GIS partial discharge type identification method and system based on deep learning network

Abstract

The invention relates to the field of power equipment state monitoring and fault diagnosis, and provides a GIS partial discharge identification method and system based on a deep learning network. According to the method, time synchronization and variational mode decomposition are carried out on multichannel partial discharge signals, energy entropy is calculated in a sliding time window, a candidate feature sequence is constructed, correlation screening based on category cost and learnable gating with channel health degree, normalized SNR and phase coverage degree as input are introduced to obtain a feature subset, and the candidate feature sequence is constructed. A one-dimensional convolutional neural network is combined with cost-sensitive focus loss and temperature scaling to carry out training and confidence coefficient calibration. In online reasoning, a node-level recognition result and a rechecking task are generated by utilizing multi-channel logit weighted fusion, consistency judgment, bounded multiplicative up-regulation and single-channel confidence and stability driven threshold self-adaption, and an alarm closed loop is formed based on a rechecking time limit. According to the method, the partial discharge identification precision and the confidence reliability are improved, and the misinformation and missing report risks are reduced.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis, specifically a GIS partial discharge type identification method and system based on deep learning networks. Background Technology

[0002] Gas-insulated switchgear (GIS) is widely used in high-voltage and ultra-high-voltage power grids. Internal insulation defects in GIS can trigger partial discharges. The type of partial discharge is closely related to the location and morphology of the defect, and the identification results of the partial discharge type are of great significance for defect location and maintenance decisions.

[0003] Existing methods for detecting and identifying partial discharge in GIS typically fall into two categories: Partial discharge signals are processed using time-domain, frequency-domain, or time-frequency-domain methods to extract features such as peak value, energy, pulse count, phase distribution, frequency band energy ratio, and entropy indices. These features are then classified using support vector machines, decision trees, or ensemble learning models. However, these methods rely on manually designed features, making it difficult to balance robustness and recognition accuracy under complex operating conditions.

[0004] By utilizing convolutional neural networks, recurrent neural networks, and converter networks to perform end-to-end modeling of partial discharge spectra or waveforms, the feature representation capability can be improved. Existing methods often only focus on single-channel signals, or in multi-channel scenarios, they only use simple averaging or majority voting methods for fusion. Furthermore, they lack systematic modeling of identification confidence, which can easily lead to overconfident and erroneous conclusions, making it difficult to provide maintenance personnel with interpretable alarm levels and verification basis.

[0005] In actual GIS engineering sites, multi-channel partial discharge sensors are complexly deployed, with differences in noise levels, hardware aging, and phase coverage between channels. Traditional methods often fail to adequately consider channel health and signal-to-noise ratio differences, and do not utilize confidence stability to adaptively adjust the inference threshold, resulting in poor multi-channel fusion performance and high false alarm and false negative rates. Summary of the Invention

[0006] The purpose of this invention is to provide a solution to one of the aforementioned problems existing in the prior art. Specifically, this invention is achieved through the following technical solution: A method for identifying partial discharge types in GIS based on deep learning networks includes the following steps: Step 1: Time synchronization and power frequency phase alignment are performed on the partial discharge signals acquired by multiple discharge signal acquisition channels. Variational mode decomposition is performed under given decomposition parameters to obtain the modal components of each channel. Step 2: Calculate the energy entropy of the modal components within the sliding time window, and splice them together according to time and power frequency phase to form a candidate feature sequence; Step 3: Based on the category cost, perform relevance screening. Input the channel health, normalized SNR and phase coverage into the learnable gating network to gating the candidate feature sequences and obtain the feature subset. Step 4: Construct a one-dimensional convolutional neural network, take a subset of features as input, train the recognition model using a cost-sensitive focus loss function, perform temperature scaling calibration on the recognition output of the validation set, obtain the post-calibration confidence, and determine the inference threshold. Step 5: Perform steps 1 to 3 to obtain a feature subset from the signal to be identified. Input this subset into the identification model to obtain the partial discharge type label, channel confidence, and channel category logit for each channel. In the multi-channel case, determine the channel weights based on channel health, normalized SNR, and channel type. Perform weighted fusion of the channel category logit to obtain the fused category logit. Calculate the fusion confidence based on the fused category logit, and determine consistency based on the maximum and second-largest values ​​of the fusion confidence and a consistency threshold. If consistency is achieved, multiplicatively increase the fusion confidence using a preset bounded multiplicative coefficient. Generate node-level identification results or verification tasks based on the relationship between the increased fusion confidence and the inference threshold. If consistency is not achieved, generate verification tasks. In the single-channel case, calculate the confidence stability from the channel confidence within multiple sliding time windows. Combine the channel health and normalized SNR to adjust the inference threshold within a preset adaptive range. Generate node-level identification results or verification tasks based on the relationship between the adjusted channel confidence and the inference threshold. Step 6: Bind the partial discharge type label and confidence level of the node-level identification results to time, power frequency phase, and channel-to-physical location mapping to form node-level positioning information, handling suggestions, and early warning records; set a review time limit for the review task, and output a low confidence alarm if the review task is not closed within the review time limit.

[0007] Furthermore, the process of time synchronization and power frequency phase alignment of partial discharge signals acquired from multiple discharge signal acquisition channels, and performing variational mode decomposition under given decomposition parameters to obtain the modal components of each channel, includes: Cross-correlation peak time delay estimation is performed on partial discharge signals acquired from multiple discharge signal acquisition channels. Based on the time delay estimation results, the partial discharge signals of each channel are resampled and made consistent. The partial discharge signals of each channel after resampling are divided into multiple sliding time windows according to Hamming windows, and mirror extension is performed on each sliding time window. Median filtering is performed on the samples in each sliding time window to remove spikes and mean filtering to remove DC. Abnormal sliding time windows are removed according to amplitude thresholds or energy thresholds. Under the given number of modes, bandwidth penalty factor and initial value of center frequency, variational mode decomposition is performed on each sliding time window of each channel to obtain modal components.

[0008] Furthermore, the calculation of energy entropy for modal components within a sliding time window, and the splicing of candidate feature sequences according to time and power frequency phase, includes: Within each sliding time window, the energy entropy of each modal component obtained in step one is calculated to form an energy entropy sequence for each channel. The energy entropies are arranged according to the time order and power frequency phase order corresponding to the sliding time window. The energy entropies corresponding to different power frequency phases within the same power frequency cycle are spliced ​​together to form a candidate feature sequence jointly encoded by time and power frequency phase.

[0009] Furthermore, the relevance screening based on category cost involves inputting channel health, normalized SNR, and phase coverage into a learnable gating network to gate candidate feature sequences, obtaining a feature subset, including: The category cost is determined based on the misclassification cost and sample size for each partial discharge type. The correlation index, which is weighted by the category cost, is used to calculate the correlation between each feature in the candidate feature sequence and the partial discharge type label. Features with correlation below the correlation threshold are removed to obtain the candidate feature sequence after correlation screening. A learnable gating network is constructed with channel health, normalized SNR, and phase coverage as inputs. The output is the gating coefficient corresponding to the dimension of the candidate feature sequence after correlation screening. The gating coefficient is multiplied dimension by dimension of the candidate feature sequence after correlation screening and compared with the gating threshold. The feature components below the gating threshold are truncated to obtain the feature subset.

[0010] Furthermore, the construction of a one-dimensional convolutional neural network, using a subset of features as input, training the recognition model with a cost-sensitive focus loss function, performing temperature scaling calibration on the validation set recognition output, obtaining the calibrated confidence level, and determining the inference threshold, includes: A one-dimensional convolutional neural network is used as the recognition model structure. Forward and backward propagation are performed on the feature subset. The cost-sensitive focus loss function is used for parameter updates. The class cost in the cost-sensitive focus loss function is set according to the class cost in step three. After training, the logit of each class output by the recognition model and the true label are recorded on the validation set. Temperature scaling calibration is performed by optimizing the temperature parameters to ensure that the post-calibration confidence meets the preset confidence reliability index. The inference threshold is selected based on the post-calibration confidence distribution and the partial discharge type label distribution.

[0011] Furthermore, in the multi-channel scenario, channel weights are determined based on channel health, normalized SNR, and channel type. The channel category logits are weighted and fused to obtain a fusion category logit. Fusion confidence is calculated based on the fusion category logit, and consistency is determined based on the maximum and second-largest fusion confidence values ​​and a consistency threshold. When consistency is achieved, the fusion confidence is multiplicatively increased by a preset bounded multiplicative coefficient, including: For each channel, the recognition model outputs the category logit and channel confidence. Channel health, normalized SNR, and channel type are converted into channel weights through a preset mapping relationship. The category logits of each channel are linearly weighted according to the channel weights to obtain the fusion category logit, and the fusion confidence is calculated based on the fusion category logit. The maximum and second maximum values ​​of the fusion confidence are determined, and the difference between the maximum and second maximum values ​​is calculated. The difference is compared with a consistency threshold. If the difference is greater than or equal to the consistency threshold, the multi-channel results are considered consistent; if the difference is less than the consistency threshold, the multi-channel results are considered inconsistent. When the multi-channel results are consistent, the fusion confidence is multiplicatively increased according to a preset bounded multiplicative coefficient. When the multi-channel results are inconsistent, no multiplicative increase is performed. The value range of the preset bounded multiplicative coefficient is limited to the interval greater than 1 and less than or equal to the upper limit of the increase coefficient.

[0012] Furthermore, in the single-channel case, the confidence stability is calculated from the channel confidence within multiple sliding time windows, and the inference threshold is adjusted within a preset adaptive range based on channel health and normalized SNR, including: The variance or range of the maximum class channel confidence score for a single channel across multiple consecutive sliding time windows is calculated to obtain the confidence score volatility. Confidence score stability is then calculated based on the confidence score volatility. Upper and lower stability thresholds are set, as are upper and lower channel quality thresholds. When the confidence score stability is higher than the upper stability threshold and both channel health and normalized SNR are higher than the upper channel quality threshold, the inference threshold adjustment is set to a negative value to lower the inference threshold. When the confidence score stability is lower than the lower stability threshold or both channel health and normalized SNR are lower than the lower channel quality threshold, the inference threshold adjustment is set to a positive value to raise the inference threshold. The absolute value of the inference threshold adjustment does not exceed a preset adaptive range. Channel quality is a weighted combination of channel health and normalized SNR.

[0013] Furthermore, the node-level identification results are bound to partial discharge type labels and confidence levels with time, power frequency phase, and channel-to-physical location mapping to form node-level positioning information, handling suggestions, and early warning records; a review time limit is set for the review task, and a low-confidence alarm is output when the review task is not closed within the review time limit, including: For the generated node-level identification results, the partial discharge type label, confidence level, time information, power frequency phase information, and channel-to-physical location mapping information are combined into node-level positioning information. Based on the node-level identification results and node-level positioning information, handling suggestions and early warning records are generated. A review time limit is set for each generated review task. Within the review time limit, the operation and maintenance personnel or the automatic review module provide a review conclusion. After the review conclusion is confirmed, the node-level identification results and handling suggestions of the corresponding node are updated. If the review task exceeds the review time limit and no review conclusion is generated, a low confidence alarm is output for the corresponding node and recorded as an unclosed-loop alarm event.

[0014] A GIS partial discharge type identification system based on deep learning network, applying the aforementioned GIS partial discharge type identification method based on deep learning network, includes: a signal acquisition and synchronization module, a variational mode decomposition module, a feature generation and gating screening module, a deep learning identification module, a multi-channel logit fusion and consistency judgment module, a single-channel threshold adaptive judgment module, a location and treatment suggestion module, a review task management module, and a data processing module; The signal acquisition and synchronization module, variational mode decomposition module, feature generation and gating filtering module, deep learning recognition module, multi-channel logit fusion and consistency judgment module, single-channel threshold adaptive judgment module, positioning and handling suggestion module, and review task management module are respectively connected to the data processing module. The signal acquisition and synchronization module is used to synchronize the time and align the power frequency phase of partial discharge signals acquired from multiple discharge signal acquisition channels; The variational mode decomposition module is used to perform variational mode decomposition on the time-synchronized partial discharge signal under given decomposition parameters to obtain the mode components of each channel; The feature generation and gating filtering module is used to calculate the modal component energy entropy within a sliding time window, form a candidate feature sequence, perform correlation filtering based on category cost, input channel health, normalized SNR and phase coverage into the learnable gating network, and output a feature subset. The deep learning recognition module is used to construct a one-dimensional convolutional neural network, train the recognition model using a cost-sensitive focus loss function, perform temperature scaling calibration on the recognition output of the validation set, output the calibrated confidence and inference threshold, and output the partial discharge type label, channel confidence and channel category logit during the inference stage. The multi-channel logit fusion and consistency determination module is used to calculate channel weights based on the determined channel health, normalized SNR and channel type in the case of multiple channels, perform weighted fusion of logits of each channel category, determine consistency based on fusion confidence and consistency threshold, multiplicatively increase the fusion confidence according to the preset bounded multiplicative coefficient when consistency is achieved, and generate node-level recognition results or review tasks according to the relationship between fusion confidence and inference threshold; when consistency is not achieved, a review task is generated. The single-channel threshold adaptive determination module is used to calculate the confidence stability based on the channel confidence change within multiple sliding time windows in the case of a single channel, and adjust the inference threshold within a preset adaptive range in combination with channel health and normalized SNR, and generate node-level recognition results or review tasks based on the relationship between the adjusted channel confidence and the inference threshold. The location and treatment suggestion module is used to bind the partial discharge type label and confidence level in the node-level identification results with time, power frequency phase and channel to physical location mapping, and output node-level location information, treatment suggestions and early warning records; The review task management module is used to set review time limits for review tasks, receive review conclusions, and output low confidence alarms when a review task is not closed within the review time limit; the data processing module is used to store, forward, and manage the data generated by each module.

[0015] Preferably, the determination of channel weights based on channel health, normalized SNR, and channel type includes: Channel health is calculated based on historical noise levels, hardware status records, and manual inspection results, and is used to characterize the reliability of the channel during partial discharge detection. Normalized SNR is calculated based on channel signal energy and channel noise energy, mapping the signal-to-noise ratio to a preset range through a linear function, and is used to participate in channel weight calculation and inference threshold adjustment. Phase coverage is calculated based on the occurrence range and frequency of partial discharge events on the power frequency phase within multiple power frequency cycles, and is used to characterize the effective sampling coverage within the power frequency cycle, and serves as the input to the learnable gating network.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: By constructing a learnable gating network using channel health, normalized SNR, and phase coverage, features from channels with low health, low SNR, and insufficient phase coverage are suppressed at the feature level, reducing the interference of noise and abnormal channels on the deep learning model.

[0017] A cost-sensitive focus loss function is used to handle class imbalance and important defect classes, and temperature scaling is used to calibrate the output confidence level, making the confidence level value more consistent with the true accuracy.

[0018] In the case of multiple channels, the channel category logit is weighted and fused. The fusion confidence is adjusted by combining consistency judgment and bounded multiplicative adjustment. This avoids confusion between reliable and poor channels caused by simple averaging, and improves the confidence of the consistency results through bounded multiplicative adjustment.

[0019] In the single-channel scenario, a confidence stability index is introduced, and the inference threshold is adaptively adjusted within a preset range in combination with channel health and normalized SNR, taking into account both false alarm rate and false negative rate.

[0020] A closed-loop process is constructed to manage node-level identification results, location information, handling suggestions, and review tasks. Low-confidence alarms are output for samples that are not closed within the review time limit, which facilitates operation and maintenance personnel to track and manage potential risks and improves the engineering availability and security of the system. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a GIS partial discharge type identification method based on deep learning networks. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0023] Example 1 like Figure 1 As shown, a method for identifying partial discharge types in GIS based on deep learning networks includes the following steps: Step 1: Time synchronization and power frequency phase alignment are performed on the partial discharge signals acquired by multiple discharge signal acquisition channels. Variational mode decomposition is performed under given decomposition parameters to obtain the modal components of each channel. Step 2: Calculate the energy entropy of the modal components within the sliding time window, and splice them together according to time and power frequency phase to form a candidate feature sequence; Step 3: Based on the category cost, perform relevance screening. Input the channel health, normalized SNR and phase coverage into the learnable gating network to gating the candidate feature sequences and obtain the feature subset. Step 4: Construct a one-dimensional convolutional neural network, take a subset of features as input, train the recognition model using a cost-sensitive focus loss function, perform temperature scaling calibration on the recognition output of the validation set, obtain the post-calibration confidence, and determine the inference threshold. Step 5: Perform steps 1 to 3 on the signal to be identified to obtain a feature subset. Input this subset into the identification model to obtain the partial discharge type label, channel confidence score, and channel category logit for each channel. In the case of multiple channels, determine the channel weights based on channel health, normalized SNR, and channel type. Perform weighted fusion of the channel category logits to obtain the fused category logit; where the channel category logit... The original category score represents the channel category; the fusion category logit represents the original category score obtained by weighted summation of all channel category logits; the fusion confidence is calculated based on the fusion category logit, and consistency is determined according to the maximum and second largest fusion confidence values ​​and the consistency threshold. When consistency is achieved, the fusion confidence is multiplicatively increased by a preset bounded multiplicative coefficient. Node-level recognition results or review tasks are generated according to the relationship between the increased fusion confidence and the inference threshold. When consistency is not achieved, a review task is generated. In the case of a single channel, the confidence stability is calculated from the channel confidence within multiple sliding time windows. The inference threshold is adjusted within a preset adaptive range by combining channel health and normalized SNR. Node-level recognition results or review tasks are generated according to the relationship between the adjusted channel confidence and the inference threshold. Here, the logit is the original category score output by the classification model. Step 6: Bind the partial discharge type label and confidence level of the node-level identification results to time, power frequency phase, and channel-to-physical location mapping to form node-level positioning information, handling suggestions, and early warning records; set a review time limit for the review task, and output a low confidence alarm if the review task is not closed within the review time limit.

[0024] Specifically, partial discharge sensors are deployed at several measurement points on the GIS equipment. These sensors can include ultra-high frequency sensors, extra-high frequency sensors, or high-frequency current transformers. Each sensor corresponds to a discharge signal acquisition channel. The acquisition system synchronously samples each channel to obtain multi-channel raw partial discharge signals, specifically including: The acquisition system uses a sampling frequency Synchronous sampling is performed on each channel, with a sampling interval of: Let the number of sampling points for each channel during the analysis period be . Then the first Each channel at the sampling time The discrete signal at point is denoted as: in This represents the number of channels.

[0025] One channel is selected as a reference channel. Using the partial discharge signal of the reference channel as a benchmark, cross-correlation calculations are performed between the signals of other channels and the reference channel signal. The maximum peak position of the cross-correlation sequence is found, and the time offset corresponding to the maximum peak position is used as the estimated time delay of each channel relative to the reference channel. Specifically, this includes: One channel is selected as a reference channel. Using the partial discharge signal of the reference channel as a benchmark, cross-correlation calculations are performed between the signals of other channels and the reference channel signal. The maximum peak position of the cross-correlation sequence is found, and the time offset corresponding to the maximum peak position is used as the estimated time delay of each channel relative to the reference channel. Specifically, this includes: Let the reference channel index be... Its signal is For any non-reference channel Reference Channel Perform discrete cross-correlation calculations: in This represents the discrete time delay.

[0026] Find the position of the maximum peak in the cross-correlation sequence: Then the first The estimated time delay of the channel relative to the reference channel is: When only sampling point-level alignment is required, the aligned discrete signal can be represented as: When subsampling precision is required, interpolation is performed on the original signal. Let the interpolation operator be... Then the aligned continuous-time signal is: Through the aforementioned cross-correlation time synchronization, the same partial discharge event in different channels is aligned on the time axis, providing a time consistency basis for subsequent joint processing of multiple channels.

[0027] Using the power frequency synchronization signal or the power frequency reference signal extracted from the voltage record, the power frequency phase position is calculated for each sampling point. Specifically, this includes: assuming the power frequency is... (For example, 50Hz), the power frequency period is: Set reference time Corresponding to the zero point of the power frequency phase, at any sampling time The corresponding power frequency phase is: in, Indicates the power frequency period The modulo operation; subsequently, the start time of the sliding time window will be... Mapped to the corresponding power frequency phase Furthermore, the characteristics of each time window are arranged in phase order within the same power frequency cycle to achieve phase alignment of the partial discharge signal within the power frequency cycle.

[0028] After completing time synchronization and power frequency phase alignment, the alignment signal for each channel is... Divide the time window into segments. Let the length of the sliding time window be... The sliding step size is , No. The starting index for each time window is: in This represents the total number of time windows.

[0029] Use a length of Hanming Window: This represents the sample number within the sliding time window; then the... Channel 1 The window function weighted signal within each sliding time window is: Before performing variational mode decomposition, each sliding time window signal can be mirrored and extended, that is, the signal is extended using symmetrical samples at both ends of the time window to reduce the edge effect of variational mode decomposition.

[0030] Median filtering is applied to the samples within each sliding time window to remove spikes and isolated sharp point interference; the average value of the samples within the sliding time window is calculated, and subtracted from the average value to remove DC interference; the sliding time window is judged to be abnormal based on the amplitude threshold or energy threshold, and abnormal sliding time windows exceeding the threshold are removed, specifically including: Let the median filter window length be... (Odd number), window half-width: The median-filtered samples are: in Let be the offset index relative to the current position; calculate the . Channel 1 Mean of the median filtered signal within each time window: And perform DC removal processing: After DC removal, the signal within the time window is symmetrically distributed around zero, which is beneficial for the stable calculation of subsequent energy-related characteristics. (Regarding the time window signal after DC removal...) Calculate the amplitude and energy indices. The amplitude index is defined as: The energy index is defined as: Under normal operating conditions, the maximum absolute value and energy distribution range of samples within the sliding time window are statistically analyzed, and an amplitude threshold is selected. and energy threshold When the time window satisfies: If the sliding time window is deemed abnormal, it will be removed from subsequent variational mode decomposition and feature extraction processes. Sliding time windows not marked as abnormal will retain the signal after median filtering and DC removal. This serves as the input for variational mode decomposition. Through the above preprocessing, the effective components of partial discharge are preserved, while isolated spike interference, DC drift, and obvious abnormal pulses are suppressed, improving the reliability of subsequent analysis.

[0031] Given the number of modes, bandwidth penalty factor, and initial value of center frequency, variational mode decomposition is performed for each effective sliding time window of each channel, specifically including: Given a number of modes Bandwidth penalty factor and the initial values ​​of the center frequencies of each mode. Under these conditions, variational mode decomposition (VMD) is performed for each channel and each effective sliding time window. In discrete form, the th... Channel 1 The signal within a time window can be approximately represented as the sum of its modal components: in For the first Channel 1 In the time window, the first The modal component within the time window The amplitude at each sampling point.

[0032] In continuous-time form, for the signal The VMD optimization model can be written as: in The unit impulse function, This represents the convolution operation. Represents the differential operator with respect to time. As a bandwidth penalty factor, it is adjusted by... Control the degree of contraction of each mode frequency band.

[0033] Energy is statistically analyzed for each channel, each sliding time window, and each modal component. Channel 1 Time Window The energy of a modal component is defined as: The total energy of all modal components within this time window is: Modal energy percentage is defined as: An energy entropy index is constructed based on the modal energy ratio to describe the degree of energy dispersion among different modes. Channel 1 The energy entropy of a time window is defined as: The more uniform the energy distribution, the better. The larger the value; when energy is concentrated in a few modes, The smaller the value, the more it reflects the complexity of the partial discharge signal in different modes.

[0034] In embodiments of the present invention, candidate features can be constructed in the following manner: In one implementation, energy entropy is used... As the first Channel 1 The scalar characteristics of the time window; in another implementation, the modal energy proportion vector is used. As the feature vector of this time window, the two can also be used in combination.

[0035] According to the starting time sequence corresponding to the sliding time window, the features within each time window are arranged by mode number, and the starting time of the time window is mapped to the power frequency phase position. The features corresponding to different power frequency phases within the same power frequency cycle are spliced ​​together in phase order to form a candidate feature sequence that simultaneously contains time sequence and power frequency phase information.

[0036] The category cost is set according to the number of partial discharge types in the training data and the engineering risk level. For partial discharge types with fewer samples and higher engineering risks, a larger category cost is assigned; for partial discharge types with more samples and lower risk levels, a smaller category cost is assigned. Specifically, this includes: Let the number of partial discharge types be The type index is When constructing the training dataset, the number of training samples for each type is counted. Let the total number of training samples be... and define: Sample scarcity is defined as: An engineering risk assessment was conducted for each type of partial discharge, and the engineering risk level was normalized to... The higher the risk The larger.

[0037] Considering both sample scarcity and engineering risk level, a category value is set: in As the sample scarcity weight, This represents the engineering risk weight. Partial discharge types with smaller sample sizes and higher engineering risks correspond to larger weights. Types with larger sample sizes and lower engineering risks correspond to smaller sample sizes. This allows the training process to focus more on minority and high-risk classes.

[0038] Let the dimension of the candidate feature vector be... , No. Each feature component is denoted as Training set, number 1 Each sample has a feature vector: The corresponding tag is , For each feature dimension and each type of partial discharge ,statistics: No. Feature mean of class samples: ;No. The characteristic variance of the sample class: The overall mean of all samples: Employ a correlation metric similar to Fisher's discriminant criterion: in To prevent small constants with a denominator of zero. The larger the value, the higher the value. The eigencomponent for the th feature component The higher the degree of distinction between the class and the whole.

[0039] The relevance score for each partial discharge type is weighted by category value: For each feature dimension By summing the weighted relevance contributions of all categories, we obtain the category cost weighted relevance index: Set correlation threshold ,when At that time, remove the first Each characteristic component; when At that time, the first one is reserved. Each feature component is used to obtain a candidate feature sequence that has been filtered by correlation, thereby reducing redundant features while ensuring discriminative ability.

[0040] Channel health is calculated based on historical noise levels, channel hardware status records, and manual inspection results. For the first... Each channel is defined as follows: Noise Level Score The lower and more stable the noise level, The larger the score, the better; hardware status score The fewer hardware defect alarms, the better. The larger the score, the higher the score for manual inspection. If the inspection is normal Approximately 1. Assume a weight. ,and: The channel health status is defined as: aisle The signal-to-noise ratio is defined by power as follows: in The effective signal power of the channel. For noise power. Given a lower limit for signal-to-noise ratio. and upper limit The normalized SNR is defined as: Divide one power frequency cycle into There are several phase intervals, with interval indices as follows: The first term obtained by dividing the power frequency period Each phase interval is numbered. The occurrence of partial discharge events in each phase interval within multiple power frequency cycles is statistically analyzed. For the first... Channel, define indicator variables: Phase coverage is defined as: Build based on channel health Normalized SNR and phase coverage The input is a learnable gating network. The three components are combined to form the input vector: Taking a two-layer fully connected neural network as an example, the output of the first layer is: Second layer output gating coefficient vector: in: This is the weight matrix; It is the bias vector; For the Sigmoid function: ; For the gating coefficient vector consistent with the feature dimension, each component .

[0041] Let the feature components after correlation screening be... ,in For channel index, For indexing time windows, Indexed by feature dimension. The gated features are: Further set the gate threshold Hard-gating the gated features: The set of retained feature components The feature subsets are arranged in order of time window to form a feature time series, which serves as the input to the one-dimensional convolutional neural network recognition model. The gating network and the recognition model are jointly optimized during training, enabling the gating strategy to adaptively adjust the contribution weights of each channel's features based on channel quality.

[0042] A one-dimensional convolutional neural network is used as the recognition model structure. The feature subset is input into the one-dimensional convolutional neural network in time series form. The time series features are extracted step by step through multiple convolutional layers, nonlinear activation layers and pooling layers. Finally, the category logit corresponding to each partial discharge type is output through a fully connected layer.

[0043] For the For each training sample, the network outputs a logit vector as follows: The Softmax probability is: The real label is Let the predicted probability of the true class of this sample be: To improve the identification performance of minority and high-risk partial discharge types, a cost-sensitive focus loss function is constructed by introducing category cost and category balance coefficient on the basis of standard focus loss. Let the focusing parameters be... The category balance coefficient is The category cost value is Then the first The loss for each sample is: The average loss over the entire training set is: The parameters of the one-dimensional convolutional neural network and the gated network are jointly updated through backpropagation algorithm and gradient optimization until the loss function converges or the preset number of training rounds is reached.

[0044] After training, the logit output by the recognition model and the true label are recorded on an independent validation set. Temperature scaling is applied to the logit to calibrate the confidence level. Let the temperature parameter be... The scaled logit is: The probability after calibration is: The prediction category is: The corresponding confidence level is: By adjusting the temperature parameters This ensures that the average predicted confidence level of samples within a given confidence interval on the validation set is as close as possible to the actual accuracy, satisfying the pre-set reliability index. After calibration, a fixed temperature parameter is used during the inference phase. Convert the logit to a calibrated confidence level.

[0045] The relationship between the calibrated confidence level and the true label was statistically analyzed on the validation set by scanning different confidence thresholds. We examined the corresponding false positive rate, false negative rate, and overall performance indicators, and selected a confidence threshold that takes into account both the false positive rate and the false negative rate as the inference threshold.

[0046] Based on channel health Normalized SNR And establish channel weight mapping relationships based on channel type. Channel type reflects the geometric relationship between the sensor and key parts of the device under test, and the channel type is mapped to a weight correction factor. .

[0047] First, calculate the basic weights based on channel health and normalized SNR: in This is the weighting parameter. Then, it is multiplied by the channel type correction factor to obtain the unnormalized weights: Normalize all channels to obtain channel weights: Extract a subset of features from the signal to be identified using the steps described above and input it into the trained recognition model to obtain the class logit vector for each channel. .

[0048] For each type of partial discharge The fused logit is obtained by weighting and summing the logits of each channel according to their respective channel weights. The fusion confidence score is obtained by inputting the fusion logit into Softmax through temperature scaling: Let the maximum fusion confidence and its category be: The confidence level of the largest fusion is: The difference between the two is: Let the consistency threshold be... ,when When the results from multiple channels are consistent, it is determined that the results from all channels are consistent; when When the results from multiple channels are inconsistent, it is determined that the results are not consistent.

[0049] To enhance the confidence of highly consistent results when multiple channels yield consistent outcomes, a bounded multiplicative adjustment mechanism is introduced. Let the multiplicative adjustment coefficient be... ,satisfy The maximum fusion confidence level after adjustment is: When the results from multiple channels are inconsistent, do not perform multiplicative up-adjustment, let: The adjusted fusion confidence and inference threshold Compare: When the results from multiple channels are consistent and At that time, the node-level identification results are output, and the partial discharge type corresponding to the maximum fusion confidence is used as the node-level identification label; when If the results from multiple channels are inconsistent, a review task will be generated.

[0050] When only one channel is available or only one effective channel remains among multiple channels, an adaptive inference threshold adjustment mechanism based on confidence stability and channel quality is introduced to improve decision reliability. Let the maximum channel confidence of this single channel within multiple consecutive sliding time windows be: Its average value is: Confidence level volatility can be expressed using standard deviation: Let the maximum permissible volatility be The confidence stability is defined as: The smaller the confidence level fluctuation, the greater the stability. The closer it is to 1.

[0051] Channel quality is characterized by a weighted combination of channel health and normalized SNR: in These are the weight parameters.

[0052] Set a stability threshold Stability threshold Channel quality threshold Channel quality threshold and the maximum magnitude of the inference threshold adjustment. Define the inference threshold adjustment amount: The adjusted inference threshold is: In the case of a single channel, if the channel confidence level satisfy: If the output is a node-level identification result, then the node-level identification result will be output; otherwise, a review task will be generated.

[0053] The generated node-level identification results are combined with the partial discharge type label, final confidence level, time information, power frequency phase information, and channel-to-physical location mapping information to form node-level positioning information. Based on the node-level positioning information, corresponding handling suggestions are generated for different partial discharge types and locations, such as arranging power outages for related equipment maintenance, increasing inspection frequency, or strengthening online monitoring, and corresponding early warning records are generated in the operation and maintenance system.

[0054] A review time limit is set for each generated review task, configured based on the importance of the equipment and the operation and maintenance strategy. Within the review time limit, operation and maintenance personnel or the automatic review module review the original signal, features, and identification results, and provide a review conclusion. The review conclusion may include: confirming the original identification result, correcting the partial discharge type label, or marking the event as invalid, etc. After the review conclusion is confirmed, the node-level identification result and handling suggestions for the corresponding node are updated accordingly, realizing closed-loop alarm management.

[0055] If a review task fails to generate a review conclusion within the review time limit, a low-confidence alarm will be output to the corresponding node, and the event will be recorded as an unclosed-loop alarm event for subsequent tracking and statistics by the operation and maintenance management system.

[0056] Example 2 A GIS partial discharge type identification system based on deep learning network, applying the aforementioned GIS partial discharge type identification method based on deep learning network, includes: a signal acquisition and synchronization module, a variational mode decomposition module, a feature generation and gating screening module, a deep learning identification module, a multi-channel logit fusion and consistency judgment module, a single-channel threshold adaptive judgment module, a location and treatment suggestion module, a review task management module, and a data processing module; The signal acquisition and synchronization module, variational mode decomposition module, feature generation and gating filtering module, deep learning recognition module, multi-channel logit fusion and consistency judgment module, single-channel threshold adaptive judgment module, positioning and handling suggestion module, and review task management module are respectively connected to the data processing module. The signal acquisition and synchronization module is used to synchronize the time and align the power frequency phase of partial discharge signals acquired from multiple discharge signal acquisition channels; The variational mode decomposition module is used to perform variational mode decomposition on the time-synchronized partial discharge signal under given decomposition parameters to obtain the mode components of each channel; The feature generation and gating filtering module is used to calculate the modal component energy entropy within a sliding time window, form a candidate feature sequence, perform correlation filtering based on category cost, input channel health, normalized SNR and phase coverage into the learnable gating network, and output a feature subset. The deep learning recognition module is used to construct a one-dimensional convolutional neural network, train the recognition model using a cost-sensitive focus loss function, perform temperature scaling calibration on the recognition output of the validation set, output the calibrated confidence and inference threshold, and output the partial discharge type label, channel confidence and channel category logit during the inference stage. The multi-channel logit fusion and consistency determination module is used to calculate channel weights based on channel health, normalized SNR and channel type in multi-channel scenarios, perform weighted fusion of logits for each channel category, determine consistency based on fusion confidence and consistency threshold, multiplicatively increase fusion confidence by a preset bounded multiplicative coefficient when consistency is achieved, and generate node-level recognition results or review tasks based on the relationship between fusion confidence and inference threshold; and generate review tasks when consistency is not achieved. The single-channel threshold adaptive determination module is used to calculate the confidence stability based on the channel confidence change within multiple sliding time windows in the case of a single channel, and adjust the inference threshold within a preset adaptive range in combination with channel health and normalized SNR, and generate node-level recognition results or review tasks based on the relationship between the adjusted channel confidence and the inference threshold. The location and treatment suggestion module is used to bind the partial discharge type label and confidence level in the node-level identification results with time, power frequency phase and channel to physical location mapping, and output node-level location information, treatment suggestions and early warning records; The review task management module is used to set review time limits for review tasks, receive review conclusions, and output low confidence alarms when a review task is not closed within the review time limit; the data processing module is used to store, forward, and manage the data generated by each module.

[0057] Example 3 This embodiment provides a specific implementation of the method and system of the present invention deployed on a 220kV GIS bay, which is used to illustrate the selection process of each indicator and parameter and the entire online identification process.

[0058] Three partial discharge detection channels, denoted as Channel 1, Channel 2, and Channel 3, are arranged in the GIS bus bay and circuit breaker bay of a 220kV substation. Channel 1 is located on the bus side and uses an ultra-high frequency partial discharge sensor; Channel 2 is located near the circuit breaker and uses an ultra-high frequency partial discharge sensor; Channel 3 is located near the cable terminal and uses a high frequency current transformer.

[0059] The data acquisition device synchronously samples three channels at a sampling frequency of 125 MHz and a single acquisition duration of one second. The power frequency synchronization signal is provided through a voltage transformer.

[0060] Channels 2 and 3 are cross-correlated with the sampled signals of channel 1 to find the maximum peak position of the cross-correlation sequence and obtain the time offset of channel 2 and channel 3 relative to channel 1.

[0061] Based on this time offset, channel 2 and channel 3 are time-aligned by interpolation resampling to ensure that the same partial discharge pulse in the three channels is aligned as much as possible in terms of sampling point number.

[0062] The power frequency phase corresponding to each sampling point is obtained by using zero-point detection of the power frequency voltage signal. The data collected every second is divided into several complete power frequency cycles, and a power frequency phase mark is added to each sample point.

[0063] The signal after resampling for each channel is divided into multiple sliding time windows using a Hamming window with a length of 1,000 points and a step size of 500 points.

[0064] For each sliding time window, a mirror extension is first performed, extending the sample sequence symmetrical to the original sample at both ends of the time window to alleviate the edge effect of variational mode decomposition.

[0065] Within the sliding time window, median filtering is performed using a sliding window of fixed length of five points. The median value after sorting within each of the five points is used as the output at that position, and the original sample value is replaced with this output to remove isolated cusp interference. The average value of all samples within the sliding time window is calculated, and each sample value is subtracted from this average value to achieve DC-DC demeaning. Based on historical normal operating data, the typical range of sample absolute values ​​is statistically analyzed, and the upper limit of the sample absolute value is set as the amplitude threshold, while the upper limit of the energy accumulation value within the window is set as the energy threshold. The absolute value of all samples within the current sliding time window is compared point by point with the amplitude threshold. If any sample absolute value exceeds the amplitude threshold, the sliding time window is marked as abnormal. Simultaneously, the energy accumulation value within the sliding time window is calculated and compared with the energy threshold. When the energy accumulation value exceeds the energy threshold, the sliding time window is also marked as abnormal. Sliding time windows marked as abnormal do not participate in subsequent variational mode decomposition.

[0066] In this embodiment, the number of modes is set to six, the bandwidth penalty factor is set to a medium level, and the initial center frequencies are selected from six center frequencies according to the uniform frequency band coverage method. Variational mode decomposition is performed on each channel for each normal sliding time window, outputting six modal components. The obtained modal components are stored according to the channel number, time window number, and mode number.

[0067] For each channel, each sliding time window, and each modal component, calculate the energy of that modal component within that sliding time window. The energy is calculated by summing the squared amplitudes of the samples within the sliding time window point by point, and the sum is taken as the energy of that modal component.

[0068] The energies of the six modal components within the same sliding time window are normalized, and the energy of each modality is divided by the sum of the energies of the six modes to obtain the six energy percentages. The energy percentages reflect the relative energy contribution of each modal component within the current sliding time window.

[0069] An energy distribution is constructed based on the energy proportions, and an entropy value describing the uniformity of the energy distribution is calculated as an energy entropy index. The more dispersed the energy distribution, the closer the energy proportions of multiple modes are, and the larger the energy entropy value; the more concentrated the energy distribution, the more a few modes occupy the majority of the energy, and the smaller the energy entropy value. An energy entropy value is obtained for each channel and each sliding time window.

[0070] The start time of each sliding time window is associated with the power frequency phase. The energy entropy indices obtained from multiple sliding time windows of the same channel are arranged and spliced ​​according to the time order and the power frequency phase order to form a candidate feature sequence jointly encoded by time and power frequency phase.

[0071] In this embodiment, the partial discharge types include five categories: conductor tip discharge, floating potential discharge, particle discharge, air gap discharge, and noise.

[0072] When constructing the training dataset, the number of samples for each type of partial discharge was counted, and an engineering risk level was given based on operational experience. For example, air gap discharge and particle discharge, which may lead to insulation breakdown, were classified as high-risk, floating potential discharge as medium-risk, and conductor tip discharge and noise as low-risk.

[0073] A category cost is assigned to each type of partial discharge based on the sample size and risk level. Types with fewer samples and higher risk levels are assigned larger category costs, while types with more samples and lower risk levels are assigned smaller category costs, forming a mapping table from type to category cost value.

[0074] For each feature component in the candidate feature sequence, the training data is divided into samples according to the partial discharge type. The numerical distribution of the feature component within each type is statistically analyzed and compared with the distribution of the feature component under other types to obtain the correlation score of the feature component to each partial discharge type.

[0075] Multiply the relevance score of each type by the corresponding category cost to obtain the weighted relevance contribution. Summate or take the weighted average of the weighted relevance contributions of all types to obtain the category cost-weighted relevance index for that feature component. Repeat the above process for all feature components in the candidate feature sequence to form a set of category cost-weighted relevance indices.

[0076] A relevance threshold is pre-selected, and feature components whose category cost weighted relevance index is lower than the threshold are removed to obtain a candidate feature sequence after relevance screening.

[0077] Channel health is calculated separately for each of the three channels. Channel health consists of three parts: historical noise level score, hardware condition score, and manual inspection score.

[0078] Historical noise level score is calculated by statistically analyzing background noise energy during nighttime or maintenance power outages; the lower the noise energy, the higher the score. Hardware status score is assessed based on the number of faults and alarms recorded in the equipment operation log for that channel; fewer fault records result in a higher score. Manual inspection score is given based on the results of on-site inspections of connectors, sensor installation, and grounding conditions. All three scores are normalized to a range of zero to one and summed according to preset weights to obtain the channel health index.

[0079] Normalized SNR is obtained by calculating the ratio of signal energy to background noise energy during partial discharge activity. The original SNR is then compared with a lower threshold and an upper threshold. If the original SNR is below the lower threshold, it is mapped to zero; if it is above the upper threshold, it is mapped to one; and if it is in between, it is mapped to a value between zero and one in a linear proportion.

[0080] Phase coverage is obtained by statistically analyzing the range of power frequency phases in which partial discharge events occur over multiple power frequency cycles within a certain time period. The power frequency cycle is divided into several equally wide phase intervals, and the number of partial discharge events occurring in each phase interval is counted. The ratio of the number of phase intervals with a greater than zero occurrence to the total number of phase intervals is the phase coverage.

[0081] A two-layer fully connected neural network is used as a learnable gating network. The inputs are channel health, normalized SNR, and phase coverage. The output is a gating coefficient vector with the same dimension as the candidate feature sequence after correlation screening. Internally, the neural network maps the three inputs to a set of gating coefficients with the same number of feature components through linear transformation and nonlinear activation. The gating coefficients are restricted to the interval between zero and one.

[0082] The gating coefficients are multiplied dimension-by-dimensionally by the candidate feature sequences filtered for correlation to obtain the gated feature values. A gating threshold is set, and components with gated feature values ​​less than the threshold are set to zero, while other components are retained, resulting in a feature subset. The learnable gated network can be used in conjunction with a one-dimensional convolutional neural network for parameter updates during the training phase.

[0083] A subset of features is used as input samples, with each sample labeled as one of the five types of partial discharge mentioned above. A recognition model is constructed using three one-dimensional convolutional layers plus a fully connected classification layer.

[0084] Training is performed using a cost-sensitive focus loss function. The focus loss includes class cost weights, which are set according to the aforementioned class cost mapping table, giving higher weight to partial discharge types with fewer samples and higher risk levels in the loss calculation.

[0085] After training, an independent validation set is selected, and each sample in the validation set is input into the network. The output logit of each partial discharge type category and the true label are recorded.

[0086] Without changing other network parameters, a single adjustable temperature parameter is used to scale the logit for all categories. The temperature parameter is searched starting from an initial value. For each candidate temperature parameter, the scaled logit is converted into a confidence score, and the difference between the average predicted confidence score and the true accuracy across multiple confidence score intervals is calculated. By comparing the magnitude of these differences, the temperature parameter that results in the smallest overall difference is selected as the temperature scaling calibration result.

[0087] During the inference phase, the class logit output by the network is scaled using the temperature parameter before being converted into a calibrated confidence level. The false positive rate and false negative rate corresponding to different confidence thresholds are calculated on the validation set, and one confidence threshold is selected as the inference threshold. In this embodiment, the inference threshold is set to 0.8 to balance false positive control and the detection rate of important defect types.

[0088] During normal operation, all three channels are available. The above steps are performed on the partial discharge signals acquired within a certain time period to obtain feature subsets of channel 1, channel 2, and channel 3. These subsets are then input into the recognition model to obtain the channel confidence score and channel category logit for each channel.

[0089] Channel health, normalized SNR, and channel type are used to calculate channel weights. Channel 1 is close to the busbar, channel 2 is close to the circuit breaker, and channel 3 is close to the cable terminal. Based on criticality and historical performance, channel 2 has the highest health, followed by channel 1, and channel 3 has a relatively lower health. Regarding normalized SNR, under current operating conditions, channel 2 has the highest normalized SNR, followed by channel 3, and channel 1 is slightly lower. Through a preset mapping relationship, channel health and normalized SNR are combined into a base weight value, and combined with a channel type correction factor to obtain the three channel weights. After normalization, the sum of the three weights is equal to one.

[0090] For each partial discharge type, the category logits of the three channels are weighted and summed according to their channel weights to obtain the fused category logit for that partial discharge type. This process is repeated for all partial discharge types to form a set of fused category logits. The fused category logits are converted into fused confidence scores, and the maximum and second-largest fused confidence scores are identified. The difference between the two scores is calculated and compared with a consistency threshold. In this embodiment, the consistency threshold is set to 0.2.

[0091] When the difference is greater than or equal to 0.2, the fusion results are considered consistent, and the maximum fusion confidence is multiplicatively increased according to a preset bounded multiplicative coefficient. The bounded multiplicative coefficient is limited to the range of 1 to 1.2; in this embodiment, the multiplicative increase coefficient is 1.1. When the increased fusion confidence is greater than or equal to the inference threshold of 0.8, a node-level recognition result is generated; when the increased fusion confidence is less than the inference threshold, a review task is generated. When the difference is less than 0.2, the fusion results are considered inconsistent, and a review task is directly generated.

[0092] During certain runtime periods, for example, channel 3 may be temporarily out of service due to maintenance, leaving only channels 1 and 2 available, or a channel may be blocked by the system due to extremely low health, leaving only one channel available. When only one channel is available, the maximum channel confidence score of that channel within multiple consecutive sliding time windows is recorded, and the confidence score fluctuation is calculated.

[0093] In this embodiment, when mapping confidence volatility to confidence stability, a monotonically decreasing mapping is used: the smaller the confidence volatility, the closer the confidence stability is to the upper limit of one; the larger the confidence volatility, the closer the confidence stability is to the lower limit of zero. The upper threshold for stability is set to 0.7, the lower threshold for stability is set to 0.3, and the upper and lower thresholds for channel quality are 0.7 and 0.3, respectively.

[0094] When the confidence stability is greater than 0.7 and the channel quality is greater than 0.7, the inference threshold adjustment is set to a negative value, for example, -0.05, adjusting the inference threshold from 0.8 to 0.75. When the confidence stability is less than 0.3 or the channel quality is less than 0.3, the inference threshold adjustment is set to a positive value, for example, 0.05, adjusting the inference threshold from 0.8 to 0.85. The preset adaptive amplitude is limited to 0.05, and in this embodiment, the absolute value of the adjustment does not exceed this limit.

[0095] For single-channel samples, when the adjusted channel confidence is greater than or equal to the adjusted inference threshold, a node-level recognition result is generated; when the adjusted channel confidence is less than the adjusted inference threshold, a verification task is generated.

[0096] The system combines the generated node-level identification results with the physical location mapping table in the GIS equipment structure model to obtain the interval, equipment unit, and relative position corresponding to a certain partial discharge event. The partial discharge type label, confidence level, timestamp, power frequency phase, and channel-to-physical location mapping information are combined into node-level positioning information.

[0097] Based on different partial discharge types and confidence levels, corresponding handling recommendations are generated. For example, for high-confidence air gap discharge, it is recommended to arrange a power outage for maintenance in a short period of time. For medium-confidence particulate discharge, it is recommended to increase the frequency of inspections and shorten the cycle of the next partial discharge test. The results are then written into the early warning record.

[0098] For all generated review tasks, the system sets a review timeframe, such as 24 hours. Within this timeframe, operations and maintenance personnel provide review conclusions through secondary analysis, on-site verification, or other auxiliary testing methods. After the review conclusions are provided, the system updates the corresponding node's tag and handling recommendations. For tasks that are not reviewed within the timeframe, the system automatically outputs a low-confidence alarm for the corresponding node and marks the alarm event as an unclosed-loop event for separate statistics and tracking by operations and maintenance personnel.

[0099] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying partial discharge types in GIS based on deep learning networks, characterized in that, Includes the following steps: Step 1: Time synchronization and power frequency phase alignment are performed on the partial discharge signals acquired by multiple discharge signal acquisition channels. Variational mode decomposition is performed under given decomposition parameters to obtain the modal components of each channel. Step 2: Calculate the energy entropy of the modal components within the sliding time window, and splice them together according to time and power frequency phase to form a candidate feature sequence; Step 3: Based on the category cost, perform relevance screening. Input the channel health, normalized SNR and phase coverage into the learnable gating network to gating the candidate feature sequences and obtain the feature subset. Step 4: Construct a one-dimensional convolutional neural network, take a subset of features as input, train the recognition model using a cost-sensitive focus loss function, perform temperature scaling calibration on the recognition output of the validation set, obtain the post-calibration confidence, and determine the inference threshold. Step 5: Perform steps 1 to 3 to obtain a feature subset from the signal to be identified, input it into the identification model, and obtain the partial discharge type label, channel confidence, and channel category logit for each channel. In the case of multiple channels, determine the channel weights based on channel health, normalized SNR, and channel type. Perform weighted fusion of the channel category logit to obtain the fusion category logit. Calculate the fusion confidence based on the fusion category logit, and determine consistency based on the maximum and second-largest values ​​of the fusion confidence and the consistency threshold. When consistency is achieved, multiplicatively increase the fusion confidence according to a preset bounded multiplicative coefficient. Generate node-level identification results or verification tasks according to the relationship between the increased fusion confidence and the inference threshold. Generate verification tasks when consistency is not achieved. In the case of a single channel, the confidence stability is calculated from the channel confidence within multiple sliding time windows. The inference threshold is adjusted within a preset adaptive range by combining the channel health and normalized SNR. Node-level recognition results or verification tasks are generated according to the relationship between the adjusted channel confidence and the inference threshold. Step 6: Bind the partial discharge type label and confidence level of the node-level identification results to time, power frequency phase, and channel-to-physical location mapping to form node-level positioning information, handling suggestions, and early warning records; set a review time limit for the review task, and output a low confidence alarm if the review task is not closed within the review time limit.

2. The method for identifying partial discharge types in GIS based on deep learning networks according to claim 1, characterized in that, The process of time synchronization and power frequency phase alignment of partial discharge signals acquired from multiple discharge signal acquisition channels, followed by variational mode decomposition under given decomposition parameters to obtain the modal components of each channel, includes: Cross-correlation peak time delay estimation is performed on partial discharge signals acquired from multiple discharge signal acquisition channels. Based on the time delay estimation results, the partial discharge signals of each channel are resampled and made consistent. The partial discharge signals of each channel after resampling are divided into multiple sliding time windows according to Hamming windows, and mirror extension is performed on each sliding time window. Median filtering is performed on the samples in each sliding time window to remove spikes and mean filtering to remove DC. Abnormal sliding time windows are removed according to amplitude thresholds or energy thresholds. Under the given number of modes, bandwidth penalty factor and initial value of center frequency, variational mode decomposition is performed on each sliding time window of each channel to obtain modal components.

3. The method for identifying partial discharge types in GIS based on deep learning networks according to claim 1, characterized in that, The method of calculating the energy entropy of modal components within a sliding time window and splicing them together according to time and power frequency phase to form a candidate feature sequence includes: Within each sliding time window, the energy entropy of each modal component obtained in step one is calculated to form an energy entropy sequence for each channel. The energy entropies are arranged according to the time order and power frequency phase order corresponding to the sliding time window. The energy entropies corresponding to different power frequency phases within the same power frequency cycle are spliced ​​together to form a candidate feature sequence jointly encoded by time and power frequency phase.

4. The GIS partial discharge type identification method based on deep learning network according to claim 1, characterized in that, The relevance screening based on category cost involves inputting channel health, normalized SNR, and phase coverage into a learnable gating network to gate candidate feature sequences, obtaining a feature subset, including: The category cost is determined based on the misclassification cost and sample size for each partial discharge type. The correlation index, which is weighted by the category cost, is used to calculate the correlation between each feature in the candidate feature sequence and the partial discharge type label. Features with correlation below the correlation threshold are removed to obtain the candidate feature sequence after correlation screening. A learnable gating network is constructed with channel health, normalized SNR, and phase coverage as inputs. The output is the gating coefficient corresponding to the dimension of the candidate feature sequence after correlation screening. The gating coefficient is multiplied dimension by dimension of the candidate feature sequence after correlation screening and compared with the gating threshold. The feature components below the gating threshold are truncated to obtain the feature subset.

5. The GIS partial discharge type identification method based on deep learning network according to claim 1, characterized in that, The construction of a one-dimensional convolutional neural network, using a feature subset as input, trains the recognition model using a cost-sensitive focus loss function, performs temperature scaling calibration on the recognition output of the validation set, obtains the post-calibration confidence, and determines the inference threshold, including: A one-dimensional convolutional neural network is used as the recognition model structure. Forward and backward propagation are performed on the feature subset. The cost-sensitive focus loss function is used for parameter updates. The class cost in the cost-sensitive focus loss function is set according to the class cost in step three. After training, the logit of each class output by the recognition model and the true label are recorded on the validation set. Temperature scaling calibration is performed by optimizing the temperature parameters to ensure that the post-calibration confidence meets the preset confidence reliability index. The inference threshold is selected based on the post-calibration confidence distribution and the partial discharge type label distribution.

6. The method for identifying partial discharge types in GIS based on deep learning networks according to claim 1, characterized in that, In the multi-channel scenario, channel weights are determined based on channel health, normalized SNR, and channel type. A weighted fusion of channel category logits is performed to obtain a fusion category logit. Fusion confidence is calculated based on the fusion category logit, and consistency is determined based on the maximum and second-largest fusion confidence values ​​and a consistency threshold. When consistency is achieved, the fusion confidence is multiplicatively increased by a preset bounded multiplicative coefficient, including: For each channel, the recognition model outputs the category logit and channel confidence. Channel health, normalized SNR, and channel type are converted into channel weights through a preset mapping relationship. The category logits of each channel are linearly weighted according to the channel weights to obtain the fusion category logit, and the fusion confidence is calculated based on the fusion category logit. The maximum and second maximum values ​​of the fusion confidence are determined, and the difference between the maximum and second maximum values ​​is calculated. The difference is compared with a consistency threshold. If the difference is greater than or equal to the consistency threshold, the multi-channel results are considered consistent; if the difference is less than the consistency threshold, the multi-channel results are considered inconsistent. When the multi-channel results are consistent, the fusion confidence is multiplicatively increased according to a preset bounded multiplicative coefficient. When the multi-channel results are inconsistent, no multiplicative increase is performed. The value range of the preset bounded multiplicative coefficient is limited to the interval greater than 1 and less than or equal to the upper limit of the increase coefficient.

7. The GIS partial discharge type identification method based on deep learning network according to claim 1, characterized in that, In the single-channel scenario, the confidence stability is calculated from the channel confidence within multiple sliding time windows, and the inference threshold is adjusted within a preset adaptive range based on channel health and normalized SNR, including: The variance or range of the maximum class channel confidence score for a single channel across multiple consecutive sliding time windows is calculated to obtain the confidence score volatility. Confidence score stability is then calculated based on the confidence score volatility. Upper and lower stability thresholds are set, as are upper and lower channel quality thresholds. When the confidence score stability is higher than the upper stability threshold and both channel health and normalized SNR are higher than the upper channel quality threshold, the inference threshold adjustment is set to a negative value to lower the inference threshold. When the confidence score stability is lower than the lower stability threshold or both channel health and normalized SNR are lower than the lower channel quality threshold, the inference threshold adjustment is set to a positive value to raise the inference threshold. The absolute value of the inference threshold adjustment does not exceed a preset adaptive range. Channel quality is a weighted combination of channel health and normalized SNR.

8. The GIS partial discharge type identification method based on deep learning network according to claim 1, characterized in that, The node-level identification results are bound to partial discharge type labels and confidence levels with time, power frequency phase, and channel-to-physical location mapping to form node-level positioning information, handling suggestions, and early warning records. Set a review time limit for the review task. If the review task is not closed within the review time limit, output a low confidence alarm, including: For the generated node-level identification results, the partial discharge type label, confidence level, time information, power frequency phase information, and channel-to-physical location mapping information are combined into node-level positioning information. Based on the node-level identification results and node-level positioning information, handling suggestions and early warning records are generated. A review time limit is set for each generated review task. Within the review time limit, the operation and maintenance personnel or the automatic review module provide a review conclusion. After the review conclusion is confirmed, the node-level identification results and handling suggestions of the corresponding node are updated. If the review task exceeds the review time limit and no review conclusion is generated, a low confidence alarm is output for the corresponding node and recorded as an unclosed-loop alarm event.

9. A GIS partial discharge type identification system based on deep learning networks, characterized in that, The GIS partial discharge type identification method based on deep learning network according to any one of claims 1-8 includes: a signal acquisition and synchronization module, a variational mode decomposition module, a feature generation and gating screening module, a deep learning identification module, a multi-channel logit fusion and consistency judgment module, a single-channel threshold adaptive judgment module, a location and treatment suggestion module, a review task management module, and a data processing module. The signal acquisition and synchronization module, variational mode decomposition module, feature generation and gating filtering module, deep learning recognition module, multi-channel logit fusion and consistency judgment module, single-channel threshold adaptive judgment module, positioning and handling suggestion module, and review task management module are respectively connected to the data processing module. The signal acquisition and synchronization module is used to synchronize the time and align the power frequency phase of partial discharge signals acquired from multiple discharge signal acquisition channels; The variational mode decomposition module is used to perform variational mode decomposition on the time-synchronized partial discharge signal under given decomposition parameters to obtain the mode components of each channel; The feature generation and gating filtering module is used to calculate the modal component energy entropy within a sliding time window, form a candidate feature sequence, perform correlation filtering based on category cost, input channel health, normalized SNR and phase coverage into the learnable gating network, and output a feature subset. The deep learning recognition module is used to construct a one-dimensional convolutional neural network, train the recognition model using a cost-sensitive focus loss function, perform temperature scaling calibration on the recognition output of the validation set, output the calibrated confidence and inference threshold, and output the partial discharge type label, channel confidence and channel category logit during the inference stage. The multi-channel logit fusion and consistency determination module is used to calculate channel weights based on the determined channel health, normalized SNR and channel type in the case of multiple channels, perform weighted fusion of logits of each channel category, determine consistency based on fusion confidence and consistency threshold, multiplicatively increase the fusion confidence according to the preset bounded multiplicative coefficient when consistency is achieved, and generate node-level recognition results or review tasks according to the relationship between fusion confidence and inference threshold; when consistency is not achieved, a review task is generated. The single-channel threshold adaptive determination module is used to calculate the confidence stability based on the channel confidence change within multiple sliding time windows in the case of a single channel, and adjust the inference threshold within a preset adaptive range in combination with channel health and normalized SNR, and generate node-level recognition results or review tasks based on the relationship between the adjusted channel confidence and the inference threshold. The location and treatment suggestion module is used to bind the partial discharge type label and confidence level in the node-level identification results with time, power frequency phase and channel to physical location mapping, and output node-level location information, treatment suggestions and early warning records; The review task management module is used to set review time limits for review tasks, receive review conclusions, and output low confidence alarms when a review task is not closed within the review time limit; the data processing module is used to store, forward, and manage the data generated by each module.

10. A GIS partial discharge type identification system based on a deep learning network according to claim 9, characterized in that, The method of determining channel weights based on channel health, normalized SNR, and channel type includes: Channel health is calculated based on historical noise levels, hardware status records, and manual inspection results, and is used to characterize the reliability of the channel during partial discharge detection. Normalized SNR is calculated based on channel signal energy and channel noise energy, mapping the signal-to-noise ratio to a preset range through a linear function, and is used to participate in channel weight calculation and inference threshold adjustment. Phase coverage is calculated based on the occurrence range and frequency of partial discharge events on the power frequency phase within multiple power frequency cycles, and is used to characterize the effective sampling coverage within the power frequency cycle, and serves as the input to the learnable gating network.

Images