A partial discharge diagnosis method and system for GIS equipment

Abstract

The application discloses a partial discharge diagnosis method and system of GIS equipment, and relates to the technical field of power equipment fault diagnosis, and solves the problem of insufficient diagnosis precision of the partial discharge of the GIS equipment. The application embodiment aligns the UHF electromagnetic radiation signal and the AE signal by taking the partial discharge event as an anchor point, solves the cross-modal mismatch problem of different signals, introduces a cross-attention mechanism to realize cross-modal fine-grained interaction enhancement, can reduce the mutual pollution of mismatched noise, and improves the discriminability, stability and interpretability of shared representation. Combined with a double-branch private expert model and a gating network, task-level information controlled sharing and dynamic distribution are realized, noise transmission and feature pollution between tasks can be effectively reduced, and the stability and generalization performance of the two-task joint modeling are improved. By using the method, the diagnosis precision of the partial discharge of the GIS equipment can be improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault diagnosis technology, and in particular to a method and system for diagnosing partial discharge in GIS equipment. Background Technology

[0002] Gas-insulated switchgear (GIS), as a fully enclosed switchgear control device, is widely used in substations due to its excellent insulation and arc-extinguishing performance. Its operational reliability is crucial for the safety and stability of the power grid. Due to manufacturing processes and long-term operation, defects such as floating potential components, metal particle contamination, or air bubbles inside insulators may occur within GIS, inducing corona discharge or partial discharge (PD). Under continuous discharge, this can lead to insulation degradation or even breakdown. Since the discharge mechanisms and signal morphologies corresponding to different defects vary significantly, and the discharge intensity and development stage directly affect the equipment's risk level, accurately identifying PD types and assessing their severity is key to developing maintenance strategies and improving the operational reliability of GIS.

[0003] Because single-information diagnostic methods are limited by information loss due to feature compression, the mainstream approach currently employs multi-source information fusion for diagnosis. Existing strategies mostly remain at the feature-level concatenation or decision-level weighting: the former often compresses each modal signal into a small number of statistics before fusion, making it difficult to characterize fine-grained cross-modal correlations; the latter superimposes judgments at the result level, lacking explicit interaction and alignment mechanisms. Both of these lead to insufficient stability and interpretability of the fused output representation, resulting in insufficient accuracy of the diagnostic results.

[0004] Therefore, a method and system for diagnosing partial discharge in GIS equipment is needed. Summary of the Invention

[0005] To address the insufficient accuracy of existing technologies in diagnosing partial discharge in GIS equipment, this invention provides a method and system for diagnosing partial discharge in GIS equipment, which can improve the accuracy of partial discharge diagnosis. The specific technical solution is as follows: In a first aspect, embodiments of this application provide a method for diagnosing partial discharge in a GIS device, comprising: The method acquires the Ultra-High Frequency (UHF) electromagnetic radiation signal and Acoustic Emission (AE) signal inside the GIS device between a first time before the partial discharge event and a second time after the event. Using a cross-attention mechanism, a first feature of the UHF electromagnetic radiation signal and a second feature of the AE signal are fused to obtain a shared representation vector. The UHF electromagnetic radiation signal and the AE signal are input into a private expert model to obtain a first private representation vector and a second private representation vector output by the model. The private expert model includes a first branch and a second branch. The first branch outputs the first private representation vector based on the UHF electromagnetic radiation signal and identifies the type of the partial discharge event. The second branch outputs the second private representation vector based on the AE signal and assesses the severity of the partial discharge event. Using a gating network, the method diagnoses the type and severity of the partial discharge in the GIS device based on the shared representation vector, the first private representation vector, and the second private representation vector.

[0006] Preferably, before fusing the first feature of the UHF electromagnetic radiation signal and the second feature of the AE signal through a cross-attention mechanism to obtain a shared representation vector, the method further includes: extracting features from the UHF electromagnetic radiation signal based on a temporal convolutional network by dilated convolution and residual connections to obtain the first feature; and extracting the second feature based on a convolutional recurrent network; wherein the convolutional subnetwork of the convolutional recurrent network is used to extract the local morphology and energy distribution features of the AE signal, and the recurrent subnetwork of the convolutional recurrent network is used to characterize the temporal dependence and attenuation evolution features of the AE signal during propagation.

[0007] Preferably, the first branch includes the temporal convolutional network, and the second branch includes the recurrent convolutional network.

[0008] Preferably, the gating network includes a first gating network and a second gating network; the method of diagnosing the partial discharge type and severity of the GIS device through the gating network based on the shared representation vector, the first private representation vector, and the second private representation vector includes: constructing a first input vector of the gating network based on the shared representation vector and the first private representation vector; calculating a first fusion weight based on the first input vector; calculating a first fusion vector through the first gating network based on the first fusion weight, the shared representation vector, and the first private representation vector, the first fusion vector indicating the type of the partial discharge event; constructing a second input vector of the gating network based on the shared representation vector and the second private representation vector; calculating a second fusion weight based on the second input vector; calculating a second fusion vector through the second gating network based on the second fusion weight, the shared representation vector, and the second private representation vector, the second fusion vector indicating the severity of the partial discharge event; and inputting the first fusion vector and the second fusion vector into a decoder to obtain the type and severity of the partial discharge event output by the decoder.

[0009] Preferably, the calculation of the first fusion weight based on the first input vector includes: inputting the first input vector into the fully connected layer of the gated network to obtain an intermediate output representation; and performing normalization calculation on the intermediate output representation to obtain the first fusion weight.

[0010] Preferably, before fusing the first feature of the UHF electromagnetic radiation signal and the second feature of the AE signal through the cross-attention mechanism to obtain a shared representation vector, the method further includes: training and optimizing the cross-attention mechanism, the private expert model, and the gating network through a total loss function; wherein the total loss function includes a weighted sum of type recognition loss and severity assessment loss.

[0011] Preferably, the weights of the type identification loss and the severity assessment loss are updated using a gradient conflict-aware Pareto tradeoff strategy.

[0012] Secondly, embodiments of this application provide a partial discharge diagnostic system for GIS equipment, applied to the method described in the first aspect, the system comprising: The acquisition module is used to acquire the ultra-high frequency (UHF) electromagnetic radiation signal and acoustic emission (AE) signal inside the GIS device between the first time before the partial discharge event and the second time after the partial discharge event. The shared representation module is used to fuse the first feature of the UHF electromagnetic radiation signal and the second feature of the AE signal through a cross-attention mechanism to obtain a shared representation vector. A private characterization module is used to input the UHF electromagnetic radiation signal and the AE signal into a private expert model to obtain a first private characterization vector and a second private characterization vector output by the private expert model. The private expert model includes a first branch and a second branch. The first branch is used to output the first private characterization vector based on the UHF electromagnetic radiation signal and to identify the type of the partial discharge event. The second branch is used to output the second private characterization vector based on the AE signal and to assess the severity of the partial discharge event. The diagnostic module is used to diagnose the type and severity of partial discharge of the GIS device through a gated network, based on the shared representation vector, the first private representation vector, and the second private representation vector.

[0013] Thirdly, embodiments of this application provide a computing device, including: a memory for storing a program; and a processor for loading the program to execute the method as described in the first aspect.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the method described in the first aspect.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: By aligning UHF electromagnetic radiation signals and AE signals using partial discharge events as anchor points, the cross-modal mismatch problem of different signals is solved; the introduction of a cross-attention mechanism to achieve fine-grained cross-modal interaction enhancement can reduce mutual contamination of mismatch noise and improve the discriminability, stability, and interpretability of shared representations. Combining a dual-branch private expert model and a gated network enables controlled sharing and dynamic allocation of task-level information, effectively reducing noise transmission and feature contamination between tasks and improving the stability and generalization performance of joint modeling of the two tasks. Using the method of this application, the diagnostic accuracy of partial discharge in GIS equipment can be improved. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0017] Figure 1 A schematic flowchart illustrating a partial discharge diagnosis method for a GIS device provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a partial discharge diagnostic system for a GIS device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computing device provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0020] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0021] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] To address the problem of insufficient diagnostic accuracy of partial discharge in gas-insulated switchgear (GIS) equipment using traditional methods, this invention provides a method and system for diagnosing partial discharge in GIS equipment, which can improve the diagnostic accuracy of partial discharge in GIS equipment.

[0023] Please see Figure 1 , Figure 1 This application provides a flowchart illustrating a partial discharge diagnosis method for a GIS device, which is applied to a computing device. Figure 1 As shown, the method includes: Step 101: The computing device acquires the ultra-high frequency (UHF) electromagnetic radiation signal and acoustic emission (AE) signal inside the GIS device between the first time before the partial discharge event and the second time after the partial discharge event.

[0024] The computing device can be a server or a terminal, specifically a server or terminal within the power system, or a server or terminal located outside the power system and communicating with it. The computing device can acquire relevant data about GIS equipment within the power system via wired or wireless communication.

[0025] The computing device can acquire data collected by various sensors deployed inside, on or near the GIS equipment to detect whether a partial discharge event has occurred in the GIS equipment.

[0026] Specifically, when a partial discharge event occurs in a GIS device, physical quantities such as UHF electromagnetic radiation signals and AE signals are generated. The electromagnetic waves excited by the partial discharge in the 300MHz-3GHz range can be collected by a UHF sensor, and the sensor can capture the electromagnetic radiation signals in this frequency band. The acoustic wave signals of the stress waves generated by the partial discharge propagating through the metal shell of the GIS can also be collected by an AE piezoelectric sensor.

[0027] Preferably, the computing device can use algorithms such as threshold denoising and wavelet transform denoising to remove continuous steady-state noise, such as continuous vibration noise caused by equipment operation; then, using the background noise when there is no discharge as the baseline, the signal collected corresponding to the discharge event is compared with the baseline, and the difference is retained as the denoised UHF electromagnetic radiation signal and AE signal.

[0028] Preferably, the computing device can set the discharge judgment threshold to 3 to 5 times the amplitude of the background noise; if the amplitude of the acquired signal exceeds the discharge judgment threshold N times consecutively, it is determined that the current situation is a suspected partial discharge, where N is an integer greater than 1; then, the time domain characteristics and phase characteristics of the acquired signal are calculated. If the time domain characteristics of the acquired signal are a pulse-type non-steady-state signal, or the phase characteristics show a symmetrical or asymmetrical double-peak discharge phase distribution, then it is determined that the current GIS device has experienced a partial discharge event.

[0029] After a partial discharge event is confirmed, the computing device can use the trigger time of the partial discharge event as a reference to collect the UHF electromagnetic radiation signal and AE signal from the first time before the trigger time to the second time after the trigger time.

[0030] The trigger time for a partial discharge event is the moment when the data first exceeds the discharge determination threshold among N consecutive data points.

[0031] The first and second times can be set based on experience.

[0032] Specifically, the computing device can use a pre-trigger caching function to cache sensor signals collected over a period of time recently, but not write them to storage; once the trigger time is determined, the corresponding signal from the first time before the trigger time to the current time is retrieved from the cache.

[0033] For example, the initial time is 5~20ms, which needs to cover the rising edge of the discharge pulse. The specific time can be adjusted according to the sensor sampling rate. The higher the sampling rate, the shorter the pre-trigger time can be.

[0034] Specifically, a single discharge event may be accompanied by multiple pulses, and it is necessary to continuously collect the signal until it completely decays to the background noise level to ensure data integrity.

[0035] For example, the second time is 20~50ms.

[0036] For example, the total duration from the first time point to the second time point is less than 100ms to avoid data redundancy caused by excessive length.

[0037] In another possible implementation, the computing device can continuously cache or store the collected data, and when the trigger time is determined, the computing device can directly intercept the corresponding UHF electromagnetic radiation signal and AE signal based on the first time and the second time.

[0038] It is understandable that the UHF electromagnetic radiation signal and AE signal acquired by the computing device are aligned based on the acquisition timestamp, and the acquisition time accuracy of the two is consistent. If there is a difference in the timestamp, the corresponding data can be supplemented by interpolation.

[0039] In this embodiment, an event-level alignment strategy is established based on discharge event triggering to map signals of different modes to the same reference to form paired samples, so as to solve the cross-modal mismatch problem under the differences in sampling rate, triggering method and propagation path of different detection methods.

[0040] Step 102: The computing device fuses the first feature of the UHF electromagnetic radiation signal and the second feature of the AE signal through a cross-attention mechanism to obtain a shared representation vector.

[0041] The computing device can construct a parallel modal feature extractor, and construct parallel coding branches for the differences between UHF electromagnetic radiation signals and AE signals in terms of frequency range, propagation mechanism and noise morphology. It uses a feature extraction network adapted to modal characteristics to perform feature extraction and robust representation learning on each input segment, and outputs two modal representation vectors, thereby reducing the information loss caused by traditional feature engineering compression and improving cross-condition robustness.

[0042] Preferably, the computing device can extract features from the UHF electromagnetic radiation signal based on a temporal convolutional network by using dilated convolution and residual connections to obtain the first feature; and extract the second feature based on a convolutional recurrent network.

[0043] UHF electromagnetic radiation signals are characterized by strong temporal sequence, multi-scale features including microsecond-level pulse leading edges and millisecond-level attenuation; computing devices can capture the long temporal sequence and multi-scale features of UHF electromagnetic radiation signals through temporal convolutional networks.

[0044] The computing device can extract the UHF electromagnetic radiation signal of a single discharge event into a fixed-length time series, normalize it, and then input it into a temporal convolutional network.

[0045] For example, a temporal convolutional network includes an input layer, a dilated convolutional layer 1, an activation layer, a dilated convolutional layer 2, a residual connection layer 1, a dilated convolutional layer 3, a residual connection layer 2, and a global average pooling layer. After the input layer acquires the preprocessed UHF signal, the dilated convolutional layer 1 convolves the UHF signal with three kernels at a dilation rate of 1, resulting in three feature maps, each corresponding to a fine-grained feature. The activation layer sets negative feature values ​​to 0, preserving effective features. The dilated convolutional layer 2 convolves the output of the activation layer with three kernels at a dilation rate of 2, capturing medium-scale features and obtaining six feature maps. The residual connection layer 1 adds the original signal from the input layer to the output of the dilated convolutional layer 2 after dimension matching, ensuring that the original features are not lost. The dilated convolutional layer 3 convolves the output of the residual connection layer 1 with three kernels at a dilation rate of 4, capturing long-term temporal features and obtaining twelve feature maps. The residual connection layer 2 adds the output of the residual connection layer 1 to the output of the dilated convolutional layer 3. The global average pooling layer calculates the average of the twelve feature maps after residual fusion, obtaining a 128-dimensional first feature.

[0046] Among them, the AE signal is a stress wave generated by the discharge. It has both the local waveform of a pulse and the characteristics of temporal attenuation and multiple reflections when propagating in the GIS shell. The computing device can analyze it through a convolutional recurrent network.

[0047] Specifically, the convolutional recurrent network includes a convolutional subnetwork and a recurrent subnetwork. The convolutional subnetwork of the convolutional recurrent network is used to extract the local morphology and energy distribution features of the AE signal, and the recurrent subnetwork of the convolutional recurrent network is used to characterize the temporal dependence and attenuation evolution features of the AE signal during propagation.

[0048] For example, the convolutional subnetwork includes an input layer, convolutional layer 1, batch normalization layer, activation layer, convolutional layer 2, and max pooling layer. The input layer receives the preprocessed AE signal; convolutional layer 1 scans the AE signal with 16 convolutional kernels at a stride of 2, and each convolutional kernel outputs a feature map, resulting in 16 feature maps, each corresponding to a local morphological feature; the batch normalization layer and activation layer normalize and activate these 16 feature maps; convolutional layer 2 scans the activation layer output with 8 convolutional kernels at a stride of 2, outputting 32 feature maps, each corresponding to the energy distribution of a time window; the max pooling layer reduces the dimensionality of the 32 feature maps, outputting a feature matrix and retaining the key features with the highest energy.

[0049] For example, the recurrent subnetwork includes a Long Short-Term Memory (LSTM) layer, a Dropout layer, and a fully connected layer. The LSTM layer converts the feature matrix output by the convolutional subnetwork into a format acceptable to the LSTM network, and then performs bidirectional LSTM computation, including capturing the positive temporal dependency from discharge to decay from the 1st time step to the Pth time step; and capturing the inverse dependency from decay to discharge from the Pth time step to the 1st time step; the Dropout layer randomly discards 20% of the features to avoid overfitting; the fully connected layer compresses the temporal features output by the Dropout layer into a P-dimensional second feature, which simultaneously contains local energy features and temporal dependency features.

[0050] Then, the computing device can construct a multi-modal feature interaction fusion representation module, design an interaction fusion structure, retain the two-way representation after interaction enhancement during fusion, provide structured input for the subsequent multi-task expert framework, and avoid interaction loss and noise contamination.

[0051] The interactive fusion structure employs a cross-attention mechanism: using one modal representation as a query term and another modal representation as a key-value pair, it explicitly learns the cross-modal correspondence and complementary enhancement weights for the data of the same partial discharge event; the fusion output retains the interactively enhanced representation to reduce information loss caused by early compression and reduce the risk of mutual contamination of noise introduced by mismatch, thereby improving the stability and interpretability of the fusion representation.

[0052] Specifically, the computing device can perform two rounds of cross-attention calculation using the first feature and the second feature of the same partial discharge event as input. In the first round, the computing device uses the first feature as the query vector and the second feature as the key vector and value vector, and calculates the similarity between the query vector and the key vector; then, based on the similarity, the value vector is weighted to obtain the enhanced feature; then, the enhanced feature and the first feature are residually connected to obtain the enhanced first feature.

[0053] The second round follows the same principle: using the second feature as the query vector and the first feature as the key and value vectors, similar calculations are performed to obtain the enhanced second feature. Then, the enhanced first and enhanced second features are concatenated to obtain the shared representation vector.

[0054] In this embodiment, UHF–AE event-level aligned samples are constructed using discharge events as anchor points, which can solve the cross-modal mismatch problem caused by differences in sampling rate, triggering method, and propagation path among different detection methods. Furthermore, heterogeneous parallel representation learning using temporal convolutional networks and convolutional recurrent networks is employed, and cross-attention is introduced to achieve fine-grained cross-modal interaction enhancement. Compared to existing feature splicing or voting fusion methods, this scheme can explicitly model modal correspondences, reduce mutual contamination by mismatch noise, and improve the discriminativeness, stability, and interpretability of the fused representation.

[0055] Step 103: The computing device inputs the UHF electromagnetic radiation signal and the AE signal into the private expert model to obtain the first private representation vector and the second private representation vector output by the private expert model.

[0056] Among them, computing devices can construct a private expert and task adaptive fusion framework for joint tasks. Based on the private expert model, the task adaptive fusion gating network mechanism is integrated to enable different tasks to learn their specific discriminative information. The shared representation and task private representation are adaptively fused through gating, thereby reducing noise interference and negative transfer caused by multi-task shared modeling and improving the stability and generalization performance of joint modeling.

[0057] The proprietary expert model comprises a first branch and a second branch, each learning specific representation information required for different tasks. The first branch outputs a first proprietary representation vector based on the UHF electromagnetic radiation signal, and is used to identify the type of partial discharge event. The second branch outputs a second proprietary representation vector based on the AE signal, and is used to assess the severity of the partial discharge event.

[0058] The calculation expression for the first branch is: The calculation expression for the second branch is: ,in, This indicates a private expert network for a specific task. Indicates the severity of the task's private expert network. , The aligned UHF electric field radiation signal and AE signal. This is the first private representation vector. This is the second private representation vector.

[0059] Preferably, the first branch includes the temporal convolutional network, and the second branch includes the recurrent convolutional network.

[0060] Model reuse can reduce the need for storage space and the amount of computation required for training and optimization.

[0061] Step 104: The computing device diagnoses the type and severity of partial discharge of the GIS device through the gating network, based on the shared representation vector, the first private representation vector, and the second private representation vector.

[0062] The computing device can construct a task-adaptive fusion gating network mechanism. Based on the different dependencies of various tasks on shared and private information, independent gating fusion networks are constructed for type recognition and severity assessment tasks, respectively, to integrate the shared representation vectors. Adaptively fuse the private expert representation of the corresponding task, and use the fusion result as the input to the decoder of that task.

[0063] Preferably, the computing device can construct a first input vector of the gating network based on the shared representation vector and the first private representation vector; calculate the first fusion weight based on the first input vector; calculate a first fusion vector through the first gating network based on the first fusion weight, the shared representation vector, and the first private representation vector, the first fusion vector being used to indicate the type of the partial discharge event; construct a second input vector of the gating network based on the shared representation vector and the second private representation vector; calculate the second fusion weight based on the second input vector; calculate a second fusion vector through the second gating network based on the second fusion weight, the shared representation vector, and the second private representation vector, the second fusion vector being used to indicate the severity of the partial discharge event; input the first fusion vector and the second fusion vector into a decoder to obtain the type and severity of the partial discharge event output by the decoder.

[0064] For any task The gating fusion process can be represented as: ; ; ; in, Let be the input vector of the gated network. The Hadamard product is used to provide consistency information between shared and private representations; To integrate weights, satisfy ; For the task The fusion vector.

[0065] Preferably, the computing device can use the first input vector The intermediate output representation is obtained by inputting the fully connected layer of this gated network. Then, the intermediate output is represented. conduct The first fusion weight is obtained by normalization calculation. .

[0066] in, and These are the weight matrix and polarization matrix of the fully connected layer of the first gated network.

[0067] The decoders correspond to two different tasks, respectively using , The corresponding output is obtained from the input, namely the type and severity of the partial discharge event. The decoder's calculation expression is: ; in, The type of partial discharge event, This is for the first decoder operation; The severity of a partial discharge event can be categorized into severity levels. This is for the second decoder operation.

[0068] It is understandable that the output of the gating network is a high-dimensional feature representation after multimodal fusion, rather than the final result required by the task. The role of the decoder is to map the abstract features to the specific task space, while solving the problems of feature dimension mismatch and semantic gap.

[0069] Understandably, the decoder is pre-trained using a training set that combines the fused vectors and the classification results.

[0070] By dynamically allocating the contributions of shared and private representations through gating networks, the propagation of cross-task noise in the shared path can be suppressed, and feature contamination caused by mutual constraints between different tasks can be avoided, thereby enhancing the robustness and interpretability of task output.

[0071] In this embodiment, based on the shared representation vector, task-specific expert branches are constructed for type recognition and severity assessment, respectively. The shared representation and the private representation are then fused through a task-adaptive gating network and input into the corresponding decoder, achieving controlled sharing and dynamic allocation of task-level information. Compared to the information fragmentation and deployment overhead caused by independent modeling, or the gradient interference problem easily encountered in traditional shared multi-task structures, this scheme effectively reduces noise propagation and feature contamination between tasks, improving the stability and generalization performance of joint modeling of two tasks.

[0072] Before the method of this application is implemented or put into practical application, the computing device can construct a multi-objective collaborative optimization mechanism for joint tasks, so that the two tasks can achieve stable collaborative convergence under the same network framework, avoid the single-task bias and performance at odds caused by fixed weights, and solve the problems of loss scale difference and gradient conflict in multi-task learning during training.

[0073] Preferably, the computing device can train and optimize the cross-attention mechanism, the private expert model, and the gating network using a total loss function; wherein the total loss function includes a weighted sum of type recognition loss and severity assessment loss.

[0074] The computing device unifies the type recognition loss and severity assessment loss through the total loss function and the multi-objective weight adaptive update strategy to form an objective function for joint training. The joint loss enables end-to-end collaborative optimization of shared representations, task private experts and gating fusion modules.

[0075] Specifically, multi-objective collaborative optimization refers to enabling two tasks to mutually reinforce each other. For example, discharge type features can assist in severity judgment, and vice versa, ultimately achieving better performance for both tasks than training them individually. From acquiring the raw UHF electromagnetic radiation and AE signals to outputting the type and severity of partial discharge events, the parameters of the temporal convolutional network, convolutional recurrent network, cross-attention mechanism, private expert model, and gating network are all synchronously updated through backpropagation of the total loss function, achieving end-to-end collaborative optimization rather than training each module separately.

[0076] The expression for the total loss function includes: ; ; ; ; in, For the total loss, , For the weight of the loss term, For type recognition loss, To assess the severity of the loss, N is the number of samples, i is the sample index, and c is the total number of partial discharge types. To predict the probability that the i-th sample belongs to the c-th class for the model, Let i be the true label of the i-th sample. for loss, Let i be the true severity level of the i-th sample. Let be the severity level predicted by the model for the i-th sample.

[0077] Preferably, the weights of the type identification loss and the severity assessment loss are updated using a gradient conflict-aware Pareto tradeoff strategy.

[0078] By measuring the consistency or conflict of gradient directions across different tasks and updating the weight allocation accordingly, the training process automatically suppresses mutually constraining gradient components and strengthens the optimization contribution of complementary directions, thereby achieving a more stable multi-task collaborative convergence effect.

[0079] To address the convergence instability and performance trade-offs caused by differences in loss scale and conflicting gradient directions during joint training, this solution employs a Pareto weight adaptive update mechanism based on gradient consistency or conflict metrics to dynamically balance type loss and severity loss. Compared to fixed or empirical weight settings, this strategy suppresses mutually constraining gradient components and strengthens the optimization contributions of complementary directions, enabling the model to obtain a more balanced multi-task solution and improving training stability and cross-condition adaptability.

[0080] In this embodiment, the cross-modal mismatch problem of different signals is solved by aligning UHF electromagnetic radiation signals and AE signals using partial discharge events as anchor points. A cross-attention mechanism is introduced to achieve fine-grained cross-modal interaction enhancement, which reduces mutual contamination of mismatch noise and improves the discriminativeness, stability, and interpretability of shared representations. Combining a dual-branch private expert model and a gated network enables controlled sharing and dynamic allocation of task-level information, effectively reducing noise transmission and feature contamination between tasks and improving the stability and generalization performance of joint modeling of the two tasks. Using the method of this application, the diagnostic accuracy of partial discharge in GIS equipment can be improved.

[0081] In this embodiment, by using event-level alignment and cross-attention interactive modeling, the correspondence and complementary information between UHF and AE signals can be explicitly mined, reducing cross-modal mismatch and noise contamination, and improving the stability and interpretability of multi-source fusion representation. Through task-private experts and task-adaptive gating fusion mechanisms, dynamic allocation can be achieved between shared information and task-specific information, suppressing multi-task negative migration and performance at odds, and simultaneously improving the accuracy of type identification and the consistency of severity assessment under complex noise backgrounds and multiple operating conditions. At the same time, a unified framework replaces the independent deployment of multiple models, reducing the complexity of engineering implementation and computational resource overhead, and enhancing the deployment feasibility and generalization ability in online monitoring scenarios.

[0082] The method provided in the embodiments of this application has been described above. The system provided in the embodiments of this application will be described below.

[0083] Please see Figure 2, Figure 2 This is a schematic diagram of the structure of a partial discharge diagnostic system for a GIS device provided in an embodiment of this application, as shown below. Figure 2 As shown, the system 20 includes: The acquisition module 201 is used to acquire the ultra-high frequency (UHF) electromagnetic radiation signal and acoustic emission (AE) signal inside the GIS device between the first time before the partial discharge event and the second time after the partial discharge event. The shared representation module 202 is used to fuse the first feature of the UHF electromagnetic radiation signal and the second feature of the AE signal through a cross-attention mechanism to obtain a shared representation vector. The private characterization module 203 is used to input the UHF electromagnetic radiation signal and the AE signal into a private expert model to obtain a first private characterization vector and a second private characterization vector output by the private expert model. The private expert model includes a first branch and a second branch. The first branch is used to output the first private characterization vector based on the UHF electromagnetic radiation signal and to identify the type of the partial discharge event. The second branch is used to output the second private characterization vector based on the AE signal and to assess the severity of the partial discharge event. The diagnostic module 204 is used to diagnose the type and severity of partial discharge of the GIS device through a gating network, based on the shared representation vector, the first private representation vector, and the second private representation vector.

[0084] Preferably, the system 20 further includes an extraction module 205, which is used to extract features from the UHF electromagnetic radiation signal based on a temporal convolutional network by dilated convolution and residual connections to obtain the first feature; and to extract the second feature based on a convolutional recurrent network; wherein the convolutional subnetwork of the convolutional recurrent network is used to extract the local morphology and energy distribution features of the AE signal, and the recurrent subnetwork of the convolutional recurrent network is used to characterize the temporal dependence and attenuation evolution features of the AE signal during propagation.

[0085] Preferably, the first branch includes the temporal convolutional network, and the second branch includes the recurrent convolutional network.

[0086] Preferably, the diagnostic module 204 is specifically configured to: construct a first input vector of the gating network based on the shared representation vector and the first private representation vector; calculate the first fusion weight based on the first input vector; calculate a first fusion vector through the gating network based on the first fusion weight, the shared representation vector, and the first private representation vector, the first fusion vector being used to indicate the type of the partial discharge event; construct a second input vector of the gating network based on the shared representation vector and the second private representation vector; calculate the second fusion weight based on the second input vector; calculate a second fusion vector through the gating network based on the second fusion weight, the shared representation vector, and the second private representation vector, the first fusion vector being used to indicate the severity of the partial discharge event; and input the first fusion vector and the second fusion vector into the decoder to obtain the type and severity of the partial discharge event output by the decoder.

[0087] Preferably, the diagnostic module 204 is specifically used to input the first input vector into the fully connected layer of the gated network to obtain an intermediate output representation; and to perform normalization calculation on the intermediate output representation to obtain the first fusion weight.

[0088] Preferably, the system 20 further includes a training optimization module 206 for training and optimizing the cross-attention mechanism, the private expert model, and the gating network using a total loss function; wherein the total loss function includes a weighted sum of type recognition loss and severity assessment loss.

[0089] Preferably, the weights of the type identification loss and the severity assessment loss are updated using a gradient conflict-aware Pareto tradeoff strategy.

[0090] The partial discharge diagnostic system for GIS equipment provided in this application can be understood by referring to the relevant content in the foregoing method embodiment section, and will not be repeated here.

[0091] like Figure 3 As shown, Figure 3 This is a schematic diagram of a possible logical structure of a computing device provided in an embodiment of this application. The computing device 30 includes a processor 301, a communication interface 302, a memory 303, and a bus 304. The processor 301, the communication interface 302, and the memory 303 are interconnected via the bus 304. In an embodiment of this application, the processor 301 is used to control and manage the operation of the computing device 30. For example, the processor 301 is used to execute... Figure 1 The steps in the embodiments and / or other processes used in the techniques described herein. Communication interface 302 is used to support communication by computing device 30. Memory 303 is used to store program code and data of computing device 30.

[0092] The processor 301 can be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. The processor can also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a digital signal processor and a microprocessor, etc. The bus 304 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0093] In another embodiment of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the above-described... Figure 1 The method described in the embodiments.

[0094] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0095] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0096] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0097] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0098] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0099] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for diagnosing partial discharge in GIS equipment, characterized in that, The method includes: Acquire the ultra-high frequency (UHF) electromagnetic radiation signal and acoustic emission (AE) signal inside the GIS device between a first time before the partial discharge event and a second time after the partial discharge event; By using a cross-attention mechanism, the first feature of the UHF electromagnetic radiation signal and the second feature of the AE signal are fused to obtain a shared representation vector; The UHF electromagnetic radiation signal and the AE signal are input into a private expert model to obtain a first private representation vector and a second private representation vector output by the private expert model. The private expert model includes a first branch and a second branch. The first branch is used to output the first private representation vector based on the UHF electromagnetic radiation signal and is used to identify the type of the partial discharge event. The second branch is used to output the second private representation vector based on the AE signal and is used to assess the severity of the partial discharge event. The partial discharge type and severity of the GIS equipment are diagnosed through a gating network based on the shared representation vector, the first private representation vector, and the second private representation vector.

2. The method according to claim 1, characterized in that, Before fusing the first feature of the UHF electromagnetic radiation signal and the second feature of the AE signal through a cross-attention mechanism to obtain a shared representation vector, the method further includes: Based on a temporal convolutional network, the first feature is obtained by extracting features from the UHF electromagnetic radiation signal through dilated convolution and residual connection. The second feature is extracted based on a convolutional recurrent network; wherein, the convolutional subnetwork of the convolutional recurrent network is used to extract the local morphology and energy distribution features of the AE signal, and the recurrent subnetwork of the convolutional recurrent network is used to characterize the temporal dependence and attenuation evolution features of the AE signal during propagation.

3. The method according to claim 2, characterized in that, The first branch includes the temporal convolutional network, and the second branch includes the recurrent convolutional network.

4. The method according to claim 1, characterized in that, The gating network includes a first gating network and a second gating network; the process of diagnosing the partial discharge type and severity of the GIS equipment through the gating network, based on the shared representation vector, the first private representation vector, and the second private representation vector, includes: The first input vector of the gated network is constructed based on the shared representation vector and the first private representation vector. The first fusion weight is calculated based on the first input vector; A first fusion vector is calculated using the first gating network based on the first fusion weight, the shared representation vector, and the first private representation vector. The first fusion vector is used to indicate the type of the partial discharge event. The second input vector of the gated network is constructed based on the shared representation vector and the second private representation vector. The second fusion weight is calculated based on the second input vector; Using the second gating network, a second fusion vector is calculated based on the second fusion weight, the shared representation vector, and the second private representation vector. The second fusion vector is used to indicate the severity of the partial discharge event. The first fusion vector and the second fusion vector are input into the decoder to obtain the type and severity of the partial discharge event output by the decoder.

5. The method according to claim 4, characterized in that, The calculation of the first fusion weight based on the first input vector includes: The first input vector is input into the fully connected layer of the gated network to obtain an intermediate output representation; The intermediate output representation is normalized to obtain the first fusion weight.

6. The method according to claim 1, characterized in that, Before fusing the first feature of the UHF electromagnetic radiation signal and the second feature of the AE signal through a cross-attention mechanism to obtain a shared representation vector, the method further includes: The cross-attention mechanism, the private expert model, and the gating network are trained and optimized using a total loss function; wherein the total loss function includes a weighted sum of type recognition loss and severity assessment loss.

7. The method according to claim 6, characterized in that, The weights of the type identification loss and the severity assessment loss are updated using a gradient conflict-aware Pareto tradeoff strategy.

8. A partial discharge diagnostic system for GIS equipment, characterized in that, The system, applied to the method of any one of claims 1-7, comprises: The acquisition module is used to acquire the ultra-high frequency (UHF) electromagnetic radiation signal and acoustic emission (AE) signal inside the GIS device between a first time before the partial discharge event and a second time after the partial discharge event. The shared representation module is used to fuse the first feature of the UHF electromagnetic radiation signal and the second feature of the AE signal through a cross-attention mechanism to obtain a shared representation vector; A private characterization module is used to input the UHF electromagnetic radiation signal and the AE signal into a private expert model to obtain a first private characterization vector and a second private characterization vector output by the private expert model. The private expert model includes a first branch and a second branch. The first branch is used to output the first private characterization vector based on the UHF electromagnetic radiation signal and to identify the type of the partial discharge event. The second branch is used to output the second private characterization vector based on the AE signal and to assess the severity of the partial discharge event. The diagnostic module is used to diagnose the type and severity of partial discharge of the GIS device through a gating network, based on the shared representation vector, the first private representation vector, and the second private representation vector.

9. A computing device, characterized in that, include: Memory, used to store programs; A processor for loading the program to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the method of any one of claims 1-7.

Images