Cable discharge signal blind separation and enhancement processing method based on adversarial network

Abstract

The invention relates to the technical field of cable asset health management and predictive maintenance, and discloses a cable discharge signal blind separation and enhancement processing method based on an adversarial network, and the method comprises the steps: building a multi-modal monitoring data set through collecting mixed signals and environment data in cable operation; blind separation of discharge signals is realized by using the generative adversarial network, and prior information is not needed; identifying the number of potential signal sources through covariance analysis and double-criterion estimation; iterative optimization and signal enhancement are carried out in combination with a graph neural network and variational reasoning; and finally, through multiple cross validation and quality correction, an enhanced signal with high reliability is output. According to the method, the signal processing technology is deeply fused with asset management, risk prediction and operation and maintenance decision, weak discharge signals can be effectively separated and enhanced under the condition of low signal-to-noise ratio, the accuracy and reliability of cable early fault diagnosis are improved, and credible data support is provided for cable asset health state assessment, risk prediction and operation and maintenance decision.

Description

Technical Field

[0001] This invention relates to the field of cable signal processing technology, and more specifically, to a method for blind separation and enhancement of cable discharge signals based on adversarial networks. Background Technology

[0002] With the widespread application of power cable systems in smart grids, their operational reliability directly affects the safety and stability of the entire power system. During long-term service, the insulation material of cables is susceptible to partial discharge due to the influence of electric fields, thermal stress, and environmental factors. Early, weak partial discharge signals are often important precursors to insulation degradation. However, in actual monitoring environments, the acquired signals are usually a mixture of multiple sources, including weak partial discharge pulses, strong background noise (such as power frequency interference and electromagnetic interference), electrical transient processes, and other non-discharge interference signals. These signals often suffer from severe time-frequency overlap, extremely low signal-to-noise ratios, and a lack of prior information, making effective separation difficult. Existing blind separation methods often rely on the periodicity of the modulated signal, but cannot adapt to the non-periodic, transient, wideband, and energy-dispersed signal characteristics of cable partial discharge. Furthermore, conventional filtering methods are prone to signal distortion or loss of key features when frequency components are highly aliased, especially for early discharge signals with weak amplitude and short duration, posing a risk of missed detection and misjudgment, failing to meet the requirements of modern asset management for data reliability, predictability, and intelligent decision-making.

[0003] Therefore, how to achieve high-fidelity blind separation of early discharge signals of cables in a complex noise background without prior knowledge support, and further perform adaptive enhancement and reliability verification to output high-quality signals that can be used for asset health management, risk warning and operation and maintenance decision-making, has become an urgent technical problem to be solved. Summary of the Invention

[0004] This invention provides a method for blind separation and enhancement of cable discharge signals based on adversarial networks. It solves the technical problem in the prior art that it is difficult to achieve high-fidelity blind separation of early cable discharge signals in complex noise backgrounds without prior knowledge support, and further perform adaptive enhancement and reliability verification to output high-quality signals that can be used for asset health management, risk warning and operation and maintenance decision-making.

[0005] This invention provides a method for blind separation and enhancement of cable discharge signals based on adversarial networks, comprising: acquiring mixed signal and environmental data of the cable, performing preprocessing and normalization to construct a signal set to be separated; constructing and organizing a clean discharge signal sample library and a noise sample library based on the signal set to be separated to build a generative adversarial network; inputting the signal set to be separated into the generative adversarial network, extracting multi-scale features of the mixed signal through the encoder of the generator, and then restoring and outputting the corresponding preliminary separated signal through the decoder; constructing a covariance matrix and performing eigenvalue decomposition on the preliminary separated signal to obtain the theoretical boundary of the noise eigenvalues, and analyzing the difference features between the signal subspace and the noise subspace in the eigenvalue sequence to obtain an estimate of the number of signal sources in the mixed signal; determining the number of nodes in the graph structure based on the estimated number of signal sources, obtaining multiple Euclidean distances between nodes based on the time-frequency features of the preliminary separated signal, and constructing a representation. A similarity matrix of spatial and temporal correlation is generated. Based on this similarity matrix, spatial dependencies and temporal evolution patterns of node features are extracted to generate graph feature vectors. Using these graph feature vectors, a hierarchical variation distribution is constructed, and a posterior probability distribution for signal separation is calculated based on spatial and temporal correlation. The optimized target separation signal is then recursively optimized by maximizing the posterior probability distribution. A preliminary signal quality assessment is performed on the target separation signal to select high-quality separation results. Multidimensional feature extraction and adaptive amplification are then applied to the selected target separation signal to obtain the enhanced signal. Multiple cross-validation is performed on the target separation signal and the enhanced signal, and the separation effect is verified by calculating signal quality indices such as signal-to-noise ratio, improvement ratio, and cross-correlation coefficient. Finally, the verified separation effect is evaluated for signal quality, and low-quality signal regions are corrected to obtain the final enhanced signal.

[0006] Further, mixed signal and environmental data of the cable are acquired, and preprocessed and normalized to construct a signal set to be separated. This includes: acquiring mixed signal data containing discharge signals using ultra-high frequency sensors, current sensors, and partial discharge detection devices deployed at key nodes of the cable line; acquiring environmental parameters using temperature and humidity sensors and electromagnetic field strength detectors; denoting the mixed signal data as xi, where i represents the sampling time, and xi includes weak partial discharge pulse signals, strong background noise, and electrical transient signals; and performing preprocessing operations on the mixed signal data, including using bandpass filtering to remove DC and ultra-high frequency components, and using normalization to unify the signal amplitude range. Time synchronization between the mixed signal and environmental data is achieved through timestamp alignment. The normalized mixed signal data is divided into multiple signal segments according to a fixed time window, with each segment serving as an independent sample. The environmental data is mean-processed according to the corresponding time window to form environmental feature vectors paired with the signal segments. The signal segments undergo multi-channel conversion, converting the time-domain waveform of each signal segment into a frequency-domain spectrum, and combining it with the environmental feature vectors to generate a joint feature matrix. The joint feature matrix has a dimension of 4×m, where m is the number of sampling points in the signal segment. All joint feature matrices are combined to form a signal set to be separated, denoted as X=[x1, x2, ..., x...]. n ], where n is the total number of signal segments. Further, a pure discharge signal sample library and a noise sample library are pre-constructed and organized to build a generative adversarial network, including: simulating different types of partial discharges (point discharge, surface discharge, internal discharge, and suspension discharge) under a controlled laboratory environment; collecting pure discharge signal samples under shielding measures and low-noise conditions to construct a pure discharge signal sample library; collecting pure noise samples containing power frequency interference, electromagnetic field interference, and white noise under different environmental conditions, analyzing their spectral characteristics and probability distribution to construct a noise sample library; superimposing pure discharge signals with various types of noise according to different signal-to-noise ratios and mixing ratios to generate mixed signal samples with known components; classifying and storing all samples according to multiple dimensions such as discharge type, noise type, signal-to-noise ratio, and environmental conditions; and adding multi-dimensional representations such as time-domain waveforms, frequency-domain spectrograms, and time-frequency diagrams to each sample to construct a standardized sample library S=[s1, s2, ..., s...]. k ], where k is the total number of samples; a generative adversarial network architecture is trained based on the pure discharge signal sample library and the noise sample library; the generator in the generative adversarial network adopts an encoder and decoder architecture, and the discriminator adopts a deep convolutional neural network architecture; by training the generative adversarial network, the generator learns the mapping relationship for extracting pure discharge signals from mixed signals, and the network parameters of the generator and the discriminator are alternately optimized during the training process until the weighted sum of the adversarial loss and the reconstruction loss reaches the preset convergence condition.

[0007] Further, the set of signals to be separated is input into the generative adversarial network (GAN). The encoder of the generator extracts multi-scale features of the mixed signal, and the decoder then reconstructs and outputs the corresponding preliminary separated signal. This includes: inputting the mixed signal segments from the set of signals to be separated one by one into the trained GAN; the GAN performs multi-layer convolution operations on the input mixed signal through the encoder of the generator, extracting multi-scale features at different levels, including low-frequency basic features, mid-frequency texture features, and high-frequency detail features; fusing the multi-scale features in the feature space to generate a fused feature vector; the fused feature vector contains the separation representation of the discharge signal and noise signal in the mixed signal; the separation representation represents the compressed encoding form of the original time-domain signal in a high-dimensional feature space, with a dimension smaller than the original signal dimension, preserving the key semantic information of the signal and the discriminative features required for separation; inputting the fused feature vector into the decoder, which performs layer-by-layer upsampling through transposed convolution operations, outputting an abstract representation of the feature space; the... Abstract representation means that the original mixed signal is compressed, refined and highly generalized in a high-dimensional feature space, retaining the essential discriminative features and key information of the discharge signal and noise signal, while discarding redundant details, and expressing the information structure of the original signal in a low dimension. The abstract representation is mapped from the high-dimensional feature space back to the original time-domain signal space, gradually restoring it to the time-domain signal waveform. During the decoding process, skip connections are used to concatenate the features of the corresponding layer of the encoder with the features of the current layer of the decoder, retaining the detailed information of the signal. Boundary effects are removed and smoothed on the time-domain signal waveform to obtain the preliminary separated signal. The preliminary separated signal identifies and extracts the discharge signal component from the mixed signal and suppresses the noise component. A sliding window strategy is used to segment the long-time sequence signal, with each window length set to multiple sampling points and a preset overlap rate between windows. The generator outputs the corresponding preliminary separated signal segment for the mixed signal of each window. The multiple separation results of the overlapping area are fused by a weighted average method to obtain a complete and continuous preliminary separated signal.

[0008] Further, a covariance matrix is ​​constructed and eigenvalue decomposition is performed on the initially separated signal to obtain the theoretical boundary of the noise eigenvalues. The difference between the signal subspace and the noise subspace in the eigenvalue sequence is analyzed to obtain an estimate of the number of signal sources in the mixed signal. This includes: converting the initially separated signal into a multi-channel time-series matrix; each row of the multi-channel time-series matrix represents a multi-channel observation at a given time, and each column represents the time-series data of one channel; constructing a covariance matrix based on the multi-channel time-series matrix to obtain the covariance between any two channels; the covariance matrix is ​​a symmetric positive definite matrix, and its dimension is equal to the number of channels; performing eigenvalue decomposition on the covariance matrix to obtain the eigenvalue sequence [λ1≥λ2≥…≥λ MIn this process, M represents the number of channels. The energy percentage of each eigenvalue is calculated based on the eigenvalue sequence to obtain the eigenvalue energy sequence. Based on the eigenvalue sequence and the number of rows and columns of the multi-channel time-series matrix, the probability density value of the noise eigenvalue is calculated. The upper and lower boundaries of the noise eigenvalue are obtained through the probability density value. By comparing the relationship between each eigenvalue in the eigenvalue sequence and the upper and lower boundaries of the noise, the signal subspace and noise subspace are distinguished in conjunction with the eigenvalue energy sequence. Based on the upper and lower boundaries of the noise eigenvalue, the difference characteristics between the signal subspace and noise subspace in the eigenvalue sequence are analyzed, and the interval between adjacent eigenvalues ​​is calculated to construct an eigenvalue interval sequence. An energy gradient decision criterion is constructed based on the eigenvalue energy sequence. The energy gradient is obtained, and a dynamic energy threshold is set. When the energy gradient is greater than the dynamic energy threshold, it indicates that there is an energy jump between the eigenvalue at the current time and the eigenvalue at the next time. Therefore, the eigenvalue at the current time is determined from the perspective of energy distribution. The value corresponds to the first signal component; the first signal component is the energy-dominant signal component, representing the main discharge signal with a large energy proportion in the mixed signal; an interval decision criterion is constructed based on the feature value interval sequence, and an interval threshold is set. When the median of the feature value interval is greater than the preset interval threshold, the feature value at the current moment is determined to correspond to the second signal component from the perspective of interval distribution; the second signal component is the interval abrupt signal component, representing the secondary discharge signal or weak early discharge signal with a small energy proportion in the mixed signal but with interval jump characteristics in the feature value distribution; the energy gradient decision criterion identifies strong discharge signal sources, and the interval decision criterion identifies weak discharge signal sources. The two criteria complement each other to comprehensively capture signal sources of different intensities; the decision results of the energy gradient decision criterion and the interval decision criterion are assigned confidence weights wa and wb, respectively, and the decision values ​​are weighted and fused. The weighted decision value results are fused, and the number of feature values ​​that meet the decision conditions is counted to obtain the estimated number of signal sources.Further, the number of nodes in the graph structure is determined based on the estimated number of signal sources. The Euclidean distance between nodes is obtained based on the time-frequency characteristics of the initially separated signal. A similarity matrix characterizing spatial and temporal correlation is constructed, including: determining the number of nodes in the graph structure based on the estimated number of signal sources; mapping each signal source component in the initially separated signal to a node in the graph structure; performing time-frequency analysis on the initially separated signal to obtain the time spectrum of the signal, and extracting the time-frequency characteristics of the signal component corresponding to each node; the time-frequency characteristics include instantaneous frequency, time-frequency energy distribution, and time-frequency clustering; calculating the similarity between any two nodes. The first Euclidean distance between the time-frequency feature vectors of the corresponding signal components is calculated; a spatial correlation metric is constructed based on the first Euclidean distance, and a spatial correlation weight is obtained; the larger the correlation weight value, the more similar the two nodes are in the feature space; a second Euclidean distance is calculated for the time-frequency feature vectors of the same node at adjacent times; a temporal correlation metric is constructed based on the second Euclidean distance, and a temporal correlation weight value is calculated using a temporal kernel function; the spatial correlation weight value and the temporal correlation weight value are fused to construct a comprehensive similarity matrix; and the similarity matrix is ​​normalized to obtain a normalized similarity matrix.

[0009] Further, the similarity matrix is ​​input into a pre-constructed hybrid graph convolutional network to extract the spatial dependencies and temporal evolution patterns of node features, generating graph feature vectors. These graph feature vectors are then input into a preset variational inference model to construct a hierarchical distribution of changes and calculate the posterior probability distribution for signal separation based on spatial and temporal correlations. Recursive iterative optimization is performed using the maximized posterior probability distribution to output the optimized target separation signal. This includes inputting the normalized similarity matrix and the node feature matrix of the initial separation signal into a preset hybrid graph convolutional network. The hybrid graph convolutional network contains multiple graph convolutional layers, each performing message passing and feature aggregation operations. In the first graph convolutional layer, the set of neighboring nodes of the i-th node is determined based on the normalized similarity matrix, the feature information of its neighboring nodes is aggregated, and the updated node features are calculated. Convolution operations are performed on the node features in the time dimension to capture the temporal evolution of the node features, and temporal convolutional features are calculated. Spatial graph convolutional features and temporal graph convolutional features are fused to obtain hybrid graph convolutional features, which are then subjected to feature dimensionality reduction and nonlinear transformation through a fully connected layer to generate graph feature vectors. The graph feature vectors are converted into mapping feature vectors through a mapping network, where the mapping feature vectors represent the mean and logarithmic variance of the hierarchical variation distribution, respectively. Based on the mapping feature vectors and the preliminary separation signal... A hierarchical distribution of hierarchical variations is constructed by combining a global distribution representing overall dependencies and a local distribution representing independent features. The global distribution is a single Gaussian distribution, and the local distribution is a product of the Gaussian distributions corresponding to each initially separated signal. Spatial and temporal features are extracted from the initially separated signals. The Euclidean distance between the spatial features of any two signals is calculated to obtain a spatial correlation metric, and the Euclidean distance between the temporal features of the same signal at adjacent times is calculated to obtain a temporal correlation metric. A spatial correlation probability is constructed based on the spatial correlation metric, and a temporal correlation probability is constructed based on the temporal correlation metric. The global distribution is then... The posterior probability distribution is obtained by multiplying the distribution, the local distribution, the spatial correlation probability, and the temporal correlation probability. Recursive iterative optimization is performed by maximizing the lower bound of the posterior probability distribution. During the iteration, latent variables are sampled from the hierarchical variation distribution, and mapped back to the signal space through the decoder network to obtain the reconstructed signal. The reconstruction loss between the reconstructed signal and the original preliminary separation signal is calculated, and the reconstruction loss measures the signal reconstruction quality. The divergence loss between the hierarchical variation distribution and the prior distribution is also calculated, and the divergence loss constrains the distribution of the latent variables. The two losses are weighted and combined to obtain the total loss, and all network parameters are updated using a backpropagation algorithm to minimize the total loss.The iterative optimization process is repeated until the convergence condition that the change in total loss is less than a preset threshold or the maximum number of iterations is reached. The optimized target separation signal is then output. This target separation signal is the reconstructed signal obtained in the last iteration, and it has a higher signal-to-noise ratio and more accurate signal source separation compared to the initial separation signal.

[0010] Further, a preliminary signal quality assessment is performed on the target separated signal to select high-quality separation results. Then, multi-dimensional feature extraction and adaptive amplification signal enhancement processing are applied to the selected target separated signal, including: performing a preliminary signal quality assessment on the target separated signal, calculating the signal-to-noise ratio (SNR) of the input mixed signal and the SNR of the target separated signal; the SNR reflects the degree of signal quality improvement achieved by the separation process; based on the two SNR values, a correlation coefficient is further calculated between the target separated signal and a pre-stored reference clean signal; the correlation coefficient reflects the similarity between the separated signal and the ideal clean signal; a quality assessment threshold is set according to the SNR and the correlation coefficient. When the SNR is greater than a preset threshold and the correlation coefficient is greater than a preset threshold, a quality assessment threshold is set. When the correlation coefficient threshold is reached, it indicates that the target separation signal has both a good signal-to-noise ratio improvement and a high correlation with the clean signal, and is judged as a high-quality separation result. Target separation signals that meet the dual threshold conditions are selected and transmitted to the next enhancement processing step. Low-quality signals that do not meet the conditions are marked and returned to the initial signal separation or iterative optimization separation step. Multi-dimensional feature extraction is performed on the selected high-quality target separation signals to construct a multi-dimensional feature parameter set. Specifically, time-domain feature parameters such as pulse amplitude, pulse width, rise time, fall time, and pulse leading edge steepness are extracted in the time domain; frequency-domain feature parameters such as the main frequency component, spectral centroid, and high-frequency component proportion are extracted in the frequency domain using Fast Fourier Transform; and wavelet packet transform is used in both the time and frequency domains to extract various... The time-domain and frequency-domain characteristic parameters of the energy entropy and instantaneous frequency change rate of the time-domain and frequency-domain sub-bands are used. Adaptive gain control technology is employed to enhance the amplitude of the multidimensional characteristic parameter set. Specifically, the energy distribution of the target separated signal is calculated based on the pulse amplitude, and the peak factor is calculated based on the peak value and root mean square (RMS). An adaptive gain factor is obtained through the energy distribution and peak factor. Based on the adaptive gain factor and the pulse amplitude, a differentiated gain strategy is applied to signal components with different amplitude characteristics to obtain an amplitude-enhanced signal. Specifically, for early discharge signal components with low amplitude, high-gain amplification is used to increase the amplitude to a preset amplitude range; for high-amplitude discharge signal components, low-gain or linear amplification is used to avoid signal saturation distortion, thus obtaining an amplitude-enhanced signal. Multi-domain feature enhancement is performed based on the amplitude enhancement signal. Specifically, in the time domain, the rising and falling edges of the pulse signal are enhanced using edge sharpening techniques, targeting the rise and fall times of the time-domain feature parameters. A high-pass filter is used to convolve the amplitude enhancement signal to obtain the time-domain enhanced signal. Frequency domain enhancement is then performed based on the time-domain enhanced signal. Specifically, the proportion of the dominant frequency component and high-frequency component in the frequency-domain feature parameters is used to enhance the characteristic frequency band of the discharge signal while suppressing residual noise frequency bands. The time-domain enhanced signal is then subjected to frequency-domain filtering, and the frequency-domain enhanced signal is obtained through inverse Fourier transform. The time-domain and frequency-domain enhanced signals are input into a trained U-Net network, outputting the final enhanced target separation signal.The final enhanced signal, through comprehensive quality screening, multi-dimensional feature extraction, adaptive amplitude enhancement, multi-domain feature enhancement, and deep learning enhancement, boasts advantages such as high signal-to-noise ratio, clear discharge characteristics, and good time-frequency localization properties.

[0011] Further, multiple cross-validation is performed on the target separated signal and the enhancement processing result. The separation effect is verified by calculating signal quality indicators such as signal-to-noise ratio (SNR), improvement ratio, and cross-correlation coefficient. The signal quality of the separation effect is then evaluated based on these signal quality indicators, and low-quality signal regions are corrected to obtain the final enhanced signal. This includes: performing multiple cross-validation on the final enhanced signal and the enhancement processing result to obtain verification results. The verification results include multiple verification scores. First, signal quality indicators are verified to obtain the SNR of the final enhanced signal, the improvement ratio of the SNR relative to the original mixed signal, and the cross-correlation coefficient with the laboratory reference pure signal. These three indicators are used as the first verification score. Based on the... The final enhanced signal is subjected to spectral characteristic verification. The spectrum is obtained through Fast Fourier Transform (FFT), and the spectral distribution is analyzed to identify residual noise frequency bands. The energy of the noise frequency bands and the total energy are obtained, yielding the noise frequency band energy percentage. When the noise frequency band energy percentage is less than a preset threshold, the noise suppression effect is considered good. Based on the noise energy percentage, a spectral purity index is obtained as the second verification score. Based on the spectral characteristic verification, a set of discharge characteristic parameters is extracted from the final enhanced signal through physical mechanism verification. This set includes key parameters such as the peak value, pulse width, and repetition frequency of the discharge pulse. These discharge characteristic parameters are matched with a physical characteristic database of known discharge types to check whether the peak value range, pulse width range, and frequency range conform to the current cable insulation state. Based on the physical mechanism verification, multi-sensor data fusion verification is performed to obtain the target separation signals collected by N sensors distributed at different monitoring points. The cross-correlation coefficient between any two signals is calculated, and a cross-correlation matrix is ​​constructed. The spatial consistency of the signal separation results is evaluated using multi-point cross-validation technology. The average cross-correlation coefficient is calculated, and when the average cross-correlation coefficient is greater than a preset threshold, the spatial consistency is considered good, and the spatial consistency index is used as the fourth verification score. Based on the multi-sensor verification, trend verification is performed in conjunction with the historical discharge database. The historical discharge characteristic parameter sequence related to the current monitoring point is extracted from the historical database, and the... The current set of discharge characteristic parameters is compared with historical sequences in a time-series analysis to obtain the slope and acceleration of the change trend of the characteristic parameters; it is checked whether the current discharge event is a continued development of a known discharge source or a newly emerging discharge defect, and the trend consistency score is obtained as the fifth verification score; the five verification scores in the verification results are assigned confidence weights {w1, w2, w3, w4, w5}, where each weight is determined according to the reliability and importance of the verification item and the sum of each weight is 1; a weighted comprehensive confidence score is obtained through the confidence weights and verification scores; the comprehensive confidence score reflects the credibility of the final enhanced signal in five dimensions: signal quality, spectral purity, physical mechanism, spatial consistency, and temporal trend;A deep quality assessment is performed based on the comprehensive confidence score to obtain the quality assessment result. Based on the quality assessment result, low-quality signal regions are identified according to the activation intensity distribution of the feature maps of each layer of the encoder and the signal quality index. Regions with activation intensity below a threshold or time-frequency points with local signal-to-noise ratios in the feature maps are marked as a set of low-quality regions. The fused features are compared with the feature representation of the original final enhanced signal to obtain the comparison result. Based on the comparison result, the correction amount is calculated through residual learning. A correction enhancement factor is applied to the low-quality regions to calculate the strengthening correction amount. When the correction enhancement factor is greater than 1, it is adaptively determined according to the degree of quality defect in the region, and the worse the quality of the region, the greater the correction intensity. For normal quality regions, the original correction amount is maintained. The fused features corrected for low-quality regions are input to the decoder's output layer. The output layer performs convolution operations using convolution kernels and maps back to the time-domain signal space after passing through an activation function. Post-processing is then applied to the time-domain signal waveform, including smoothing with a low-pass filter to remove high-frequency noise spikes, followed by boundary correction to eliminate boundary effects, resulting in the final enhanced signal. This final enhanced signal combines the advantages of multiple cross-validation, deep quality assessment, multi-scale feature fusion, and targeted correction of low-quality regions, exhibiting high signal-to-noise ratio, high spectral purity, conformity to physical mechanisms, good spatial consistency, and reasonable temporal trends. It serves as a reliable basis for diagnosing early cable discharge faults and is output to subsequent application modules for diagnostic tasks such as discharge type identification, severity assessment, and precise location of the discharge source.

[0012] A computer-readable storage medium for storing computer-readable instructions that, when read by a computer, enable the execution of a blind separation and enhancement processing method for cable discharge signals based on an adversarial network.

[0013] The beneficial effects of this invention are as follows: By employing generative adversarial networks (GANs) to achieve preliminary blind separation of mixed signals, this invention eliminates the need for prior signal information, thus solving the problem of insufficient separation capability of traditional methods under conditions of low signal-to-noise ratio and severe spectral overlap. Through covariance matrix eigenvalue decomposition and dual-criteria collaborative estimation, it achieves comprehensive identification of strong signals and weak early signal sources. Recursive optimization combining graph neural networks and variational inference improves the accuracy and robustness of signal separation. Multidimensional feature extraction and differentiated enhancement strategies effectively highlight the characteristics of early discharge signals. Finally, through five-fold cross-validation and a low-quality region targeted correction mechanism, it ensures the high fidelity and engineering reliability of the output signal, thereby achieving accurate extraction and enhancement of early partial discharge signals in cables and improving the reliability and practicality of power cable fault early warning. Furthermore, the high-quality enhanced signal output by this invention and its associated multidimensional evaluation indicators can be directly integrated into an asset health management system, supporting status visualization, risk classification, predictive maintenance work order generation, and optimized allocation of maintenance resources, significantly improving the intelligence level and economic benefits of cable asset management. Attached Figure Description

[0014] Figure 1 This is a schematic flowchart of a cable discharge signal blind separation and enhancement processing method based on adversarial networks provided in an embodiment of the present invention. Detailed Implementation

[0015] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0016] like Figure 1 As shown, at least one embodiment of the present invention discloses a method for blind separation and enhancement processing of cable discharge signals based on adversarial networks, including: The technical solution described in this application has significant application value and economic benefits in the fields of power asset management, operation and maintenance decision support, and commercial operation optimization, specifically reflected in the following aspects: First, it reduces losses from unplanned power outages and improves economic indicators of power supply reliability. By detecting cable insulation defects early and accurately, it provides early warnings and arranges planned maintenance before the fault develops into a serious breakdown, effectively avoiding large-scale unplanned power outages caused by sudden faults, significantly improving power supply reliability indicators (such as the System Average Outage Time Index (SAIDI) and the System Average Outage Frequency Index (SAIFI), and reducing direct economic losses caused by power outages (including power loss, user compensation, and socio-economic impacts). In a market-oriented power environment, unplanned power outages can lead to regulatory penalties, reputational damage, and customer loss for power supply companies. This method, by advancing the fault warning cycle by several weeks to several months, provides sufficient maintenance windows for the operation and maintenance department, transforming passive emergency repairs into proactive prevention, fundamentally changing the risk management model of cable assets. According to statistics from the power industry, reducing unplanned power outages by one hour can save power supply companies hundreds of thousands to millions of yuan in comprehensive costs (depending on load level and user type). This method is particularly effective in high-reliability scenarios such as urban core areas, industrial parks, and data centers.

[0017] Secondly, this method optimizes maintenance budgets and resource allocation to achieve refined operation and maintenance management. Based on the precise discharge location results, discharge type identification results, discharge magnitude assessment results, and insulation condition ratings output by this method, operation and maintenance management departments can transition from traditional periodic preventive maintenance to a condition-based predictive maintenance (CBM) management model. Specifically, for cables identified as normal or attenuated in the diagnostic report, the maintenance cycle can be extended or maintenance investment reduced to avoid resource waste caused by over-maintenance; for cables identified as abnormal or hazardous, maintenance resources, spare parts, and professional personnel can be prioritized, achieving precise allocation of limited resources. The three-dimensional spatial positioning information (latitude, longitude, and burial depth) output by this method can directly guide on-site operations, reducing fault location time and lowering excavation costs and traffic impact costs. Furthermore, based on trend analysis of historical diagnostic data and machine learning prediction models, the method enables the scientific preparation of annual maintenance plans, optimized management of spare parts inventory (reducing inventory backlog and stockout risks), and rational scheduling of maintenance teams, thereby optimizing maintenance budget allocation and improving the efficiency of operation and maintenance fund utilization. According to statistics from practical cases in power companies, this method can reduce cable maintenance costs by 20%-40%, increase spare parts inventory turnover by more than 30%, and increase maintenance personnel utilization by more than 25%, creating significant cost savings and management benefits for enterprises.

[0018] Third, extending the lifespan of cable assets enhances their overall lifecycle value. Power cables, as high-value fixed assets (the cost of a single high-voltage cable can reach millions to tens of millions of yuan), directly impact a company's return on assets and depreciation policy. Under traditional maintenance methods, weak early discharge signals are difficult to detect in time, and the continued development of insulation defects eventually leads to permanent damage and cable scrapping, resulting in a significantly shorter actual cable lifespan than the design life (typically 30-40 years). This method achieves effective detection in the early stages of insulation degradation (such as the initial development stage of electrical treeing and the micro-discharge stage with partial discharge levels <100pC), creating conditions for timely intervention measures (such as local insulation reinforcement, reducing operating voltage, and improving the operating environment), slowing down the insulation aging process, and preventing defects from developing from a reversible state to irreversible permanent damage. Based on insulation aging theory and engineering experience, early intervention can extend cable life by 5-10 years, equivalent to reducing asset replacement investment by 20%-30%. For large-scale cable networks (such as urban distribution networks containing thousands to tens of thousands of kilometers of cables), the asset preservation and appreciation benefits brought by this lifespan extension effect can reach hundreds of millions to billions of yuan. In addition, extending asset life can help optimize corporate financial statements (reduce depreciation expenses and increase net asset value), improve debt-to-equity ratio, and enhance financing capabilities, which has a positive impact on the financial health and business reputation of power companies.

[0019] Fourth, this method enhances the intelligent management level of the power grid and supports the digital transformation strategy. It can achieve data interconnection and business integration with existing information platforms of power companies, such as SCADA (Supervisory Control and Data Acquisition), PMS (Production Management System), EAM (Enterprise Asset Management System), and GIS (Geographic Information System), to build an integrated management platform for panoramic cable status perception, intelligent diagnosis, and decision support. Specifically, the diagnostic reports output by this method (containing structured data such as fault type, location coordinates, risk level, confidence level, and maintenance recommendations) can be automatically pushed to the asset management system through standardized interfaces (such as RESTful API and OPC UA protocol), triggering business processes such as automatic work order generation, automatic maintenance task assignment, and automatic material requisition. This achieves fully automated closed-loop management from signal acquisition and intelligent diagnosis to business response, significantly improving management efficiency and response speed. Meanwhile, historical diagnostic data can be stored in the enterprise data warehouse or data lake, supporting big data analysis, business intelligence (BI) report generation, and management dashboard visualization. This provides management with decision support information such as an overview of cable asset health, fault trend prediction, maintenance effectiveness evaluation, and equipment manufacturer quality comparison, driving the transformation of power grid asset management from experience-driven to data-driven, and from passive response to proactive prediction, thereby enhancing the enterprise's core competitiveness in the digital age. Furthermore, this method can also provide power regulatory authorities with power supply reliability monitoring data, insurance companies with asset risk assessment basis, and equipment manufacturers with product quality feedback, demonstrating broad cross-industry application value.

[0020] Fifth, it supports standardized applications across multiple scenarios, reducing cross-regional deployment costs. This method has been validated through practical applications in various complex environments, including urban cable tunnels, direct-buried cable lines, submarine cables, rail transit traction cables, wind farm collector lines, and dedicated cables in industrial parks. It demonstrates excellent scenario adaptability and robustness under different voltage levels (10kV-500kV), different laying methods (direct burial, pipeline, tunnel, overhead, underwater), different insulation types (XLPE, oil-paper, rubber), and different noise environments (strong electromagnetic interference, high humidity, high temperature). This method employs standardized data interfaces, modular software architecture, and a universal hardware platform, enabling rapid replication and deployment across regions and voltage levels, significantly reducing the costs of repetitive development and customized implementation for enterprises in different business scenarios. For large power groups operating across provinces and countries, this method supports unified technical standards, unified data specifications, and unified diagnostic processes, facilitating standardized technical management, shared operation and maintenance experience, and centralized data assets at the group level, thereby improving the overall management level and collaborative efficiency of the group. Furthermore, standardized deployment reduces personnel training costs, spare parts commonality costs, and system maintenance costs, creating favorable conditions for large-scale enterprise application. From a business promotion perspective, the standardized nature of this method also provides a foundation for business model innovations such as technology output, product sales, and service-oriented operations, demonstrating promising market prospects and commercial potential.

[0021] Sixth, this method promotes the upgrading of fault diagnosis and location technologies, empowering advanced analytical applications. The high-quality, clean discharge signals output by this method (after blind separation and enhancement processing, the signal-to-noise ratio is improved by 30-50dB, and waveform distortion is <5%) provide a reliable data foundation for subsequent advanced diagnostic tasks. Specifically, the clean signals can be directly used for high-value diagnostic functions such as accurate discharge type identification (achieving accurate differentiation of tip discharge, surface discharge, internal discharge, and floating discharge through a deep learning classifier, with an accuracy rate >95%), precise three-dimensional spatial location of discharge sources (positioning error ±5 meters), quantitative assessment of insulation status (outputting insulation aging degree scores and remaining life prediction ranges), and discharge development trend prediction (predicting the discharge magnitude change trajectory over the next 3-12 months based on time series analysis). This provides strong technical support for comprehensive cable health status assessment, in-depth fault root cause tracing, and scientific formulation of maintenance strategies. Based on this, advanced analytical tools such as cable asset health index models, cable remaining life prediction models, and cable fault probability assessment models can be further constructed to support risk-based asset management decisions, prioritize the allocation of maintenance resources to high-risk assets, and maximize the return on investment (ROI) of asset management. Furthermore, this method promotes the upgrading of fault diagnosis technology from traditional threshold judgment and expert experience to advanced technologies such as deep learning, artificial intelligence, and big data analysis, which facilitates technological progress and industrial upgrading in the field of power equipment condition monitoring. It is of great strategic significance for improving the intelligence level of my country's power equipment and enhancing the international competitiveness of the power industry.

[0022] In summary, the technical solution described in this application not only solves the technical problem of effectively detecting early discharge signals of cables in complex noise environments, but also creates significant management value and economic benefits in areas such as full life-cycle management of power assets, scientific operation and maintenance decision-making, optimization of business operations, and digital transformation of enterprises. It is an intelligent data processing system and method specifically applicable to administrative management, business operations, financial optimization, asset supervision, and risk prediction in the power industry, and has broad application prospects and promotional value.

[0023] When an abnormal cable discharge signal is detected, blind separation and enhancement processing are performed to obtain a detectable cable discharge signal, which is then used for the analysis and evaluation of the current signal state. Specifically:

[0024] Step 1: Acquire the mixed signal of the cable, and perform preprocessing and normalization to construct the signal set to be separated;

[0025] The acquisition of mixed signal and environmental data of the cable refers to the collaborative collection of raw observation signals by multiple types of sensors deployed at key nodes of the cable line (such as joints, terminals, and intermediate sections). Ultra-high frequency (UHF) sensors are used to capture electromagnetic wave signals radiated during partial discharge. Their frequency response range is typically set between 300MHz and 3GHz, effectively capturing transient high-frequency pulse components generated in the early stages of discharge. Current sensors, using high-frequency current transformers (HFCTs) installed on the cable grounding wire, are used to detect partial discharge pulse current flowing through the grounding wire, exhibiting good resistance to power frequency interference and high sensitivity. The partial discharge detection device integrates pulse detection, amplification, and digitization functions to achieve preliminary extraction and recording of weak discharge signals. These three types of sensors simultaneously acquire signals from electromagnetic field, conduction current, and comprehensive characteristic dimensions, respectively, forming a multi-channel mixed signal input. Temperature and humidity sensors are used to monitor changes in temperature and relative humidity in the environment where cable joints or terminals are located. Fluctuations in temperature and humidity can alter the dielectric properties of insulation materials, thus affecting the propagation characteristics of discharge signals and background noise levels. Electromagnetic field strength detectors are used to quantify the amplitude of low-frequency electromagnetic interference in the surrounding space (such as 50Hz and its harmonic interference caused by nearby power equipment operation), assisting in the identification of external electromagnetic pollution sources. All sensors are equipped with high-precision time synchronization modules (such as clock synchronization units supported by GPS or the IEEE 1588 protocol) to ensure time alignment accuracy of data from different sources is better than 1μs, used for dynamic compensation of noise characteristics in subsequent modeling. Specifically, acquiring mixed signal and environmental data from the cable, and performing preprocessing and normalization to construct the dataset to be separated refers to the process of acquiring raw observation data through multi-source sensor collaboration during cable operation, considering the characteristics of weak partial discharge signals, susceptibility to noise, and significant influence from environmental factors. This data is then systematically preprocessed and structured to form a standard data format suitable for the input requirements of deep learning models. This process provides a high-quality, highly consistent data foundation for subsequent blind separation and enhancement methods based on generative adversarial networks, ensuring the stability and generalization ability of model training. Let the mixed signal data be denoted as x. i Where i represents the sampling time, x i It contains weak partial discharge pulse signals, strong background noise, and electrical transient signals; among which, x iThis represents the multi-channel mixed signal vector acquired at the i-th sampling time, where each dimension corresponds to the real-time sampled value of a sensor channel. This signal vector mainly contains three types of components: first, partial discharge pulse signals excited by defects inside the cable (such as air gaps, impurities, and concentrated electric field areas at the ends), characterized by nanosecond-level rise times, short durations (tens to hundreds of nanoseconds), and weak amplitudes (often below a few millivolts); second, strong background noise, including continuous power frequency interference (50 / 60Hz and its harmonics), white noise, radio frequency interference, and broadband electromagnetic noise caused by switching operations, whose energy is often much higher than the discharge signal; and third, electrical transient signals, such as voltage / current transients caused by circuit breaker opening and closing, and sudden load changes, whose waveforms are complex and may overlap with the discharge signal in the time domain, causing confusion. Therefore, x i Essentially, it is a highly aliased, non-stationary, and low signal-to-noise ratio composite signal. Preprocessing of the mixed signal data includes using bandpass filtering to remove DC and ultra-high frequency components, normalization to unify the signal amplitude range, and timestamp alignment to achieve time synchronization between the mixed signal and environmental data. The bandpass filter is designed based on the main frequency band distribution characteristics of the partial discharge signal, typically with a passband range of 10kHz to 100MHz, to filter out DC drift and extremely low-frequency interference below 10kHz, as well as irrelevant RF noise or system noise above 100MHz, while retaining the main energy concentration area of ​​the discharge pulse. The filtering method can be a digital infinite impulse response (IIR) or finite impulse response (FIR) filter, with the order and delay characteristics selected based on the actual hardware platform resources. Normalization uses minimum and maximum scaling or Z-score normalization to map each channel signal to a uniform numerical range (e.g., [-1,1] or [0,1]), avoiding the dominance of certain channels in the model training process due to differences in sensor sensitivity. Timestamp alignment aligns the time stamp headers of all sensor output data packets with a unified time base (UTC time or local synchronous clock). If slight asynchrony exists, linear interpolation or spline interpolation methods are used for resampling compensation to ensure strict correspondence between the mixed signal and environmental parameters within the same time window. The normalized mixed signal data is divided into multiple signal segments according to a fixed time window, with each segment serving as an independent sample. Environmental data is mean-valued according to the corresponding time window to form environmental feature vectors paired with the signal segments. The fixed time window length can be set based on a trade-off between signal periodicity and computational efficiency, typically between 10ms and 1s, which covers multiple potential discharge events while facilitating batch training. For example, setting the window length to 100ms and the sliding step size to 50ms (i.e., 50% overlap) divides a long continuous signal into a series of short segments {S1, S2, ..., S...}. nEach segment serves as an independent training sample for subsequent modeling. For environmental parameters (such as temperature T, humidity H, and electromagnetic field strength E), due to their slow changes, their arithmetic mean is taken within each time window to generate a corresponding environmental feature vector. This vector is then paired with the k-th signal segment and stored to form a data pair of signal segment environmental context for subsequent joint modeling and analysis. Multi-channel conversion is performed on the signal segments, transforming the time-domain waveform of each segment into a frequency-domain spectrum. This spectrum is then combined with the environmental feature vector to generate a joint feature matrix. The joint feature matrix has a dimension of 4×m, where m is the number of sampling points in the signal segment. Multi-channel conversion refers to expanding the original time-domain signal from a single representation to a multi-modal joint representation. Specifically, for each signal segment Si, firstly, its original time-domain waveform is used as the first channel; secondly, a fast Fourier transform is performed on the segment to obtain its amplitude spectrum as the second channel; thirdly, its short-time Fourier transform (STFT) time-frequency spectrum is calculated and integrated along the time axis to obtain the time-frequency energy distribution curve as the third channel; finally, combined with the environmental feature vector generated in the previous steps, its three scalar elements are expanded into a constant sequence of length m, which is then superimposed on the matrix as the fourth channel.

[0026] This design enables the deep learning model to simultaneously perceive the temporal dynamics, frequency structure, time-frequency evolution trend, and external environmental conditions of the signal during the same forward propagation process, enhancing the model's robustness to external interference and its ability to distinguish real discharge signals. All joint feature matrices are combined to form the dataset to be separated, denoted as X=[x1, x2, ..., x...]. n ], where n is the total number of signal segments. The dataset X to be separated is the joint feature matrix [x1, x2, ..., xn] after all the aforementioned processing. nThe dataset, arranged in sequence, serves as the input tensor for the generative adversarial network (GAN). This dataset not only preserves the key dynamic information of the original signal but also integrates spectral features and environmental context, forming a high-dimensional input space rich in semantic information. In practical applications, this dataset can be divided into training, validation, and test sets, typically in a ratio of 7:2:1 or 8:1:1, for model training and performance evaluation. This application achieves systematic acquisition, cleaning, alignment, and multimodal fusion of mixed signals from partial discharge in cables and environmental parameters, constructing a structurally sound and information-rich dataset to be separated. The use of a multi-source sensor collaborative acquisition mechanism solves the problem of interference susceptibility in traditional single-channel detection; the introduction of environmental feature vectors and the construction of a joint feature matrix enables the model to distinguish between environmentally induced noise and real discharge signals; and the use of segmentation and normalization improves data processing efficiency and model training stability. Therefore, this method significantly improves the quality and usability of the input data, providing solid data support for subsequent blind separation and enhancement based on GANs, and effectively addressing the technical challenge of signal separation under low signal-to-noise ratio conditions.

[0027] Step 2: Based on the set of signals to be separated, construct a generative adversarial network;

[0028] This study simulates four different types of partial discharge—point discharge, surface discharge, internal discharge, and floating discharge—under a controlled laboratory environment. By setting up typical defect models, a gradually increasing AC voltage is applied in a high-voltage test system to induce specific types of partial discharge phenomena: point discharge is achieved by placing a metal needle-shaped electrode near the insulating material; surface discharge is achieved by creating contaminated or damp areas on the insulator surface to form surface creepage paths; internal discharge is excited by pre-fabricated air gaps or impurity cavities within the solid insulating medium; and floating discharge simulates the potential floating state caused by ungrounded metal components. Pure signal samples of these discharge types are collected under shielding and low-noise conditions to construct a pure discharge signal sample library covering the most common early fault modes in cable systems. Pure noise samples containing power frequency interference, electromagnetic field interference, and white noise under different environmental conditions are collected, and their spectral characteristics and probability distributions are analyzed to construct a noise sample library. Pure discharge signals are superimposed with various types of noise according to different signal-to-noise ratios and mixing ratios to generate mixed signal samples with known components. All samples are classified and stored according to multiple dimensions such as discharge type, noise type, signal-to-noise ratio, and environmental conditions. Each sample is further characterized by time-domain waveforms, frequency-domain spectra, and time-frequency plots to construct a standardized sample library. Based on the pure discharge signal sample library and the noise sample library, a generative adversarial network (GAN) architecture is designed. The generator adopts an encoder-decoder architecture, and the discriminator adopts a deep convolutional neural network architecture. Through adversarial training, the generator learns the mapping relationship for extracting pure discharge signals from the mixed signals. During training, the network parameters of the generator and discriminator are alternately optimized until the weighted sum of the adversarial loss and reconstruction loss reaches the preset convergence condition. The collected pure discharge signal samples must be collected in an electromagnetically shielded room to minimize external electromagnetic interference. Simultaneously, a low-noise preamplifier is used in conjunction with an ultra-high frequency sensor or a high-frequency current transformer to capture weak pulse signals at a sampling rate of no less than 500 MS / s to ensure complete preservation of the transient characteristics of the discharge signal. After detrending processing, the collected raw data is labeled with discharge type, applied voltage level, temperature, humidity, and other environmental parameters, and stored in a database in a unified format, forming an initial clean discharge signal sample library. For the construction of the noise sample library, a combination of on-site measurement and controlled injection is used. Power frequency interference mainly originates from the power system's own carrier communication, transformer excitation harmonics, and coupling interference from nearby operating equipment, and can be acquired through long-term monitoring using sensors deployed in substations or distribution rooms. Electromagnetic field interference comes from non-power frequency sources such as wireless communication base stations, transient switching operations, and frequency converters, and is collected using broadband electromagnetic field probes in typical industrial environments. White noise, as an idealized random process model, can be digitally synthesized to generate random sequences conforming to a Gaussian or generalized stable distribution, supplementing training needs in extremely low signal-to-noise ratio scenarios.Each noise sample undergoes spectral analysis (e.g., power spectral density (PSD) estimation) and statistical characteristic modeling (e.g., kurtosis, skewness, autocorrelation function), and its source environment, intensity level, and duration are labeled for later on-demand retrieval. Furthermore, to generate mixed signal samples suitable for supervised training, the aforementioned pure discharge signals are dynamically superimposed with various noise types according to a preset signal-to-noise ratio range (e.g., -10dB to 10dB) and mixing ratio. For example, in a mixed sample, the main amplifier power supply can be set to internal discharge, and the background noise can consist of 60% power frequency interference + 30% electromagnetic interference + 10% white noise, with an SNR of −5dB. This process supports multiple combination strategies, thereby constructing a highly diverse set of "mixed-pure" paired samples, providing a clear learning objective for the generative adversarial network. Each mixed signal sample not only preserves its original time-domain waveform but also simultaneously calculates its frequency domain spectrum (via Fast Fourier Transform (FFT)) and time-frequency plot (e.g., Short-Time Fourier Transform (STFT) or Continuous Wavelet Transform (CWT)), among other representations. This multi-dimensional representation enables the model to capture signal features in different domains, enhancing its ability to perceive weak discharge signals against complex noise backgrounds. All samples are organized according to a four-dimensional labeling system: discharge type (category), noise type combination, signal-to-noise ratio range, and environmental conditions (temperature, humidity, electromagnetic background), and are finally integrated into a structured, standardized sample library S=[s1, s2, ..., s]. k [ ], where k is the total number of samples, typically greater than 100,000, to meet the large data volume requirements of deep learning. Based on this, a dedicated generative adversarial network architecture is designed for blind separation tasks. The generator adopts an encoder-decoder structure, where the encoder consists of multiple convolutional layers, progressively extracting the spatial-temporal features of the input mixed signal and compressing them into the latent space; the decoder then gradually restores the signal details through transposed convolution or upsampling operations, outputting a preliminarily separated discharge signal. This structure helps preserve signal edge information and pulse morphology, avoiding excessive smoothing.

[0029] The discriminator employs a deep convolutional neural network architecture, specifically a stack of ResNet, DenseNet, or Inception modules, to determine whether the generator's output signal closely approximates the distribution of a real, pure discharge signal. The discriminator receives two types of input: real, pure signals from a sample library and pseudo-signals from the generator. A binary classification loss guides the generator to continuously optimize its output quality. During adversarial training, the parameters of the generator and discriminator are alternately optimized. In the discriminator training phase, batches of real, pure discharge signals and generator output signals are input to train the discriminator to accurately distinguish between the two. In the generator training phase, a mixed signal is input to the generator, and the generator's output is fed into the discriminator. The generator is trained using a weighted sum of adversarial and reconstruction losses, enabling it to extract the essential features of the discharge signal from the mixed signal. An attention mechanism and conditional input technique are introduced, allowing the generator to adaptively adjust its separation strategy based on the temporal characteristics of the input signal and environmental parameters until a preset convergence condition is met. Specifically, the generator employs an encoder and decoder structure; the encoder extracts multi-scale features of the input signal through a multi-layer convolutional neural network; the decoder restores the signal to the time domain layer by layer through transposed convolution; the discriminator adopts a deep convolutional neural network architecture, extracts the discriminative features of the signal through multi-layer convolution and pooling operations, and outputs a probability value representing the authenticity of the input signal; during adversarial training, the parameters of the generator and discriminator are alternately optimized. In the discriminator training phase, a batch of real pure discharge signals and a batch of generated signals output by the generator are input to train the discriminator to accurately distinguish between them. In the generator training phase, a mixed signal is input to the generator and the generator output is sent to the discriminator. The generator is trained by weighting the adversarial loss and reconstruction loss to extract the essential features of the discharge signal from the mixed signal; an attention mechanism and conditional input technology are introduced to enable the generator to adaptively adjust the separation strategy according to the time domain features of the input signal and environmental parameters until the preset convergence condition is reached. This application achieves the following: A high-quality training sample library with precise annotations and broad coverage of operating conditions is constructed through systematic collection of laboratory simulated discharges and real-world environmental noise; data availability and model understanding capabilities are improved by combining multi-dimensional signal representation and structured classification management; a GAN architecture based on encoder-decoder and deep convolutional discriminator is designed, and the generator's ability to reproduce weak signal features is enhanced through adversarial training mechanisms. By employing a controlled experimental and real-world environmental data fusion modeling method, the problem of separation failure under low signal-to-noise ratio conditions caused by the lack of effective prior knowledge in traditional methods is solved. This significantly improves the accuracy and fidelity of early discharge signal identification in cable operating environments with severe spectral overlap and complex noise, thus providing a reliable data foundation for subsequent fault diagnosis.

[0030] Step 3: Input the set of signals to be separated into the generative adversarial network, and output the corresponding preliminary separation signal;

[0031] Furthermore, the mixed signal segments from the dataset to be separated are input one by one into the trained generative adversarial network (GAN). The GAN performs multi-layer convolution operations on the input mixed signal through the generator's encoder, extracting multi-scale features including low-frequency basic features, mid-frequency texture features, and high-frequency detail features at different levels. The mixed signal segment refers to a time-series signal segment obtained from the original cable monitoring data by dividing it into fixed time windows. Its length can be flexibly set according to the actual sampling rate and signal duration, for example, it can be 1024 or 2048 sampling points. This signal segment serves as the input unit of the GAN, entering the trained generator module. The generator adopts an encoder-decoder architecture, where the encoder consists of multiple stacked convolutional layers. The kernel size of each layer is typically set to 3×3 or 5×5, with a stride of 1 or 2. Combined with batch normalization and activation functions (such as LeakyReLU), it achieves a layer-by-layer abstract representation of the input signal. During the encoding process, shallow convolutional layers primarily capture high-frequency detail features of the signal, such as the steepness of the rising edge of partial discharge pulses and subtle structures like transient transition points. Intermediate layers extract mid-frequency texture features, reflecting the distribution pattern, repetition frequency, and waveform consistency of the pulse group. Deeper networks focus on low-frequency fundamental features, characterizing the overall trend of the signal, energy concentration areas, and potential periodic components. This hierarchical extraction mechanism enables the model to simultaneously perceive key discriminative information of weak discharge signals and the statistical characteristics of background noise, forming a multi-layered fingerprint characterization of the mixed signal. Multi-scale feature extraction can be further enhanced by dilated convolution or pyramid pooling structures to improve robustness against broadband interference. Multi-scale features are fused in the feature space to generate a fused feature vector. This fused feature vector contains separate representations of the discharge and noise signals in the mixed signal. The separate representation represents the compressed encoding form of the original time-domain signal in a high-dimensional feature space, with a dimension smaller than the original signal dimension, preserving the key semantic information and discriminative features required for separation. Multi-scale feature fusion can be achieved through channel concatenation, weighted summation, or attention mechanisms. For example, an adaptive spatial-channel attention module (such as CBAM) can be used to dynamically allocate feature weights for each layer, highlighting feature channels and temporal locations sensitive to discharge signals. The fused high-dimensional features are fed into a bottleneck layer, where they are compressed into a low-dimensional vector, i.e., the fused feature vector, through a fully connected layer or global average pooling operation. This vector has a significantly lower dimension than the original signal, for example, it can be compressed to 1 / 8 to 1 / 16 of the original length, but retains the core semantic information of the discharge signal, such as pulse polarity, amplitude range, and temporal clustering, forming the basic representation for subsequent signal reconstruction. This compression process is essentially a nonlinear dimensionality reduction, which can remove redundant noise while maintaining the distinguishability of discharge events.The fused feature vector is input to the decoder, which performs layer-by-layer upsampling through transposed convolution operations, outputting an abstract representation of the feature space. This abstract representation represents the compression, refinement, and high-level generalization of the original mixed signal in the high-dimensional feature space, retaining the essential discriminative features and key information of the discharge signal and noise signal while discarding redundant details, expressing the information structure of the original signal in a low-dimensional way. The decoder consists of multiple transposed convolutional layers (also called deconvolutional layers), each performing an upsampling operation to gradually restore the spatial resolution of the signal. The size of the transposed convolutional kernel is generally consistent with the corresponding layer of the encoder, with a stride of 2 to achieve double upsampling. Combined with batch normalization and the ReLU activation function, stable gradient propagation is ensured. In this process, the fused feature vector is first reshaped into a suitable tensor shape, and then expanded layer by layer to generate a series of progressively refined intermediate feature maps. These feature maps together constitute the "abstract representation," a highly refined version of the original mixed signal in the high-dimensional latent space. Although it no longer directly corresponds to the original waveform, it contains the essential structure of the discharge signal and the direction of noise suppression, serving as a crucial intermediate state for achieving accurate separation. The abstract representation is mapped back from the high-dimensional feature space to the original time-domain signal space, gradually restoring it to a time-domain signal waveform. This mapping process is achieved through a final layer of transposed convolution or ordinary convolution, with the output channel number set to 1 (single-channel signal). A linear function or Tanh is used as the activation function to ensure continuous output signal amplitude and no nonlinear distortion. The final output time-domain waveform serves as the candidate result for the initial signal separation, with a duration consistent with the input signal segment. This mapping process achieves a regression from the semantic compression space to the physical signal space, fulfilling the core function of denoising translation. During decoding, skip connections are used to concatenate the features of the corresponding layer of the encoder with the features of the current layer of the decoder, preserving the signal's detailed information. A skip connection directly transmits the feature map output from the i-th layer of the encoder to the (n-1)-th layer of the decoder and concatenates it with the upsampled result of that layer along the channel dimension. This design effectively alleviates the information attenuation problem in deep networks, especially helping to preserve high-frequency details, such as the spikes and steep edges of discharge pulses—key features easily lost during downsampling. By introducing shallow details, the decoder can accurately reconstruct local abrupt changes while restoring the overall contour, avoiding feature blurring caused by signal smoothing. This structure, inspired by the U-Net architecture, has been proven to have excellent detail preservation capabilities in image segmentation tasks and is suitable for the fine reconstruction of weak pulse signals in this application.

[0032] The time-domain signal waveform is smoothed and edge-effects removed to obtain a preliminary separated signal. Boundary effects arise from the lack of sufficient context at signal edges during convolution operations, often manifesting as waveform oscillations or distortions at the beginning and end. To eliminate these artifacts, mirror extension preprocessing of the input signal can be used, or the central effective region can be truncated after output (e.g., removing 5% of the initial and final sampling points). Furthermore, applying a lightweight smoothing filter (such as a Savitzky-Golay filter or moving average filter) to the output waveform can further suppress high-frequency glitches without affecting the pulse's main structure. The processed signal is the "preliminary separated signal," whose main component is the estimated discharge signal component, with noise components significantly suppressed. The preliminary separated signal identifies and extracts the discharge signal component from the mixed signal and suppresses noise components; the quality of the preliminary separated signal depends on the generator's learning ability. With sufficient training, the generator learns to prioritize outputting components in the mixed signal that conform to the statistical characteristics of the discharge signal (e.g., narrow pulses, double-exponential decay waveforms, energy concentration in specific frequency bands, etc.), while suppressing components with white noise, power frequency harmonics, or random interference characteristics as noise. This process requires no prior signal source model and is a data-driven blind separation method, particularly suitable for applications with unknown discharge types and complex noise environments. A sliding window strategy is used to segment long-term time-series signals, with each window having multiple sampling points and a preset overlap rate between windows. The generator outputs a preliminary separated signal segment from the mixed signal of each window. A weighted average method is used to fuse the multiple separation results in the overlapping areas to obtain a complete and continuous preliminary separated signal. The sliding window strategy is used to handle long-term signals exceeding the model input length limit. The window length can be set according to hardware resources and real-time requirements, typically 1024-4096 points; the overlap rate is generally set to 30%-50% to ensure sufficient information redundancy between adjacent segments. For each signal segment within a window, the generator independently outputs the corresponding preliminary separation result. When reconstructing the complete signal, a Hanning window or triangular window weighted average is used for the overlapping areas to ensure a smooth and natural fusion transition, avoiding artificial breakpoints or amplitude jumps introduced by segmentation. This strategy achieves seamless processing of monitoring signals from cables of arbitrary length, ensuring the applicability and stability of the system during long-term operation. This application achieves the construction of a low-dimensional implicit representation rich in discriminative information by extracting and fusing multi-scale features from mixed signals; it realizes end-to-end mapping from noisy signals to discharge signals by leveraging an encoder-decoder architecture and a skip connection mechanism; and it overcomes the limitation of model input length by combining a sliding window and an overlapping weighted fusion strategy to obtain continuous and stable preliminary separation signals. This method fully utilizes the powerful nonlinear modeling capabilities of deep neural networks, effectively improving the detectability of early discharge signals in cables under low signal-to-noise ratio environments without requiring prior knowledge, providing a high-quality input foundation for subsequent refined separation and enhancement.

[0033] Step 4: Perform eigenvalue decomposition on the preliminary separated signal to obtain an estimate of the number of signal sources in the mixed signal. Further, construct a covariance matrix for the preliminary separated signal and perform eigenvalue decomposition to obtain the theoretical boundary of the noise eigenvalues. Analyze the differences between the signal subspace and noise subspace in the eigenvalue sequence to obtain an estimate of the number of signal sources in the mixed signal. This method includes converting the preliminary separated signal into a multi-channel time-series matrix; each row of the multi-channel time-series matrix represents a multi-channel observation at a given time, and each column represents the time-series data of one channel. The construction of the multi-channel time-series matrix is ​​the foundation for signal subspace modeling. Essentially, it expresses continuous signal observations in the time dimension in matrix form, facilitating subsequent statistical feature extraction. For example, in a cable partial discharge monitoring system, if signals are simultaneously acquired by multiple UHF sensors, each sensor channel corresponds to a column of the matrix, and the sampling time points constitute the data arrangement in the row direction. This matrix can be represented as... ,in For the length of time, For the number of channels, The number field is real to ensure sufficient sample support for subsequent covariance analysis. A covariance matrix is ​​constructed based on the multi-channel time series matrix to obtain the covariance between any two channels; the covariance matrix is ​​a symmetric positive definite matrix with dimensions equal to the number of channels. The covariance matrix... This matrix reflects the linear correlation between signal channels. Its diagonal elements represent the autocovariance (i.e., energy) of each channel, while the off-diagonal elements reflect the degree of cross-correlation between different channels. Due to spatial coupling effects and propagation delays in real signals, this matrix can capture the response patterns of potential signal sources on multiple sensors. The matrix is ​​symmetric and positive semidefinite, and its stability can be guaranteed in numerical calculations through singular value decomposition or Cholesky decomposition. To improve robustness, the original data can be mean-reduced before construction, and a sliding window averaging strategy can be used to reduce the impact of transient interference. Eigenvalue decomposition of the covariance matrix yields the eigenvalue sequence. Where m is the number of channels, the energy proportion of each eigenvalue is calculated based on the eigenvalue sequence to obtain the eigenvalue energy sequence. Here, eigenvalue decomposition... Obtain the orthogonal eigenvector matrix and the set of eigenvalues ​​arranged in descending order , `<conjugate>` is the conjugate transpose operator, representing the conjugate transpose of matrix U, UH. Each eigenvalue... This represents the signal energy intensity along the direction of the corresponding eigenvector, with the energy percentage being... ,in, Let j be the j-th eigenvalue in the eigenvalue sequence, forming a normalized energy distribution sequence. Strong signal sources typically correspond to larger eigenvalues, while noise-dominated subspaces exhibit multiple smaller and closer eigenvalues. This step realizes the mapping from the original observation space to the principal component space, providing a mathematical basis for distinguishing between signals and noise.

[0034] Based on the eigenvalue sequence and the number of rows and columns of the multi-channel time-series matrix, the probability density values ​​of noise eigenvalues ​​are calculated. The upper and lower boundaries of the noise eigenvalues ​​are obtained through these probability density values. By comparing the relationship between each eigenvalue in the eigenvalue sequence and the upper and lower boundaries of the noise, and combining this with the eigenvalue energy sequence, the signal subspace and noise subspace are distinguished. Specifically, a Marchenko-Pastur distribution model—the eigenvalue distribution under the pure noise assumption—is established using random matrix theory. Its shape is determined by the number of sampling points. With the number of channels The ratio determines this. By fitting this theoretical distribution, the expected upper and lower bounds of the eigenvalues ​​in the case of noise alone can be derived. When the measured eigenvalues ​​exceed this range, they are considered to likely originate from a real signal source. This boundary can be used to initially screen eigenvalues ​​that significantly deviate from the noise distribution as a candidate set for the signal subspace. Furthermore, combining energy proportion information can further eliminate pseudo-signal components with low energy but slightly above the boundary, improving the accuracy of discrimination. Based on the upper and lower boundaries of the noise eigenvalues, the differences between the signal subspace and the noise subspace in the eigenvalue sequence are analyzed, the interval between adjacent eigenvalues ​​is calculated, and an eigenvalue interval sequence is constructed. The eigenvalue interval is defined as... ,in, Let i be the i-th eigenvalue in the eigenvalue sequence arranged in descending order. For the (i+1)th feature value, the interval is... This interval sequence is used to quantify the energy difference between adjacent principal components. Under ideal white noise, eigenvalues ​​should be densely distributed with uniform spacing; however, the presence of discrete signal sources will cause significant "faults" at specific locations, leading to sudden increases in local spacing. This interval sequence effectively reveals structural anomalies hidden within a continuous energy decay trend, and is particularly suitable for detecting early discharge signal sources with weak energy but independent spatial patterns. By smoothing and filtering the interval sequence and performing peak detection, potential jump points can be identified. An energy gradient decision criterion is constructed based on the eigenvalue energy sequence. The energy gradient is obtained, and a dynamic energy threshold is set. When the energy gradient is greater than the dynamic energy threshold, it indicates an energy jump between the current eigenvalue and the eigenvalue at the next time step. From the perspective of energy distribution, the eigenvalue at the current time corresponds to the first signal component; the first signal component is the energy-dominant signal component, representing the main discharge signal with a large energy proportion in the mixed signal. The energy gradient is defined as the rate of change of the proportion of adjacent energy. ,in, This represents the energy percentage of the (m+1)th channel, reflecting abrupt changes in energy concentration. Dynamic threshold. It can adaptively adjust according to the signal-to-noise ratio of the current environment, for example, by using ,in, The global gradient standard deviation, This is the sensitivity coefficient (e.g., 1.5-2.5). When At that time, the judgment of the first Each eigenvalue belongs to the signal subspace, and the corresponding main discharge signal (such as internal air gap discharge) dominates in the covariance domain due to its strong amplitude. This criterion is good at capturing high-energy signal sources, but it is easily triggered by large noise fluctuations at extremely low signal-to-noise ratios.

[0035] An interval decision criterion is constructed based on the eigenvalue interval sequence. An interval threshold is set. When the median of the eigenvalue intervals is greater than the preset interval threshold, the eigenvalue at the current moment is determined to correspond to the second signal component from the perspective of interval distribution. The second signal component is an interval abrupt signal component, representing a minor discharge signal or a weak early discharge signal with a small energy proportion in the mixed signal but exhibiting interval jump characteristics in the eigenvalue distribution. The interval threshold can be set as the 95th quantile of the theoretical noise interval distribution or obtained based on statistical learning from historical data. This criterion focuses on the sparsity of the eigenvalue spectrum rather than its intensity. Therefore, even if a discharge signal has weak energy (such as weak surface discharge caused by initial insulation aging), as long as it has an independent propagation path or excitation mode in space, it will still cause observable interval jumps in the eigenvalue sequence. This mechanism is particularly suitable for identifying early defect signals that are difficult to detect using traditional energy detection methods, enhancing the system's early warning capability. The energy gradient decision criterion identifies strong discharge signal sources, while the interval decision criterion identifies weak discharge signal sources. These two criteria complement each other to comprehensively capture signal sources of varying intensities. The decision results of the energy gradient and interval decision criteria are assigned confidence weights wa and wb, respectively, and the weighted decision values ​​are then fused. The number of eigenvalues ​​satisfying the decision conditions is counted to obtain an estimate of the number of signal sources. The weights are... and It can be dynamically configured according to the application scenario, such as appropriately increasing the power level in a strong interference environment. To enhance sensitivity to weak signals; and to prioritize trust when multiple strong sources are known to exist. The fusion method can employ hard-decision counting or soft-decision integration (such as weighted score accumulation), ultimately outputting an integer estimate of the number of signal sources. This estimated value serves as a key prior parameter for subsequent graph neural network modeling, directly influencing the node count and topology construction. As an optional implementation, a Bayesian inference framework can be introduced to probabilistically fuse the two pieces of evidence, further enhancing the reliability of the estimation. This application achieves a high-precision estimation of the number of real signal sources in the initial separation signal. By employing an eigenvalue analysis method based on the covariance matrix and innovatively proposing a dual-criteria collaborative decision mechanism of energy gradient and interval change, it can effectively identify energy-dominated strong discharge signal sources and keenly perceive weak early signal sources with characteristic interval jumps, solving the problem of overestimation or missed detection that easily occurs in complex electromagnetic environments using the traditional single threshold method. Especially under common operating conditions in cable operation sites, such as multiple types of concurrent discharges and low signal-to-noise ratio superimposed noise, this method significantly improves the completeness and robustness of signal source identification. Therefore, the obtained signal source count estimate can accurately guide subsequent graph structure modeling and variational optimization processes, ensuring the technical consistency and reliability of the entire blind separation process.

[0036] Step 5: Determine the number of nodes in the graph structure based on the estimated number of signal sources, obtain multiple Euclidean distances between nodes based on the time-frequency characteristics of the initially separated signals, and construct a similarity matrix. Further, the number of nodes in the graph structure is determined based on the estimated number of signal sources. Euclidean distances between nodes are obtained based on the time-frequency characteristics of the initially separated signals. A similarity matrix representing spatial and temporal correlations is constructed, including: determining the number of nodes in the graph structure based on the estimated number of signal sources; mapping each signal source component in the initially separated signal to a node in the graph structure; performing time-frequency analysis on the initially separated signal to obtain the time spectrum of the signal and extracting the time-frequency characteristics of the corresponding signal component for each node; the time-frequency characteristics include instantaneous frequency, time-frequency energy distribution, and time-frequency clustering; calculating the first Euclidean distance between the time-frequency feature vectors of the corresponding signal components of any two nodes; constructing a spatial correlation metric based on the first Euclidean distance and obtaining spatial correlation weights; a larger correlation weight value indicates greater similarity between the two nodes in the feature space; calculating a second Euclidean distance for the time-frequency feature vectors of the same node at adjacent times; constructing a temporal correlation metric based on the second Euclidean distance and calculating the temporal correlation weight value using a temporal kernel function; fusing the spatial correlation weight value and the temporal correlation weight value to construct a comprehensive similarity matrix; and normalizing the similarity matrix to obtain a normalized similarity matrix. Determining the number of nodes in the graph structure based on the estimated number of signal sources refers to obtaining the estimated number of signal sources. This number serves as the total number of graph nodes in the graph neural network model. It is not pre-set or empirically determined, but rather adaptively estimated from the mixed signal using a data-driven approach. This ensures the graph structure accurately reflects the actual number of discharge signal sources, avoiding topological distortion caused by overfitting or underfitting. For example, in a cable monitoring scenario, if three independent discharge sources are estimated, a graph structure with three nodes is constructed, each corresponding to a potential discharge signal component. Mapping each signal source component in the initially separated signal to a node in the graph structure means that each signal component initially separated by the generative adversarial network is abstracted as a vertex on the graph. This mapping method allows subsequent modeling of the interaction relationships between signal components using a graph neural network. Each node not only carries its own signal waveform information but also establishes connections with other nodes through edges, thus forming a signal dependency network with semantic structure. Time-frequency analysis of the initially separated signal to obtain its time spectrum is a crucial step in achieving multi-dimensional feature extraction. Methods such as short-time Fourier transform, continuous wavelet transform, or synchronous compression transform can be used to convert the one-dimensional time-domain signal into a two-dimensional time-frequency plane representation. This time-frequency spectrum can simultaneously reveal the dynamic characteristics of signal frequency changes over time, and is particularly suitable for the analysis of non-stationary, transient partial discharge signals. The time-frequency characteristics of the signal components corresponding to each node are extracted, specifically including three core indicators: instantaneous frequency, time-frequency energy distribution, and time-frequency clustering. Instantaneous frequency describes the changing trend of the signal's dominant frequency component at a given moment and can be estimated by the frequency position of the energy peak in the time-frequency spectrum. Time-frequency energy distribution characterizes the breadth and concentration of the signal's energy distribution across the entire time-frequency plane, and can be used to distinguish between broadband noise and narrowband discharge signals. Time-frequency clustering measures the compactness of signal energy in the time-frequency plane; high clustering usually corresponds to effective discharge events with strong impulses and short durations, while low clustering may indicate diffuse interference or background noise.

[0037] The first Euclidean distance between the time-frequency eigenvectors of corresponding signal components of any two nodes is calculated to quantify the similarity of different signal sources in the feature space. Assume nodes... and nodes The time-frequency eigenvectors are respectively and Then the first Euclidean distance is: ;in, The distance is the square of the Euclidean norm. The smaller this distance is, the closer the two signal components are in terms of time-frequency characteristics, and the more likely they are to originate from the same type of discharge mechanism or have similar propagation paths. Constructing a spatial correlation measure based on the first Euclidean distance transforms the original distance into a correlation weight that reflects the strength of the association between nodes. A common method is to use a Gaussian kernel function for nonlinear mapping. ;in, For scale parameters, The natural exponential function is used to describe the process of exponential growth or decay. It is a distribution related to the probability density function and can be adaptively adjusted according to the statistical characteristics of the dataset. A larger weight value indicates that two nodes are more similar in the feature space, and a stronger connection should be established in the graph structure. The second Euclidean distance is calculated for the time-frequency eigenvectors of the same node at adjacent times to evaluate the time-evolution stationarity of a single signal source. Let the nodes... At any moment and The time-frequency eigenvectors are respectively and Then the second Euclidean distance is: This distance reflects the degree of drastic change in signal characteristics over time. A smaller distance indicates good temporal continuity, consistent with the slow evolution of signals in real physical processes; conversely, a larger distance may indicate abrupt changes or interference. A temporal correlation metric is constructed based on the second Euclidean distance, and a temporal correlation weight value is calculated using a temporal kernel function. An exponentially decaying kernel function can be selected. ;in, This is an adjustment coefficient. This weight reflects the smoothness of the signal itself in the time dimension, serving as an important basis for self-loops or time edges in graph structures, and can also be used for temporal neighborhood aggregation in weighted graph convolution. Spatial correlation weights and temporal correlation weights are fused to construct a comprehensive similarity matrix. This expresses the horizontal (across channels) and vertical (across time) dependencies between signal components in a unified form. A feasible fusion strategy is weighted summation: ;in, This can be obtained by averaging the time-series weights within the shared time window. As a balancing factor, the emphasis can be adjusted according to the application scenario. This similarity matrix forms the basis of the adjacency matrix of the graph structure. The similarity matrix is ​​then normalized to obtain a normalized similarity matrix. Normalization helps eliminate biases caused by differences in node degree, improves the training stability of the graph neural network, and ensures that subsequent graph convolution operations meet the basic premises of spectral graph theory. The above technical features are interconnected and synergistic; the signal source quantity estimation provides the node cardinality for the graph structure, ensuring reasonable modeling granularity; time-frequency analysis provides highly discriminative feature representations for each node; spatial and temporal distances respectively characterize horizontal similarity and vertical continuity; the construction of correlation weights realizes the transformation from the original distance to the graph connection strength; finally, fusion and normalization form a standardized graph topology prior, laying the structural foundation for subsequent information propagation and feature aggregation in the graph neural network. This application realizes the construction of a dynamic graph structure that reflects both the feature similarity between signal sources and their temporal evolution by establishing a two-dimensional (spatial and temporal) similarity model based on time-frequency features for initially separated signals. This graphical structure not only enhances the signal separation model's ability to resolve multi-source aliasing signals in complex electromagnetic environments, but also strengthens the physical consistency and temporal coherence of the separation results. Therefore, it solves the problem that traditional blind separation methods struggle to model deep dependencies between signals, significantly improving the accuracy and robustness of early cable discharge signal identification under weak signal conditions.

[0038] Step 6: Extract node features based on the similarity matrix to generate a graph feature vector. Using this graph feature vector, construct a posterior probability distribution. Perform recursive iterative optimization by maximizing the posterior probability distribution to output the optimized target separation signal. Further, conduct a preliminary signal quality assessment of the target separation signal, select high-quality separation results, and perform multi-dimensional feature extraction and adaptive amplification signal enhancement processing on the selected target separation signal. This process first uses quantitative indicators to determine the reliability of the optimized target separation signal to ensure that subsequent enhancement operations only apply to effective signal components. Second, a comprehensive signal representation system is constructed based on multi-domain feature analysis. Finally, an intelligent gain strategy is used to achieve differentiated amplification, significantly improving the observability of weak discharge signals while preserving the integrity of strong signals. The entire process follows a progressive logic of "evaluation—analysis—enhancement," balancing signal fidelity and feature highlighting. A preliminary signal quality assessment is performed on the target separated signal by calculating the signal-to-noise ratio (SNR) of the input mixed signal and the SNR of the target separated signal. The SNR reflects the degree to which the separation process improves signal quality. Based on these two SNR values, the correlation coefficient between the target separated signal and a pre-stored reference clean signal is further calculated. The correlation coefficient reflects the similarity between the separated signal and the ideal clean signal. A quality assessment threshold is set based on the SNR and correlation coefficient. When both the SNR and correlation coefficient are greater than a preset threshold, it indicates that the target separated signal has both a good SNR improvement and a high correlation with the clean signal, and is thus judged as a high-quality separation result. Target separated signals that meet the dual threshold conditions are selected and transferred to the next enhancement processing step. Low-quality signals that do not meet the conditions are marked and returned to the initial signal separation or iterative optimization separation step. The preliminary signal quality assessment refers to measuring the reliability of the target separated signal using objective and quantifiable statistical indicators. The signal-to-noise ratio (SNR), as one of the basic indicators, reflects the energy advantage of the effective components in the signal relative to noise. Specifically, the signal-to-noise ratio (SNR) of the input mixed signal represents the degree of noise pollution in the original observation environment, while the SNR of the target separated signal reflects the denoising capability of the blind separation algorithm. The difference between the two is the SNR improvement, directly reflecting the quality of the separation performance. Furthermore, a correlation coefficient between the target separated signal and a reference clean signal is introduced to measure the consistency of the waveform morphology. This reference clean signal can be derived from a standard laboratory discharge model, historical high SNR samples, or an ideal pulse sequence generated through simulation. A high correlation coefficient means that the separation result not only has concentrated energy but also a temporal structure close to the real discharge waveform, possessing physical rationality. Based on this, a dual-threshold mechanism is set: the first threshold is the lower limit of the SNR (e.g., 10 dB), and the second threshold is the correlation coefficient threshold (e.g., 0.8).Only signal segments that simultaneously meet both conditions are recognized as high-quality separation results and proceed to the next stage of the enhancement process; otherwise, they are considered low-quality outputs, which the system marks and feeds back to the preceding modules (such as the retraining stage of the generative adversarial network or the iterative optimization loop of graph convolution) to achieve closed-loop correction. This screening mechanism avoids ineffective enhancement of erroneous or distorted signals, improving the robustness of the overall system. As an optional implementation, quality assessment can also introduce a dynamic threshold adjustment mechanism, adaptively adjusting the judgment criteria based on the electromagnetic interference level or cable load status of the current operating environment. For example, in strong power frequency interference scenarios, the signal-to-noise ratio threshold can be appropriately reduced while the correlation weight is increased, thereby maintaining a balance between system sensitivity and accuracy. Multidimensional feature extraction is performed on the selected high-quality target separation signals to construct a multidimensional feature parameter set. Specifically, time-domain feature parameters such as pulse amplitude, pulse width, rise time, fall time, and pulse leading-edge steepness are extracted. In the frequency domain, frequency-domain feature parameters such as the dominant frequency component, spectral centroid, and the proportion of high-frequency components are extracted using Fast Fourier Transform. Furthermore, in both the time and frequency domains, wavelet packet transform is used to extract the energy entropy and instantaneous frequency change rate of each time and frequency sub-band. This multidimensional feature extraction aims to characterize the essential properties of the discharge signal from different dimensions, providing a decision-making basis for subsequent adaptive enhancement. The time-domain characteristics focus on the geometric properties of the pulse waveform: pulse amplitude refers to the maximum absolute amplitude of a single discharge pulse, reflecting the energy intensity of partial discharge; pulse width is defined as the time span from the start point to the end point of the pulse, usually expressed as full width at half maximum (FWHM), which can be used to distinguish different types of discharge modes; rise time and fall time are the time required for the pulse to rise from 10% peak value to 90% peak value and from 90% to 10% peak value, respectively, and are important bases for judging the type of discharge source (such as internal discharge vs. surface discharge); the steepness of the pulse leading edge, i.e., the rise slope, reflects the intensity of discharge development and is often positively correlated with the severity of insulation defects. The frequency-domain characteristics reveal the frequency composition of the signal: the frequency point where the energy is most concentrated is determined after obtaining the spectrum of the dominant frequency component through Fast Fourier Transform, which helps to identify the resonant response corresponding to a specific discharge mechanism; the centroid of the spectrum represents the weighted average frequency position of the signal energy distribution, calculated using the formula [formula missing]. Here, fi represents the frequency point, and Pi represents its corresponding power, which can be used to monitor the drift trend of the discharge frequency band. The proportion of high-frequency components refers to the proportion of energy above a certain cutoff frequency (e.g., 50MHz) to the total energy, suitable for determining whether broadband radiation caused by early weak discharge exists. Furthermore, combining wavelet packet transform to achieve more refined time-frequency joint analysis involves decomposing the signal into multiple orthogonal sub-bands, calculating the energy entropy of each sub-band (lower entropy indicates more concentrated energy and more ordered signal), and extracting the instantaneous frequency change rate of each sub-band to capture dynamic modulation behavior in non-stationary signals, particularly suitable for tracking intermittent discharge signals in complex backgrounds. As an alternative, Hilbert-Huang transform can be used instead of wavelet packets for nonlinear non-stationary signal analysis, or Mel frequency cepstral coefficients can be used to simulate human ear perception characteristics for feature dimensionality reduction. An adaptive gain control technique is employed to enhance the amplitude of a multi-dimensional feature parameter set. Specifically, the energy distribution of the target separated signal is calculated based on the pulse amplitude, and the peak factor is calculated based on the peak value and root mean square (RMS). An adaptive gain factor is obtained using the energy distribution and peak factor. Based on the adaptive gain factor and pulse amplitude, differentiated gain strategies are applied to signal components with different amplitude characteristics to obtain an amplitude-enhanced signal. For early discharge signal components with low amplitude, high-gain amplification is used to increase the amplitude to a preset range. For high-amplitude discharge signal components, low-gain or linear amplification is used to avoid signal saturation distortion, resulting in an amplitude-enhanced signal. The adaptive gain control technique is an intelligent amplification mechanism based on the signal's inherent characteristics. Its core lies in dynamically generating a gain factor based on the extracted multi-dimensional features, rather than using a fixed amplification factor.

[0039] Specifically: First, the overall energy distribution of the signal is estimated using the pulse amplitude sequence, and the local energy density can be calculated using the sliding window method. Then, the peak value and root mean square value are combined to calculate the gust factor, which reflects the presence of sudden strong pulses in the signal. High energy distribution but low gust factor may indicate continuous noise, while low energy but high gust factor typically corresponds to sparse and weak discharge events. A mapping function is constructed based on the two parameters, outputting an adaptive gain factor G. For example, a nonlinear function can be designed: Where E is the normalized energy and C is the normalized peak factor. and Adjustable weighting coefficients ensure greater gain for weak signals. A differentiated gain strategy is then implemented: for low-amplitude signal components (e.g., less than 3σ, where σ is the standard deviation of background noise), high-gain amplification (e.g., gain factor ≥ 10) is applied to increase their amplitude to a preset analysis range (e.g., ±1V) for subsequent feature detection and classification; for high-amplitude signal components (e.g., greater than 10σ), low-gain (e.g., 1-2 times) or linear gain is used to prevent the signal from exceeding the ADC's dynamic range and causing clipping distortion; a smooth gain curve (e.g., the Sigmoid function) is used in the transition region to achieve seamless switching. This strategy effectively solves the problem that traditional linear amplification cannot highlight weak signals under low signal-to-noise ratio conditions, and is prone to saturation in high-amplitude regions, achieving maximum visualization enhancement while maintaining fidelity. Multi-domain feature enhancement is performed based on the amplitude enhancement signal. Specifically, in the time domain, edge sharpening techniques are used to enhance the rising and falling edges of the pulse signal, targeting the rise and fall times of the time-domain feature parameters. A high-pass filter is used to convolve the amplitude enhancement signal to obtain the time-domain enhanced signal. Frequency domain enhancement is then performed based on this time-domain enhanced signal. An adaptive filter is used to enhance the characteristic frequency bands of the discharge signal while suppressing residual noise bands, targeting the dominant frequency component and the proportion of high-frequency components in the frequency-domain feature parameters. Frequency-domain filtering is applied to the time-domain enhanced signal, and an inverse Fourier transform is used to obtain the frequency-domain enhanced signal. The time-domain and frequency-domain enhanced signals are input into a trained U-Net network, outputting the final enhanced target separation signal. The final enhanced signal combines the advantages of quality screening, multi-dimensional feature extraction, adaptive amplitude enhancement, multi-domain feature enhancement, and deep learning enhancement, exhibiting high signal-to-noise ratio, clear discharge characteristics, and good time-frequency localization properties. The multi-domain feature enhancement further strengthens the expressiveness of key features based on amplitude enhancement. At the time domain, the focus is on enhancing the sharpness of pulse edges using edge sharpening techniques such as gradient operators (Sobel, Laplacian) or morphological filtering to highlight the transient characteristics of rising / falling edges. A high-pass filter (e.g., Butterworth type, with a cutoff frequency of 10–50MHz) is used to convolve the signal, suppressing low-frequency drift and baseline fluctuations while preserving high-frequency details to generate a time-domain enhanced signal. At the frequency domain, the emphasis is on spectrum reconstruction: an adaptive filter is designed whose passband center frequency automatically tracks the dominant frequency component, and whose bandwidth dynamically adjusts according to the proportion of high-frequency components. This enhances the discharge characteristic frequency band (e.g., 30–80MHz) while effectively suppressing residual noise frequency bands (e.g., 50Hz and its harmonics). The filtering operation is performed in the frequency domain by performing an FFT on the time-domain enhanced signal, multiplying it by the filter response function in the frequency domain, and then performing an inverse Fourier transform to restore the frequency-domain enhanced signal. Finally, the signals enhanced in both the time and frequency domains are concatenated into a multi-channel input tensor and fed into a pre-trained U-Net neural network.This network features an encoder-decoder structure and a skip connection mechanism, enabling it to fuse multi-scale contextual information, further remove artifacts, fill in missing details, and output the final enhanced target separation signal. The U-Net model can be trained using supervised learning, with noisy separation signals as input and corresponding high signal-to-noise ratio (SNR) clean signals as labels. The loss function employs a combination of methods (e.g., L1 loss + perceptual loss) to ensure the output signal is both faithful to the original and discriminative. As an alternative, a Transformer-based architecture can be used to replace U-Net, utilizing a self-attention mechanism to model long-range dependencies, making it particularly suitable for global optimization of long-term discharge signals. This application achieves quality discrimination and intelligent enhancement of target separation signals. By employing a dual-threshold screening mechanism based on signal-to-noise ratio and correlation coefficient, erroneously separated components are effectively excluded from the enhancement process. The extraction of multi-dimensional physical features from the time and frequency domains provides ample information support for subsequent processing. The adoption of an adaptive gain control strategy based on energy distribution and peak factor enables focused amplification of weak early discharge signals without compromising the integrity of strong signals. Furthermore, the combination of classical signal processing methods and deep learning models achieves a closed-loop enhancement from local feature enhancement to global structural optimization. Therefore, the final output signal not only exhibits a significantly improved signal-to-noise ratio but also clearer discharge pulse morphology and more prominent characteristic frequencies, possessing excellent time-frequency localization characteristics. This makes it a reliable data foundation for early cable fault diagnosis, supporting advanced applications such as discharge type identification, severity assessment, and discharge source location.

[0040] Step 7: Perform a preliminary signal quality assessment on the target separated signal, screen out high-quality separation results, and perform signal enhancement processing on the screened target separated signal to obtain the enhancement result. The preliminary signal quality assessment is accomplished by calculating the signal-to-noise ratio (SNR) improvement of the target separated signal relative to the original mixed signal, and its Pearson correlation coefficient with the reference clean signal. A dual threshold is set: if the SNR improvement exceeds a certain threshold and the correlation coefficient is higher than a preset value, it is determined to be a high-quality separation result and enters the enhancement process; otherwise, it is marked as a low-quality signal and returned to the previous module for reprocessing. For high-quality signals, perform multi-dimensional feature extraction: extract pulse amplitude, width, rise / fall time, and leading-edge steepness in the time domain; obtain the dominant frequency component, spectral centroid, and high-frequency component proportion in the frequency domain using Fast Fourier Transform (FFT); and extract the energy entropy and instantaneous frequency change rate of each sub-band using wavelet packet transform in the time-frequency domain to form a multi-dimensional feature parameter set. Adaptive amplification is implemented based on this set: calculate the energy distribution according to the pulse amplitude and peak factor to determine the adaptive gain factor. For low-amplitude early discharge signals, high-gain amplification (such as nonlinear gain functions) is used to bring them into a recognizable range; for high-amplitude signals, low-gain or linear amplification is used to avoid saturation distortion and achieve differentiated enhancement. Further multi-domain feature enhancement is performed: in the time domain, edge sharpening techniques are used to enhance the rising and falling edges of the pulse; in the frequency domain, adaptive filters are designed to enhance characteristic frequency bands and suppress residual noise; finally, the processed signal is input into a trained U-Net network for end-to-end repair, outputting a signal with enhanced amplitude and clearer details.

[0041] Step 8: Perform multiple cross-validation on the target separation signal and the enhancement processing result to obtain the verified separation effect; evaluate the signal quality of the verified separation effect and correct the low-quality signal region to obtain the final enhanced signal;

[0042] Furthermore, multiple cross-validation is performed on the target separated signal and the enhancement processing result. The separation effect is verified by calculating signal quality indicators such as signal-to-noise ratio (SNR), improvement ratio, and cross-correlation coefficient. Signal quality evaluation is then conducted based on the verification of the separation effect using these signal quality indicators, and low-quality signal regions are corrected to obtain the final enhanced signal. First, multiple cross-validation is performed on the final enhanced signal and the enhancement processing result to obtain a verification result containing multiple validation scores. This process aims to independently evaluate the reliability and authenticity of the signal from multiple dimensions, avoiding the risk of misjudgment caused by a single indicator. The multiple cross-validation includes the following five core verifications: verifying signal quality indicators, obtaining the SNR of the final enhanced signal, the improvement ratio of the SNR relative to the original mixed signal, and the cross-correlation coefficient with the laboratory reference pure signal; these three indicators are used as the first validation score. The SNR measures the ratio of the effective component to the noise power in the target signal. It is calculated by using the energy of the target enhanced signal as the numerator and the estimated energy of the residual noise segment as the denominator, expressed in logarithmic form. The improvement ratio is defined as the difference between the enhanced SNR and the original mixed signal SNR, reflecting the net gain capability of the algorithm throughout the entire process. The cross-correlation coefficient characterizes the waveform similarity between the current enhanced signal and the known standard discharge signal waveform. The closer its value is to 1, the closer the separated signal is to the ideal state. These three indicators together constitute the basic quality assessment system at the signal level, which is suitable for judging whether the signal has achieved effective denoising and feature preservation in a statistical sense. For example, in a certain implementation scenario, when the signal-to-noise ratio of the original mixed signal is 8dB, if the enhanced signal reaches more than 6dB and the cross-correlation coefficient is greater than 0.85, it can be preliminarily determined that the signal has engineering usability. Based on the final enhanced signal, the spectral characteristics are verified. The spectrum is obtained through fast Fourier transform, the spectral distribution is analyzed to identify the residual noise frequency band, and the noise frequency band energy and total energy are obtained to obtain the noise frequency band energy ratio. When the noise frequency band energy ratio is less than a preset threshold, the noise suppression effect is judged to be good. The spectral purity index is obtained based on the noise energy ratio as the second verification score. Among them, fast Fourier transform converts the time domain signal to the frequency domain to form an amplitude spectrum or power spectral density map. By designating the frequency bands containing typical interference sources (such as power frequency harmonics at 50 / 150 / 250Hz, switching power supply noise bands, etc.) as residual noise monitoring zones, the proportion of signal energy within these zones to the total signal energy is calculated, yielding the noise frequency band energy percentage. For example, if this percentage is below 5%, the spectral purity is considered high, and a high score is awarded; conversely, a higher percentage results in a deduction. This step aims to detect the presence of incompletely suppressed periodic or broadband noise components, providing quantitative constraints, particularly for artifacts or mode collapse phenomena that may be introduced by generative adversarial networks. Furthermore, wavelet packet decomposition can be used instead of FFT for more refined sub-band energy analysis, improving frequency localization accuracy.Based on spectral characteristic verification and physical mechanism verification, a set of discharge characteristic parameters is extracted from the final enhanced signal, including key parameters such as peak value, pulse width, and repetition frequency of the discharge pulse. These discharge characteristic parameters are then matched with a physical characteristic database of known discharge types to check whether the peak value range, width range, and frequency range conform to the current cable insulation state and environmental conditions. The physical mechanism conformity is used as the third verification score. Discharge characteristic parameters are crucial for identifying partial discharge types. For example, internal air gap discharge typically exhibits narrow pulses (width <10ns) and high repetition frequencies (greater than 1kHz), while surface discharge presents wider pulses (20–50ns) and lower repetition rates. A physical characteristic database is constructed by setting empirical parameter ranges for different discharge modes, and the judgment threshold is dynamically adjusted based on actual operating voltage, temperature, humidity, and other environmental conditions to achieve context-aware rationality verification. If a separated signal has a high signal-to-noise ratio but its pulse width reaches 200ns and has no significant polarity difference, it does not conform to any typical discharge model and should be considered an abnormal result. This step prevents the generation of false signals due to overfitting in deep learning. Based on the verification of physical mechanisms, multi-sensor data fusion verification is performed. Target separation signals are acquired from N sensors distributed at different monitoring points. The cross-correlation coefficient between any two signals is calculated, and a cross-correlation matrix is ​​constructed. The spatial consistency of the signal separation results is evaluated using multi-point cross-validation technology. The average cross-correlation coefficient is calculated, and spatial consistency is considered good when the average cross-correlation coefficient is greater than a preset threshold. This spatial consistency index is used as the fourth verification score. N UHF sensors are distributed along the cable path, recording the propagation response of the same discharge source at different locations. Due to the attenuation and time delay of electromagnetic wave propagation, the signals received by each sensor have slight differences in amplitude and phase, but the time series of their core discharge events should remain highly synchronized. The cross-correlation coefficient between pairs of signals is calculated using the sliding window method, and an N×N cross-correlation matrix is ​​constructed. The mean of the off-diagonal elements is then used as a measure of overall spatial consistency. For example, when the average cross-correlation coefficient is higher than 0.7, it indicates that multiple observation points have captured consistent discharge activity, enhancing the confidence in the signal's authenticity. This method can also be combined with a time-of-arrival (TOA) positioning algorithm to infer the consistency of the discharge source's location, further strengthening the verification logic. Based on multi-sensor verification, trend verification is performed using a historical discharge database. Historical discharge characteristic parameter sequences related to the current monitoring point are extracted from the database. The current set of discharge characteristic parameters is compared with the historical sequences over time to obtain the slope and acceleration of the characteristic parameter changes. It is then checked whether the current discharge event is a continuation of a known discharge source or a newly emerging discharge defect. The trend consistency score is used as the fifth verification score. The historical discharge database stores the evolution trajectory of discharge characteristics at the same measuring point or in a nearby area over several past periods.By establishing time series models (such as ARIMA, LSTM, etc.), the expected discharge intensity and evolution rate at the current moment are predicted and compared with measured values. If the current pulse peak shows an exponential growth trend and the acceleration of change is significantly greater than the historical average, it suggests a potential risk of accelerated degradation. If a new discharge mode suddenly appears (such as a change from internal discharge to surface discharge), it may also indicate new faults such as structural damage or moisture absorption. The trend consistency score is dynamically assigned accordingly to distinguish between normal fluctuations and true deterioration. This step achieves a leap from "static snapshot" to "dynamic evolution," making the diagnosis more forward-looking. Five verification scores are assigned confidence weights {w1, w2, w3, w4, w5}, where each weight is determined based on the reliability and importance of the verification item and satisfies w1+w2+w3+w4+w5=1. For example, in the early stages of a newly built line where historical data is lacking, the weights can be appropriately reduced. Weighting, increasing and The weight of the term should be considered; however, in aging cables that have been in operation for a long time, the influence of the trend term should be given priority. A weighted comprehensive confidence score is obtained by weighted summation, which comprehensively reflects the credibility of the final enhanced signal in five dimensions: signal quality, spectral purity, physical mechanism, spatial consistency, and temporal trend.

[0043] A deep quality assessment is performed based on the comprehensive confidence score to obtain the quality assessment results. Specifically, a pre-trained autoencoder structure is used to re-encode the final enhanced signal, and the activation intensity distribution in the feature maps of each layer of the encoder is observed. For regions with weak activation (such as ReLU activation values ​​below a set threshold τ), combined with the local signal-to-noise ratio map, they are marked as a set of low-quality signal regions. These regions often correspond to parts with weak signals, blurred structures, or those susceptible to reconstruction errors, such as the starting edge of an early weak discharge pulse or overlapping segments under dense noise interference. The fused features (i.e., the high-level abstract representation before decoding) are compared point-by-point with the feature representation obtained from the original final enhanced signal through the same encoding path to obtain a residual feature map. Based on this residual map, an initial correction amount Δ0 is calculated using a residual learning mechanism. Subsequently, for each spatial-temporal unit (such as a pixel on the time-frequency plane) in the set of low-quality regions, a correction enhancement factor α greater than 1 is introduced. Its value is adaptively determined according to the degree of quality defect in the region, so that regions with lower signal-to-noise ratios receive greater correction. For normal-quality regions, the original correction amount remains unchanged (i.e., α=1). This generates an enhanced and corrected fusion feature F'. The fusion feature F', corrected for low-quality regions, is input to the output layer of the decoder. The output layer performs convolution operations using convolution kernels and is mapped back to the time-domain signal space using activation functions (such as Sigmoid or Tanh) to generate an updated time-domain signal waveform. Post-processing is performed on this waveform: first, a low-pass filter (with a cutoff frequency slightly higher than the discharge signal's main frequency) is used for smoothing to remove high-frequency noise spikes; then, boundary correction is performed, such as using mirror extension combined with window function weighting to eliminate the boundary effects introduced by segmented processing, ultimately outputting a continuous, complete, and high-quality enhanced signal. This application achieves comprehensive and reliable verification and local fine-grained repair of the separation and enhancement results of early discharge signals in cables. The multiple cross-validation mechanism, from five independent perspectives—statistical, spectral, physical, spatial, and temporal—mutually corroborates each other, significantly reducing the probability of misjudgment; the comprehensive confidence score provides a unified quantitative decision-making basis; and the targeted correction strategy based on feature map activation analysis and residual learning accurately repairs weak links without destroying the global signal structure. Therefore, the final output enhanced signal not only has a high signal-to-noise ratio and clear discharge characteristics, but also exhibits excellent quality in terms of spectral purity, physical rationality, spatial consistency and development trend. It can serve as an authoritative basis for early discharge fault diagnosis of cables and can be widely used in key tasks such as intelligent identification of discharge type, graded assessment of insulation degradation, and precise location of discharge source.

[0044] A computer-readable storage medium is provided for storing computer-readable instructions that, when read by a computer, enable the execution of a blind separation and enhancement processing method for cable discharge signals based on adversarial networks, as described above. This application provides a computer-readable storage medium for automating the execution of a blind separation and enhancement processing method for cable discharge signals based on adversarial networks. The storage medium stores computer program instructions that, when loaded and executed by a processor, drive a computing device to complete the entire process from mixed signal acquisition, preprocessing, blind separation to signal enhancement and quality assessment.

[0045] Application Scenario 1: Multi-point Monitoring of Urban High-Voltage Cable Tunnels: In urban underground cable tunnels, multiple 110kV to 220kV high-voltage cables are typically laid in parallel, with a spacing of only 0.5 to 1 meter between each cable. Key locations prone to partial discharge, such as cable joints and terminals, are densely distributed, forming multiple potential discharge sources. The tunnel also contains strong power frequency interference and electromagnetic radiation from operating equipment such as substations and switchgear. Simultaneously, the large current-carrying capacity of the cables leads to significant temperature rises, with ambient temperatures reaching 40℃ to 60℃ and humidity fluctuating between 30% and 90%. The coupling of these factors makes the monitoring environment extremely complex. Traditional monitoring methods struggle to accurately identify and locate multiple simultaneously existing weak discharge signals and are easily affected by electromagnetic interference from nearby cables and operating equipment, leading to misjudgments and missed detections. To address the characteristics of multi-point monitoring scenarios in urban high-voltage cable tunnels, the following targeted adjustments are made to the basic technical solution: In the data acquisition phase, an ultra-high frequency sensor array is deployed every 20 to 50 meters along the cable tunnel. Each array contains 3 to 5 sensors, forming a spatial three-dimensional monitoring network, with a total of 30 to 50 monitoring points (N). The IEEE 1588 precision clock synchronization protocol is used to ensure that the time alignment accuracy of each sensor is better than 100 ns, supporting high-precision time-delay ranging. In the preprocessing phase, to address the signal spatial aliasing problem caused by the parallel operation of multiple cables, array signal processing technology is introduced. Beamforming algorithms are used to spatially filter the received signals from each sensor array, initially suppressing interference signals from non-target directions and improving spatial resolution. The accuracy was improved to within 0.3 meters. During the generative adversarial network training phase, the pure discharge signal sample library was expanded, and mixed discharge samples simulating multiple cables operating in parallel were added. The sample signal-to-noise ratio range was set from -15dB to 5dB, covering extremely low signal-to-noise ratio scenarios. For cases where the power frequency harmonic interference intensity is 3 to 5 times the original signal amplitude, a frequency domain constraint term was added to the discriminator loss function to enhance the generator's ability to suppress power frequency and its harmonic components. During the signal source estimation phase, since multiple cables may simultaneously have discharge sources, the dynamic energy threshold parameters α=1.2 and β=0.5 of the energy gradient decision criterion were adjusted to reduce threshold sensitivity and avoid misclassifying multiple signal sources with similar energies as a single source. Simultaneously, the weight w of the interval decision criterion was increased. b=0.6, enhancing the detection capability of weak early discharge signals; in the graph structure modeling stage, a hierarchical graph structure is constructed based on the spatial topology of the cable tunnel, dividing the tunnel into several monitoring sections. Nodes in each section constitute a subgraph, and subgraphs are connected through boundary nodes. A hierarchical hybrid graph convolutional network is used to extract the local dependencies of signals within sections and the global correlation patterns of signals across sections, reducing computational complexity; in the multi-sensor fusion verification stage, an N×N cross-correlation matrix is ​​constructed using a densely arranged sensor array. Multidimensional scaling analysis technology is used to map the high-dimensional cross-correlation relationship to a two-dimensional plane, intuitively presenting the spatial distribution pattern of discharge sources. Combined with the time difference of arrival positioning algorithm and the triangulation principle, the three-dimensional spatial precise positioning of discharge sources is achieved, with a positioning accuracy better than 0.5 meters; a multi-target tracking algorithm based on particle filtering is introduced to continuously track the spatiotemporal evolution trajectory of multiple discharge sources in the tunnel, distinguishing the continuous development of known discharge sources from newly emerging discharge defects, generating a discharge source evolution trend map, and providing data support for formulating differentiated maintenance strategies.

[0046] Implementation steps: First, deploy the sensor array and configure the clock synchronization in the target cable tunnel according to the design plan. Verify the consistency of the amplitude-frequency characteristics of each sensor channel through on-site calibration tests to ensure that the normalization error is less than 3%. Next, collect mixed signal data under two working conditions: normal cable operation and artificially applied partial discharge defects. Construct a dedicated training sample library containing multiple parallel cables and strong power frequency interference, with a total number of samples of no less than 200,000. Then, train a generative adversarial network model optimized for this scenario. The training cycle is 5,000 to 8,000 rounds. Convergence is determined when the signal-to-noise ratio improvement ratio of the generator output signal is stable above 12dB and the cross-correlation coefficient with the reference signal is greater than 0.85. Finally, deploy the trained model to the tunnel monitoring system to receive data in real time. The mixed signals collected by the sensor array undergo a full-process processing including blind separation, signal source estimation, graph optimization, and signal enhancement, with a single processing delay controlled within 2 seconds. The separated and enhanced signals are subjected to five-dimensional multi-cross-validation, with a comprehensive confidence score threshold set at 0.75. Signals exceeding the threshold are considered valid discharge events, triggering a three-dimensional spatial positioning program. The positioning results are expressed as latitude and longitude coordinates and linear distance relative to the tunnel entrance, and are marked in real time on the electronic map of the monitoring center. A database of discharge characteristic parameters for each cable and joint in the tunnel is established based on continuous monitoring data, and a trend analysis report is automatically generated every 24 hours. The deviation of the current discharge intensity from the historical baseline is compared, and an early warning signal is issued when the deviation rate exceeds 30% or the acceleration exceeds twice the historical average. This technical solution has achieved excellent results in a 220kV cable tunnel application in a certain city. Under harsh conditions where the power frequency harmonic interference intensity is 3.5 times that of the original signal and the signal-to-noise ratio is as low as -12dB, it successfully separated 5 independent discharge sources, improved the signal-to-noise ratio of the separated signal to 14.3dB, achieved a cross-correlation coefficient of 0.88 with the laboratory standard signal, and had a spectral purity index of 96.2%. Through multi-sensor fusion positioning technology, it has avoided large-scale power outages caused by sudden faults.

[0047] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for blind separation and enhancement processing of cable discharge signals based on adversarial networks, characterized in that, include: Acquire the mixed signal from the cable, perform preprocessing and normalization, and construct the signal set to be separated; Generative adversarial networks are built based on the set of signals to be separated. The set of signals to be separated is input into the generative adversarial network, and the corresponding preliminary separation signal is output. Based on the preliminary separated signal, eigenvalue decomposition is performed to obtain an estimate of the number of signal sources in the mixed signal; The number of nodes in the graph structure is determined based on the estimated number of signal sources. Multiple Euclidean distances between nodes are obtained based on the time-frequency characteristics of the initially separated signals, and a similarity matrix is ​​constructed. Node features are extracted based on the similarity matrix to generate graph feature vectors; The posterior probability distribution is constructed using the graph feature vectors, and the optimized target separation signal is output by recursively iterating through the maximized posterior probability distribution. A preliminary signal quality assessment is performed on the target separation signal to select high-quality separation results, and the selected target separation signal is then subjected to signal enhancement processing to obtain the enhancement processing result. Multiple cross-validations were performed on the target separation signal and the enhancement processing results to verify the separation effect. The signal quality of the verification separation effect is evaluated, and the low-quality signal regions are corrected to obtain the final enhanced signal.

2. The method for blind separation and enhancement of cable discharge signals based on adversarial networks according to claim 1, characterized in that, The mixed signal from the cable is acquired, preprocessed, and normalized to construct a set of signals to be separated, including: Mixed signal data containing discharge signals are collected by deploying ultra-high frequency sensors, current sensors, and partial discharge detection devices at key nodes of cable lines; and environmental parameters are obtained by using temperature and humidity sensors and electromagnetic field strength detectors. The mixed signal data is denoted as xi, where i represents the sampling time, and xi includes weak partial discharge pulse signal, strong background noise and electrical transient signal; The mixed signal data is preprocessed, including using bandpass filtering to remove DC and UHF components, using normalization to unify the signal amplitude range, and using timestamp alignment to achieve time synchronization between the mixed signal and environmental data. The normalized mixed signal data is divided into multiple signal segments according to a fixed time window. Each segment is used as an independent sample. The environmental data is mean-processed according to the corresponding time window to form an environmental feature vector paired with the signal segment. The signal segments are subjected to multi-channel conversion, and the time-domain waveform of each signal segment is converted into a frequency-domain spectrum. A joint feature matrix is ​​generated by combining the environmental feature vector. The dimension of the joint feature matrix is ​​4×m, where m is the number of sampling points of the signal segment. Combine all joint feature matrices to form the signal set to be separated.

3. The method for blind separation and enhancement of cable discharge signals based on adversarial networks according to claim 1, characterized in that, Based on the set of signals to be separated, a generative adversarial network is built, including: In a controlled laboratory environment, we simulated different types of partial discharge, including tip discharge, surface discharge, internal discharge, and suspension discharge. We collected pure discharge signal samples through shielding measures and in a low-noise environment to build a pure discharge signal sample library. We collected pure noise samples containing power frequency interference, electromagnetic field interference, and white noise under different environmental conditions, analyzed their spectral characteristics and probability distribution, and constructed a noise sample library. According to different signal-to-noise ratios and mixing ratios, the pure discharge signal is superimposed with various types of noise to generate mixed signal samples with known components. All samples are classified and stored according to multiple dimensions such as discharge type, noise type, signal-to-noise ratio and environmental conditions. Each sample is appended with a multi-dimensional representation of time-domain waveform, frequency-domain spectrogram, and time-frequency plot to construct a standardized sample library S=[s1, s2, ..., s...]. k ], where k is the total number of samples; Based on the pure discharge signal sample library and the noise sample library, a generative adversarial network architecture is trained. The generator in the generative adversarial network adopts an encoder-decoder architecture, while the discriminator adopts a deep convolutional neural network architecture. By training the generative adversarial network, the generator learns the mapping relationship for extracting pure discharge signals from mixed signals. During the training process, the network parameters of the generator and the discriminator are alternately optimized until the weighted sum of the adversarial loss and the reconstruction loss reaches the preset convergence condition.

4. The method for blind separation and enhancement processing of cable discharge signals based on adversarial networks according to claim 1, characterized in that, The set of signals to be separated is input into the generative adversarial network, and the corresponding preliminary separation signal is output, including: The mixed signal segments in the signal set to be separated are input one by one into the trained generative adversarial network. The generative adversarial network performs multi-layer convolution operations on the input mixed signal through the encoder of the generator, and extracts multi-scale features including low-frequency basic features, mid-frequency texture features and high-frequency detail features at different levels. The multi-scale features are fused in the feature space to generate a fused feature vector; the fused feature vector contains separate representations of the discharge signal and the noise signal in the mixed signal; The fused feature vector is input to the decoder, which performs layer-by-layer upsampling through transpose convolution to output an abstract representation of the feature space; the abstract representation is then mapped back from the high-dimensional feature space to the original time-domain signal space, gradually restoring it to a time-domain signal waveform. During the decoding process, skip connections are used to concatenate the features of the corresponding layer of the encoder with the features of the current layer of the decoder, thus preserving the detailed information of the signal. Boundary effects are removed and the time-domain signal waveform is smoothed to obtain a preliminary separated signal; The preliminary separation signal identifies and extracts the discharge signal component from the mixed signal and suppresses the noise component; A sliding window strategy is used to segment long-term sequence signals. Each window has a length of multiple sampling points and a preset overlap rate between windows. The generator outputs a preliminary separated signal segment for the mixed signal of each window. The multiple separation results of the overlapping area are fused by a weighted average method to obtain a complete and continuous preliminary separated signal.

5. The method for blind separation and enhancement of cable discharge signals based on adversarial networks according to claim 1, characterized in that, Based on the preliminary separated signal, eigenvalue decomposition is performed to obtain an estimate of the number of signal sources in the mixed signal, including: The initial separated signal is converted into a multi-channel time series matrix; each row of the multi-channel time series matrix represents the multi-channel observation value at a time moment, and each column represents the time series data of one channel; A covariance matrix is ​​constructed based on the multi-channel time series matrix to obtain the covariance between any two channels; the covariance matrix is ​​a symmetric positive definite matrix with a dimension equal to the number of channels; The covariance matrix is ​​decomposed into eigenvalues ​​to obtain the eigenvalue sequence [λ1≥λ2≥…≥λ]. M ], where M is the number of channels, and the energy percentage of each feature value is calculated based on the feature value sequence to obtain the feature value energy sequence; Based on the eigenvalue sequence and the number of rows and columns of the multi-channel time series matrix, the probability density value of the noise eigenvalue is calculated; The upper and lower boundaries of noise feature values ​​are obtained through the probability density values. The relationship between each feature value in the feature value sequence and the upper and lower boundaries of the noise is compared. The signal subspace and noise subspace are distinguished and determined by combining the feature value energy sequence. Based on the upper and lower boundaries of the noise feature values, the difference features between the signal subspace and the noise subspace in the feature value sequence are analyzed, the interval between adjacent feature values ​​is calculated, and a feature value interval sequence is constructed. Specifically, an energy gradient decision criterion is constructed based on the feature value energy sequence to obtain the energy gradient and set a dynamic energy threshold. When the energy gradient is greater than the dynamic energy threshold, it indicates that there is an energy jump between the feature value at the current moment and the feature value at the next moment. Then, from the perspective of energy distribution, the feature value at the current moment corresponds to the first signal component. An interval decision criterion is constructed based on the feature value interval sequence and an interval threshold is set. When the median of the feature value interval is greater than the preset interval threshold, the feature value at the current moment corresponds to the second signal component from the perspective of interval distribution. The decision results of the energy gradient decision criterion and the interval decision criterion are assigned confidence weights w1 and w2 respectively, and the decision values ​​are weighted and fused. The weighted decision value results are then fused, and the number of feature values ​​that meet the decision conditions is counted to obtain the estimated number of signal sources.

6. The method for blind separation and enhancement of cable discharge signals based on adversarial networks according to claim 1, characterized in that, The number of nodes in the graph structure is determined based on the estimated number of signal sources. Multiple Euclidean distances between nodes are obtained based on the time-frequency characteristics of the initially separated signals. A similarity matrix is ​​then constructed, including: The number of nodes in the graph structure is determined based on the estimated number of signal sources; each signal source component in the initially separated signal is mapped to a node in the graph structure. Time-frequency analysis is performed on the initially separated signal to obtain the time spectrum of the signal and extract the time-frequency features of the signal component corresponding to each node; the time-frequency features include instantaneous frequency, time-frequency energy distribution, and time-frequency clustering. Calculate the first Euclidean distance between the time-frequency eigenvectors of the corresponding signal components of any two nodes; A spatial correlation metric is constructed based on the first Euclidean distance, and a spatial correlation weight is obtained; the larger the correlation weight value, the more similar the two nodes are in the feature space. Calculate the second Euclidean distance for the time-frequency eigenvectors of the same node at adjacent times; A temporal correlation metric is constructed based on the second Euclidean distance, and a temporal correlation weight value is calculated using a temporal kernel function. The spatial relevance weight value and the temporal relevance weight value are fused to construct a comprehensive similarity matrix; and the similarity matrix is ​​normalized to obtain a normalized similarity matrix.

7. The method for blind separation and enhancement of cable discharge signals based on adversarial networks according to claim 4, characterized in that, Node features are extracted based on the similarity matrix to generate graph feature vectors; Using the graph feature vectors, a posterior probability distribution is constructed. Recursive iterative optimization is then performed by maximizing this posterior probability distribution, resulting in an optimized target separation signal, including: The normalized similarity matrix and the node feature matrix of the preliminary separated signal are input into a preset hybrid graph convolutional network; the hybrid graph convolutional network contains multiple graph convolutional layers, and each graph convolutional layer performs message passing and feature aggregation operations; In the first graph convolutional layer, the set of neighboring nodes of the i-th node is determined based on the normalized similarity matrix, the feature information of its neighboring nodes is aggregated, and the updated node features are calculated. The node features are convolved in the time dimension to capture the temporal evolution of the node features and calculate the temporal convolution features. Spatial graph convolutional features and temporal graph convolutional features are fused to obtain hybrid graph convolutional features. Feature dimensionality reduction and nonlinear transformation are performed through fully connected layers to generate graph feature vectors. The graph feature vectors are converted into mapped feature vectors through a mapping network, whereby the mapped feature vectors represent the mean and log-variance of the hierarchical variation distribution, respectively. A hierarchical variation distribution is constructed based on the mapped feature vector and the initial separation signal. The hierarchical variation distribution is constructed by combining a global distribution representing the overall dependency relationship and a local distribution representing the independent features. The global distribution is a single Gaussian distribution, and the local distribution is a product of the Gaussian distributions corresponding to each initial separation signal. Spatial and temporal features are extracted from the initially separated signals. The Euclidean distance between the spatial features of any two signals is calculated to obtain the spatial correlation measure. The Euclidean distance between the temporal features of the same signal at adjacent times is calculated to obtain the temporal correlation measure. A spatial correlation probability is constructed based on the spatial correlation metric, and a temporal correlation probability is constructed based on the temporal correlation metric. The posterior probability distribution is obtained by multiplying the global distribution, the local distribution, the spatial correlation probability, and the temporal correlation probability; recursively iteratively optimizes the posterior probability distribution by maximizing the lower bound of evidence. During the iteration process, latent variables are sampled from the hierarchical change distribution, and the latent variables are mapped back to the signal space through the decoder network to obtain the reconstructed signal; Calculate the reconstruction loss between the reconstructed signal and the original preliminary separation signal, the reconstruction loss measuring the signal reconstruction quality; and calculate the divergence loss between the hierarchical change distribution and the prior distribution, the divergence loss constraining the distribution of the latent variables; weight the two losses to obtain the total loss, and update all network parameters through the backpropagation algorithm to minimize the total loss; The iteration optimization process is repeated. When the convergence condition that the change in total loss is less than a preset threshold is met or the maximum number of iterations is reached, the iteration stops and the optimized target separation signal is output. The target separation signal is the reconstructed signal obtained in the last iteration.

8. The method for blind separation and enhancement processing of cable discharge signals based on adversarial networks according to claim 5, characterized in that, A preliminary signal quality assessment is performed on the target separation signal to select high-quality separation results. Then, signal enhancement processing is applied to the selected target separation signals to obtain the enhancement results, including: A preliminary signal quality assessment is performed on the target separated signal, and the signal-to-noise ratio (SNR) of the input mixed signal and the SNR of the target separated signal are calculated; the SNR reflects the degree to which the separation process improves the signal quality. Based on two signal-to-noise ratios, the correlation coefficient between the target separated signal and the pre-stored reference pure signal is further calculated; the correlation coefficient reflects the degree of similarity between the separated signal and the ideal pure signal. A quality assessment threshold is set based on the signal-to-noise ratio and the correlation coefficient. When the signal-to-noise ratio is greater than the preset threshold and the correlation coefficient is greater than the preset correlation coefficient threshold, it indicates that the target separation signal has both a good signal-to-noise ratio improvement and is highly correlated with the clean signal, and is judged as a high-quality separation result. Target separation signals that meet the dual threshold conditions are selected and transmitted to the next enhancement processing step. Low-quality signals that do not meet the conditions are marked and returned to the initial signal separation or iterative optimization separation step. Multidimensional feature extraction is performed on the selected high-quality target separation signals to construct a multidimensional feature parameter set; Specifically, time-domain feature parameters such as pulse amplitude, pulse width, rise time, fall time, and pulse leading edge steepness are extracted in the time domain; frequency-domain feature parameters such as the dominant frequency component, spectral centroid, and proportion of high-frequency components are extracted in the frequency domain using fast Fourier transform; and time-domain and frequency-domain feature parameters such as energy entropy and instantaneous frequency change rate of each time-domain and frequency-domain sub-band are extracted using wavelet packet transform in both the time and frequency domains. Adaptive gain control technology is used to enhance the amplitude of the multidimensional feature parameter set; Specifically, the energy distribution of the target separation signal is calculated based on the pulse amplitude, the peak factor is calculated based on the peak value and the root mean square, and the adaptive gain factor is obtained through the energy distribution and the peak factor. Based on the adaptive gain factor and the pulse amplitude, a differentiated gain strategy is applied to signal components with different amplitude characteristics to obtain an amplitude-enhanced signal; Specifically, for the early discharge signal component with low amplitude, high-gain amplification is used to increase the amplitude to the preset amplitude range; for the discharge signal component with high amplitude, low-gain or linear amplification is used to avoid signal saturation distortion and obtain an amplitude-enhanced signal. Multi-domain feature enhancement is performed based on the amplitude enhancement signal; wherein, in the time domain, the rising and falling edges of the pulse signal are enhanced by edge sharpening technology for the rise and fall times in the time domain feature parameters; and a high-pass filter is used to perform convolution operation on the amplitude enhancement signal to obtain the time domain enhancement signal; Frequency domain enhancement is performed based on the time-domain enhancement signal; wherein, the proportion of the main frequency component and high-frequency component in the frequency domain feature parameters is used to enhance the characteristic frequency band of the discharge signal while suppressing the residual noise frequency band through an adaptive filter; the time-domain enhancement signal is subjected to frequency domain filtering processing, and the frequency domain enhancement signal is obtained through inverse Fourier transform; The time-domain and frequency-domain enhanced signals are input into the trained U-Net network, and the final enhanced target separation signal is output.

9. The method for blind separation and enhancement of cable discharge signals based on adversarial networks according to claim 6, characterized in that, Multiple cross-validations were performed on the target separation signal and the enhancement processing results to verify the separation effect. The signal quality of the verified separation effect is evaluated, and low-quality signal regions are corrected to obtain the final enhanced signal, including: Multiple cross-validation is performed on the final enhanced signal and the enhancement processing result to obtain the validation result; the validation result includes multiple validation scores. Each verification score in the verification result is assigned a credibility weight, wherein each weight is determined based on the reliability and importance of the verification item and the sum of the weights is 1. A weighted comprehensive confidence score is obtained by combining confidence weights and verification scores; the comprehensive confidence score reflects the credibility of the final enhanced signal in five dimensions: signal quality, spectral purity, physical mechanism, spatial consistency, and temporal trend. A deep quality assessment is performed based on the comprehensive confidence score to obtain the quality assessment results. Based on the quality assessment results, low-quality signal regions are identified according to the activation intensity distribution of the feature maps of each layer of the encoder and the signal quality index. Regions with activation intensity below the threshold or time-frequency points with local signal-to-noise ratio in the feature maps are marked as a set of low-quality regions. Specifically, the fused features are compared with the feature representation of the original final enhanced signal to obtain a comparison result; based on the comparison result, the correction amount is calculated through residual learning; a correction enhancement factor is applied to low-quality regions to calculate the enhanced correction amount, wherein when the correction enhancement factor is greater than 1, it is adaptively determined according to the degree of quality defect in the region, and the worse the quality of the region, the greater the correction intensity; for normal quality regions, the original correction amount is maintained. The fused features, corrected for low-quality regions, are input to the output layer of the decoder. The output layer performs convolution operations using convolution kernels and maps back to the time-domain signal space using an activation function. The time-domain signal waveform is then post-processed by smoothing it with a low-pass filter to remove high-frequency noise spikes, followed by boundary correction to eliminate boundary effects, resulting in the final enhanced signal.

10. A computer-readable storage medium, characterized in that, Used to store computer-readable instructions, which, when read by a computer, enable the execution of a cable discharge signal blind separation and enhancement processing method based on an adversarial network as described in any one of claims 1-9.

Images