Local Discharge Signal Recognition Method, System and Storage Medium Based on Deep Learning

Abstract

A method, system and storage medium for identifying partial discharge signals based on deep learning, which relates to the field of computer systems based on specific computational models. The method includes: dividing multi-source monitoring data into a high-frequency signal group and a low-frequency signal group; respectively inputting corresponding feature information into a first deep neural network and a second deep neural network to obtain a high-frequency feature vector and a low-frequency feature vector; generating a signal attenuation compensation factor based on the low-frequency feature vector to perform distortion compensation on the high-frequency feature vector to obtain a high-frequency corrected vector; calculating the feature correlation degree between the high-frequency corrected vector and the low-frequency feature vector, and when the feature correlation degree is lower than a preset correlation threshold, extracting the transient feature information of the high-frequency corrected vector and fusing the low-frequency feature vector to obtain a low-frequency corrected vector; inputting the high-frequency corrected vector and the low-frequency corrected vector into a partial discharge recognition model to obtain the partial discharge type and recognition confidence. Implementing this application can improve the recognition accuracy of partial discharge signals.

Description

Technical Field

[0001] This application relates to the field of computer systems based on specific computational models, and particularly to a method, system, and storage medium for identifying partial discharge signals based on deep learning. Background Art

[0002] With the continuous expansion of the scale of the power system and the increasing complexity of the equipment operation environment, the importance of partial discharge detection of power equipment has become increasingly prominent. As an important means to evaluate the insulation state of high-voltage electrical equipment, partial discharge detection is of great significance for ensuring the safe and stable operation of the power system. Especially in key equipment such as large substations, GIS equipment, and power cables, accurately identifying and promptly detecting partial discharge faults plays a crucial role in preventing major accidents and realizing predictive maintenance of equipment.

[0003] In related technologies, partial discharge detection of power equipment mainly adopts a fixed threshold judgment method, that is, by setting a predetermined detection parameter threshold, an alarm is triggered when the detected signal exceeds the set threshold. For example, in the detection of GIS equipment, discharge signals are collected through UHF sensors, and amplitude and frequency thresholds are set for fault judgment; in the detection of transformers, acoustic sensing and electrical quantity measurement are combined, and fixed criteria such as acoustic signal intensity and electrical quantity amplitude are respectively set for fault identification. The detection device analyzes the single or multiple signals collected respectively, and judges whether there is a partial discharge fault in the equipment by comparing with the preset threshold.

[0004] However, there are information islands in the data monitoring of various signals in related technologies. In complex fault modes, such as when the monitoring sensitivity in a certain area decreases, the recognition accuracy of partial discharge signals will decrease. Summary of the Invention

[0005] This application provides a method, system, and storage medium for identifying partial discharge signals based on deep learning, which is used to improve the recognition accuracy of partial discharge signals.

[0006] In a first aspect, the present application provides a method for identifying partial discharge signals based on deep learning, which is applied to a partial discharge identification system. The method includes: receiving multi-source monitoring data of partial discharge, and dividing the multi-source monitoring data into a high-frequency signal group and a low-frequency signal group according to a preset discrimination frequency; respectively extracting the feature information of the high-frequency signal group and the low-frequency signal group, and respectively inputting the corresponding feature information into a first deep neural network and a second deep neural network to obtain a high-frequency feature vector and a low-frequency feature vector; generating a signal attenuation compensation factor based on the low-frequency feature vector, and performing distortion compensation on the high-frequency feature vector according to the signal attenuation compensation factor to obtain a high-frequency corrected vector; calculating the feature correlation degree between the high-frequency corrected vector and the low-frequency feature vector, and when the feature correlation degree is lower than a preset correlation threshold, extracting the transient feature information of the high-frequency corrected vector; fusing the transient feature information and the low-frequency feature vector to obtain a low-frequency corrected vector; inputting the high-frequency corrected vector and the low-frequency corrected vector into a partial discharge identification model to obtain the partial discharge type and the identification confidence level.

[0007] In the above embodiment, the partial discharge identification system divides the multi-source monitoring data into a high-frequency and a low-frequency signal group, extracts feature vectors through a deep neural network, generates an attenuation compensation factor using the low-frequency feature vector to perform distortion compensation on the high-frequency feature vector, and then extracts transient feature information according to the feature correlation degree; through multi-level feature fusion and optimization, high-precision identification of partial discharge signals is achieved; effectively solving the problem of low identification accuracy of traditional fixed threshold methods in complex environments and improving the reliability of partial discharge detection.

[0008] Combined with some embodiments of the first aspect, in some embodiments, the steps of respectively extracting the feature information of the high-frequency signal group and the low-frequency signal group, and respectively inputting the corresponding feature information into a first deep neural network and a second deep neural network to obtain a high-frequency feature vector and a low-frequency feature vector specifically include: performing time-frequency transformation on the high-frequency signal group and the low-frequency signal group to obtain time-frequency feature maps; performing hierarchical convolutional operations on the time-frequency feature maps based on a convolutional neural network to obtain feature tensors; performing self-supervised learning encoding on the feature tensors to generate feature representation matrices; inputting the feature representation matrices into the corresponding first deep neural network and second deep neural network respectively to obtain a high-frequency feature vector and a low-frequency feature vector.

[0009] In the above embodiment, the partial discharge identification system performs time-frequency transformation on the high-frequency and low-frequency signal groups to obtain time-frequency feature maps, then extracts feature tensors through hierarchical convolutional operations of a convolutional neural network, and performs self-supervised learning encoding to generate feature representation matrices; this multi-level feature extraction method fully exploits the time-frequency domain feature information of the signals, improves the feature expression ability, and makes the discharge type identification more accurate.

[0010] In some embodiments in combination with some embodiments of the first aspect, the step of inputting the high-frequency correction vector and the low-frequency correction vector into the partial discharge recognition model to obtain the partial discharge type and the recognition confidence specifically includes: inputting the high-frequency correction vector into a long short-term memory network to obtain a time series feature sequence; inputting the low-frequency correction vector into a Transformer encoder to obtain an attention feature sequence; calculating a similarity matrix of the time series feature sequence and the attention feature sequence through contrastive learning, and generating the partial discharge type and the recognition confidence based on the similarity matrix.

[0011] In the above embodiments, the partial discharge recognition system inputs the high-frequency correction vector into a long short-term memory network to extract time series features, inputs the low-frequency correction vector into a Transformer encoder to obtain attention features, and then calculates a similarity matrix through contrastive learning; this deep learning architecture makes full use of the sequence modeling ability to achieve accurate recognition of the discharge type and reliable confidence evaluation.

[0012] In some embodiments in combination with some embodiments of the first aspect, before the step of fusing the transient feature information and the low-frequency feature vector to obtain the low-frequency correction vector, the method further includes: extracting the discharge pulse period from the high-frequency correction vector to obtain a period feature sequence; performing wavelet decomposition on the period feature sequence to obtain multi-scale components; performing noise reduction on the multi-scale components based on adaptive filtering to obtain a reconstructed signal; extracting the feature information in the reconstructed signal to obtain the transient feature information.

[0013] In the above embodiments, the partial discharge recognition system extracts the discharge pulse period feature from the high-frequency correction vector, obtains multi-scale components through wavelet decomposition, and performs noise reduction processing using adaptive filtering; this signal processing method effectively extracts the essential features of the discharge signal, suppresses the influence of interference noise, and improves the purity and reliability of the feature information.

[0014] In some embodiments in combination with some embodiments of the first aspect, after the step of extracting the feature information in the reconstructed signal to obtain the transient feature information, it specifically includes: calculating the time-frequency spectrogram of the reconstructed signal to obtain an energy distribution sequence; performing peak detection on the energy distribution sequence to determine the transient feature points; segmenting the signal based on the transient feature points to obtain transient feature intervals, and generating transient feature information according to the transient feature intervals.

[0015] In the above embodiments, the partial discharge recognition system calculates the time-frequency spectrogram of the reconstructed signal to obtain the energy distribution, performs peak detection to determine the feature points and segments to extract the feature intervals; this analysis method based on energy features can accurately capture the transient features of the discharge signal.

[0016] In some embodiments in combination with some embodiments of the first aspect, after the step of inputting the high-frequency correction vector and the low-frequency correction vector into the partial discharge recognition model to obtain the partial discharge type and the recognition confidence, the method further includes: performing feature fusion on the high-frequency correction vector and the low-frequency correction vector to obtain a fused feature; calculating the correlation coefficient between the fused feature and a preset partial discharge template to obtain a matching degree matrix; screening out the feature pattern with the largest matching degree from the matching degree matrix to obtain a recognition result; and updating the neural network parameters of the partial discharge recognition model according to the recognition result.

[0017] In the above embodiments, the partial discharge recognition system performs feature fusion on the high- and low-frequency correction vectors, obtains the recognition result through matching analysis with the preset template, and updates the model parameters accordingly; this adaptive optimization mechanism continuously improves the recognition ability of the model, enables the system to continuously learn and improve, and enhances the user experience.

[0018] In some embodiments in combination with some embodiments of the first aspect, the step of updating the neural network parameters of the partial discharge recognition model according to the recognition result specifically includes: calculating the confidence distribution of the recognition result to obtain a feedback weight coefficient; performing backpropagation on the neural network based on the feedback weight coefficient to obtain a gradient value; adjusting the convolutional layer parameters according to the gradient value to obtain optimized parameters, and writing the optimized parameters into the partial discharge recognition model.

[0019] In the above embodiments, the partial discharge recognition system calculates the feedback weight based on the confidence distribution of the recognition result and adjusts the network parameters through backpropagation; this dynamic optimization mechanism enables the model to continuously adapt to new discharge feature patterns and continuously improves the recognition accuracy.

[0020] In a second aspect, an embodiment of the present application provides a partial discharge recognition system, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, and the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the partial discharge recognition system to execute the method described in the first aspect and any possible implementation manner in the first aspect.

[0021] In a third aspect, an embodiment of the present application provides a computer program product containing instructions, which, when the computer program product runs on a partial discharge recognition system, enables the partial discharge recognition system to execute the method described in the first aspect and any possible implementation manner in the first aspect.

[0022] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, including instructions, which, when the instructions run on a partial discharge recognition system, enable the partial discharge recognition system to execute the method described in the first aspect and any possible implementation manner in the first aspect.

[0023] Understandably, the partial discharge recognition system provided in the second aspect above, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the method provided in the embodiments of the present application. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding method, and will not be elaborated here.

[0024] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0025] 1. Due to the adoption of the high-low frequency separation and deep learning feature extraction methods for multi-source monitoring data, as well as the attenuation compensation and transient feature fusion technologies based on low-frequency feature vectors, it is possible to make full use of the complementary advantages of signals in different frequency bands, realize the accurate extraction and fusion of signal features, effectively solve the problems of incomplete single-frequency band feature extraction and recognition deviation caused by signal attenuation in the prior art, and thus achieve high-precision and high-reliability recognition of partial discharge signals, improving the accuracy and reliability of power equipment condition monitoring.

[0026] 2. Due to the adoption of signal processing methods such as discharge pulse period extraction, wavelet multi-scale decomposition, and adaptive filtering, it is possible to effectively separate the discharge feature information from complex original signals, and through multi-level signal reconstruction and feature extraction, effectively solve the problems of easy interference of discharge signals and inaccurate feature extraction in the prior art, and thus achieve the precise capture and reliable characterization of discharge features, improving the anti-interference ability of feature extraction and the stability of recognition.

[0027] 3. Due to the adoption of optimization mechanisms such as feature fusion, template matching, and network parameter adaptive update, it is possible to comprehensively utilize multi-source feature information for recognition and judgment, and improve the recognition performance through continuous model optimization, effectively solve the problems of poor model adaptability and low reliability of recognition results in the prior art, and thus achieve the optimization and performance improvement of the recognition model, enhancing the practicability and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flowchart of a method for recognizing partial discharge signals based on deep learning in the embodiments of the present application;

[0029] Figure 2 is another flowchart of a method for recognizing partial discharge signals based on deep learning in the embodiments of the present application;

[0030] Figure 3 is a schematic structural diagram of an entity device of a partial discharge recognition system in the embodiments of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The terms used in the following embodiments of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. As used in the specification of this application, the singular forms "a", "an", "above-mentioned", "the", and "this" are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in this application refers to any or all possible combinations including one or more of the listed items.

[0032] Hereinafter, the terms "first" and "second" are only for descriptive purposes and cannot be construed as implying or suggesting relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of this application, unless otherwise stated, the meaning of "a plurality" is two or more.

[0033] For ease of understanding, the application scenarios of the embodiments of this application are introduced below.

[0034] During the long-term operation of GIS equipment in a large substation distribution room, partial discharge faults may occur. Operation and maintenance personnel need to detect and diagnose these faults in a timely manner to prevent equipment damage. Traditional partial discharge monitoring systems use sensors with a single frequency band and are easily affected by external interference and signal attenuation, resulting in false alarms or missed alarms. Especially in a complex electromagnetic environment, environmental noise, crosstalk signals, etc. will seriously affect the detection accuracy. In addition, different types of partial discharges have different characteristic manifestations in different frequency bands, and the detection method with a single frequency band is difficult to comprehensively capture the discharge characteristics, reducing the reliability of fault diagnosis.

[0035] In the related art, the monitoring and identification of partial discharge can be achieved by adopting a single-frequency band signal acquisition and fixed threshold detection method. The scenario of using the deep learning-based partial discharge signal identification method in the related art is introduced below.

[0036] A power company uses an existing partial discharge monitoring system to conduct on-line monitoring of a transformer. The system uses a signal processing method with a fixed threshold and a simple pattern recognition algorithm. When partial discharge occurs in the transformer due to internal insulation defects, due to the attenuation and distortion during signal transmission, the high-frequency characteristics are severely distorted. At the same time, the system cannot effectively distinguish corona discharge generated during normal operation from dangerous internal discharge, resulting in false alarms. Maintenance personnel need to conduct frequent on-site inspections, increasing the operation and maintenance costs. Moreover, the system cannot adapt to the dynamic changes in the equipment operation state, and the performance of the recognition model gradually degrades.

[0037] By adopting the partial discharge signal recognition method based on deep learning in the embodiments of the present application, accurate and reliable discharge type recognition is achieved through high-low frequency feature fusion and an adaptive compensation mechanism, which not only improves the detection sensitivity but also reduces the false alarm rate. The following introduces the scenarios where the partial discharge signal recognition method based on deep learning in the present application is used.

[0038] A certain intelligent substation uses the multi-source partial discharge recognition system proposed in the present application for equipment monitoring. The system simultaneously collects high-frequency and low-frequency signals, and realizes feature extraction and fault recognition through adaptive signal processing and deep learning algorithms. When a floating potential discharge occurs in a certain gas-insulated switchgear (GIS), the system can simultaneously analyze the transient features in the high-frequency signal and the continuous features in the low-frequency signal to accurately identify the discharge type. Through feature fusion and an online learning mechanism, the system can continuously optimize the recognition model and improve the diagnostic accuracy. Even in the case of severe signal interference, it can still maintain high detection reliability.

[0039] It can be seen that by adopting the deep learning and online optimization methods in the embodiments of the present application, while realizing intelligent fault diagnosis, it can also effectively solve problems such as signal attenuation distortion and environmental interference, and thus realize the continuous optimization of the monitoring system.

[0040] For ease of understanding, the following describes the process of the method provided in this embodiment in combination with the above scenarios. Please refer to Figure 1 , which is a schematic flowchart of a partial discharge signal recognition method based on deep learning in the embodiments of the present application.

[0041] S101. Receive multi-source monitoring data of partial discharge, and divide the multi-source monitoring data into a high-frequency signal group and a low-frequency signal group according to a preset discrimination frequency.

[0042] Among them, partial discharge refers to the local breakdown discharge phenomenon that occurs in the insulating medium of electrical equipment; multi-source monitoring data refers to the discharge signal data simultaneously collected by various sensing devices such as ultra-high frequency sensors, acoustic sensors, and ground current sensors; the preset discrimination frequency refers to the frequency threshold used to distinguish high-frequency and low-frequency signals, which is set between 1 MHz and 100 MHz; the high-frequency signal group refers to the signal set with a frequency higher than the preset discrimination frequency, mainly including the transient pulse characteristics generated by the discharge; the low-frequency signal group refers to the signal set with a frequency lower than the preset discrimination frequency, mainly including the continuous characteristics during the discharge process.

[0043] When the partial discharge recognition system detects partial discharge in power equipment, it needs to classify the signals collected by multiple sensors. Specifically, the partial discharge recognition system first receives the original signal data collected by various sensors and preprocesses these data through time domain synchronization and amplitude normalization. Then, according to the preset frequency demarcation point (such as 10 MHz), a band-pass filter is used to separate the signal into two groups: high-frequency and low-frequency. The high-frequency signal group mainly contains the fast transient components generated during discharge, while the low-frequency signal group contains the slow-varying characteristics of the discharge process. This grouping method can adopt corresponding processing strategies according to the signal characteristics of different frequency bands.

[0044] In some embodiments, the signal frequency band division can be achieved in various ways: Optionally, the wavelet transform method is adopted. First, the original signal is decomposed at multiple scales to obtain wavelet coefficients in different frequency bands, and then the wavelet coefficients are recombined into high-frequency and low-frequency signal groups according to the preset discrimination frequency. The specific steps include: performing discrete wavelet transform on the original signal to obtain approximation coefficients and detail coefficients; determining the recombination layer according to the preset frequency; reconstructing the detail coefficients higher than the recombination layer into high-frequency signals, and reconstructing the approximation coefficients lower than the recombination layer into low-frequency signals. Optionally, the digital filtering method is adopted, and a band-pass filter is designed to separate the frequency bands of the original signal. The specific steps include: designing the cut-off frequency and transition band parameters of the high-pass and low-pass filters according to the preset discrimination frequency; performing high-pass and low-pass filtering on the original signal respectively to obtain the preliminary separation results; ensuring the integrity of the separated signals through energy compensation. It can be understood that other signal processing methods can also be adopted to achieve frequency band division, such as Fourier transform, Hilbert transform, etc., which are not limited here.

[0045] S102. Extract the characteristic information of the high-frequency signal group and the low-frequency signal group respectively, and input the corresponding characteristic information into the first deep neural network and the second deep neural network respectively to obtain high-frequency feature vectors and low-frequency feature vectors.

[0046] Among them, the characteristic information represents the characteristics of the signal, including time domain, frequency domain, and time-frequency domain characteristics; the first deep neural network and the second deep neural network refer to deep learning models designed according to the signal characteristics of different frequency bands; the feature vector refers to the high-dimensional feature representation after being mapped by the neural network.

[0047] The partial discharge recognition system needs to extract effective features from signals of different frequency bands. Specifically, the partial discharge recognition system first performs time-frequency analysis on the high-frequency signal group to extract transient features such as peak value, rise time, oscillation frequency, etc.; performs statistical analysis on the low-frequency signal group to extract steady-state features such as mean value, variance, spectral energy, etc. Then, these features are respectively input into a specially designed deep neural network, and through multi-layer non-linear transformation, the original features are mapped to a more discriminative feature space to obtain a feature vector with a fixed dimension. This two-stream feature extraction architecture can make full use of the characteristics of signals in different frequency bands.

[0048] In some embodiments, feature extraction and mapping can be achieved in various ways: Optionally, a time-frequency joint analysis method is adopted to construct a multi-scale feature extraction network. The specific steps include: calculating the time-frequency diagram of the signal; designing a multi-scale convolutional layer to extract local features; integrating features through a pooling layer and a fully connected layer. Optionally, an autoencoder structure is adopted to realize unsupervised learning of features. The specific steps include: designing an encoder and a decoder network; training the network through reconstruction loss; extracting the output of the encoder as a feature vector. It can be understood that other feature learning methods such as graph neural networks, attention mechanisms, etc. can also be adopted, which are not limited herein.

[0049] S103. Generate a signal attenuation compensation factor based on the low-frequency feature vector, and perform distortion compensation on the high-frequency feature vector according to the signal attenuation compensation factor to obtain a high-frequency corrected vector.

[0050] Among them, the low-frequency feature vector represents a set of feature data extracted from the low-frequency signal group; the signal attenuation compensation factor refers to a correction coefficient used to correct the attenuation distortion of the high-frequency signal, which is a real number between 0 and 1; the distortion compensation represents the process of restoring and correcting the signal through the compensation factor; the high-frequency corrected vector refers to the high-frequency feature vector after attenuation compensation, which more accurately reflects the characteristics of the original discharge signal.

[0051] After obtaining the high- and low-frequency feature vectors, the partial discharge recognition system needs to compensate for the attenuation distortion of the high-frequency signal. Specifically, the partial discharge recognition system first analyzes the energy distribution and spectral characteristics of the low-frequency feature vector to establish a signal transmission attenuation model. Based on this model, the attenuation degree of different frequency bands is calculated to generate a corresponding compensation factor matrix. Then, the compensation factor matrix is dimensionally matched and mathematically operated with the high-frequency feature vector to achieve amplitude and phase compensation of the high-frequency signal, and finally a corrected high-frequency feature vector is obtained. This compensation mechanism can effectively restore the original features of the high-frequency signal.

[0052] It should be noted that the distortion compensation process adopts a non-linear mapping compensation mechanism. First, an attenuation transfer function model is constructed based on the low-frequency feature vector, and the amplitude and phase attenuation characteristics during the signal transmission process are obtained through polynomial fitting. For amplitude attenuation, an exponential attenuation model related to frequency A(f) = A0*exp(-αf) is established, where α is the attenuation coefficient and is estimated through the energy spectrum of the low-frequency signal. For phase distortion, a linear phase model related to frequency φ(f) = 2πfτ + φ0 is adopted, where τ is the group delay and is estimated through the phase spectrum of the low-frequency signal. Then, the compensation factor K(f) = 1 / A(f)*exp(-jφ(f)) is calculated, and the high-frequency feature vector is compensated through frequency-domain complex multiplication. For example, if the amplitude attenuation at a certain frequency point is 0.6 and the phase delay is π / 3, the amplitude of the compensation factor at this frequency point is 1.67 and the phase is -π / 3. Channel non-linear characteristics also need to be considered in distortion compensation. The amplitude-frequency and phase-frequency characteristics of the channel are fitted through piecewise polynomial functions, and different compensation coefficients are used in different frequency bands. For the strongly non-linear region, a harmonic compensation term is introduced to correct signal distortion. After compensation, the causality of the signal is tested through Hilbert transform to ensure that the compensation result is physically realizable.

[0053] In some embodiments, attenuation compensation can be achieved in various ways: Optionally, an adaptive compensation method is adopted to dynamically estimate the attenuation degree based on the statistical characteristics of the low-frequency feature vector. The specific steps include: calculating the power spectral density of the low-frequency feature vector; establishing a transfer function model based on the energy attenuation law; and determining the optimal compensation factor through iterative optimization. Optionally, a deep learning method is adopted to train a compensation network to automatically learn the signal attenuation characteristics. The specific steps include: constructing a feature mapping network; inputting the low-frequency features for forward propagation; and updating the network parameters backward based on the reconstruction error to obtain a compensation model. It can be understood that other compensation methods can also be adopted, such as frequency-domain equalization, time-domain interpolation, etc., which are not limited here.

[0054] S104. Calculate the feature correlation degree between the high-frequency correction vector and the low-frequency feature vector, and extract the transient feature information of the high-frequency correction vector when the feature correlation degree is lower than the preset correlation threshold.

[0055] Among them, the feature correlation degree represents the similarity degree between two feature vectors, which is expressed by the correlation coefficient or distance metric; the preset correlation threshold refers to the standard value for judging feature correlation; the transient feature information refers to the fast-changing features during the discharge process, such as the rise time, peak value, duration, etc.

[0056] The partial discharge recognition system needs to evaluate the consistency of high-frequency and low-frequency features and extract supplementary features when necessary. Specifically, the partial discharge recognition system first calculates the correlation between the high-frequency corrected vector and the low-frequency feature vector using various correlation measurement methods, including the Pearson correlation coefficient, etc. When the correlation is lower than the preset threshold, it indicates that there is inconsistency between the two sets of features and further analysis is required. At this time, the system will extract transient process features from the high-frequency corrected vector, including amplitude, duration, oscillation characteristics, etc., as feature supplements.

[0057] It should be noted that the feature correlation calculation can adopt the multi-dimensional correlation analysis method. First, standardize the high-frequency corrected vector and the low-frequency feature vector to eliminate the influence of dimension. Then calculate the generalized cross-correlation function of the two vectors: R(τ)=E[X(t)Y(t + τ)], where E represents the expectation operation. Evaluate the correlation of features through the maximum cross-correlation coefficient and the correlation delay. At the same time, calculate the canonical correlation coefficient to examine the correlation strength of features in different dimensions. The two vectors are transformed through canonical transformation to obtain new variables U = aX and V = bY, and the transformation matrices a and b are optimized to maximize the correlation coefficient between U and V.

[0058] In some embodiments, feature correlation analysis can be achieved in various ways: Optionally, adopt the statistical method to calculate the multi-dimensional correlation index of the feature vector. The specific steps include: calculating the covariance matrix between vectors; extracting the main correlation components; setting a dynamic threshold for judgment. Optionally, adopt the distance measurement method to calculate the vector distance in the feature space. The specific steps include: selecting an appropriate distance measurement criterion; calculating the Euclidean distance or cosine distance between vectors; mapping the distance to the correlation interval. It can be understood that other correlation analysis methods, such as information entropy, fuzzy similarity, etc., can also be adopted, which are not limited here.

[0059] S105. Integrate the transient feature information and the low-frequency feature vector to obtain a low-frequency corrected vector.

[0060] Among them, feature fusion refers to the process of comprehensively integrating feature information from different sources or types; the low-frequency corrected vector refers to the optimized feature vector after integrating the transient features, which contains more complete discharge feature information.

[0061] The partial discharge recognition system needs to effectively integrate the extracted transient features with the original low-frequency features. Specifically, the partial discharge recognition system first performs scale normalization on the transient features and the low-frequency features to make them comparable. Then adopt a feature fusion algorithm, such as weighted summation, feature splicing, etc., to integrate the two types of features. The importance weights of various features will be considered during the fusion process, and the physical meaning of the features will be maintained. Finally, a low-frequency corrected vector containing complete information is obtained.

[0062] In some embodiments, the feature fusion process is carried out in an adaptive weighting manner. First, the transient feature information and the low-frequency feature vector are unified in dimension, and both are mapped to a feature space of the same dimension through principal component analysis. Then, the weight coefficients of each feature component are calculated based on the information entropy criterion, and the feature component with a larger information entropy obtains a higher weight. Next, the weighted features are subjected to feature transformation through a non-linear mapping function (such as the hyperbolic tangent function) to enhance the expression ability of the features. Finally, the transformed features are concatenated in series to form a fused feature vector. For example, for a certain discharge signal, the transient features include 10-dimensional features such as a rise time of 2 μs and a peak value of 1.5 mV, and the low-frequency feature vector includes 15-dimensional features such as spectral energy and phase. Through the above fusion process, a 25-dimensional low-frequency correction vector is obtained.

[0063] In some embodiments, feature fusion can be achieved in various ways: Optionally, a multi-level fusion strategy is adopted to combine at different feature levels. The specific steps include: feature hierarchical alignment; intra-layer feature fusion; inter-layer feature integration. Optionally, an attention mechanism is adopted for adaptive fusion according to the feature importance. The specific steps include: calculating the feature attention weight; feature weighted combination; feature vector reconstruction. It can be understood that other fusion methods, such as deep feature fusion, multi-modal fusion, etc., can also be adopted, which are not limited herein.

[0064] S106. Input the high-frequency correction vector and the low-frequency correction vector into the partial discharge recognition model to obtain the partial discharge type and the recognition confidence.

[0065] Among them, the partial discharge recognition model refers to a deep learning model used to recognize the discharge type; the discharge types include various forms such as corona discharge, surface discharge, and floating discharge; the recognition confidence represents the degree of certainty of the model for the recognition result.

[0066] The partial discharge recognition system needs to recognize the discharge type based on the corrected feature vector. Specifically, the partial discharge recognition system first integrates the high-frequency and low-frequency correction vectors into an input feature matrix according to a preset format. Then, the feature matrix is input into a pre-trained deep learning model, and the model includes network structures such as multiple layers of convolution and recurrence. The probability distributions of different discharge types are obtained through forward calculation, and the type with the highest probability is selected as the recognition result, and the corresponding confidence value is output. This confidence reflects the reliability of the recognition result and can be used for subsequent decision-making.

[0067] It should be noted that the partial discharge recognition model adopts a hierarchical deep learning architecture. The input layer receives high-frequency and low-frequency correction vectors and extracts local feature patterns through a multi-layer convolutional neural network. The convolutional layer uses convolutional kernels of different scales (such as 3×3, 5×5) to perform convolutional operations on the input features to extract multi-scale features. The pooling layer compresses the feature dimension through the maximum pooling operation and retains the significant features. The fully connected layer maps the features to the sample category space, and the output layer uses the Softmax function to calculate the probability distribution of each category. The model optimizes the network parameters through the backpropagation algorithm, and the loss function adopts the cross-entropy loss. When training the model, a dynamic learning rate strategy is used. A larger learning rate is used at the beginning of training to converge quickly, and the learning rate is reduced in the later stage to fine-tune the parameters.

[0068] In some embodiments, the discharge type recognition can be achieved in multiple ways: Optionally, an ensemble learning method is adopted to combine the prediction results of multiple base models. The specific steps include: constructing multiple heterogeneous classifiers; independently performing feature learning and classification; and fusing the results by voting or weighting. Optionally, a multi-task learning framework is adopted to simultaneously predict the discharge type and other related attributes. The specific steps include: designing a shared feature extraction layer; constructing multiple task branches; and jointly optimizing the loss functions of different tasks. It can be understood that other classification methods, such as semi-supervised learning, transfer learning, etc., can also be adopted, which are not limited herein.

[0069] In the above embodiments, the basic feature extraction and recognition processes are mainly introduced. In practical applications, the network structure and optimization strategy can also be adjusted according to specific scenarios to further improve the system performance. The scenarios of this embodiment are supplemented below.

[0070] To further improve the system performance, the substation has optimized and upgraded the partial discharge recognition system. The system has added a feature extraction module based on Transformer, which can better capture the temporal dependence relationship of the signal. At the same time, a contrast learning mechanism is introduced to improve the discriminability of the feature representation. In actual operation, the system can not only accurately identify single-type discharge faults, but also diagnose composite discharge faults. In addition, the system has the ability of knowledge transfer, can quickly adapt to the monitoring needs of new devices, and reduces the amount of labeled data required for model training. Through distributed deployment and collaborative learning, data sharing and model optimization among multiple substations are realized.

[0071] After combining the above scenarios, the following further describes the more specific process of the method provided in this embodiment. Please refer to Figure 2 , which is another process schematic diagram of the partial discharge signal recognition method based on deep learning in the embodiments of the present application.

[0072] S201. Receive multi-source monitoring data of partial discharge, and divide the multi-source monitoring data into a high-frequency signal group and a low-frequency signal group according to a preset discrimination frequency.

[0073] Referring to step S101, the partial discharge recognition system will perform signal frequency grouping.

[0074] S202. Extract the feature information of the high-frequency signal group and the low-frequency signal group respectively, and input the corresponding feature information into the first deep neural network and the second deep neural network respectively to obtain a high-frequency feature vector and a low-frequency feature vector.

[0075] Referring to step S102, the partial discharge recognition system will construct a high-frequency feature vector and a low-frequency feature vector.

[0076] In some embodiments, the partial discharge recognition system will perform time-frequency transformation and convolution operation when extracting the feature vector, that is, the partial discharge recognition system will perform time-frequency transformation on the high-frequency signal group and the low-frequency signal group to obtain a time-frequency feature map; based on a convolutional neural network, perform hierarchical convolution operation on the time-frequency feature map to obtain a feature tensor; perform self-supervised learning coding on the feature tensor to generate a feature representation matrix; input the feature representation matrix into the corresponding first deep neural network and second deep neural network respectively to obtain a high-frequency feature vector and a low-frequency feature vector.

[0077] Among them, the time-frequency transformation refers to a mathematical transformation method for converting a time-domain signal into a time-frequency domain, such as a short-time Fourier transform or a wavelet transform; the time-frequency feature map refers to a two-dimensional representation of a signal on the time-frequency plane; the hierarchical convolution operation refers to a process of extracting multi-level features of a feature map using different convolution kernels; the feature tensor refers to a multi-dimensional feature data structure obtained after convolution operation; the self-supervised learning coding refers to a coding method that can learn feature representations without labeled data; the feature representation matrix refers to a two-dimensional feature data set obtained after coding.

[0078] After the partial discharge recognition system obtains the high- and low-frequency signal groups, it needs to perform feature extraction and representation learning. Specifically, the partial discharge recognition system first performs a short-time Fourier transform or a continuous wavelet transform on the two groups of signals respectively to generate a two-dimensional feature map reflecting the time-frequency characteristics of the signals. Then use a multi-layer convolutional neural network to extract local feature patterns through convolution kernels of different scales to form a multi-channel feature tensor. Next, adopt self-supervised methods such as contrast learning or reconstruction learning to learn more discriminative feature representations. Finally, input these features into a specially designed deep neural network respectively, and obtain feature vectors of a fixed dimension through non-linear transformation. The whole process realizes the automatic extraction and representation from the original signal to high-dimensional features.

[0079] It should be noted that the deep neural network here includes the first and second deep neural networks. During the training phase, these two networks respectively receive the feature representation matrix as input and are supervised and trained using labeled discharge data. The training objective is to minimize the difference between the predicted output and the true label, and cross-entropy or mean squared error is used as the loss function. The network structure contains multiple fully connected layers, each followed by a non-linear activation function and a normalization layer. In practical applications, the two networks respectively process the feature representations of different frequency bands, and map the high-dimensional features to fixed-dimensional feature vectors through non-linear transformation.

[0080] Contrastive learning is based on a contrastive learning model. During the training phase, the model simultaneously receives the time series feature sequence and the attention feature sequence, and is trained by constructing positive and negative sample pairs. The training objective is to maximize the similarity between positive sample pairs and minimize the similarity with negative sample pairs, and the InfoNCE loss function is used for optimization. The model contains a feature projection head and a similarity calculation module. During use, the model calculates the similarity matrix between different feature sequences, and generates the recognition result and confidence evaluation of the discharge type based on the similarity distribution.

[0081] In some embodiments, feature extraction and representation learning can be implemented in various ways: Optionally, a multi-scale time-frequency analysis method is adopted. The specific steps include: designing time-frequency atoms of multiple scales; calculating the correlation coefficient between the signal and the atoms; constructing the time-frequency energy distribution; and extracting time-frequency features. Optionally, an adaptive feature learning method is adopted. The specific steps include: constructing a contrastive learning loss function; designing a data augmentation strategy; training a feature extractor; and verifying the quality of the feature representation. It can be understood that other feature learning methods can also be adopted, such as variational autoencoders, graph neural networks, etc., which are not limited here.

[0082] S203. Generate a signal attenuation compensation factor based on the low-frequency feature vector, and perform distortion compensation on the high-frequency feature vector according to the signal attenuation compensation factor to obtain a high-frequency corrected vector.

[0083] Referring to step S103, the partial discharge recognition system will perform distortion compensation on the high-frequency feature vector.

[0084] S204. Calculate the feature correlation degree between the high-frequency corrected vector and the low-frequency feature vector, and extract the transient feature information of the high-frequency corrected vector when the feature correlation degree is lower than the preset correlation threshold.

[0085] Referring to step S104, the partial discharge recognition system will extract the transient feature information of the high-frequency corrected vector.

[0086] S205. Extract the discharge pulse period from the high-frequency corrected vector to obtain a period feature sequence.

[0087] Among them, the discharge pulse period represents the time interval between adjacent discharge pulses in the partial discharge signal; the period feature sequence refers to the time series data describing the recurrence pattern of the discharge pulses; the high-frequency correction vector represents the set of high-frequency feature data after attenuation compensation.

[0088] After obtaining the high-frequency correction vector, the partial discharge recognition system needs to analyze the time characteristics of the discharge process. Specifically, the partial discharge recognition system first performs a time-domain analysis on the high-frequency correction vector to locate the time points of the discharge pulses through threshold detection. Then, it calculates the time difference between adjacent pulses to form a period sequence. Statistical analysis is performed on this sequence, including mean, variance, distribution characteristics, etc., to construct a period feature sequence. Such a sequence can reflect the dynamic change law of the discharge process.

[0089] In some embodiments, the period extraction can be achieved in various ways: Optionally, the adaptive threshold method is used to determine the pulse position and calculate the period. The specific steps include: calculating the signal envelope; setting the dynamic detection threshold; locating the pulse peak points; and statistically analyzing the adjacent peak intervals. Optionally, the spectral analysis method is used to extract the period features. The specific steps include: calculating the signal autocorrelation function; performing periodogram analysis; and extracting the main period components. It can be understood that other period analysis methods, such as wavelet transform, Hilbert transform, etc., can also be used, which are not limited here.

[0090] S206. Perform wavelet decomposition on the period feature sequence to obtain multi-scale components.

[0091] Among them, wavelet decomposition represents the process of performing multi-resolution analysis on a signal using wavelet basis functions; multi-scale components refer to the signal components at different frequency scales; wavelet basis functions refer to a family of finite support functions that satisfy specific mathematical properties.

[0092] The partial discharge recognition system needs to perform multi-scale analysis on the period feature sequence. Specifically, the partial discharge recognition system first selects a suitable wavelet basis function (such as Daubechies wavelet) and determines the decomposition level. Then, it recursively decomposes the period feature sequence to obtain the approximation coefficients and detail coefficients at different scales. These coefficients reflect the energy distribution and local characteristics of the signal in different frequency bands, forming a multi-scale component set.

[0093] In some embodiments, wavelet decomposition can be achieved in various ways: Optionally, orthogonal wavelet transform is used to achieve lossless decomposition of the signal. The specific steps include: selecting an orthogonal wavelet basis; constructing a filter bank; performing multi-layer decomposition; and saving the coefficients of each layer. Optionally, wavelet packet decomposition is used to obtain a more detailed frequency band division. The specific steps include: constructing a wavelet packet tree; calculating the node energy; selecting the optimal basis; and extracting the feature coefficients. It can be understood that other decomposition methods, such as dual-tree complex wavelet, adaptive wavelet, etc., can also be used, which are not limited here.

[0094] S207. Denoise the multi-scale components based on adaptive filtering to obtain a reconstructed signal.

[0095] Among them, adaptive filtering refers to a signal processing method that can automatically adjust filtering parameters according to signal characteristics; denoising refers to the process of suppressing or removing unnecessary noise components in a signal; a reconstructed signal refers to a signal that is re-synthesized after denoising processing.

[0096] After obtaining the multi-scale components, the partial discharge recognition system needs to perform signal optimization processing. Specifically, the partial discharge recognition system first estimates the signal-to-noise ratio of each scale component and establishes an adaptive threshold function. Then, according to the noise characteristics of different scales, the filter parameters are dynamically adjusted to denoise each component. Finally, the processed scale components are weighted and reconstructed to obtain an optimized signal. This adaptive processing method can better preserve the effective features of the signal.

[0097] It should be noted that the adaptive filtering process in this case realizes signal optimization based on the least mean square error criterion. First, estimate the signal-to-noise ratio of each scale component and construct an adaptive Wiener filter. The frequency response function of the filter is dynamically adjusted according to the local statistical characteristics of the signal, maintaining a high passband gain in the frequency band with strong signal energy and reducing the gain in the frequency band dominated by noise. The filter coefficients are updated in real time through the recursive least squares algorithm to track changes in signal characteristics. The filtered scale components are synthesized into a complete signal through the wavelet reconstruction algorithm. For example, for a certain scale component, if the signal-to-noise ratio in the local interval is 15 dB, the filter gain in the corresponding frequency band is about 0.9, while in the interval with a signal-to-noise ratio of only 5 dB, the gain is reduced to about 0.3.

[0098] In some embodiments, adaptive denoising can be achieved in multiple ways: Optionally, use empirical mode decomposition combined with the adaptive threshold method. The specific steps include: decomposing the signal into intrinsic mode functions; calculating the energy ratio of each mode function; setting an adaptive threshold according to the energy ratio; reconstructing the optimized signal. Optionally, use the Kalman filtering method to achieve dynamic tracking filtering of the signal. The specific steps include: establishing a state space model; predicting the state estimate; updating the gain matrix; fusing the measurement values. It can be understood that other denoising methods can also be used, such as Wiener filtering, particle filtering, etc., which are not limited here.

[0099] S208. Extract the characteristic information in the reconstructed signal to obtain transient characteristic information.

[0100] Among them, the reconstructed signal represents the time-domain signal re-synthesized after noise reduction processing; the feature information refers to the key parameters reflecting the discharge characteristics in the signal; the transient feature information represents the set of features that change rapidly during the discharge process, including amplitude features, time features, and frequency features; feature extraction represents the process of obtaining effective features from the signal; the feature parameters refer to the quantitative indicators used to describe the signal characteristics.

[0101] After obtaining the reconstructed signal, the partial discharge recognition system needs to extract the transient features reflecting the discharge characteristics. Specifically, the partial discharge recognition system first performs time-domain analysis on the reconstructed signal to extract amplitude features, including the maximum value, root mean square value, peak factor, etc. Then it analyzes the time features of the signal, including the rise time, duration, oscillation period, etc. Next, it performs frequency-domain analysis to extract features such as the main frequency component, frequency band width, and energy distribution. Finally, these features are standardized to form a unified transient feature vector. These feature information comprehensively describe the transient characteristics of the discharge process and provide a basis for subsequent fault diagnosis.

[0102] In some embodiments, the transient feature extraction can be achieved in various ways: Optionally, a parametric feature extraction method is adopted. The specific steps include: designing a feature extraction operator; calculating time-domain statistics; extracting frequency-domain parameters; and constructing a feature vector. Optionally, a non-parametric feature extraction method is adopted. The specific steps include: calculating the signal envelope; extracting morphological features; analyzing local waveforms; and generating feature descriptors. It can be understood that other feature extraction methods can also be adopted, such as wavelet energy features, fractal features, etc., which are not limited here.

[0103] In some embodiments, the partial discharge recognition system performs transient recognition based on the energy distribution, that is, the partial discharge recognition system calculates the time-frequency spectrogram of the reconstructed signal to obtain the energy distribution sequence; performs peak detection on the energy distribution sequence to determine the transient feature points; segments the signal based on the transient feature points to obtain the transient feature intervals, and generates transient feature information according to the transient feature intervals.

[0104] Among them, the time-frequency spectrogram represents the distribution diagram of the signal energy in the time-frequency plane; the energy distribution sequence refers to the sequence of the signal energy changing with time; the transient feature points represent the moment points when the signal changes suddenly; the transient feature intervals refer to the time intervals containing the transient process; the transient feature information includes feature parameters such as amplitude and duration.

[0105] After obtaining the reconstructed signal, the partial discharge recognition system needs to extract transient features. Specifically, the partial discharge recognition system first uses the sliding window short-time Fourier transform to calculate the time-frequency spectrogram of the signal and analyze the distribution characteristics of energy in the time-frequency domain. Then, the adaptive threshold algorithm is applied to the energy distribution sequence to detect the energy mutation points and determine the possible occurrence times of transient events. Based on these feature points, the signal is divided into multiple intervals, and the feature parameters of each interval are extracted, including the maximum amplitude, rise time, oscillation frequency, etc., to form a complete description of transient features.

[0106] It should be noted that the feature extraction process integrates time-domain and frequency-domain analysis. In the time domain, the mutation points of the signal are detected through adaptive threshold, and the threshold is set as the local mean plus 3 times the standard deviation. The morphological features of the detected mutation interval are calculated, including peak value, rise time, duration, etc. In the frequency domain, wavelet packet decomposition is used to obtain the energy distribution of different frequency bands, and the optimal wavelet basis is selected to make the sparse representation of the signal optimal. Then, features such as the frequency band energy ratio and spectral entropy are calculated based on the wavelet coefficients. Finally, the time-domain and frequency-domain features are reduced in dimension through principal component analysis, and the principal components with a cumulative contribution rate exceeding 95% are retained.

[0107] In some embodiments, transient feature extraction can be achieved in multiple ways: Optionally, a multi-resolution analysis method is adopted. The specific steps include: performing wavelet packet decomposition; selecting the optimal basis function; locating the feature points; extracting local features. Optionally, a sparse representation method is adopted. The specific steps include: constructing an over-complete dictionary; solving the sparse coefficients; reconstructing the transient components; extracting the feature parameters. It can be understood that other feature extraction methods can also be adopted, such as empirical mode decomposition, variational mode decomposition, etc., which are not limited herein.

[0108] S209. Integrate the transient feature information and the low-frequency feature vector to obtain a low-frequency correction vector.

[0109] Referring to step S105, the partial discharge recognition system will correct the low-frequency feature vector.

[0110] S210. Input the high-frequency correction vector and the low-frequency correction vector into the partial discharge recognition model to obtain the partial discharge type and the recognition confidence level.

[0111] Referring to step S106, the partial discharge recognition system will determine the partial discharge type and the recognition confidence level based on the model.

[0112] In some embodiments, the partial discharge recognition system processes data based on multiple models. That is, the partial discharge recognition system inputs the high-frequency correction vector into the long short-term memory network to obtain a time-series feature sequence; inputs the low-frequency correction vector into the Transformer encoder to obtain an attention feature sequence; calculates the similarity matrix between the time-series feature sequence and the attention feature sequence through contrastive learning, and generates the partial discharge type and recognition confidence based on the similarity matrix.

[0113] Among them, the long short-term memory network represents a recurrent neural network structure that can handle long-term dependencies; the time-series feature sequence refers to the sequence data reflecting the time-evolution characteristics of the signal; the Transformer encoder refers to the feature encoder based on the self-attention mechanism; the attention feature sequence represents the feature sequence weighted by attention; the similarity matrix refers to the correlation measurement result between different feature sequences.

[0114] After obtaining the correction vector, the partial discharge recognition system needs to perform time-series feature analysis and pattern recognition. Specifically, the partial discharge recognition system first inputs the high-frequency correction vector into the LSTM network, captures the long-term and short-term dependencies of the signal through the gating mechanism, and generates a feature sequence containing time-series dynamic information. At the same time, it inputs the low-frequency correction vector into the Transformer encoder, uses the multi-head self-attention mechanism to learn the correlation between features, and obtains a feature sequence with global context. Then, it calculates the similarity matrix between the two feature sequences through the contrastive learning framework and analyzes their matching degree. Finally, it designs a classification decision rule based on the similarity matrix and outputs the most likely discharge type and its confidence.

[0115] It should be noted that the partial discharge recognition model structure adopts a hybrid deep learning network. After the input layer, there are parallel CNN and LSTM branches: the CNN branch uses residual connections to extract spatial features, and the LSTM branch captures time-series dependencies. Then, the features of the two branches are fused through the attention mechanism, and the attention weights are calculated by the softmax function. The fused features are sent to the fully connected layer, and dropout is used to prevent overfitting. The output layer uses focal loss to handle the problem of sample imbalance and assigns higher weights to difficult-to-classify samples.

[0116] During the training phase of the long short-term memory network (LSTM) model, the high-frequency correction vector sequence is used as the input and trained with time-series annotation data. The training objective is to minimize the sequence prediction error, and the mean square error or sequence correlation is used as the loss function. The model includes three gating units, namely the input gate, forget gate, and output gate, and a memory unit. These components work together to learn the long-term dependencies in the sequence data. In practical applications, the model receives the high-frequency correction vector, processes the sequence information through the gating mechanism and the memory unit, and outputs a feature sequence containing time-series dynamic features.

[0117] In addition, during the training phase, the Transformer encoder model uses low-frequency correction vectors as input data and is trained in a self-supervised learning manner. During the training process, a masked language model task and a multi-head self-attention mechanism are used, and the optimization objective is to maximize the mutual information of the feature representations. The model architecture includes multiple layers of self-attention layers and feed-forward neural network layers, and each layer is equipped with residual connections and layer normalization. When in use, the model captures global dependencies in the input sequence by calculating the attention weight matrix and generates an attention feature sequence with context awareness.

[0118] In some embodiments, sequence feature analysis can be achieved in various ways: Optionally, a bidirectional attention fusion method is adopted. The specific steps include: constructing a bidirectional LSTM network; calculating temporal attention weights; fusing bidirectional features; generating an enhanced feature sequence. Optionally, a hierarchical contrast learning method is adopted. The specific steps include: designing a multi-level feature extractor; constructing a multi-granularity contrast loss; optimizing the feature representation; learning the discriminant boundary. It can be understood that other sequence analysis methods can also be adopted, such as gated recurrent units, temporal convolutional networks, etc., which are not limited here.

[0119] S211. Perform feature fusion on the high-frequency correction vector and the low-frequency correction vector to obtain a fused feature.

[0120] Among them, feature fusion refers to the process of comprehensively integrating feature information from different sources; the fused feature refers to the new feature representation obtained through feature fusion; the feature weight refers to the importance degree of different features in the fusion process.

[0121] The partial discharge recognition system needs to effectively integrate high- and low-frequency features. Specifically, the partial discharge recognition system first performs dimension alignment and normalization processing on the two feature vectors. Then, the feature importance weights are calculated, which can be based on information gain or feature correlation analysis. Finally, methods such as weighted fusion or non-linear mapping are used to combine the two feature vectors into a unified feature representation. This fusion strategy can make full use of the complementary information in different frequency bands.

[0122] In some embodiments, feature fusion can be achieved in various ways: Optionally, a feature fusion method using an attention mechanism is adopted. The specific steps include: calculating feature attention weights; generating an attention map; weighted fusing features; optimizing the attention parameters. Optionally, a feature fusion method using an autoencoder is adopted. The specific steps include: constructing an encoder network; learning the feature mapping relationship; decoding and reconstructing features; extracting bottleneck layer features. It can be understood that other fusion methods can also be adopted, such as graph feature fusion, hierarchical fusion, etc., which are not limited here.

[0123] S212. Calculate the correlation coefficient between the fused feature and the preset partial discharge template to obtain a matching degree matrix.

[0124] Among them, the preset partial discharge template represents a set of standard characteristic patterns of known various discharge types; the correlation coefficient is a statistic that measures the similarity degree between two feature vectors; the matching degree matrix is a correlation coefficient matrix between the fusion feature and each template.

[0125] The partial discharge recognition system needs to evaluate the matching degree between the fusion feature and the standard template. Specifically, the partial discharge recognition system first loads the preset discharge template library, which contains the typical characteristic patterns of different types of discharges. Then, it uses various correlation measurement methods, such as Pearson correlation coefficient, cosine similarity, etc., to calculate the correlation coefficients between the fusion feature and each template. These coefficients are organized in matrix form to reflect the matching relationship between the feature and each type of template.

[0126] In some embodiments, the correlation calculation can be implemented in various ways: Optionally, a multi-scale correlation analysis method is adopted. The specific steps include: constructing a feature pyramid; calculating multi-scale correlation coefficients; weighted fusion of the correlation degrees; generating a matching degree sequence. Optionally, a kernel correlation analysis method is adopted. The specific steps include: selecting a kernel function; calculating a kernel matrix; solving a feature mapping; calculating a kernel correlation coefficient. It can be understood that other correlation analysis methods, such as mutual information, canonical correlation analysis, etc., can also be adopted, which are not limited herein.

[0127] S213. Screen out the characteristic pattern with the maximum matching degree from the matching degree matrix to obtain the recognition result.

[0128] Among them, the characteristic pattern represents a typical characteristic combination of a certain type of discharge; the maximum matching degree refers to the maximum correlation degree in the correlation coefficient matrix; the recognition result includes the discharge type and the corresponding confidence information.

[0129] The partial discharge recognition system needs to determine the final recognition result based on the matching degree matrix. Specifically, the partial discharge recognition system first normalizes the matching degree matrix to make the matching degrees between different types comparable. Then, it uses a maximum value detection algorithm to find the characteristic pattern with the highest matching degree and its corresponding discharge type. At the same time, it calculates the difference between the maximum matching degree and the sub-optimal matching degree as the reliability index of the recognition result. If multiple patterns have similar matching degrees, the system will further analyze the distinctiveness of the features to improve the recognition accuracy.

[0130] In some embodiments, the optimal mode selection can be achieved in multiple ways: Optionally, a multi-criteria decision-making method is used for mode selection. The specific steps include: setting multiple evaluation indicators; calculating the comprehensive score; establishing a decision matrix; and selecting the optimal solution. Optionally, a probability inference method is used to determine the optimal mode. The specific steps include: constructing a probabilistic graphical model; calculating the posterior probability; maximum likelihood estimation; and outputting the recognition result. It can be understood that other decision-making methods, such as fuzzy inference, evidence theory, etc., can also be used, which are not limited herein.

[0131] S214. Update the neural network parameters of the partial discharge recognition model according to the recognition result.

[0132] Among them, the neural network parameters refer to the learnable parameters such as weights and biases in the deep learning model; parameter update refers to the process of adjusting the model parameters according to new samples; and the learning rate refers to the step coefficient of parameter update.

[0133] The partial discharge recognition system needs to continuously optimize the model performance through online learning. Specifically, the partial discharge recognition system first compares the recognition result with the actual label and calculates the recognition error. Then, based on the error gradient, the backpropagation algorithm is used to update the parameters of each layer of the neural network. The historical gradient information is considered during the parameter update process, the learning rate is dynamically adjusted, and regularization techniques are used to prevent overfitting. This incremental learning mechanism enables the model to adapt to the dynamic changes of discharge characteristics.

[0134] It should be noted that the parameter update of the partial discharge recognition model adopts a dynamic adaptive optimization strategy. First, the sample weights are calculated based on the confidence distribution output by the model. Samples with lower confidence obtain higher weights during parameter update. Then, the gradient of the weighted loss function with respect to the network parameters is calculated, and the gradient calculation uses reverse automatic differentiation technology. The optimizer uses the Adam algorithm, combines the momentum term and the adaptive learning rate, accelerates convergence, and avoids local optima. An L2 regularization term is introduced during parameter update to prevent overfitting, and the regularization coefficient is dynamically adjusted according to the number of training rounds. For example, when the recognition confidence of a certain sample is lower than 0.6, its weight coefficient will be increased to 1.5 times, prompting the model to pay more attention to such difficult-to-classify samples.

[0135] In some embodiments, parameter update can be achieved in multiple ways: Optionally, an adaptive learning rate optimization algorithm is used. The specific steps include: calculating the parameter gradient; estimating the first and second moments of the gradient; adaptively adjusting the learning rate; and updating the model parameters. Optionally, a transfer learning method is used to update the model. The specific steps include: freezing the parameters of the base layer; fine-tuning the top-level network; incrementally training new samples; and verifying the model performance. It can be understood that other parameter optimization methods, such as meta-learning, lifelong learning, etc., can also be used, which are not limited herein.

[0136] In some embodiments, the partial discharge recognition system optimizes the model based on feedback. That is, the partial discharge recognition system calculates the confidence distribution of the recognition results to obtain the feedback weight coefficient; performs backpropagation on the neural network based on the feedback weight coefficient to obtain the gradient value; adjusts the convolutional layer parameters according to the gradient value to obtain the optimized parameters, and writes the optimized parameters into the partial discharge recognition model.

[0137] Among them, the confidence distribution represents the prediction probability distribution of the model for different categories; the feedback weight coefficient refers to the weight factor used to adjust the parameter update amplitude; the gradient value represents the partial derivative of the loss function with respect to the network parameters; the optimized parameters refer to the updated neural network weight parameters.

[0138] After obtaining the recognition results, the partial discharge recognition system needs to optimize the model parameters. Specifically, the partial discharge recognition system first analyzes the probability distribution output by the model, calculates the uncertainty of the prediction results, and generates the weight coefficient reflecting the prediction reliability. Then, these weights are combined with the prediction errors, and the gradients of each layer's parameters are calculated through the backpropagation algorithm. Based on the calculated gradient values, an adaptive optimization algorithm is used to update the network parameters, especially the weights and biases of the convolutional layer. Finally, the optimized parameters are saved into the model to complete the online update of the model.

[0139] In some embodiments, parameter optimization can be achieved in multiple ways: Optionally, a dynamic learning rate adjustment method is adopted. The specific steps include: estimating the gradient variance; calculating the adaptive step size; updating the momentum term; adjusting the learning rate. Optionally, a hierarchical parameter optimization method is adopted. The specific steps include: calculating the layer sensitivity; designing the hierarchical learning rate; updating the parameters layer by layer; verifying the optimization effect. It can be understood that other optimization methods can also be adopted, such as second-order optimization, meta-learning, etc., which are not limited here.

[0140] In the embodiments of the present application, due to the adoption of the multi-source signal collaborative processing and deep feature learning methods, the transient characteristics of high-frequency signals and the continuous characteristics of low-frequency signals can be utilized simultaneously, effectively solving the problems of low accuracy and poor reliability of traditional single-frequency band detection methods in complex environments, and thus realizing the accurate recognition of partial discharges and early fault warnings. Through the feature fusion mechanism and the adaptive compensation algorithm, the system can effectively cope with signal attenuation and environmental interference, and improve the detection sensitivity. The online optimization method based on deep learning enables the system to have the ability of continuous learning and can adapt to the dynamic changes of the equipment operation state. The multi-level feature extraction and recognition strategy not only improves the diagnostic accuracy of the system, but also realizes the recognition of composite discharge faults, providing reliable technical support for the condition monitoring and preventive maintenance of power equipment.

[0141] The partial discharge recognition system in the embodiments of the present invention application will be described from the perspective of hardware processing. Please refer to Figure 3, which is a schematic structural diagram of an entity device in the partial discharge recognition system according to an embodiment of the present application.

[0142] It should be noted that Figure 3 The structure of the partial discharge recognition system shown is only an example and should not impose any limitations on the functions and usage scope of the embodiments of the present invention.

[0143] As Figure 3 shown, the partial discharge recognition system includes a CPU 301, which can perform various appropriate actions and processes according to the program stored in the ROM 302 or the program loaded into the RAM 303 from the storage section 308, such as executing the method described in the above embodiments. In the RAM 303, various programs and data required for system operation are also stored. The CPU 301, ROM 302, and RAM 303 are connected to each other via a bus 304. The I / O interface 305 is also connected to the bus 304.

[0144] The following components are connected to the I / O interface 305: an input section 306 including an audio input device, a button switch, etc.; an output section 307 including a liquid crystal display (LCD), an audio output device, an indicator light, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. The drive 310 is also connected to the I / O interface 305 as needed. A removable medium 311, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 310 as needed so that a computer program read from it can be installed into the storage section 308 as needed.

[0145] Specifically, according to an embodiment of the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present invention includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication section 309, and / or installed from the removable medium 311. When the computer program is executed by the CPU 301, various functions defined in the present invention are executed.

[0146] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present invention. Among them, each block in the flowchart or block diagram may represent a module, a program segment, or a part of code, and the above-mentioned module, program segment, or part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the block may occur in a different order from that marked in the accompanying drawings.

[0147] Specifically, the partial discharge recognition system of this embodiment includes a processor and a memory. A computer program is stored on the memory. When the computer program is executed by the processor, it implements the partial discharge signal recognition method based on deep learning provided in the above-mentioned embodiment.

[0148] On the other hand, the present invention also provides a computer-readable storage medium. This storage medium may be included in the partial discharge recognition system described in the above-mentioned embodiment; or it may exist separately and not be assembled into the partial discharge recognition system. The above storage medium carries one or more computer programs. When the one or more computer programs are executed by a processor of a partial discharge recognition system, the partial discharge recognition system is enabled to implement the partial discharge signal recognition method based on deep learning provided in the above-mentioned embodiment.

[0149] As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present application.

[0150] As used in the above embodiments, depending on the context, the term "when..." may be interpreted to mean "if...", or "after...", or "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if detecting (the stated condition or event)" may be interpreted to mean "if determining...", or "in response to determining...", or "when detecting (the stated condition or event)", or "in response to detecting (the stated condition or event)".

Claims

1. A method for identifying partial discharge signals based on deep learning, characterized in that, Applied to a partial discharge recognition system, the method includes: Receiving multi-source monitoring data of partial discharge, and dividing the multi-source monitoring data into a high-frequency signal group and a low-frequency signal group according to a preset discrimination frequency; Respectively extracting the feature information of the high-frequency signal group and the low-frequency signal group, and respectively inputting the corresponding feature information into a first deep neural network and a second deep neural network to obtain a high-frequency feature vector and a low-frequency feature vector; Generating a signal attenuation compensation factor based on the low-frequency feature vector, and performing distortion compensation on the high-frequency feature vector according to the signal attenuation compensation factor to obtain a high-frequency corrected vector; Calculating the feature correlation degree between the high-frequency corrected vector and the low-frequency feature vector, and when the feature correlation degree is lower than a preset correlation threshold, extracting the transient feature information of the high-frequency corrected vector; Fusing the transient feature information and the low-frequency feature vector to obtain a low-frequency corrected vector; Inputting the high-frequency corrected vector and the low-frequency corrected vector into a partial discharge recognition model to obtain the partial discharge type and the recognition confidence.

2. The method according to claim 1, wherein The step of respectively extracting the feature information of the high-frequency signal group and the low-frequency signal group, and respectively inputting the corresponding feature information into a first deep neural network and a second deep neural network to obtain a high-frequency feature vector and a low-frequency feature vector specifically includes: Performing time-frequency transformation on the high-frequency signal group and the low-frequency signal group to obtain a time-frequency feature map; Performing hierarchical convolution operations on the time-frequency feature map based on a convolutional neural network to obtain a feature tensor; Performing self-supervised learning encoding on the feature tensor to generate a feature representation matrix; Inputting the feature representation matrix into the corresponding first deep neural network and second deep neural network to respectively obtain the high-frequency feature vector and the low-frequency feature vector.

3. The method according to claim 1, characterized in that The step of inputting the high-frequency corrected vector and the low-frequency corrected vector into a partial discharge recognition model to obtain the partial discharge type and the recognition confidence specifically includes: Inputting the high-frequency corrected vector into a long short-term memory network to obtain a time series feature sequence; Inputting the low-frequency corrected vector into a Transformer encoder to obtain an attention feature sequence; Calculating a similarity matrix between the time series feature sequence and the attention feature sequence through contrastive learning, and generating the partial discharge type and the recognition confidence based on the similarity matrix.

4. The method according to claim 1, characterized in that, Before the step of fusing the transient feature information and the low-frequency feature vector to obtain a low-frequency corrected vector, the method further includes: Extracting a discharge pulse period from the high-frequency corrected vector to obtain a period feature sequence; Performing wavelet decomposition on the period feature sequence to obtain multi-scale components; Performing noise reduction on the multi-scale components based on adaptive filtering to obtain a reconstructed signal; Extracting the feature information in the reconstructed signal to obtain transient feature information.

5. The method according to claim 4, wherein After the step of extracting the feature information in the reconstructed signal to obtain transient feature information, it specifically includes: Calculating the time-frequency spectrogram of the reconstructed signal to obtain an energy distribution sequence; Performing peak detection on the energy distribution sequence to determine transient feature points; Segment the signal based on the transient feature points to obtain transient feature intervals, and generate transient feature information according to the transient feature intervals.

6. The method according to claim 1, wherein After the step of inputting the high-frequency correction vector and the low-frequency correction vector into the partial discharge recognition model to obtain the partial discharge type and the recognition confidence, the method further includes: Perform feature fusion on the high-frequency correction vector and the low-frequency correction vector to obtain a fused feature; Calculate the correlation coefficient between the fused feature and a preset partial discharge template to obtain a matching degree matrix; Select the feature pattern with the largest matching degree from the matching degree matrix to obtain an identification result; Update the neural network parameters of the partial discharge recognition model according to the identification result.

7. The method according to claim 6, wherein The step of updating the neural network parameters of the partial discharge recognition model according to the identification result specifically includes: Calculate the confidence distribution of the identification result to obtain a feedback weight coefficient; Perform backpropagation on the neural network based on the feedback weight coefficient to obtain a gradient value; Adjust the convolutional layer parameters according to the gradient value to obtain optimized parameters, and write the optimized parameters into the partial discharge recognition model.

8. A partial discharge recognition system, characterized in that, The partial discharge recognition system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the partial discharge recognition system to execute the method according to any one of claims 1-7.

9. A computer-readable storage medium, comprising instructions, characterized in that, When the instruction runs on the partial discharge recognition system, the partial discharge recognition system is enabled to execute the method according to any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product runs on the partial discharge recognition system, the partial discharge recognition system is enabled to execute the method according to any one of claims 1-7.

Images