A transformer sound signal processing method based on CEEMDAN

Abstract

The application discloses a transformer sound signal processing method based on CEEMDAN, collects a noisy transformer sound signal for processing, obtains a signal matrix and carries out pretreatment, obtains a pretreated signal, uses a correlation coefficient and a permutation entropy algorithm to screen the pretreated signal, inputs the screened signal into a CEEMDAN model for denoising processing, inputs the denoised signal into a Deep Focusing U-Net model for processing, obtains a target source signal, constructs a transformer target voiceprint library, and registers and trains the target source signal with a ResNet34 model, the trained model is used for processing the target source signal and comparing with the registered signal, when the comparison result is higher than a judgment threshold, the target source signal is input into the voiceprint library for identification, and the category of the transformer target sound source is output. The method combines the correlation coefficient and the permutation entropy to screen the signal component, and improves the accuracy of voiceprint identification.

Description

Technical Field

[0001] This invention belongs to the field of sound signal processing technology, specifically relating to a transformer sound signal processing method based on CEEMDAN. Background Technology

[0002] As core equipment in the power system, real-time monitoring of the transformer's operating status is crucial for ensuring the safety and stability of the power grid. Sound signals, serving as an important "barometer" of transformer operating status, can directly reflect the equipment's health condition, such as stable power frequency noise during normal operation, short-circuit impact noise during fault conditions, and abnormal noises from loose clamps. Therefore, analyzing transformer sound signals to achieve status monitoring has become a key technological direction in the field of power equipment operation and maintenance.

[0003] However, acquiring transformer sound signals under actual operating conditions faces severe challenges. The on-site environment contains a large amount of interfering noise, including environmental noise (such as wind and rain, birdsong, and human voices) and power system interference (such as bus corona noise and cooling system noise). When these interfering signals are superimposed on the target sound signal from the transformer, refraction and reflection during propagation result in an extremely low signal-to-noise ratio (SNR) for the acquired mixed signal. This low SNR directly affects the accuracy of subsequent voiceprint recognition, causing traditional acoustic monitoring systems to frequently misjudge and fail to reliably distinguish between normal operating sounds, fault sounds, and interference sounds, severely limiting the practicality of transformer condition monitoring.

[0004] To address noise interference, various signal denoising techniques have been proposed in the industry, among which Empirical Mode Decomposition (EMD) and Ensemble Empirical Mode Decomposition (EEMD) are commonly used methods. However, these two methods have significant drawbacks when processing complex, non-stationary audio signals from transformers: firstly, they are prone to mode aliasing, meaning that signal components of different frequencies are incorrectly decomposed into the same intrinsic mode function (IMF), resulting in the inability to effectively separate the effective signal from noise; secondly, they suffer from boundary effects, where distortion at both ends of the signal during decomposition affects the overall denoising effect, making it difficult to meet the requirements for high-purity signal extraction.

[0005] In the target signal separation stage, traditional blind source separation techniques (such as Fast Independent Component Analysis, FastICA) rely on the assumption of signal independence and are only applicable to "positive definite" scenarios where the number of sensors is not less than the number of signal sources. However, the types of noise sources in transformer sites are random and complex (up to dozens), and the number of sensors deployed is often limited by the installation environment, frequently resulting in an "underdetermined" state (number of sensors < number of signal sources). In this case, traditional blind source separation algorithms cannot effectively handle "ill-conditioned" mixed signals, resulting in low target source signal separation accuracy and difficulty in extracting pure transformer sound signals.

[0006] Furthermore, existing voiceprint recognition models lack robustness in complex noisy environments. Traditional machine learning models (such as support vector machines and random forests) are heavily reliant on feature extraction; when the input signal is noisy, feature discriminative power decreases, leading to a significant reduction in the model's accuracy in classifying normal transformer sounds and fault sounds. Although some studies have introduced deep learning models, most of these models are not optimized for the characteristics of transformer sound signals, failing to fully capture the deep features of the signals and thus making it difficult to meet the high-precision requirements of power systems for equipment condition monitoring.

[0007] In summary, current transformer audio signal processing technologies have shortcomings in denoising effectiveness, underdetermined blind source separation capability, and robustness of voiceprint recognition. There is an urgent need for a comprehensive processing method that can achieve high signal-to-noise ratio denoising, accurate target signal separation, and reliable voiceprint recognition in high-interference environments, so as to improve the accuracy and reliability of transformer condition monitoring. Summary of the Invention

[0008] To address the problems in the related technologies, this application provides a transformer-based audio signal processing method based on CEEMDAN, which solves the problems mentioned in the background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a transformer audio signal processing method based on CEEMDAN, comprising the following steps:

[0010] Step S1: Collect and process the noisy transformer sound signals to obtain a mixed signal matrix to be processed;

[0011] Step S2: Preprocess the mixed signal matrix to obtain the preprocessed sound signal;

[0012] Step S3: The preprocessed audio signal is processed using the correlation coefficient and permutation entropy algorithm to obtain a preliminary preprocessed audio signal. The preliminary preprocessed audio signal is then denoised based on the CEEMDAN model to obtain a denoised audio signal.

[0013] Step S4: Input the denoised audio signal into the Deep Focusing U-Net model and process it through underdetermined blind source separation to obtain the target source signal;

[0014] Deep Focusing U-Net model processing procedure:

[0015] The denoised audio signal is sequentially input into five downsampling blocks for processing, resulting in downsampling feature maps from the first to the fifth layer. These downsampling feature maps are then linearly interpolated to obtain linearly interpolated feature maps from the first to the fifth layer. These linearly interpolated feature maps are then input into five upsampling blocks via a first skip connection. Each upsampling block fuses the linearly interpolated feature maps input from the first skip connection through an upsampling operation, resulting in upsampling feature maps from the first to the fifth layer. Finally, these upsampling feature maps are added one by one via a second skip connection and an upsampling operation to obtain the target source signal.

[0016] Step S5: Construct a target acoustic signature library for transformers and register the target acoustic source signals and train a ResNet34 deep learning model. The trained ResNet34 deep learning model processes the target acoustic source signals and compares the processing results with the target acoustic source signals. If the comparison result is higher than the set judgment threshold, the target acoustic source signal is retained. The retained target acoustic source signal is input into the target acoustic signature library for transformers for identification, and the final target acoustic source category of the transformer is output.

[0017] Furthermore, the specific process for obtaining the preprocessed audio signal is as follows:

[0018] The analog signals of the mixed signal matrix to be processed are converted into digital signals according to the Nelster criterion.

[0019] The amplitude of the noisy transformer-controlled sound signal is replaced by the numerical value of the digital signal. Specifically:

[0020] The amplitude range of the entire noisy transformer sound signal is normalized. After normalization, it is divided into quantization intervals with fixed intervals. The amplitude within the quantization interval is replaced by the value of the digital signal to obtain the normalized and quantized digital signal.

[0021] Based on the normalized and quantized digital signal, a high-pass filter operation is used to enhance the high-frequency components of the noisy transformer sound signal to obtain the enhanced high-frequency components.

[0022] Based on enhancing high-frequency components, a discrete Fourier transform is performed on the noisy transformer sound signal to obtain a preprocessed sound signal.

[0023] Furthermore, the specific process for initially screening and preprocessing audio signals is as follows:

[0024] Gaussian white noise is added to the preprocessed audio signal to construct the signal to be separated. Empirical mode decomposition (EMD) is performed on the constructed signal to be separated to obtain multiple intrinsic mode function (IMF) components. The average of the multiple IMF components is calculated to obtain the first IMF component. The first IMF component is subtracted from the preprocessed audio signal to obtain the residual component at time t. ;

[0025] right Update until Stop when the conditions are met, specifically:

[0026] right Add Gaussian white noise again and perform empirical mode decomposition to update. From the update Extract three consecutive points as follows , , When satisfied > and > At that time, It is a local maximum point; when it satisfies < and < At that time, It is a local minimum point; when Greater than or less than both sides and hour, It is identified as a maximum or minimum point; when updating If there is only one maximum or minimum point, the condition is met and the process stops. Consider it as a new intrinsic mode function (IMF); The signal value of the residual component at time 1; The signal value of the residual component at the second time step; The signal value of the residual component at time 3;

[0027] Power spectrum analysis of the IMF using short-time Fourier transform yields the k-th intrinsic mode function component. The dominant frequency fmax is the kth eigenmode function component. The main frequency fmax and the set cutoff frequency In comparison, when it is less than the cutoff frequency The decomposition ends, and the k-th intrinsic mode function component is output. Conversely, the decomposition continues;

[0028] By combining correlation coefficient and permutation entropy algorithms, The audio signals are then filtered to obtain pre-processed audio signals.

[0029] Calculate the correlation coefficient separately. The correlation coefficient with the corresponding preprocessed audio signal represents:

[0030] ;

[0031] In the formula, For the k-th intrinsic mode function component The degree of linear correlation between y(t) and y(t); Let be the mean of y(t); This represents the total number of intrinsic mode function components; The value of the k-th eigenmode function component at time t; y(t) is the average value of the k-th intrinsic mode function component; y(t) is the preprocessed audio signal at time t.

[0032] In the permutation entropy algorithm, the k-th eigenmode function component is traversed. The time series data is processed by extracting embedding vectors using a sliding window approach, and then the elements of the embedding vectors are searched and sorted to generate a permutation pattern. Statistical analysis of each permutation pattern Calculate the frequency of occurrence and the permutation pattern Probability distribution, representing:

[0033] ;

[0034] In the formula, For the k-th intrinsic mode function component The probability of a certain permutation pattern occurring; For the k-th intrinsic mode function component, the th The number of times a certain permutation pattern appears; It is the length of the time series; For the embedded dimension; For time delay;

[0035] The permutation entropy of the intrinsic mode function components represents:

[0036] ;

[0037] In the formula, For the first The permutation entropy of the intrinsic mode function components; For embedding dimension The total number of permutation patterns; It is the natural logarithm function;

[0038] A combined screening strategy based on correlation coefficient and permutation entropy algorithms ,express:

[0039] ;

[0040] In the formula, For the first Screening strategy for intrinsic mode function components; The threshold for the correlation coefficient; The threshold for permutation entropy;

[0041] when Below or Higher than , indicating the first Each intrinsic mode function component There is too much noise, so the first one is discarded. Each intrinsic mode function component ;

[0042] when Higher than or equal to and Higher than , indicating the first Each intrinsic mode function component Noise mixing, response , After wavelet thresholding denoising, the first denoising step is retained. Each intrinsic mode function component ;

[0043] when Higher than or equal to and Less than or equal to Keep the first one directly Each intrinsic mode function component This is the initial screening and preprocessing of sound signals.

[0044] Furthermore, the specific process of obtaining the noise-reduced audio signal is as follows:

[0045] The wavelet transform is performed on the pre-processed audio signal to obtain the wavelet decomposition coefficients, which are represented as follows:

[0046] ;

[0047] In the formula, For the first Layer wavelet decomposition coefficients; For the first Wavelet decomposition coefficients corresponding to the effective signal of the layer; For the first Wavelet decomposition coefficients corresponding to layer noise signals;

[0048] Set a critical threshold For the Layer wavelet decomposition coefficients Processing yields the first... estimated values ​​of wavelet decomposition coefficients By minimizing differences Obtain the best estimate; use the best estimate to perform wavelet reconstruction, and the obtained estimated signal is the signal after wavelet denoising;

[0049] get ,express:

[0050] ;

[0051] In the formula, For the processed first Layer wavelet decomposition coefficients; It is a symbolic function; The wavelet decomposition coefficients of the j-th level The upper limit of the absolute value; Critical threshold The lower limit; It is a quadratic interpolation polynomial;

[0052] Quadratic interpolation polynomial ,express:

[0053] ;

[0054] based on The denoised audio signal is obtained by reconstructing the signal after wavelet denoising from the effective signal, and is represented as follows:

[0055] ;

[0056] In the formula, The audio signal after noise reduction at time t; The total number of signals after wavelet denoising; For the first The wavelet-denoised signal at time t of layer ; The total number of valid signals; For the first Wavelet decomposition coefficients corresponding to the effective signal at time t in layer t.

[0057] Furthermore, the Deep Focusing U-Net model consists of an encoder and a decoder;

[0058] The encoder consists of a first downsampling block, a second downsampling block, a third downsampling block, a fourth downsampling block, and a fifth downsampling block;

[0059] The decoder consists of a first upsampling block, a second upsampling block, a third upsampling block, a fourth upsampling block, and a fifth upsampling block;

[0060] The encoder processes the data sequentially through the first downsampling block, the second downsampling block, the third downsampling block, the fourth downsampling block, and the fifth downsampling block to obtain the first-layer downsampling feature map, the second-layer downsampling feature map, the third-layer downsampling feature map, the fourth-layer downsampling feature map, and the fifth-layer downsampling feature map, respectively. This is the downsampled feature map of the i-th layer; ;

[0061] The feature map of the i-th layer is downsampled using linear interpolation. To process, meaning:

[0062] ;

[0063] In the formula, For the first Feature map of the layer after linear interpolation; This is a convolution operation; This is a downsampling operation; For the first Layer downsampling feature map; After processing by convolution operation, the first Layer downsampling feature map;

[0064] in, include , , , , ; This is the feature map of the first layer after linear interpolation; This is the feature map of the second layer after linear interpolation; This is the feature map of the third layer after linear interpolation; This is the feature map of the fourth layer after linear interpolation. This is the feature map of the fifth layer after linear interpolation.

[0065] Furthermore, the specific process of obtaining the upsampled feature maps from the first to the fifth layer is as follows:

[0066] The feature maps of the first, second, third, fourth, and fifth layers after linear interpolation are fused into the first upsampling block via a first skip connection to obtain the first layer upsampling feature map.

[0067] The feature maps of the first, second, third, and fourth layers after linear interpolation are input into the second upsampling block through the first jump connection.

[0068] The first-layer upsampled feature map is input into the second upsampled block through upsampling; the second upsampled block fuses the first-layer upsampled feature map, the second-layer upsampled feature map, the third-layer upsampled feature map, and the fourth-layer upsampled feature map with the first-layer upsampled feature map to obtain the second-layer upsampled feature map.

[0069] The feature maps of the first, second, and third layers after linear interpolation are connected to the third upsampling block via the first jump connection.

[0070] The second-layer upsampled feature map is input into the third upsampled block through upsampling; the third upsampled block fuses the first-layer feature map after linear interpolation, the second-layer feature map after linear interpolation, and the third-layer feature map after linear interpolation with the second-layer upsampled feature map to obtain the third-layer upsampled feature map;

[0071] The feature maps of the first layer after linear interpolation and the feature maps of the second layer after linear interpolation are input into the fourth upsampling block through the first jump connection;

[0072] The third-layer upsampled feature map is input into the fourth upsampled block through upsampling; the fourth upsampled block fuses the first-layer feature map after linear interpolation, the second-layer feature map after linear interpolation, and the third-layer upsampled feature map to obtain the fourth-layer upsampled feature map;

[0073] The feature map of the first layer after linear interpolation is input into the fifth upsampling block through the first skip connection; the upsampled feature map of the fourth layer is input into the fifth upsampling block through upsampling; the feature map of the first layer after linear interpolation and the upsampled feature map of the fourth layer are fused through the fifth upsampling block to obtain the upsampled feature map of the fifth layer.

[0074] Furthermore, the specific process of obtaining the target source signal is as follows:

[0075] The first-layer upsampled feature map is input into the second upsampled block through upsampling. The first-layer upsampled feature map is input into the second upsampled block through the second skip connection. The second upsampled block fuses the first-layer upsampled feature map input through upsampling, the first-layer upsampled feature map input through the second skip connection, and the second upsampled feature map to obtain the second fused upsampled map.

[0076] The second fused upsampled image is input into the third upsampled block through upsampling. The first layer upsampled feature map and the second fused upsampled image are input into the third upsampled block through the second skip connection. The third upsampled block fuses the second fused upsampled image, the first layer upsampled feature map input through the second skip connection, the second fused upsampled image and the third layer upsampled feature map to obtain the third fused upsampled image.

[0077] The third fused upsampled image is input to the fourth upsampled block through upsampling. The first layer upsampled feature map, the second fused upsampled image, and the third fused upsampled image are input to the fourth upsampled block through the second jump connection. The fourth upsampled block fuses the third fused upsampled image, the first layer upsampled feature map, the second fused upsampled image, the third fused upsampled image, and the fourth layer upsampled feature map input through the second jump connection to obtain the fourth fused upsampled image.

[0078] The fourth fused upsampled image is input to the fifth upsampled block through upsampling. The first layer upsampled feature image, the second fused upsampled image, the third fused upsampled image, and the fourth fused upsampled image are input to the fifth upsampled block through the second jump connection. The fifth upsampled block fuses the fourth fused upsampled image, the first layer upsampled feature image, the second fused upsampled image, the third fused upsampled image, the fourth fused upsampled image, and the fifth layer upsampled feature image to obtain the fifth fused upsampled image. The fifth fused upsampled image is then upsampled and used as the final output, which is the target source signal.

[0079] Furthermore, the specific process of step S5 is as follows:

[0080] Construct the target voiceprint library;

[0081] Registering an audio file in the target voiceprint library is equivalent to registering the target sound source signal. After the registered target sound source signal is framed, pre-emphasized, and windowed, it is converted into a frequency domain signal by short-time Fourier transform. The voiceprint features of the frequency domain signal are extracted and then converted into target audio features by the trained ResNet34 deep learning model.

[0082] After preprocessing, short-time Fourier transform, and voiceprint feature extraction, the target source signal is converted into audio features by a trained ResNet34 deep learning model. The audio features are compared with the target audio features one by one to obtain similarity results. The similarity results are judged by a set judgment threshold. If the similarity result is higher than the judgment threshold, the target source signal is retained and input into the target voiceprint database for recognition. The output transformer determines the category of the final target sound source. If the similarity result is lower than the judgment threshold, it is a non-target source signal and is then removed from the target source signal.

[0083] Furthermore, the training phase of the ResNet34 deep learning model:

[0084] The registered target sound source signals are preprocessed in batches to extract feature vectors. Batch preprocessing includes framing, pre-emphasis, windowing, short-time Fourier transform, and Mel frequency cepstral coefficients. The feature vectors are then fed into the ResNet34 deep learning model for training to obtain the trained ResNet34 deep learning model.

[0085] Compared with existing technologies, the present invention has the following advantages:

[0086] (1) This invention filters the intrinsic mode function components by combining correlation coefficient and permutation entropy algorithm, and performs noise reduction by wavelet thresholding method, which effectively improves the purity of the signal and significantly improves the accuracy of voiceprint recognition. In particular, it shows strong robustness in high noise environment and effectively avoids misjudgment caused by noise interference.

[0087] (2) This invention introduces the Deep Focusing U-Net model; through the underdetermined blind source separation technology in the Deep Focusing U-Net model, the extraction accuracy of the target source signal can be significantly improved. The Deep Focusing U-Net model effectively enhances the attention to the upsampled feature map by introducing residual skip connections, thereby improving the continuity and completeness of feature extraction and accurately acquiring the target source signal. Attached Figure Description

[0088] Figure 1 This is a flowchart of the method of the present invention.

[0089] Figure 2 This is a flowchart of the Deep Focusing U-Net model of the present invention. Detailed Implementation

[0090] like Figure 1 As shown, the present invention provides a technical solution: a transformer audio signal processing method based on CEEMDAN, comprising:

[0091] Step S1: Collect and process the noisy transformer sound signals to obtain a mixed signal matrix to be processed;

[0092] Step S2: Preprocess the mixed signal matrix to obtain the preprocessed sound signal;

[0093] Step S3: The preprocessed audio signal is processed using the correlation coefficient and permutation entropy algorithm to obtain a preliminary preprocessed audio signal. The preliminary preprocessed audio signal is then denoised based on the CEEMDAN model to obtain a denoised audio signal.

[0094] Step S4: Input the denoised audio signal into the Deep Focusing U-Net model and process it through underdetermined blind source separation to obtain the target source signal;

[0095] Deep Focusing U-Net model processing procedure:

[0096] The denoised audio signal is sequentially input into five downsampling blocks for processing, resulting in downsampling feature maps from the first to the fifth layer. These downsampling feature maps are then linearly interpolated to obtain linearly interpolated feature maps from the first to the fifth layer. These linearly interpolated feature maps are then input into five upsampling blocks via a first skip connection. Each upsampling block fuses the linearly interpolated feature maps input from the first skip connection through an upsampling operation, resulting in upsampling feature maps from the first to the fifth layer. Finally, these upsampling feature maps are added one by one via a second skip connection and an upsampling operation to obtain the target source signal.

[0097] Step S5: Construct a target acoustic signature library for transformers and register the target acoustic source signals and train a ResNet34 deep learning model. The trained ResNet34 deep learning model processes the target acoustic source signals and compares the processing results with the target acoustic source signals. If the comparison result is higher than the set judgment threshold, the target acoustic source signal is retained. The retained target acoustic source signal is input into the target acoustic signature library for transformers for identification, and the final target acoustic source category of the transformer is output.

[0098] The specific process of obtaining the mixed signal matrix in step S1 is as follows:

[0099] Multiple microphones are placed around the transformer to collect the transformer's operating sound and environmental noise, which is the sound signal of the noisy transformer.

[0100] The noisy transformer sound signal is weighted and mixed according to the target signal-to-noise ratio (e.g., 10 dB) to generate a noisy mixed signal;

[0101] The noisy mixed signals are aligned by time and combined into a mixed signal matrix to be processed. The rows of the mixed signal matrix to be processed represent time samples, and the columns of the mixed signal matrix to be processed represent different microphone channels.

[0102] Among them, the mixed signal matrix to be processed is preprocessed, including sampling, quantization, pre-emphasis, and frame windowing;

[0103] Sampling involves converting the analog signals of the mixed signal matrix into digital signals according to the Nelster criterion.

[0104] Quantization involves using the numerical value of a digital signal to replace the amplitude of a noisy transformer-controlled sound signal. Specifically:

[0105] The amplitude range of the entire noisy transformer sound signal is normalized. After normalization, it is divided into quantization intervals with fixed intervals. The amplitude within the quantization interval is replaced by the value of the digital signal to obtain the normalized and quantized digital signal. The number of quantization intervals is 8 bits or 16 bits, with 16 bits being more commonly used. For human hearing, the difference in amplitude intervals of 1 / 65536 of the noisy transformer sound signal is imperceptible.

[0106] The pre-emphasis is based on the normalized and quantized digital signal. A high-pass filter is used to enhance the high-frequency components of the noisy transformer sound signal to obtain the enhanced high-frequency components.

[0107] Pre-emphasis is a high-pass filtering operation. Since the high-frequency components of the noisy transformer sound signal are lost much more than the low-frequency components during propagation, pre-emphasis enhances the high-frequency components of the noisy transformer sound signal to compensate for the unequal loss of high-frequency and low-frequency components during propagation. This makes the important information in the noisy transformer sound signal more prominent, which is helpful for subsequent feature extraction and analysis.

[0108] Framing and windowing is a basic step in processing noisy transformer audio signals. Since transformer audio signals are usually non-stationary (i.e., their statistical characteristics change over time), but noisy transformer audio signals are short-term stationary, meaning they can be considered stationary for a short period of time. Therefore, in order to perform effective spectrum analysis of noisy transformer audio signals, the noisy transformer audio signals are usually divided into several short time periods (frames) to reduce boundary effects.

[0109] Based on enhancing high-frequency components, when framing the noise signal of the sound signal of the noisy transformer, there is generally overlap between the two frames in order to ensure the continuity between the signals of adjacent frames.

[0110] Taking human voice recognition and speech signals as examples, human voice recognition typically uses 20-30ms as a frame. If the frame length is too short, the feature vector of the noise-transformed sound signal will have poor representativeness; if the frame length is too long, the speech signal will change too much, affecting the accuracy of the feature vector.

[0111] Compared to speech signals, the sound signal of a noisy transformer has higher stability and fewer abrupt changes. Therefore, the frame length of the sound signal of a noisy transformer can be appropriately increased to obtain higher accuracy. However, an excessively long frame length will seriously affect the recognition speed. Here, the frame length of the sound signal of a noisy transformer is set to 500ms. In addition, considering that the sound signal of a noisy transformer is relatively stable under the same operating conditions and that the continuity between two frames is good, an overlap rate of 40% is adopted.

[0112] The overlap rate is set to 40% because the mixed signal matrix to be processed needs to be subjected to Discrete Fourier Transform in subsequent operations. However, direct Discrete Fourier Transform will produce large distortion. Therefore, the sound signal with noise transformer is combined with Hamming window function to perform point-by-point operation to obtain the windowed processing result. Based on the windowed processing result, Discrete Fourier Transform is performed on the mixed signal matrix to be processed to obtain the preprocessed sound signal.

[0113] The specific process for obtaining the preliminary screened and preprocessed audio signals is as follows:

[0114] Gaussian white noise is added to the preprocessed audio signal to construct the signal to be separated. Empirical mode decomposition (EMD) is performed on the constructed signal to be separated to obtain multiple intrinsic mode function (IMF) components. The average of the multiple IMF components is calculated to obtain the first IMF component. The first IMF component is subtracted from the preprocessed audio signal to obtain the residual component at time t. ;

[0115] right Update until Stop when the conditions are met, specifically:

[0116] right Add Gaussian white noise again and perform empirical mode decomposition to update. From the update Extract three consecutive points as follows , , When satisfied > and > At that time, It is a local maximum point; when it satisfies < and < At that time, It is a local minimum point; when Greater than or less than both sides and hour, It is identified as a maximum or minimum point; when updating If there is only one maximum or minimum point, the condition is met and the process stops. Consider it as a new intrinsic mode function (IMF); The signal value of the residual component at time 1; The signal value of the residual component at the second time step; The signal value of the residual component at time 3;

[0117] Power spectrum analysis of the IMF using short-time Fourier transform yields the k-th intrinsic mode function component. The dominant frequency fmax is the kth eigenmode function component. The main frequency fmax and the set cutoff frequency In comparison, when it is less than the cutoff frequency The decomposition ends, and the k-th intrinsic mode function component is output. Conversely, the decomposition continues;

[0118] In the process of empirical mode decomposition and The constraints are expressed as follows:

[0119] (1);

[0120] (2);

[0121] In the formula, Preprocess the audio signal at time t; Let K be the intrinsic mode function component of the i-th layer; K represents the total number of intrinsic mode function components.

[0122] Among them, the correlation coefficient and permutation entropy algorithm are combined to... The screening process was conducted to remove signals with low correlation and high signal complexity (potentially containing significant noise). Remove the pre-processed audio signal and obtain a preliminary filtered signal.

[0123] Calculate the correlation coefficient separately. Preliminary noise reduction is performed based on the correlation coefficient with the corresponding preprocessed audio signal; when A smaller correlation coefficient indicates that... Factors with a high signal-to-noise ratio should be discarded, while those with a low signal-to-noise ratio should be retained. Taking all factors into consideration, factors with a correlation coefficient less than 0.3 should be selected. give up; The formula for calculating the correlation coefficient with the corresponding preprocessed audio signal is shown in equation (3):

[0124] (3);

[0125] In the formula, For the k-th intrinsic mode function component The degree of linear correlation between y(t) and y(t); Let be the mean of y(t); The value of the k-th eigenmode function component at time t; It is the average value of the k-th intrinsic mode function component;

[0126] The calculation steps of the permutation entropy algorithm PE are as follows:

[0127] In the permutation entropy algorithm, the k-th eigenmode function component is traversed. The time series data is processed by extracting embedding vectors using a sliding window approach, and then the elements of the embedding vectors are searched and sorted to generate a permutation pattern. Statistical analysis of each permutation pattern Calculate the frequency of occurrence and the permutation pattern Probability distribution, representing:

[0128] (4);

[0129] In the formula, For the k-th intrinsic mode function component The probability of a certain permutation pattern occurring; For the k-th intrinsic mode function component, the th The number of times a permutation pattern appears; M is a counting function used to count the number of times a permutation pattern appears; It is the length of the time series; For the embedded dimension; For time delay;

[0130] The permutation entropy of the intrinsic mode function components represents:

[0131] (5);

[0132] In the formula, For the first The permutation entropy of the intrinsic mode function components; For embedding dimension The total number of permutation patterns; It is the natural logarithm function;

[0133] A screening algorithm coupled with correlation coefficient and permutation entropy algorithm is used to screen the k-th intrinsic mode function components. By performing screening, the correlation coefficient can effectively identify the k-th intrinsic mode function component with a high signal-to-noise ratio. The permutation entropy algorithm can quantify the complexity of a signal; the combination of the two can more accurately identify and remove noise components, and can effectively improve the signal-to-noise ratio while maintaining the characteristics of the pre-processed sound signal.

[0134] A combined screening strategy based on correlation coefficient and permutation entropy algorithms ,express:

[0135] (6);

[0136] In the formula, For the first Screening strategy for intrinsic mode function components; The threshold for the correlation coefficient; The threshold for permutation entropy;

[0137] when Below or Higher than , indicating the first Each intrinsic mode function component The first one was deemed to have too much noise and was therefore discarded. Each intrinsic mode function component ;

[0138] when Higher than or equal to and Higher than , indicating the first Each intrinsic mode function component Noise mixing, response , After wavelet thresholding denoising, the first denoising step is retained. Each intrinsic mode function component ;

[0139] when Higher than or equal to and Less than or equal to Keep the first one directly Each intrinsic mode function component This is the initial screening and preprocessing of sound signals.

[0140] The specific process for obtaining the noise-reduced audio signal is as follows:

[0141] Wavelet thresholding has a good effect on noise reduction of pre-processed audio signals. After wavelet transform, the noise component of the pre-processed audio signal is mainly concentrated in the high-frequency component. The noise is evenly distributed on a large number of wavelet coefficients with small amplitude, while the effective signal is distributed on a small number of wavelet coefficients with large amplitude. Therefore, wavelet transform can be used to separate the effective signal and noise signal of the pre-processed audio signal.

[0142] Wavelet thresholding is used to perform wavelet transform on the pre-processed audio signal to obtain wavelet decomposition coefficients. The wavelet decomposition coefficients consist of two parts: one part represents the wavelet decomposition coefficients corresponding to the effective signal, and the other part represents the wavelet decomposition coefficients corresponding to the noise signal. This is represented as follows:

[0143] (7);

[0144] In the formula, For the first Layer wavelet decomposition coefficients; For the first Wavelet decomposition coefficients corresponding to the effective signal of the layer; For the first Wavelet decomposition coefficients corresponding to layer noise signals;

[0145] Set a critical threshold For the Layer wavelet decomposition coefficients Processing yields the first... estimated values ​​of wavelet decomposition coefficients By minimizing differences Obtain the best estimate; use the best estimate to perform wavelet reconstruction, and the obtained estimated signal is the signal after wavelet denoising;

[0146] Critical threshold The formula for the lower limit T is as follows:

[0147] (8);

[0148] In the formula, For noise variance; The signal sampling length;

[0149] The existing hard thresholding method is a simple and direct denoising method. When the wavelet decomposition coefficients of the j-th level are... The absolute value is less than the hard thresholding method, so it is directly set to zero. When the wavelet decomposition coefficient of the j-th layer is less than the absolute value, it is set to zero. The absolute value remains unchanged when it is greater than that of the hard threshold method; the hard threshold method is simple and effective, suitable for preliminary screening of pre-processed audio signals where noise is relatively obvious; however, the hard threshold method has a significant drawback: due to the fact that the absolute value of the first digit remains unchanged when it is greater than that of the second digit, the hard threshold method remains unchanged when it is greater than that of the third digit, the hard threshold method remains unchanged when it is greater than that of the third digit, the hard threshold method remains unchanged when it is greater than that of the third digit, the hard threshold method remains unchanged when it is greater than that of the third digit, the hard threshold method remains unchanged when it is greater than that of the third digit, the hard threshold method remains unchanged when it is greater than that of the third digit, the hard threshold method remains unchanged when it is greater than that of the third digit, the hard threshold method remains unchanged when it is greater than that of the fourth digit, the hard threshold method remains unchanged when it is greater than that of the third digit, the hard threshold method remains unchanged when it is greater than that of the third digit, the hard threshold method remains unchanged when it is greater than that of the fourth digit, the hard threshold method remains unchanged when it is greater than that of the third digit, the hard threshold method remains unchanged when it is greater than that of the fourth digit, the hard threshold method remains unchanged when it is greater than that of the fifth ... fourth Layer wavelet decomposition coefficients The processing of is relatively rigid, which may introduce artifacts (i.e., sudden changes or discontinuities) during signal reconstruction;

[0150] In the existing soft threshold method, when the absolute value of the wavelet decomposition coefficient at the j-th layer is less than the soft threshold, the wavelet decomposition coefficient at the j-th layer is also set to zero. When the absolute value of the wavelet decomposition coefficient at the j-th layer is greater than the soft threshold, the absolute value of the wavelet decomposition coefficient at the j-th layer is retained after subtracting the soft threshold; this method makes the preliminary screening and preprocessing of the sound signal smoother; the soft threshold method can effectively reduce the artifacts during the reconstruction of the preliminary screening and preprocessing of the sound signal through smoothing. Since the wavelet decomposition coefficient at the j-th layer is reduced during the denoising process, the detail components of the preliminary screening and preprocessing of the sound signal may be weakened;

[0151] This paper uses the quadratic polynomial interpolation method to overcome the shortcomings of the hard threshold method and the soft threshold method;

[0152] Obtain , which represents:

[0153] (9);

[0154] In the formula, is the wavelet decomposition coefficient at the -th layer after processing; is the sign function; is the upper threshold of the absolute value of the wavelet decomposition coefficient at the j-th layer ; is the quadratic interpolation polynomial;

[0155] The quadratic interpolation polynomial , which represents:

[0156] (10);

[0157] Among them, when the absolute value of the wavelet decomposition coefficient at the j-th layer is between T and q (i.e., T⩽∣ ∣<q), the quadratic polynomial interpolation method is operated to make the between T and q smoothly transition, avoiding the discontinuity problem caused by the sudden change processing in the hard threshold method, thus overcoming the discontinuity problem of the hard threshold method; and when , the wavelet decomposition coefficient at the j-th layer after processing directly takes , without the operation of subtracting the soft threshold as in the existing soft threshold method, avoiding the fixed offset brought by the soft threshold method and making up for the deficiency of the soft threshold method;

[0158] based on The denoised audio signal is obtained by reconstructing the signal after wavelet denoising from the effective signal, and is represented as follows:

[0159] (11);

[0160] In the formula, The audio signal after noise reduction at time t; The total number of signals after wavelet denoising; For the first The wavelet-denoised signal at time t of layer ; The total number of valid signals; For the first Wavelet decomposition coefficients corresponding to the effective signal at time t in layer t.

[0161] The Deep Focusing U-Net model consists of an encoder and a decoder.

[0162] The encoder consists of a first downsampling block, a second downsampling block, a third downsampling block, a fourth downsampling block, and a fifth downsampling block;

[0163] The decoder consists of a first upsampling block, a second upsampling block, a third upsampling block, a fourth upsampling block, and a fifth upsampling block;

[0164] like Figure 2 As shown; the denoised audio signal is processed sequentially through the first downsampling block, the second downsampling block, the third downsampling block, the fourth downsampling block and the fifth downsampling block in the encoder to obtain the first layer downsampling feature map, the second layer downsampling feature map, the third layer downsampling feature map, the fourth layer downsampling feature map and the fifth layer downsampling feature map respectively; This is the downsampled feature map of the i-th layer; ,express:

[0165] (12);

[0166] In the formula, This is a downsampling operation; This is the downsampled feature map of the (i-1)th layer;

[0167] The feature map of the i-th layer is downsampled using linear interpolation. To process, meaning:

[0168] (13);

[0169] In the formula, For the first Feature map of the layer after linear interpolation; This is a convolution operation; For the first Layer downsampling feature map; After processing by convolution operation, the first Layer downsampling feature map;

[0170] in, include , , , , ; This is the feature map of the first layer after linear interpolation; This is the feature map of the second layer after linear interpolation; This is the feature map of the third layer after linear interpolation; This is the feature map of the fourth layer after linear interpolation. This is the feature map of the fifth layer after linear interpolation.

[0171] The specific process for obtaining the upsampled feature maps from the first to the fifth layer is as follows:

[0172] The feature maps of the first, second, third, fourth, and fifth layers after linear interpolation are fused into the first upsampling block via a first skip connection to obtain the first layer upsampling feature map.

[0173] The feature maps of the first, second, third, and fourth layers after linear interpolation are input into the second upsampling block through the first jump connection.

[0174] The first-layer upsampled feature map is input into the second upsampled block through upsampling; the second upsampled block fuses the first-layer upsampled feature map, the second-layer upsampled feature map, the third-layer upsampled feature map, and the fourth-layer upsampled feature map with the first-layer upsampled feature map to obtain the second-layer upsampled feature map.

[0175] The feature maps of the first, second, and third layers after linear interpolation are connected to the third upsampling block via the first jump connection.

[0176] The second-layer upsampled feature map is input into the third upsampled block through upsampling; the third upsampled block fuses the first-layer feature map after linear interpolation, the second-layer feature map after linear interpolation, and the third-layer feature map after linear interpolation with the second-layer upsampled feature map to obtain the third-layer upsampled feature map;

[0177] The feature maps of the first layer after linear interpolation and the feature maps of the second layer after linear interpolation are input into the fourth upsampling block through the first jump connection;

[0178] The third-layer upsampled feature map is input into the fourth upsampled block through upsampling; the fourth upsampled block fuses the first-layer feature map after linear interpolation, the second-layer feature map after linear interpolation, and the third-layer upsampled feature map to obtain the fourth-layer upsampled feature map;

[0179] The feature map of the first layer after linear interpolation is input into the fifth upsampling block through the first skip connection; the upsampled feature map of the fourth layer is input into the fifth upsampling block through upsampling; the feature map of the first layer after linear interpolation and the upsampled feature map of the fourth layer are fused through the fifth upsampling block to obtain the upsampled feature map of the fifth layer.

[0180] The specific process of obtaining the target source signal is as follows:

[0181] The first-layer upsampled feature map is input into the second upsampled block through upsampling. The first-layer upsampled feature map is input into the second upsampled block through the second skip connection. The second upsampled block fuses the first-layer upsampled feature map input through upsampling, the first-layer upsampled feature map input through the second skip connection, and the second upsampled feature map to obtain the second fused upsampled map.

[0182] The second fused upsampled image is input into the third upsampled block through upsampling. The first layer upsampled feature map and the second fused upsampled image are input into the third upsampled block through the second skip connection. The third upsampled block fuses the second fused upsampled image, the first layer upsampled feature map input through the second skip connection, the second fused upsampled image and the third layer upsampled feature map to obtain the third fused upsampled image.

[0183] The third fused upsampled image is input to the fourth upsampled block through upsampling. The first layer upsampled feature map, the second fused upsampled image, and the third fused upsampled image are input to the fourth upsampled block through the second jump connection. The fourth upsampled block fuses the third fused upsampled image, the first layer upsampled feature map, the second fused upsampled image, the third fused upsampled image, and the fourth layer upsampled feature map input through the second jump connection to obtain the fourth fused upsampled image.

[0184] The fourth fused upsampled image is input to the fifth upsampled block through upsampling. The first layer upsampled feature image, the second fused upsampled image, the third fused upsampled image, and the fourth fused upsampled image are input to the fifth upsampled block through the second jump connection. The fifth upsampled block fuses the fourth fused upsampled image, the first layer upsampled feature image, the second fused upsampled image, the third fused upsampled image, and the fourth fused upsampled image with the fifth layer upsampled feature image to obtain the fifth fused upsampled image. The fifth fused upsampled image is then upsampled and used as the final output, which is the target source signal.

[0185] To further enhance the continuity and completeness of feature extraction, a second skip connection is introduced. This second skip connection, through a unique residual skip connection strategy, strengthens the weights of the fused upsampled map, thereby optimizing the Deep Focusing U-Net model's focus on the fused upsampled map during training. The second skip connection not only enhances the propagation capability but also allows for effective fusion of the first-layer upsampled feature map of the first upsampled block with the fifth-layer upsampled feature map of the fifth upsampled block. In fact, through continuous skip connection operations, the first-layer upsampled feature map of the first upsampled block can be connected to all layers above the second-layer to fifth-layer upsampled blocks. This continuous inter-layer connection ensures that the upsampled feature map is fully propagated and integrated throughout the network. This multi-layered connection structure helps strengthen the deep learning capabilities of the Deep Focusing U-Net model and enhances its ability to extract complex features.

[0186] As the depth of the Deep Focusing U-Net model increases, the number of such consecutive connections also increases accordingly. This means that the deeper the layers, the greater the weight of the upsampled feature maps. This design choice enhances the DeepFocusing U-Net model's ability to capture upsampled feature maps. By increasing the weight of the upsampled feature maps, the DeepFocusing U-Net model can pay more attention to the representation of upsampled feature maps during training, thereby improving its ability to understand and analyze audio signals after denoising in complex conditions.

[0187] Using underdetermined blind source separation based on the Deep Focusing model has significant advantages for processing denoised audio signals:

[0188] (1) Enhance the Deep Focusing U-Net model’s ability to capture upsampled feature maps: The Deep Focusing U-Net model increases the weight of upsampled feature maps through residual skip connections, making the Deep Focusing U-Net model pay more attention to upsampled feature maps during training.

[0189] (2) Improve the generalization ability of Deep Focusing U-Net model: The second skip connection not only enables the first layer upsampled feature map of the first upsampled block to be propagated and strengthened in the deep layer, but also enhances the multi-scale fusion capability of the features; enabling Deep Focusing U-Net model to extract and integrate information at different scales, which is very beneficial for the features of different frequency bands in the denoised audio signal; (3) Through the continuous connection between layers and the expansion of the upsampled feature map, the audio signal after transformer denoising often contains a variety of complex spectral components and noise interference; through the design of the second skip connection, Deep Focusing U-Net model can effectively adjust the size of the upsampled feature map of different layers, so that the first layer upsampled feature map of the first upsampled block can be fully utilized in the second upsampled block to the fifth upsampled block; this enhanced Deep Focusing U-Net model skip connection ensures that Deep Focusing U-Net model can fully utilize the various features in the audio signal after transformer denoising and obtain the target source signal more accurately.

[0190] The specific process of step S5 is as follows:

[0191] Through the power transformer acoustic signature simulation laboratory, the pure and noise-free frequency of the transformer operation on site is simulated. It is divided into normal operation audio, fault operation audio, and interference noise. Based on the division of normal operation audio, fault operation audio, and interference noise, a target acoustic signature library is constructed.

[0192] Normal operating audio covers the audio corresponding to the following voltage levels, namely, normal operating audio at voltage levels of 110kV, 220kV, 500kV, and 800kV. Fault operating audio includes the following types: short circuit impact, loose clamps, partial discharge, abnormal noise from the cooler, etc. Interference noise includes: human voices, birdsong, whistles, thunder, wind and rain, busbar noise, cooling noise, etc.

[0193] Registering an audio file in the target voiceprint library is equivalent to registering the target sound source signal. After the registered target sound source signal is framed, pre-emphasized, and windowed, it is converted into a frequency domain signal by short-time Fourier transform. The voiceprint features of the frequency domain signal are extracted and then converted into target audio features by the trained ResNet34 deep learning model.

[0194] After preprocessing, short-time Fourier transform, and voiceprint feature extraction, the target source signal is converted into audio features by a trained ResNet34 deep learning model. The audio features are compared with the target audio features one by one to obtain similarity results. The similarity results are judged by a set judgment threshold. If the similarity result is higher than the judgment threshold, the target source signal is retained and input into the target voiceprint database for recognition. The output transformer determines the category of the final target sound source. If the similarity result is lower than the judgment threshold, it is a non-target source signal and is then removed from the target source signal.

[0195] The ResNet34 deep learning model training phase includes:

[0196] The registered target sound source signals are preprocessed in batches to extract feature vectors. Batch preprocessing includes framing, pre-emphasis, windowing, short-time Fourier transform, and Mel frequency cepstral coefficients. The feature vectors are then fed into the ResNet34 deep learning model for training to obtain the trained ResNet34 deep learning model.

[0197] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A transformer sound signal processing method based on CEEMDAN, characterized in that, The method comprises the following steps: Step S1: collecting and processing the noise-containing transformer sound signal to obtain a mixed to-be-processed signal matrix; Step S2: preprocessing the mixed to-be-processed signal matrix to obtain a preprocessed sound signal; Step S3: processing the preprocessed sound signal by using a correlation coefficient and permutation entropy algorithm to obtain a preliminarily screened preprocessed sound signal, and denoising the preliminarily screened preprocessed sound signal based on a CEEMDAN model to obtain a denoised audio signal; Step S4: inputting the denoised audio signal into a Deep Focusing U-Net model for underdetermined blind source separation processing to obtain a target source signal; The Deep Focusing U-Net model processing process comprises the following steps: The denoised audio signal is sequentially input into five downsampling blocks for processing to obtain first-layer to fifth-layer downsampling feature maps, the first-layer to fifth-layer downsampling feature maps are processed by linear interpolation to obtain first-layer to fifth-layer linear interpolation processed feature maps, the first-layer to fifth-layer linear interpolation processed feature maps are input into five upsampling blocks through first jump connection, each layer of the upsampling blocks fuses the linear interpolation processed feature maps input through the first jump connection through upsampling operation to obtain first-layer to fifth-layer upsampling feature maps, and the first-layer to fifth-layer upsampling feature maps are added one by one through second jump connection and upsampling operation to obtain the target source signal; The Deep Focusing U-Net model comprises an encoder and a decoder; The encoder comprises a first downsampling block, a second downsampling block, a third downsampling block, a fourth downsampling block and a fifth downsampling block; The decoder comprises a first upsampling block, a second upsampling block, a third upsampling block, a fourth upsampling block and a fifth upsampling block; The noise-reduced audio signal is sequentially processed by a first downsampling block, a second downsampling block, a third downsampling block, a fourth downsampling block and a fifth downsampling block in the encoder to obtain a first layer downsampled feature map, a second layer downsampled feature map, a third layer downsampled feature map, a fourth layer downsampled feature map and a fifth layer downsampled feature map, respectively; for the i-th layer downsampled feature map; ; using linear interpolation method on the down-sampled feature map of the i-th layer is processed to indicate: ； In the formula, is the first layer after linear interpolation processing of the feature map; is a convolution operation; is a down-sampling operation; is the first layer down-sampled feature map; is the first layer down-sampled feature map after convolution operation processing; wherein, comprising , , , , ; is a feature map of the first layer after linear interpolation processing; is a feature map of the second layer after linear interpolation processing; is a feature map of the third layer after linear interpolation processing; is a feature map of the fourth layer after linear interpolation processing; is a feature map of the fifth layer after linear interpolation processing; Step S5: constructing a transformer target voiceprint library, registering the target source signal and training a ResNet34 deep learning model, processing the target source signal by using the trained ResNet34 deep learning model, comparing a processing result with the target source signal, retaining the target source signal when a comparison result is higher than a set judgment threshold, inputting the retained target source signal into the transformer target voiceprint library for identification, and outputting a category of the final target source of the transformer.

2. The transformer sound signal processing method based on CEEMDAN according to claim 1, characterized in that: The specific process of obtaining the preprocessed sound signal comprises the following steps: According to the Nyquist criterion, the analog signal of the mixed to-be-processed signal matrix is converted into a digital signal; The numerical value of the digital signal is used to replace the amplitude of the noise-containing transformer sound signal, specifically as follows: The amplitude interval of the entire noise-containing transformer sound signal is normalized, and after normalization, the amplitude interval is divided into fixed interval quantization intervals, the amplitude value of the amplitude in the quantization interval is replaced by the numerical value of the digital signal to obtain a normalized and quantized digital signal; Based on the normalized and quantized digital signal, a high-pass filtering operation is performed to enhance the high-frequency component of the noise-containing transformer sound signal to obtain an enhanced high-frequency component; Based on the enhanced high-frequency component, the noise-containing transformer sound signal is subjected to discrete Fourier transform to obtain the preprocessed sound signal.

3. The transformer sound signal processing method based on CEEMDAN according to claim 2, characterized in that: A specific process for obtaining the preprocessed sound signal for preliminary screening is as follows: The Gaussian white noise is added to the preprocessed sound signal to construct a signal to be separated, the signal to be separated is decomposed by empirical mode decomposition to obtain a plurality of intrinsic mode function components, the plurality of intrinsic mode function components are averaged to obtain a first intrinsic mode function component, and the preprocessed sound signal is subtracted by the first intrinsic mode function component to obtain a residual component at a tth moment ; The update is performed on until the condition is met, in particular: To Add Gaussian white noise again, and carry out empirical mode decomposition operation update , extract three consecutive points from the update , , , respectively ; when > and > , then is a local maximum point; when < and < , then is a local minimum point; when is greater than or less than both sides and , is identified as a maximum point or a minimum point; when the update has only one maximum point or minimum point, the condition is met to stop, and is regarded as a new intrinsic mode function IMF; is the signal value of the residual component at the first time; is the signal value of the residual component at the second time; is the signal value of the residual component at the third time; The power spectrum analysis of the short-time Fourier transform of the IMF obtains the main frequency fmax of the kth IMF component The main frequency fmax of the kth IMF component is compared with the set cutoff frequency When the main frequency fmax of the kth IMF component is less than the cutoff frequency , the decomposition ends and the kth IMF component is output, otherwise, the decomposition continues. The sound signal is screened by combining the correlation coefficient and permutation entropy algorithm to obtain a preliminary screening pretreated sound signal; Calculate the correlation coefficient separately. The correlation coefficient with the corresponding preprocessed audio signal represents: ； wherein is the kth eigenmode function component is the linear correlation degree between y(t) and is the mean value of y(t); is the total number of eigenmode function components; is the value of the kth eigenmode function component at the tth time instant; is the mean value of the kth eigenmode function component; y(t) is the preprocessed sound signal at the tth time instant; In the permutation entropy algorithm, the k-th eigenmode function component is traversed. The time series data is processed by extracting embedding vectors using a sliding window approach, and then the elements of the embedding vectors are searched and sorted to generate a permutation pattern. Statistical analysis of each permutation pattern Calculate the frequency of occurrence and the permutation pattern Probability distribution, representing: ； wherein, is the probability of the kth eigen mode function component of the mth arrangement pattern to appear; is the number of the kth eigen mode function component of the mth arrangement pattern to appear; is the length of the time series; is the embedding dimension; is the time delay;​​ The permutation entropy value of the intrinsic mode function component is represented by: ； wherein is the permutation entropy value of the mth eigemodal function component; is the total number of permutation patterns based on the embedding dimension ; and is the natural logarithm function. Combination screening strategy based on correlation coefficient and permutation entropy algorithm , denotes: ； In the formula, is the threshold value of the correlation coefficient; is the screening strategy of the mth intrinsic modal function component; is the threshold value of the correlation coefficient; is the threshold value of the permutation entropy; When Below Or Above , indicates the noise of the th eigenmode function component , discard the th eigenmode function component ; When above or equal to and above , the first eigenmode function component mixing noise, should deal with , wavelet threshold de-noising processing to retain the first eigenmode function component ; When higher than or equal to and lower than or equal to , the first eigenmode function component is directly reserved , that is, the preliminary screening pretreatment sound signal.

4. The transformer sound signal processing method based on CEEMDAN according to claim 3, characterized in that: A specific process for obtaining the denoised audio signal is as follows: Wavelet decomposition coefficients are obtained by performing wavelet transform on the preprocessed sound signal for preliminary screening, and are represented by: ； wherein is the first layer wavelet decomposition coefficient; is the first layer wavelet decomposition coefficient corresponding to the effective signal; is the first layer wavelet decomposition coefficient corresponding to the noise signal; Setting a critical threshold To the first Layer wavelet decomposition coefficient Processing, get the first Wavelet decomposition coefficient estimate , by minimizing the difference Get the best estimate; using the best estimate for wavelet reconstruction, get the estimated signal is the signal after wavelet denoising; obtained , indicating: ； wherein is the processed jth layer wavelet decomposition coefficient is the processed jth layer wavelet decomposition coefficient is a sign function is the absolute value of the jth layer wavelet decomposition coefficient is an upper threshold of the absolute value of the jth layer wavelet decomposition coefficient is a lower threshold of the critical threshold is a lower threshold of the critical threshold is a quadratic interpolation polynomial Quadratic interpolation polynomial denotes: ； based on The denoised audio signal is obtained by reconstructing the wavelet denoised signal and the effective signal, and is represented as: ； In the formula, is the audio signal after noise reduction at the tth moment; is the total number of signals after wavelet denoising; is the audio signal after wavelet denoising at the tth moment of the lth layer; is the signal after wavelet denoising at the tth moment of the lth layer; is the total number of effective signals; is the audio signal after wavelet denoising at the tth moment of the lth layer; is the wavelet decomposition coefficient corresponding to the effective signal at the tth moment of the lth layer.

5. The CEEMDAN-based transformer sound signal processing method according to claim 4, characterized in that: A specific process for obtaining the upsampled feature maps of the first layer to the fifth layer is as follows: The first layer feature map after linear interpolation processing, the second layer feature map after linear interpolation processing, the third layer feature map after linear interpolation processing, the fourth layer feature map after linear interpolation processing, and the fifth layer feature map after linear interpolation processing are input into the first upsampling block through the first skip connection for fusion to obtain the first layer upsampled feature map; The first layer feature map after linear interpolation processing, the second layer feature map after linear interpolation processing, the third layer feature map after linear interpolation processing, and the fourth layer feature map after linear interpolation processing are input into the second upsampling block through the first skip connection; The first layer upsampled feature map is input into the second upsampling block through upsampling; the first layer feature map after linear interpolation processing, the second layer feature map after linear interpolation processing, the third layer feature map after linear interpolation processing, and the fourth layer feature map after linear interpolation processing are fused with the first layer upsampled feature map through the second upsampling block to obtain the second layer upsampled feature map; The first layer feature map after linear interpolation processing, the second layer feature map after linear interpolation processing, and the third layer feature map after linear interpolation processing are input into the third upsampling block through the first skip connection; The second layer upsampled feature map is input into the third upsampling block through upsampling; the first layer feature map after linear interpolation processing, the second layer feature map after linear interpolation processing, and the third layer feature map after linear interpolation processing are fused with the second layer upsampled feature map through the third upsampling block to obtain the third layer upsampled feature map; The first layer feature map after linear interpolation processing and the second layer feature map after linear interpolation processing are input into the fourth upsampling block through the first skip connection; The third layer upsampled feature map is input into the fourth upsampling block through upsampling; the first layer feature map after linear interpolation processing and the second layer feature map after linear interpolation processing are fused with the third layer upsampled feature map through the fourth upsampling block to obtain the fourth layer upsampled feature map; The first layer feature map after linear interpolation processing is input into the fifth upsampling block through the first skip connection; the fourth layer upsampled feature map is input into the fifth upsampling block through upsampling; the first layer feature map after linear interpolation processing is fused with the fourth layer upsampled feature map through the fifth upsampling block to obtain the fifth layer upsampled feature map.

6. The CEEMDAN-based transformer sound signal processing method according to claim 5, characterized in that: A specific process for obtaining the target source signal is as follows: The first layer up-sampling feature map is input into the second up-sampling block through up-sampling, the first layer up-sampling feature map is input into the second up-sampling block through the second jump connection, the first layer up-sampling feature map input through up-sampling and the first layer up-sampling feature map input through the second jump connection are fused with the second layer up-sampling feature map through the second up-sampling block to obtain a second fused up-sampling map; The second fused up-sampling map is input into the third up-sampling block through up-sampling, the first layer up-sampling feature map and the second fused up-sampling map are input into the third up-sampling block through the second jump connection, the second fused up-sampling map input through up-sampling, the first layer up-sampling feature map input through the second jump connection and the second fused up-sampling map are fused with the third layer up-sampling feature map through the third up-sampling block to obtain a third fused up-sampling map; The third fused up-sampling map is input into the fourth up-sampling block through up-sampling, the first layer up-sampling feature map, the second fused up-sampling map and the third fused up-sampling map are input into the fourth up-sampling block through the second jump connection, the third fused up-sampling map input through up-sampling, the first layer up-sampling feature map input through the second jump connection, the second fused up-sampling map and the third fused up-sampling map are fused with the fourth layer up-sampling feature map through the fourth up-sampling block to obtain a fourth fused up-sampling map; The fourth fused up-sampling map is input into the fifth up-sampling block through up-sampling, the first layer up-sampling feature map, the second fused up-sampling map, the third fused up-sampling map and the fourth fused up-sampling map are input into the fifth up-sampling block through the second jump connection, the fourth fused up-sampling map input through up-sampling, the first layer up-sampling feature map input through the second jump connection, the second fused up-sampling map, the third fused up-sampling map and the fourth fused up-sampling map are fused with the fifth layer up-sampling feature map through the fifth up-sampling block to obtain a fifth fused up-sampling map; the fifth fused up-sampling map is processed through up-sampling as the final output, that is, the target source signal.

7. The CEEMDAN-based transformer sound signal processing method of claim 6, wherein: The specific process of step S5 is: Constructing a target voiceprint library; Registering an audio file in the target voiceprint library, that is, registering a target sound source signal, after frame division, pre-emphasis, windowing, short-time Fourier transform, converting into a frequency domain signal, extracting a voiceprint feature of the frequency domain signal, and converting into a target audio feature through a trained ResNet34 deep learning model; After preprocessing, short-time Fourier transform and voiceprint feature extraction, the target source signal is converted into an audio feature through the trained ResNet34 deep learning model; The audio feature and the target audio feature are compared one by one to obtain a similarity result, and a judgment threshold is set to judge the similarity result; if the similarity result is higher than the judgment threshold, the target source signal is retained and input into the target voiceprint library for identification, and the category of the final target sound source of the transformer is output; if the similarity result is lower than the judgment threshold, the target source signal is a non-target source signal and is eliminated.

8. The CEEMDAN-based transformer sound signal processing method of claim 7, wherein: ResNet34 deep learning model training stage: The registration target sound source signal is batch preprocessed, and a feature vector of the preprocessed registration target sound source signal is extracted, and the batch pre-processing includes framing, pre-emphasis, windowing, short-time Fourier transform, and mel-frequency cepstral coefficient; The feature vector is sent into a ResNet34 deep learning model for training, and a trained ResNet34 deep learning model is obtained.

Images