Pump machine health monitoring method based on voiceprint characteristics

Abstract

The invention belongs to the technical field of power equipment health monitoring, and aims to solve the problems that a current equipment health monitoring method is easily interfered by pump operation and external environment noise, the fault diagnosis accuracy is poor and the like. The invention discloses a pump health monitoring method based on voiceprint features. The method comprises the steps of collecting voiceprint data during operation of a pump, performing separation processing on the collected voiceprint data, removing background voiceprint data, preprocessing the voiceprint data, extracting voiceprint feature coefficients, and performing voiceprint recognition and fault diagnosis according to the extracted voiceprint feature coefficients. The operation efficiency of the equipment can be improved, the maintenance cost is reduced, and the production loss caused by faults is reduced.

Description

Technical Field

[0001] The present application belongs to the technical field of power equipment health monitoring, and in particular, relates to a pump health monitoring method based on voiceprint features. Background Art

[0002] In the modern industrial field, the stable operation of power equipment is essential for the continuity and safety of production. As an important power equipment, the pump plays a key role in energy conversion and power output. However, due to factors such as long-term operation, complex working environment and wear of components, the performance and health of the steam engine may gradually decline or even fail.

[0003] Traditional methods of pump monitoring usually rely on regular manual inspections and measurements of physical parameters such as temperature and pressure. Although these methods can provide information about the status of the pump to a certain extent, they have many limitations. For example, manual inspections are inefficient and prone to omissions, while physical parameter measurements often only reflect the partial status of the equipment.

[0004] With the continuous development of acoustic technology and signal processing technology, monitoring methods based on soundprint features have gradually attracted people's attention. As a direct reflection of the operating status of the equipment, sound contains rich information. The sound generated by the pump during operation is closely related to the structure of the equipment, the movement state of the components, and the flow of the working medium. By collecting, analyzing and processing these soundprint features, real-time, comprehensive and accurate monitoring of the health status of the pump can be achieved. Summary of the invention

[0005] The purpose of this application is to provide a pump health monitoring method based on voiceprint features to solve the problems existing in current equipment health monitoring methods, such as susceptibility to interference from pump operation and external environmental noise and poor fault diagnosis accuracy.

[0006] In order to achieve the above objectives, this application provides the following technical solutions:

[0007] A pump health monitoring method based on voiceprint features, comprising:

[0008] S1. Collect the voiceprint data when the pump is running;

[0009] S2. Separate and process the collected voiceprint data and remove background voiceprint data;

[0010] S3, preprocessing the voiceprint data to extract voiceprint feature coefficients;

[0011] S4. Perform voiceprint recognition and fault diagnosis based on the extracted voiceprint feature coefficients.

[0012] In some embodiments, the voiceprint data during the operation of the pump is collected through a microphone on the pump equipment, and the collected data is transmitted to the smart gateway through an audio line.

[0013] In some embodiments, a three-way microphone is used to monitor the diaphragm pump, compressor and valve of the pump machine respectively, collect the voiceprint data during the operation of the pump machine in real time, and transmit the voiceprint data to the smart gateway device through the audio cable.

[0014] In some embodiments, S2 includes:

[0015] S211, centralizing each observation vector of the collected mixed voiceprint data;

[0016] S212, calculate the covariance matrix C of the centralized data; perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalue λ i and the eigenvector u i ; Use eigenvalues ​​and eigenvectors to transform the data and obtain whitened data:

[0017]

[0018] Where Λ is the eigenvalue diagonal matrix, U is the eigenvector matrix, and z i is the i-th row vector of the matrix Z after whitening;

[0019] All the centralized observation vectors are combined into a data matrix X′, where each row is a centralized observation vector. The formula for calculating the covariance matrix is:

[0020]

[0021] Where N is the number of observation vectors;

[0022] S213, randomly select a unit vector w1 as the initial value of the weight vector;

[0023] S214, updating the weight vector by maximizing non-Gaussianity;

[0024] S215, randomly select a weight vector w2, where the weight vector w2 is orthogonal to w1;

[0025] S216. Use the obtained weight vector to project the whitened data onto the independent component. The projection formula is:

[0026] s1=w1 T Z

[0027] s2=w2 T Z

[0028] Where Z is the data matrix after whitening, and its row vector is z i , s1 and s2 are independent components after separation.

[0029] In some embodiments, in S211, three microphones are used, and the voiceprint data collected by each microphone is an observation vector. For each observation vector x i , calculate its mean Then perform the centralization operation:

[0030]

[0031] In the formula, x i ′ is the observation vector after centering.

[0032] In some embodiments, the step of generating a random unit vector in S213 is as follows:

[0033] Generate a random vector v, each of which is drawn from a standard normal distribution. The dimension of the vector is the same as the dimension of the data. If the dimension of the data is d, then generate a random vector of length d. Each v i From the normal distribution Extract from

[0034] To normalize a random vector v to a unit vector w1, divide v by its Euclidean norm:

[0035] The Euclidean norm

[0036] In some embodiments, S3 includes:

[0037] S311, using a first-order high-pass filter to perform pre-emphasis processing on the fault voiceprint data;

[0038] S312, dividing the pre-emphasized fault voiceprint data into frames according to the preset time length of each frame;

[0039] S313, multiply each frame of data by the Hamming window function:

[0040]

[0041] Where w(n) is the value of the window function at the nth sample point, n is the sample point index in the frame, from 0 to N-1, and N is the length of the frame;

[0042] S314, convert the fault voiceprint data from the time domain signal to the frequency domain signal, perform fast Fourier transform (FFT) on each frame signal y(n), and obtain the frequency domain representation:

[0043]

[0044] Where N is the frame length;

[0045] S315, calculate the power spectrum of each frame spectrum

[0046] S316. Convert the linear frequency f of the spectrum in S315 to Mel frequency:

[0047]

[0048] Then construct a set of triangular filters to get the Mel spectrum:

[0049] S(m)=∑ k P(k)·H m (k)

[0050] In the formula, H m (k) is the frequency response of the m-th Mel filter at the K-th frequency point;

[0051] S317, performing discrete cosine transform on the logarithmic Mel spectrum logS(m) to obtain the required voiceprint feature parameters.

[0052] In some embodiments, S317 includes:

[0053] Perform discrete cosine transform (DCT) on the logarithmic Mel spectrum and retain the first K coefficients:

[0054]

[0055] Where C(n) is the nth voiceprint feature coefficient, M is the number of Mel filters, is the DCT basis function.

[0056] In some embodiments, S4 includes:

[0057] S411, collect and prepare voiceprint data in normal working state and different fault conditions, and extract voiceprint characteristic coefficients as features through S2 and S3;

[0058] S412, performing Z-score normalization processing on the voiceprint feature coefficient;

[0059] S413, selecting a support vector machine as a classifier, and using a polynomial kernel function to process nonlinear relationships;

[0060] S414, training the SVM model using the training set, and adjusting the parameters of the polynomial kernel function and the penalty term parameters by means of grid search and cross validation;

[0061] S415, using the test set to evaluate the classification performance of the trained SVM model, analyzing the accuracy, recall and other evaluation indicators of the model;

[0062] S416, the voiceprint feature data Z-score is normalized and then input into the pre-trained model.

[0063] In some embodiments, in S412, the standardization processing step includes:

[0064] SS1. Calculate the mean μ of each voiceprint feature coefficient j and standard deviation σ j ;

[0065] For the coefficient j in the voiceprint feature coefficient, calculate the mean and standard deviation of all samples on this coefficient using the following formula:

[0066]

[0067] Where N is the total number of samples, x ij is the jth voiceprint feature coefficient of the i-th sample;

[0068] SS2, apply the Z-score standardization formula to convert each voiceprint feature coefficient x ij Normalized to z ij , the formula is as follows:

[0069]

[0070] In the formula, z ij is the jth voiceprint feature coefficient of the i-th sample after standardization.

[0071] Compared with the prior art, the pump health monitoring method based on voiceprint features provided by this application has the following beneficial effects:

[0072] This application has the advantages of being non-invasive, highly real-time, and capable of detecting early faults, and is expected to provide a more efficient and reliable means for health monitoring of pumps, thereby improving the operating efficiency of the equipment, reducing maintenance costs, and reducing production losses caused by faults.

[0073] This application significantly improves the accuracy and reliability of fault diagnosis, further enhances the real-time nature of monitoring, and builds a powerful and efficient technical support system for the health monitoring of pumps, which is conducive to timely detection and proper handling of safety hazards of pumps, and effectively ensures the safety and stability of production work.

[0074] This application can timely discover faults during the operation of the pump, eliminating the serious lag of traditional detection methods, and can effectively improve the working stability of the equipment. In the process of using the technical solution of this application, the installation and use process will not affect the normal operation of the pump, and it has great potential for promotion and use in other power facilities in society, and has good social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solution of the present application, the following is a brief introduction to the drawings required for the technical description.

[0076] Figure 1 A flow chart of the pump health monitoring method based on voiceprint features provided in this application. DETAILED DESCRIPTION

[0077] The following is further explained in detail through specific implementation methods.

[0078] like Figure 1 As shown, the present application provides a pump health monitoring method based on voiceprint features, comprising the following steps:

[0079] S1. Collect the sound print data of the pump during operation through the microphone. The microphone is fixed on the pump by magnetic suction to collect the sound print data during the operation of the pump. The collected data is transmitted to the smart gateway through the audio line.

[0080] S2. Separate and process the collected voiceprint data and remove background voiceprint data. The intelligent gateway device separates the pre-processed voiceprint data, separates different voiceprint components, and removes background voiceprint data.

[0081] S3. Preprocess the voiceprint data and extract the voiceprint feature coefficients. The intelligent gateway device performs preprocessing operations on the target voiceprint data and extracts the voiceprint feature coefficients therein.

[0082] S4. Perform voiceprint recognition and fault diagnosis based on the extracted voiceprint characteristic coefficients. The intelligent gateway device performs voiceprint recognition and fault diagnosis based on the extracted voiceprint characteristic coefficients, determines the operating status of the pump equipment, and identifies possible fault types.

[0083] This application collects the soundprint data of the pump during operation through a microphone; uses the soundprint data processing algorithm provided by this application to separate and extract the soundprint feature parameters, and combines it with a classification model based on a vector machine (SVM) to achieve efficient monitoring of the health status of the pump. This application significantly improves the accuracy and reliability of fault diagnosis, further enhances the real-time nature of monitoring, and builds a powerful and efficient technical support system for the health monitoring of the pump, which is conducive to timely detection and proper handling of safety hazards of the pump, and effectively ensures the safety and stability of production work.

[0084] In one embodiment, the sound print data of the pump during operation is collected by means of a microphone fixed on the pump device. The sound print data is a mixture of the target sound print data and the sound print data of the background noise of the pump during operation.

[0085] In one embodiment, a three-way microphone is used to monitor the diaphragm pump, compressor and valve of the pump respectively, collect the voiceprint data during the operation of the pump in real time, and transmit the voiceprint data to the intelligent gateway device through the audio line.

[0086] This application collects a large amount of voiceprint data of the pump in normal operation and common fault conditions and extracts voiceprint feature coefficients, and uses these data to train a fault diagnosis model. During real-time monitoring, the extracted voiceprint feature coefficients will be passed into the model for analysis. Through model analysis, the current health status of the pump can be accurately determined, providing strong support for real-time monitoring of the health status of the pump.

[0087] Specifically, in step S2, the purpose of separating the voiceprint is to separate the target voiceprint data and the pump operation background noise voiceprint data from the mixed voiceprint data, and the specific steps include the following:

[0088] S211, centralize each observation vector of the collected mixed voiceprint data, that is, subtract their respective means to eliminate the mean offset of the data.

[0089] Each observation vector of the voiceprint data refers to a discrete sample of the sound signal obtained during the operation of the pump. The number of observation vectors is related to the number of microphones. The voiceprint data collected by each microphone is an observation vector. In this embodiment, three microphones are used, so there are three observation vectors.

[0090] For each observation vector x i , calculate its mean Then perform the centralization operation:

[0091]

[0092] In the formula, x i ′ is the observation vector after centering.

[0093] S212, calculate the covariance matrix C of the centralized data; perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalue λ i and the eigenvector u i ; Use eigenvalues ​​and eigenvectors to transform the data and obtain whitened data:

[0094]

[0095] Where Λ is the eigenvalue diagonal matrix, U is the eigenvector matrix, and z i is the i-th row vector of the matrix Z after whitening.

[0096] All the centralized observation vectors are combined into a data matrix X′, where each row is a centralized observation vector. The formula for calculating the covariance matrix is:

[0097]

[0098] Where N is the number of observation vectors.

[0099] The purpose of this step is to transform the voiceprint data into a form with unit variance and no correlation.

[0100] S213: Randomly select a unit vector w1 as the initial value of the weight vector. This weight vector will be used in the subsequent iterative update process.

[0101] Specifically, the steps to generate a random unit vector are as follows:

[0102] Generates a random vector v, each component of which is drawn from a standard normal distribution (mean 0, variance 1). The dimension of the vector is the same as the dimension of the data. If the dimension of the data is d, then generates a random vector of length d Each v i From the normal distribution Extracted from .

[0103] Normalize the random vector v to the unit vector w1. The specific operation is: divide v by its Euclidean norm (that is, the length of the vector): The Euclidean norm

[0104] S214, update the weight vector by maximizing non-Gaussianity to achieve the independence of sound components. Calculate the projection u i =w1 T z i , where z i is the whitened data. Using nonlinear function To update the weight vector: Normalize the updated weight vector In the formula, g′(u i ) is g(u i ), represents mathematical expectation.

[0105] Repeat this step until the change in the Euclidean norm is less than a preset threshold, then it is considered converged, the iteration is stopped, and the following steps are continued; otherwise, the iteration is continued.

[0106] S215. Similarly, in steps S213 and S214, a weight vector w2 is randomly selected. The weight vector w2 should be orthogonal to w1. In step S214, w2 is orthogonalized after each update to ensure that the weight vectors w2 and w1 are orthogonal to each other to improve the separation effect.

[0107] S216, use the final weight vector to project the whitened data onto the independent component. The specific calculation projection formula is s1=w1 T Z,s2=w2 T Z, where Z is the data matrix after whitening, and its row vector is z i , s1 and s2 are independent components after separation, thus achieving the separation of voiceprint components.

[0108] After completing the voiceprint separation process, the separated independent components s1 and s2 are subjected to frequency domain analysis, the power spectrum density is calculated, and the spectrum characteristics of the normal operation voiceprint of the pump are compared to distinguish the target voiceprint data from the operation and background noise voiceprint data.

[0109] Specifically, in step S3, the voiceprint data is preprocessed to extract the voiceprint characteristic coefficients of the target voiceprint data. The intelligent gateway device performs preprocessing operations on the collected voiceprint data, including pre-emphasis, framing and windowing of the data.

[0110] S311, use a first-order high-pass filter to pre-emphasize the fault voiceprint data, and its transfer function is:

[0111] H(z)=1-αz -1

[0112] In the formula, the value of α is usually between 0.9 and 1.

[0113] In this example, in order to better emphasize the high frequency part during the pre-emphasis process while maintaining the stability and practicality of the system, α is set to 0.97.

[0114] S312, according to the preset time length of each frame, the pre-emphasized fault voiceprint data is divided into frames, and a certain overlap is maintained between frames, so as to ensure that the information between adjacent frames has a certain continuity. For example, the pre-emphasized fault voiceprint data is divided into 20 milliseconds as a frame, and the overlap between frames is set to 10 milliseconds, so as to ensure that the information between adjacent frames has a certain continuity.

[0115] S313, multiply each frame of data by the Hamming window function: This reduces spectrum leakage and improves the accuracy and stability of spectrum analysis.

[0116] Where w(n) is the value of the window function at the nth sample point, n is the sample point index in the frame, from 0 to N-1, and N is the length of the frame, that is, the width of the window.

[0117] S314, convert the fault voiceprint data from the time domain signal to the frequency domain signal, perform fast Fourier transform (FFT) on each frame signal y(n), and obtain the frequency domain representation:

[0118]

[0119] Where N is the frame length.

[0120] S315, calculate the power spectrum of each frame spectrum To reflect the energy distribution of the signal.

[0121] S316, converting the linear frequency f of the spectrum in step S315 into Mel frequency: Then a set of triangular filters (Mel filters) are constructed to cover the Mel frequency range and applied to the power spectrum to obtain the Mel spectrum:

[0122] S(m)=∑ k P(k)·H m (k)

[0123] In the formula, H m (k) is the frequency response of the m-th Mel filter at the K-th frequency point.

[0124] S317, performing discrete cosine transform on the logarithmic Mel spectrum logS(m) to obtain the required voiceprint feature parameters.

[0125] Perform discrete cosine transform (DCT) on the logarithmic Mel spectrum and retain the first K coefficients:

[0126]

[0127] Where C(n) is the nth voiceprint feature coefficient, M is the number of Mel filters, is the DCT basis function.

[0128] These K coefficients are combined into a feature vector for subsequent fault monitoring and identification tasks.

[0129] The specific calculation process of the 0th order coefficient is as follows:

[0130]

[0131] The 0th order coefficient is actually the sum of all logarithmic Mel spectra.

[0132] Specifically, in the process of implementing the pump fault soundprint diagnosis in step S4, it is necessary to compare the separated and extracted soundprint feature coefficients with the existing soundprint feature coefficients of the pump in normal working state and under different fault conditions to confirm the operating state of the pump.

[0133] S411, collect and prepare voiceprint data in normal working state and different fault conditions, extract voiceprint feature coefficients as features through steps S2 and S3. Divide the collected sample voiceprint feature coefficients into training set and test set to prepare for the subsequent steps of training and evaluating the model.

[0134] S412. Perform Z-score normalization on the voiceprint feature coefficients to ensure that each feature coefficient has a similar numerical range to improve the stability and convergence speed of model training.

[0135] The specific standardization processing steps are as follows:

[0136] SS1. Calculate the mean μ of each voiceprint feature coefficient j and standard deviation σ j .

[0137] For the coefficient j in the voiceprint feature coefficient, calculate the mean and standard deviation of all samples on this coefficient using the following formula:

[0138]

[0139]

[0140] Where N is the total number of samples, x ij is the jth voiceprint feature coefficient of the ith sample.

[0141] SS2, apply the Z-score standardization formula to convert each voiceprint feature coefficient x ij Normalized to z ij , the formula is as follows:

[0142]

[0143] In the formula, zij is the jth voiceprint feature coefficient of the i-th sample after standardization.

[0144] S413, select support vector machine (SVM) as a classifier, and use polynomial kernel function to handle nonlinear relationship. Specify parameters of polynomial kernel, such as degree of polynomial, kernel function coefficient and penalty term parameter, to optimize the fitting ability and generalization ability of the model.

[0145] S414, training the SVM model using the training set, and adjusting the parameters of the polynomial kernel function and the penalty term parameters by means of grid search and cross-validation methods to optimize the performance and generalization ability of the model;

[0146] S415, use the test set to evaluate the classification performance of the trained SVM model, analyze the model's accuracy, recall and other evaluation indicators to verify the generalization ability and robustness of the model on unseen data.

[0147] S416. During the actual monitoring process, when the voiceprint feature data is collected, the data is Z-score standardized and then input into the pre-trained model. The model can quickly and accurately determine whether the pump has a fault and identify the specific type of fault, so as to make corresponding maintenance or treatment measures in time.

[0148] In the actual monitoring process, the new data collected needs to be Z-score standardized using the mean and standard deviation calculated from the training set.

[0149] In summary, this application proposes a pump health detection method based on voiceprint features, which can promptly detect fault problems during the operation of the pump, eliminate the serious shortcomings of traditional detection methods, and measure the health status of the pump in real time, effectively improving the working stability of the equipment. In the process of using this application, the installation and use process does not affect the normal operation of the pump. It has great potential for promotion and use to other power facilities in the society and has good social benefits. In the future, more sensor modules will be combined with this system to realize a more comprehensive pump operation monitoring function, further improve the efficiency of pump inspections and operation and maintenance, and provide more scientific data basis for relevant managers to carry out targeted management and control, which will help improve the maintenance efficiency, reliability and stability of the equipment.

[0150] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed in the present application should be covered within the protection scope of the present application.

Claims

1. A pump health monitoring method based on voiceprint features, characterized in that: include: S1. Collect the voiceprint data when the pump is running; S2. Separate and process the collected voiceprint data and remove background voiceprint data; S3, preprocessing the voiceprint data to extract voiceprint feature coefficients; S4. Perform voiceprint recognition and fault diagnosis based on the extracted voiceprint feature coefficients.

2. The pump health monitoring method based on voiceprint features according to claim 1 is characterized in that: The voiceprint data during the operation of the pump is collected through the microphone on the pump equipment, and the collected data is transmitted to the smart gateway through the audio cable.

3. The pump health monitoring method based on voiceprint features according to claim 1 is characterized in that: A three-way microphone is used to monitor the diaphragm pump, compressor and valve of the pump machine respectively, collect the voiceprint data during the operation of the pump machine in real time, and transmit the voiceprint data to the intelligent gateway device through the audio cable.

4. The pump health monitoring method based on voiceprint features according to claim 1 is characterized in that S2 include: S211, centralizing each observation vector of the collected mixed voiceprint data; S212, calculate the covariance matrix C of the centralized data; perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalue λ i and the eigenvector u i ; Use eigenvalues ​​and eigenvectors to transform the data and obtain whitened data: Where Λ is the eigenvalue diagonal matrix, U is the eigenvector matrix, and z i is the i-th row vector of the matrix Z after whitening; All the centralized observation vectors are combined into a data matrix X′, where each row is a centralized observation vector. The formula for calculating the covariance matrix is: Where N is the number of observation vectors; S213, randomly select a unit vector w1 as the initial value of the weight vector; S214, updating the weight vector by maximizing non-Gaussianity; S215, randomly select a weight vector w2, where the weight vector w2 is orthogonal to w1; S216. Use the obtained weight vector to project the whitened data onto the independent component. The projection formula is: s1=w1 T WITH <h2 style=";text-align:left;direction:ltr">s2 = w2<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> Z Where Z is the data matrix after whitening, and its row vector is z i , s1 and s2 are independent components after separation.

5. The pump health monitoring method based on voiceprint features according to claim 4 is characterized in that: In S211, three microphones are used, and the voiceprint data collected by each microphone is an observation vector. For each observation vector x i , calculate its mean Then perform the centralization operation: In the formula, x i ′ is the observation vector after centering.

6. The pump health monitoring method based on voiceprint features according to claim 4 is characterized in that: The steps for generating random unit vectors in S213 are as follows: Generate a random vector v, each of which is drawn from a standard normal distribution. The dimension of the vector is the same as the dimension of the data. If the dimension of the data is d, then generate a random vector of length d. Each v i From the normal distribution Extract from To normalize a random vector v to a unit vector w1, divide v by its Euclidean norm: The Euclidean norm 7. The pump health monitoring method based on voiceprint features according to claim 1 is characterized in that S3 include: S311, using a first-order high-pass filter to perform pre-emphasis processing on the fault voiceprint data; S312, dividing the pre-emphasized fault voiceprint data into frames according to the preset time length of each frame; S313, multiply each frame of data by the Hamming window function: Where w(n) is the value of the window function at the nth sample point, n is the sample point index in the frame, from 0 to N-1, and N is the length of the frame; S314, convert the fault voiceprint data from the time domain signal to the frequency domain signal, perform fast Fourier transform (FFT) on each frame signal y(n), and obtain the frequency domain representation: Where N is the frame length; S315, calculate the power spectrum of each frame spectrum S316. Convert the linear frequency f of the spectrum in S315 to Mel frequency: Then construct a set of triangular filters to get the Mel spectrum: S(m)=∑ k P(k)·H m (k) In the formula, H m (k) is the frequency response of the m-th Mel filter at the K-th frequency point; S317, performing discrete cosine transform on the logarithmic Mel spectrum logS(m) to obtain the required voiceprint feature parameters.

8. The pump health monitoring method based on voiceprint features according to claim 7 is characterized in that: S317 includes: Perform discrete cosine transform (DCT) on the logarithmic Mel spectrum and retain the first K coefficients: Where C(n) is the nth voiceprint feature coefficient, M is the number of Mel filters, is the DCT basis function.

9. The pump health monitoring method based on voiceprint features according to claim 1 is characterized in that S4 include: S411, collect and prepare voiceprint data in normal working state and different fault conditions, and extract voiceprint characteristic coefficients as features through S2 and S3; S412, performing Z-score normalization processing on the voiceprint feature coefficient; S413, selecting a support vector machine as a classifier, and using a polynomial kernel function to process nonlinear relationships; S414, training the SVM model using the training set, and adjusting the parameters of the polynomial kernel function and the penalty term parameters by means of grid search and cross validation; S415, using the test set to evaluate the classification performance of the trained SVM model, analyzing the accuracy, recall and other evaluation indicators of the model; S416, the voiceprint feature data Z-score is normalized and then input into the pre-trained model.

10. The pump health monitoring method based on voiceprint features according to claim 9 is characterized in that: In S412, the standardization processing step includes: SS1. Calculate the mean μ of each voiceprint feature coefficient j and standard deviation σ j ; For the coefficient j in the voiceprint feature coefficient, calculate the mean and standard deviation of all samples on this coefficient using the following formula: Where N is the total number of samples, x ij is the jth voiceprint feature coefficient of the i-th sample; SS2, apply the Z-score standardization formula to convert each voiceprint feature coefficient x ij Normalized to z ij , the formula is as follows: In the formula, z ij is the jth voiceprint feature coefficient of the i-th sample after standardization.

Images