A method and system for noise reduction and identification of transformer voiceprint signals in complex environments

Abstract

The present invention belongs to the technical field of transformer monitoring, and discloses a method and system for denoising and identifying transformer voiceprint signals in a complex environment. The method selects a wavelet basis function to decompose the transformer voiceprint signal, and decomposes the transformer voiceprint signal into wavelet packet coefficient matrices at different frequencies and time scales; inputs the obtained wavelet packet coefficient matrix into an anisotropic noise factor algorithm, and calculates the low-rank matrix and noise covariance of the transformer voiceprint signal; reconstructs the transformer voiceprint signal after denoising according to the calculated low-rank matrix and noise covariance of the transformer voiceprint signal; inputs the transformer voiceprint signal after denoising into a voiceprint signal identification network, and identifies the operating state of the transformer. The present invention can effectively remove noise components, effectively improve the signal-to-noise ratio, obtain a pure transformer voiceprint signal, and improve the identification accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer monitoring, and in particular relates to a method and system for denoising and identifying transformer voiceprint signals in a complex environment. Background Art

[0002] The transformer will generate vibration and noise during operation, and the transformer voiceprint signal contains rich information about the equipment's operating status. By analyzing the transformer voiceprint signal, equipment fault diagnosis and operating status monitoring can be performed. However, the transformer voiceprint signal is often affected by environmental noise, electromagnetic interference and structural noise, which reduces the accuracy of the fault characteristic signal. Therefore, denoising the transformer voiceprint signal is a key link to ensure the accuracy of diagnosis. The traditional time-frequency domain denoising method has poor adaptability to complex noise, which is easy to cause the loss of effective signals or remove some fixed frequency band noise. However, under the complex working conditions of the transformer, its noise spectrum is uncertain, and traditional filtering is difficult to effectively distinguish between fault signals and noise bands. Therefore, the existing transformer voiceprint signal denoising method has many deficiencies in adaptability, real-time performance, characteristic signal protection and model dependence. There is an urgent need for a new method to achieve a more adaptive, real-time and effective characteristic signal protection denoising technology, so as to more accurately perform equipment fault diagnosis and status monitoring. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for denoising and identifying transformer voiceprint signals in a complex environment, so as to effectively extract effective information from transformer voiceprint signals under complex working conditions and interference from multiple noise sources, while minimizing dependence on prior knowledge to enhance the versatility and generalizability of the method.

[0004] The technical solution of the present invention to implement the above purpose is as follows: A method for noise reduction and identification of transformer voiceprint signals in a complex environment, the steps are as follows:

[0005] Step S1: Select transformer voiceprint signal data and establish a transformer voiceprint signal database;

[0006] Step S2: Select a wavelet basis function to decompose the transformer voiceprint signal, and decompose the transformer voiceprint signal into wavelet packet coefficient matrices at different frequencies and time scales;

[0007] Step S3: input the obtained wavelet packet coefficient matrix into the factor anisotropic noise (FAAN) algorithm to calculate the low-rank matrix and noise covariance of the transformer voiceprint signal;

[0008] Step S4: reconstructing the transformer voiceprint signal according to the low-rank matrix and noise covariance of the transformer voiceprint signal to obtain a de-noised transformer voiceprint signal;

[0009] Use Adaptive Factor Optimize the wavelet packet coefficient matrix C:

[0010] ;

[0011] = ;

[0012] In the formula, is the noise covariance matrix The square root of The low-rank matrix representing the transformer voiceprint signal is the main component of the transformer voiceprint signal; The optimized wavelet packet coefficient matrix suppresses the noise and enhances the main components of the transformer voiceprint signal;

[0013] Use the optimized wavelet packet coefficient matrix to perform inverse wavelet packet transform to obtain the transformer voiceprint signal after noise reduction:

[0014] ;

[0015] In the formula, Represents the transformer voiceprint signal after noise reduction, Represents the inverse wavelet packet transform, which transforms the wavelet packet coefficient matrix in the frequency domain Convert to time domain signal;

[0016] Step S5: inputting the transformer voiceprint signal after noise reduction into a voiceprint signal recognition network to identify the operating status of the transformer.

[0017] Further preferably, the specific process of step S3 is as follows:

[0018] Step S3-1: Calculate the sample covariance matrix of the wavelet packet coefficient matrix C:

[0019] ;

[0020] In the formula, is the sample covariance matrix, represents the transpose of the wavelet packet coefficient matrix C, n is the number of samples;

[0021] Step S3-2: Assume that the noise covariance matrix Q is a diagonal matrix and normalize the sample covariance matrix:

[0022] ;

[0023] In the formula, represents the normalized sample covariance matrix; is the inverse matrix of the square root of the noise covariance matrix Q;

[0024] Step S3-3: Perform eigendecomposition on the normalized sample covariance matrix to obtain the eigenvector matrix and eigenvalue diagonal matrix:

[0025] ;

[0026] In the formula, represents the eigenvector matrix, represents the transposed matrix of the eigenvector matrix; Represents the eigenvalue diagonal matrix, each eigenvalue of the eigenvalue diagonal matrix represents The size of the expansion in the direction of the corresponding eigenvector;

[0027] Step S3-4: Select a low-rank structure to construct a low-rank matrix :

[0028] ;

[0029] In the formula, is the low-rank matrix of the transformer voiceprint signal; Represents the eigenvector matrix The first r eigenvectors of ; represents the eigenvalue diagonal matrix The first r eigenvalues ​​of ; is the identity matrix.

[0030] Further optimization is performed by using an improved ESC algorithm to optimize the iteration of the low-rank matrix. The improved ESC algorithm is based on the ESC algorithm. In the development stage, a dynamic refraction reverse learning mechanism is introduced to disturb the crowd so that it jumps out of the local optimal solution. The position update formula in the development stage is:

[0031] ,

[0032] ,

[0033] ,

[0034] In the formula, It is the position of the elite pool member. is the scaling factor of the dynamic refraction back-learning mechanism, for The fitness of for Adaptability; represents the current position of the i-th individual in the z-th dimension, represents the upper bound of the z-th dimension, represents the lower bound of the z-th dimension, represents the new position of the i-th individual in the z-th dimension; is a parameter that takes the value 0 or 1 with equal probability, determined by the Bernoulli distribution; is another parameter that takes the value 0 or 1 with equal probability, determined by a Bernoulli distribution; is an adaptive Levy weight; is another adaptive Levy weight; is the position of an individual randomly selected from the population; t represents the current iteration number, Indicates the maximum number of iterations.

[0035] Further preferably, the types of transformer soundprint signals include: heavy overload, DC bias, loose clamps, abnormal sound of cooler, partial discharge and normal state.

[0036] Further preferably, in step S2, a 4-layer db4 basis function is selected to decompose the transformer voiceprint signal to obtain a wavelet packet coefficient matrix ,in, is the wavelet packet coefficient of the g-th node in the j-th layer, j={1,2,…,M}, g={1,2,…,G}, M is the number of decomposition layers, and G is the number of nodes.

[0037] The present invention also provides a transformer voiceprint signal noise reduction and identification system in a complex environment, comprising the following modules:

[0038] A data acquisition module is used to select transformer voiceprint signal data and establish a transformer voiceprint signal database;

[0039] A signal decomposition module is used to select a wavelet basis function to decompose the transformer voiceprint signal, and decompose the transformer voiceprint signal into wavelet packet coefficient matrices at different frequencies and time scales;

[0040] A noise reduction module is used to input the obtained wavelet packet coefficient matrix into the factor anisotropic noise (FAAN) algorithm to calculate the low-rank matrix and noise covariance of the transformer voiceprint signal;

[0041] A signal reconstruction module is used to reconstruct the transformer voiceprint signal according to the low-rank matrix and noise covariance of the transformer voiceprint signal to obtain a de-noised transformer voiceprint signal;

[0042] The identification module is used to input the transformer voiceprint signal after noise reduction into the voiceprint signal identification network to identify the operating status of the transformer.

[0043] The present invention takes into account the aliasing of various noises in the actual substation environment, combines the factor anisotropy of noise (FAAN) with the wavelet packet algorithm, decomposes and reconstructs the wavelet packet coefficients extracted by wavelet packet decomposition, replaces the traditional threshold method in the low-frequency part, and estimates the covariance matrix through the expectation method, thereby effectively removing the noise component. An improved optimization algorithm is used to accelerate the iteration of the covariance matrix to obtain the optimal solution, which can effectively improve the signal-to-noise ratio and obtain a pure transformer voiceprint signal; and the pure transformer voiceprint signal is input into the CNN-LSTM network for classification and identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the present invention.

[0045] Figure 2 Flowchart for improving the ESC algorithm.

[0046] Figure 3 Schematic diagram of the voiceprint signal recognition network. DETAILED DESCRIPTION

[0047] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0048] like Figure 1 As shown in the figure, a method for noise reduction and identification of transformer voiceprint signals in a complex environment is shown in the figure. The steps are as follows:

[0049] Step S1: Select transformer voiceprint signal data and establish a transformer voiceprint signal database. The types of transformer voiceprint signals include: heavy overload, DC bias, loose clamps, abnormal cooler sound, partial discharge and normal state;

[0050] Step S2: Select appropriate wavelet basis functions to decompose the transformer voiceprint signal, and decompose the transformer voiceprint signal into wavelet packet coefficient matrices at different frequency and time scales, so as to effectively separate the scale characteristics of the transformer voiceprint signal. According to the characteristics of the transformer voiceprint signal, select 4 layers of db4 basis functions to decompose the transformer voiceprint signal and obtain the wavelet packet coefficient matrix ,in, is the wavelet packet coefficient of the g-th node in the j-th layer, j={1,2,…,M}, g={1,2,…,G}, M is the number of decomposition layers, G is the number of nodes;

[0051] Step S3: input the obtained wavelet packet coefficient matrix into the factor anisotropic noise (FAAN) algorithm to calculate the low-rank matrix and noise covariance of the transformer voiceprint signal;

[0052] Step S4: reconstructing the transformer voiceprint signal according to the low-rank matrix and noise covariance of the transformer voiceprint signal to obtain a de-noised transformer voiceprint signal;

[0053] Use Adaptive Factor Optimize the wavelet packet coefficient matrix:

[0054] ;

[0055] = ;

[0056] In the formula, is the noise covariance matrix The square root of The low-rank matrix representing the transformer voiceprint signal is the main component of the transformer voiceprint signal; The optimized wavelet packet coefficient matrix suppresses the noise and enhances the main components of the transformer voiceprint signal;

[0057] Use the optimized wavelet packet coefficient matrix to perform inverse wavelet packet transform to obtain the transformer voiceprint signal after noise reduction:

[0058]

[0059] In the formula, Represents the transformer voiceprint signal after noise reduction, Represents the inverse wavelet packet transform, which transforms the wavelet packet coefficient matrix in the frequency domain Convert to time domain signal;

[0060] Step S5: inputting the transformer voiceprint signal after noise reduction into a voiceprint signal recognition network to identify the operating status of the transformer.

[0061] The process of calculating the low-rank matrix and noise covariance of the transformer voiceprint signal by the FAAN algorithm in this embodiment is as follows:

[0062] Step S3-1: Calculate the sample covariance matrix of the wavelet packet coefficient matrix C:

[0063] ;

[0064] In the formula, is the sample covariance matrix, It represents the transpose of the wavelet packet coefficient matrix C, n is the number of samples, and the sample covariance matrix reflects the data structure of the wavelet packet coefficient matrix C, which is used to describe the correlation of the signal in the frequency band.

[0065] Step S3-2: Assume that the noise covariance matrix Q is a diagonal matrix and initialize . are the 1st, 2nd, …, mth elements of the diagonal matrix, respectively, and the sample covariance matrix is ​​normalized:

[0066] ;

[0067] In the formula, Represents the normalized sample covariance matrix for eigendecomposition; It is the inverse matrix of the square root of the noise covariance matrix Q, which is used to normalize the sample covariance matrix to make the noise uniformly distributed.

[0068] Step S3-3: Perform eigendecomposition on the normalized sample covariance matrix to obtain the eigenvector matrix and eigenvalue diagonal matrix:

[0069] ;

[0070] In the formula, represents the eigenvector matrix, represents the transposed matrix of the eigenvector matrix; Represents the eigenvalue diagonal matrix, each eigenvalue of the eigenvalue diagonal matrix represents The size of the expansion in the direction of the corresponding eigenvector.

[0071] Step S3-4: Select a low-rank structure to construct a low-rank matrix , the main components of the transformer voiceprint signal are obtained:

[0072] ;

[0073] In the formula, is a low-rank matrix of the transformer voiceprint signal, used to represent the main components of the transformer voiceprint signal; Represents the eigenvector matrix The first r eigenvectors of ; represents the eigenvalue diagonal matrix The first r eigenvalues ​​are used to represent the main component size of the transformer voiceprint signal; is the unit matrix, which is used to subtract 1 from the eigenvalue to adjust the deviation of the eigenvalue.

[0074] Since the FAAN algorithm solves the noise covariance matrix and the low-rank matrix through multiple iterations, the convergence speed may be affected. Therefore, the improved ESC algorithm is used in this method to optimize the iteration of the low-rank matrix.

[0075] The ESC algorithm (optimization method based on crowd evacuation behavior) simulates the behavior of the crowd during the evacuation process, represents the initial position of the individuals in the crowd as a potential B matrix, evaluates the quality of each solution by evaluating the individual position, and thus obtains the best B matrix, and then obtains the best denoising effect. The principle of the ESC algorithm is as follows Figure 2 shown.

[0076] First, initialize a population, and each individual is described by an expression of a D-dimensional vector:

[0077] ,

[0078] ,

[0079] In the formula, represents the position of the i-th individual, represents the current position of the i-th individual in the z-th dimension, represents the upper bound of the z-th dimension, represents the lower bound of the z-th dimension, is a random number between 0 and 1 used to generate the initial position;

[0080] After initializing the population, use the fitness function The fitness of each individual is evaluated. The population is then sorted in ascending order according to fitness, and the top individuals are stored in the elite pool E:

[0081] ,

[0082] In the formula, Indicates the maximum number of potential solutions found in the elite pool;

[0083] The ESC algorithm models an iterative process to reflect the evolution of crowd behavior during the evacuation process. The algorithm classifies the actions of individuals into three categories: calm, gathered, or panic, based on different behavioral responses during the evacuation process:

[0084] ,

[0085] In the formula, represents the panic degree in the crowd at the tth iteration, t represents the current iteration number, Indicates the maximum number of iterations; is pi;

[0086] when When the population is in the exploration stage, the population is divided into a calm group, a gathering group, and a panic group according to their fitness levels. Specifically, the population is sorted in ascending order of fitness, and the individuals are divided into three groups in proportion. = 0.15, the proportion of individuals in the gathering group β = 0.35, and the proportion of individuals in the panic group γ = 0.5, corresponding to the calm group, the gathering group, and the panic group, respectively, and then iterate among the three groups of individuals until ,The iteration ends and enters the next stage:

[0087] 1) Individuals in the calm group behave rationally and move toward the center move:

[0088] ,

[0089] In the formula, represents the new position of the i-th individual in the z-th dimension; is a parameter that takes the value 0 or 1 with equal probability, determined by the Bernoulli distribution; is an adaptive Levy weight; is a randomly generated position represented by the following formula:

[0090] ,

[0091] In the formula, represents a random reference point in position generation; s is a random perturbation parameter, and its value is (0, 1).

[0092] 2) In the clustered group, individuals update their positions according to the influence of the calm and panic groups’ behaviors. The position update formula of the clustered group is as follows:

[0093] ,

[0094] In the formula, is another parameter that takes the value 0 or 1 with equal probability, determined by a Bernoulli distribution; is another adaptive Levy weight; is the position of an individual randomly drawn from the population, Positions for members of the elite pool;

[0095] 3) In the panic group, affected by the potential exits and random instructions from other individuals, panic-driven individuals explore the solution space more irregularly. The position update formula of the panic group is as follows:

[0096] ,

[0097] In the formula, Positions for members of the elite pool;

[0098] when , it enters the development phase, in which all individuals are considered calm. The focus shifts to fine-tuning positions based on the best solution determined so far. The individuals optimize their positions by approaching members of the elite pool, which represents possible safe exits and the best solutions determined in previous iterations, as well as individuals randomly selected from the population, gradually generating the optimal solution. In order to prevent the obtained optimal value from falling into the local optimal solution scenario, a dynamic refraction reverse learning mechanism is introduced on the basis of the original algorithm to perturb the population and make it jump out of the local optimal solution. The position update formula in the development phase is:

[0099] ,

[0100] ,

[0101] ,

[0102] In the formula, It is the position of the elite pool member. is the scaling factor of the dynamic refraction back-learning mechanism, for The fitness of for The fitness of .

[0103] When the maximum number of iterations is reached, the best individual is output and the algorithm ends.

[0104] In view of the characteristics of transformer voiceprint signals, step 5 of this embodiment uses an improved CNN-LSTM network as a voiceprint signal recognition network, such as Figure 3 As shown in the figure, the voiceprint signal recognition network is composed of 3×3 convolution, the first CBAM module, the first maximum pooling layer, 8×8 convolution, the second maximum pooling layer, the second CBAM module, a flattening layer, two consecutive LSTM networks, a fully connected layer and a softmax function.

[0105] The convolutional layer can capture local patterns in the signal and the texture of the spectrum. Convolution kernels of different sizes (3×3 and 8×8 convolution) can extract features at multiple scales. The features in the signal may vary at different time scales or frequency scales:

[0106] In order to enhance the feature extraction capability of the CNN network, the CBAM module is introduced. After the audio signal is processed by the CNN network, the CBAM module will calculate the attention weight for each channel and spatial position.

[0107] Since the convolution outputs three-dimensional data and the LSTM network requires one-dimensional input, the data is first flattened before the LSTM network. A flattening layer is added to the output of the CNN network, and then the first LSTM network is added to learn the time series features of the data. The second LSTM network further processes the time series features from the first layer and extracts the long-term dependencies of the audio signal. The output of the fully connected layer is converted into a high-dimensional representation, and finally the softmax function is used to output the classification result.

[0108] The improved CNN-LSTM network structure has obvious advantages in transformer voiceprint signal recognition through refined feature extraction and attention mechanism enhancement, and can improve recognition accuracy and robustness.

[0109] Another embodiment provides a transformer voiceprint signal noise reduction and identification system in a complex environment, including the following modules:

[0110] A data acquisition module is used to select transformer voiceprint signal data and establish a transformer voiceprint signal database;

[0111] A signal decomposition module is used to select a wavelet basis function to decompose the transformer voiceprint signal, and decompose the transformer voiceprint signal into wavelet packet coefficient matrices at different frequencies and time scales;

[0112] A noise reduction module is used to input the obtained wavelet packet coefficient matrix into the factor anisotropic noise (FAAN) algorithm to calculate the low-rank matrix and noise covariance of the transformer voiceprint signal;

[0113] A signal reconstruction module is used to reconstruct the transformer voiceprint signal according to the low-rank matrix and noise covariance of the transformer voiceprint signal to obtain a de-noised transformer voiceprint signal;

[0114] The identification module is used to input the transformer voiceprint signal after noise reduction into the voiceprint signal identification network to identify the operating status of the transformer.

[0115] Another embodiment provides an electronic device, including a processor, a memory, and a computer program stored in the memory, wherein the computer program, when executed by the processor, implements the above-mentioned method for reducing noise and identifying transformer voiceprint signals in a complex environment.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for noise reduction and identification of transformer voiceprint signals in a complex environment, characterized in that: Here are the steps: Step S1: Select transformer voiceprint signal data and establish a transformer voiceprint signal database; Step S2: Select a wavelet basis function to decompose the transformer voiceprint signal, and decompose the transformer voiceprint signal into wavelet packet coefficient matrices at different frequencies and time scales; Step S3: Input the obtained wavelet packet coefficient matrix into the anisotropic noise factor algorithm to calculate the low-rank matrix and noise covariance of the transformer voiceprint signal; use the improved ESC algorithm to optimize the iteration of the low-rank matrix. The improved ESC algorithm is based on the ESC algorithm. In the development stage, the dynamic refraction reverse learning mechanism is introduced to disturb the crowd so that it jumps out of the local optimal solution. The position update formula in the development stage is: ， ， ， In the formula, It is the position of the elite pool member. is the scaling factor of the dynamic refraction back-learning mechanism, for The fitness of for Adaptability; represents the current position of the i-th individual in the z-th dimension, represents the upper bound of the z-th dimension, represents the lower bound of the z-th dimension, represents the new position of the i-th individual in the z-th dimension; is a parameter that takes the value 0 or 1 with equal probability, determined by the Bernoulli distribution; is another parameter that takes the value 0 or 1 with equal probability, determined by a Bernoulli distribution; is an adaptive Levy weight; is another adaptive Levy weight; is the position of an individual randomly drawn from the population; t represents the current iteration number, Indicates the maximum number of iterations; Step S4: reconstructing the transformer voiceprint signal according to the low-rank matrix and noise covariance of the transformer voiceprint signal to obtain a de-noised transformer voiceprint signal; Use Adaptive Factor Optimize the wavelet packet coefficient matrix C: ； = ； In the formula, is the noise covariance matrix The square root of The low-rank matrix representing the transformer voiceprint signal is the main component of the transformer voiceprint signal; The optimized wavelet packet coefficient matrix suppresses the noise and enhances the main components of the transformer voiceprint signal; Use the optimized wavelet packet coefficient matrix to perform inverse wavelet packet transform to obtain the transformer voiceprint signal after noise reduction: ； In the formula, Represents the transformer voiceprint signal after noise reduction, Represents the inverse wavelet packet transform, which transforms the wavelet packet coefficient matrix in the frequency domain Convert to time domain signal; Step S5: inputting the transformer voiceprint signal after noise reduction into a voiceprint signal recognition network to identify the operating status of the transformer.

2. The method for transformer voiceprint signal noise reduction and identification in a complex environment according to claim 1 is characterized in that: The specific process of step S3 is as follows: Step S3-1: Calculate the sample covariance matrix of the wavelet packet coefficient matrix C: ； In the formula, is the sample covariance matrix, represents the transpose of the wavelet packet coefficient matrix C, n is the number of samples; Step S3-2: Assume that the noise covariance matrix Q is a diagonal matrix and normalize the sample covariance matrix: ； In the formula, represents the normalized sample covariance matrix; is the inverse matrix of the square root of the noise covariance matrix Q; Step S3-3: Perform eigendecomposition on the normalized sample covariance matrix to obtain the eigenvector matrix and eigenvalue diagonal matrix: ； In the formula, represents the eigenvector matrix, represents the transposed matrix of the eigenvector matrix; Represents the eigenvalue diagonal matrix, each eigenvalue of the eigenvalue diagonal matrix represents The size of the expansion in the direction of the corresponding eigenvector; Step S3-4: Select a low-rank structure to construct a low-rank matrix : ； In the formula, is the low-rank matrix of the transformer voiceprint signal; Represents the eigenvector matrix The first r eigenvectors of ; represents the eigenvalue diagonal matrix The first r eigenvalues ​​of ; is the identity matrix.

3. The method for transformer voiceprint signal noise reduction and identification in a complex environment according to claim 1 is characterized in that: Transformer soundprint signal types include: heavy overload, DC bias, loose clamps, abnormal cooler noise, partial discharge and normal status.

4. The method for transformer voiceprint signal noise reduction and identification in a complex environment according to claim 1 is characterized in that: In step S2, a 4-layer db4 basis function is selected to decompose the transformer voiceprint signal to obtain the wavelet packet coefficient matrix ,in, is the wavelet packet coefficient of the g-th node in the j-th layer, j={1,2,…,M}, g={1,2,…,G}, M is the number of decomposition layers, and G is the number of nodes.

5. A system for implementing the transformer voiceprint signal noise reduction and identification method in a complex environment as claimed in claim 1, characterized in that: Includes the following modules: A data acquisition module is used to select transformer voiceprint signal data and establish a transformer voiceprint signal database; A signal decomposition module is used to select a wavelet basis function to decompose the transformer voiceprint signal, and decompose the transformer voiceprint signal into wavelet packet coefficient matrices at different frequencies and time scales; A noise reduction module is used to input the obtained wavelet packet coefficient matrix into the anisotropic noise factor algorithm to calculate the low-rank matrix and noise covariance of the transformer voiceprint signal; A signal reconstruction module is used to reconstruct the transformer voiceprint signal according to the low-rank matrix and noise covariance of the transformer voiceprint signal to obtain a de-noised transformer voiceprint signal; The identification module is used to input the transformer voiceprint signal after noise reduction into the voiceprint signal identification network to identify the operating status of the transformer.

Images