Adaptive wavelet optimization and feature extraction method and system for transformer acoustic signals

Abstract

This application discloses an adaptive wavelet optimization and feature extraction method and system for transformer acoustic signals. The method includes: calling the Pywt wavelet analysis library to decompose the original signal according to the decomposition layer number J to obtain multi-layer detail signals; for each layer of detail signals, performing the following: introducing the M-estimator to improve the noise variance calculation model, calculating the standard deviation and a unified monitoring threshold; constructing a dynamic threshold based on the criterion of minimizing the approximation error of the denoised signal; designing a correction factor; correcting the wavelet coefficients of each layer of detail signals; reconstructing the pure signal using an inverse decomposition method; dividing the pure signal into multiple short-time signals; extracting the MFCC feature vector of each short-time signal; using the support vector machine recursive feature elimination method to weightedly screen the MFCC feature vectors; and combining the principal component analysis algorithm to compress the feature vector dimension and generate the final feature matrix. This application solves the problems of the contradiction between noise suppression and signal fidelity in traditional methods, and the low recognition rate caused by high-dimensional feature redundancy.

Description

Technical Field

[0001] The present application relates to the technical field of transformer operating status monitoring, and in particular to a method and system for adaptive wavelet optimization and feature extraction of transformer acoustic signals. Background Art

[0002] As a key component of power transmission and distribution, transformer operation is directly related to the safety, stability, and efficiency of the power grid. Therefore, implementing real-time, accurate transformer operation monitoring and noise control is crucial for ensuring the normal operation of the power system, improving energy efficiency, and reducing environmental pollution.

[0003] With the rapid development of the power industry and the continuous advancement of smart grid construction, the technical requirements for transformer operation monitoring and noise control are constantly increasing. Transformer operating status acoustic monitoring technology collects and analyzes acoustic field signals during transformer operation to extract key characteristic parameters for early warning and diagnosis of transformer faults. However, current acoustic monitoring technology still has many drawbacks, such as low recognition accuracy, susceptibility to environmental noise interference, and unsatisfactory noise reduction, making it difficult to meet the needs of modern power systems for intelligent and refined monitoring. Summary of the Invention

[0004] In view of one of the defects in the prior art, the purpose of this application is to provide a method and system for adaptive wavelet optimization and feature extraction of transformer acoustic signals.

[0005] In a first aspect of the present application, a method for adaptive wavelet optimization and feature extraction of transformer acoustic signals is provided, comprising:

[0006] Collect the original sound signal;

[0007] Call the Pywt wavelet analysis library to decompose the original signal according to the decomposition layer number J to obtain multi-layer detail signals;

[0008] For each layer of detail signal, perform the following steps:

[0009] The M-estimator is introduced to improve the noise variance calculation model, calculate the standard deviation and the unified monitoring threshold;

[0010] Based on the standard deviation and the unified monitoring threshold, a dynamic threshold based on a criterion of minimizing the approximation error of the denoised signal is constructed;

[0011] Design correction factor;

[0012] Based on the correction factor and the dynamic threshold, the wavelet coefficients of the detail signal of each layer are corrected;

[0013] Based on the modified wavelet coefficients, the pure signal is reconstructed using the inverse decomposition method;

[0014] Dividing the clean signal into multiple short-time signals;

[0015] Extracting a Mel-frequency cepstral coefficient feature vector of each of the short-time signals;

[0016] Using a support vector machine recursive feature elimination method to weightedly screen the Mel-frequency cepstral coefficient feature vector;

[0017] Combined with the principal component analysis algorithm, the dimension of the eigenvector after weighted screening is compressed to generate the final feature matrix.

[0018] Optionally, the introduction of an M-estimator to improve the noise variance calculation model, calculate the standard deviation and a unified monitoring threshold, includes:

[0019] The M estimator is introduced into the standard deviation calculation, and the outlier interference is suppressed by the robust loss function. The standard deviation of the detail signal of the j-th layer is calculated as:

[0020] ;

[0021] in is the robust loss function, i.e. Huber function;

[0022] Cd j,i is the detail coefficient after wavelet decomposition, N is the number of detail coefficients in the jth layer, j is the index of the decomposition layer of the original signal, and its value range is 1 to J;

[0023] σ 0 is the initial scale parameter, σ0=1.4826×MAD(Cd j,i ), MAD stands for median absolute deviation;

[0024] μ is the location parameter, μ=median(Cd j,i ), median means median;

[0025] Based on the standard deviation, a unified monitoring threshold of the j-th layer detail signal is obtained:

[0026] .

[0027] Optionally, constructing a dynamic threshold based on a criterion for minimizing the approximation error of the denoised signal includes:

[0028] Arrange the absolute values ​​of the detail coefficients of the jth layer in ascending order as a vector P=[CD j,1 ,CD j,2 ,…,CD j,N ]

[0029] Calculate each element r of the risk vector R i :

[0030] ;

[0031] The smallest element r in R min As the approximation error, record its corresponding CD min , k It is the sequence number used to traverse each element in the detail coefficient vector of the jth layer, where the value of k ranges from 1 to i; where i is the vector P=[CD j,1 ,CD j,2 ,…,CD j,N ], whose value range is from 1 to N, where N is the number of detail coefficients in the j-th layer;

[0032] Calculate the adaptive threshold of the detail coefficient of the jth layer:

[0033] ;

[0034] of It is the abbreviation of adaptive, which is used to identify the threshold as an adaptive threshold;

[0035] Based on the unified monitoring threshold and the adaptive threshold, the dynamic threshold of the j-th layer is obtained:

[0036] ;

[0037] in, is the average value of the detail coefficient of the jth layer:

[0038] ;

[0039] in Stands for average, which is used to represent the average value of the detail coefficient of the jth layer. is the minimum energy level of the j-th layer detail coefficient:

[0040] .

[0041] Optionally, the design correction factor for:

[0042] ;

[0043] Among them, B j is the absolute value of the detail coefficient of the jth layer of the detail signal, represents the median, Represents the absolute value of the kth detail coefficient of the jth layer detail signal, k It is the sequence number used to traverse each element in the detail coefficient vector of the j-th layer. The value range of k here is 1 to N, and N is the number of detail coefficients of the j-th layer.

[0044] Optionally, based on the correction factor and the dynamic threshold, correcting the wavelet coefficients of each layer of detail signal includes:

[0045] ;

[0046] in is the wavelet estimation coefficient, is the wavelet coefficient, k It is the serial number used to traverse each element in the detail coefficient vector of the jth layer. The value range of k here is 1 to N. is the dynamic threshold of the jth layer, is the designed correction factor, and N is the number of detail coefficients in the jth layer.

[0047] Optionally, dividing the clean signal into multiple short-time signals includes:

[0048] The clean signal is divided into M frame signals, wherein:

[0049] ;

[0050] Where M is the number of frames, n is the length of the clean signal, L is the frame length, and d is the overlap ratio;

[0051] Hamming window processing is applied to the M-frame framed signal, and point-by-point window multiplication is performed to eliminate spectrum leakage caused by frame edge truncation and obtain the optimized short-time signal.

[0052] Optionally, extracting a Mel-frequency cepstral coefficient feature vector of each short-time signal includes:

[0053] Performing a fast Fourier transform on a single short-time signal and obtaining a modulus value thereof;

[0054] use p A filter is used to transform the short-time signal into the Mel frequency domain through Mel filtering, and we get p parameters

[0055] ;

[0056] Where U is the number of FFT points; is the short-time signal after fast Fourier transform; are the filter parameters, It is an index variable used to traverse the fast Fourier transform FFT result, and its value range is 0 to U-1; here i and j represent the short-time signal after transforming to the Mel frequency domain. The corresponding index of each parameter ranges from 1 to ;

[0057] Will Perform a discrete cosine transform:

[0058] ;

[0059] based on Get MFCC feature vector.

[0060] Optionally, the weighted screening of the Mel-frequency cepstral coefficient feature vectors using a support vector machine recursive feature elimination method includes:

[0061] Use the set of all MFCC features as the initialization feature set F；

[0062] Use the current feature set F to train an SVM classifier and obtain a weight vector;

[0063] Calculate the importance score of each MFCC feature based on the weight vector;

[0064] Sort by importance score to get the sorted list L;

[0065] From the feature set F Remove the least important features and update the feature set F；

[0066] Repeat the above steps and iterate until the termination condition is met to obtain the final feature set;

[0067] The MFCC feature vectors of the final feature set are weighted to obtain weighted feature vectors.

[0068] Optionally, the principal component analysis algorithm is combined to compress the dimension of the weighted filtered eigenvector to generate a final feature matrix, including:

[0069] The new eigenvector is generated by linearly combining the dimensional components of the original eigenvector so that the dimensional components of the new eigenvector are independent of each other and the variance is maximized;

[0070] Find the linear combination F1 with the largest variance among all linear combinations, which is called the first principal component; then find the linear combination F2 with the largest variance among the linear combinations that are unrelated to F1, which is called the second principal component;

[0071] In this way, the principal components of each order that are independent of each other and have decreasing variances are obtained to form the final feature matrix, which is used to input the classification model for fault identification.

[0072] In a second aspect of the present application, an adaptive wavelet optimization and feature extraction system for transformer acoustic signals is provided, comprising:

[0073] Acquisition module: collects the original sound signal;

[0074] Decomposition module: calls the Pywt wavelet analysis library to decompose the original signal collected by the acquisition module according to the decomposition layer number J to obtain multi-layer detail signals;

[0075] Feature extraction module: For each layer of detail signal obtained by the decomposition module, perform the following operations:

[0076] The M-estimator is introduced to improve the noise variance calculation model, calculate the standard deviation and the unified monitoring threshold;

[0077] Based on the standard deviation and the unified monitoring threshold, a dynamic threshold based on a criterion of minimizing the approximation error of the denoised signal is constructed;

[0078] Design correction factor;

[0079] Based on the correction factor and the dynamic threshold, the wavelet coefficients of the detail signal of each layer are corrected;

[0080] Based on the modified wavelet coefficients, the pure signal is reconstructed using the inverse decomposition method;

[0081] Dividing the clean signal into multiple short-time signals;

[0082] Extracting a Mel-frequency cepstral coefficient feature vector of each of the short-time signals;

[0083] Using a support vector machine recursive feature elimination method to weightedly screen the Mel-frequency cepstral coefficient feature vector;

[0084] Combined with the principal component analysis algorithm, the dimension of the eigenvector after weighted screening is compressed to generate the final feature matrix.

[0085] The adaptive wavelet optimization and feature extraction method for transformer acoustic signals provided in this application adopts the M-estimator method to calculate the model, the dynamic threshold function and combines the adaptive factor for denoising, and uses the support vector machine recursive feature elimination method (SVM-RFE) and the principal component analysis (PCA) algorithm to extract features, effectively improving the accuracy of transformer fault identification, solving the problems of the contradiction between noise suppression and signal fidelity in traditional methods, and the low recognition rate caused by high-dimensional feature redundancy. The technical effect can be applied to the online diagnosis system of transformer faults such as mechanical looseness and winding deformation.

[0086] Other technical effects brought about by the additional features will be further explained in the corresponding embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0088] Figure 1 4 is a flow chart of a method for adaptive wavelet optimization and feature extraction of transformer acoustic signals according to an exemplary embodiment;

[0089] Figure 2 is a noise preprocessing flowchart according to an exemplary embodiment;

[0090] Figure 3 1 is a schematic diagram of multi-level wavelet decomposition according to an exemplary embodiment;

[0091] Figure 4 FIG4 is a diagram showing a process of obtaining MFCC according to an exemplary embodiment;

[0092] Figure 5 A feature optimization flowchart according to an exemplary embodiment is shown;

[0093] Figure 6 The figure is a block diagram of a system for adaptive wavelet optimization and feature extraction of transformer acoustic signals according to an exemplary embodiment. DETAILED DESCRIPTION

[0094] The present application is described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present application, but are not intended to limit the present application in any form. It should be noted that, without departing from the concept of the present application, a number of variations and improvements may be made by those skilled in the art, and these all fall within the scope of protection of the present application. Parts not described in detail in the following examples may be implemented using existing technologies.

[0095] Traditional methods have problems such as the contradiction between noise suppression and signal fidelity, and low recognition rate due to high-dimensional feature redundancy. Based on the above problems, the present invention provides an adaptive wavelet optimization and feature extraction method for transformer acoustic signals to solve the above problems.

[0096] Reference Figure 1 、 Figure 2 and Figure 3 As shown, in one embodiment of the present application, a method for adaptive wavelet optimization and feature extraction of transformer sound signals may include the following steps:

[0097] S1, collects the original sound signal;

[0098] S2, calling the Pywt wavelet analysis library, performs multi-scale decomposition on the original signal according to the decomposition layer number J to obtain multi-layer detail signals;

[0099] Specifically, Pywt is an open-source wavelet transform library for Python. It provides various wavelet transform algorithms and tools for signal processing, image processing, data analysis, and other fields. It can also perform multi-scale decomposition of sound signals and extract features of different frequency components.

[0100] In wavelet transforms, a signal is decomposed into approximate (low-frequency) and detail (high-frequency) components of different scales. The detail signal refers to the sub-signal corresponding to the high-frequency portion during the decomposition process. It represents detailed information such as local mutations, edges, or noise in the signal.

[0101] The detail coefficient (Cd) is the coefficient value corresponding to the detail signal after wavelet decomposition. These coefficients are the numerical results of the signal projected onto the high-frequency subspace by the wavelet basis function, reflecting the detail intensity of the signal at a specific scale and direction.

[0102] After obtaining multiple layers of detail signals in steps S1-S2, for each layer of detail signals, perform the following steps S3-S11:

[0103] S3, introduces the M estimator to improve the noise variance calculation model, calculates the standard deviation and the unified monitoring threshold;

[0104] S4, based on the standard deviation and the unified monitoring threshold, construct a dynamic threshold based on the criterion of minimizing the approximation error of the denoised signal;

[0105] S5, design correction factor;

[0106] S6, based on the correction factor and dynamic threshold, correct the wavelet coefficients of each layer of detail signal;

[0107] S7, reconstruct the denoised wavelet using the inverse decomposition method to obtain a pure signal;

[0108] S8, divides the pure signal into several short-time signals;

[0109] S9, extracting the Mel-frequency cepstral coefficient feature vector for each short-time signal;

[0110] S10, using support vector machine recursive feature elimination method to perform weighted screening on Mel-frequency cepstral coefficient feature vectors;

[0111] S11, combining the principal component analysis algorithm to compress the dimensions of the weighted and filtered eigenvectors to generate the final eigenmatrix.

[0112] It's important to note that each layer of detail signal corresponds to a final feature matrix—the final result of feature extraction—which can be input into a classification model for fault diagnosis. Specifically, the final feature matrix corresponding to each layer of detail signal is input into the classification model to determine whether the detail signal at that layer is faulty. Once all layers of detail signals have been identified, the model has completed a complete traversal of the original acquired signal.

[0113] In this application, the M-estimator method is used for model calculation and dynamic threshold function setting, combined with an adaptive factor to achieve denoising. Furthermore, a support vector machine recursive feature elimination (SVM-RFE) method and a principal component analysis (PCA) algorithm are used for feature extraction. This method effectively improves the accuracy of transformer fault identification, resolving the contradiction between noise suppression and signal fidelity in traditional methods, as well as the low recognition rate caused by high-dimensional feature redundancy.

[0114] In order to better collect sound signals, in some specific embodiments of the present application, when implementing the above S1, the original sound signal can be collected by the following steps:

[0115] S1.1, select sensor.

[0116] A wide-band acoustic sensor is used with a frequency response range of 20 Hz to 20 kHz and a sampling rate of 44.1 kHz to ensure coverage of the typical frequency bands of transformer mechanical noise and partial discharge acoustic signals.

[0117] S1.2, signal acquisition.

[0118] Deploy sensors near the transformer body and radiator, collect continuous acoustic signals for at least 10 minutes, and save them as WAV format files.

[0119] Sensors installed in the transformer's main body and on its heat sink area accurately capture key acoustic signals, such as mechanical vibration and partial discharge, during operation. These signals directly reveal fault signs such as core loosening and winding deformation. Long-term signal acquisition ensures that fault characteristics (such as intermittent discharge sounds) are fully captured, providing sufficient data support for subsequent analysis. Furthermore, the deployment of multiple sensors, combined with sound source localization technology, effectively distinguishes the noise generated by the transformer main body from that generated by the heat sink, thereby enhancing the accuracy of fault diagnosis.

[0120] Of course, in some specific embodiments, the collected original signal may be subjected to filtering preprocessing.

[0121] Specifically, a Butterworth bandpass filter can be used to filter out power frequency interference and high-frequency electromagnetic noise, and the cutoff frequency range of the filter is set in the range of 20 Hz to 20 kHz. The signal obtained after filtering is recorded as .

[0122] Next, the filtered signal Carry out normalization processing to eliminate the difference in sensor gain. The formula is:

[0123] ;

[0124] μ x is the filtered signal The mean of all sampling points, σ x is the filtered signal The standard deviation is used to characterize the fluctuation range of the signal amplitude.

[0125] In the aforementioned embodiments of this application, the purpose of pre-filtering is to decouple the 50 / 60 Hz fundamental frequency and its harmonics (such as 100 Hz and 150 Hz) from the acoustic signal in the power system, eliminating the interference of these frequency components on the valid signal through filtering technology. This step also removes high-frequency electromagnetic interference above 20 kHz (such as switching power supply noise and radio frequency interference), while retaining the main energy frequency band of the transformer mechanical noise, ensuring the effectiveness of the subsequent feature extraction process.

[0126] It should be noted that the pre-processing method is not limited to a single method, and only the above-mentioned filtering process may be performed, or only the normalization process may be performed.

[0127] In the preprocessing of the above-mentioned embodiments of the present application, the normalization process mainly solves the following two problems: first, due to the differences in the sensitivity of different sensors, the amplitudes of the collected signals are not uniform, and the normalization process can standardize the data; second, normalization helps to eliminate the dimensional influence of the amplitude, thereby avoiding weight deviations due to differences in the range of feature values ​​when using classification models (such as support vector machines SVM).

[0128] In an embodiment of the present application, S2 calls the Pywt wavelet analysis library to perform multi-scale decomposition on the original signal according to the decomposition layer number J to obtain multi-layer detail signals for multi-resolution analysis.

[0129] like Figure 3 As shown, in a specific embodiment, when the original signal is subjected to multi-scale decomposition according to the decomposition layer number J:

[0130] First layer (when J=1):

[0131] The original signal X is split into (First level approximation) + (First level of detail). It is the "simplified trend" of X. It is the "details / high frequency components" that are filtered out by X.

[0132] Second layer (when J=2):

[0133] Use the approximate signal of the first layer Continue to decompose and get (Second level approximation) + (Second level of detail). It is a trend towards simplification. yes The details in the Lower frequency detail (larger scale).

[0134] Nth layer (when J=N, where N is greater than 2): Repeat the above process until it is decomposed to the Nth layer, and finally obtain Can (Nth layer approximation) + Cdn (Nth layer details).

[0135] Exemplary:

[0136] Select the wavelet basis function as Symlet 5.

[0137] The number of decomposition layers J=5 ensures that low-frequency features and high-frequency noise are effectively separated.

[0138] Perform multi-scale decomposition on the original signal to obtain the detail coefficient Cd of each layer j,i .

[0139] In the embodiments of the present application, the original signal is decomposed into different frequency channels by wavelet multi-scale decomposition technology, thereby realizing multi-resolution analysis. Each layer of detail signal corresponds to a specific frequency band feature (for example, Capture high frequency noise, Capturing edges, etc.). Combined with its time-frequency localization properties, this technique not only accurately locates signal mutation points but also effectively separates components of different frequencies, making it particularly suitable for processing non-stationary signals. The number of decomposition layers J determines the granularity of the features. Shallow decomposition helps capture macroscopic trends in the signal, while deep decomposition facilitates the extraction of microscopic anomalies.

[0140] Traditional wavelet thresholding algorithms use the mean value of sample data to calculate the standard deviation. However, when the sample data has a polarization distribution or outliers appear, the standard deviation cannot accurately reflect the overall situation of the sample data. To overcome this problem, the embodiment of the present application introduces the M estimator into the standard deviation calculation and suppresses outlier interference through a robust loss function.

[0141] In order to accurately reflect the overall situation of the sample data, such as Figure 2 As shown, in some specific implementations of the present application, in the above step S3, the M estimator is introduced to improve the noise variance calculation model, and the standard deviation and the unified monitoring threshold are calculated. The following steps can be adopted:

[0142] S3.1, calculate the parameter position and initial scale parameters:

[0143] Position parameter μ=median(Cd j,i );

[0144] Initial scale parameter σ0 = 1.4826 × MAD (Cd j,i ); MAD represents median absolute deviation, median represents median; Cd j,i Represents the i-th detail coefficient of the j-th layer detail signal;

[0145] S3.2, based on the above position parameter μ, calculate the robust loss function (Huber function) :

[0146] ;

[0147] Among them, c is a constant, the position parameter μ is the j-layer wavelet detail coefficient Cd j,i The median of , its physical meaning is to make a robust estimate of the “center position” of the detail coefficient of the j-th layer, replacing the mean in the traditional method to suppress outlier interference.

[0148] For example, the parameter c of the Huber function is 1.345, which ensures 95% estimation efficiency for normal data.

[0149] S3.3, based on the initial scale parameter σ0 in S3.1 and the loss function in S3.2 , iteratively calculate the standard deviation of the detail signal σ j (j is the index of the decomposition layer of the original signal, ranging from 1 to J):

[0150] ;

[0151] The iteration termination condition is that the change in the estimated values ​​of two adjacent iterations is less than the set threshold: that is, the difference between the standard deviation of the j-th layer detail signal calculated by the k+1th iteration and the kth iteration is less than the set threshold, for example: , subscript j represents the number of wavelet decomposition layers, and N represents the detail coefficients after the jth wavelet decomposition layer. The total number of represents the robust loss function (Huber function).

[0152] S3.4, unified threshold calculate:

[0153] ;

[0154] N is the number of detail coefficients in the jth layer.

[0155] In the above embodiments of the present application, by combining the robust median with MAD (median absolute deviation) and introducing the Huber function for parameter estimation, this denoising process effectively solves the problem that the traditional mean-standard deviation model is easily biased by outliers.

[0156] The threshold is the critical value for distinguishing noise signals from useful signals. If the threshold is set too small, the interference signal may not be removed completely. If the threshold is set too large, the useful signal may be removed. In the process of traditional wavelet transform decomposing the signal, the noise in the wavelet detail coefficient decreases as the number of decomposition layers increases. Using a unified threshold strategy for denoising will damage the useful signal. In order to overcome this problem, S4 of the embodiment of the present application is based on the standard deviation and a unified monitoring threshold to construct a dynamic threshold based on the criterion of minimizing the approximation error of the denoised signal. The following steps can be adopted:

[0157] S4.1 Risk vector calculation.

[0158] Arrange the absolute values ​​of the detail coefficients of the jth layer in ascending order as a vector P=[CD j,1 ,CD j,2 ,…,CD j,N ].

[0159] Calculate each element ri of the risk vector R. The risk vector R is a dynamic threshold selection tool constructed based on statistical estimation theory. Its essence is to quantify the expected loss of signal denoising at different thresholds, thereby achieving data-driven optimal threshold decision. Specifically, the calculation of each element ri is as follows:

[0160] ;

[0161] Take the minimum value r in R min , record its corresponding CD min .

[0162] in, j Indicates the number of layers of wavelet decomposition, which is used to distinguish detail coefficients at different levels. Different layers reflect different frequency characteristics of the signal. i It is the j The layer detail coefficients are arranged in ascending order and used to calculate the risk vector elements ri The index ranges from 1 toN , corresponding to different positions after arrangement. k is the index of summation operation. N represents the total number of detail coefficients of the jth layer, which is the ascending order vector P=[CD j,1 ,CD j,2 ,…,CD j,N ] length.

[0163] Specifically, in this formula, CD j,k are the absolute values ​​of detail coefficients in ascending order. CD j,i When it is a large value coefficient, it is in the summation term and the remainder A higher proportion in ri The value is relatively small (i.e., the risk is lower). R The minimum value in r min and its corresponding CD min , the final dynamic threshold will tend to retain values ​​greater than CD min The large-value coefficient of is used to avoid it being mistakenly identified as noise and removed.

[0164] S4.2, Calculate the adaptive threshold :

[0165] ;

[0166] of It is the abbreviation of adaptive, which is used to identify the threshold as an adaptive threshold;

[0167] S4.3, Threshold selection rules:

[0168] Calculate the average value of the detail coefficient of layer j :

[0169] ;

[0170] in Stands for average, which is used to represent the average value of the detail coefficient of the jth layer.

[0171] Calculating minimum energy levels :

[0172] ;

[0173] Dynamic Threshold choose:

[0174] ;

[0175] Specifically, when When , it means that the noise energy of this layer is low, and a unified threshold is selected. .when When , it indicates that the noise energy of this layer is high, and a unified threshold is selected With adaptive threshold The minimum value among , that is, the hybrid threshold is adopted. By adaptively selecting the threshold according to different noise scenarios (low noise layer, high noise layer), it is possible to better preserve weak signals while suppressing strong noise.

[0176] In the above-mentioned embodiments of the present application, dynamic threshold selection is adopted. By selecting the minimum risk vector to optimize the threshold parameters, the problem of residual noise or over-smoothing of the signal caused by the fixed threshold in traditional wavelet denoising is effectively solved. This method has three core advantages: First, by using the risk vector to quantify the cost of different threshold selections, intelligent noise-signal separation is achieved, and the optimal threshold point is automatically located, avoiding the blindness of manual parameter adjustment; second, scene adaptive selection is adopted to dynamically adjust the threshold type according to the noise energy level (a unified threshold is used for the low-noise layer, and a mixed threshold is used for the high-noise layer), taking into account both high-frequency strong noise suppression and low-frequency weak signal retention; third, by giving priority to retaining large-value coefficients, pulse-type fault characteristics such as mechanical looseness are effectively protected to prevent them from being accidentally deleted.

[0177] Since the wavelet coefficients of each decomposition layer are quite different, the soft and hard threshold compromise method with fixed parameters does not have adaptive characteristics and cannot effectively solve the problem of effective waveform deformation and damage. Therefore, according to the law that the wavelet detail coefficient changes with the number of decomposition layers, the adaptive correction factor is used in step S5 of the embodiment of the present application. 。

[0178] In some specific embodiments of the present application, step S5 designs the correction factor , the following formula can be used:

[0179] ;

[0180] Among them, B j is the absolute value of the detail coefficient of the jth layer of the detail signal, represents the median, Indicates the absolute value of the kth detail coefficient of the jth layer detail signal, k It is the sequence number used to traverse each element in the detail coefficient vector of the j-th layer. The value range of k here is 1 to N, and N is the number of detail coefficients of the j-th layer.

[0181] Calculate the optimal parameter, i.e., the correction factor, based on the detail coefficient of each decomposition layer , improving the adaptability, robustness and accuracy of detail coefficient correction at different decomposition layers.

[0182] Specifically, B j is the absolute value of the detail coefficient of the jth layer of the detail signal. In the high-frequency layer dominated by noise, the noise accounts for a high proportion in the detail coefficient. The correction factor calculated by this formula is Will increase, and then strengthen the threshold contraction in the optimized j-th layer wavelet coefficient estimation function, so as to achieve effective suppression of strong noise. In the low-frequency layer dominated by the signal, the useful signal accounts for a high proportion in the detail coefficient. The correction factor calculated by this formula will decrease, thereby weakening the threshold contraction in the optimized j-th layer wavelet coefficient estimation function, retaining more useful coefficients and avoiding waveform distortion.

[0183] Therefore, the optimized j-th layer wavelet coefficient estimation function is:

[0184] ;

[0185] in is the wavelet estimation coefficient, is the wavelet coefficient, k It is the serial number used to traverse each element in the detail coefficient vector of the jth layer. The value range of k here is 1 to N. is the dynamic threshold of the jth layer, is the designed correction factor, and N is the number of detail coefficients in the jth layer.

[0186] Call the inverse wavelet transform function (waverec) of the Pywt library to reconstruct the signal and output the denoised time domain waveform.

[0187] The above embodiment of the present application significantly improves the performance of wavelet denoising by introducing a correction factor. The traditional soft and hard threshold compromise method uses fixed parameters, which is difficult to adapt to the characteristic differences of wavelet coefficients in different decomposition layers, and easily leads to residual noise in the high-frequency layer and loss of useful signals in the low-frequency layer. This embodiment is based on the amplitude of the decomposition layer detail coefficient. Dynamically calculating adaptive correction factors enables multi-scale adaptive noise reduction. In high-frequency layers dominated by noise, the correction factor is increased to strengthen threshold contraction, effectively suppressing strong noise. In low-frequency layers dominated by signals, the correction factor is reduced to retain more useful coefficients and avoid waveform distortion. The optimized coefficient estimation function combines the advantages of soft and hard thresholds, suppressing noise through soft thresholding and reducing reconstruction artifacts through hard thresholding, while achieving "precise noise reduction" through adaptive contraction.

[0188] When the above-mentioned embodiment of the present application reconstructs the signal through the inverse wavelet transform of the Pywt library, the optimization effect of each layer is transferred to the time domain waveform, and ultimately achieves the effects of improving the signal-to-noise ratio, improving the fault feature retention rate, and reducing the signal distortion rate in practical applications. It is particularly suitable for non-stationary signal processing and fault diagnosis scenarios, and provides a more reliable preprocessing tool for equipment status monitoring.

[0189] In order to make the reconstructed clean signal have the characteristic of short-term stability, in some specific embodiments of the present application, S8, dividing the clean signal into several short-term signal segments, can adopt the following steps:

[0190] S8.1 framing

[0191] Framing divides a continuous audio signal into multiple short "frames" to ensure short-term signal stability. The frame length L and overlap ratio k must balance signal analysis accuracy and computational efficiency. Too long a frame length L can compromise short-term stability, while too short a frame length can prevent sufficient feature extraction. An appropriate overlap ratio k can avoid sudden changes between adjacent frames and improve signal continuity.

[0192] Specifically, the number of frames The calculation formula is:

[0193] ;

[0194] Where n is the total signal length, L is the frame length, and d is the overlap ratio.

[0195] Exemplary:

[0196] Frame length L = 400ms: The corresponding number of sampling points is 44.1kHz × 0.4s = 1764 points (when the sampling rate is 44.1kHz, 1ms corresponds to approximately 44.1 sampling points);

[0197] Overlap rate d = 50%: the number of overlapping points in adjacent frames is 1764 × 50% = 882 points;

[0198] The total signal length n=10s: the corresponding number of sampling points is 44.1kHz×10s=441000 points.

[0199] frame

[0200] S8.2 Hamming window applied:

[0201] Window function:

[0202] ;

[0203] N refers to the length of the frame.

[0204] Each frame signal is windowed point by point to eliminate spectrum leakage caused by frame edge truncation.

[0205] In the aforementioned embodiment of the present application, S8.1 employs a 400-millisecond frame length and 50% overlapping framing to continuously capture signal stationarity and features over short time periods; S8.2 utilizes a Hamming window combined with an edge fading function to effectively suppress spectrum leakage. The combination of these two techniques overcomes the two major challenges of "non-stationary interference" and "spectral diffusion" encountered in long-term signal processing, thereby providing high-quality, low-noise time-frequency domain input signals for subsequent feature extraction steps.

[0206] In order to extract the features of the pure signal, MFCC features are extracted for each frame signal. In some specific implementations of the present application, S9 extracts the Mel frequency cepstral coefficient feature vector for each short-time signal, such as Figure 4 As shown, the following steps can be taken:

[0207] S9.1, Fast Fourier Transform FFT

[0208] In this embodiment, a 2048-point FFT is performed on each frame signal to obtain the spectrum modulus value |X( )∣, is the index variable used to traverse the Fast Fourier Transform FFT results.

[0209] Specifically, a total of 2048 sampling points are set here, mainly based on the following special design:

[0210] First, using 2048 sampling points optimizes frequency resolution. Specifically, the frequency resolution corresponding to 2048 sampling points is 44100 Hz divided by 2048, or approximately 21.53 Hz. This is sufficient to distinguish the characteristic frequency band of transformer core loosening, which ranges from 100 to 500 Hz.

[0211] Secondly, the Fast Fourier Transform (FFT) algorithm is most efficient when using a power of 2 point number. 2048 points (that is, 2 to the power of 11) can increase the calculation speed by more than three times compared to a non-power of 2 point number (such as 2000 points).

[0212] Finally, considering the frame length matching problem, the sub-frame length is set to 400 milliseconds (equivalent to 1764 sampling points). 2048 points is larger than 1764 points, thus avoiding the spectrum distortion problem caused by excessive zero padding.

[0213] S9.2, Filter Design

[0214] In this embodiment, 26 Mel triangle filters are designed, and the center frequencies are evenly distributed according to the Mel scale:

[0215] ;

[0216] The frequency domain response formula of the triangular filter Hm(k) is:

[0217] ;

[0218] m: filter number (mth Mel filter); : Frequency point index (the first frequency point of FFT output) frequency points); f(m): center frequency of the m-th filter (Mel scale mapping value).

[0219] Specifically, an optimized configuration of 26 Mel-triangle filters is employed, precisely adapting the nonlinear frequency distribution (dense at low frequencies and sparse at high frequencies) to the human auditory perception characteristics. This effectively resolves the contradiction in selecting the right number of filters in the traditional sense: fewer than 26 filters can easily lead to missed detection of high-frequency discharge features, while more than 26 filters unnecessarily increase the computational load. This number design focuses on the core frequency band of transformer faults (100Hz-4kHz), simultaneously capturing mechanical vibration and discharge noise characteristics. This ensures complete coverage of fault features while achieving efficient feature extraction with minimal computational cost. After compression into 12-dimensional Mel-Frequency Cepstral Coefficients (MFCCs) using the Discrete Cosine Transform (DCT), these filters achieve high diagnostic accuracy while meeting the real-time requirements of industrial sites, providing a lightweight feature characterization solution for transformer condition monitoring.

[0220] S9.3 Logarithmic Energy and DCT Transform:

[0221] Calculate the output energy of each filter:

[0222] ;

[0223] To m i Perform discrete cosine transform (DCT) and take the first 12 lowest dimensional coefficients as MFCC features:

[0224] ;

[0225] In this application, a MFCC feature extraction method combining Mel triangle filters with DCT cosine transform compression was used to successfully compress transformer acoustic signals into 12-dimensional, highly discriminative features. This method effectively addresses three major issues: inefficient high-dimensional data processing, the vulnerability of fault signatures to noise, and a mismatch with human ear perception. Actual tests have demonstrated that this technology achieves high fault recognition rates and real-time frame processing capabilities on embedded devices.

[0226] Each dimensional component of the MFCC feature vector contributes differently to voiceprint recognition. Most components contribute significantly to voiceprint recognition, while others contribute less, and a few components can even reduce the recognition rate. To address this issue, S10 in this embodiment of the present application uses the SVM-RFE (Support Vector Machine Recursive Feature Elimination) feature selection method to optimize the feature vector by recursively eliminating features.

[0227] like Figure 5 As shown, in some specific embodiments of the present application, S10, using a support vector machine recursive feature elimination method to perform weighted screening on the Mel-frequency cepstral coefficient feature vector, can be performed using the following steps:

[0228] S10.1, Initialization:

[0229] Assume that the original MFCC feature dimension is m=12 and the feature set F={1,2,…,12}.

[0230] S10.2, iterative elimination:

[0231] Round 1: Use all features of F=12 to train a linear SVM and calculate the weight vector w. Based on the weight vector, calculate the importance score of each MFCC feature and sort them from high to low according to the absolute value of the importance score to obtain the sorted list L=[w1,w2,…,w 12 ].

[0232] Eliminate the weakest feature: remove the last-ranked feature in L (e.g. w9 corresponds to the 9th-dimensional MFCC) and update the feature set .

[0233] Round 2: Retrain SVM based on F=11-dimensional features, update L, and continue to eliminate the weakest features.

[0234] After multiple rounds of elimination iterations, the iteration termination condition is reached. The termination condition is: the iteration is stopped when the number of features drops to a preset threshold (for example, the feature dimension is 6) or the classification accuracy drops by more than 5%.

[0235] Specifically, the above linear SVM classifier training process is as follows:

[0236] (1) Data preparation and standardization

[0237] Input data: MFCC feature matrix X ∈R n×m , where R represents the set of real numbers, n is the number of samples (the total number of frames after framing), m is the MFCC feature dimension.

[0238] Label assignment: Each sample corresponds to a transformer state label y∈{+1(fault), (normal)}.

[0239] Data normalization: Z-score normalization is performed on each MFCC feature dimension:

[0240] ;

[0241] in, μ j For the j dimensional mean, σ j is the standard deviation, Indicates the i The first sample j dimensional MFCC eigenvalues, i is the index number of the number of samples, j is the index number of the MFCC eigenvalue dimension, and its range gradually decreases with the number of iterations. In this embodiment, it gradually decreases from [1, 12] to [1, 6].

[0242] (2) Linear SVM model training

[0243] Optimization goal: Solve the following convex quadratic programming problem:

[0244] ;

[0245] Constraints:

[0246] , ;

[0247] Among them, C is the penalty coefficient, which controls the balance between classification error and interval, ξ i is a slack variable that allows some samples to be misclassified, b is the bias term of the classification hyperplane, w is the weight vector, and n is the number of samples.

[0248] Solving the dual problem:

[0249] By using the Lagrange multiplier method, the above convex quadratic programming problem is transformed into a dual problem and the Lagrange multiplier is solved. :

[0250] ;

[0251] , ;

[0252] Support vector determination: Corresponding samples x i is the support vector, The goal of solving the dual problem is the Lagrange multiplier ;y i 、y j Represents sample x i、 x j The corresponding label; n represents the number of samples.

[0253] Weight vector calculation: According to the solution of the dual problem , weight vector w It is composed of linear combinations of support vectors:

[0254] ;

[0255] Specifically, to calculate the importance score of each MFCC feature, the following steps can be used:

[0256] (1) Define the physical meaning of the weight vector:

[0257] Weight component w j :Corresponding to j The contribution direction and strength of dimensional MFCC features to classification decisions.

[0258] Positive weight: When the feature value increases, it tends to be predicted as a positive class (such as a fault).

[0259] Negative weight: When the feature value increases, it tends to be predicted as a negative class (such as normal).

[0260] Absolute value: directly reflects the importance of the feature. The larger the absolute value, the stronger the classification decision dependence.

[0261] (2) Feature importance scoring formula:

[0262] No. j Importance score of MFCC features for:

[0263] ;

[0264] x ij For the i The first sample j dimensional MFCC eigenvalues, i is the index number of the number of samples, j is the index number of the MFCC eigenvalue dimension, and its range gradually decreases with the number of iterations. In this embodiment, j is gradually reduced from [1, 12] to [1, 6].

[0265] S10.3 Feature weighting strategy:

[0266] The final feature set F = {2, 5, 7, 8, 10, 11} is retained. The final feature set is the MFCC features retained after iterative elimination in step S10.2 above. The weights are normalized by the absolute value of the weights. :

[0267] ;

[0268] The above-mentioned embodiment of this application uses SVM-RFE to weightedly filter MFCC feature vectors, automatically identifying transformer fault-sensitive frequency bands. This compresses the original feature dimensions by nearly half, significantly reducing computational effort while improving both fault feature retention and classification accuracy. The filtered feature subsets exhibit enhanced noise immunity, and feature importance ranking intuitively reflects the physical fault mechanism, balancing real-time diagnostic requirements with engineering interpretability.

[0269] A higher MFCC feature vector dimension can ensure that the noise signal features are fully extracted. However, if the dimension of the feature vector is too high, it will take a lot of time and the computational complexity will increase accordingly. Therefore, this application introduces the principal component analysis (PCA) algorithm to reduce the dimension of the calculated high-dimensional MFCC feature vector, while ensuring that the noise characteristics of the transformer are accurately obtained. In some specific implementations of this application, S11 combines the principal component analysis algorithm to compress the dimension of the weighted and screened feature vector to generate a low-redundancy feature matrix. The following steps can be used:

[0270] S11.1, linearly combine the dimensional components of the original eigenvector to generate a new eigenvector, so that the dimensional components of the new eigenvector are independent of each other and the variance is as large as possible.

[0271] S11.2, find the linear combination F1 with the largest variance among all linear combinations, called the first principal component; then find the linear combination F2 with the largest variance among the linear combinations uncorrelated with F1, called the second principal component... and so on, find the principal components of each order that are uncorrelated and have decreasing variances.

[0272] Specifically, the steps of principal component analysis are as follows:

[0273] Suppose there is a matrix X, which is composed of n eigenvectors, and the dimension of each eigenvector is m, then X can be expressed as:

[0274] ;

[0275] Find the correlation matrix R of X , and calculate the correlation matrix R The eigenvalues ​​and eigenvectors of :

[0276] ;

[0277] Calculate the correlation matrix R The eigenvalue of λ 1, λ 2,…, λ m And the eigenvectors u1,u2,…,u corresponding to each eigenvalue m .

[0278] Calculate the variance contribution rate And the cumulative variance contribution rate , and then determine the number of principal components:

[0279] ;

[0280] ;

[0281] The cumulative variance contribution rate can be used to determine the number of principal components. Generally, the number of principal components q is selected so that the cumulative variance contribution rate is greater than 75%, which is also the dimension of the eigenvector after dimensionality reduction.

[0282] Calculate the new eigenvectors: The q eigenvectors corresponding to the q principal components are U = [ u 1 , u 2 ,…,u q ], then the new eigenvector matrix Z can be expressed as: Z=XU.

[0283] For example, the weighted feature matrix X weighted ∈R n×6 Normalize by columns:

[0284] ;

[0285] μ j For the j dimensional mean, σ j is the standard deviation.

[0286] Covariance matrix decomposition: Calculate the covariance matrix (that is, the correlation matrix is ​​the covariance matrix) , solve its eigenvalue and eigenvectors u1,u2,…,u6.

[0287] Selection of the number of principal components: Calculation of cumulative variance contribution rate:

[0288] ;

[0289] Set the threshold η(q) ≥ 85%. If λ1 + λ2 = 80%, then q = 3 should be selected (assuming λ3 contributes 5%).

[0290] Feature projection and dimensionality reduction: Extract the eigenvectors U=[u1,u2,u3] corresponding to the first q=3 principal components.

[0291] Generate a low-dimensional feature matrix: Z=XU.

[0292] In the above embodiment of the present application, the support vector machine recursive feature elimination method (SVM-RFE) is used to perform weighted screening on the MFCC feature vector, and the principal component analysis (PCA) algorithm is combined to compress the feature dimension, eliminate redundant feature parameters, reduce the model calculation complexity, and at the same time improve the feature's discrimination for transformer faults, providing high-purity feature input for fault identification.

[0293] Through the optimized feature parameter extraction and noise reduction joint processing process, the acoustic characteristics of transformer faults such as mechanical looseness and winding deformation can be more accurately characterized, the misjudgment rate can be reduced, and high recognition accuracy can be maintained under complex working conditions, which is superior to the traditional single feature processing method.

[0294] Based on the same technical concept, in other embodiments, an adaptive wavelet optimization and feature extraction system 100 for transformer sound signals is provided. Figure 6 Shown, including:

[0295] Acquisition module 110: collects original sound signals;

[0296] Decomposition module 120: calls the Pywt wavelet analysis library to decompose the original signal collected by the acquisition module according to the decomposition layer number J to obtain multi-layer detail signals;

[0297] Feature extraction module 130: performs the following operations on each layer of detail signal obtained by the decomposition module:

[0298] The M-estimator is introduced to improve the noise variance calculation model, calculate the standard deviation and the unified monitoring threshold;

[0299] Based on the standard deviation and the unified monitoring threshold, a dynamic threshold based on the criterion of minimizing the approximation error of the denoised signal is constructed;

[0300] Design correction factor;

[0301] Based on the correction factor and dynamic threshold, the wavelet coefficients of each layer of detail signal are corrected;

[0302] Based on the modified wavelet coefficients, the pure signal is reconstructed using the inverse decomposition method;

[0303] Divide the pure signal into multiple short-time signals;

[0304] Extract the Mel-frequency cepstral coefficient feature vector of each short-time signal;

[0305] The support vector machine recursive feature elimination method is used to weightedly screen the Mel-frequency cepstral coefficient feature vectors;

[0306] Combined with the principal component analysis algorithm, the dimension of the eigenvector after weighted screening is compressed to generate the final feature matrix.

[0307] The specific implementation techniques of each module / unit in the above example of the present application may refer to the corresponding steps of the adaptive wavelet optimization and feature extraction method for transformer sound signals in the above embodiment, which will not be repeated here.

[0308] The above describes some specific embodiments of the present application. It should be understood that the present application is not limited to the specific embodiments described above, and those skilled in the art may make various variations or modifications within the scope of the claims, which do not affect the substantive content of the present application. The above preferred features may be used in any combination as long as they do not conflict with each other.

Claims

1. An adaptive wavelet optimization and feature extraction method for transformer acoustic signals, characterized in that: include: Collect the original sound signal; Call the Pywt wavelet analysis library to decompose the original signal according to the decomposition layer number J to obtain multi-layer detail signals; For each layer of detail signal, perform the following steps: The M-estimator is introduced to improve the noise variance calculation model, calculate the standard deviation and the unified monitoring threshold; Based on the standard deviation and the unified monitoring threshold, a dynamic threshold based on a criterion of minimizing the approximation error of the denoised signal is constructed; Design correction factor; Based on the correction factor and the dynamic threshold, the wavelet coefficients of the detail signal of each layer are corrected; Based on the modified wavelet coefficients, the pure signal is reconstructed using the inverse decomposition method; Dividing the clean signal into multiple short-time signals; Extracting a Mel-frequency cepstral coefficient feature vector of each of the short-time signals; Using a support vector machine recursive feature elimination method to weightedly screen the Mel-frequency cepstral coefficient feature vector; Combined with the principal component analysis algorithm, the dimension of the weighted filtered eigenvector is compressed to generate the final feature matrix; The introduction of the M-estimator to improve the noise variance calculation model, calculate the standard deviation and the unified monitoring threshold, includes: The M estimator is introduced into the standard deviation calculation, and the outlier interference is suppressed by the robust loss function. The standard deviation of the detail signal of the j-th layer is calculated as: Among them, ρ(·) is the robust loss function, namely the Huber function; Cd j,i is the detail coefficient after wavelet decomposition, N is the number of detail coefficients in the jth layer, j is the index of the decomposition layer of the original signal, and its value range is 1 to J; σ0 is the initial scale parameter, σ0=1.4826×MAD(Cd j,i ), MAD stands for median absolute deviation; μ is the position parameter, μ=median(Cd j,i ), median means median; Based on the standard deviation, a unified monitoring threshold of the j-th layer detail signal is obtained: The method of constructing a dynamic threshold based on a criterion of minimizing the approximation error of a denoised signal based on the standard deviation and a unified monitoring threshold includes: Arrange the absolute values ​​of the detail coefficients of the jth layer in ascending order as a vector P = [CD j,1 ,CD j,2 ,…,CD j,N ] Calculate each element r of the risk vector R i : The smallest element r in R min As the approximation error, record its corresponding CD min , k is the serial number used to traverse each element in the j-th layer detail coefficient vector, where the value of k ranges from 1 to i; r i where i is the vector P=[CD j,1 ,CD j,2 ,…,CD j,N ], whose value range is from 1 to N, where N is the number of detail coefficients in the j-th layer; Calculate the adaptive threshold of the j-th layer detail coefficient: λ a,j The a is the abbreviation of adaptive, which is used to identify the threshold as an adaptive threshold; Based on the unified monitoring threshold and the adaptive threshold, the dynamic threshold of the j-th layer is obtained: Among them, P a,j is the average value of the detail coefficient of the jth layer: P a,j The a in it stands for average, which is used to represent the average value of the detail coefficient of the jth layer, ρ N,j is the minimum energy level of the j-th layer detail coefficient:

2. The method for adaptive wavelet optimization and feature extraction of transformer acoustic signals according to claim 1, characterized in that: The correction factor a designed j for: Among them, B j is the absolute value of the detail coefficient of the jth layer of the detail signal, nedian represents the median, CD j,k Represents the absolute value of the kth detail coefficient of the jth layer detail signal, where k is the index used to traverse each element in the jth layer detail coefficient vector. The value of k here ranges from 1 to N, where N is the number of detail coefficients in the jth layer.

3. The method for adaptive wavelet optimization and feature extraction of transformer acoustic signals according to claim 1, characterized in that: The step of correcting the wavelet coefficients of each layer of detail signal based on the correction factor and the dynamic threshold comprises: in is the wavelet estimation coefficient, w j,k is the wavelet coefficient, k is the serial number used to traverse each element in the detail coefficient vector of the jth layer, where the value of k ranges from 1 to N, and λ j is the dynamic threshold of the jth layer, a j is the designed correction factor, and N is the number of detail coefficients in the jth layer.

4. The method for adaptive wavelet optimization and feature extraction of transformer acoustic signals according to claim 1, characterized in that: The step of dividing the clean signal into a plurality of short-time signals comprises: The clean signal is divided into M frame signals, wherein: Where M is the number of frames, n is the length of the clean signal, L is the frame length, and d is the overlap ratio; Hamming window processing is applied to the M-frame framed signal, and point-by-point window multiplication is performed to eliminate spectrum leakage caused by frame edge truncation and obtain the optimized short-time signal.

5. The method for adaptive wavelet optimization and feature extraction of transformer acoustic signals according to claim 1, characterized in that: The step of extracting a Mel-frequency cepstral coefficient feature vector of each short-time signal comprises: Performing a fast Fourier transform on a single short-time signal and obtaining a modulus value thereof; Use p filters to transform the short-time signal into the Mel frequency domain through Mel filtering to obtain p parameters Where U is the number of FFT points; X(y') is the short-time signal after fast Fourier transform; H i (y') is the filter parameter, y' is the index variable used to traverse the fast Fourier transform (FFT) results, and its value range is 0 to U-1; here i and j represent the corresponding indexes of the p parameters obtained after the short-time signal is transformed into the Mel frequency domain, and their value ranges are both 1 to p; M i Perform a discrete cosine transform: Obtain MFCC feature vector based on c(i).

6. The method for adaptive wavelet optimization and feature extraction of transformer acoustic signals according to claim 1, characterized in that: The weighted screening of the Mel-frequency cepstral coefficient feature vectors using a support vector machine recursive feature elimination method includes: Use the set of all MFCC features as the initialization feature set F; Use the current feature set F to train an SVM classifier and obtain a weight vector; Calculate the importance score of each MFCC feature based on the weight vector; Sort by importance score to get the sorted list L; Remove the least important feature from the feature set F and update the feature set F; Repeat the above steps and iterate until the termination condition is met to obtain the final feature set; The MFCC feature vectors of the final feature set are weighted to obtain weighted feature vectors.

7. The method for adaptive wavelet optimization and feature extraction of transformer acoustic signals according to claim 1, characterized in that: The principal component analysis algorithm is combined to compress the dimension of the weighted filtered eigenvector to generate the final feature matrix, including: The new eigenvector is generated by linearly combining the dimensional components of the original eigenvector so that the dimensional components of the new eigenvector are independent of each other and the variance is maximized; Find the linear combination F1 with the largest variance among all linear combinations, which is called the first principal component; then find the linear combination F2 with the largest variance among the linear combinations that are unrelated to F1, which is called the second principal component; In this way, the principal components of each order that are independent of each other and have decreasing variances are obtained to form the final feature matrix, which is used to input the classification model for fault identification.

8. An adaptive wavelet optimization and feature extraction system for transformer acoustic signals, characterized in that: include: Acquisition module: collects the original sound signal; Decomposition module: calls the Pywt wavelet analysis library to decompose the original signal collected by the acquisition module according to the decomposition layer number J to obtain multi-layer detail signals; Feature extraction module: For each layer of detail signal obtained by the decomposition module, perform the following operations: The M-estimator is introduced to improve the noise variance calculation model, calculate the standard deviation and the unified monitoring threshold; Based on the standard deviation and the unified monitoring threshold, a dynamic threshold based on a criterion of minimizing the approximation error of the denoised signal is constructed; Design correction factor; Based on the correction factor and the dynamic threshold, the wavelet coefficients of the detail signal of each layer are corrected; Based on the modified wavelet coefficients, the pure signal is reconstructed using the inverse decomposition method; Dividing the clean signal into multiple short-time signals; Extracting a Mel-frequency cepstral coefficient feature vector of each of the short-time signals; Using a support vector machine recursive feature elimination method to weightedly screen the Mel-frequency cepstral coefficient feature vector; Combined with the principal component analysis algorithm, the dimension of the weighted filtered eigenvector is compressed to generate the final feature matrix; The introduction of the M-estimator to improve the noise variance calculation model, calculate the standard deviation and the unified monitoring threshold, includes: The M estimator is introduced into the standard deviation calculation, and the outlier interference is suppressed by the robust loss function. The standard deviation of the detail signal of the j-th layer is calculated as: Among them, ρ(·) is the robust loss function, namely the Huber function; Cd j,i is the detail coefficient after wavelet decomposition, N is the number of detail coefficients in the jth layer, j is the index of the decomposition layer of the original signal, and its value range is 1 to J; σ0 is the initial scale parameter, σ0=1.4826×MAD(Cd j,i ), MAD stands for median absolute deviation; μ is the position parameter, μ=median(Cd j,i ), median means median; Based on the standard deviation, a unified monitoring threshold of the j-th layer detail signal is obtained: The method of constructing a dynamic threshold based on a criterion of minimizing the approximation error of a denoised signal based on the standard deviation and a unified monitoring threshold includes: Arrange the absolute values ​​of the detail coefficients of the jth layer in ascending order as a vector P = [CD j,1 ,CD j,2 ,…,CD j,N ] Calculate each element r of the risk vector R i : The smallest element r in R min As the approximation error, record its corresponding CD min , k is the serial number used to traverse each element in the j-th layer detail coefficient vector, where the value of k ranges from 1 to i; r i where i is the vector P=[CD j,1 ,CD j,2 ,…,CD j,N ], whose value range is from 1 to N, where N is the number of detail coefficients in the j-th layer; Calculate the adaptive threshold of the j-th layer detail coefficient: λ a,j The a is the abbreviation of adaptive, which is used to identify the threshold as an adaptive threshold; Based on the unified monitoring threshold and the adaptive threshold, the dynamic threshold of the j-th layer is obtained: Among them, P a,j is the average value of the detail coefficient of the jth layer: P a,j The a in it stands for average, which is used to represent the average value of the detail coefficient of the jth layer, ρ N,j is the minimum energy level of the j-th layer detail coefficient:

Images