CN117995202A - Transformer core looseness voiceprint recognition method based on feature fusion and PSO-SVM - Google Patents

Abstract

The invention discloses a transformer core looseness voiceprint recognition method based on feature fusion and PSO-SVM, which comprises the steps of obtaining a new feature value MGCC by fusing an acoustic signal extraction Mel cepstrum coefficient and a gamma cepstrum coefficient, extracting a previous ten-dimensional coefficient by a main analysis method as a feature vector for expressing acoustic signal information, inputting the feature vector into an SVM support vector machine, and training to obtain a core looseness recognition model; by carrying out no-load experiments on a certain 10kV transformer and respectively collecting voiceprints under different tightness degrees of the iron core, the invention carries out identification classification on the voiceprints based on the PSO-SVM, and effectively improves the identification rate of the loosening faults of the iron core.

Description

Transformer core loosening soundprint recognition method based on feature fusion and PSO-SVM

Technical Field

The present invention relates to the technical field of transformer fault detection, and in particular to a transformer core loosening voiceprint recognition method based on feature fusion and PSO-SVM.

Background technique

As one of the important links in the power grid, power transformers play an important role and play a huge role in voltage and current conversion and voltage stabilization. During the standard operation of power transformers, their windings and cores may cause mechanical failures in long-term operation. This change in geometry may cause increased vibration of the windings, resulting in mechanical degradation of the solid insulation. Most of the causes of transformer damage are due to failures in the transformer windings, cores, and main insulation. Therefore, monitoring the looseness of the transformer core is of great significance for maintaining the stability of the power grid.

Among the diagnostic methods for transformer core loosening, vibration signal monitoring method and oil testing gas detection method are currently the most mainstream. However, the collection of vibration signals requires vibration sensors, and its location has very strict requirements. A small offset will result in a large error. In addition, a certain number of vibration sensors are required to measure vibration signals for large transformers, which greatly increases the cost and makes it difficult to be widely used in actual environments. The oil testing method cannot achieve real-time monitoring and may have certain limitations for certain faults that require timely response. Acoustic signals can achieve non-contact large-scale monitoring with relatively low cost, and can also achieve online monitoring.

The existing technologies related to the present invention mainly include the following patents related to transformer core loosening soundprint recognition:

Patent "A method and device for identifying loose transformer core components based on CNN+LSTM" Application No./Patent No.: 2021111409170.1, publicly designed a method and device for identifying loose transformer iron components based on CNN+LSTM, mainly based on the CNN+LSTM network model to analyze abnormal noise on the audio signal and give a judgment result. Patent "A method and system for judging loose transformer core using sound detection" Application No./Patent No.: 202111354851.2, publicly designed a method and system for judging loose transformer core using sound detection, by comparing and separating the spectrum similarity between the transformer core sound signals in normal and loose core conditions to judge whether the core is loose. Patent "A method for detecting abnormal transformer soundprint" Application No./Patent No.: 202110872885.4, proposed a denoising model, extracted Mel spectrum features from the denoised sound signal, and used the detection model G-MADE to score the features, and finally judged whether the transformer was correct based on the score.

The above patent considers that voiceprint anomalies have certain reference significance for transformer core loosening faults, but the voiceprint feature values used are mostly relatively simple, and the Mel filter has energy leakage problems during use. At present, there is little original data of transformer acoustic signals, and the accumulation of public data sets is also relatively difficult. In addition, neural network algorithms have local minima and generalization capabilities. Therefore, it is necessary to design a transformer core loosening voiceprint recognition method based on feature fusion and PSO-SVM to solve the above problems.

Summary of the invention

The technical problem to be solved by the present invention is to provide a transformer core loose soundprint recognition method based on feature fusion and PSO-SVM, which obtains new feature parameters by fusing Mel spectrum coefficients and gammatone spectrum coefficients, and uses particle swarm algorithm to optimize the fault diagnosis model parameters of SVM that can solve the classification problem under small samples, thereby improving the fault recognition accuracy.

In order to achieve the above technical effects, the technical solution adopted by the present invention is:

The transformer core loosening soundprint recognition method based on feature fusion and PSO-SVM includes:

S1, no-load test measured the soundprint signal of the transformer core under different loose conditions;

S2, extract the characteristic values of the transformer voiceprint signal based on the Mel cepstral coefficient and the gammatone cepstral coefficient; obtain the fusion feature parameters by vector concatenation, and use the principal component analysis method to reduce the dimension of the fusion feature parameters;

S3, based on the fusion eigenvalues, uses the particle swarm algorithm to optimize the support vector machine parameters, and obtains the optimal kernel function and penalty factor through the optimization algorithm;

S4, a part of the audio obtained under different working conditions of the transformer is used as a training set for training, and the other part is used as a test set for testing to obtain a fault diagnosis model;

S5, fault diagnosis is performed based on the transformer core loose fault diagnosis model.

Preferably, in step S2, extracting the transformer voiceprint signal feature values based on the Mel cepstral coefficients and the gammatone cepstral coefficients, and obtaining the fusion feature parameters by vector concatenation includes:

The transformer soundprint measured in step S1 is preprocessed including pre-emphasis, framing and windowing. The window function uses a Hamming window. The time domain framed signal is transformed by fast Fourier transform to obtain a linear spectrum, which is then passed through a Mel filter group and a gammatone filter group respectively to calculate its logarithmic energy. Finally, the required characteristic parameters are obtained by discrete cosine transform.

Furthermore, the Mel-frequency cepstral coefficients and gamma-tone cepstral coefficients are used to fuse the transformer voiceprint signals, and the principal component analysis method is used to select the top k-dimensional eigenvalues with a contribution rate of 100% to form new fused eigenvalues. Then, the particle swarm algorithm is used to optimize the support vector machine parameters to construct a fault diagnosis model, which improves the accuracy and robustness of transformer core loosening faults.

Preferably, in step S2, the fusion feature parameters are obtained by vector concatenation, and the dimensionality reduction of the fusion feature parameters by using the principal component analysis method includes:

S201, standardize the data and zero-mean each row of the matrix so that the mean of each row is 0;

S202, find the covariance matrix to measure the correlation and variation between attributes;

S203, the eigenvectors are sorted into a matrix according to the size of the eigenvalues from large to small, and then the first k rows are taken as a new matrix, where k is the target dimension after dimensionality reduction;

S204, selecting k features that best reflect the information carried by the acoustic signal according to the cumulative contribution rate; the cumulative contribution rate formula is: Where λ _i represents the corresponding eigenvalue in the matrix, and n represents the number of eigenvalues.

Preferably, in step S3, based on the feature fusion value, obtaining the optimal kernel function and penalty factor through an optimization algorithm includes:

S301, population parameter initialization, setting population size and number of iterations;

S302, calculating the fitness of particles in each iteration, and selecting the particle with the highest fitness as the global optimal solution;

S303, updating the position and velocity of each individual particle according to whether the global optimal solution and the positions and velocities of other individual particles meet the accuracy requirements and the number of iterations requirements;

S304, repeatedly executing the determination of the optimal solution based on the position and velocity of each individual particle, and updating the position and velocity of each example based on the new optimal solution until the conditions are met, thereby obtaining the optimal parameters;

S305, constructing a transformer core loose fault diagnosis model based on the fusion feature value and the optimal parameter support vector machine and training it with the training set.

Furthermore, the present invention builds a transformer core loosening voiceprint recognition system based on feature fusion and PSO-SVM, which can effectively realize the recognition and monitoring of transformer core loosening faults. The system includes a data processing module, a model training module and an identification test module:

The data processing module is used to process the collected audio to obtain the required characteristic values;

The model training module is used to build the PSO-SVM model and train the PSO-SVM model according to the data of the fault samples;

The identification test module is used to perform fault diagnosis based on the above transformer core loose fault diagnosis model.

The beneficial effects of the transformer core loosening voiceprint recognition method based on feature fusion and PSO-SVM provided by the present invention are as follows:

The solution provided by the present invention has high accuracy in fault monitoring and identification. The voiceprint signals of the transformer core under different loose conditions are measured through no-load experiments. The characteristic values of the transformer voiceprint signals are extracted based on the Mel cepstral coefficients and the gammatone cepstral coefficients. The fused feature parameters are obtained by vector splicing. The principal component analysis method is used to reduce the dimension of the fused feature parameters. Finally, the support vector machine parameters are optimized by the particle swarm optimization algorithm.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flowchart of transformer core loosening identification based on PSO+SVM provided by the present invention;

FIG2 is a time domain diagram of a transformer normal operation soundprint in a second embodiment of the present invention;

FIG3 is a time domain diagram of the soundprint of a transformer core loosening operation in the second embodiment of the present invention;

FIG4 is a flow chart based on feature fusion of the present invention;

FIG5 is a flow chart of MFCC extraction of the present invention;

FIG6 is a diagram showing the cumulative contribution rate of each dimension of the fused eigenvalues of the present invention;

7 is a schematic diagram of the particle swarm algorithm for optimizing support vector machine parameters according to the present invention;

FIG8 is a three-dimensional diagram of the acoustic print characteristic value of the transformer core after fusion of the present invention under normal conditions;

9 is a three-dimensional diagram of the transformer core acoustic print characteristic value preload force 1.2Fn after fusion of the present invention;

10 is a three-dimensional diagram of the transformer core acoustic pattern characteristic value preload force 0.5Fn after fusion of the present invention;

FIG11 is a diagram showing the model recognition results of the present invention;

FIG. 12 is a schematic diagram of a system framework provided by the system of the present invention.

Detailed ways

Embodiment 1:

As shown in Figure 1, the transformer core loosening voiceprint recognition method based on feature fusion and PSO-SVM includes:

S1, no-load test measured the soundprint signal of the transformer core under different loose conditions;

S2, extract the characteristic values of the transformer voiceprint signal based on the Mel cepstral coefficient and the gammatone cepstral coefficient; obtain the fusion feature parameters by vector concatenation, and use the principal component analysis method to reduce the dimension of the fusion feature parameters;

S3, based on the fusion eigenvalues, uses the particle swarm algorithm to optimize the support vector machine parameters, and obtains the optimal kernel function and penalty factor through the optimization algorithm;

S4, a part of the audio obtained under different working conditions of the transformer is used as a training set for training, and the other part is used as a test set for testing to obtain a fault diagnosis model;

S5, fault diagnosis is performed based on the transformer core loose fault diagnosis model.

Preferably, in step S2, extracting the transformer voiceprint signal feature values based on the Mel cepstral coefficients and the gammatone cepstral coefficients, and obtaining the fusion feature parameters by vector concatenation includes:

The transformer soundprint measured in step S1 is preprocessed including pre-emphasis, framing and windowing. The window function uses a Hamming window. The time domain framed signal is transformed by fast Fourier transform to obtain a linear spectrum, which is then passed through a Mel filter group and a gammatone filter group respectively to calculate its logarithmic energy. Finally, the required characteristic parameters are obtained by discrete cosine transform.

As shown in Figure 4, the Mel cepstral coefficients and gammatone cepstral coefficients are used to fuse the transformer voiceprint signals, and the principal component analysis method is used to select the first k-dimensional eigenvalues with a contribution rate of 100% to form new fused eigenvalues. Subsequently, the particle swarm algorithm is used to optimize the support vector machine parameters to construct a fault diagnosis model, which improves the accuracy and robustness of transformer core loosening faults.

Preferably, in step S2, the fusion feature parameters are obtained by vector concatenation, and the dimensionality reduction of the fusion feature parameters by using the principal component analysis method includes:

S201, standardize the data and zero-mean each row of the matrix so that the mean of each row is 0;

S202, find the covariance matrix to measure the correlation and variation between attributes;

S203, the eigenvectors are sorted into a matrix according to the size of the eigenvalues from large to small, and then the first k rows are taken as a new matrix, where k is the target dimension after dimensionality reduction;

S204, selecting k features that best reflect the information carried by the acoustic signal according to the cumulative contribution rate; the cumulative contribution rate formula is: Where λ _i represents the corresponding eigenvalue in the matrix, and n represents the number of eigenvalues.

Preferably, in step S3, based on the feature fusion value, obtaining the optimal kernel function and penalty factor through an optimization algorithm includes:

S301, population parameter initialization, setting population size and number of iterations;

S302, calculating the fitness of particles in each iteration, and selecting the particle with the highest fitness as the global optimal solution;

S303, updating the position and velocity of each individual particle according to whether the global optimal solution and the positions and velocities of other individual particles meet the accuracy requirements and the number of iterations requirements;

S304, repeatedly executing the determination of the optimal solution based on the position and velocity of each individual particle, and updating the position and velocity of each example based on the new optimal solution until the conditions are met, thereby obtaining the optimal parameters;

S305, constructing a transformer core loose fault diagnosis model based on the fusion feature value and the optimal parameter support vector machine and training it with the training set.

Furthermore, the present invention builds a transformer core loosening voiceprint recognition system based on feature fusion and PSO-SVM, which can effectively realize the recognition and monitoring of transformer core loosening faults. The system includes a data processing module, a model training module and an identification test module:

The data processing module is used to process the collected audio to obtain the required characteristic values;

The model training module is used to build the PSO-SVM model and train the PSO-SVM model according to the data of the fault samples;

The identification test module is used to perform fault diagnosis based on the above transformer core loose fault diagnosis model.

Embodiment 2:

In a specific implementation, the specific process of transformer core loosening voiceprint recognition using the transformer core loosening voiceprint recognition method based on feature fusion and PSO-SVM provided by the present invention is as follows:

As shown in FIG1 , this embodiment detects transformer core loosening faults through voiceprint signals, including the following steps:

1. Obtain acoustic signals from transformers operating under three working conditions;

2. Pre-emphasis processing of the acoustic signal;

3. The signal is framed;

4. Add Hamming window to the signal;

5. Extract Mel-frequency cepstral coefficients MFCC and gamma-frequency cepstral coefficients GFCC from the acoustic signal and perform vector concatenation;

6. Reduce the dimension of the fused eigenvalues through PCA;

7. Optimize the support vector machine parameters based on the particle swarm optimization algorithm;

8. Obtain the final recognition model through the training set.

The transformer sound is mainly generated by the vibration of the transformer body and the vibration of the components installed on the transformer body. The vibration noise of the transformer body generally comes from the magnetostriction of the iron core and the leakage magnetic force between the iron core laminations. With the improvement of modern iron core lamination technology, the vibration caused by the leakage magnetic force can be ignored. As the transformer runs for an increasing amount of time, the tightness of the iron core will also change. In this embodiment, MFCC and GFCC based on the human cochlear auditory model are selected as characteristic values, through which more soundprint information can be obtained to diagnose the looseness of the transformer core.

As shown in Figures 2 and 3, this is a time domain diagram of the transformer operation soundprint collected, and the spectrum diagram is analyzed.

As shown in FIG5 , in this example, the GFCC feature parameters extracted in step 5 are specifically:

Step 1: Perform fast Fourier transform on the pre-processed acoustic signal;

Step 2: Pass through the Gammatone filter;

Step 3: Perform logarithmic compression to obtain signal energy;

Step 4: Finally obtain the GFCC characteristic parameters by performing discrete cosine transformation on the signal;

Similarly, replace the Gammatone filter with the Mel filter and follow the same steps to obtain the MFCC feature parameters.

Furthermore, in step 6, linear vector splicing is performed by selecting two characteristic parameters.

As shown in FIG6 , the cumulative contribution rate of the first 10 dimensions has reached 100% and is selected as the principal component, so as to obtain fusion feature parameters that retain the information carried by the acoustic signal to a great extent and have low redundancy.

Furthermore, the optimization of the support vector machine parameters based on the particle swarm optimization algorithm is specifically as follows:

1) Initialize population parameters, set population size and number of iterations;

2) Calculate the fitness of the particles in each iteration and select the particle with the highest fitness as the global optimal solution;

3) Update the position and velocity of each individual particle based on whether the global optimal solution and the positions and velocities of other individual particles meet the accuracy requirements and the number of iterations required;

4) Repeat the process of updating the position and velocity of each particle to determine the optimal solution, and update the position and velocity of each example based on the new optimal solution until the conditions are met to obtain the optimal parameters.

In this example, fused eigenvalues are used to extract transformer voiceprint eigenvalues, which has the characteristics of high clean audio recognition and better robustness. The fused eigenvalues are reduced in dimension according to PCA, which solves the complexity caused by the combination of two feature parameters, effectively reduces the dimension and redundancy of the data, and retains the main features and information of the data. The three-dimensional graphs are shown in Figures 8 to 10.

Finally, this example uses the particle swarm algorithm to optimize two parameters of the support vector machine. The selection of the penalty factor C and the kernel function g has a great influence on the SVM classification accuracy. By optimizing these two parameters, the accuracy of the diagnosis system can be improved.

Furthermore, the specific process of feature parameter selection is as follows:

The specific process of MFCC extraction is as follows:

1) Perform fast Fourier transform on the preprocessed voiceprint signal, transform the preprocessed time domain frame signal by fast Fourier transform (FFT) to obtain a linear spectrum, then pass it through an m-dimensional Mel filter, and calculate its logarithmic energy through formula (1).

s(m)＝ln( _Xn (k) ^2Hm ( _k )); (1)

Where m represents the dimension of the Mel filter, and in this paper, m=13. Xn(k) represents the linear spectrum after FFT transformation;

H _m (k) is the filter parameter:

Where: f(m) is the center frequency of the filter; finally, s(m) is transformed by discrete cosine equation (3) to obtain the required MFCC feature parameters:

2) Further, the voiceprint signal is preprocessed including pre-emphasis, framing and windowing, specifically:

Pre-emphasis: Pre-emphasis is used to increase the high-frequency components of the voiceprint signal to prevent the loss of high-frequency information due to attenuation during the voiceprint propagation process. The expression is:

H(z)＝1-μz ^-1 ； (4)

Where: μ is the pre-emphasis coefficient, which is usually between 0.9375 and 0.97. This paper takes the value of 0.95.

Framing: Voiceprint signals have short-term stability. In order to ensure the continuity and autocorrelation between frames, framing is required. At the same time, if the frame length is too long, the signal changes will obviously affect the accuracy of the feature vector. If the frame length is too short, some information will be lost. In this paper, N= per frame is 25ms and the overlap rate is 50%.

Windowing: Adding a window function to each frame can avoid excessive distortion. This paper selects the Hamming window with less frequency leakage as the added window function, and its formula is:

Where N is the length of the Hamming window.

3) Further, PCA is used to reduce the dimension of the fused feature values. The specific method is as follows:

Zero-mean is performed on each row of the matrix, that is, the mean value of the row in which it is located is subtracted from each element, so that the mean of each row is 0. Eliminate the offset of the data so that the data is distributed near the origin of the coordinate system;

Find the covariance matrix to measure the correlation and variation between attributes;

The eigenvectors are sorted into a matrix according to the size of the eigenvalues from large to small, and then the first k rows are taken as a new matrix, where k is the target dimension after dimensionality reduction;

In order to ensure that the information contained in the voiceprint is fully extracted while reducing the computational complexity, it is necessary to calculate the cumulative contribution rate of each dimensional feature vector and select k features that best reflect the information carried by the acoustic signal based on the cumulative contribution rate.

The cumulative contribution rate formula is:

Where λ _i represents the corresponding eigenvalue in the matrix, and n represents the number of eigenvalues.

As shown in Figure 6, the contribution rate of the eigenvalues of each dimension is shown. According to the image, the cumulative contribution rate of MGCC has reached 100% in the 10th dimension, indicating that the first 10 dimensions already contain all the information of the voiceprint, so the eigenvalues of the first 10 dimensions are taken as the feature vectors that can characterize the signal.

Furthermore, the specific process of optimizing the support vector machine parameters by the particle swarm optimization algorithm is as follows:

As shown in Figure 7, the particle swarm optimization algorithm (PSO) is an intelligent algorithm with high solution accuracy that searches for the optimal solution in the entire environment by updating the particle position and speed strategy. It simulates the foraging behavior of birds in nature and searches for the area around the bird closest to the food; compares the birds to particles, determines their position by speed, and obtains the optimal solution through random particle swarm iteration. Specifically:

1. Initialize all particles, that is, assign values to their speed and position, and set the individual's historical best pBest as the current position, and the best individual in the group as the current gBest;

2. In each generation of evolution, calculate the fitness function value of each particle;

3. If the current fitness function value is better than the historical optimal value, update pBest;

4. If the current fitness function value is better than the global historical optimal value, update gBest;

5. The speed update formula and position update formula of particle swarm algorithm are:

in represents the velocity vector of particle i in the kth iteration, / > represents the d-dimensional position vector of particle i in the k-th iteration; ω represents the inertia weight, which is generally taken as 0.9.

r ₁ , r ₂ are random numbers in the interval [0,1].

c ₁ is the individual learning factor, c ₂ is the group learning factor, and usually c ₁ = c ₂ = 2.

Furthermore, the specific process of the transformer core loosening fault diagnosis model based on particle swarm optimization support vector machine is as follows:

The flow chart of this example is shown in Figure 1 and Figure 2. The MFCC and GFCC feature values are extracted from the collected voiceprint signal, and the final fusion feature value is obtained by selecting the cumulative contribution rate. Using the cross-validation method, the data set is divided into a training set and a validation set, and then the model performance is trained and evaluated under different parameter combinations. The trained model is tested with the validation set to obtain the classification accuracy as the performance evaluation indicator of the classifier.

As shown in Figure 11, the recognition accuracy of the fault model, in this specific example, the average accuracy reaches 95.238%. Under normal circumstances, the model recognition accuracy reaches 96.15%, when the preload force is set to 0.5F _N , the recognition accuracy is 95.12%, and when the preload force is set to 1.2F _N , the recognition accuracy is 94.8%.

As shown in FIG12 , it is a schematic diagram of the system framework provided by the present system, which includes: a data processing module, a model training module, and a recognition test module; wherein:

The data processing module is used to process the collected audio to obtain the required feature values;

The model training module is used to build the PSO-SVM model and train the PSO-SVM model according to the data of the fault samples;

The identification test module is used to perform fault diagnosis based on the transformer core loose fault diagnosis model.

The above-mentioned core loosening soundprint recognition system is perfect and has high accuracy for fault monitoring and recognition. The soundprint signals of transformer cores under different loose conditions are measured through no-load experiments. The characteristic values of transformer soundprint signals are extracted based on Mel cepstral coefficients and gammatone cepstral coefficients. The fused feature parameters are obtained by vector splicing. The principal component analysis method is used to reduce the dimension of the fused feature parameters. Finally, the support vector machine parameters are optimized by the particle swarm optimization algorithm.

Claims (5)

1. The transformer core looseness voiceprint recognition method based on feature fusion and PSO-SVM is characterized by comprising the following steps:

S1, measuring voiceprint signals of a transformer iron core under different loosening conditions by an idle load experiment;

s2, extracting characteristic values of voiceprint signals of the transformers respectively based on the Mel cepstrum coefficient and the gamma cepstrum coefficient; obtaining fusion characteristic parameters through vector splicing, and reducing the dimension of the fusion characteristic parameters by using a principal component analysis method;

S3, optimizing parameters of a support vector machine by using a particle swarm algorithm based on the fusion characteristic values, and acquiring an optimal kernel function and a penalty factor by using an optimizing algorithm;

s4, training one part of the acquired audio frequency of the transformer under different working conditions as a training set, and testing the other part of the audio frequency as a testing set to obtain a fault diagnosis model;

s5, performing fault diagnosis based on the transformer core looseness fault diagnosis model.

2. The method for identifying the loose voiceprint of the transformer core based on feature fusion and PSO-SVM according to claim 1, wherein in step S2, extracting the characteristic values of the voiceprint signals of the respective transformers based on the mel cepstrum coefficient and the gamma cepstrum coefficient, and obtaining the fusion characteristic parameters through vector splicing comprises:

Preprocessing the voiceprint of the transformer measured in the step S1, wherein the preprocessing comprises pre-emphasis, framing and windowing, a Hamming window is selected as a window function, a linear spectrum is obtained from a time framing signal through fast Fourier transformation, and then the linear spectrum is respectively passed through a Mel filter bank and a gamma pass filter bank, and logarithmic energy of the linear spectrum is obtained; and finally, obtaining the required characteristic parameters through discrete cosine change.

3. The transformer core loosening voiceprint recognition method based on feature fusion and PSO-SVM according to claim 1, wherein in step S2, fusion feature parameters are obtained through vector splicing, and the dimension reduction of the fusion feature parameters by using a principal component analysis method comprises:

S201, carrying out standardization on data to zero-average each row of the matrix so that the average value of each row is 0;

s202, solving a covariance matrix, and measuring the correlation and the change degree between attributes;

S203, the eigenvector is assembled into a matrix according to the size of the eigenvalue from large to small, and then the previous k rows are taken as a new matrix, wherein k is the target dimension after dimension reduction;

s204, selecting k characteristics which can embody the carrying information of the acoustic signals according to the accumulated contribution rate; the cumulative contribution rate formula is: where λ _i represents the corresponding eigenvalues in the matrix and n represents the number of eigenvalues.

4. The method for identifying the loosening voiceprint of the transformer core based on the feature fusion and the PSO-SVM according to claim 1, wherein in the step S3, obtaining the optimal kernel function and the penalty factor through the optimizing algorithm based on the feature fusion value comprises the following steps:

S301, initializing population parameters, and setting population scale and iteration times;

s302, calculating the fitness of particles in each iteration, and selecting the highest particles as a global optimal solution;

S303, updating the position and the speed of each particle individual according to whether the global optimal solution and the positions and the speeds of other particle individuals meet the precision requirement and the iteration frequency requirement;

s304, repeatedly executing the update of the position and the speed of each particle individual to determine an optimal solution, and updating the position and the speed of each example based on the new optimal solution until the condition is met, so as to obtain optimal parameters;

s305, building based on the fusion characteristic value and the optimal parameter support vector machine and training by using a training set to obtain a transformer core looseness fault diagnosis model.

5. The transformer core looseness voiceprint recognition system based on the feature fusion and the PSO-SVM is characterized by comprising a data processing module, a model training module and a recognition test module, wherein the system executes the transformer core looseness voiceprint recognition method based on the feature fusion and the PSO-SVM according to claims 1-4;

the data processing module is used for processing the collected audio to obtain a required characteristic value;

The model training module is used for building the PSO-SVM model and training the PSO-SVM model according to the data of the fault sample;

the identification test module is used for carrying out fault diagnosis based on the transformer core looseness fault diagnosis model.