A Substation Acoustic Signal Partial Discharge Diagnosis Method Based on Improved Matching Pursuit Network

Abstract

This application provides a method for diagnosing partial discharge (PD) of substation acoustic signals based on an improved matched pursuit network, relating to the field of substation acoustic signal PD diagnosis technology. The method includes segmenting and preprocessing the initial time-domain signal to obtain preprocessed time-domain segmented signals; performing time-frequency conversion on both noise-free and noisy analog signals to obtain noisy and noise-free time-frequency matrices; constructing multiple training sample pairs; inputting the noisy time-frequency matrix into an improved matched pursuit network to obtain an estimated feature matrix; calculating the loss value between the estimated feature matrix and the noise-free time-frequency matrix; optimizing the network parameters of the improved matched pursuit network to obtain an optimized improved matched pursuit network; inputting the real signal time-frequency matrix into the optimized improved matched pursuit network to obtain and identify a sparse time-frequency feature matrix, thus obtaining the PD type identification result of the current preprocessed time-domain segmented signal; improving the reliability and generalization ability of the diagnosis.

Description

Technical Field

[0001] This application generally relates to the field of substation acoustic signal partial discharge diagnosis technology, and specifically to a substation acoustic signal partial discharge diagnosis method based on an improved matching pursuit network. Background Technology

[0002] Partial discharge in substations is a major sign of insulation degradation in power equipment. Partial discharge diagnosis using acoustic signals offers advantages such as non-contact measurement and early warning. Existing methods for diagnosing partial discharge using acoustic signals in substations typically include signal acquisition, preprocessing, time-frequency feature extraction, and type identification. This involves using microphone arrays to collect acoustic signals from the surrounding area; performing preliminary noise reduction through filtering; extracting features using time-frequency analysis methods such as short-time Fourier transform or synchronous Fourier squeeze transform; and finally, identifying the type of partial discharge using support vector machines or Softmax classifiers. Some studies have attempted to introduce matching pursuit algorithms for sparse time-frequency representation or use convolutional neural networks to learn features from time-frequency spectra as image input.

[0003] However, existing sparse time-frequency representation methods based on matching pursuit require manually designed dictionaries adapted to different types of partial discharges (PDs). Substation noise environments are complex and PD types are diverse, making it difficult for manual dictionaries to balance generalization ability and feature extraction accuracy, leading to the easy extraction of noise components when the noise is strong. On the other hand, deep learning-based diagnostic methods require a large number of labeled real PD samples for training, but PD events in substations are scarce and labeling costs are high, making it difficult to obtain real labeled data, which leads to difficulties in model training and limited generalization ability. Summary of the Invention

[0004] In view of the above-mentioned defects or deficiencies in the prior art, it is desirable to provide a substation acoustic signal partial discharge diagnosis method based on an improved matching pursuit network to improve the reliability and generalization ability of diagnosis.

[0005] This application provides a method for diagnosing partial discharge of acoustic signals in substations based on an improved matching pursuit network, comprising the following steps: The initial time-domain signal of the substation is acquired, and the initial time-domain signal is segmented and preprocessed to obtain the preprocessed time-domain segmented signal. Based on the physical mechanism of partial discharge acoustic radiation, noiseless analog signals and noisy analog signals corresponding to different partial discharge types are constructed, and time-frequency conversion is performed on the noisy analog signals and the noiseless analog signals respectively to obtain noisy time-frequency matrices and noiseless time-frequency matrices. Using the noisy time-frequency matrix as input samples and the noise-free time-frequency matrix as output labels, multiple training sample pairs are constructed. The noisy time-frequency matrix in the training sample pairs is then input into the improved matching pursuit network to obtain the estimated feature matrix. Based on the structural similarity loss function, the loss value between the estimated feature matrix and the noise-free time-frequency matrix in the training sample pair is calculated, and the network parameters of the improved matching pursuit network are optimized according to the loss value to obtain the optimized improved matching pursuit network. The preprocessed time-domain segmented signal is subjected to time-frequency conversion to obtain the real signal time-frequency matrix, and the real signal time-frequency matrix is ​​input into the optimized improved matching pursuit network to obtain a sparse time-frequency feature matrix. Based on the classifier, the sparse time-frequency feature matrix is ​​identified to obtain the partial discharge type identification result of the preprocessed time-domain segmented signal.

[0006] According to the technical solution provided in this application, the initial time-domain signal is segmented and preprocessed to obtain a preprocessed time-domain segmented signal, specifically including the following steps: The initial time-domain signal is divided according to a preset segment length to obtain multiple time-domain segmented signals; The DC component is removed from the time-domain segmented signal to obtain the removed time-domain segmented signal; The removed time-domain segmented signal is subjected to high-pass filtering to obtain the preprocessed time-domain segmented signal.

[0007] According to the technical solution provided in this application, based on the physical mechanism of partial discharge acoustic radiation, noise-free analog signals corresponding to different types of partial discharge and noisy analog signals corresponding to the noise-free analog signals are constructed, specifically including the following steps: Obtain the damped oscillation model corresponding to the preset partial discharge type, and generate a noise-free analog signal based on the damped oscillation model; The simulated environmental noise signal of the substation is acquired, and the sum of the simulated environmental noise signal and the noise-free simulated signal is used as the noisy simulated signal corresponding to each of the preset partial discharge types.

[0008] According to the technical solution provided in this application, time-frequency conversion is performed on the preprocessed time-domain segmented signal to obtain the real signal time-frequency matrix, specifically including the following steps: The preprocessed time-domain segmented signal is subjected to Hilbert transform to obtain the envelope spectrum signal; Perform a short-time Fourier transform on the envelope spectrum signal to obtain the time-frequency matrix; The range of the frequency band of interest is determined based on the partial discharge resonant frequency, and the elements in the time-frequency matrix that are located within the range of the frequency band of interest are extracted to obtain the real signal time-frequency matrix.

[0009] According to the technical solution provided in this application, based on the structural similarity loss function, the loss value between the estimated feature matrix and the noise-free time-frequency matrix in the training sample pair is calculated, specifically including the following steps: Calculate the mean, variance, and covariance of the estimated feature matrix and the noise-free time-frequency matrix, respectively. Based on the structural similarity loss function, a negative structural similarity index is calculated according to the mean, the variance, and the covariance, and the negative structural similarity index is used as the loss value.

[0010] According to the technical solution provided in this application, the improved matching pursuit network includes multiple stacked matching pursuit units, each of which is used to obtain an estimated feature matrix; Obtaining the sparse time-frequency feature matrix includes the following steps: Batch normalization is performed on the noisy time-frequency matrix in the training sample pair to obtain the batch-normalized noisy time-frequency matrix; The noisy time-frequency matrix after batch normalization is subjected to dilation convolution and activation function processing to obtain the effective feature matrix; The difference between the noisy time-frequency matrix and the effective feature matrix in the training sample pair is used as the updated residual matrix and output to the next matching and tracking unit; Calculate the product of the effective feature matrix extracted by each matching tracking unit and its corresponding weight gain, and sum all the products to obtain the estimated feature matrix.

[0011] According to the technical solution provided in this application, the network parameters of the improved matching pursuit network are optimized based on the loss value to obtain the optimized improved matching pursuit network, specifically including the following steps: Based on the optimizer, the network parameters of the improved matching pursuit network are iteratively updated with the goal of minimizing the loss value; the network parameters include at least the convolution kernel weights, batch normalization parameters, and weight gain of each unit in each matching pursuit unit. When the loss value converges or reaches the preset number of iterations, training stops, and the optimized and improved matching tracking network is obtained.

[0012] According to the technical solution provided in this application, based on a classifier, the sparse time-frequency feature matrix is ​​identified to obtain the partial discharge type identification result of the preprocessed time-domain segmented signal. Specifically, this includes the following steps: Flatten the sparse time-frequency feature matrix into a one-dimensional feature vector; The one-dimensional feature vector is input into the classifier to obtain the probability that the real signal belongs to each preset partial discharge type, and the preset partial discharge type corresponding to the highest probability is taken as the partial discharge type identification result of the current preprocessed time-domain segmented signal.

[0013] According to the technical solution provided in this application, the method further includes the following steps: Acquire test signals under different signal-to-noise ratio conditions; The test signal is converted from time to frequency to obtain the test signal time-frequency matrix; The time-frequency matrix of the test signal is input into the optimized and improved matching pursuit network to obtain the corresponding sparse time-frequency feature matrix. The partial discharge type identification results of each test signal are obtained by identifying the sparse time-frequency feature matrix corresponding to the time-frequency matrix of the test signal based on the classifier. Based on the partial discharge type identification results of each test signal, the identification accuracy under each signal-to-noise ratio condition is determined; the identification accuracy is used to verify the robustness of the optimized improved matching pursuit network under different noise environments.

[0014] As can be seen from the above technical solution, this application has at least the following beneficial effects: This application provides a method for diagnosing partial discharge (PD) signals in substations based on an improved matching pursuit network. The method includes: acquiring the initial time-domain signal of the substation and performing segmented preprocessing on the initial time-domain signal to obtain preprocessed segmented time-domain signals; based on the physical mechanism of PD acoustic radiation, constructing noiseless analog signals corresponding to different PD types and noisy analog signals corresponding to the noiseless analog signals, and performing time-frequency conversion on the noisy analog signals and the noiseless analog signals respectively to obtain noisy time-frequency matrices and noiseless time-frequency matrices; using the noisy time-frequency matrix as input samples and the noiseless time-frequency matrix as output labels to construct multiple training sample pairs, and assigning labels to the noisy time-frequency matrices in the training sample pairs. The noise time-frequency matrix is ​​input into the improved matching pursuit network to obtain the estimated feature matrix. Based on the structural similarity loss function, the loss value between the estimated feature matrix and the noise-free time-frequency matrix in the training sample pair is calculated. Based on the loss value, the network parameters of the improved matching pursuit network are optimized to obtain the optimized improved matching pursuit network. The preprocessed time-domain segmented signal is subjected to time-frequency conversion to obtain the real signal time-frequency matrix. The real signal time-frequency matrix is ​​input into the optimized improved matching pursuit network to obtain the sparse time-frequency feature matrix. Based on the classifier, the sparse time-frequency feature matrix is ​​identified to obtain the partial discharge type identification result of the current preprocessed time-domain segmented signal.

[0015] This application constructs noise-free and noisy simulated signals for different partial discharge (PD) types based on the physical mechanism of PD radiation. It does not rely on actual PD labeled data, and the simulated signals closely match the radiation characteristics of real PD, providing high-quality, highly adaptable training samples for the improved matching pursuit network. This effectively enhances the model's generalization ability, enabling stable operation under complex substation conditions with varying noise intensities and PD types. Secondly, a structural similarity loss function is used to calculate the loss between the estimated feature matrix and the noise-free time-frequency matrix output by the network. This loss function accurately matches the spatial distribution and amplitude characteristics of the time-frequency spectrum, providing a higher fit to the real PD characteristics compared to traditional mean square error loss functions. Combined with an optimizer for iterative optimization of the network parameters, the consistency between the sparse time-frequency feature matrix output by the optimized improved matching pursuit network and the real PD feature matrix is ​​ensured, improving the accuracy of feature extraction. Furthermore, an end-to-end complete processing flow was constructed, from initial time-domain signal acquisition and segmented preprocessing to analog signal construction and time-frequency conversion, and then to improved matching pursuit network feature extraction and classifier type identification. The parameters and processing logic of each link are coordinated, which not only simplifies the operation process of substation partial discharge diagnosis and improves the diagnostic efficiency, but also makes the feature extraction process strongly correlated with the physical mechanism of partial discharge, solves the "black box" problem of traditional deep learning models, and improves the interpretability of the model and the credibility of the diagnostic results. Attached Figure Description

[0016] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of a substation acoustic signal partial discharge diagnosis method based on an improved matching pursuit network.

[0018] Figure 2 This is the time-domain waveform of the original local radio signal.

[0019] Figure 3 This is the time-domain waveform of the DC-DC-removed local output signal.

[0020] Figure 4 This is the time-domain waveform of the partial discharge signal after high-pass filtering.

[0021] Figure 5 This is the original STFT time-frequency diagram.

[0022] Figure 6 This is a sparse time-frequency plot after MPNet feature extraction.

[0023] Figure 7 This is a comparison chart showing the accuracy of this method in identifying substation partial discharge types with traditional STFT+SVM and FSST+SVM, respectively. Detailed Implementation

[0024] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] like Figure 1 As shown, this application provides a method for diagnosing partial discharge of acoustic signals in substations based on an improved matching pursuit network, including the following steps: S100: Acquire the initial time-domain signal of the substation and perform segmented preprocessing on the initial time-domain signal to obtain the preprocessed time-domain segmented signal.

[0027] One method for acquiring the initial time-domain signal in a substation is to use a high-sensitivity microphone array. Acquisition points are placed 1-3 meters around key equipment with high partial discharge (PD) generation, such as transformers and GIS systems, to collect continuous PD signals as the initial time-domain signal. Compared to a single microphone, this method can more comprehensively capture the spatial characteristics of the PD signal and reduce signal loss. The acquisition duration and frequency can be set according to actual needs.

[0028] The process of segmenting the initial time-domain signal into preprocessed time-domain segments includes the following steps: The initial time-domain signal is divided into multiple time-domain segmented signals according to a preset segment length; The DC component is removed from the time-domain segmented signal to obtain the removed time-domain segmented signal; The time-domain segmented signal after removal is subjected to high-pass filtering to obtain the preprocessed time-domain segmented signal.

[0029] It should be noted that the initial time-domain signal is a continuous long sequence. Directly performing subsequent time-frequency conversion and network feature extraction on this long sequence would result in excessive computation and low processing efficiency. Furthermore, since the partial discharge (PD) signal is a pulse-like signal, segmentation is necessary to more accurately capture local PD pulse characteristics and avoid the masking of local PD features by irrelevant signals in the long signal, thus facilitating accurate extraction of PD-related information. Here, the preset segment length can be set according to actual needs. Using the preset segment length as a window, the continuous initial time-domain signal is divided into several independent time-domain segment signals of equal length, each with a corresponding time interval. The number of segments can be calculated based on the sampling duration and segment length.

[0030] Partial discharge (PD) signals acquired at substation sites often have a fixed DC component superimposed due to factors such as circuit drift of the acquisition equipment and environmental electromagnetic interference. This manifests as an overall deviation of the signal waveform from the zero baseline. This DC component is considered invalid interference and can mask the true amplitude characteristics of the PD pulse, leading to baseline fluctuations during subsequent time-frequency conversion and affecting feature extraction accuracy. Therefore, it is necessary to remove the DC component first to return the signal to the zero baseline. Specifically, the DC component is calculated and removed for each time-domain segment signal. The average amplitude of the time-domain segment signal at all sampling points within a preset segment length is calculated, and then the average amplitude is subtracted from the amplitude of the time-domain segment signal at each sampling point within the preset segment length. The calculation formula is as follows: ; In the formula, The time-domain segmented signal after removal. For time-domain segmented signals, This represents the total number of data collection points within a preset segment length. For the first k Within the preset segment length, the first... i The initial time-domain signal corresponding to each acquisition point t The sampling time at the collection point. , For the duration of data collection, .

[0031] The above In the formula, The sampling frequency.

[0032] In the complex environment of substations, there are a large number of low-frequency noises below 1kHz (such as mechanical vibration of equipment, ambient noise, and operating noise of auxiliary equipment such as fans / oil pumps). These noises have large amplitudes and wide distributions, and will be superimposed on the partial discharge sound signal (mainly high-frequency characteristics). Even after removing the DC component, the low-frequency noise will still interfere with the partial discharge pulse characteristics. Therefore, it is necessary to filter out the low-frequency invalid noise through high-pass filtering and retain only the high-frequency signal containing partial discharge characteristics. A Butterworth high-pass filter can be used to filter the segmented time-domain signal after filtering. The Butterworth high-pass filter has the advantages of a flat frequency response within the passband and no obvious ripple, avoiding amplitude distortion of the effective partial discharge signal during filtering. Simultaneously, based on the frequency characteristics of the substation partial discharge signal (e.g., the corona discharge resonant frequency of 8kHz), a pre-set filter cutoff frequency allows partial discharge-related signals with frequencies higher than the cutoff frequency to pass through, while suppressing low-frequency noise below the cutoff frequency. The resulting pre-processed segmented time-domain signal retains only the effective characteristics of high-frequency partial discharge, with no DC offset and no significant low-frequency noise interference, significantly improving signal quality. It can be directly used as the input signal for subsequent time-frequency conversion, ensuring the accuracy of subsequent diagnostic processes from the source. Here, the filter cutoff frequency can be set to 1kHz.

[0033] For example, such as Figure 2 As shown, the original signal exhibits a significant DC offset (average of approximately 0.2 Pa), with the overall waveform deviating from the zero baseline. Simultaneously, low-frequency ambient noise below 1 kHz (amplitude approximately 0.15 Pa) causes significant baseline fluctuations in the waveform, partially masking the partial discharge pulse characteristics. This illustration visually reflects the quality issues of the original signal under the complex noise environment of a substation, validating the necessity of subsequent DC removal and filtering preprocessing. Figure 3 As shown, by subtracting the average amplitude of each signal segment, the signal mean drops to approximately -0.003 Pa, successfully returning to the zero baseline, and the DC component is effectively eliminated. However, baseline fluctuations caused by low-frequency noise (amplitude approximately 0.15 Pa) are still visible in the figure, indicating that simply removing DC cannot completely suppress environmental noise interference, and further high-pass filtering is still required. Figure 4 The time-domain waveform of the partial discharge signal after processing by a Butterworth high-pass filter with a cutoff frequency of 1 kHz is shown. Figure 3 In comparison, the amplitude of low-frequency noise below 1kHz decreased from 0.15Pa to 0.02Pa, demonstrating a significant noise suppression effect; the partial discharge pulse characteristics (peak value 0.8~1.0Pa) were clearly highlighted, and the signal quality was greatly improved. This figure verifies the ability of high-pass filtering to effectively filter out low-frequency noise while preserving high-frequency partial discharge characteristics, providing a high-quality input signal for subsequent time-frequency conversion.

[0034] S200. Based on the physical mechanism of partial discharge acoustic radiation, construct noiseless analog signals and noisy analog signals corresponding to different partial discharge types, and perform time-frequency conversion on the noisy analog signals and the noiseless analog signals respectively to obtain the noisy time-frequency matrix and the noiseless time-frequency matrix.

[0035] Based on the physical mechanism of partial discharge acoustic radiation, noise-free analog signals and noisy analog signals corresponding to different types of partial discharge are constructed. This process includes the following steps: Obtain the damped oscillation model corresponding to the preset partial discharge type, and generate a noise-free analog signal based on the damped oscillation model; The simulated environmental noise signal of the substation is acquired, and the sum of the simulated environmental noise signal and the noise-free simulated signal is used as the noisy simulated signal corresponding to each preset partial discharge type.

[0036] It should be noted that the preset partial discharge types include at least corona discharge, surface discharge, and internal discharge. The acoustic radiation process of these three types of partial discharge all exhibits damped oscillation characteristics. When a partial discharge occurs, it generates a pulsed acoustic signal, and the amplitude of this signal decays exponentially over time. The damped oscillation model can accurately characterize this physical characteristic. The core parameters of the model differ for different partial discharge types to distinguish the signal characteristics of different partial discharges. For example, the damping ratio corresponding to corona discharge is 0.05, the damping ratio corresponding to surface discharge is 0.08, and the damping ratio corresponding to internal discharge is 0.12.

[0037] Calculate the noise-free analog signal using the following formula: ; In the formula, For the first k Preset partial discharge type within a preset segment length type Noiseless analog signal, preset partial discharge type , For corona discharge, For surface discharge, For internal discharge, For the first k The number of partial discharges within a preset segment length is determined by the partial discharge occurrence rate. For the first k Within the preset segment length, the first... i Preset partial discharge type of each collection point type The corresponding amplitude, Preset partial discharge type type The corresponding damping ratio, Preset partial discharge type type The corresponding partial discharge signal resonant frequency is determined by the material of the equipment. For the firstk Within the preset segment length, the first... i The partial discharge occurrence time at each collection point satisfies and , , For the first k The end time of each preset segment length, , For the first k The start time of a preset segment length, Preset partial discharge type type The corresponding initial phase of the partial discharge signal, It is a unit step function.

[0038] In the actual operating environment of a substation, partial discharge (PD) signals are masked by the superposition of noise from multiple devices (such as transformer mechanical vibration and GIS equipment operating noise) and environmental noise (such as on-site noise and low-frequency noise from fans / oil pumps). Noise-free simulated signals cannot reflect the real signal scenario. Therefore, by superimposing simulated environmental noise, a noisy simulated signal highly similar to the actual substation scenario is constructed. This allows the network to learn the ability to separate PD features from complex noise during the training phase, ensuring strong robustness of the trained network in real substation environments. Specifically, the simulated environmental noise signal is a mixed noise that closely matches the actual substation environment, covering typical substation noise frequency bands (such as 800Hz low-frequency environmental noise and 3kHz equipment operating noise) and noise amplitude characteristics. It can also simulate noise scenarios with different signal-to-noise ratios (such as -10dB and -5dB, typical low signal-to-noise ratio scenarios in substations). Based on the noise distribution pattern of the substation, a simulated environmental noise signal with the same duration and sampling frequency as the initial time domain signal is generated to ensure the synchronization of signal superposition. Then, the simulated environmental noise signal is superimposed with the noiseless simulated signal of the corresponding preset partial discharge type to obtain the noisy simulated signal of each preset partial discharge type under different noise scenarios.

[0039] The generation method of simulated environmental noise signal is as follows: First, collect typical noise samples from substations (including transformer mechanical vibration noise, GIS equipment operation noise, low-frequency noise from fans / oil pumps, etc.) and extract their power spectral density characteristics; then, based on Gaussian white noise, generate a noise sequence with the same statistical characteristics through a well-designed filter (frequency response matching the measured noise spectrum); finally, adjust the noise amplitude according to the required signal-to-noise ratio conditions (such as -10dB, -5dB, 0dB, 5dB) to generate a simulated environmental noise signal with the same time length and sampling frequency as the noiseless simulated signal.

[0040] The formula for calculating a noisy analog signal is: ; In the formula, For the firstk A noisy analog signal with a preset segment length, For the first k A noise-free analog signal with a preset segment length, For the first k A simulated noise signal with a preset segment length.

[0041] Then, Hilbert demodulation and short-time Fourier transform (STFT) operations are performed on the noiseless analog signal and the noisy analog signal respectively. First, the envelope spectrum signals of the noiseless analog signal and the noisy analog signal are obtained by Hilbert demodulation, focusing on the amplitude variation characteristics of the partial discharge signal. Then, the envelope spectrum signal is transformed into a two-dimensional time-frequency matrix (the horizontal axis is time and the vertical axis is frequency) by STFT time-frequency transformation. At the same time, the frequency band of interest is determined according to the partial discharge resonant frequency, and the elements in the time-frequency matrix located in the frequency band of interest are extracted and the noise of irrelevant frequencies is removed. Finally, the noisy time-frequency matrix and the noiseless time-frequency matrix are obtained, and the time-domain signal characteristics are transformed into time-frequency characteristics that are easier for the network to learn.

[0042] Since the operation process is exactly the same for noiseless analog signals and noisy analog signals, the above calculation process will now be introduced using noisy analog signals as an example.

[0043] The formula for calculating the envelope spectrum of a noisy analog signal is: ; In the formula, For the first k The envelope spectrum signal corresponding to a noisy analog signal with a preset segment length. For the first k A noisy analog signal with a preset segment length, j The imaginary unit, for Hilbert transform.

[0044] The above , It is the integral variable.

[0045] The formula for calculating the noisy time-frequency matrix is: ; In the formula, For the Hanning window function, , ( (otherwise it is 0). is the complex exponential term of the Fourier transform.

[0046] The expression for determining the frequency band of interest based on the partial discharge resonant frequency is as follows: ; In the formula, The minimum endpoint value of the frequency band of interest. The maximum endpoint value of the frequency band of interest. This represents the minimum resonant frequency corresponding to all preset partial discharge types. This represents the maximum resonant frequency corresponding to all preset partial discharge types.

[0047] S300. Using the noisy time-frequency matrix as the input sample and the noiseless time-frequency matrix as the output label, multiple training sample pairs are constructed. The noisy time-frequency matrix in the training sample pair is then input into the improved matching pursuit network to obtain the estimated feature matrix.

[0048] Using a noisy time-frequency matrix as input samples, the characteristics of noisy partial discharge signals in a real substation are simulated. A noiseless time-frequency matrix is ​​used as output labels, which represent the pure partial discharge characteristics under ideal conditions. The two are combined in pairs to obtain training sample pairs, and the training sample pairs cover three types of partial discharge: corona, surface, and internal, to ensure the generalization ability of the network.

[0049] Furthermore, the improved matching pursuit network includes multiple stacked matching pursuit units, each of which is used to obtain the estimated feature matrix; Obtaining the sparse time-frequency feature matrix includes the following steps: Batch normalization is performed on the noisy time-frequency matrix in the training sample pair to obtain the batch-normalized noisy time-frequency matrix; The noisy time-frequency matrix after batch normalization is subjected to dilation convolution and activation function processing to obtain the effective feature matrix; The difference between the noisy time-frequency matrix and the effective feature matrix in the training sample pair is used as the updated residual matrix and output to the next matching tracking unit; Calculate the product of the effective feature matrix extracted by each matching tracking unit and its corresponding weight gain, and sum all the products to obtain the estimated feature matrix.

[0050] It should be noted that, for a batch of input noisy time-frequency matrices, the mean and variance of all noisy time-frequency matrices within that batch are calculated. A linear transformation is then used to map the element values ​​of each noisy time-frequency matrix to a standard distribution with a mean of 0 and a variance of 1. Simultaneously, a learnable scaling factor is introduced. and offset factor This allows the network to adaptively adjust the numerical distribution according to training needs, avoiding the loss of feature information caused by standardization; the expression for the noisy time-frequency matrix after batch normalization is: ; In the formula, For the first kA noisy time-frequency matrix with preset segment lengths after batch normalization. For batch normalization processing operations, Scaling factor As the offset factor, This is the mean of all noisy time-frequency matrices within the batch. Let V be the variance of all noisy time-frequency matrices within the batch. (To prevent small constants with a denominator of 0 in the expression).

[0051] Batch normalization ensures that noisy time-frequency matrices from different batches and scenarios have uniform numerical distribution characteristics, significantly improving the stability and convergence speed of network training. At the same time, it eliminates the interference of amplitude scale in the time-frequency matrix, allowing the network to focus more on the spatial distribution of time-frequency features (such as partial discharge feature ridges) rather than simply the amplitude magnitude, thus improving the targeting of feature extraction.

[0052] Dilated convolution introduces a dilation rate on top of ordinary convolution, allowing the convolution kernel to sample at intervals along the time and frequency dimensions of the noisy time-frequency matrix after batch normalization, thus expanding the receptive field (e.g., doubling the receptive field when the dilation rate is 2). This accurately captures the continuous spatial distribution of partial discharge feature ridges while avoiding small-scale interference from local noise. Secondary batch normalization is performed on the output of dilated convolution again to eliminate the numerical distribution shift caused by the dilated convolution operation, ensuring the stability of feature extraction. Leaky ReLU activation is used instead of the hard threshold activation of traditional ReLU, preserving small gradients for negative numerical features, avoiding feature death problems during network training, and fully retaining the detailed information of partial discharge features. Finally, through this combination of operations, an effective feature matrix is ​​obtained, which contains only the effective partial discharge features separated from the noisy time-frequency matrix, and the noisy features are initially filtered out.

[0053] Then, the noisy time-frequency matrix and the effective feature matrix are used to obtain an updated residual matrix. The updated residual matrix is ​​used to represent the noise feature matrix remaining after extracting the effective partial discharge features. The updated residual matrix is ​​input into the next matching pursuit unit to achieve layer-by-layer suppression of noise features, allowing subsequent matching pursuit units to further separate the unextracted fine features of partial discharge from the remaining noise, ultimately achieving maximum noise suppression and complete extraction of partial discharge features.

[0054] by n Taking the first matching tracking unit as an example, each matching tracking unit extracts the effective feature matrix through dilated convolution. i The formulas for feature extraction and residual update of each matching tracking unit are as follows: ; ; In the formula, For the first k The preset segment length is obtained by the first The effective feature matrix extracted by each matching tracking unit For the first The feature function of each matching tracking unit consists of dilated convolution, batch normalization, and Leaky ReLU activation. This refers to dilation convolution and batch normalization operations. It is a linear rectified activation function. For the first The updated residual matrix of each matching tracking unit reflects the remaining noise after extracting the effective feature matrix. For the first k The preset segment length is in the first... The residual matrix of the matching tracking unit input.

[0055] The partial discharge features extracted by different matching tracking units have different feature contributions (e.g., the first matching tracking unit extracts the main features, and subsequent matching tracking units extract the fine features). By using learnable weight gain to assign adaptive weights to the effective feature matrix of each matching tracking unit, the feature contribution of each unit is quantified. Then, the estimated time-frequency feature matrix is ​​obtained by weighted fusion, which integrates the partial discharge features (main features and fine features) extracted from each layer and achieves feature sparsity, so that the final estimated feature matrix retains only the core partial discharge features and completely eliminates noise interference.

[0056] Assign a learnable weight gain to each matching tracking unit The learnable weight gain can be adaptively updated through network training, and it satisfies... The magnitude of the learnable weight gain value is used to characterize the contribution of the effective features of the corresponding matching and tracking unit.

[0057] The global sparse partial discharge temporal feature matrix is ​​calculated using the following formula: ; In the formula, For the first k A global sparse partial discharge temporal feature matrix with a preset segment length.

[0058] Then, based on batch normalization and ReLU activation, the estimated feature matrix is ​​obtained: ; In the formula, For the first k The estimated feature matrix of a preset segment length, For the first kBatch normalization of global sparse partial discharge temporal feature matrices with preset segment lengths.

[0059] Learnable weight gain allows the network to adaptively focus on high-contribution partial discharge (PD) features and eliminate low-contribution redundant features, achieving adaptive sparsity of features. Weighted fusion integrates the main and detailed PD features extracted by each matching pursuit unit, ensuring that the final estimated time-frequency feature matrix contains both the core main features of PD and retains the details of fine features, significantly improving the completeness and discriminability of the features. Experimental data show that after this step, noise energy is suppressed by more than 60%, the PD feature ridges are continuous and the energy is highly concentrated, and the sparsity is significantly improved.

[0060] Breaking through the limitations of traditional matching pursuit algorithms that rely on manually designed dictionaries, this study achieves accurate and automatic extraction of partial discharge (PD) time-frequency features in complex noise environments through adaptive feature extraction using deep learning. Simultaneously, the residual decomposition architecture of the stacked matching pursuit units ensures a high degree of correlation between feature extraction and the physical mechanism of PD, solving the problem of poor interpretability in traditional deep learning models. The final extracted sparse time-frequency feature matrix not only fully preserves the core time-frequency features of PD (such as resonant frequency and feature ridges) but also completely eliminates interference from complex noise in substations.

[0061] For example, such as Figure 5 As shown in the figure, the energy is dispersed within the frequency band of interest from 5 to 13 kHz. Noise (3 kHz equipment noise and 800 Hz ambient noise) causes the partial discharge characteristic ridge near 8 kHz to be broken and discontinuous, with the maximum energy value being only about 0.3 Pa·s. This figure intuitively reflects the limitations of traditional time-frequency analysis methods in complex noise environments, with poor feature extraction robustness and partial discharge characteristics easily masked by noise. Figure 6 This paper presents sparse time-frequency plots obtained by feature extraction from the same noisy partial discharge signal using an improved matched pursuit network (MPNet). Figure 5 In comparison, noise energy was suppressed by more than 60% (noise region energy decreased from 0.2 Pa·s to 0.08 Pa·s); the partial discharge feature ridge near 8 kHz was continuous and complete, with highly concentrated energy (maximum value increased to 0.5 Pa·s), and the sparsity of time-frequency features was significantly improved. This figure verifies MPNet's core ability to extract effective features and suppress noise interference layer by layer through stacked matching pursuit units.

[0062] S400. Based on the structural similarity loss function, calculate the loss value between the estimated feature matrix and the noise-free time-frequency matrix in the training sample pair, and optimize the network parameters of the improved matching pursuit network according to the loss value to obtain the optimized improved matching pursuit network.

[0063] The loss between the estimated feature matrix and the noise-free time-frequency matrix in the training sample pair is calculated based on the structural similarity loss function, specifically including the following steps: Calculate the mean, variance, and covariance of the estimated feature matrix and the noise-free time-frequency matrix, respectively; Based on the structural similarity loss function, a negative structural similarity index is calculated according to the mean, variance, and covariance, and the negative structural similarity index is used as the loss value.

[0064] It should be noted that the mean reflects the overall amplitude level of the matrix, corresponding to brightness in structural similarity. It represents the overall energy magnitude of the partial discharge feature in the time-frequency space. When calculated, the arithmetic mean of the amplitudes of all elements of the matrix is ​​taken, which can reflect the overall energy deviation between the two matrices. The variance reflects the amplitude fluctuation of the elements within the matrix, corresponding to contrast in structural similarity. It represents the difference in amplitude variation of the partial discharge feature ridge. When calculated, the mean is used to calculate the mean of the squared deviations of each element, which can reflect the matching degree of feature fluctuations between the two matrices. The covariance reflects the degree of linear correlation between the elements of the two matrices, corresponding to structure in structural similarity. It represents the consistency of the distribution of the partial discharge feature in the time-frequency space. If the covariance value is large, it indicates that the amplitude variation trend of the features of the two matrices at the same time and frequency position is consistent, and the spatial distribution of the partial discharge feature ridge is more matched.

[0065] The formulas for calculating the mean, variance, and covariance are as follows: ; ; ; In the formula, To estimate the mean of the characteristic matrix, To estimate the variance of the characteristic matrix, To estimate the covariance of the characteristic matrix and the noise-free time-frequency matrix, It is a noise-free time-frequency matrix.

[0066] The mean and variance of the estimated characteristic matrix and the noise-free time-frequency matrix are calculated in the same way, and will not be repeated here.

[0067] The expression for the structural similarity loss function is: ; In the formula, The loss value. To the number of training sample pairs, To estimate the mean of the characteristic matrix, The mean of the noise-free time-frequency matrix. To estimate the variance of the characteristic matrix, The variance of the noise-free time-frequency matrix, To estimate the covariance of the characteristic matrix and the noise-free time-frequency matrix, , All of these are constants for calculating steady-state losses. , , This represents the maximum value of the noise-free time-frequency matrix.

[0068] The final loss value is obtained by calculating the structural similarity index for each individual training sample pair, then taking the arithmetic mean of the indices for all training sample pairs, and finally taking the negative value. This calculation method supports batch training and is compatible with batch normalization and batch gradient descent of the network. After each iteration of network parameters, the loss value of a batch of samples is quickly calculated, providing the optimizer with a precise gradient descent direction, allowing the network to quickly converge to the optimal state with a high degree of feature structure matching.

[0069] Furthermore, based on the loss value, the network parameters of the improved matching pursuit network are optimized to obtain the optimized improved matching pursuit network, specifically including the following steps: Based on the optimizer, the network parameters of the improved matching pursuit network are iteratively updated with the goal of minimizing the loss value. The network parameters include at least the convolution kernel weights, batch normalization parameters, and weight gain of each unit in each matching pursuit unit. Training stops when the loss value converges or reaches the preset number of iterations, resulting in an optimized and improved matching tracking network.

[0070] Among them, the optimizer type is, for example, the Adam optimizer. The kernel weights in each matching pursuit (MP) unit are the core parameters for dilated convolution to extract partial discharge time-frequency features. The value of the kernel weights directly determines the ability of dilated convolution to distinguish between partial discharge features and noise features in the time-frequency matrix. Updating this parameter allows the convolution kernel to continuously adapt to the spatial distribution of partial discharge features and accurately capture the ridge lines of partial discharge features. The batch normalization parameter is the scaling factor and offset factor in the batch normalization operation. This parameter determines the numerical standardization scale of the time-frequency matrix. Updating this parameter allows the network to adaptively adjust the standardization rules according to the amplitude distribution of partial discharge features, avoiding the loss of partial discharge feature details due to over-standardization.

[0071] Specifically, a batch of training sample pairs is taken and input into the improved matching pursuit network. After forward propagation, the estimated feature matrix is ​​obtained. The loss value of this batch is calculated according to the aforementioned structural similarity loss function to measure the difference between the estimated feature matrix and the true label. Then, backpropagation is performed based on the optimizer to calculate the gradient values ​​of the loss value with respect to all core parameters in the network. The gradient direction represents which direction the parameter adjustment can reduce the loss value. The optimizer makes small adjustments to the convolution kernel weights, batch normalization parameters, and weight gains based on the gradient values ​​and the preset initial learning rate to update the parameters. The above steps are repeated to input all training sample pairs into the improved matching pursuit network in batches, realizing multi-round iterative updates of network parameters. After each round of updates, the network's loss value will gradually decrease, and the feature extraction capability will gradually improve. When the global training loss of the improved matching pursuit network decreases by less than a preset threshold and remains stable without significant fluctuations over multiple iterations, the loss is considered to have converged. At this point, the parameters of the improved matching pursuit network have been adjusted to the optimal range, and the structural similarity between the estimated feature matrix it can extract and the true partial-amplitude feature matrix has reached an extremely high level. Continuing training will not significantly improve the network performance and may even lead to overfitting. Here, the preset threshold is, for example, 10. -5 Alternatively, training can be stopped once a preset number of iterations has been reached. This prevents the network from getting stuck in an infinite training loop while ensuring that the network receives sufficient training and possesses basic feature extraction capabilities. Here, the preset number of iterations is, for example, 1000.

[0072] S500: Perform time-frequency conversion on the preprocessed time-domain segmented signal to obtain the real signal time-frequency matrix, and input the real signal time-frequency matrix into the optimized improved matching pursuit network to obtain the sparse time-frequency feature matrix.

[0073] The process of performing time-frequency conversion on the preprocessed time-domain segmented signal to obtain the real signal time-frequency matrix includes the following steps: The Hilbert transform is performed on the preprocessed time-domain segmented signal to obtain the envelope spectrum signal; The time-frequency matrix is ​​obtained by performing a short-time Fourier transform on the envelope spectrum signal; The range of the frequency band of interest is determined based on the partial discharge resonant frequency, and the elements in the time-frequency matrix that are located within the range of the frequency band of interest are extracted to obtain the actual signal time-frequency matrix.

[0074] The calculation process here is the same as described above, and will not be repeated here.

[0075] For the preprocessed time-domain segmented signal, a three-step operation of Hilbert transform, short-time Fourier transform, and extraction of the frequency band of interest is used to transform the one-dimensional time-domain partial discharge signal into a two-dimensional time-frequency matrix that is easier to learn by the improved matching pursuit network. At the same time, the effective frequency band of the partial discharge feature is focused and irrelevant frequency noise is removed, providing accurate and suitable network input for subsequent sparse time-frequency feature extraction.

[0076] S600: Based on the classifier, the sparse time-frequency feature matrix is ​​identified to obtain the partial discharge type identification result of the current preprocessed time-domain segmented signal.

[0077] The process involves identifying the sparse time-frequency feature matrix based on a classifier to obtain the partial discharge type identification result of the preprocessed time-domain segmented signal. This includes the following steps: Flatten the sparse time-frequency feature matrix into a one-dimensional feature vector; The one-dimensional feature vector is input into the classifier to obtain the probability that the real signal belongs to each preset partial discharge type, and the preset partial discharge type corresponding to the highest probability is taken as the partial discharge type identification result of the current preprocessed time-domain segmented signal.

[0078] Here, the sparse time-frequency feature matrix is ​​flattened, for example, using the classic matrix vectorization (vec) method. The sparse time-frequency feature matrix is ​​arranged in a fixed row-major or column-major order, and all elements of the sparse time-frequency feature matrix are sequentially arranged into a one-dimensional feature vector, which is either a one-dimensional row vector or column vector. Then, the probability of each preset partial discharge type is calculated according to the following formula: ; In the formula, Let be the probability that the k-th real signal belongs to the c-th preset partial discharge type, where c=1 represents corona discharge, c=2 represents surface discharge, and c=3 represents internal discharge. The classifier outputs a linear signal of the k-th real signal in the c-th class of the preset partial discharge type. , The weight vector for the predefined local emission type of class c, For matrix flattening operation, For the bias term of the predefined partial discharge type of type c, For the preset partial discharge type index ( ), e It is a natural exponential function.

[0079] The preset partial discharge type corresponding to the highest probability is taken as the final partial discharge type identification result of the preprocessed time-domain segmented signal.

[0080] Furthermore, this method also includes the following steps: Acquire test signals under different signal-to-noise ratio conditions; The test signal is converted from time to frequency to obtain the time-frequency matrix of the test signal; The time-frequency matrix of the test signal is input into the optimized and improved matching pursuit network to obtain the corresponding sparse time-frequency feature matrix. The partial discharge type identification results of each test signal are obtained by identifying the sparse time-frequency feature matrix corresponding to the time-frequency matrix of the test signal based on the classifier. Based on the partial discharge type identification results of each test signal, the identification accuracy under each signal-to-noise ratio condition is determined; the identification accuracy is used to verify the robustness of the optimized and improved matched pursuit network under different noise environments.

[0081] Specifically, firstly, based on the damped oscillation physical models corresponding to the three preset partial discharge types—corona discharge, surface discharge, and internal discharge—noise-free test signals are generated to ensure that the partial discharge characteristics of the test signals are consistent with those of real substation partial discharge signals. Secondly, typical mixed noise in a substation (including equipment mechanical noise, low-frequency environmental noise, etc.) is simulated. By adjusting the amplitude of the noise, it is superimposed with the pure partial discharge signal to generate test signals with different signal-to-noise ratio (SNR) gradients. The SNR range typically covers the actual noise range of the substation, for example, -10dB to 5dB, with a step size of 5dB, where -10dB represents strong noise and 5dB represents weak noise. Finally, sufficient test samples are configured for each SNR and each preset partial discharge type (e.g., 200 test samples for each preset partial discharge type) to avoid deviations in verification results due to insufficient sample size and to ensure the objectivity of accuracy statistics.

[0082] Then, for the test signals at various signal-to-noise ratios, a Hilbert transform is performed to obtain the envelope spectrum signal, a short-time Fourier transform (STFT) is performed to obtain the full-band time-frequency matrix, and the frequency band of interest is truncated according to the partial discharge resonant frequency, finally obtaining the time-frequency matrix of the test signal corresponding to each test signal. Throughout the entire process, the parameters of the Hilbert transform, the window function (Hanning window) of the STFT, the window length, the overlap rate, and the range of the frequency band of interest are all consistent with those in the model training and real signal diagnosis stages, without any parameter adjustments.

[0083] Then, the time-frequency matrices of the test signals at each signal-to-noise ratio are input one by one into the improved matching pursuit network that has been trained and has fixed parameters. The time-frequency matrices of the test signals are processed according to the operation process of batch normalization, dilated convolution, residual update, and weighted feature fusion, and finally the sparse time-frequency feature matrix corresponding to each test signal is output.

[0084] Finally, the sparse time-frequency feature matrix corresponding to the test signal is flattened into a one-dimensional feature vector and input into the trained Softmax classifier. The Softmax classifier outputs the probability value of each test signal belonging to one of the three types of partial discharges through linear mapping and probability normalization, and the partial discharge type corresponding to the highest probability is taken as the recognition result of the test signal. The parameters of the Softmax classifier also remain fixed after training without any adjustment. For all test samples at each signal-to-noise ratio (SNR), the ratio of correctly identified samples to the total number of test samples is used as the recognition accuracy. To reduce random errors, multiple tests can be performed for each SNR, and the average value can be taken to obtain the final recognition accuracy under each SNR condition. If a high recognition accuracy can be maintained across the entire SNR range (especially in high-noise environments with low SNR, such as -5dB to -10dB) (e.g., 95.2% accuracy at -5dB and 82.3% accuracy at -10dB), and the recognition accuracy does not fluctuate drastically, it indicates that the current model has strong anti-interference ability, good environmental adaptability, and meets the robustness requirements. If the recognition accuracy drops significantly at low SNR, it indicates that the current model has insufficient noise suppression ability and poor robustness, and the network needs to be optimized.

[0085] For example, such as Figure 7 As shown, the accuracy of partial discharge (PD) type identification in this invention (MPNet+Softmax) was compared with that of two traditional methods (STFT+SVM, FSST+SVM) at different signal-to-noise ratios (-10dB~5dB). The experiment used three types of PD signals: corona, surface, and internal, with 200 samples for each type. The tests were performed five times at each PD level, and the average was taken. At a low PD level of -10dB, the accuracy of this method was 82.3%, significantly higher than STFT+SVM (58.1%) and FSST+SVM (65.7%). At a typical substation noise level of -5dB, the accuracy of this method was 95.2%, meeting engineering application requirements, while the comparative methods were only 70.4% and 78.9%. At a high PD level of 5dB, the accuracy of this method was 98.7%, still leading by 8-10 percentage points.

[0086] This figure fully demonstrates the robustness of our method in complex noise environments, which is mainly attributed to the strong correlation between the sparse time-frequency features extracted by MPNet and the physical mechanism of partial discharge.

[0087] Furthermore, it should be noted that the technical solution of this application has several feasible variations in its specific implementation. For example, in the simulated sample construction stage, in addition to the generation method based on the physical model, a small amount of real labeled data combined with data augmentation (such as signal noise addition and time-frequency plot cropping) can be used to expand the training samples. Another example is that the dilated convolution in the MP unit can be replaced with convolution with an attention mechanism or multi-scale convolution to focus on or capture features at different scales. Yet another example is that the final classifier can also use traditional classifiers such as Support Vector Machines (SVM). However, the above variations may differ from the preferred embodiment of this invention in terms of computational complexity, feature redundancy, or final diagnostic performance. The combination of simulated signal construction based on the physical model, stacked MPNet based on dilated convolution, and the Softmax classifier described in this embodiment has comprehensive advantages in balancing model performance, complexity, and interpretability.

[0088] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for diagnosing partial discharge of acoustic signals in substations based on an improved matching pursuit network, characterized in that, Includes the following steps: The initial time-domain signal of the substation is acquired, and the initial time-domain signal is segmented and preprocessed to obtain the preprocessed time-domain segmented signal. Based on the physical mechanism of partial discharge acoustic radiation, noiseless analog signals and noisy analog signals corresponding to different partial discharge types are constructed, and time-frequency conversion is performed on the noisy analog signals and the noiseless analog signals respectively to obtain noisy time-frequency matrices and noiseless time-frequency matrices. Using the noisy time-frequency matrix as input samples and the noise-free time-frequency matrix as output labels, multiple training sample pairs are constructed. The noisy time-frequency matrix in the training sample pairs is then input into the improved matching pursuit network to obtain the estimated feature matrix. Based on the structural similarity loss function, the loss value between the estimated feature matrix and the noise-free time-frequency matrix in the training sample pair is calculated, and the network parameters of the improved matching pursuit network are optimized according to the loss value to obtain the optimized improved matching pursuit network. The preprocessed time-domain segmented signal is subjected to time-frequency conversion to obtain the real signal time-frequency matrix, and the real signal time-frequency matrix is ​​input into the optimized improved matching pursuit network to obtain a sparse time-frequency feature matrix. Based on the classifier, the sparse time-frequency feature matrix is ​​identified to obtain the partial discharge type identification result of the preprocessed time-domain segmented signal.

2. The method for diagnosing partial discharge of acoustic signals in substations based on an improved matching pursuit network according to claim 1, characterized in that, The initial time-domain signal is segmented and preprocessed to obtain a preprocessed time-domain segmented signal. This process includes the following steps: The initial time-domain signal is divided according to a preset segment length to obtain multiple time-domain segmented signals; The DC component is removed from the time-domain segmented signal to obtain the removed time-domain segmented signal; The removed time-domain segmented signal is subjected to high-pass filtering to obtain the preprocessed time-domain segmented signal.

3. The method for diagnosing partial discharge of acoustic signals in substations based on an improved matched pursuit network according to claim 1, characterized in that, Based on the physical mechanism of partial discharge acoustic radiation, noise-free analog signals corresponding to different types of partial discharge and noisy analog signals corresponding to the noise-free analog signals are constructed, specifically including the following steps: Obtain the damped oscillation model corresponding to the preset partial discharge type, and generate a noise-free analog signal based on the damped oscillation model; The simulated environmental noise signal of the substation is acquired, and the sum of the simulated environmental noise signal and the noise-free simulated signal is used as the noisy simulated signal corresponding to each of the preset partial discharge types.

4. The method for diagnosing partial discharge of acoustic signals in substations based on an improved matched pursuit network according to claim 1, characterized in that, The preprocessed time-domain segmented signal is subjected to time-frequency conversion to obtain the real signal time-frequency matrix, specifically including the following steps: The preprocessed time-domain segmented signal is subjected to Hilbert transform to obtain the envelope spectrum signal; Perform a short-time Fourier transform on the envelope spectrum signal to obtain the time-frequency matrix; The range of the frequency band of interest is determined based on the partial discharge resonant frequency, and the elements in the time-frequency matrix that are located within the range of the frequency band of interest are extracted to obtain the real signal time-frequency matrix.

5. A method for diagnosing partial discharge of acoustic signals in substations based on an improved matching pursuit network, as described in claim 1, is characterized in that... Based on the structural similarity loss function, the loss value between the estimated feature matrix and the noise-free time-frequency matrix in the training sample pair is calculated, specifically including the following steps: Calculate the mean, variance, and covariance of the estimated feature matrix and the noise-free time-frequency matrix, respectively. Based on the structural similarity loss function, a negative structural similarity index is calculated according to the mean, the variance, and the covariance, and the negative structural similarity index is used as the loss value.

6. The method for diagnosing partial discharge of acoustic signals in substations based on an improved matching pursuit network according to claim 1, characterized in that, The improved matching pursuit network includes multiple stacked matching pursuit units, each of which is used to obtain an estimated feature matrix; Obtaining the sparse time-frequency feature matrix includes the following steps: Batch normalization is performed on the noisy time-frequency matrix in the training sample pair to obtain the batch-normalized noisy time-frequency matrix; The noisy time-frequency matrix after batch normalization is subjected to dilation convolution and activation function processing to obtain the effective feature matrix; The difference between the noisy time-frequency matrix and the effective feature matrix in the training sample pair is used as the updated residual matrix and output to the next matching and tracking unit; Calculate the product of the effective feature matrix extracted by each matching tracking unit and its corresponding weight gain, and sum all the products to obtain the estimated feature matrix.

7. A method for diagnosing partial discharge of acoustic signals in substations based on an improved matched pursuit network, as described in claim 6, is characterized in that... Based on the loss value, the network parameters of the improved matching pursuit network are optimized to obtain the optimized improved matching pursuit network, specifically including the following steps: Based on the optimizer, the network parameters of the improved matching pursuit network are iteratively updated with the goal of minimizing the loss value; the network parameters include at least the convolution kernel weights, batch normalization parameters, and weight gain of each unit in each matching pursuit unit. When the loss value converges or reaches the preset number of iterations, training stops, and the optimized and improved matching tracking network is obtained.

8. The method for diagnosing partial discharge of acoustic signals in substations based on an improved matched pursuit network according to claim 1, characterized in that, Based on the classifier, the sparse time-frequency feature matrix is ​​identified to obtain the partial discharge type identification result of the preprocessed time-domain segmented signal. This specifically includes the following steps: Flatten the sparse time-frequency feature matrix into a one-dimensional feature vector; The one-dimensional feature vector is input into the classifier to obtain the probability that the real signal belongs to each preset partial discharge type, and the preset partial discharge type corresponding to the highest probability is taken as the partial discharge type identification result of the current preprocessed time-domain segmented signal.

9. A method for diagnosing partial discharge of acoustic signals in substations based on an improved matched pursuit network according to claim 1, characterized in that, The method further includes the following steps: Acquire test signals under different signal-to-noise ratio conditions; The test signal is converted from time to frequency to obtain the test signal time-frequency matrix; The time-frequency matrix of the test signal is input into the optimized and improved matching pursuit network to obtain the corresponding sparse time-frequency feature matrix. The partial discharge type identification results of each test signal are obtained by identifying the sparse time-frequency feature matrix corresponding to the time-frequency matrix of the test signal based on the classifier. Based on the partial discharge type identification results of each test signal, the identification accuracy under each signal-to-noise ratio condition is determined; the identification accuracy is used to verify the robustness of the optimized improved matching pursuit network under different noise environments.

Images