Partial discharge signal identification method and system based on deep learning, and storage medium

Abstract

The invention discloses a partial discharge signal identification method and system based on deep learning and a storage medium, and relates to the field of computer systems based on a specific calculation model, and the method comprises the steps: dividing multi-source monitoring data into a high-frequency signal group and a low-frequency signal group; respectively inputting corresponding feature information to the first deep neural network and the 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 to obtain a high-frequency correction vector; calculating the feature correlation degree of the high-frequency correction vector and the low-frequency feature vector, and when the feature correlation degree is lower than a preset correlation threshold value, extracting transient feature information of the high-frequency correction vector, and fusing the low-frequency feature vector to obtain a low-frequency correction vector; and inputting the high-frequency correction vector and the low-frequency correction vector to a partial discharge identification model to obtain a partial discharge type and identification confidence. According to the invention, the identification accuracy of the partial discharge signal can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer systems based on specific computing models, and in particular to a local discharge signal recognition method, system and storage medium based on deep learning. Background Art

[0002] As the scale of power systems continues to expand and the equipment operating environment becomes increasingly complex, the importance of partial discharge detection of power equipment has become increasingly prominent. As an important means of evaluating the insulation status of high-voltage electrical equipment, partial discharge detection is of great significance for ensuring the safe and stable operation of power systems. In particular, in key equipment such as large substations, GIS equipment, and power cables, accurate identification and timely detection of partial discharge faults play a key role in preventing major accidents and achieving 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 detection signal exceeds the set threshold. For example, in GIS equipment detection, the discharge signal is collected by a UHF sensor, and the amplitude and frequency thresholds are set for fault judgment; in transformer detection, acoustic sensing and electrical quantity measurement are combined to set fixed criteria such as acoustic signal strength and electrical quantity amplitude for fault identification. The detection equipment analyzes the collected single or multiple signals separately, and judges whether the equipment has a partial discharge fault by comparing them with the preset threshold.

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

[0005] The present application provides a partial discharge signal recognition method, system and storage medium based on deep learning, which are used to improve the recognition accuracy of partial discharge signals.

[0006] In a first aspect, the present application provides a local discharge signal recognition method based on deep learning, which is applied to a local discharge recognition system, and the method includes: receiving multi-source monitoring data of local discharge, and dividing the multi-source monitoring data into a high-frequency signal group and a low-frequency signal group according to a preset distinction frequency; extracting feature information of the high-frequency signal group and the low-frequency signal group respectively, and inputting the corresponding feature information into a first deep neural network and a second deep neural network respectively 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 correction vector; calculating the feature correlation between the high-frequency correction vector and the low-frequency feature vector, and extracting the transient feature information of the high-frequency correction vector when the feature correlation is lower than a preset correlation threshold; fusing the transient feature information and the low-frequency feature vector to obtain a low-frequency correction vector; inputting the high-frequency correction vector and the low-frequency correction vector into a local discharge recognition model to obtain a local discharge type and recognition confidence.

[0007] In the above embodiment, the partial discharge identification system divides the multi-source monitoring data into high-frequency and low-frequency signal groups, extracts feature vectors through a deep neural network, uses the low-frequency feature vectors to generate attenuation compensation factors to compensate for the distortion of the high-frequency feature vectors, and then extracts transient feature information based on feature correlation; through multi-level feature fusion and optimization, high-precision identification of partial discharge signals is achieved; the problem of low identification accuracy of the traditional fixed threshold method in complex environments is effectively solved, and the reliability of partial discharge detection is improved.

[0008] In combination with some embodiments of the first aspect, in some embodiments, the steps of respectively extracting feature information of the high-frequency signal group and the low-frequency signal group, and respectively inputting the corresponding feature information into the first deep neural network and the second deep neural network to obtain high-frequency feature vectors and low-frequency feature vectors specifically include: performing time-frequency transformation on the high-frequency signal group and the low-frequency signal group to obtain a time-frequency feature map; performing layered 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; and inputting the feature representation matrix into the corresponding first deep neural network and the second deep neural network to obtain high-frequency feature vectors and low-frequency feature vectors, respectively.

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

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

[0011] In the above embodiment, the partial discharge identification system inputs the high-frequency correction vector into the long short-term memory network to extract the timing features, inputs the low-frequency correction vector into the Transformer encoder to obtain the attention features, and then calculates the similarity matrix through contrastive learning; this deep learning architecture makes full use of the sequence modeling capabilities to achieve accurate identification of discharge types and reliable confidence assessment.

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

[0013] In the above embodiment, the partial discharge identification system extracts the discharge pulse period characteristics from the high-frequency correction vector, obtains multi-scale components through wavelet decomposition, and uses adaptive filtering for noise reduction processing; this signal processing method effectively extracts the essential characteristics of the discharge signal, suppresses the influence of interference noise, and improves the purity and reliability of the characteristic information.

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

[0015] In the above embodiment, the partial discharge identification system calculates the time-frequency spectrum of the reconstructed signal to obtain the energy distribution, performs peak detection to determine the characteristic points and extracts the characteristic intervals in segments; this analysis method based on energy characteristics can accurately capture the transient characteristics of the discharge signal.

[0016] In combination with some embodiments of the first aspect, in some embodiments, after the step of inputting a high-frequency correction vector and a low-frequency correction vector into a partial discharge identification model to obtain a partial discharge type and an identification 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 matrix; screening out a feature pattern with the highest matching degree from the matching matrix to obtain an identification result; and updating the neural network parameters of the partial discharge identification model according to the identification result.

[0017] In the above embodiment, the partial discharge identification system performs feature fusion on the high- and low-frequency correction vectors, obtains the identification results 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 improves the user experience.

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

[0019] In the above embodiment, the partial discharge identification system calculates the feedback weight based on the confidence distribution of the identification result and adjusts the network parameters through back propagation; this dynamic optimization mechanism enables the model to continuously adapt to new discharge characteristic patterns and continuously improve the identification accuracy.

[0020] In a second aspect, an embodiment of the present application provides a partial discharge identification system, which 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 identification system to perform the method described in the first aspect and any possible implementation method of the first aspect.

[0021] In a third aspect, an embodiment of the present application provides a computer program product comprising instructions. When the above-mentioned computer program product runs on a partial discharge identification system, the above-mentioned partial discharge identification system executes the method described in the first aspect and any possible implementation method of the first aspect.

[0022] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, comprising instructions. When the instructions are executed on a partial discharge identification system, the partial discharge identification system executes the method described in the first aspect and any possible implementation method of the first aspect.

[0023] It can be understood that the partial discharge identification system provided in the second aspect, 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 embodiment of the present application. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding method, which will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. Due to the adoption of high- and low-frequency separation and deep learning feature extraction methods of multi-source monitoring data, as well as attenuation compensation and transient feature fusion technology based on low-frequency feature vectors, it is possible to fully utilize the complementary advantages of signals in different frequency bands to achieve accurate extraction and fusion of signal features, effectively solving the problems of incomplete feature extraction of a single frequency band and identification deviation caused by signal attenuation in the existing technology, thereby achieving high-precision and high-reliability identification of partial discharge signals and improving the accuracy and reliability of power equipment status monitoring.

[0025] 2. Due to the use of signal processing methods such as discharge pulse period extraction, wavelet multi-scale decomposition and adaptive filtering, the discharge feature information can be effectively separated from the complex original signal, and through multi-level signal reconstruction and feature extraction, the problems of discharge signals being susceptible to interference and inaccurate feature extraction in the existing technology are effectively solved, thereby achieving accurate capture and reliable characterization of discharge characteristics, and improving the anti-interference ability of feature extraction and the stability of recognition.

[0026] 3. Due to the adoption of optimization mechanisms such as feature fusion, template matching and adaptive update of network parameters, it is possible to comprehensively utilize multi-source feature information for recognition and judgment, and improve recognition performance through continuous model optimization, effectively solving the problems of poor model adaptability and low reliability of recognition results in existing technologies, thereby achieving optimization and performance improvement of the recognition model and enhancing the practicality and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of a partial discharge signal recognition method based on deep learning in an embodiment of the present application; Figure 2 is another flow chart of the partial discharge signal recognition method based on deep learning in an embodiment of the present application; Figure 3 It is a schematic diagram of the structure of a physical device of the partial discharge identification system in the embodiment of the present application. DETAILED DESCRIPTION

[0028] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to be used as limitations to the present application. As used in the specification of the present application, the singular expressions "one", "a kind of", "above", "the" and "this" are intended to also include plural expressions, unless there is a clear indication to the contrary in the context. It should also be understood that the term "and / or" used in the present application refers to any or all possible combinations comprising one or more of the listed items.

[0029] In the following, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as suggesting or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features, and in the description of the embodiments of the present application, unless otherwise specified, "plurality" means two or more.

[0030] For ease of understanding, the application scenarios of the embodiments of the present application are introduced below.

[0031] GIS equipment in the power distribution room of a large substation may experience partial discharge faults during long-term operation. 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 single-frequency sensors, which are easily affected by external interference and signal attenuation, resulting in false alarms or missed alarms. Especially in complex electromagnetic environments, environmental noise, crosstalk signals, etc. will seriously affect the accuracy of detection. In addition, different types of partial discharges have different characteristic manifestations in different frequency bands. The detection method of a single frequency band is difficult to fully capture the discharge characteristics, reducing the reliability of fault diagnosis.

[0032] In the related art, partial discharge monitoring and identification can be achieved by adopting a single frequency band signal acquisition and a fixed threshold detection method. The following introduces a scenario of using a partial discharge signal identification method based on deep learning in the related art.

[0033] An electric power company uses an existing partial discharge monitoring system to monitor transformers online. The system uses a fixed threshold signal processing method and a simple pattern recognition algorithm. When partial discharge occurs in the transformer due to internal insulation defects, the high-frequency characteristics are severely distorted due to attenuation and distortion during signal transmission. At the same time, the system cannot effectively distinguish between corona discharge generated by normal operation and dangerous internal discharge, which will cause false alarms. Maintenance personnel need to conduct frequent on-site inspections, which increases operation and maintenance costs. In addition, the system cannot adapt to the dynamic changes in the operating status of the equipment, and the performance of the recognition model gradually degrades.

[0034] The partial discharge signal recognition method based on deep learning in the embodiment of the present application is adopted to achieve accurate and reliable discharge type recognition through high- and low-frequency feature fusion and adaptive compensation mechanism, which not only improves the detection sensitivity but also reduces the false alarm rate. The following introduces the scenario of using the partial discharge signal recognition method based on deep learning in the present application.

[0035] A smart substation uses the multi-source partial discharge identification system proposed in this application for equipment monitoring. The system collects high-frequency and low-frequency signals at the same time, and realizes feature extraction and fault identification through adaptive signal processing and deep learning algorithms. When a suspended potential discharge occurs in a certain combination electrical appliance, 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 online learning mechanisms, the system can continuously optimize the recognition model and improve the diagnostic accuracy. Even when the signal is severely interfered, it can still maintain a high detection reliability.

[0036] It can be seen that the deep learning and online optimization methods in the embodiments of the present application can not only realize intelligent fault diagnosis, but also effectively solve problems such as signal attenuation distortion and environmental interference, thereby achieving continuous optimization of the monitoring system.

[0037] For ease of understanding, the following describes the process of the method provided by this implementation in combination with the above scenario. Figure 1 , which is a flow chart of a local discharge signal recognition method based on deep learning in an embodiment of the present application.

[0038] S101 , 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 distinction frequency.

[0039] Among them, local discharge refers to the local breakdown discharge phenomenon that occurs in the insulating medium of power equipment; multi-source monitoring data refers to the discharge signal data collected simultaneously by multiple sensing devices such as ultra-high frequency sensors, acoustic sensors, and ground current sensors; the preset distinction frequency represents the frequency threshold used to distinguish high-frequency and low-frequency signals, which is set between 1MHz and 100MHz; the high-frequency signal group refers to a set of signals with a frequency higher than the preset distinction frequency, which mainly includes the transient pulse characteristics generated by the discharge; the low-frequency signal group refers to a set of signals with a frequency lower than the preset distinction frequency, which mainly includes the continuity characteristics during the discharge process.

[0040] When the partial discharge identification system detects partial discharge in power equipment, it needs to classify and process the signals collected by multiple sensors. Specifically, the partial discharge identification system first receives the original signal data collected by various sensors, and performs time domain synchronization and amplitude normalization preprocessing on these data. Then, according to the pre-set frequency demarcation point (such as 10MHz), a bandpass 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, and the low-frequency signal group contains the slow-changing characteristics of the discharge process. This grouping method can adopt corresponding processing strategies according to the signal characteristics of different frequency bands.

[0041] In some embodiments, the frequency band division of the signal can be achieved in a variety of ways: optionally, using a wavelet transform method, first performing multi-scale decomposition on the original signal to obtain wavelet coefficients of different frequency bands, and then reorganizing the wavelet coefficients into high- and low-frequency signal groups according to a preset distinction frequency. The specific steps include: performing a discrete wavelet transform on the original signal to obtain approximate coefficients and detail coefficients; determining the number of reorganization layers according to a preset frequency; reconstructing detail coefficients higher than the number of reorganization layers into high-frequency signals, and reconstructing approximate coefficients lower than the number of reorganization layers into low-frequency signals. Optionally, using a digital filtering method, a bandpass filter is designed to separate the frequency bands of the original signal. The specific steps include: designing the cutoff frequency and transition band parameters of high-pass and low-pass filters according to the preset distinction frequency; performing high-pass and low-pass filtering on the original signal to obtain preliminary separation results; and ensuring the integrity of the separated signal through energy compensation. It is understandable that other signal processing methods can also be used to achieve frequency band division, such as Fourier transform, Hilbert transform, etc., which are not limited here.

[0042] S102, extracting feature information of the high-frequency signal group and the low-frequency signal group respectively, and inputting 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.

[0043] Among them, the feature information represents the characteristics of the signal, including time domain, frequency domain and time-frequency domain features; the first deep neural network and the second deep neural network refer to deep learning models designed for the characteristics of signals in different frequency bands; the feature vector refers to the high-dimensional feature representation after neural network mapping.

[0044] The partial discharge identification system needs to extract effective features from signals in different frequency bands. Specifically, the partial discharge identification 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.; and performs statistical analysis on the low-frequency signal group to extract steady-state features, such as mean, variance, spectral energy, etc. These features are then input into a specially designed deep neural network, and through multi-layer nonlinear transformations, the original features are mapped to a more discriminative feature space to obtain a fixed-dimensional feature vector. This dual-stream feature extraction architecture can make full use of the characteristics of signals in different frequency bands.

[0045] In some embodiments, feature extraction and mapping can be achieved in a variety of ways: Optionally, a time-frequency joint analysis method is used to construct a multi-scale feature extraction network. Specific steps include: calculating the time-frequency graph of the signal; designing a multi-scale convolution layer to extract local features; integrating features through pooling layers and fully connected layers. Optionally, an autoencoder structure is used to achieve unsupervised learning of features. Specific steps include: designing encoder and decoder networks; training the network through reconstruction loss; extracting encoder output as a feature vector. It is understandable that other feature learning methods can also be used, such as neural networks, attention mechanisms, etc., which are not limited here.

[0046] 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 correction vector.

[0047] 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 the 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 refers to the process of restoring and correcting the signal through the compensation factor; the high-frequency correction vector refers to the high-frequency feature vector after attenuation compensation, which more accurately reflects the characteristics of the original discharge signal.

[0048] After obtaining the high- and low-frequency feature vectors, the partial discharge identification system needs to compensate for the attenuation distortion of the high-frequency signal. Specifically, the partial discharge identification system first analyzes the energy distribution and spectral characteristics of the low-frequency feature vectors and establishes a signal transmission attenuation model. Based on this model, the attenuation degree of different frequency bands is calculated to generate the corresponding compensation factor matrix. The compensation factor matrix is ​​then dimensionally matched and mathematically operated with the high-frequency feature vector to achieve amplitude and phase compensation of the high-frequency feature, and finally obtain the corrected high-frequency feature vector. This compensation mechanism can effectively restore the original characteristics of the high-frequency signal.

[0049] It should be noted that the distortion compensation process adopts a nonlinear mapping compensation mechanism. First, an attenuation transfer function model is constructed based on the low-frequency eigenvector, and the amplitude and phase attenuation characteristics during signal transmission are obtained through polynomial fitting. For amplitude attenuation, a frequency-dependent exponential attenuation model A(f)=A0*exp(-αf) is established, where α is the attenuation coefficient, which is estimated by the low-frequency signal energy spectrum. For phase distortion, a frequency-dependent linear phase model φ(f)=2πfτ+φ0 is adopted, where τ is the group delay, which is estimated by the low-frequency signal phase spectrum. Then the compensation factor K(f)=1 / A(f)*exp(-jφ(f)) is calculated, and the high-frequency eigenvector is compensated by frequency domain complex multiplication. For example, if the amplitude attenuation of a certain frequency point is 0.6 and the phase delay is π / 3, the compensation factor amplitude of the frequency point is 1.67 and the phase is -π / 3. Distortion compensation also needs to consider the nonlinear characteristics of the channel. The amplitude-frequency and phase-frequency characteristics of the channel are fitted by piecewise polynomial functions, and different compensation coefficients are used in different frequency bands. For strong nonlinear regions, harmonic compensation terms are introduced to correct signal distortion. After compensation, the causality of the signal is verified by Hilbert transform to ensure that the compensation results are physically feasible.

[0050] In some embodiments, attenuation compensation can be achieved in a variety of ways: Optionally, an adaptive compensation method is used 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; determining the optimal compensation factor through iterative optimization. Optionally, a deep learning method is used to train the compensation network to automatically learn the signal attenuation characteristics. The specific steps include: constructing a feature mapping network; inputting low-frequency features for forward propagation; and reversely updating the network parameters based on the reconstruction error to obtain a compensation model. It is understandable that other compensation methods can also be used, such as frequency domain equalization, time domain interpolation, etc., which are not limited here.

[0051] S104, calculating the feature correlation between the high-frequency correction vector and the low-frequency feature vector, and extracting the transient feature information of the high-frequency correction vector when the feature correlation is lower than a preset correlation threshold.

[0052] Among them, feature correlation indicates the similarity between two feature vectors, which is expressed by correlation coefficient or distance metric; preset correlation threshold refers to the standard value for judging feature correlation; transient feature information refers to the rapidly changing characteristics during the discharge process, such as rise time, peak value, duration, etc.

[0053] The partial discharge identification system needs to evaluate the consistency of high-frequency and low-frequency features and extract supplementary features when necessary. Specifically, the partial discharge identification system first uses a variety of correlation measurement methods to calculate the correlation between the high-frequency correction vector and the low-frequency feature vector, including the Pearson correlation coefficient. When the correlation is lower than the preset threshold, it means that there is an 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 correction vector, including amplitude, duration, oscillation characteristics, etc., as supplementary features.

[0054] It should be noted that the feature correlation calculation can adopt the multidimensional correlation analysis method. First, the high-frequency correction vector and the low-frequency feature vector are standardized to eliminate the dimension effect. Then the generalized cross-correlation function of the two vectors is calculated: R (τ) = E [X (t) Y (t + τ)], where E represents the expected operation. The correlation of the features is evaluated by the maximum cross-correlation coefficient and the correlation delay. At the same time, the canonical correlation coefficient is calculated to examine the correlation strength of the features in different dimensions. The two vectors are transformed by the 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.

[0055] In some embodiments, feature correlation analysis can be implemented in a variety of ways: optionally, a statistical method is used to calculate the multidimensional 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, a distance measurement method is used 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 a correlation interval. It is understandable that other correlation analysis methods, such as information entropy, fuzzy similarity, etc., can also be used, which are not limited here.

[0056] S105 , fusing transient feature information and low-frequency feature vectors to obtain low-frequency correction vectors.

[0057] Among them, feature fusion refers to the process of integrating feature information from different sources or types; the low-frequency correction vector refers to the optimized feature vector after the transient features are integrated, which contains more complete discharge feature information.

[0058] The partial discharge identification system needs to effectively fuse the extracted transient features with the original low-frequency features. Specifically, the partial discharge identification system first normalizes the scale of the transient features and the low-frequency features to make them comparable. Then, feature fusion algorithms such as weighted summation and feature splicing are used to integrate the two types of features. The importance weights of various features are considered during the fusion process, and the physical meaning of the features is maintained. Finally, a low-frequency correction vector containing complete information is obtained.

[0059] In some embodiments, the feature fusion process is performed in an adaptive weighted manner. First, the dimensions of the transient feature information and the low-frequency feature vector are unified, and the two are mapped to the feature space of the same dimension through principal component analysis. Then, the weight coefficient of each feature component is calculated based on the information entropy criterion, and the feature component with the larger information entropy obtains a higher weight. Then, the weighted features are transformed through a nonlinear mapping function (such as a hyperbolic tangent function) to enhance the expressive power of the features. Finally, the transformed features are connected in series to form a fused feature vector. For example, for a certain discharge signal, the transient features include 10-dimensional features such as rise time 2μs and peak value 1.5mV, and the low-frequency feature vector includes 15-dimensional features such as spectrum energy and phase. A 25-dimensional low-frequency correction vector is obtained through the above fusion process.

[0060] In some embodiments, feature fusion can be achieved in a variety of ways: Optionally, a multi-level fusion strategy is used to combine features at different levels. Specific steps include: feature layer alignment; intra-layer feature fusion; inter-layer feature integration. Optionally, an attention mechanism is used to perform adaptive fusion based on feature importance. Specific steps include: calculating feature attention weights; feature weighted combination; feature vector reconstruction. It is understandable that other fusion methods can also be used, such as deep feature fusion, multimodal fusion, etc., which are not limited here.

[0061] S106, inputting the high-frequency correction vector and the low-frequency correction vector into the partial discharge identification model to obtain the partial discharge type and identification confidence.

[0062] Among them, the partial discharge recognition model represents a deep learning model used to identify discharge types; discharge types include corona discharge, surface discharge, suspended discharge and other forms; recognition confidence represents the degree of confidence of the model in the recognition results.

[0063] The partial discharge identification system needs to identify the discharge type based on the corrected feature vector. Specifically, the partial discharge identification system first integrates the high-frequency and low-frequency correction vectors into an input feature matrix according to a preset format. The feature matrix is ​​then input into a pre-trained deep learning model, which contains multi-layer convolution, loop and other network structures. The probability distribution of different discharge types is obtained through forward calculation, the type with the highest probability is selected as the identification result, and the corresponding confidence value is output. The confidence reflects the reliability of the identification result and can be used for subsequent decision-making.

[0064] It should be noted that the partial discharge identification 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 convolution layer uses convolution kernels of different scales (such as 3×3, 5×5) to perform convolution operations on the input features and extract multi-scale features. The pooling layer compresses the feature dimensions and retains significant features through the maximum pooling operation. 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 back-propagation algorithm, and the loss function uses the cross-entropy loss. The dynamic learning rate strategy is used during model training. A larger learning rate is used for rapid convergence in the early stage of training, and the learning rate is reduced in the later stage to fine-tune the parameters.

[0065] In some embodiments, discharge type recognition can be achieved in a variety of ways: Optionally, an integrated learning method is used to combine the prediction results of multiple basic models. Specific steps include: constructing multiple heterogeneous classifiers; independently performing feature learning and classification; fusing the results by voting or weighting. Optionally, a multi-task learning framework is used to simultaneously predict the discharge type and other related attributes. Specific steps include: designing a shared feature extraction layer; constructing multiple task branches; and jointly optimizing the loss functions of different tasks. It is understandable that other classification methods, such as semi-supervised learning, transfer learning, etc., can also be used, which are not limited here.

[0066] In the above embodiment, the basic feature extraction and recognition process is mainly introduced. In practical applications, the network structure and optimization strategy can also be adjusted according to specific scenarios to further improve system performance. The following is a supplement to the scenario of this embodiment.

[0067] To further improve system performance, the substation optimized and upgraded the partial discharge identification system. The system added a Transformer-based feature extraction module to better capture the temporal dependencies of the signal. At the same time, a contrastive learning mechanism was introduced to improve the discriminability of feature representation. In actual operation, the system can not only accurately identify a single type of discharge fault, but also diagnose complex discharge faults. In addition, the system has the ability to transfer knowledge and can quickly adapt to the monitoring needs of new equipment, reducing the amount of labeled data required for model training. Through distributed deployment and collaborative learning, data sharing and model optimization between multiple substations are achieved.

[0068] After combining the above scenarios, the following is a more detailed description of the process of the method provided by this implementation. Figure 2 , is another flow chart of the local discharge signal recognition method based on deep learning in an embodiment of the present application.

[0069] S201 , 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 distinction frequency.

[0070] Referring to step S101 , the partial discharge identification system groups signal frequencies.

[0071] S202, extracting feature information of the high-frequency signal group and the low-frequency signal group respectively, and inputting 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.

[0072] Referring to step S102 , the partial discharge identification system constructs a high-frequency feature vector and a low-frequency feature vector.

[0073] In some embodiments, the partial discharge identification system performs time-frequency transformation and convolution operations when extracting feature vectors, that is, the partial discharge identification system performs time-frequency transformation on the high-frequency signal group and the low-frequency signal group to obtain a time-frequency feature map; performs layered convolution operations on the time-frequency feature map based on a convolutional neural network to obtain a feature tensor; performs self-supervised learning encoding on the feature tensor to generate a feature representation matrix; and inputs the feature representation matrix into the corresponding first deep neural network and second deep neural network to obtain high-frequency feature vectors and low-frequency feature vectors, respectively.

[0074] Among them, time-frequency transform refers to the mathematical transformation method of converting time domain signals into time-frequency domain, such as short-time Fourier transform or wavelet transform; time-frequency feature map refers to the two-dimensional representation of the signal on the time-frequency plane; layered convolution operation refers to the process of multi-level feature extraction of feature maps using different convolution kernels; feature tensor refers to the multi-dimensional feature data structure obtained after convolution operation; self-supervised learning coding refers to the coding method that can learn feature representation without label data; feature representation matrix refers to the two-dimensional feature data set obtained after encoding.

[0075] After obtaining the high- and low-frequency signal groups, the partial discharge identification system needs to perform feature extraction and representation learning. Specifically, the partial discharge identification system first performs short-time Fourier transform or continuous wavelet transform on the two groups of signals to generate a two-dimensional feature map that reflects the time-frequency characteristics of the signal. Then, a multi-layer convolutional neural network is used to extract local feature patterns through convolution kernels of different scales to form a multi-channel feature tensor. Then, self-supervisory methods such as contrast learning or reconstruction learning are used to learn to obtain more discriminative feature representations. Finally, these features are input into a specially designed deep neural network, and fixed-dimensional feature vectors are obtained through nonlinear transformation. The entire process realizes the automatic extraction and representation of high-dimensional features from raw signals.

[0076] It should be noted that the deep neural network here includes the first and second deep neural networks. In the training stage, the two networks receive the feature representation matrix as input respectively, and use the labeled discharge data for supervised training. The training goal is to minimize the difference between the predicted output and the true label, and the cross entropy or mean square error is used as the loss function. The network structure contains multiple fully connected layers, each followed by a nonlinear activation function and a normalization layer. In practical applications, the two networks process the feature representation of different frequency bands respectively, and map the high-dimensional features into feature vectors of fixed dimensions through nonlinear transformations.

[0077] Contrastive learning is based on the contrastive learning model. During the training phase, the model receives both the temporal feature sequence and the attention feature sequence, and is trained by constructing positive and negative sample pairs. The training goal 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 includes 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 results and confidence evaluation of the discharge type based on the similarity distribution.

[0078] In some embodiments, feature extraction and representation learning can be achieved in a variety of ways: Optionally, a multi-scale time-frequency analysis method is used. The specific steps include: designing time-frequency atoms of multiple scales; calculating the correlation coefficient between the signal and the atom; constructing a time-frequency energy distribution; extracting time-frequency features. Optionally, an adaptive feature learning method is used. The specific steps include: constructing a contrastive learning loss function; designing a data enhancement strategy; training a feature extractor; and verifying the quality of feature representation. It is understandable that other feature learning methods, such as variational autoencoders, graph neural networks, etc., can also be used, which are not limited here.

[0079] 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 correction vector.

[0080] Referring to step S103 , the partial discharge identification system performs distortion compensation on the high-frequency feature vector.

[0081] S204, calculating the feature correlation between the high-frequency correction vector and the low-frequency feature vector, and extracting transient feature information of the high-frequency correction vector when the feature correlation is lower than a preset correlation threshold.

[0082] Referring to step S104 , the partial discharge identification system extracts transient characteristic information of the high-frequency correction vector.

[0083] S205, extracting the discharge pulse period from the high-frequency correction vector to obtain a periodic characteristic sequence.

[0084] Among them, the discharge pulse period represents the time interval between adjacent discharge pulses in the local discharge signal; the periodic characteristic sequence refers to the time series data describing the regularity of the repeated occurrence of discharge pulses; and the high-frequency correction vector represents the set of high-frequency characteristic data after attenuation compensation.

[0085] After obtaining the high-frequency correction vector, the partial discharge identification system needs to analyze the time characteristics of the discharge process. Specifically, the partial discharge identification system first performs time domain analysis on the high-frequency correction vector and locates the time point of the discharge pulse through threshold detection. Then the time difference between adjacent pulses is calculated to form a periodic sequence. The sequence is statistically analyzed, including the mean, variance, distribution characteristics, etc., to construct a periodic feature sequence. This sequence can reflect the dynamic change law of the discharge process.

[0086] In some embodiments, period extraction can be achieved in a variety of ways: Optionally, an adaptive threshold method is used to determine the pulse position and calculate the period. Specific steps include: calculating the signal envelope; setting a dynamic detection threshold; locating the pulse peak point; and counting the intervals between adjacent peaks. Optionally, a spectral analysis method is used to extract periodic features. Specific steps include: calculating the signal autocorrelation function; performing periodogram analysis; and extracting the main periodic components. It is understandable that other periodic analysis methods, such as wavelet transform, Hilbert transform, etc., can also be used, which are not limited here.

[0087] S206, performing wavelet decomposition on the periodic characteristic sequence to obtain multi-scale components.

[0088] Among them, wavelet decomposition refers to the process of multi-resolution analysis of signals using wavelet basis functions; multi-scale components refer to signal components at different frequency scales; wavelet basis functions refer to a family of finite support functions that satisfy specific mathematical properties.

[0089] The partial discharge identification system needs to perform multi-scale analysis on the periodic feature sequence. Specifically, the partial discharge identification system first selects a suitable wavelet basis function (such as Daubechies wavelet) and determines the number of decomposition layers. Then, the periodic feature sequence is recursively decomposed to obtain approximate coefficients and detail coefficients of different scales. These coefficients reflect the energy distribution and local characteristics of the signal in different frequency bands, forming a set of multi-scale components.

[0090] In some embodiments, wavelet decomposition can be achieved in a variety of ways: Optionally, orthogonal wavelet transform is used to achieve lossless decomposition of the signal. Specific steps include: selecting an orthogonal wavelet basis; constructing a filter bank; performing multi-layer decomposition; and saving coefficients of each layer. Optionally, wavelet packet decomposition is used to obtain a more detailed frequency band division. Specific steps include: constructing a wavelet packet tree; calculating node energy; selecting an optimal basis; and extracting characteristic coefficients. It is understandable that other decomposition methods, such as dual-tree complex wavelets, adaptive wavelets, etc., can also be used, which are not limited here.

[0091] S207 , performing noise reduction on the multi-scale components based on adaptive filtering to obtain a reconstructed signal.

[0092] Among them, adaptive filtering refers to a signal processing method that can automatically adjust the filtering parameters according to the signal characteristics; noise reduction refers to the process of suppressing or removing unnecessary noise components in the signal; and reconstructed signal refers to the signal resynthesized after noise reduction processing.

[0093] After obtaining the multi-scale components, the partial discharge identification system needs to perform signal optimization processing. Specifically, the partial discharge identification 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 perform noise reduction processing on each component. Finally, the processed scale components are weighted and reconstructed to obtain the optimized signal. This adaptive processing method can better maintain the effective characteristics of the signal.

[0094] It should be noted that the adaptive filtering process here achieves signal optimization based on the minimum mean square error criterion. First, the signal-to-noise ratio of each scale component is estimated, and an adaptive Wiener filter is constructed. 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 noise-dominated frequency band. The filter coefficients are updated in real time through a recursive least squares algorithm to track changes in signal characteristics. The scale components after filtering are synthesized into a complete signal through a wavelet reconstruction algorithm. For example, for a certain scale component, if the signal-to-noise ratio of the local interval is 15dB, the filter gain of the corresponding frequency band is about 0.9, while in the interval where the signal-to-noise ratio is only 5dB, the gain is reduced to about 0.3.

[0095] In some embodiments, adaptive noise reduction can be achieved in a variety of ways: Optionally, empirical mode decomposition combined with adaptive threshold method is used. 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, the Kalman filter method is used to achieve dynamic tracking filtering of the signal. Specific steps include: establishing a state space model; predicting state estimation; updating the gain matrix; fusing measurement values. It is understandable that other noise reduction methods, such as Wiener filtering, particle filtering, etc., can also be used, which are not limited here.

[0096] S208. Extract characteristic information from the reconstructed signal to obtain transient characteristic information.

[0097] Among them, the reconstructed signal refers to the time domain signal re-synthesized after noise reduction processing; the characteristic information refers to the key parameters in the signal that reflect the discharge characteristics; the transient characteristic information refers to the set of features that change rapidly during the discharge process, including amplitude features, time features and frequency features; feature extraction refers to the process of obtaining effective features from the signal; the characteristic parameters refer to the quantitative indicators used to describe the signal characteristics.

[0098] After obtaining the reconstructed signal, the partial discharge identification system needs to extract the transient features that reflect the discharge characteristics. Specifically, the partial discharge identification 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 analyze the time characteristics of the signal, including rise time, duration, oscillation period, etc. Then perform frequency domain analysis to extract features such as main frequency components, bandwidth, energy distribution, etc. Finally, these features are standardized to form a unified transient feature vector. These feature information comprehensively describes the transient characteristics of the discharge process and provides a basis for subsequent fault diagnosis.

[0099] In some embodiments, transient feature extraction can be achieved in a variety of ways: Optionally, a parametric feature extraction method is used. 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 used. The specific steps include: calculating a signal envelope; extracting morphological features; analyzing local waveforms; and generating feature descriptors. It is understandable that other feature extraction methods, such as wavelet energy features, fractal features, etc., can also be used, which are not limited here.

[0100] In some embodiments, the partial discharge identification system will perform transient identification based on energy distribution, that is, the partial discharge identification system will calculate the time-frequency spectrum of the reconstructed signal to obtain an energy distribution sequence; perform 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 based on the transient feature intervals.

[0101] Among them, the time-frequency spectrum diagram represents the distribution diagram of signal energy on the time-frequency plane; the energy distribution sequence refers to the sequence of signal energy changes over time; the transient feature point represents the moment when the signal undergoes a sudden change; the transient feature interval refers to the time interval containing the transient process; the transient feature information includes characteristic parameters such as amplitude and duration.

[0102] After obtaining the reconstructed signal, the partial discharge identification system needs to extract transient features. Specifically, the partial discharge identification system first uses the sliding window short-time Fourier transform to calculate the time-frequency spectrum of the signal and analyze the distribution characteristics of energy in the time-frequency domain. Then, an adaptive threshold algorithm is applied to the energy distribution sequence to detect energy mutation points and determine the possible time when transient events occur. 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 transient feature description.

[0103] It should be noted that the feature extraction process integrates time domain and frequency domain analysis. In the time domain, the mutation point of the signal is detected by adaptive threshold, and the threshold is set to 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 optimize the sparse representation of the signal. Then, features such as 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 by principal component analysis, and the principal components with a cumulative contribution rate of more than 95% are retained.

[0104] In some embodiments, transient feature extraction can be achieved in a variety of ways: Optionally, a multi-resolution analysis method is used. The specific steps include: performing wavelet packet decomposition; selecting the optimal basis function; locating feature points; extracting local features. Optionally, a sparse representation method is used. The specific steps include: constructing an overcomplete dictionary; solving sparse coefficients; reconstructing transient components; extracting feature parameters. It is understandable that other feature extraction methods, such as empirical mode decomposition, variational mode decomposition, etc., can also be used, which are not limited here.

[0105] S209: Fusing the transient feature information and the low-frequency feature vector to obtain a low-frequency correction vector.

[0106] Referring to step S105 , the partial discharge identification system will correct the low-frequency feature vector.

[0107] S210, inputting a high-frequency correction vector and a low-frequency correction vector into a partial discharge identification model to obtain a partial discharge type and an identification confidence.

[0108] Referring to step S106 , the partial discharge identification system determines the partial discharge type and identification confidence based on the model.

[0109] In some embodiments, the partial discharge identification system performs data processing according to multiple models, that is, the partial discharge identification system inputs the high-frequency correction vector into the long short-term memory network to obtain a temporal feature sequence; inputs the low-frequency correction vector into the Transformer encoder to obtain an attention feature sequence; calculates the similarity matrix of the temporal feature sequence and the attention feature sequence through comparative learning, and generates the partial discharge type and identification confidence based on the similarity matrix.

[0110] 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 that reflects 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 refers to the feature sequence weighted by attention; the similarity matrix refers to the correlation measurement result between different feature sequences.

[0111] After obtaining the correction vector, the partial discharge identification system needs to perform time series feature analysis and pattern recognition. Specifically, the partial discharge identification 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, the low-frequency correction vector is input into the Transformer encoder, and the multi-head self-attention mechanism is used to learn the correlation between features to obtain a feature sequence with global context. Then, the similarity matrix of the two feature sequences is calculated through the comparative learning framework to analyze their matching degree. Finally, the classification decision rule is designed based on the similarity matrix to output the most likely discharge type and its confidence.

[0112] It should be noted that the structure of the partial discharge identification model adopts a hybrid deep learning network. After the input layer, there are parallel CNN and LSTM branches: the CNN branch uses residual connection to extract spatial features, and the LSTM branch captures temporal dependencies. Then the features of the two branches are fused through the attention mechanism, and the attention weight is 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 deal with the problem of sample imbalance and give higher weights to difficult-to-classify samples.

[0113] During the training phase, the Long Short-Term Memory (LSTM) model takes a high-frequency correction vector sequence as input and is trained using time series labeled data. The training goal is to minimize the sequence prediction error, using mean square error or sequence correlation as the loss function. The model contains three gating units: input gate, forget gate, and output gate, as well as a memory unit. These components work together to learn long-term dependencies in sequence data. In practical applications, the model receives high-frequency correction vectors, processes sequence information through gating mechanisms and memory units, and outputs a feature sequence containing time series dynamic features.

[0114] In addition, during the training phase, the Transformer encoder model uses low-frequency correction vectors as input data and is trained using self-supervised learning. During the training process, the masked language model task and multi-head self-attention mechanism are used, and the optimization goal is to maximize the mutual information of feature representation. The model architecture contains multiple layers of self-attention layers and feedforward neural network layers, each with residual connections and layer normalization. When in use, the model captures the global dependencies in the input sequence by calculating the attention weight matrix and generates a context-aware attention feature sequence.

[0115] In some embodiments, sequence feature analysis can be achieved in a variety of ways: Optionally, a bidirectional attention fusion method is used. The specific steps include: building a bidirectional LSTM network; calculating temporal attention weights; fusing bidirectional features; generating enhanced feature sequences. Optionally, a hierarchical contrastive learning method is used. The specific steps include: designing a multi-level feature extractor; building a multi-granularity contrast loss; optimizing feature representation; learning discriminant boundaries. It is understandable that other sequence analysis methods, such as gated recurrent units, temporal convolutional networks, etc., can also be used, which are not limited here.

[0116] S211, performing feature fusion on the high-frequency correction vector and the low-frequency correction vector to obtain fused features.

[0117] Among them, feature fusion refers to the process of comprehensively integrating feature information from different sources; fused features refer to the new feature representation obtained through feature fusion; feature weight refers to the importance of different features in the fusion process.

[0118] The partial discharge identification system needs to effectively integrate high-frequency and low-frequency features. Specifically, the partial discharge identification system first aligns the dimensions of the two feature vectors and normalizes them. Then, the feature importance weights are calculated, which can be based on information gain or feature correlation analysis. Finally, weighted fusion or nonlinear mapping methods are used to combine the two feature vectors into a unified feature representation. This fusion strategy can make full use of the complementary information of different frequency bands.

[0119] In some embodiments, feature fusion can be achieved in a variety of ways: Optionally, a feature fusion method using an attention mechanism is used. The specific steps include: calculating feature attention weights; generating an attention map; weighted fusion features; optimizing attention parameters. Optionally, a feature fusion method using an autoencoder is used. The specific steps include: constructing an encoder network; learning feature mapping relationships; decoding and reconstructing features; extracting bottleneck layer features. It is understandable that other fusion methods can also be used, such as feature fusion, hierarchical fusion, etc., which are not limited here.

[0120] S212, calculating the correlation coefficient between the fusion feature and the preset partial discharge template to obtain a matching degree matrix.

[0121] Among them, the preset partial discharge template represents a set of standard feature patterns of various known discharge types; the correlation coefficient refers to a statistic that measures the similarity between two feature vectors; and the matching degree matrix refers to the correlation coefficient matrix between the fused features and each template.

[0122] The partial discharge identification system needs to evaluate the matching degree between the fused features and the standard templates. Specifically, the partial discharge identification system first loads the preset discharge template library, which contains the typical characteristic patterns of different types of discharges. Then, a variety of correlation measurement methods, such as Pearson correlation coefficient and cosine similarity, are used to calculate the correlation coefficient between the fused features and each template. These coefficients are organized into a matrix form to reflect the matching relationship between the features and each type of template.

[0123] In some embodiments, correlation calculation can be implemented in a variety of ways: Optionally, a multi-scale correlation analysis method is used. The specific steps include: constructing a feature pyramid; calculating a multi-scale correlation coefficient; weighted fusion correlation; generating a matching degree sequence. Optionally, a kernel correlation analysis method is used. The specific steps include: selecting a kernel function; calculating a kernel matrix; solving a feature map; calculating a kernel correlation coefficient. It is understandable that other correlation analysis methods, such as mutual information, canonical correlation analysis, etc., can also be used, which are not limited here.

[0124] S213, selecting the feature pattern with the greatest matching degree from the matching degree matrix to obtain a recognition result.

[0125] Among them, the characteristic pattern represents the typical characteristic combination of a certain type of discharge; the maximum matching degree refers to the maximum correlation in the correlation coefficient matrix; and the recognition result includes the discharge type and the corresponding confidence information.

[0126] 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, the maximum value detection algorithm is used to find the characteristic pattern with the highest matching degree and its corresponding discharge type. At the same time, the difference between the maximum matching degree and the suboptimal matching degree is calculated as the reliability indicator of the recognition result. If multiple patterns have similar matching degrees, the system will further analyze the distinguishability of the features to improve the accuracy of recognition.

[0127] In some embodiments, the optimal mode selection can be achieved in a variety of 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; selecting the optimal solution. Optionally, a probabilistic reasoning method is used to determine the optimal mode. The specific steps include: constructing a probabilistic graphical model; calculating the posterior probability; maximum likelihood estimation; outputting the recognition result. It is understandable that other decision-making methods, such as fuzzy reasoning, evidence theory, etc., can also be used, which are not limited here.

[0128] S214. Update the neural network parameters of the partial discharge identification model according to the identification result.

[0129] Among them, neural network parameters refer to learnable parameters such as weights and biases in deep learning models; parameter update refers to the process of adjusting model parameters according to new samples; and learning rate refers to the step coefficient of parameter update.

[0130] The partial discharge identification system needs to continuously optimize the model performance through online learning. Specifically, the partial discharge identification system first compares the identification result with the actual label and calculates the identification error. Then, based on the error gradient, the back propagation algorithm is used to update the parameters of each layer of the neural network. The historical gradient information is considered in the parameter update process, the learning rate is dynamically adjusted, and regularization technology is used to prevent overfitting. This incremental learning mechanism enables the model to adapt to the dynamic changes of discharge characteristics.

[0131] It should be noted that the parameter update of the partial discharge identification model adopts a dynamic adaptive optimization strategy. First, the sample weight is calculated based on the confidence distribution of the model output. The sample with lower confidence will get a higher weight when the parameter is updated. Then the gradient of the weighted loss function with respect to the network parameters is calculated, and the gradient calculation adopts the reverse automatic differentiation technology. The optimizer uses the Adam algorithm, combined with the momentum term and adaptive learning rate, to accelerate convergence and avoid local optimality. The L2 regularization term is introduced during parameter update to prevent overfitting, and the regularization coefficient is dynamically adjusted with the training round. For example, when the recognition confidence of a 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-distinguish samples.

[0132] In some embodiments, parameter updates can be achieved in a variety of ways: Optionally, an adaptive learning rate optimization algorithm is used. The specific steps include: calculating parameter gradients; estimating the first-order moment and second-order moment of the gradient; adaptively adjusting the learning rate; updating model parameters. Optionally, a transfer learning method is used to update the model. The specific steps include: freezing the base layer parameters; fine-tuning the top-level network; incrementally training new samples; and verifying model performance. It is understandable that other parameter optimization methods, such as meta-learning, lifelong learning, etc., can also be used, which are not limited here.

[0133] In some embodiments, the partial discharge identification system will perform model optimization based on feedback, that is, the partial discharge identification system will calculate the confidence distribution of the identification result to obtain the feedback weight coefficient; perform backpropagation on the neural network based on the feedback weight coefficient to obtain the gradient value; adjust the convolution layer parameters according to the gradient value to obtain the optimized parameters, and write the optimized parameters into the partial discharge identification model.

[0134] Among them, the confidence distribution represents the model's predicted probability distribution 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; and the optimization parameter refers to the updated neural network weight parameter.

[0135] After obtaining the identification results, the partial discharge identification system needs to optimize the model parameters. Specifically, the partial discharge identification system first analyzes the probability distribution of the model output, calculates the uncertainty of the prediction results, and generates weight coefficients that reflect the reliability of the prediction. These weights are then combined with the prediction error, and the gradients of the parameters of each layer are calculated through the back propagation 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 layers. Finally, the optimized parameters are saved in the model to complete the online update of the model.

[0136] In some embodiments, parameter optimization can be achieved in a variety of ways: Optionally, a dynamic learning rate adjustment method is used. 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 used. The specific steps include: calculating the layer sensitivity; designing the hierarchical learning rate; updating the parameters layer by layer; verifying the optimization effect. It is understandable that other optimization methods, such as second-order optimization, meta-learning, etc., can also be used, which are not limited here.

[0137] In the embodiment of the present application, due to the use of multi-source signal collaborative processing and deep feature learning methods, it is possible to simultaneously utilize the transient characteristics of high-frequency signals and the continuous characteristics of low-frequency signals, effectively solving the problems of low accuracy and poor reliability of traditional single-band detection methods in complex environments, thereby achieving accurate identification of partial discharges and early warning of faults. Through the feature fusion mechanism and adaptive compensation algorithm, the system can effectively deal with signal attenuation and environmental interference and improve detection sensitivity. The online optimization method based on deep learning enables the system to have continuous learning capabilities and can adapt to dynamic changes in the operating status of the equipment. The multi-level feature extraction and recognition strategy not only improves the diagnostic accuracy of the system, but also realizes the identification of complex discharge faults, providing reliable technical support for status monitoring and preventive maintenance of power equipment.

[0138] The following describes the partial discharge identification system in the embodiment of the present invention from the perspective of hardware processing. Figure 3, which is a schematic diagram of a physical device structure of a local discharge identification system in an embodiment of the present application.

[0139] It should be noted that Figure 3 The structure of the partial discharge identification system shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0140] like Figure 3 As shown, the partial discharge identification 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 from the storage part 308 to the RAM 303, such as executing the method described in the above embodiment. Various programs and data required for system operation are also stored in the RAM 303. The CPU 301, the ROM 302 and the RAM 303 are connected to each other via a bus 304. The I / O interface 305 is also connected to the bus 304.

[0141] 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) and 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. A 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 therefrom is installed into the storage section 308 as needed.

[0142] In particular, 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 part 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.

[0143] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. Each box in the flowchart or block diagram may represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box may also occur in an order different from that marked in the accompanying drawings.

[0144] Specifically, the partial discharge identification system of this embodiment includes a processor and a memory, and a computer program is stored in the memory. When the computer program is executed by the processor, the partial discharge signal identification method based on deep learning provided in the above embodiment is implemented.

[0145] As another aspect, the present invention further provides a computer-readable storage medium, which may be included in the partial discharge identification system described in the above embodiment; or may exist independently without being assembled into the partial discharge identification system. The above storage medium carries one or more computer programs, and when the above one or more computer programs are executed by a processor of the partial discharge identification system, the partial discharge identification system implements the partial discharge signal identification method based on deep learning provided in the above embodiment.

[0146] As described 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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, 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 embodiments of the present application.

[0147] As used in the above embodiments, the term "when..." may be interpreted to mean "if..." or "after..." or "in response to determining..." or "in response to detecting...", depending on the context. Similarly, the phrases "upon determining..." or "if (the stated condition or event) is detected" may be interpreted to mean "if determining..." or "in response to determining..." or "upon detecting (the stated condition or event)" or "in response to detecting (the stated condition or event)", depending on the context.

Claims

1. A method for identifying partial discharge signals based on deep learning, characterized in that: Applied to a partial discharge identification system, the method comprises: 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 distinction frequency; Extracting feature information of the high-frequency signal group and the low-frequency signal group respectively, and inputting 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; 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 correction vector; Calculating the feature correlation between the high-frequency correction vector and the low-frequency feature vector, and extracting transient feature information of the high-frequency correction vector when the feature correlation is lower than a preset correlation threshold; Fusion of the transient feature information and the low-frequency feature vector to obtain a low-frequency correction vector; The high-frequency correction vector and the low-frequency correction vector are input into a partial discharge identification model to obtain a partial discharge type and an identification confidence.

2. The method according to claim 1, characterized in that The steps of respectively extracting feature information of the high-frequency signal group and the low-frequency signal group, and respectively inputting the corresponding feature information into the first deep neural network and the 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 a time-frequency feature graph; Performing a layered convolution operation 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; The feature representation matrix is ​​input into the corresponding first deep neural network and the second deep neural network to obtain the high-frequency feature vector and the low-frequency feature vector respectively.

3. The method according to claim 1, characterized in that The step of inputting the high-frequency correction vector and the low-frequency correction vector into the partial discharge identification model to obtain the partial discharge type and identification 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 the Transformer encoder to obtain an attention feature sequence; The similarity matrix of the temporal feature sequence and the attention feature sequence is calculated by contrastive learning, and the local discharge type and recognition confidence are generated 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 correction vector, the method further includes: Extracting the discharge pulse period from the high-frequency correction vector to obtain a periodic characteristic sequence; Performing wavelet decomposition on the periodic characteristic sequence to obtain multi-scale components; Performing noise reduction on the multi-scale components based on adaptive filtering to obtain a reconstructed signal; Extract characteristic information from the reconstructed signal to obtain transient characteristic information.

5. The method according to claim 4, characterized in that After the step of extracting the characteristic information from the reconstructed signal to obtain transient characteristic information, the method further comprises: Calculating a time-frequency spectrum of the reconstructed signal to obtain an energy distribution sequence; Performing peak detection on the energy distribution sequence to determine transient characteristic points; Signal segmentation is performed based on the transient feature points to obtain transient feature intervals, and transient feature information is generated according to the transient feature intervals.

6. The method according to claim 1, characterized in that After the step of inputting the high-frequency correction vector and the low-frequency correction vector into the partial discharge identification model to obtain the partial discharge type and identification confidence, the method further includes: Performing feature fusion on the high-frequency correction vector and the low-frequency correction vector to obtain a fusion feature; Calculate the correlation coefficient between the fusion feature and the preset partial discharge template to obtain a matching matrix; Selecting the feature pattern with the greatest matching degree from the matching degree matrix to obtain a recognition result; The neural network parameters of the partial discharge identification model are updated according to the identification result.

7. The method according to claim 6, characterized in that The step of updating the neural network parameters of the partial discharge identification model according to the identification result specifically includes: Calculating the confidence distribution of the recognition result to obtain a feedback weight coefficient; Perform back propagation on the neural network based on the feedback weight coefficient to obtain a gradient value; The convolution layer parameters are adjusted according to the gradient value to obtain optimized parameters, and the optimized parameters are written into the partial discharge identification model.

8. A partial discharge identification system, characterized in that: The partial discharge identification 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 identification system to execute the method described in any one of claims 1 to 7.

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

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

Images