characteristic loudness of the response signal of the leaky speaker. Fig.12 This is a graph showing the change in the proportion of the energy of each frequency band to the total energy of the entire frequency band of the time-varying characteristic loudness of the response signal of a small sound speaker.

[0094] right Figure 9-12 Calculate the energy information entropy, the result is as follows Figure 13-Figure 16 As shown; Fig.13 is the time-varying characteristic loudness frequency band energy information entropy weight map of the response signal of a normal speaker, Fig.14 is the time-varying characteristic loudness frequency band energy information entropy weight map of the response signal of the touch circle speaker, Fig.15 is the time-varying characteristic loudness frequency band energy information entropy weight map of the response signal of the leaky speaker, Fig.16 It is the time-varying characteristic loudness frequency band energy information entropy weight map of the response signal of the small sound speaker.

[0095] Due to the nonlinear spectrum resolution of the human ear, the frequency band energy value in the time-varying characteristic loudness is fitted by information entropy to obtain the weight coefficient of the frequency band energy, such as Figure 13-Figure 16 In order to better reflect its nonlinear characteristics, the time-varying characteristic loudness is multiplied by the weight coefficient to obtain the time-varying characteristic loudness under the weight, as shown in Figure 14-Figure 20 As shown, Fig.17 is the time-varying characteristic loudness diagram of the response signal of a normal speaker under information entropy weight, Fig.18 is the time-varying characteristic loudness diagram of the response signal of the touch circle speaker under the information entropy weight, Fig.19 is the time-varying characteristic loudness graph of the response signal of the leaky speaker under the information entropy weight, Fig. 20 It is the time-varying characteristic loudness diagram of the response signal of the small sound speaker under the information entropy weight.

[0096] S3: Through the two-dimensional principal component analysis method, extract the main information of the time-varying characteristic loudness spectrum, eliminate redundant information, and obtain the reduced-dimensional feature matrix B; assign the labels corresponding to the four different abnormal types of speaker response signals, namely normal, ring contact, air leakage, and low sound, to 1, 2, 3, and 4 respectively, combine the features and labels into a sample matrix, and divide the samples into a training set and a test set. Specifically, the following steps are included:

[0097] S31: Centralize the time-varying characteristic loudness spectra of all speaker response signals. There are M image samples in total, and the jth image sample is recorded as matrix A j , the average image of all samples is The covariance matrix G of the obtained image is t :

[0098]

[0099] S32: The covariance matrix G of the time-varying characteristic loudness spectrum image t Input into the eig function in MATLAB to get the image covariance matrix G t The eigenvalue λ k and the corresponding eigenvector s k ;

[0100] S33: The feature vector s k Sort in descending order according to the size of their corresponding eigenvalues;

[0101] S34: Select the main components of the time-varying characteristic loudness spectrum when the cumulative contribution rate is greater than 90%, and take the eigenvectors corresponding to the first d eigenvalues ​​as the characteristic subspace ω, s k is the feature vector, 1≤k≤d, d is the number of features corresponding to the cumulative contribution rate greater than 90%, ω=(s 1 ,s 2 ,...s d );

[0102] S35: According to the characteristic vector s k Construct an image feature matrix for feature extraction. For the time-varying feature loudness spectrum s(n,z i ) The corresponding image matrix A can obtain the projection feature vector Y k , Y k Represents the feature vector s k The corresponding projection feature vector, k is the feature number variable, d is the feature number corresponding to the cumulative contribution rate greater than 90%, Y k =As k (k=1,2,...,d);

[0103] S36: Construct a feature matrix B based on the projected feature vector, where B is a z×d-dimensional matrix, Y 1 ,Y 2 ,...,Y d are projection eigenvectors, B=(Y 1 ,Y 2 ,...,Y d );

[0104] S37: Assign labels corresponding to the speaker response signals in four different states as 1, 2, 3 and 4 respectively, form a sample matrix with features and labels, and divide it into a training set and a test set. The training set label is consistent with the training data set, and the test set label is consistent with the test data set.

[0105] S4: Select training samples corresponding to labels 1 and 2, 1 and 3, 1 and 4, 2 and 3, 2 and 4, and 3 and 4 as a training set, and obtain 6 training sets. For each training set, construct an unknown SVM binary classification model, and obtain the optimal hyperparameters by cross-validating the training samples, thereby obtaining 6 trained SVM binary classification models. The following steps are included: S41: Construct an unknown SVM binary classification model, wherein the SVM binary classification model is:

[0106] g(x)=wT Φ(x)+b;

[0107] Where x is the sample feature, x=B; Φ(x) is a mapping function that maps the sample feature x to a high dimension, and w, b are parameters to be determined; S42: construct a target optimization function based on the SVM binary classification model:

[0108]

[0109] c is a hyperparameter, which balances the sample fitting and the predictive ability of the test sample; y i is the label of sample i, x i To represent the sample characteristics of sample i, the number of samples is m; ξ i is the relaxation factor, ξ i =max(0,1-y i (w·x i +b), each sample has a corresponding slack variable, which represents the degree to which the sample does not satisfy the constraints;

[0110] S43: For the target optimization function, introduce Lagrange multipliers to solve and obtain the optimal classification hyperplane parameters where α=(α 1 ,α 2 ,...,α i ) is the Lagrange multiplier, x s is a component α in the Lagrange multiplier s The corresponding sample features, y s is the label corresponding to the sample feature;

[0111] S44: The calculated w 0 and b 0 Substituting into the nonlinear SVM classifier model expression:

[0112]

[0113] In the formula, K(x i ,x)=Φ T (x i )Φ(x) is the kernel function, and the Gaussian kernel function K(x i ,x)=exp(-γ||x i -x|| 2 ), γ is a hyperparameter;

[0114] S45: taking sample features and sample labels as sample input matrix, taking 70% to 80% of sample features as training samples for cross-validation, inputting the entire matrix into Classification Learner in MATLAB, selecting Gaussian nonlinear SVM classifier model, setting the values ​​of hyperparameters c and γ, selecting training samples corresponding to labels 1 and 2, 1 and 3, 1 and 4, 2 and 3, 2 and 4, 3 and 4 as a group of training sets, and performing machine learning on each group of training sets;

[0115] S46: Change the values ​​of hyperparameters c and γ, repeat step S45 until the preset number of times is reached, find the hyperparameters c and γ corresponding to the maximum accuracy as the optimal hyperparameters, and set the model with these parameters as the optimal SVM binary classification model;

[0116] S47: Input the remaining samples of each group of training samples into the optimal SVM binary classification model obtained in step S46 for classification, and compare the classification results with the corresponding labels. When the accuracy of the classifier is higher than the preset value, it is determined that the SVM binary classification model training is completed, thereby obtaining 6 trained SVM binary classification models.

[0117] S5: Input the sample features extracted from the speaker response signal to be identified into the six SVM binary classification models in step S3, each model outputs a label, and the category corresponding to the label with the most occurrences is determined as the abnormal type of the speaker response signal. Specifically, the following steps are included:

[0118] S51: The sample features extracted from the speaker response signal to be identified are expressed as W;

[0119] S52: Input W into the obtained 6 SVM binary classification models. After the 6 SVM binary classification models have completed the classification, the category corresponding to the label with the most occurrences is determined as the abnormal type of the speaker response signal. The classification results are shown in Table 1, where the first row is the actual category and the first column is the predicted category.

[0120] Table 1 Classification results

[0121] normal Touch Circle Leakage Small sound normal 98.67% 0.00% 0.00% 1.33% Touch Circle 0.40% 99.20% 0.40% 0.00% Leakage 0.40% 0.40% 97.20% 2.00% Small sound 1.60% 0.00% 0.00% 98.40%

[0122] 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.

[0123] 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.

[0124] 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 speaker abnormal sound classification method based on auditory perception combined with principal component analysis, characterized by: The following steps are involved: S1: Use a microphone to collect the sound signal of the speaker. Suppose the sampling frequency of the microphone is f s , in Hz; the duration of the sound signal is T, in seconds; according to the German DIN45631 / A1 standard, the time-varying characteristic loudness spectrum v(n,z i ), where n refers to the number of points at which the duration of the sound signal T is divided into 2ms increments; i means that the 24 critical frequency bands Bark are divided into 240 sub-bands according to 0.1 Bark, i is the corresponding frequency band number, i = 1, 2, ..., 240; z i Bark's frequency scale with 24 critical bands for psychoacoustics; S2: Calculate the information entropy weight coefficient of each frequency band and obtain the time-varying characteristic loudness spectrum s(n,z i ); S3: Through the two-dimensional principal component analysis method, the main information of the time-varying characteristic loudness spectrum is extracted, and the redundant information is eliminated to obtain the feature matrix B after dimensionality reduction; the labels corresponding to the four different abnormal types of speaker response signals, namely normal, ring contact, air leakage, and low sound, are assigned to 1, 2, 3, and 4 respectively, and the features and labels are combined into a sample matrix, and the samples are divided into a training set and a test set; S4: Select training samples corresponding to labels 1 and 2, 1 and 3, 1 and 4, 2 and 3, 2 and 4, and 3 and 4 as a training set, and obtain 6 training sets. For each training set, construct an unknown SVM binary classification model, and obtain the optimal hyperparameters by cross-validating the training samples, thereby obtaining 6 trained SVM binary classification models. S5: Input the sample features extracted from the speaker response signal to be identified into the six SVM binary classification models in step S3, each model outputs a label, and the category corresponding to the label with the most occurrences is determined as the abnormal type of the speaker response signal; Calculating the information entropy weight coefficient of each frequency band in step S2 specifically includes the following steps: S21: Obtain the time-varying characteristic loudness spectrum v(n,z i )Energy E(z) in each frequency band i ): S22: Calculate the proportion of each frequency band to the total energy P(z i ): S23: Calculate the information entropy weight coefficient g(z i ): g(z i )=-P(z i )logP(from i ); In step S2, the time-varying characteristic loudness spectrum s(n, z i ) Specifically, it uses the nonlinear effect of human hearing to obtain the time-varying characteristic loudness s(n,z i ), that is, the time-varying characteristic loudness spectrum v(n,z i ) multiplied by the information entropy weight coefficient g(z i ), the specific calculation formula is as follows: s(n,z i )=v(n,z i )g(z i ); The step S3 specifically comprises the following steps: S31: Centralize the time-varying characteristic loudness spectra of all speaker response signals. There are M image samples in total, and the jth image sample is recorded as matrix A j , the average image of all samples is The covariance matrix G of the obtained image is t : S32: The covariance matrix G of the time-varying characteristic loudness spectrum image t Input into the eig function in MATLAB to get the image covariance matrix G t The eigenvalue λ k and the corresponding eigenvector s k ; S33: The feature vector s k Sort in descending order according to the size of their corresponding eigenvalues; S34: Select the main components of the time-varying characteristic loudness spectrum when the cumulative contribution rate is greater than 90%, and take the eigenvectors corresponding to the first d eigenvalues ​​as the characteristic subspace ω, s k is the feature vector, 1≤k≤d, d is the number of features corresponding to the cumulative contribution rate greater than 90%, ω=(s1,s2,...s d ); S35: According to the characteristic vector s k Construct an image feature matrix for feature extraction. For the time-varying feature loudness spectrum s(n,z i ) The corresponding image matrix A can obtain the projection feature vector Y k , Y k Represents the feature vector s k The corresponding projection feature vector, k is the feature number variable, d is the feature number corresponding to the cumulative contribution rate greater than 90%, Y k =As k , k=1,2,...,d; S36: Construct feature matrix B, Y1, Y2, ..., Y according to the projected feature vector d are projection eigenvectors, B=(Y1,Y2,...,Y d ); S37: assigning labels corresponding to the speaker response signals in four different states to 1, 2, 3 and 4 respectively, combining the features and labels into a sample matrix, and dividing the matrix into a training set and a test set, wherein the labels of the training set are consistent with the training data set, and the labels of the test set are consistent with the test data set; The step S4 comprises the following steps: S41: construct an unknown SVM binary classification model, wherein the SVM binary classification model is: g(x)=w T Φ(x)+b; Where x is the sample feature, x = B; Φ(x) is the mapping function that maps the sample feature x to a high dimension, and w, b are the parameters to be determined; S42: Constructing a target optimization function according to the SVM binary classification model: s.t.y q (w T Φ(x q )+b)≥1-ξ q ; q=1,2...m c is a hyperparameter, y q is the label of sample q, x q To represent the sample characteristics of sample q, the number of samples is m; ξ q is the relaxation factor, ξ q =max(0,1-y q (w·x q +b), each sample has a corresponding slack variable, which represents the degree to which the sample does not satisfy the constraints; S43: For the target optimization function, introduce Lagrange multipliers to solve and obtain the optimal classification hyperplane parameters where α1,α2,...,α q is the Lagrange multiplier, x s is a component α in the Lagrange multiplier s The corresponding sample features, y s is the label corresponding to the sample feature; S44: Substitute the calculated w0 and b0 into the nonlinear SVM classifier model expression to obtain: In the formula, K(x q ,x)=Φ T (x q )Φ(x) is the kernel function, and the Gaussian kernel function K(x q ,x)=exp(-γ||x q -x|| 2 ), γ is a hyperparameter; S45: taking sample features and sample labels as sample input matrix, taking 70% to 80% of sample features as training samples for cross-validation, inputting the entire matrix into Classification Learner in MATLAB, selecting Gaussian nonlinear SVM classifier model, setting the values ​​of hyperparameters c and γ, selecting training samples corresponding to labels 1 and 2, 1 and 3, 1 and 4, 2 and 3, 2 and 4, 3 and 4 as a group of training sets, and performing machine learning on each group of training sets; S46: Change the values ​​of hyperparameters c and γ, repeat step S45 until the preset number of times is reached, find the hyperparameters c and γ corresponding to the maximum accuracy as the optimal hyperparameters, and set the model with these parameters as the optimal SVM binary classification model; S47: Input the remaining samples of each group of training samples into the optimal SVM binary classification model obtained in step S46 for classification, and compare the classification results with the corresponding labels. When the accuracy of the classifier is higher than the preset value, it is determined that the SVM binary classification model training is completed, thereby obtaining 6 trained SVM binary classification models.

2. The speaker abnormal sound classification method based on auditory perception combined with principal component analysis according to claim 1 is characterized by: The step S5 includes the following steps: S51: The sample features extracted from the speaker response signal to be identified are expressed as W; S52: W is input into the obtained 6 SVM binary classification models. After the 6 SVM binary classification models have completed the classification, the category corresponding to the label with the largest number of occurrences is determined as the abnormal type of the speaker response signal.

Citation Information

Patent Citations

  • Speaker online pure voice failure diagnosis method

    CN101170843A

  • Efficient spectrum sensing method based on support vector machine

    CN112422213A