A switch cabinet partial discharge pulse current discharge type identification method

Abstract

This invention provides a method for identifying the discharge type of partial discharge pulse current in switchgear, belonging to the field of current detection technology based on deep learning. It simultaneously acquires pulse waveforms from the target switchgear and adjacent switchgear, and also acquires transient ground voltage signals, UHF signals, and acoustic emission signals from the target switchgear. Through event alignment and quality assessment, a phase-resolved partial discharge image tensor, a phase statistical vector, an auxiliary time-frequency tensor, and a data quality vector are generated. Based on the phase-resolved partial discharge image branch, the phase statistical branch, and the auxiliary sensing branch, phase periodic feature vectors, phase statistical feature vectors, and auxiliary sensing feature vectors are obtained. A quality gating coefficient is generated from the data quality vector, and the three vectors are weighted and fused. Finally, an evidence-based deep learning classification module generates a predicted discharge type category. This invention achieves unified processing of discharge feature extraction, multi-source information fusion, discharge type identification, and high-risk sample retesting, while possessing both high identification accuracy and anti-interference capability.

Description

Technical Field

[0001] This invention belongs to the field of current detection technology based on deep learning, and particularly relates to a method for identifying the discharge type of partial discharge pulse current in switchgear. Background Technology

[0002] Switchgear is used for the collection, distribution, control, and fault isolation of electrical energy in power distribution systems. Its internal busbars, circuit breaker contacts, cable terminals, and insulating supports are subjected to the combined effects of electric fields, thermal stress, mechanical vibration, and environmental humidity over long periods. When air gaps, cracks, contamination, moisture, conductor tips, or loose connections appear in the insulation, partial discharge may occur in localized areas, resulting in incomplete insulation penetration. This phenomenon is called partial discharge. Partial discharge generates short-duration, fast-rising-edge pulse currents. If the discharge type is not identified in time, insulation defects may continue to develop, causing flashovers, phase-to-phase short circuits, or power outages. Therefore, accurate, stable, and reliable identification of partial discharge pulse currents in switchgear is crucial for assessing the insulation condition and for proper operation and maintenance.

[0003] Phase-resolved partial discharge (PD) mapping organizes detection results according to power frequency phase, discharge pulse amplitude, and pulse occurrence frequency, converting a large number of transient discharge pulses into two-dimensional images with phase and amplitude distribution characteristics. Different discharge types typically exhibit different phase clustering intervals, amplitude dispersion, and pulse cluster morphology. For example, corona discharge usually forms a relatively concentrated pulse band in a specific polarity half-cycle, while suspended discharge typically forms discrete high-amplitude pulse clusters in both positive and negative half-cycles, and surface discharge may form a wider phase distribution region as surface charge accumulates. Therefore, phase-resolved PD mapping has become an important data format for identifying partial discharge types in switchgear.

[0004] Currently, the identification of partial discharge pulse current discharge types in switchgear mainly includes the following methods: (1) Identification methods based on phase-resolved partial discharge images or artificial phase statistical features. These methods typically use convolutional neural networks to extract spectral features or classify based on statistical parameters such as the number and amplitude of pulses in different phase intervals. However, ordinary convolution does not consider the periodic continuity between 0 degrees and 360 degrees, which easily severs the discharge features at the phase boundary and makes it difficult to characterize the phase span and amplitude dispersion of the discharge clusters separately; artificial phase statistical features rely on pre-set statistical parameters and phase intervals, and have limited adaptability to changes in acquisition conditions. Therefore, existing methods still do not fully utilize the physical laws of phase period and phase statistical information; (2) Discharge type identification method based on multi-sensor information fusion. This type of method usually directly splices or performs fixed weighted fusion of information from multiple sources such as high-frequency current transformers, transient ground voltage, ultra-high frequency and acoustic emission. However, the data quality of different sensors will change with synchronization status, signal-to-noise ratio, saturation degree, baseline drift, frame loss rate and sensor effective status. Fixed fusion method cannot dynamically adjust the contribution of each information source according to the current sample quality. Low quality or invalid auxiliary signals are easy to interfere with the identification results. Therefore, the existing multi-source identification methods still have insufficient anti-interference ability and adaptability to information source missingness under complex acquisition conditions. (3) Decision-making methods that directly output recognition results based on the maximum category score. This type of method usually takes the category corresponding to the maximum predicted probability as the final result. However, the maximum category probability is difficult to reflect whether the classification evidence is sufficient, nor can it determine whether there is a conflict between different information branches. When the collected data is abnormal, noisy, or the main and auxiliary information are inconsistent, the model may still output a definite category. It lacks an effective rejection mechanism for low-confidence samples and abnormal inputs. Low-confidence results may be directly used for subsequent operation and maintenance, affecting the reliability of the recognition results.

[0005] In summary, while existing methods for identifying partial discharge pulse current discharge types in switchgear can achieve discharge type identification under certain conditions, they generally suffer from problems such as insufficient utilization of phase periodic features, susceptibility of multi-source information fusion to low-quality or missing auxiliary signals, and inadequate reliability assessment of classification results and ability to reject abnormal samples. Summary of the Invention

[0006] To address the above problems, this invention provides a method for identifying the discharge type of partial discharge pulse current in switchgear, comprising the following steps: S1. Simultaneously acquire the pulse waveforms of the target cabinet and adjacent cabinets, and acquire the transient ground voltage signal, UHF signal, and acoustic emission signal of the target cabinet. Perform preprocessing on the pulse waveforms of the target cabinet and adjacent cabinets to generate phase-resolved partial discharge image tensors. At the same time, generate phase statistical vectors according to phase intervals. Perform time-frequency transformation on the transient ground voltage signal, UHF signal, and acoustic emission signal to generate auxiliary time-frequency tensors, and calculate the data quality vector. S2, input the phase-resolved partial discharge image tensor, phase statistics vector, and auxiliary time-frequency tensor into the phase-resolved partial discharge image branch, phase statistics branch, and auxiliary sensing branch respectively to obtain the phase period feature vector, phase statistics feature vector, and auxiliary sensing feature vector. S3 generates quality gating coefficients from the data quality vector, and performs weighted fusion of the three vectors obtained in S2 to form a multi-source fusion feature vector; then, the evidence deep learning classification end generates the total evidence strength and discharge type probability vectors, and determines the category corresponding to the maximum category probability as the current discharge type prediction category.

[0007] Preferably, the process also includes the following: Based on the phase-resolved partial discharge image branch, phase statistics branch, and auxiliary sensing branch, the phase period probability distribution, phase statistical probability distribution, and auxiliary sensing probability distribution are obtained simultaneously. The uncertainty of insufficient evidence and the uncertainty of primary-secondary conflict are calculated by combining the total evidence strength. Then, the classification threshold, uncertainty threshold, and quality index threshold obtained by grid search on the validation set results are combined to make a credibility judgment on the current discharge type prediction category, and output the credibility classification, low credibility rejection or abnormal input rejection and corresponding processing prompts.

[0008] Preferably, S1 specifically includes: First, a high-frequency current transformer is deployed in the target cabinet to construct a pulse channel, and a high-frequency current transformer is deployed in the adjacent cabinet to construct an adjacent cabinet pulse channel. The high-frequency current transformer collects partial discharge pulses and samples the pulse waveforms of the target cabinet and the adjacent cabinet. Simultaneously, a transient ground voltage sensor, an ultra-high frequency sensor, and an acoustic emission sensor are deployed in the target cabinet to collect transient ground voltage signals, ultra-high frequency signals, and acoustic emission signals, respectively, and record the common trigger marker. Second, fixed transmission delay correction and waveform preprocessing are performed on the pulse waveforms of the target cabinet and the adjacent cabinet, and repeated triggers are merged to form a synchronization event, and a phase-resolved partial discharge image tensor is generated. Third, the pulse density and average normalized amplitude of the target cabinet pulse waveform are statistically analyzed according to the phase interval to generate a phase statistical vector. Finally, based on the common trigger marker, time-frequency transformation is performed on the transient ground voltage signal, ultra-high frequency signal, and acoustic emission signal corresponding to the synchronization event to generate an auxiliary time-frequency tensor, and the data quality vector is calculated.

[0009] Preferably, in step S1, firstly, the fixed transmission delay of the target cabinet pulse channel and the adjacent cabinet pulse channel is calibrated and subtracted from the original trigger time. Baseline removal and bandpass filtering are performed on the target cabinet pulse waveform and the adjacent cabinet pulse waveform to obtain the filtered waveform. Secondly, with the trigger time of any valid target cabinet pulse channel as the center, the corresponding pulses of other pulse channels are searched within the synchronization window, and the sampling timestamps corresponding to the trigger times are saved. Thirdly, repeated triggers are merged based on time proximity, polarity compatibility, and cross-channel waveform correlation. When all three conditions are met, a unified event number is assigned to form a synchronization event. Finally, a partial discharge event corresponding to the amplitude of the filtered waveform on the target cabinet pulse channel exceeding the trigger threshold is recorded as the target cabinet pulse. A grayscale phase-resolved partial discharge image is generated based on the power frequency phase mark, normalized amplitude, and occurrence frequency of the target cabinet pulse, and a phase-resolved partial discharge image tensor is composed of a batch size B.

[0010] Preferably, in step S1, firstly, the transient ground voltage signal, the down-converted UHF signal, and the acoustic emission signal corresponding to the synchronization event are subjected to time-frequency transformation, normalization, and stacking to form an auxiliary time-frequency tensor; secondly, the phase-locked confidence, the effective channel ratio of the high-frequency current transformer in the target cabinet, the normalized signal-to-noise ratio, the saturation rate, the baseline drift, the frame drop rate, the effective number of cycles, the effective state of the transient ground voltage, the effective state of the UHF signal, and the effective state of the acoustic emission are calculated and arranged in a fixed order to form a data quality vector.

[0011] Preferably, S2 specifically includes: based on the phase-resolved partial discharge image branch, phase periodic feature vectors and phase periodic probability distributions are obtained through phase periodic filling, convolutional pooling, bidirectional convolution, residual feature extraction, and global average pooling; based on the phase statistics branch, the phase statistics vectors are restored into phase statistics sequences, and phase statistics feature vectors and phase statistics probability distributions are obtained through linear mapping, cyclic position encoding, Transformer encoding, and attention pooling; based on the auxiliary sensing branch, auxiliary sensing feature vectors and auxiliary sensing probability distributions are obtained through depthwise separable convolution, effective source mask pooling, and fully connected mapping.

[0012] Preferably, the auxiliary sensing branch specifically involves: inputting the transient ground voltage signal, UHF signal, and acoustic emission signal channels of the auxiliary time-frequency tensor into a depthwise separable convolutional unit to obtain primary features of the transient ground voltage signal, primary features of the UHF signal, and primary features of the acoustic emission signal, respectively; the depthwise separable convolution first extracts spatial features independently for each input channel, and then uses a 1×1 convolution to complete a lightweight convolution for channel mixing; secondly, effective source mask pooling is used: let The effective states of transient ground voltage signal, UHF signal, and acoustic emission signal are represented respectively. After the feature map of the missing source is set to zero, the effective auxiliary features of transient ground voltage signal, UHF signal, and acoustic emission signal are obtained by averaging the three channel features along the spatial position. Then, the three types of effective auxiliary features are concatenated and mapped to the auxiliary sensing feature vector through a fully connected layer. At the same time, the auxiliary sensing probability distribution is output through the fully connected auxiliary classification head at the end of the branch.

[0013] Preferably, in step S3, the multi-source fusion feature vector is input into the evidence deep learning classification terminal and mapped to a non-negative class evidence vector; the Dirichlet distribution parameters of each class are calculated according to each class component of the class evidence vector, and the total evidence strength is obtained by summing the Dirichlet distribution parameters of each class; finally, a discharge type probability vector is formed according to the proportion of each class Dirichlet distribution parameter to the total evidence strength, the maximum value in the discharge type probability vector is read as the maximum class probability of the sample, and the class corresponding to the maximum class probability is determined as the current discharge type prediction class.

[0014] Preferably, the discharge type probability vector, insufficient evidence uncertainty threshold, and primary-secondary conflict uncertainty threshold for each sample in the validation set are obtained. The discharge type category probability threshold is obtained through grid search with the weighted sum of the macro-average F1 value and the correct reception rate as the objective. The two types of uncertainty thresholds, insufficient evidence uncertainty threshold and primary-secondary conflict uncertainty threshold, are calibrated. Based on the data quality vectors of normal and abnormal samples in the validation set, the effective channel ratio threshold, phase-locked confidence threshold, minimum effective cycle number, and saturation rate threshold for the high-frequency current transformer in the target cabinet are determined. Finally, based on the data quality vector, discharge type probability vector, insufficient evidence uncertainty, and primary-secondary conflict uncertainty of the sample to be tested, the acquisition validity and primary-secondary consistency judgments are performed sequentially. If both judgment conditions are met, the predicted current discharge type category obtained in S3 is taken as the final result; otherwise, a low-confidence rejection or abnormal input rejection is output, along with corresponding engineering processing prompts.

[0015] Preferably, the predicted current discharge type category includes other interference, floating discharge, surface discharge, and corona discharge.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) A phase period multi-branch prediction network is constructed based on the phase physical law of phase-resolved partial discharge. In view of the problem that existing methods do not make full use of the phase period physical law and phase statistical information, this network uses phase period filling to maintain the continuity between 0 degrees and 360 degrees, separates the phase span and amplitude dispersion through bidirectional convolution, and extracts the phase statistical law by cyclic position encoding. At the same time, independent auxiliary branches are set to encode the transient ground voltage, UHF and acoustic emission time-frequency features, so that phase-resolved partial discharge images, phase statistics and multi-sensor observations can be jointly modeled while maintaining the physical meaning, thereby enhancing the ability to characterize the phase distribution, amplitude diffusion and multi-source features of different discharge types, and improving the accuracy and stability of discharge type identification. (2) Construct a quality-gated multi-source fusion network oriented towards channel quality. To address the problem that existing methods lack anti-interference capability and adaptability to missing information sources under complex acquisition conditions, this network generates quality gating coefficients corresponding to the three branches based on the synchronization state, signal-to-noise ratio, saturation level, frame drop rate, and effective status of auxiliary sources of the current sample. It then uses multi-head cross-attention to fuse phase periodic feature vectors, phase statistical feature vectors, and auxiliary sensing feature vectors. Branches with poor quality or invalidity are suppressed. When all auxiliary sources are unavailable, the fusion classification is completed by the phase-resolved partial discharge image branch and the phase statistical branch, thereby reducing the interference of low-quality auxiliary signals on the main criterion and maintaining stable recognition when auxiliary information sources are missing. This improves the anti-interference capability under complex acquisition conditions and the adaptability to missing information sources. (3) Construct a dual-uncertainty evidence-based hierarchical judgment mechanism. In response to the task requirement of constructing a mechanism for rejecting low-confidence samples and abnormal inputs, this mechanism calculates the uncertainty of insufficient evidence based on the total evidence strength, calculates the uncertainty of primary and secondary conflicts based on the difference between the probability distributions of effective branch categories, and forms a prediction uncertainty vector; then, combined with the collection validity threshold, category probability threshold, and uncertainty threshold, it sequentially performs collection validity, category confidence, and primary and secondary consistency judgments, and outputs a credible classification, a low-confidence rejection, or an abnormal input rejection, thereby avoiding the forced classification of samples with insufficient evidence, primary and secondary information conflicts, or collection anomalies, ensuring the credibility of the identification results and the safety of switch cabinet operation and maintenance. Attached Figure Description

[0017] Figure 1 This is a complete flowchart of the overall technical route of the present invention.

[0018] Figure 2 This is a diagram of the branch network structure for phase-resolved partial discharge images in this invention.

[0019] Figure 3 This is a diagram of the phase statistics branch network structure of the present invention.

[0020] Figure 4 This is a diagram of the auxiliary sensing branch network structure of the present invention.

[0021] Figure 5 This is a diagram of the quality-gated multi-source fusion network structure of the present invention.

[0022] Figure 6 This is a phase-resolved partial discharge map corresponding to the suspended discharge type in the self-collected data in this embodiment of the invention.

[0023] Figure 7 The image shows the phase-resolved partial discharge patterns corresponding to the surface discharge types in the self-collected data in this embodiment of the invention.

[0024] Figure 8 This is a training and validation curve of a self-built multi-source dataset in an embodiment of the present invention.

[0025] Figure 9 This is a performance comparison chart of different single-source and multi-source recognition models in the embodiments of the present invention.

[0026] Figure 10 This is a confusion matrix diagram on four types of self-collected PRPD test sets in an embodiment of the present invention.

[0027] Figure 11 The figures show the ablation experiment results of three core innovations in this invention.

[0028] Figure 12 This is a diagram showing the robustness results under noise and auxiliary source absence conditions in an embodiment of the present invention. Detailed Implementation

[0029] This invention provides a method for identifying the discharge type of partial discharge pulse current in switchgear, the overall technical process of which is as follows: Figure 1 As shown: The S1 switchgear features multi-source synchronous acquisition and data characterization. It synchronously acquires high-frequency current transformer sampling waveforms from the target cabinet and adjacent cabinets, as well as transient ground voltage, UHF and acoustic emission signals, and power frequency phase reference signals from the target cabinet. Fixed transmission delay correction, baseline removal, and bandpass filtering are performed on the pulses, and repeated triggers are merged to form synchronization events. A phase-resolved partial discharge image tensor is constructed based on the power frequency phase, normalized amplitude, and frequency of occurrence of the effective pulses from the target cabinet. A phase statistical vector is constructed based on the pulse density and average amplitude within 72 phase intervals. Short-time Fourier transform, normalization, and interpolation are performed on the three types of auxiliary signals corresponding to the synchronization events to form auxiliary time-frequency tensors. A data quality vector is formed based on phase locking, channel effectiveness, signal-to-noise ratio, saturation, drift, frame loss, effective period, and the effective state of auxiliary signals.

[0030] The S2 phase period multi-branch prediction network is constructed as follows: The phase-resolved partial discharge image tensor is input into the phase-resolved partial discharge image branch, and after phase period padding, convolutional pooling, bidirectional convolution, residual feature extraction, and global average pooling, the phase period feature vector and phase period probability distribution are obtained. The phase statistical vector is restored to a phase statistical sequence, and after linear mapping, cyclic position encoding, Transformer encoding, and attention pooling, the phase statistical feature vector and phase statistical probability distribution are obtained. The auxiliary time-frequency tensor is input into the auxiliary sensing branch, and after depthwise separable convolution, effective source mask pooling, and fully connected mapping, the auxiliary sensing feature vector and auxiliary sensing probability distribution are obtained.

[0031] The S3 quality-gated multi-source fusion network is constructed and trained. Data quality vectors are input into two fully connected layers to obtain quality gating coefficients corresponding to three fixed branches. The phase periodic feature vector, phase statistical feature vector, and auxiliary sensing feature vector are weighted to obtain gated phase periodic feature vector, gated phase statistical feature vector, and gated auxiliary sensing feature vector, and the effective branches are determined. Multi-head cross-attention fusion and gated weighted pooling are performed on the three types of gated features to form a multi-source fusion feature vector. This vector is then input into the evidence deep learning classification end to obtain the total evidence strength and discharge type probability vectors, and the category corresponding to the highest category probability is determined as the predicted current discharge type category.

[0032] The S4 dual-uncertainty evidence-layered decision mechanism: First, based on the total evidence strength and the probability distributions of phase period, phase statistics, and auxiliary sensing, the uncertainty of insufficient evidence and the uncertainty of primary-secondary conflict are calculated. Second, the discharge type probability vector, the threshold of insufficient evidence uncertainty, and the threshold of primary-secondary conflict uncertainty for each sample in the validation set are obtained. The discharge type category probability threshold is obtained through grid search, and the two types of uncertainty thresholds, namely the threshold of insufficient evidence uncertainty and the threshold of primary-secondary conflict uncertainty, are calibrated. Based on the data quality vectors of normal and abnormal samples in the validation set, the effective channel ratio threshold, phase-locking confidence threshold, minimum number of effective cycles, and saturation rate threshold of the high-frequency current transformer in the target cabinet are determined. Finally, based on the data quality vector, discharge type probability vector, uncertainty of insufficient evidence, and uncertainty of primary-secondary conflict of the sample to be tested, the acquisition validity and primary-secondary consistency judgments are performed sequentially. When both judgment conditions are met, the predicted current discharge type category obtained in S3 is taken as the final result; otherwise, a low-confidence rejection or abnormal input rejection is output, and corresponding engineering processing prompts are given.

[0033] The present invention will be further described below with reference to embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0034] S1. Synchronous acquisition of multi-source data from switchgear and construction of data representation. S1-1 Self-built multi-source synchronous acquisition platform and experimental batch settings First, high-frequency current transformers (HFCTs) are installed at the three topologically inequivalent grounding points of the target cabinet to construct a pulse channel for the target cabinet. One HFCT is installed in the adjacent cabinet to construct a pulse channel for the adjacent cabinet. The target cabinet pulse channel and the adjacent cabinet pulse channel are collectively referred to as the pulse channel. The pulse channel refers to the data acquisition path formed by the HFCT and the acquisition device. The target cabinet pulse channel acquires the target cabinet pulse waveform, and the adjacent cabinet pulse channel acquires the adjacent cabinet pulse waveform. The target cabinet pulse waveform and the adjacent cabinet pulse waveform are obtained by converting the partial discharge pulse into a voltage signal using the HFCT and then sampling it with a synchronous clock by the acquisition device. Topological inequivalence means that the pulse propagation paths or coupling links corresponding to different installation points are different, providing distinguishable arrival time and attenuation information. Second, a transient earth voltage (TEV) sensor, an ultra-high frequency (UHF) sensor, and an acoustic emission sensor are installed in the target cabinet. (AE) Sensors are used to collect transient ground voltage signals, ultra-high frequency signals, and acoustic emission signals respectively; four high-frequency current transformers share a synchronous clock with 100 MHz sampling points per second and an effective bandwidth of 1 to 40 MHz; synchronous acquisition means that different sensors use the same sampling clock and a time marker that can be converted to the same time base; furthermore, transient ground voltage signals, ultra-high frequency signals, and acoustic emission signals are collected according to their respective native sampling rates, and a common trigger marker is recorded; the acquisition device records the pulse channels of the target cabinet, adjacent cabinets, and the transient ground voltage channel, ultra-high frequency channel, and acoustic emission channel. The system collects timestamps, acquisition frame numbers, and channel status, including whether the channel is enabled, whether synchronization is normal, and whether waveform recording is complete. Simultaneously, it acquires a power frequency phase reference signal. The acquisition device calculates the power frequency phase for each partial discharge pulse based on the power frequency phase reference signal and generates a power frequency phase marker. When the power frequency phase reference signal is acquired normally and the phase calculation is successful, the power frequency phase marker is recorded as a valid power frequency phase marker. Finally, it sets up floating discharge, surface discharge, corona discharge, and other interference states, and divides the system into training, verification, and test sets according to the physical cabinet, defective specimen, and acquisition batch.

[0035] S1-2 High-Frequency Current Transformer Synchronization Events and Phase-Resolved Partial Discharge Image Tensor Construction First, the fixed transmission delay of the target cabinet pulse channel and the adjacent cabinet pulse channels is calibrated and subtracted from the original trigger time. Baseline removal and bandpass filtering are performed on the target cabinet pulse waveform and the adjacent cabinet pulse waveform to obtain the filtered waveform. Second, with the trigger time of any valid target cabinet pulse channel as the center, the corresponding pulses of other pulse channels are searched within the synchronization window, and the sampling timestamps corresponding to the trigger times are saved. Third, repeated triggers are merged based on time proximity, polarity compatibility, and cross-channel waveform correlation. When all three conditions are met, a unified event number is assigned to form a synchronization event. Finally, a partial discharge event corresponding to the amplitude of the filtered waveform on the target cabinet pulse channel exceeding the trigger threshold is recorded as the target cabinet pulse. A 256×256 grayscale phase-resolved partial discharge image is generated based on the power frequency phase mark, normalized amplitude, and occurrence frequency of the target cabinet pulse, and a B×256×256×1 phase-resolved partial discharge image tensor is formed with a batch size B.

[0036] First, the fixed transmission delay of the target cabinet pulse channel and the adjacent cabinet pulse channel is calibrated, i.e., the constant delay of the cable, sensor, and acquisition channel, and subtracted from the original trigger time. Then, using the trigger time of any valid target cabinet pulse channel as the center, the corresponding pulses of other pulse channels are searched within the synchronization window. A valid target cabinet pulse channel refers to a target cabinet pulse channel whose channel status is enabled and synchronization is normal. Triggering refers to the acquisition device generating a pulse detection record when the amplitude of the filtered waveform corresponding to the pulse channel exceeds the trigger threshold, saving the sampling timestamp corresponding to the trigger time and the target cabinet pulse waveform or adjacent cabinet pulse waveform of a preset length before and after it. The filtered waveform refers to the waveform after bandpass filtering of the target cabinet pulse waveform or adjacent cabinet pulse waveform. Second, baseline removal is performed, i.e., the DC bias is estimated using the median of the segment without obvious pulses before triggering and subtracted from the entire target cabinet pulse waveform or adjacent cabinet pulse waveform. Then, bandpass filtering is performed, i.e., only the valid pulse waveforms of the target cabinet and adjacent cabinet pulse waveforms are retained. The signal within the frequency band is measured and low-frequency drift and out-of-band noise are suppressed. Next, repeated triggers are merged based on time proximity, polarity compatibility, and cross-channel waveform correlation. Time proximity indicates whether the trigger time difference after delay correction is less than the synchronization window threshold; polarity compatibility indicates whether the positive and negative directions of the main pulse after channel direction correction are the same or conform to the calibration relationship; cross-channel waveform correlation indicates the similarity between the target cabinet pulse waveform and the pulse waveforms of adjacent cabinets after alignment and normalization. When all three conditions are met, a unified event number is assigned, defined as a synchronization event. Finally, a partial discharge event corresponding to the amplitude of the filtered waveform on the target cabinet pulse channel exceeding the trigger threshold is recorded as the target cabinet pulse. A 256×256 grayscale phase-resolved partial discharge image is generated based on the power frequency phase marker, normalized amplitude, and occurrence frequency of the target cabinet pulse. A B×256×256×1 phase-resolved partial discharge image tensor is formed using a batch size B. The normalized amplitude refers to the normalized amplitude of the peak value of the filtered waveform of the target cabinet pulse.

[0037] S1-3 Phase Statistical Vector Construction First, a preset number of consecutive power frequency cycles are selected as the current phase-resolved partial discharge statistical window. This preset number is denoted as the preset cycle number. The phase-resolved partial discharge statistical window is the continuous sampling time period used to generate a phase-resolved partial discharge image. The phase axis from 0 degrees to 360 degrees is divided into 72 phase intervals. Second, within each phase interval, the proportion of the number of target cabinet pulses to the total number of target cabinet pulses within the current phase-resolved partial discharge statistical window is calculated, and the average normalized amplitude of the pulses falling into that phase interval is calculated. Third, the 72 pulse densities and 72 average amplitudes are concatenated in the same phase order. The relationship is expressed as: ; In the formula Represents the phase density sequence. This represents the average amplitude sequence; finally, the phase statistics vector is formed by stacking the samples in batches of size B to form a B×144.

[0038] S1-4 Construction of Auxiliary Time-Frequency Tensor and Data Quality Vector First, based on the common trigger marker, the transient ground voltage signal, the down-converted UHF signal, and the resampled acoustic emission signal corresponding to the synchronization event are each truncated into 512 points, and a Short-Time Fourier Transform is performed. The STFT yields three types of time-frequency amplitudes; these amplitudes are the amplitudes of the Fourier transform results of each short-time frame, where the two dimensions represent the time position and frequency range of the short-time frame, respectively; the STFT uses a sliding window to convert the waveform into a time-varying spectrum; the down-conversion and resampling are used to unify the UHF signal and acoustic emission signal to the same 512-point length and time scale as the transient ground voltage signal; secondly, the three types of time-frequency amplitudes are normalized and interpolated to 64×64, and stacked in the order of transient ground voltage signal, UHF signal, and acoustic emission signal to form an auxiliary time-frequency tensor of B×64×64×3; thirdly, the phase-locking confidence, effective target cabinet pulse channel ratio, normalized signal-to-noise ratio, saturation rate, baseline drift, frame drop rate, effective cycle count, effective state of transient ground voltage, effective state of UHF, and effective state of acoustic emission are calculated sequentially; among them, the phase-locking confidence is calculated based on the effective power frequency phase mark ratio and the root mean square error of the phase; the effective cycle count is the effective power frequency phase mark ratio and the effective target cabinet pulse channel ratio and the root mean square error of the phase. The number of complete power frequency cycles with complete power frequency phase marking and at least one valid target cabinet pulse channel; the effective target cabinet pulse channel ratio is the ratio of the number of effective target cabinet pulse channels to the total number of target cabinet pulse channels; the normalized signal-to-noise ratio is calculated based on the ratio of the signal energy of the main pulse segment to the energy of the noise segment before triggering; the saturation rate is the ratio of the number of sampling points that reach the upper or lower limit of the acquisition range to the total number of sampling points; the baseline drift is the absolute difference between the median of the baseline of the segment before triggering and the calibrated baseline; the frame loss rate is the ratio of the number of missing frames determined by the acquisition frame sequence number and sampling timestamp to the expected number of acquisition frames; the effective state of transient ground voltage, the effective state of ultra-high frequency, and the effective state of acoustic emission are determined based on whether the channel states of the transient ground voltage channel, ultra-high frequency channel, and acoustic emission channel are synchronized normally, whether the corresponding signals are recorded completely and without severe saturation, and whether the signal-to-noise ratio reaches the preset threshold, respectively, and are set to 1 if the conditions are met, otherwise 0; finally, a B×10 data quality vector is formed according to the above fixed order.

[0039] S2. Construction of Phase Periodic Multi-Branch Prediction Network S2-1 Phase Periodic Filling: First, Circular Phase Padding (CPP) is performed on the phase-resolved partial discharge image tensor. This involves copying the rightmost 8 columns of the original image to the left of the phase axis and the leftmost 8 columns to the right of the original image to maintain continuity between 0 degrees and 360 degrees during convolution calculations. Four rows of reflection padding are used above and below the amplitude axis to maintain its aperiodic boundaries. Second, 32 3×3×1 convolutional kernels are used to perform sliding weighted calculations on adjacent 3-row, 3-column regions, followed by batch normalization, ReLU activation, and 2×2 max pooling, resulting in a B×128×128×32 feature tensor. Third, 64 3×3×32 convolutional kernels are used to extract local information from each channel of the previous layer, and the above processing is repeated to obtain a B×64×64×64 primary image feature tensor. Each channel of this primary tensor collectively represents the local density, phase edges, amplitude gradient, cluster shape, and phase continuity of the pulse cluster.

[0040] S2-2 Bidirectional Convolution and Phase Periodic Feature Output: like Figure 2 As shown, firstly, the primary image feature tensors are fed in parallel into two convolutional branches. The phase direction branch uses 64 convolutional kernels with a spatial size of 1×9 and an input depth of 64, while the amplitude direction branch uses 64 convolutional kernels with a spatial size of 9×1 and an input depth of 64. The bidirectional anisotropic convolution (BAC) uses convolutional kernels extending along the phase direction and the amplitude direction to extract the phase span and amplitude diffusion of the discharge clusters, respectively. The two convolutional branches output feature tensors of size B×64×64×64, which are then concatenated along the channel direction to form a bidirectional feature tensor of size B×64×64×128. Subsequently, 128 1×1 convolutional kernels are used to weightedly combine the channel information of the two branches, and the combined features are extracted through residual blocks, outputting a phase-amplitude weighted feature of size B×64×64×128. Finally, the 64×64 spatial locations of each channel of the phase-amplitude weighted feature are averaged and layer normalization is performed to obtain a B×128 phase periodic feature vector, where each dimension represents the summed response of a feature channel. Simultaneously, at the end of this branch, a B×4 phase periodic probability distribution is output via a fully connected auxiliary classification head for S4-1 to calculate the primary-secondary conflict uncertainty.

[0041] S2-3 Phase Statistical Branching and Cyclic Position Encoding: like Figure 3As shown, firstly, the phase statistics vector is split into a 72-dimensional pulse density sequence and a 72-dimensional average amplitude sequence according to the original splicing rules. Then, the pulse density and average amplitude of the same phase interval are paired to form a B×72×2 phase statistics sequence. Secondly, the same linear layer is used on each of the 72 phase intervals to map the 2-dimensional data at each position to 128 dimensions, forming a B×72×128 phase statistics embedding feature. Thirdly, cyclic positional encoding (CPE) is performed on the phase statistics embedding feature, that is, phase position information is added to each of the 72 128-dimensional feature vectors, and the first phase interval and the 72nd phase interval are regarded as adjacent positions according to the cyclic phase relationship to maintain the continuity of phase features at 0 degrees and 360 degrees. Finally, the phase statistics embedding feature is passed through two Transformer encoding blocks in sequence. Each encoding block contains a 4-head self-attention, residual connection, layer normalization and feedforward network. Finally, attention pooling is performed on the 72 positions of the intermediate features to form a B×128 phase statistical feature vector. At the same time, a B×4 phase statistical probability distribution is output from the end of this branch via a fully connected auxiliary classification head for S4-1 to calculate the primary-secondary conflict uncertainty.

[0042] S2-4 Auxiliary Sensing Branch and Three-Branch Feature Output: like Figure 4 As shown, firstly, the transient ground voltage signal, UHF signal, and acoustic emission signal channels of the auxiliary time-frequency tensor are input into a depthwise separable convolution (DSC) unit to obtain primary features of the transient ground voltage signal, UHF signal, and acoustic emission signal, respectively. The depthwise separable convolution is a lightweight convolution that first extracts spatial features independently for each input channel and then uses a 1×1 convolution to perform channel mixing. Secondly, effective source mask pooling is used: Let... The effective states of transient ground voltage signal, UHF signal, and acoustic emission signal are represented respectively. After zeroing the feature map of the missing source corresponding to the channel, the effective auxiliary features of transient ground voltage signal, UHF signal, and acoustic emission signal are obtained by averaging the primary features of the three channels along their spatial positions. Then, the three types of effective auxiliary features are concatenated and mapped to a B×128 auxiliary sensing feature vector through a fully connected layer. At the same time, a B×4 auxiliary sensing probability distribution is output at the end of this branch through a fully connected auxiliary classification head for S4-1 to calculate the primary-auxiliary conflict uncertainty.

[0043] S3, Construction and Training of Quality-Gated Multi-Source Fusion Network The overall network structure is as follows Figure 5 As shown: S3-1: First, map the data quality vector to quality gating coefficients for three branches; second, weight the phase period feature vector, phase statistical feature vector, and auxiliary sensing feature vector respectively to obtain the gated phase period feature vector, gated phase statistical feature vector, and gated auxiliary sensing feature vector; finally, determine the effective branches by combining the quality gating coefficients and the valid data state.

[0044] S3-1 Sample-level quality gating: First, the data quality vector is input into the first fully connected layer. Ten quality metrics are weighted and combined using trainable weights, mapping each sample to a 32-dimensional quality feature. A ReLU activation function is then used to introduce non-linearity. Second, the 32-dimensional quality feature is input into the second fully connected layer. This layer has three output nodes with independent parameters. Each output node uses independent weights and biases, corresponding to the phase-resolved partial discharge branch, phase statistics branch, and auxiliary sensing branch in a fixed order. Finally, a Sigmoid activation function is used to restrict the three output values ​​to between 0 and 1, obtaining the quality gating coefficients for the three branches. Quality gating is a multiplicative coefficient that weakens or closes the corresponding branch based on the current sample quality; its calculation relationship can be expressed as: ; In the formula, This represents the weight matrix obtained during training. Represents the data quality vector. This represents the bias vector obtained during training. Next, the three gating coefficients are multiplied by the phase periodic feature vector, the phase statistical feature vector, and the auxiliary sensing feature vector, respectively, to obtain the gated phase periodic feature vector, the gated phase statistical feature vector, and the gated auxiliary sensing feature vector. When the quality gate coefficient corresponding to the phase-resolved partial discharge branch or the phase statistics branch is not less than 0.2, and there is an effective target cabinet pulse channel in the current acquisition period, the corresponding branch is determined as an effective branch; when the quality gate coefficient corresponding to the auxiliary sensing branch is not less than 0.2, and at least one of the transient ground voltage effective state, ultra-high frequency effective state and acoustic emission effective state is 1, the auxiliary sensing branch is determined as an effective branch.

[0045] S3-2 Multi-head Cross-Attention Fusion: First, the gated phase periodic feature vector, gated phase statistical feature vector, and gated auxiliary sensing feature vector are stacked into a 3×128 sequence. Using the gated phase periodic feature vector as the query and the gated phase statistical feature vector and gated auxiliary sensing feature vector as the key and value, a 4-head multi-head cross-attention (MHCA) algorithm is used to calculate the branch relationships. MHCA calculates the correlation between the phase periodic feature vector, phase statistical feature vector, and auxiliary sensing feature vector in multiple feature subspaces, with each head having a query, key, and value dimension of 32. Second, a logarithmic bias of the quality gating coefficient is added to the attention score, ensuring that low-quality branches only provide limited correction. Next, the outputs of the four attention heads are concatenated and passed through linear projection, residual connection, layer normalization, and feedforward network. Finally, gated weighted pooling is performed on the three branch positions to form a B×256 multi-source fusion feature vector, where the first 128 dimensions are the attention aggregation result and the last 128 dimensions are the quality-weighted residuals of the three original branches.

[0046] S3-3 Category Evidence and Discharge Type Probability Calculation: First, the multi-source fused feature vector is input into the Evidential Deep Learning (EDL) classification layer. This classification layer employs a fully connected layer with four output nodes, mapping the 256-dimensional fused features of each sample to a 4-dimensional raw classification output. The four output nodes correspond to other interferences, levitating discharges, surface discharges, and corona discharges in a fixed order. Second, Softplus activation is performed component-wise on the 4-dimensional raw classification output to obtain a B×4 non-negative class evidence vector. Its four components represent the strength of supporting evidence for other interferences, levitating discharges, surface discharges, and corona discharges, respectively. The non-negative class evidence for the k-th class is generated according to the following formula: ; In the formula, This represents the original real number output by the classification end for the k-th class. This represents the k-th class of non-negative evidence obtained after Softplus activation. Next, incrementing the k-th class of non-negative evidence by 1 yields the corresponding Dirichlet distribution parameters: ; The sum of the Dirichlet distribution parameters for the four categories is denoted as the total evidence strength S: ; Overall strength of evidence This represents the overall scale of all classification evidence obtained by the model for the current sample. Based on the proportion of each category parameter in the total evidence strength, the 4-dimensional discharge type probability vector is calculated according to the following formula: ; In the formula, This represents the predicted probability of the k-th discharge type. The four components of the discharge type probability vector represent the predicted probabilities of other interferences, floating discharge, surface discharge, and corona discharge, respectively, and the sum of the four components equals 1. The probability with the largest value in the discharge type probability vector is taken as the maximum class probability of the sample, and the class corresponding to the maximum probability is taken as the predicted class of the current discharge type.

[0047] S3-4 Joint Loss and Training Parameters: First, the weighted cross-entropy classification loss is calculated using the discharge type probability vector and the true class label. The weighted cross-entropy classification loss imposes a higher cost on classification errors for classes with fewer samples, thus ensuring consistency between the predicted and true classes. Secondly, Prototype Contrastive Learning (PCL) is used to calculate the prototype contrastive loss. This approach aims to bring the multi-source fusion feature vectors of similar samples closer to their corresponding class prototypes, while keeping features from different classes further apart. Furthermore, the evidence regularization loss is calculated using Kullback-Leibler divergence (KL divergence). This involves constraining the evidence obtained for incorrect classifications to prevent the model from outputting high class evidence even when it misclassifies. Subsequently, the gating consistency loss is calculated. This allows branches with better data quality to receive larger quality gating coefficients, while branches with poorer data quality or missing data receive smaller quality gating coefficients. The joint loss is expressed as: ; Finally, the Adam with Decoupled Weight Decay (AdamW) adaptive moment estimation weight decay optimizer was trained for 50 rounds. If the macro-average F1 score of the validation set did not improve for 8 consecutive rounds, the process was stopped early.

[0048] S4. Double Uncertainty Evidence-Layered Judgment Mechanism S4-1 Calculation of Prediction Uncertainty Vector: First, the total strength of evidence obtained from S3-3. Calculate the uncertainty due to insufficient evidence: ; In the formula, Indicates the uncertainty of insufficient evidence; overall strength of evidence. The larger the value, the more sufficient the classification evidence obtained by the model, and the smaller the uncertainty due to insufficient evidence. Next, read the phase period probability distribution, phase statistical probability distribution, and auxiliary sensing probability distribution corresponding to the effective branches determined in S3-1. When there are at least two effective branches, calculate the average of the pairwise Jensen-Shannon divergences between the category probability distributions of the effective branches, as the primary-auxiliary conflict uncertainty; when there are fewer than two effective branches, set the primary-auxiliary conflict uncertainty to 0.

[0049] S4-2 Classification threshold and uncertainty threshold calibration: First, obtain the discharge type probability vector, insufficient evidence uncertainty threshold, primary-secondary conflict uncertainty threshold, and data quality vector for each sample in the validation set. The probability with the highest value in the discharge type probability vector is taken as the maximum class probability for that sample, and the class corresponding to the highest probability is taken as the predicted class. Second, determine the probability thresholds for four types—other interference, floating discharge, surface discharge, and corona discharge—as well as the insufficient evidence uncertainty threshold and primary-secondary conflict uncertainty threshold, with the objective of a weighted sum of the macro-average F1 score and the correct reception rate. The macro-average F1 score is defined as the arithmetic mean of the F1 scores for each of the four discharge types. For the k-th class, the F1 score is calculated as the harmonic mean of precision and recall. ; In the formula, This represents the proportion of samples that are predicted as positive in class k that actually belong to class k. This represents the proportion of correctly predicted samples in the k-th class of true samples. The macro-average F1 score is expressed as: ; The error reception rate is defined as the proportion of incorrectly received samples to the total number of received samples, i.e.: ; For the kth class To determine the probability thresholds for other disturbances, levitational discharge, surface discharge, and corona discharge on the validation set using a grid search, the uncertainty threshold for insufficient evidence is [value missing]. The uncertainty threshold for conflict between primary and secondary components is: Specifically, the first step is to enumerate candidates: the category probability threshold is incremented by 0.05 between 0.5 and 0.99, and the uncertainty thresholds for insufficient evidence and primary / secondary conflict are incremented by 0.05 between 0 and 1, iterating through all combinations; the second step is to determine acceptance / rejection: a sample is recorded as an accepted sample if it simultaneously satisfies the following conditions: the maximum category probability is not lower than the corresponding category probability threshold, the uncertainty for insufficient evidence is not higher than, and the uncertainty for primary / secondary conflict is not higher than; otherwise, it is recorded as a rejected sample; the third step is to score: calculate the target value on the accepted samples; the fourth step is to select the optimal combination: take the combination with the largest target value as the final threshold combination, fix it, and use it in S4-3. Simultaneously, based on the quality index distribution of normal and abnormal samples in the validation set, the medians of various quality indices for the two types of samples were calculated, and the effective channel ratio threshold of the high-frequency current transformer in the target cabinet was determined using the midpoint of the two medians. Phase-locked confidence threshold Minimum number of effective cycles and saturation threshold The normal sample refers to the sample obtained under the following conditions: all target cabinet pulse channels and auxiliary sensor channels are synchronized normally, phase lock is stable, no additional noise is injected, no saturation or frame loss occurs, and all auxiliary sources are effective. The abnormal sample refers to the sample obtained under abnormal acquisition conditions such as disconnecting at least one auxiliary sensor, injecting additional noise, causing synchronization loss, channel saturation or frame loss.

[0050] S4-3 Trustworthy Classification and Rejection: First, check the effective channel ratio, phase-locked confidence level, effective cycle count, and saturation rate of the target cabinet high-frequency current transformer in the data quality vector; the effective channel ratio of the target cabinet high-frequency current transformer should not be lower than the target cabinet high-frequency current transformer effective channel ratio threshold. The phase-lock confidence level is not lower than the phase-lock confidence threshold. The number of effective cycles shall not be less than the minimum number of effective cycles. And the saturation rate is not higher than the saturation rate threshold. If the condition for valid data collection is met, the system will determine whether the data collection is valid; if any condition is not met, the system will output "abnormal input rejected". Secondly, the maximum category probability of the discharge type probability vector must not be lower than the corresponding category threshold, and the uncertainty due to insufficient evidence must not exceed [a certain threshold]. If the primary-secondary conflict uncertainty is not higher than a certain threshold, proceed to the primary-secondary consistency check; otherwise, output a low-confidence rejection. The current discharge type prediction category of S3-3 is used as the final result; otherwise, a low-confidence rejection is output. Finally, processing prompts are given according to the rejection type: for abnormal input rejection, the engineer is prompted to check the sensor connection, sampling synchronization and trigger settings and then re-acquire; for low-confidence rejection, the engineer is prompted to review the original waveform and multi-source observation data and re-acquire if necessary.

[0051] Experimental Analysis: To verify the feasibility of this invention, an experiment was conducted using a self-built multi-source synchronous acquisition platform for switchgear. The platform deployed three high-frequency current transformers in the target cabinet and one high-frequency current transformer reference channel in the adjacent cabinet, simultaneously acquiring TEV, UHF, AE, and power frequency phase signals. Suspended discharge, surface discharge, corona discharge, and other interference states were generated using replaceable defective test pieces.

[0052] The self-built dataset is organized into four categories with 2100 power frequency windows each, totaling 8400 multi-source samples. Each sample simultaneously includes high-frequency current transformer pulses from the target cabinet and adjacent cabinets, a PRPD image formed by the target cabinet pulses, phase statistics, TEV, UHF, AE time-frequency data, and channel quality records. The dataset is grouped according to physical cabinet, defective test specimens, and independent acquisition batches, with 5880 samples in the training set, 1260 in the validation set, and 1260 in the test set. Each consecutive acquisition batch and its augmented samples are included in only one subset of data to prevent leakage between adjacent windows.

[0053] Four high-frequency current transformers were synchronously sampled at 100 MHz per second, with 128 points before the peak and 384 points after the peak captured per event. TEV, UHF, and AE samples were collected at their native sampling rates and then aligned according to a common trigger marker. Training inputs included a 256×256 PRPD image tensor, a B×144 phase statistics vector, a B×64×64×3 auxiliary time-frequency tensor, and a B×10 data quality vector. The network was trained for 50 epochs with a batch size of 32. The optimizer, data partitioning, class weights, and early stopping rules remained consistent across all comparison models.

[0054] Figure 6 and Figure 7Phase-resolved partial discharge (PDD) maps corresponding to two typical discharge types—suspended discharge and surface discharge—are presented in the self-collected dataset. First, the discharge pulses in the suspended discharge map are mainly concentrated in two phase intervals: approximately 60° to 110° and approximately 240° to 290°, exhibiting an approximately symmetrical distribution across the positive and negative half-cycles, with the amplitude gradually decreasing from the center of the phase interval towards both sides. Second, the discharge pulses in the surface discharge map are mainly concentrated in two phase intervals: approximately 90° to 150° and approximately 290° to 350°, with a wider distribution range and greater amplitude dispersion. In summary, the two types of discharge exhibit distinguishable differences in phase concentration intervals, amplitude distribution ranges, and pulse density distributions. These differences form the data basis for the phase-resolved partial discharge image tensor, providing a basis for learning phase periodic features.

[0055] Figure 8 The training and validation processes are presented, where (a) is the training set loss convergence curve and (b) is the validation set loss convergence curve. Within the first 5 rounds, the training loss decreased from 1.27 to 0.34, and the validation loss decreased from 1.19 to 0.31. After the 18th round, the validation accuracy stabilized above 96%, with a final training accuracy of 98.1% and a validation accuracy of 97.0%. The training and validation curves did not continuously separate, indicating that phase periodic filling, branch random masking, prototype contrast constraint, and weight decay collectively suppressed overfitting.

[0056] like Figure 9 As shown, the combined phase periodic multi-branch prediction network and the quality-uncertainty dual-gated fusion network achieved a test accuracy of 97.2% and a macro-average F1 score of 97.1%. The accuracy of a 4-layer CNN using only the high-frequency current transformer-PRPD image was 94.9%; the accuracy of fusing high-frequency current transformer, TEV, UHF, and AE with fixed weights was 95.8%; the accuracy of a CNN directly splicing multi-source features was 96.1%; and the accuracy of ResNet18 without a phase periodic structure was 95.5%. These results indicate that the performance improvement stems from phase priors, quality-constrained auxiliary fusion, and evidence-based decision-making, rather than simply increasing the number of data channels.

[0057] Figure 10 The confusion matrix of this method on 1260 self-collected test samples is presented. The correct identification numbers for other, floating discharge, surface discharge, and corona discharge are 307, 307, 304, and 307, respectively, totaling 1225, with an accuracy of 97.2%. Residual confusion mainly occurs between low-amplitude surface discharge and corona discharge, and between strong pulse interference and floating discharge; since the number of test samples for the four categories is the same, the overall accuracy is not amplified by class imbalance.

[0058] Fusion experiment results as follows Figure 11As shown, the model's accuracy was 94.9% when using only the high-frequency current transformer-PRPD image branch. Adding phase period padding and bidirectional convolution improved the accuracy to 95.8%, indicating that the feature extraction structure designed for the PRPD phase period characteristics improved discharge mode recognition. Further adding a phase statistics branch increased the accuracy to 96.3%, demonstrating that phase statistics effectively supplemented image features. Adding TEV, UHF, and AE auxiliary sensing branches further improved the accuracy to 96.7%, indicating that multi-source auxiliary information enhanced discharge type recognition. Adding quality gating and dual uncertainty decision further improved the accuracy to 97.0%, demonstrating that dynamically adjusting the contribution of each branch based on sample quality and prediction uncertainty improves the effectiveness of multi-source fusion. Finally, adding evidence-based hierarchical decision and prototype comparison constraints achieved an accuracy of 97.2%. Overall, the gradual introduction of each functional module improved performance, validating the effectiveness of phase period feature learning, multi-source auxiliary information fusion, quality and uncertainty constraints, and the evidence-based decision mechanism.

[0059] The robustness verification results under noise and auxiliary source absence conditions are as follows: Figure 12 As shown, when the signal-to-noise ratio of the high-frequency current transformer decreases from no additional noise to 5 dB, the accuracy of this method decreases from 97.2% to 91.7%, and the accuracy of the high-frequency current transformer-PRPD single-branch CNN decreases from 94.9% to 78.6%; the average uncertainty due to insufficient evidence increases from 0.08 to 0.35. After sequentially shielding TEV, UHF, and AE, the quality gating resets the corresponding auxiliary weights to 0. When all auxiliary sources are missing, the model degenerates into the PRPD main branch, and no case of using zero-value pseudo-auxiliary signals to change the category occurs. After rejecting the 10% of samples with the highest uncertainty, the selective accuracy at 5 dB is 94.3%. Selective accuracy is the accuracy calculated on the remaining received samples after rejecting a predetermined proportion of high-uncertainty samples.

[0060] To illustrate the experimental setup, model performance, module contribution, and robustness of this invention in the task of identifying partial discharge types in switchgear, five sets of experimental data tables are provided. The data in the tables are obtained based on a self-built multi-source synchronous acquisition platform for switchgear. The training set, validation set, and test set are divided according to the physical cabinet, defective specimen, and independent acquisition batch. The same continuous acquisition batch and its enhanced samples are included in only one data subset.

[0061] Table 1 shows the main configuration and data acquisition parameters of the self-built switchgear multi-source synchronous acquisition platform of this invention. The four high-frequency current transformers share a synchronous clock of 100 MHz sampling points per second and an effective bandwidth of 1 to 40 MHz. The auxiliary source is aligned with the power frequency phase through a common trigger mark.

[0062] Table 1. Multi-source synchronous acquisition platform for partial discharge of switchgear and its main parameters

[0063] Table 2 shows a comparison of the recognition performance of different single-source and multi-source models on 1260 test samples. This invention combines a phase-period multi-branch prediction network with a quality-uncertainty dual-gated fusion network, achieving a test accuracy of 97.2% and a macro-average F1 score of 97.1%.

[0064] Table 2 Comparison of discharge type identification performance of different models on the test set

[0065] Table 3 presents the results of the module ablation experiment. After gradually introducing phase periodic filling and bidirectional convolution, phase statistical branching, auxiliary sensing branching, quality gating and double uncertainty decision, as well as evidence-layered decision, the model's test accuracy improved from 94.9% to 97.2%, verifying the positive contribution of each functional module to the recognition performance.

[0066] Table 3 Module Ablation Experiment Results

[0067] Table 4 shows the identification performance and rejection rate under different signal-to-noise ratios of high-frequency current transformers. As the added noise increases, the average uncertainty of insufficient evidence rises from 0.08 to 0.35; after rejecting the 10% of samples with the highest uncertainty, the selectivity accuracy can reach 94.3% under the 5dB condition.

[0068] Table 4. Identification performance and rejection rate under different signal-to-noise ratios of high-frequency current transformers

[0069] Table 5 shows the model degradation performance under the condition of missing auxiliary sources. When TEV, UHF, or AE are missing in sequence, the quality gating resets the corresponding auxiliary weights to 0; when all auxiliary sources are missing, the model automatically degenerates into the PRPD main branch recognition path, and there is no case of using zero-value pseudo-auxiliary signals to change the category.

[0070] Table 5. Model degradation performance under the condition of missing auxiliary sources.

[0071] Experimental results show that the proposed method for identifying the discharge type of partial discharge pulse current in switchgear, through phase periodic feature learning, multi-source quality-constrained fusion, and uncertainty closed-loop application, achieves unified processing of discharge feature extraction, multi-source information fusion, discharge type identification, and high-risk sample retesting. It has high identification accuracy, anti-interference ability, and result reliability, and can provide an algorithmic basis and technical support for intelligent diagnosis and operation and maintenance of partial discharge in switchgear.

[0072] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0073] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for identifying the discharge type of partial discharge pulse current in a switchgear, characterized in that, Includes the following processes: S1. Simultaneously acquire the pulse waveforms of the target cabinet and adjacent cabinets, and acquire the transient ground voltage signal, UHF signal, and acoustic emission signal of the target cabinet. Perform preprocessing on the pulse waveforms of the target cabinet and adjacent cabinets to generate phase-resolved partial discharge image tensors. At the same time, generate phase statistical vectors according to phase intervals. Perform time-frequency transformation on the transient ground voltage signal, UHF signal, and acoustic emission signal to generate auxiliary time-frequency tensors, and calculate the data quality vector. S2, input the phase-resolved partial discharge image tensor, phase statistics vector, and auxiliary time-frequency tensor into the phase-resolved partial discharge image branch, phase statistics branch, and auxiliary sensing branch respectively to obtain the phase period feature vector, phase statistics feature vector, and auxiliary sensing feature vector. S3 generates quality gating coefficients from the data quality vector, and performs weighted fusion of the three vectors obtained in S2 to form a multi-source fusion feature vector; then, the evidence deep learning classification end generates the total evidence strength and discharge type probability vectors, and determines the category corresponding to the maximum category probability as the current discharge type prediction category.

2. The method for identifying the discharge type of partial discharge pulse current in a switchgear as described in claim 1, characterized in that, It also includes the following processes: Based on the phase-resolved partial discharge image branch, phase statistics branch, and auxiliary sensing branch, the phase period probability distribution, phase statistical probability distribution, and auxiliary sensing probability distribution are obtained simultaneously. The uncertainty of insufficient evidence and the uncertainty of primary-secondary conflict are calculated by combining the total evidence strength. Then, the classification threshold, uncertainty threshold, and quality index threshold obtained by grid search on the validation set results are combined to make a credibility judgment on the current discharge type prediction category, and output the credibility classification, low credibility rejection or abnormal input rejection and corresponding processing prompts.

3. The method for identifying the discharge type of partial discharge pulse current in a switchgear as described in claim 1, characterized in that: S1 specifically includes: First, a high-frequency current transformer is deployed in the target cabinet to construct a pulse channel, and a high-frequency current transformer is deployed in the adjacent cabinet to construct an adjacent cabinet pulse channel. The high-frequency current transformer collects partial discharge pulses and samples the pulse waveforms of the target cabinet and the adjacent cabinet. Simultaneously, a transient ground voltage sensor, an ultra-high frequency sensor, and an acoustic emission sensor are deployed in the target cabinet to collect transient ground voltage signals, ultra-high frequency signals, and acoustic emission signals, respectively, and record the common trigger marker. Second, fixed transmission delay correction and waveform preprocessing are performed on the pulse waveforms of the target cabinet and the adjacent cabinet, and repeated triggers are merged to form a synchronization event, and a phase-resolved partial discharge image tensor is generated. Third, the pulse density and average normalized amplitude of the target cabinet pulse waveform are statistically analyzed according to the phase interval to generate a phase statistical vector. Finally, based on the common trigger marker, time-frequency transformation is performed on the transient ground voltage signal, ultra-high frequency signal, and acoustic emission signal corresponding to the synchronization event to generate an auxiliary time-frequency tensor, and the data quality vector is calculated.

4. A method for identifying the discharge type of partial discharge pulse current in a switchgear as described in claim 1 or 3, characterized in that: In S1, firstly, the fixed transmission delay of the target cabinet pulse channel and the adjacent cabinet pulse channel is calibrated and subtracted from the original trigger time. Then, baseline removal and bandpass filtering are performed on the target cabinet pulse waveform and the adjacent cabinet pulse waveform to obtain the filtered waveform. Secondly, taking the trigger time of any valid target cabinet pulse channel as the center, search for the corresponding pulse of other pulse channels within the synchronization window and save the sampling timestamp corresponding to the trigger time; Secondly, based on the time proximity, polarity compatibility, and cross-channel waveform correlation, repeated triggers are merged. When all three conditions are met, a unified event number is assigned to form a synchronous event. Finally, a partial discharge event corresponding to the filtered waveform amplitude exceeding the trigger threshold on the target cabinet pulse channel is recorded as the target cabinet pulse. A grayscale phase-resolved partial discharge image is generated based on the power frequency phase mark, normalized amplitude, and occurrence frequency of the target cabinet pulse, and a phase-resolved partial discharge image tensor is composed of a batch size B.

5. A method for identifying the discharge type of a switchgear partial discharge pulse current as described in claim 1 or 3, characterized in that: In S1, firstly, the transient ground voltage signal, the down-converted ultra-high frequency signal, and the acoustic emission signal corresponding to the synchronization event are subjected to time-frequency transformation, normalization, and stacking to form an auxiliary time-frequency tensor; Secondly, the phase-locked confidence, effective channel ratio of the high-frequency current transformer in the target cabinet, normalized signal-to-noise ratio, saturation rate, baseline drift, frame drop rate, effective cycle number, effective state of transient ground voltage, effective state of ultra-high frequency and effective state of acoustic emission are calculated and arranged into a data quality vector in a fixed order.

6. The method for identifying the discharge type of partial discharge pulse current in a switchgear as described in claim 1, characterized in that: S2 specifically includes: based on the phase-resolved partial discharge image branch, phase periodic feature vectors and phase periodic probability distributions are obtained through phase periodic filling, convolutional pooling, bidirectional convolution, residual feature extraction, and global average pooling; based on the phase statistics branch, the phase statistics vectors are restored into phase statistics sequences, and phase statistics feature vectors and phase statistics probability distributions are obtained through linear mapping, cyclic position encoding, Transformer encoding, and attention pooling; based on the auxiliary sensing branch, auxiliary sensing feature vectors and auxiliary sensing probability distributions are obtained through depthwise separable convolution, effective source mask pooling, and fully connected mapping.

7. The method for identifying the discharge type of partial discharge pulse current in a switchgear as described in claim 6, characterized in that: The auxiliary sensing branch specifically involves: inputting the transient ground voltage signal, UHF signal, and acoustic emission signal channels of the auxiliary time-frequency tensor into a depthwise separable convolutional unit to obtain primary features of the transient ground voltage signal, primary features of the UHF signal, and primary features of the acoustic emission signal, respectively; the depthwise separable convolution first extracts spatial features independently for each input channel, and then uses a 1×1 convolution to complete a lightweight convolution for channel mixing; secondly, effective source mask pooling is used: let The effective states of transient ground voltage signal, UHF signal, and acoustic emission signal are represented respectively. After the feature map of the missing source is set to zero, the effective auxiliary features of transient ground voltage signal, UHF signal, and acoustic emission signal are obtained by averaging the three channel features along the spatial position. Then, the three types of effective auxiliary features are concatenated and mapped to the auxiliary sensing feature vector through a fully connected layer. At the same time, the auxiliary sensing probability distribution is output through the fully connected auxiliary classification head at the end of the branch.

8. The method for identifying the discharge type of partial discharge pulse current in a switchgear as described in claim 1, characterized in that: In step S3, the multi-source fusion feature vector is input into the evidence deep learning classification terminal and mapped to a non-negative class evidence vector. The Dirichlet distribution parameters of each class are calculated based on the class components of the class evidence vector, and the total evidence strength is obtained by summing the Dirichlet distribution parameters of each class. Finally, a discharge type probability vector is formed based on the proportion of the Dirichlet distribution parameters of each class to the total evidence strength. The maximum value in the discharge type probability vector is read as the maximum class probability of the sample, and the class corresponding to the maximum class probability is determined as the predicted class of current discharge type.

9. The method for identifying the discharge type of partial discharge pulse current in a switchgear as described in claim 2, characterized in that: The system obtains the discharge type probability vector, insufficient evidence uncertainty threshold, and primary-secondary conflict uncertainty threshold for each sample in the validation set. It then uses a grid search to obtain the discharge type category probability threshold, targeting the weighted sum of the macro-average F1 value and the correct reception rate. Two types of uncertainty thresholds are calibrated: the insufficient evidence uncertainty threshold and the primary-secondary conflict uncertainty threshold. Based on the data quality vectors of normal and abnormal samples in the validation set, the system determines the effective channel ratio threshold, phase-locked confidence threshold, minimum effective cycle count, and saturation rate threshold for the high-frequency current transformer in the target cabinet. Finally, based on the data quality vector, discharge type probability vector, insufficient evidence uncertainty, and primary-secondary conflict uncertainty of the sample to be tested, the system sequentially performs acquisition validity and primary-secondary consistency judgments. If both judgment conditions are met, the predicted current discharge type category obtained in S3 is used as the final result; otherwise, a low-confidence rejection or abnormal input rejection is output, along with corresponding engineering processing prompts.

10. The method for identifying the discharge type of partial discharge pulse current in a switchgear as described in claim 1, characterized in that: The predicted categories for current discharge types include other interference, floating discharge, surface discharge, and corona discharge.

Images