Partial discharge type identification and diagnosis method based on multi-sensor data fusion

Abstract

The present application relates to the technical field of power equipment state monitoring and fault diagnosis, and more particularly to a partial discharge type identification and diagnosis method based on multi-sensor data fusion. Partial discharge signals of power equipment are collected to obtain phase-resolved partial discharge data corresponding to modes; the data is expanded into a one-dimensional vector along the amplitude dimension, and feature mapping is completed through one-dimensional convolution to obtain intermediate features. A one-dimensional convolution module with residual connection and an activation function is used to extract deep features from the intermediate features to obtain deep features corresponding to each mode. The deep features of each mode are spliced along the feature dimension to construct multi-modal features. The multi-modal features are input into a category perception hybrid expert module, a gating network guided by category information adaptively allocates weights of each expert subnetwork, dynamically activates the expert subnetworks related to the input category, and outputs multi-sensor fusion features. Finally, the features are input into an encoder to obtain high-discriminative features with context perception capability.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, and more specifically to a method for partial discharge type identification and diagnosis based on multi-sensor data fusion. Background Technology

[0002] Currently, various sensors such as UHF, transient ground voltage, and acoustic emission are widely used in power systems to monitor partial discharge in switchgear and transformers. Existing methods mainly include (1) Single-mode diagnosis: relying on a single sensor signal, which has poor robustness to complex noise environments. (2) Traditional feature extraction and classification methods: based on statistical features, time-frequency analysis and machine learning, which are insufficient in feature expression capabilities and difficult to capture the complex correlation of multimodal data. (3) Deep learning methods: existing studies have introduced models such as CNN and RNN, but most of them remain at the level of single-mode or simple splicing and fusion, and have failed to effectively solve the problems of alignment and adaptive feature fusion of multi-source heterogeneous signals.

[0003] Tang Ming et al. from Guizhou Luxin Technology and the Electric Power Research Institute proposed a partial discharge identification method based on multi-sensor signals and deep learning. This method acquires partial discharge signals using ultrasonic and pulse current sensors, and converts the one-dimensional time-series signals into two-dimensional time-frequency images using wavelet transform. Subsequently, the time-frequency images from different sensors are channel-stitched and input into a three-layer convolutional neural network for deep feature extraction. Finally, a fully connected layer and a Softmax classifier are used to identify the type of partial discharge. However, this method still has certain limitations in feature representation, modality fusion, and handling of imbalanced samples, making it difficult to fully adapt to the multi-modal partial discharge diagnosis needs under complex operating conditions.

[0004] Therefore, how to provide a method for partial discharge type identification and diagnosis based on multi-sensor data fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a partial discharge type identification and diagnosis method based on multi-sensor data fusion, aiming to solve the problems of traditional models that usually rely on single sensor input, lack an effective heterogeneous multi-source data fusion mechanism, and cannot fully utilize the complementary information between different sensing modalities, resulting in limited representation capabilities; in actual partial discharge monitoring, high-risk discharge samples are often scarce, and the data distribution has obvious imbalance. Traditional supervised learning methods tend to be biased towards the majority class during training, resulting in a significant decrease in the identification performance of the minority class, especially key fault types, reducing the reliability and practicality of the method in engineering applications; although deep models are good at capturing complex patterns, their inherent "black box" characteristics make it difficult to interpret the basis of predictions. In the field of power equipment monitoring, high reliability is crucial, but the lack of interpretability limits its practical deployment in actual engineering scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying and diagnosing partial discharge types based on multi-sensor data fusion includes the following steps: S1. Collect partial discharge signals from power equipment and obtain phase-resolved partial discharge (PRPD) data for the corresponding modes; S2. Unfold the PRPD data of each modality into a one-dimensional vector along the amplitude dimension, and use one-dimensional convolution to map the PRPD data unfolded along the amplitude dimension to obtain intermediate features. S3. A one-dimensional convolutional module with residual connections and LeakyReLU activation function is used to extract deep features from intermediate features to obtain the deep features of the corresponding modality. S4. Concatenate the depth features of each modality along the feature dimension to form a unified multimodal feature; S5. Input the multimodal features into the category-aware hybrid expert CMoE module, and adaptively allocate the weights of each expert sub-network through a gating network guided by category information, dynamically activate the expert sub-networks related to the input category, and output multi-sensor features. S6. Input the multi-sensor features into the Transformer encoder, capture the long-range dependencies in the feature sequence through the self-attention mechanism, realize cross-modal semantic-level feature fusion, and obtain highly discriminative features with context awareness. S7. Input the highly discriminative deep features into the classification head composed of linear layers, and finally output the discharge type.

[0007] Furthermore, S2 specifically includes: The PRPD data of each modality are unfolded into a one-dimensional vector along the amplitude dimension. The one-dimensional vectors of each modality are then input into sub-encoders with the same structure but independent parameters. Local temporal features of each modality are extracted through two consecutive layers of one-dimensional convolution, resulting in intermediate features after multimodal feature mapping. , represented as: , The modality mapper is represented as: , Where BN represents batch normalization, ReLU represents a non-linear activation function, and Conv1D represents a one-dimensional convolution operation.

[0008] Furthermore, S3 specifically includes: Based on modality alignment, preliminary feature extraction is performed on the alignment signal for each modality. The convolutional module uses the LeakyReLU activation function connected to the residual. For each modality... Extracted deep feature representation Represented as: , The feature extractor is represented as: .

[0009] Furthermore, S4 specifically includes: Depth features are independently extracted from AE, TEV, and UHF sensor signals. By fully utilizing the complementary information of the three modalities, these features are stitched together to form a unified multimodal feature set. : , in, This refers to the depth features in the AE sensor signal. Depth features in TEV sensor signals. This refers to the depth characteristics in UHF sensor signals.

[0010] Furthermore, S5 specifically includes: A category-aware expert hybrid mechanism module is constructed. Taking multimodal features as input, it dynamically integrates complementary information from heterogeneous modalities through a gating network guided by category information. This enhances intermodal interaction and improves feature representation quality. The category-aware expert hybrid mechanism module consists of a set of independently structured expert subnetworks. It consists of a learnable gating network, which adaptively assigns weights to each expert based on category information, thereby achieving dynamic category-aware feature aggregation. Let the input feature space be H, and each expert subnetwork map the input features to a shared output space Z. For a given input feature h∈H and its corresponding class label y, the output of the class-aware expert hybrid mechanism module is a weighted sum of the outputs of each expert: , in, For the output of the i-th expert, Gating weights guided by category, and satisfying .

[0011] Furthermore, to prevent the model from over-relying on category information during training and to avoid overfitting caused by direct mapping of labels to output, a partially guided strategy is adopted. Category information is introduced in a controlled manner during the training phase, while maintaining effective model operation even when label information is lacking. Specifically, a label masking mechanism is used, applying a Bernoulli random mask m ~ Bernoulli(p) to each training sample: , Where p∈[0,1] is the probability of retaining the class label. When m=0, the label is masked as a neutral sign -1, which forces the gating network to rely only on the input features for expert routing decisions, thereby improving the robustness of the model inference stage.

[0012] Furthermore, to reduce computational overhead and promote the specialization of expert subnetworks, a Top-K selection mechanism is introduced to sparsify the dense MoE architecture. The category-aware gating function is defined as: , in This is the global average pooling representation of the input feature h. Let W be the category embedding vector, W be the gating weight matrix, and K be the number of experts selected. The Top-K selection operation is defined as follows: , in Let v be the set of indices of the k largest elements in vector v.

[0013] Furthermore, each expert network uses the same architecture but with independent parameters, which can be represented as: .

[0014] Furthermore, S6 specifically includes: The fused feature representation is processed by the Transformer module to perform deep feature interaction and semantic-level fusion. The Transformer module consists of two sub-layers: a self-attention mechanism and a feedforward neural network (FFN). Self-attention sublayers are used to capture the dependencies between all positions in the input sequence, while FFN performs independent transformations on each position of the sublayer to enhance the expressive power of the feature representation. FFN consists of two fully connected layers with a ReLU activation function in between, and residual connections and layer normalization are applied around both sublayers. The output of the attention mechanism is calculated by evaluating the similarity between the query and the key to obtain attention weights, which are then applied to the value vector: , in, The dimension of the key vector is represented by the input features linearly projected into three representations in the self-attention mechanism: query Q, key K, and value V.

[0015] Furthermore, for self-attention, all three matrices come from the same input: , , , in, , , It is a learnable projection matrix.

[0016] Compared to existing technologies, this invention proposes a partial discharge type identification and diagnosis method based on multi-sensor data fusion, specifically designed for heterogeneous multi-sensor data. First, shallow convolutional layers map multi-source sensor inputs to a unified feature space to achieve modality alignment and reduce the complexity of subsequent feature extraction. Then, an improved one-dimensional convolutional module extracts deep and discriminative representations from the aligned features. A category-aware Hybrid Expert (CMoE) mechanism is designed, which dynamically activates specific expert sub-networks based on category relevance. This effectively improves the selectivity and effectiveness of cross-modal information interaction, enhancing interpretability and improving the model's adaptability in imbalanced sample distribution scenarios. An end-to-end multi-sensor data fusion network is proposed, whose structure sequentially includes feature mapping, feature extraction, multi-modal fusion, and temporal modeling. This design significantly improves the model's feature representation ability, generalization performance, and robustness in multimodal classification tasks. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the multi-sensor data fusion network for partial discharge type diagnosis according to the present invention; Figure 2 This is a schematic diagram of the experimental scenario for the switchgear of the present invention; Figure 3 This is a schematic diagram of the model accuracy training curve of the present invention; Figure 4 This is a schematic diagram of the model loss training curve of the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0020] The terms used in this application are explained as follows: AE (Acoustic Emission); TEV (Transient Earth Voltage). UHF (Ultra-High Frequency, extra high frequency); PRPD (Phase-Resolved Partial Discharge). CMoE (Category-aware Mixture of Experts).

[0021] A specific implementation method for partial discharge type identification and diagnosis based on multi-sensor data fusion is provided, such as... Figure 1As shown, three types of sensors are used to monitor the operating status of the switchgear and collect multi-sensor data. First, modality-invariant features related to partial discharge modes are extracted through feature mapping, while also compensating for the insufficient sensitivity of single sensors to certain discharge types. Specifically, the raw data of each modality is projected onto a unified feature space through a shallow convolutional module to achieve consistent representation across modalities. Subsequently, a deep feature is extracted using a convolutional module with residual connections to further enhance the representational power of the aligned features. The feature embeddings of each modality are then concatenated and input into a feature fusion module. By introducing category information into the expert routing process, category-sensitive feature extraction is achieved, and the robustness of the model is improved under imbalanced training conditions. Finally, the fused features are fed into a Transformer encoder to capture long-range dependencies and achieve semantic-level feature fusion across modalities, thereby learning highly discriminative and context-aware feature representations.

[0022] Due to the differences in the physical mechanisms of partial discharge in different insulating media, the signals acquired by acoustic emission, transient ground voltage, and ultra-high frequency sensors exhibit significant differences in physical characteristics. Different sensors have varying sensitivities to specific discharge types, leading to substantial differences in the distribution of features across different modes. If left unaddressed, these differences may introduce distribution bias during training, reducing the generalization and robustness of the features learned by the model. Therefore, designing an effective feature alignment strategy is crucial to mitigate cross-modal inconsistencies, improve the compatibility between multi-sensor data, and ultimately enhance diagnostic performance in practical partial discharge detection.

[0023] Given the strong phase periodicity of phase-resolved partial discharge (PRPD) modes, direct use in image processing can easily introduce interpolation noise and grayscale mapping noise, thus affecting the extraction of key discharge features. To address this issue and achieve fine-grained feature learning, this method expands the PRPD mode into a one-dimensional vector along the amplitude dimension, which can preserve the intrinsic features of the partial discharge signal while reducing noise interference.

[0024] In the multimodal alignment process, one-dimensional convolution is used to map the PRPD data unfolded along the amplitude dimension. Specifically, data from AE, TEV, and UHF sensors are input into three sub-encoders with the same structure but independent parameters. Local temporal features of each modality are extracted through two consecutive layers of one-dimensional convolution. This module acts as a learnable modality mapper, projecting and compressing modal features during training, thereby improving cross-modal semantic alignment capabilities. Intermediate features after multimodal feature mapping are then obtained. It can be represented as: , The modality mapper is represented as: , Where BN represents batch normalization, ReLU represents a non-linear activation function, and Conv1D represents a one-dimensional convolution operation.

[0025] Based on modality alignment, preliminary feature extraction is performed on the alignment signal for each modality. To enhance discriminative power and preserve modality-specific information, the convolutional module employs the LeakyReLU activation function connected to the residual. For modalities... Extracted deep feature representation It can be represented as: , The feature extractor is represented as: .

[0026] This design not only enhances the nonlinear representation capability of the network but also effectively alleviates the gradient vanishing and information degradation problems in deep networks. The resulting modal features, with greater consistency and discriminative power, provide a more robust feature foundation for subsequent multi-sensor fusion and intelligent decision-making.

[0027] Due to the severe class imbalance problem in partial discharge datasets, directly using standard feature extractors often leads to feature representations biased towards the majority class, thus reducing classification performance. To address this issue, we propose a class-aware hybrid expert mechanism (CMoE), which improves the model's sensitivity to the minority class by selectively activating expert subnetworks for different fault modes.

[0028] Depth features were extracted independently from AE, TEV, and UHF sensor signals. To fully utilize the complementary information of the three modalities, these features are concatenated to form a unified representation. : , CMoE module with As input, a gating network guided by category information dynamically integrates complementary information from heterogeneous modalities, thereby enhancing intermodal interaction and improving feature representation quality. This module consists of a set of independently structured expert subnetworks. It also consists of a learnable gating network. The gating network adaptively assigns weights to each expert based on category information, achieving dynamic category-aware feature aggregation.

[0029] Let the input feature space be H, and each expert subnetwork map the input features to a shared output space Z. For a given input feature h∈H and its corresponding class label y, the output of the CMoE module is a weighted sum of the outputs of each expert: , in, For the output of the i-th expert, Gating weights guided by category, and satisfying .

[0030] To prevent the model from over-relying on category information during training and to avoid overfitting caused by direct mapping of labels to output, we adopt a partially guided strategy. This involves introducing category information in a controlled manner during training, while ensuring the model can still function effectively even when label information is lacking. Specifically, this is implemented using a label masking mechanism, applying a Bernoulli random mask m ~ Bernoulli(p) to each training sample: , Where p∈[0,1] is the probability of retaining the class label. When m=0, the label is masked as a neutral sign -1, which forces the gating network to rely only on the input features for expert routing decisions, thereby improving the robustness of the model inference stage.

[0031] To reduce computational overhead and promote specialization of expert subnetworks, a Top-K selection mechanism is introduced to sparsify the dense MoE architecture. The category-aware gating function is defined as: , in This is the global average pooling representation of the input feature h. Let W be the category embedding vector, W be the gating weight matrix, and K be the number of experts selected. The Top-K selection operation is defined as follows: , in Let v be the set of indices of the k largest elements in vector v.

[0032] Each expert network uses the same architecture but has independent parameters, which can be represented as: , The proposed CMoE module integrates label information into the expert routing mechanism, enabling selective activation of experts based on semantic features and category priors. This design encourages each expert to focus on features specific to a particular category, preventing experts associated with a few categories from being overwhelmed by samples from the majority of categories. By promoting expert diversity and targeted expertise, the MoE module enhances the model's robustness when dealing with severely imbalanced PD datasets.

[0033] The multi-sensor features fused by the CMoE module contain rich information on partial discharge modes, thus requiring efficient modeling. Since the Transformer has significant advantages in capturing complex and long-range dependencies in sequence data, the fused feature representation is further processed using the Transformer to achieve deep feature interaction and semantic-level fusion.

[0034] A standard Transformer module consists of two sub-layers: a self-attention mechanism and a feedforward neural network (FFN). The self-attention sub-layer captures dependencies between all positions in the input sequence, while the FFN performs independent transformations on each position of the sub-layer to enhance the expressive power of the feature representation. Typically, the FFN consists of two fully connected layers with a ReLU activation function in between. To stabilize training and accelerate convergence, residual connections and layer normalization are applied around both sub-layers.

[0035] In self-attention mechanisms, input features The data is linearly projected into three representations: query Q, key K, and value V. The output of the attention mechanism is calculated by evaluating the similarity between the query and the key to obtain attention weights, which are then applied to the value vector. , For self-attention, all three matrices come from the same input: , , , in, , , It is a learnable projection matrix. This represents the dimension of the key vector.

[0036] By leveraging a self-attention mechanism, the Transformer effectively captures dependencies within and between fused features, enabling the model to emphasize informational patterns while suppressing redundant or noisy signals. This process helps construct highly discriminative and context-aware feature representations, thereby significantly improving the robustness and accuracy of partial discharge classification tasks.

[0037] To verify the proposed framework, such as Figure 2 As shown, a dedicated partial discharge dataset was collected on-site under typical operating conditions of a high-voltage switchgear. A multifunctional sensor capable of simultaneously acquiring acoustic emission, transient ground voltage, and ultra-high frequency signals was installed on the switchgear enclosure for non-invasive monitoring. This sensing system features a high sampling rate and wide bandwidth, accurately capturing the time-frequency characteristics of partial discharge. Data is transmitted wirelessly for processing, providing a reliable foundation for subsequent feature extraction and intelligent diagnosis.

[0038] The collected partial discharge dataset was classified into four categories: normal, corona discharge, internal discharge, and surface discharge, as shown in Table 1. Due to the sporadic nature of partial discharge events, the dataset exhibits a significant class imbalance: normal samples constitute the majority (2343), corona discharge samples are the fewest (70), while internal discharge and surface discharge samples are relatively balanced (393 and 299, respectively). This imbalance can cause the model to favor the majority class, reducing minority class recognition performance. Therefore, strategies are needed to mitigate the class imbalance and improve the model's ability to learn minority class features.

[0039] Table 1. Types and Quantities of Sampled Signals

[0040] The computing platform used in the experiment was equipped with an NVIDIA GeForce RTX3060Ti GPU, an Intel Core i5-12600KF CPU, and 16GB of memory. The model was implemented using the PyTorch 1.10.2 framework in a Python 3.8.19 environment, and the Adam optimizer was used during training. Other model-related hyperparameters are shown in Table 2.

[0041] Table 2 Hyperparameters for Model Training

[0042] K-fold cross-validation enables repeated training and validation on different subsets of data, thereby reducing evaluation bias and improving the robustness and generalization ability of model selection. To comprehensively evaluate the performance of the proposed model on an imbalanced partial discharge dataset, this study employs a 5-fold cross-validation strategy. Specifically, the training and validation sets (accounting for 70% of the dataset) are equally divided into five subsets. In each iteration, four subsets are used for training, and the remaining subset is used for validation, with the validation set rotating in each round. The remaining 30% of the data serves as an independent test set and is not used for training or validation.

[0043] Figure 3 The solid line in the diagram represents the average performance of the five-fold cross-validation, while the shaded area represents the corresponding variance, reflecting the confidence interval. For example... Figure 3 As shown, the accuracy steadily increases with training. However, in the early stages of training, the accuracy fluctuates significantly, indicating that the model performance is still unstable in the initial stages. Around the 80th training epoch, the accuracy begins to converge gradually, and the confidence interval narrows significantly, indicating that the model's stability has improved. Similarly, Figure 4 The loss curve shown also exhibits a continuous downward trend, further indicating that the model gradually converges throughout the training process.

[0044] While cross-validation demonstrates stable and reliable model performance on both the training and validation sets, this evaluation method may not fully reflect the model's generalization ability on completely unseen data. This is because cross-validation reuses different subsets of the same dataset during training and validation, potentially introducing information leakage or causing the model to overfit to the validation fold. To more rigorously evaluate the model's generalization ability, each fold of training should be further evaluated on an independent test set that was not used during training and validation. The corresponding evaluation results are summarized in Table 3.

[0045] Table 3 Performance metrics of the model under 5x cross-validation

[0046] It can be observed that on the independent test set, the model trained with the fifth fold performed best, with an accuracy of 0.9807 and an F1 score of 0.9769. In contrast, the model trained with the second fold performed worst, with an ACC of 0.9568 and an F1 score of 0.9575. Models trained with the remaining folds generally achieved scores above 0.97 on both metrics. Overall, the performance of the models across all folds remains at a high level, indicating that the proposed model has good robustness and generalization ability.

[0047] This application proposes a multi-stage feature extraction method based on one-dimensional convolution, specifically designed for heterogeneous multi-sensor data. First, shallow convolutional layers map the inputs from multiple sensors to a unified feature space, achieving modality alignment and reducing the complexity of subsequent feature extraction. Then, an improved one-dimensional convolutional module is used to extract deep and discriminative representations from the aligned features.

[0048] This application proposes a category-aware hybrid expert (CMoE) mechanism that dynamically activates specific expert subnetworks based on category relevance. This approach effectively enhances the selectivity and effectiveness of cross-modal information interaction, improving not only interpretability but also the model's adaptability in imbalanced sample distribution scenarios.

[0049] This application proposes an end-to-end multi-sensor data fusion network, the structure of which sequentially includes feature mapping, feature extraction, multimodal fusion, and temporal modeling. This design significantly improves the model's feature representation ability, generalization performance, and robustness in multimodal classification tasks.

[0050] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0051] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying and diagnosing partial discharge types based on multi-sensor data fusion, characterized in that, Includes the following steps: S1. Collect partial discharge signals from power equipment and obtain phase-resolved partial discharge (PRPD) data for the corresponding modes; S2. Unfold the PRPD data of each modality into a one-dimensional vector along the amplitude dimension, and use one-dimensional convolution to map the PRPD data unfolded along the amplitude dimension to obtain intermediate features. S3. A one-dimensional convolutional module with residual connections and LeakyReLU activation function is used to extract deep features from intermediate features to obtain the deep features of the corresponding modality. S4. Concatenate the depth features of each modality along the feature dimension to form a unified multimodal feature; S5. Input the multimodal features into the category-aware hybrid expert CMoE module, and adaptively allocate the weights of each expert sub-network through a gating network guided by category information, dynamically activate the expert sub-networks related to the input category, and output multi-sensor features. S6. Input the multi-sensor features into the Transformer encoder, capture the long-range dependencies in the feature sequence through the self-attention mechanism, realize cross-modal semantic-level feature fusion, and obtain highly discriminative features with context awareness. S7. Input the highly discriminative deep features into the classification head composed of linear layers, and finally output the discharge type.

2. The method for partial discharge type identification and diagnosis based on multi-sensor data fusion according to claim 1, characterized in that, Specifically, S2 is: The PRPD data of each modality are unfolded into a one-dimensional vector along the amplitude dimension. The one-dimensional vectors of each modality are then input into sub-encoders with the same structure but independent parameters. Local temporal features of each modality are extracted through two consecutive layers of one-dimensional convolution, resulting in intermediate features after multimodal feature mapping. , is represented as: ， The modality mapper is represented as: ， Where BN represents batch normalization, ReLU represents a non-linear activation function, and Conv1D represents a one-dimensional convolution operation.

3. The method for partial discharge type identification and diagnosis based on multi-sensor data fusion according to claim 1, characterized in that, Specifically, S3 is: Based on modality alignment, preliminary feature extraction is performed on the alignment signal for each modality. The convolutional module uses the LeakyReLU activation function connected to the residual. For each modality... Extracted deep feature representation Represented as: ， The feature extractor is represented as: 。 4. The method for partial discharge type identification and diagnosis based on multi-sensor data fusion according to claim 1, characterized in that, Specifically, S4 is: Depth features are independently extracted from AE, TEV, and UHF sensor signals. By fully utilizing the complementary information of the three modalities, these features are stitched together to form a unified multimodal feature set. : ， in, This refers to the depth features in the AE sensor signal. Depth features in TEV sensor signals. This refers to the depth characteristics in UHF sensor signals.

5. The method for partial discharge type identification and diagnosis based on multi-sensor data fusion according to claim 1, characterized in that, Specifically, S5 is: A category-aware expert hybrid mechanism module is constructed. Taking multimodal features as input, it dynamically integrates complementary information from heterogeneous modalities through a gating network guided by category information. This enhances intermodal interaction and improves feature representation quality. The category-aware expert hybrid mechanism module consists of a set of independently structured expert subnetworks. It consists of a learnable gating network, which adaptively assigns weights to each expert based on category information, thereby achieving dynamic category-aware feature aggregation. Let the input feature space be H, and each expert subnetwork map the input features to a shared output space Z. For a given input feature h∈H and its corresponding class label y, the output of the class-aware expert hybrid mechanism module is a weighted sum of the outputs of each expert: ， in, For the output of the i-th expert, Gating weights guided by category, and satisfying .

6. The method for partial discharge type identification and diagnosis based on multi-sensor data fusion according to claim 5, characterized in that, To prevent the model from over-relying on category information during training and to avoid overfitting caused by direct mapping of labels to output, a partially guided strategy is adopted. Category information is introduced in a controlled manner during the training phase, while maintaining effective model operation even when label information is lacking. Specifically, a label masking mechanism is used, applying a Bernoulli random mask m ~ Bernoulli(p) to each training sample: ， Where p∈[0,1] is the probability of retaining the class label. When m=0, the label is masked as a neutral sign -1, which forces the gating network to rely only on the input features for expert routing decisions, thereby improving the robustness of the model inference stage.

7. The method for partial discharge type identification and diagnosis based on multi-sensor data fusion according to claim 5, characterized in that, To reduce computational overhead and promote the specialization of expert subnetworks, a Top-K selection mechanism is introduced to sparsify the dense MoE architecture. The category-aware gating function is defined as: ， in This is the global average pooling representation of the input feature h. Let W be the category embedding vector, W be the gating weight matrix, and K be the number of experts selected. The Top-K selection operation is defined as follows: ， in Let v be the set of indices of the k largest elements in vector v.

8. The method for partial discharge type identification and diagnosis based on multi-sensor data fusion according to claim 5, characterized in that, Each expert network uses the same architecture but has independent parameters, which can be represented as: 。 9. The method for partial discharge type identification and diagnosis based on multi-sensor data fusion according to claim 1, characterized in that, Specifically, S6 is: The fused feature representation is processed by the Transformer module to perform deep feature interaction and semantic-level fusion. The Transformer module consists of two sub-layers: a self-attention mechanism and a feedforward neural network (FFN). Self-attention sublayers are used to capture the dependencies between all positions in the input sequence, while FFN performs independent transformations on each position of the sublayer to enhance the expressive power of the feature representation. FFN consists of two fully connected layers with a ReLU activation function in between, and residual connections and layer normalization are applied around both sublayers. The output of the attention mechanism is calculated by evaluating the similarity between the query and the key to obtain attention weights, which are then applied to the value vector: ， in, The dimension of the key vector is represented by the input features linearly projected into three representations in the self-attention mechanism: query Q, key K, and value V.

10. The method for partial discharge type identification and diagnosis based on multi-sensor data fusion according to claim 9, characterized in that, For self-attention, all three matrices come from the same input: ， ， ， in, , , It is a learnable projection matrix.

Images