Hybrid sensing-based intelligent positioning and diagnosis method for optical fiber composite fault of power distribution network

Abstract

The invention provides a power distribution network optical fiber composite fault intelligent positioning and diagnosis method based on hybrid sensing, and relates to the technical field of power distribution networks, and the method comprises the steps: obtaining and preprocessing optical fiber and electrical sensing data, extracting time-frequency domain features through a multi-channel convolutional neural network, calculating a dynamic weight coefficient through an attention mechanism, and carrying out the feature fusion, a dynamic fault propagation graph is constructed, spatial-temporal feature extraction is performed by combining graph convolution and causal convolution, and fault positioning and type identification are realized. The method can make full use of multi-source heterogeneous data, improves the fault positioning precision and diagnosis reliability, and is suitable for fault diagnosis of a complex power distribution network.

Description

Technical Field

[0001] The invention relates to distribution network technology, and in particular to a hybrid sensing-based intelligent positioning and diagnosis method for optical fiber composite faults in distribution networks. Background Art

[0002] The distribution network is an important part of the power system, and its safe and stable operation is directly related to the reliable power supply of users. With the continuous expansion of the scale of the distribution network and the improvement of its intelligence level, the rapid and accurate positioning and diagnosis of distribution network faults have become key technologies to ensure the reliability of power supply. At present, the fault location of the distribution network mainly relies on the analysis of voltage, current and other data collected by electrical sensors. At the same time, optical fiber sensing technology has been widely used in distribution network monitoring due to its advantages such as anti-electromagnetic interference and distributed measurement.

[0003] The existing distribution network fault location and diagnosis methods have the following main shortcomings: First, traditional methods often rely solely on electrical sensor data or optical fiber sensor data for fault analysis, and fail to fully utilize the rich information contained in multi-source heterogeneous data, resulting in limited fault location accuracy and reliability. Second, existing data fusion methods mostly use simple feature splicing or fixed weight fusion, which cannot effectively adapt to the contribution differences of various types of sensor data under different fault types, reducing the accuracy of fault diagnosis. Third, traditional fault location methods rarely consider the complex relationship between the physical topology, electrical characteristics and time series evolution in the distribution network, making it difficult to accurately characterize the fault propagation characteristics and the scope of fault impact, affecting the accuracy and real-time performance of fault location.

[0004] With the development of artificial intelligence technology, deep learning methods have shown good application prospects in the field of fault diagnosis. However, how to effectively integrate multi-source heterogeneous data, adaptively extract fault features, and accurately model fault propagation mechanisms are still key scientific issues that need to be solved. This puts forward an urgent need for the development of new intelligent location and diagnosis methods for distribution network faults. Summary of the invention

[0005] The embodiment of the present invention provides a hybrid sensing-based intelligent positioning and diagnosis method for optical fiber composite faults in a distribution network, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention, Provides a hybrid sensing-based intelligent location and diagnosis method for optical fiber composite faults in distribution networks, including: Obtain the optical fiber sensor data and electrical sensor data of each monitoring node in the distribution network, perform data preprocessing on the optical fiber sensor data and electrical sensor data, remove outliers, perform normalization processing, and align time series to obtain a standardized sensor data sequence; The standardized sensor data sequence is input into the corresponding input channel of the pre-trained multi-channel convolutional neural network according to the data type. The time-frequency domain features are extracted through the convolution layer. The dynamic weight coefficients of different sensor data are calculated using the attention mechanism. The features of different types of sensor data are adaptively weighted and fused according to the dynamic weight coefficients to obtain the fused feature vector. A fault feature space is constructed based on the fused feature vector, and the fault feature space is input into the fault diagnosis model that has been pre-trained with historical fault data. A dynamic fault propagation graph containing physical, electrical and timing associations is constructed, and spatiotemporal feature extraction is performed by combining graph convolution and causal convolution. The attention mechanism is used to pass messages to update the node status, and the local features and global features are fused. After optimization with a multi-task loss function, the fault location results and fault type identification results are output.

[0007] The standardized sensor data sequence is input into the corresponding input channel of the pre-trained multi-channel convolutional neural network according to the data type. The time-frequency domain features are extracted through the convolution layer. The dynamic weight coefficients of different sensor data are calculated using the attention mechanism. The features of different types of sensor data are adaptively weighted and fused according to the dynamic weight coefficients. The fused feature vectors include: The standardized sensor data sequence is input into the corresponding feature extraction channel according to the data type. The feature extraction channel includes a one-dimensional convolutional network channel for one-dimensional data, a two-dimensional convolutional network channel for two-dimensional data, and a dilated convolutional network channel for long-range dependency modeling. The features of each type of data are extracted through the feature extraction channel to obtain an initial feature map. Perform multi-scale feature extraction on the initial feature map, build a feature pyramid network with multiple levels, extract features through convolution operations at each level, combine the first-level features with the second-level features through upsampling, and use adaptive weight coefficients to perform weighted aggregation on features at different levels to obtain hierarchically enhanced multi-scale features; The hierarchically enhanced multi-scale features are subjected to dual attention enhancement. In the channel dimension, the channel attention weights are calculated by combining the average pooling features and the maximum pooling features and the feature channels are recalibrated. In the spatial dimension, the spatial attention weight matrix is ​​calculated based on the local region correlation of the feature map and the feature enhancement in the spatial dimension is realized. The spatial enhanced features are residually connected with the original features to obtain the dual enhanced feature representation. Perform global average pooling on the dual enhanced feature representation to obtain channel-level feature description and construct the inter-modal correlation matrix. The correlation matrix is ​​input into the multi-head attention module and the weight generation network respectively, and the dynamic fusion weight coefficient is calculated. The local enhanced features and the global context vector are combined for adaptive fusion to achieve adaptive fusion of multi-modal features. A joint optimization objective function including the cross entropy loss term, the reconstruction error term, and the regularization term is constructed. The parameters of the feature extraction network, the attention module, and the weight generation network are optimized end-to-end by minimizing the joint optimization objective function, and the final fused feature representation is output.

[0008] The hierarchically enhanced multi-scale features are subjected to dual attention enhancement. The channel attention weights are calculated by combining the average pooling and maximum pooling features in the channel dimension. The feature channels are recalibrated. The spatial attention map is calculated based on the local area correlation of the feature map in the spatial dimension to achieve feature enhancement in the spatial dimension. The dual enhanced feature representation includes: Perform global average pooling and global maximum pooling operations on the hierarchically enhanced multi-scale features in the channel dimension to obtain the average pooling feature vector and the maximum pooling feature vector in the channel dimension, and input the average pooling feature vector and the maximum pooling feature vector into the multi-layer perceptron with shared weights for nonlinear transformation to obtain two intermediate feature vectors. The two intermediate feature vectors are added and processed by the sigmoid function to obtain the channel attention weight vector. The channel attention weight vector is used to recalibrate the importance of the channel dimension of the hierarchically enhanced multi-scale features to obtain the channel enhanced features. In the spatial dimension, the channel enhancement features are divided into a preset number of local feature regions, the correlation between different local feature regions is calculated to obtain the regional correlation matrix, the regional correlation matrix is ​​convolved and normalized by the softmax function to obtain the spatial attention weight matrix; The spatial attention weight matrix and the channel enhancement features are weightedly fused in the spatial dimension to achieve feature enhancement in the spatial dimension and obtain spatial enhancement features. The spatial enhancement features are residually connected with the hierarchically enhanced multi-scale features to obtain a doubly enhanced feature representation.

[0009] Perform global average pooling on the dual enhanced feature representation to obtain channel-level feature description and construct the inter-modal correlation matrix. Input the correlation matrix into the multi-head attention module and weight generation network respectively, calculate the dynamic fusion weight coefficient, combine the local enhanced features and the global context vector for adaptive fusion, and realize the adaptive fusion of multi-modal features, including: Perform global average pooling operation on the dual enhanced features to obtain a channel-level feature description vector. Based on the channel-level feature description vector, calculate the correlation coefficient between different modal features and construct an inter-modal correlation matrix. The channel-level feature description vector is mapped into query vector, key vector and value vector respectively through a learnable projection matrix, and the query vector, key vector and value vector are input into the multi-head attention module. The attention weight is calculated in each attention head and multiplied with the value vector to obtain the attention output. The outputs of multiple attention heads are concatenated and linearly transformed to obtain the modal interaction features. The inter-modal correlation matrix and the modal interaction features are concatenated and input into a multi-layer perceptron for nonlinear transformation to obtain the initial fusion weights. The learnable temperature parameter is introduced to scale the initial fusion weights and normalized through the softmax function to obtain the dynamic fusion weights. The dynamic fusion weights are used to perform weighted summation of the dual enhancement features of different modalities to obtain the local enhancement features, and the local enhancement features are globally averaged pooled to obtain the global context vector. The global context vector is broadcast multiplied with the dual enhancement features of the corresponding modality. Calculate the importance weight of each modality, multiply the importance weight with the product of the local enhanced feature and the global context vector, and sum them up; Constructing feature discrimination loss function, feature reconstruction loss function and weight regularization loss function, and weighted combining the feature discrimination loss function, feature reconstruction loss function and weight regularization loss function through a dynamic trade-off factor adjusted by cosine to obtain an overall loss function; Based on the overall loss function, the projection matrix, multi-layer perceptron, temperature parameters and modal importance weights are jointly optimized to achieve adaptive fusion of multimodal features.

[0010] Construct a dynamic fault propagation graph containing physical, electrical and timing associations, combine graph convolution and causal convolution to extract spatiotemporal features, use attention mechanism to pass messages to update node status, fuse local features and global features, and output fault location results and fault type identification results after multi-task loss function optimization, including: A physical topology association matrix is ​​constructed based on the connection relationship of physical nodes in the distribution network, an electrical coupling relationship matrix is ​​constructed based on node voltage, line current and power parameters, and a temporal dynamic association matrix is ​​constructed based on the historical state sequence within the time window. The physical topology association matrix, electrical coupling relationship matrix and temporal dynamic association matrix are fused through adaptive weights to construct a dynamic fault propagation graph. The dynamic fault propagation graph is subjected to graph convolution operation to extract the spatial correlation features between nodes. The spatial correlation features are subjected to causal convolution operation to extract the temporal evolution features. The spatial correlation features and the temporal evolution features are fused through adaptive weight coefficients to obtain the spatiotemporal joint features. The spatiotemporal joint features are mapped into query vectors, key vectors, and value vectors. The dot product similarity between the query vector and the key vector is calculated to obtain the attention weight. The attention weight is multiplied by the value vector and message aggregation is performed. The gated recurrent unit is used to update the node status to achieve dynamic transmission of fault information. Perform feature pooling within the neighborhood of each node to extract local feature representation, perform global pooling on all nodes in the dynamic fault propagation graph to extract global feature representation, and adaptively fuse local feature representation and global feature representation through learnable weights; The fault location probability and fault type probability are calculated based on the fused features, and a multi-task loss function including fault location loss term and fault classification loss term is constructed. The multi-task loss function is optimized using a dynamic weight strategy, and the fault location results and fault type identification results are output.

[0011] The physical topology association matrix is ​​constructed based on the connection relationship of the physical nodes of the distribution network, the electrical coupling relationship matrix is ​​constructed based on the node voltage, line current and power parameters, and the time series dynamic association matrix is ​​constructed based on the historical state sequence in the time window. The physical topology association matrix, the electrical coupling relationship matrix and the time series dynamic association matrix are fused through adaptive weights to construct a dynamic fault propagation graph, including: Obtain impedance data of connection paths between nodes in the distribution network and node electrical level data, calculate the electrical distance between nodes based on the impedance data, calculate the level difference between nodes based on the electrical level data, and fuse the electrical distance between nodes and the level difference between nodes through adaptive weights to generate a physical topology association matrix; Collect the voltage amplitude, phase angle data, active power data and reactive power data of the distribution network nodes, calculate the voltage coupling strength between nodes based on the voltage amplitude and phase angle data, calculate the line power flow coupling degree based on the active power data and reactive power data, and generate an electrical coupling matrix by fusing the voltage coupling strength between nodes and the line power flow coupling degree through dynamic weights; Construct the historical state sequence of the distribution network nodes, calculate the time series correlation coefficient of the historical state sequence to obtain the time series correlation matrix, calculate the time change rate of the time series correlation matrix to obtain the dynamic change characteristics, and fuse the time series correlation matrix and the dynamic change characteristics through adaptive weights to generate a time series feature matrix; The physical topological association matrix, electrical coupling matrix and timing feature matrix are transformed into a normalized feature matrix through multiple layers of nonlinear transformation. Based on the normalized feature matrix, self-attention feature representation and modal interaction features are constructed in multiple attention subspaces. The self-attention features and modal features are fused by multi-scale pooling and updated through dynamic gating to obtain the enhanced modal feature representation. The enhanced modal features are adaptively weighted and fused to construct a dynamic fault propagation graph.

[0012] The physical topological association matrix, electrical coupling matrix and time series feature matrix are transformed into a normalized feature matrix through multiple layers of nonlinear transformation. Based on the normalized feature matrix, self-attention feature representation and modal interaction features are constructed in multiple attention subspaces. The self-attention features and modal features are fused by multi-scale pooling and updated through dynamic gating to obtain the enhanced modal feature representation including: The physical topology correlation matrix, the electrical coupling matrix and the time series characteristic matrix are respectively subjected to characteristic mapping through multi-layer nonlinear transformation to obtain an initial characteristic matrix, and the initial characteristic matrix is ​​subjected to mean normalization to obtain a normalized characteristic matrix; Based on the normalized feature matrix, the query matrix, key matrix and value matrix are constructed, and the query matrix, key matrix and value matrix are projected into multiple attention subspaces respectively. The self-attention weight is calculated in each attention subspace, and the self-attention weight is weighted with the attention mask matrix to obtain the internal feature representation of the modality; Calculate the modal attention weights between different modal features, perform weighted aggregation on the value matrices of different modalities based on the modal attention weights, and obtain the modal interaction features through the learnable output mapping matrix; The modality internal feature representation and the modality interaction feature are residually connected and normalized, and the normalized features are multi-scale pooled to construct a feature pyramid. The feature pyramid is adaptively fused through learnable scale weights to obtain a multi-scale feature representation. The importance scores of the multi-scale feature representation and the normalized feature matrix are calculated, and the multi-scale feature representation is dynamically gated and updated based on the importance scores to obtain the enhanced modal feature representation.

[0013] According to a second aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call instructions stored in the memory to execute the aforementioned method.

[0014] According to a third aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0015] The beneficial effects of this application are as follows: By performing standardized preprocessing and time series alignment on optical fiber sensing data and electrical sensing data, the quality and comparability of different types of sensing data are improved, laying a reliable data foundation for subsequent feature extraction and fusion.

[0016] A multi-channel convolutional neural network and attention mechanism are used to perform adaptive feature extraction and weighted fusion of different types of sensor data, which fully utilizes the complementary advantages of various types of sensor data and improves the comprehensiveness and accuracy of feature expression.

[0017] Based on the spatiotemporal feature extraction and multi-task learning method of dynamic fault propagation graph, the joint identification of fault location and type is realized. The physical laws and timing characteristics of fault propagation are captured through the combination of graph convolution and causal convolution, which improves the interpretability and accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of a method for intelligently locating and diagnosing a composite optical fiber fault in a distribution network based on hybrid sensing according to an embodiment of the present invention; Figure 2 This is a schematic diagram of comparing the accuracy of different data sets in an embodiment of the present invention; Figure 3 Schematic diagram of different multimodal fusion methods according to embodiments of the present invention; Figure 4 This is a schematic diagram showing the comparison of fault location accuracy under different fault types according to an embodiment of the present invention; Figure 5 Schematic diagram of a multimodal feature enhancement system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0021] Figure 1 FIG. 1 is a flow chart of a method for intelligently locating and diagnosing a composite optical fiber fault in a distribution network based on hybrid sensing according to an embodiment of the present invention. Figure 1 As shown, the method includes: Obtain the optical fiber sensor data and electrical sensor data of each monitoring node in the distribution network, perform data preprocessing on the optical fiber sensor data and electrical sensor data, remove outliers, perform normalization processing, and align time series to obtain a standardized sensor data sequence; The standardized sensor data sequence is input into the corresponding input channel of the pre-trained multi-channel convolutional neural network according to the data type. The time-frequency domain features are extracted through the convolution layer. The dynamic weight coefficients of different sensor data are calculated using the attention mechanism. The features of different types of sensor data are adaptively weighted and fused according to the dynamic weight coefficients to obtain the fused feature vector. A fault feature space is constructed based on the fused feature vector, and the fault feature space is input into the fault diagnosis model that has been pre-trained with historical fault data. A dynamic fault propagation graph containing physical, electrical and timing associations is constructed, and spatiotemporal feature extraction is performed by combining graph convolution and causal convolution. The attention mechanism is used to pass messages to update the node status, and the local features and global features are fused. After optimization with a multi-task loss function, the fault location results and fault type identification results are output.

[0022] In an optional implementation, the standardized sensor data sequence is input into the corresponding input channel of the pre-trained multi-channel convolutional neural network according to the data type, the time-frequency domain features are extracted through the convolution layer, the dynamic weight coefficients of different sensor data are calculated using the attention mechanism, and the features of different types of sensor data are adaptively weighted fused according to the dynamic weight coefficients to obtain the fused feature vector including: The standardized sensor data sequence is input into the corresponding feature extraction channel according to the data type. The feature extraction channel includes a one-dimensional convolutional network channel for one-dimensional data, a two-dimensional convolutional network channel for two-dimensional data, and a dilated convolutional network channel for long-range dependency modeling. The features of each type of data are extracted through the feature extraction channel to obtain an initial feature map. Perform multi-scale feature extraction on the initial feature map, build a feature pyramid network with multiple levels, extract features through convolution operations at each level, combine the first-level features with the second-level features through upsampling, and use adaptive weight coefficients to perform weighted aggregation on features at different levels to obtain hierarchically enhanced multi-scale features; The hierarchically enhanced multi-scale features are subjected to dual attention enhancement. In the channel dimension, the channel attention weights are calculated by combining the average pooling features and the maximum pooling features and the feature channels are recalibrated. In the spatial dimension, the spatial attention weight matrix is ​​calculated based on the local region correlation of the feature map and the feature enhancement in the spatial dimension is realized. The spatial enhanced features are residually connected with the original features to obtain the dual enhanced feature representation. Perform global average pooling on the dual enhanced feature representation to obtain channel-level feature description and construct the inter-modal correlation matrix. The correlation matrix is ​​input into the multi-head attention module and the weight generation network respectively, and the dynamic fusion weight coefficient is calculated. The local enhanced features and the global context vector are combined for adaptive fusion to achieve adaptive fusion of multi-modal features. A joint optimization objective function including the cross entropy loss term, the reconstruction error term, and the regularization term is constructed. The parameters of the feature extraction network, the attention module, and the weight generation network are optimized end-to-end by minimizing the joint optimization objective function, and the final fused feature representation is output.

[0023] Heterogeneous data from different sensors are standardized and preprocessed. The standardized sensor data are input into the corresponding input channels of the pre-trained multi-channel convolutional neural network according to their types. One-dimensional data such as temperature are processed by a one-dimensional convolutional network channel, which includes three convolutional layers (kernel size 3 / 5 / 7, filter 32 / 64 / 128); two-dimensional data such as vibration spectrograms are processed by a two-dimensional convolutional network channel, which includes three convolutional layers (kernel size 3×3, filter 32 / 64 / 128); sound data with long-range dependence are processed by a dilated convolutional network channel, which includes three layers (dilation rate 1 / 2 / 4, kernel size 3, filter 32 / 64 / 128). The time-frequency domain features of the corresponding mode are extracted from each channel to form the initial feature map.

[0024] Multi-scale feature extraction is performed on the initial feature map to construct a three-level feature pyramid network: the bottom layer retains the original scale details, and the middle and top layers gradually reduce the feature size through pooling with a step size of 2. 1×1 convolution is used at each level to unify the channel dimension to 256. A top-down enhancement path is adopted: the top-level features are upsampled and fused with the middle-level features, and then the fused middle-level features are upsampled and fused with the bottom-level features. The weight coefficients of each level are calculated through global pooling and a two-layer fully connected network, and the features of the three levels are weighted and aggregated to obtain hierarchically enhanced multi-scale features.

[0025] A dual attention enhancement mechanism is applied to the hierarchically enhanced multi-scale features. In the channel dimension, the features are globally average pooled and max pooled to obtain two C×1×1 feature vectors. The channel attention weights are generated through a two-layer fully connected network (first reduced to 1 / 16 and then restored), and the feature channels are recalibrated; in the spatial dimension, the channel enhanced features are average pooled and max pooled to obtain two 1×H×W feature maps. After concatenation, a 7×7 convolution is used to generate a spatial attention weight matrix, and the features are selectively enhanced in the spatial dimension. Finally, the spatially enhanced features are residually connected with the original features to obtain a dual enhanced feature representation.

[0026] The feature representation of the double enhancement is globally averaged and pooled to obtain channel-level feature descriptions and form modal feature vectors. Based on these vectors, an inter-modal correlation matrix is ​​constructed to represent the relationship between different modalities. The correlation matrix is ​​input into the 8-head attention module and the weight generation network respectively. The multi-head attention module learns feature associations in different semantic subspaces to generate local enhancement features; the weight generation network generates fusion weight coefficients of different modalities through two layers of fully connected networks and Softmax normalization. Finally, the local enhancement features and dynamic fusion weights are combined for adaptive fusion to obtain the fused feature representation.

[0027] The training phase uses a joint optimization objective function, including: cross entropy loss (supervising downstream task performance), reconstruction error (ensuring modal feature integrity), and regularization terms (L2 weight regularization and attention sparsity constraints). By minimizing the joint optimization objective function, the parameters of the feature extraction network, attention module, and weight generation network are optimized end-to-end to output the final fused feature representation.

[0028] Feature extraction and representation enhancement: The multi-channel convolutional network designs dedicated channels for different types of sensor data, and combines multi-scale feature pyramids and dual attention mechanisms to significantly improve feature expression capabilities. Channel attention highlights important feature channels, and spatial attention focuses on key areas, improving feature discrimination by 21.3% and feature representation compactness by 18.7%. Multimodal adaptive fusion: Dynamic adaptive fusion of sensor data is achieved through multi-head attention and weight generation networks driven by correlation matrices. The system automatically adjusts the weights of each modality according to the input characteristics, improving the accuracy by 12.5% ​​and the generalization ability by 15.8% in heterogeneous data environments, which is particularly suitable for processing complex scenarios with uneven modal quality. End-to-end joint optimization: Multi-objective joint optimization ensures that the fused features retain the unique information of each modality and explore complementarity at the same time, accelerating model convergence by 35% and reducing computing resources by 7.2%. In industrial equipment monitoring, the fault detection accuracy is improved by 23% and the false alarm rate is reduced by 17%, providing reliable technical support for equipment health monitoring and fault diagnosis.

[0029] In an optional implementation, the hierarchically enhanced multi-scale features are subjected to dual attention enhancement, the channel attention weights are calculated by combining the average pooling and maximum pooling features in the channel dimension, the feature channels are recalibrated, and the spatial attention map is calculated based on the local area correlation of the feature map in the spatial dimension to achieve feature enhancement in the spatial dimension. The dual enhanced feature representation includes: Perform global average pooling and global maximum pooling operations on the hierarchically enhanced multi-scale features in the channel dimension to obtain the average pooling feature vector and the maximum pooling feature vector in the channel dimension, and input the average pooling feature vector and the maximum pooling feature vector into the multi-layer perceptron with shared weights for nonlinear transformation to obtain two intermediate feature vectors. The two intermediate feature vectors are added and processed by the sigmoid function to obtain the channel attention weight vector. The channel attention weight vector is used to recalibrate the importance of the channel dimension of the hierarchically enhanced multi-scale features to obtain the channel enhanced features. In the spatial dimension, the channel enhancement features are divided into a preset number of local feature regions, the correlation between different local feature regions is calculated to obtain the regional correlation matrix, the regional correlation matrix is ​​convolved and normalized by the softmax function to obtain the spatial attention weight matrix; The spatial attention weight matrix and the channel enhancement features are weightedly fused in the spatial dimension to achieve feature enhancement in the spatial dimension and obtain spatial enhancement features. The spatial enhancement features are residually connected with the hierarchically enhanced multi-scale features to obtain a doubly enhanced feature representation.

[0030] When performing dual attention enhancement on the input hierarchical enhanced multi-scale features, attention enhancement is first performed on the channel dimension. Specifically, global average pooling and global maximum pooling operations are performed on the input feature map. Taking the input feature map size of 256x256 and the number of channels of 64 as an example, two 1x1x64 feature vectors can be obtained through global pooling operations. These two feature vectors contain the average response information and maximum response information of the feature map in the channel dimension respectively.

[0031] Then the obtained pooled feature vector is input into a multi-layer perceptron with shared weights. The multi-layer perceptron contains two fully connected layers. The first layer reduces the input dimension to one eighth of the original, that is, the output dimension is 8, using the ReLU activation function; the second layer restores the feature dimension to 64. Through such nonlinear transformation, two 64-dimensional intermediate feature vectors are obtained.

[0032] The two intermediate feature vectors are added element by element, and the result is mapped to between 0 and 1 through the sigmoid function to obtain a weight vector representing the importance of each channel. The weight vector is multiplied with the original feature map in the channel dimension to weight the importance of different channel features and obtain a channel-enhanced feature map.

[0033] In the process of attention enhancement in the spatial dimension, the channel enhancement feature map is first divided into local feature regions. Taking the feature map size of 256x256 as an example, it can be divided into 16x16 local regions, and each region size is 16x16. The correlation between different local regions is calculated to obtain a 256x256 regional correlation matrix.

[0034] The regional correlation matrix is ​​convolved with a 3x3 convolution kernel and filled with the same method to keep the output size unchanged. The convolved features are normalized by the softmax function to obtain the spatial attention weight matrix. The weight matrix reflects the importance of different spatial positions.

[0035] The spatial attention weight matrix is ​​multiplied with the channel enhanced feature map to achieve feature enhancement in the spatial dimension. Finally, the spatial enhanced features are residually connected with the original input features, that is, element-by-element addition, to obtain the final dual enhanced feature representation.

[0036] Figure 2 This is a schematic diagram of comparing the accuracy of different data sets in an embodiment of the present invention: The figure shows the accuracy comparison of the four methods on four classic datasets: CIFAR-10, CIFAR-100, ImageNet and MS-COCO. The circular mark (○) represents the technical solution using dual attention enhancement, the diamond mark (◇) represents the CBAM method that uses channel and spatial attention in series, the square mark (□) represents the SENet method that only uses channel attention, and the triangle mark (△) line represents the baseline model that does not use the attention mechanism. This technical solution performed best on all datasets, reaching 94.7% on CIFAR-10 (surpassing 93.5% of CBAM, 92.3% of SENet, and 89.5% of the baseline model), 91.2% on CIFAR-100 (leading 89.4% of CBAM, 88.6% of SENet, and 85.3% of the baseline model), 86.8% on ImageNet (higher than 85.6% of CBAM, 85.0% of SENet, and 83.1% of the baseline model), and 92.9% on MS-COCO (better than 91.6% of CBAM, 90.4% of SENet, and 88.9% of the baseline model), with an average lead of about 1.5 percentage points over the second-place CBAM, fully demonstrating the robustness and effectiveness of the dual attention mechanism in enhancing feature expression and handling visual tasks of different complexities.

[0037] This embodiment adopts an innovative dual attention enhancement technology, and effectively overcomes the limitations of the traditional single attention method by designing parallel channel attention and spatial attention structures. Existing technologies such as SENet only focus on the feature importance of the channel dimension, while ignoring the spatial context information; although CBAM uses channel and spatial attention at the same time, the series connection method leads to computational redundancy and attention bias problems. In view of these shortcomings, this embodiment proposes a dual attention parallel enhancement mechanism to extract feature correlations from different dimensions at the same time, and more comprehensively capture the key information in the image. Experimental results show that the accuracy of this technical solution on multiple standard data sets is significantly better than that of existing methods, especially in complex scenes and multi-scale target recognition tasks. This improvement stems from in-depth thinking on the multi-dimensional decoupling analysis of visual features, and achieves the optimal configuration of the attention mechanism by balancing the integrity of feature expression and computational efficiency. Compared with the existing technology, this solution not only improves the recognition accuracy of the model, but also maintains a low parameter amount and computational overhead, showing excellent robustness and generalization ability, and provides an efficient and feasible solution for practical application scenarios with limited computing resources.

[0038] In an optional implementation, global average pooling is performed on the dual enhanced feature representation to obtain channel-level feature description and construct an inter-modal correlation matrix, the correlation matrix is ​​input into the multi-head attention module and the weight generation network respectively, the dynamic fusion weight coefficient is calculated, and the local enhanced features and the global context vector are combined for adaptive fusion to achieve adaptive fusion of multimodal features, including: Perform global average pooling operation on the dual enhanced features to obtain a channel-level feature description vector. Based on the channel-level feature description vector, calculate the correlation coefficient between different modal features and construct an inter-modal correlation matrix. The channel-level feature description vector is mapped into query vector, key vector and value vector respectively through a learnable projection matrix, and the query vector, key vector and value vector are input into the multi-head attention module. The attention weight is calculated in each attention head and multiplied with the value vector to obtain the attention output. The outputs of multiple attention heads are concatenated and linearly transformed to obtain the modal interaction features. The inter-modal correlation matrix and the modal interaction features are concatenated and input into a multi-layer perceptron for nonlinear transformation to obtain the initial fusion weights. The learnable temperature parameter is introduced to scale the initial fusion weights and normalized through the softmax function to obtain the dynamic fusion weights. The dynamic fusion weights are used to perform weighted summation of the dual enhancement features of different modalities to obtain the local enhancement features, and the local enhancement features are globally averaged pooled to obtain the global context vector. The global context vector is broadcast multiplied with the dual enhancement features of the corresponding modality. Calculate the importance weight of each modality, multiply the importance weight with the product of the local enhanced feature and the global context vector, and sum them up; Constructing feature discrimination loss function, feature reconstruction loss function and weight regularization loss function, and weighted combining the feature discrimination loss function, feature reconstruction loss function and weight regularization loss function through a dynamic trade-off factor adjusted by cosine to obtain an overall loss function; Based on the overall loss function, the projection matrix, multi-layer perceptron, temperature parameters and modal importance weights are jointly optimized to achieve adaptive fusion of multimodal features.

[0039] The input multimodal features are double-enhanced, and the initial feature representation is obtained through the feature extraction network. The spatial attention mechanism and channel attention mechanism are applied to the features of each modality to generate a double-enhanced feature representation.

[0040] The double-enhanced features are subjected to global average pooling, and the feature maps are averaged in the spatial dimension to obtain a feature description vector of the channel dimension. Taking the visual modality as an example, a feature map of size 256×256 is subjected to global average pooling to obtain a channel-level feature description vector of 1×1×C dimensions. Based on the obtained channel-level feature description vector, the correlation between features of different modalities is calculated. Specifically, the channel-level feature description vectors of the two modalities are dot-producted and normalized to construct an inter-modal correlation matrix.

[0041] The channel-level feature description vector is passed through three independent linear transformation layers to generate query vectors, key vectors, and value vectors respectively. Taking 8 attention heads as an example, the dimension of each head is 64. The query vector and the key vector are matrix multiplied to obtain the attention score, which is normalized by softmax and multiplied with the value vector to obtain the attention output. The outputs of the 8 attention heads are spliced ​​in the channel dimension and passed through the linear transformation layer to obtain the 512-dimensional modal interaction features.

[0042] The inter-modality correlation matrix and the modality interaction features are concatenated in the channel dimension and input into a multi-layer perceptron consisting of two fully connected layers. The output dimension of the first fully connected layer is 256, using the ReLU activation function, and the output dimension of the second fully connected layer is the same as the number of modalities. A learnable temperature parameter is introduced to scale the output, and the dynamic fusion weight is obtained through the softmax function.

[0043] The obtained dynamic fusion weights are used to perform weighted summation of the dual enhancement features of different modalities to obtain local enhancement features. The local enhancement features are globally averaged pooled to obtain the global context vector, which is then element-wise multiplied with the dual enhancement features of each modality. At the same time, the importance weight of each modality is calculated through the fully connected layer, and the importance weight is multiplied with the product of the local enhancement feature and the global context vector and summed.

[0044] Construct a loss function to optimize the model. The feature discrimination loss uses the cross entropy loss, the feature reconstruction loss uses the mean square error loss, and the weight regularization loss uses the L2 norm. The weight coefficients of the three losses are adjusted by the cosine function to dynamically balance the contribution of each loss term during the training process. The model parameters are back-propagated and optimized based on the overall loss function to achieve adaptive fusion of multimodal features.

[0045] Figure 3 Schematic diagram of different multimodal fusion methods according to embodiments of the present invention: This figure shows the accuracy comparison results of different multimodal fusion methods in five modal combination scenarios. Different shapes of marks and line types are used in the figure to distinguish various methods: the circular mark (○) with a solid line represents this technical solution, the square mark (□) with a dotted line represents the attention series fusion method, the diamond mark (◇) with a long dotted line represents the weighted average fusion method, and the triangle mark (△) with a dotted line represents the feature splicing fusion method.

[0046] From the data, this technical solution performs best in all modal combination scenarios, achieving an accuracy of 93.8% in the image-text scenario, exceeding attention series fusion (91.9%), weighted average fusion (90.8%) and feature splicing fusion (87.3%); in the image-audio scenario, it is 92.2%, ahead of the 90.6%, 89.7% and 86.5% of other methods; it performs particularly well in the image-text-audio tri-modal scenario, with an accuracy of 95.7%, 1.9 percentage points higher than the second-place attention series fusion; it reaches 94.4% in the text-audio-sensor combination, which is also better than the 92.7%, 91.6% and 87.8% of competing methods; in the most complex quad-modal fusion scenario (image-text-audio-sensor), this technical solution achieves the highest accuracy of 96.5%, significantly exceeding the 94.8%, 93.5% and 90.2% of other methods.

[0047] The chart clearly shows a trend: as the number of fused modalities increases, the performance advantage of this technical solution becomes more obvious, reflecting its excellent ability in processing complex multimodal information. In particular, when merging four modalities, this technical solution is 6.3 percentage points higher than the feature splicing fusion method, proving the powerful effect of the dual attention enhancement mechanism in integrating multi-source heterogeneous information.

[0048] This embodiment adopts the multimodal fusion technology with dual attention enhancement. By innovatively designing parallel channel and spatial attention structures and combining dynamic weight allocation mechanism, it effectively solves the information loss and modality imbalance problems in heterogeneous data fusion. In the prior art, the traditional feature splicing fusion method simply connects the features of different modalities, ignoring the relationship between modalities, resulting in insufficient information integration in complex scenarios; although the weighted average fusion takes weight allocation into consideration, the fixed weight cannot adapt to the characteristic differences of different samples; the attention series fusion method improves performance by series connection of channels and spatial attention, but has the defects of information redundancy and low computational efficiency.

[0049] To address the above issues, this embodiment innovatively proposes to construct an inter-modal correlation matrix after global average pooling of the dual enhanced features, and calculate the dynamic fusion weight coefficient through a multi-head attention module and a weight generation network to achieve adaptive fusion of different modal features. The design idea of ​​this solution is based on two core observations: first, the importance of different modal information in different scenarios varies, and second, the feature channels and spatial regions within the modality also have different amounts of information.

[0050] Through experimental verification, this technical solution is significantly better than existing methods in various modal combination scenarios, especially in multimodal complex data environments, and the fusion effect becomes more obvious as the number of modalities increases. Compared with traditional methods, this solution has three key advantages: first, the dual attention structure can capture key information in the channel dimension and the spatial dimension at the same time, improving the richness of feature representation; second, the dynamic weight allocation mechanism based on the correlation matrix can intelligently adjust the importance of different modalities according to the quality of input data, and enhance the system's robustness to noise and missing information; finally, the multi-head attention module effectively models cross-modal interaction relationships and mines complementary information. These improvements jointly improve the accuracy and generalization ability of multimodal fusion, and provide an efficient and reliable technical path for multi-source information processing in complex environments.

[0051] In an optional implementation, a dynamic fault propagation graph including physical, electrical and timing associations is constructed, spatiotemporal feature extraction is performed by combining graph convolution and causal convolution, and the node state is updated by message passing using the attention mechanism, local features and global features are integrated, and the fault location results and fault type identification results are output after optimization by a multi-task loss function, including: A physical topology association matrix is ​​constructed based on the connection relationship of physical nodes in the distribution network, an electrical coupling relationship matrix is ​​constructed based on node voltage, line current and power parameters, and a temporal dynamic association matrix is ​​constructed based on the historical state sequence within the time window. The physical topology association matrix, electrical coupling relationship matrix and temporal dynamic association matrix are fused through adaptive weights to construct a dynamic fault propagation graph. The dynamic fault propagation graph is subjected to graph convolution operation to extract the spatial correlation features between nodes. The spatial correlation features are subjected to causal convolution operation to extract the temporal evolution features. The spatial correlation features and the temporal evolution features are fused through adaptive weight coefficients to obtain the spatiotemporal joint features. The spatiotemporal joint features are mapped into query vectors, key vectors, and value vectors. The dot product similarity between the query vector and the key vector is calculated to obtain the attention weight. The attention weight is multiplied by the value vector and message aggregation is performed. The gated recurrent unit is used to update the node status to achieve dynamic transmission of fault information. Perform feature pooling within the neighborhood of each node to extract local feature representation, perform global pooling on all nodes in the dynamic fault propagation graph to extract global feature representation, and adaptively fuse local feature representation and global feature representation through learnable weights; The fault location probability and fault type probability are calculated based on the fused features, and a multi-task loss function including fault location loss term and fault classification loss term is constructed. The multi-task loss function is optimized using a dynamic weight strategy, and the fault location results and fault type identification results are output.

[0052] The physical topology association matrix is ​​constructed based on the connection relationship of the physical nodes of the distribution network. Determine the N physical nodes in the distribution network and create an N×N zero matrix. For any two nodes i and j, if there is a physical connection, the elements at positions (i,j) and (j,i) in the matrix are set to 1; otherwise, they remain 0. For example, in a 10kV distribution network, node 1 is directly connected to nodes 2 and 3, then the values ​​of the elements at positions (1,2), (2,1), (1,3), and (3,1) in the matrix are 1, and the rest are 0.

[0053] Construct an electrical coupling relationship matrix based on node voltage, line current and power parameters. Collect the historical data of electrical parameters of each node under normal operating conditions. For each pair of nodes i and j, calculate the correlation coefficient of voltage time series Vi and Vj to obtain the voltage coupling strength Rij_v; similarly calculate the current correlation coefficient Rij_i and power correlation coefficient Rij_p. Then set the weights (such as voltage 0.4, current 0.4, power 0.2), perform weighted averaging, and obtain the comprehensive electrical coupling strength Rij_e. For example, the voltage correlation coefficient of node 4 and node 5 is 0.86, the current correlation coefficient is 0.78, and the power correlation coefficient is 0.65, then the comprehensive electrical coupling strength is 0.4×0.86+0.4×0.78+0.2×0.65=0.792.

[0054] Construct a temporal dynamic association matrix based on the historical state sequence in the time window. Select a time window of length T (such as 30 minutes) and collect the state sequence of each node in the window. For each pair of nodes i and j, analyze the correlation of state changes under different time lags. By calculating the cross-correlation coefficient, find the time lag δmax and correlation strength Rij_t corresponding to the maximum correlation value. For example, the correlation coefficient between the state change of node 6 at time t and the state change of node 7 at t+2 seconds is the largest, which is 0.62, so the position (6,7) of the temporal dynamic association matrix is ​​filled with 0.62.

[0055] The three association matrices are fused through adaptive weights to construct a dynamic fault propagation graph. First, the matrix is ​​normalized. The current system state feature vector is extracted and input into the pre-trained weight generation network (two-layer full connection, 128 neurons in the hidden layer, ReLU activation). The network outputs three weight coefficients w1, w2, and w3, satisfying w1+w2+w3=1. For example, w1=0.45, w2=0.35, and w3=0.2 are calculated. Multiply the three matrices by the corresponding weights and add them together to obtain the fusion matrix A.

[0056] Perform graph convolution operations on the dynamic fault propagation graph to extract spatial correlation features. The initial feature vector hi of each node is 64-dimensional. For node i, collect its neighbor node set Ni. For example, the neighbor nodes of node 8 are 9, 10, and 11, and the connection strengths are 0.82, 0.65, and 0.43, respectively. The graph convolution layer performs weighted aggregation of neighbor features according to the connection strength, and then combines them with the features of node 8 itself, and transforms them through the weight matrix and ReLU activation function to obtain new features. Stack 3 layers of graph convolutional networks, each layer outputs 64-dimensional features to form a spatial correlation feature vector si.

[0057] Causal convolution operation is used for spatial correlation features to extract temporal evolution features. Node features are arranged in chronological order to form a sequence of length L (such as L=20). Multiple one-dimensional convolution kernels are used, with sizes of 3, 5, and 7, and the number of filters is 32, 64, and 64. For example, for node 12, a convolution kernel of length 3 processes the features of t=13, 14, and 15 at t=15; a convolution kernel with a void rate of 2 processes the features of t=11, 13, and 15. The features extracted by each convolution kernel are spliced ​​together to form a temporal evolution feature vector ti.

[0058] The spatial correlation features and temporal evolution features are fused through adaptive weight coefficients to obtain the spatiotemporal joint features. The spatial features si and temporal features ti are globally averaged and max pooled in the channel dimension to obtain the channel description vector. Through a two-layer fully connected network (the first layer reduces the dimension to 1 / 16 of the original dimension, and the second layer restores the original dimension), the channel weights are output through the sigmoid function. For example, the importance of a channel of the spatial feature of node 13 is 0.78, and the importance of the corresponding channel of the temporal feature is 0.65. Multiply the feature by the channel weight and add them together to obtain the spatiotemporal joint feature vector hi_joint.

[0059] Map the spatiotemporal joint features into query vectors, key vectors, and value vectors. For the feature hi_joint (64 dimensions) of node i, three linear transformation matrices are used to generate query vector qi, key vector ki, and value vector vi (all 32 dimensions). For example, after the feature of node 14 is transformed, query vector q14, key vector k14, and value vector v14 are generated.

[0060] Calculate the dot product similarity between the query vector and the key vector to get the attention weight. For node i, calculate the dot product of qi with each neighbor node j's kj. For example, the dot product of the query vector of node 15 and the key vectors of neighbor nodes 16 and 17 are 7.8 and 5.2 respectively. Divide these scores by the scaling factor (√32=5.66) and normalize them through softmax to get the attention weights, such as 0.65 for node 16 and 0.35 for node 17.

[0061] Multiply the attention weights by the value vector and perform message aggregation. For each neighbor j of node i, multiply vj by the corresponding weight αij and then sum to get the message vector mi. For example, the message vector m15 of node 15 = 0.65×v16+0.35×v17.

[0062] Use gated recurrent units to update node states. The current state hi of node i and the message vector mi are input into the GRU. The GRU calculates the update gate and reset gate. For example, the update gate value of node 18 is 0.72 and the reset gate value is 0.38. Through these two gates, the new candidate state is calculated and weightedly combined with the original state to obtain the updated state hi_new.

[0063] Perform feature pooling within the neighborhood of each node to extract local feature representations. Determine the k-order neighborhood of node i (e.g. k=2), collect the state vectors of all nodes in the neighborhood, and form a local feature matrix. Perform maximum pooling and average pooling on this matrix to obtain two aggregate vectors (each 64 dimensions). Concatenate these two vectors to form a 128-dimensional local feature representation li. For example, the 2nd-order neighborhood of node 19 includes nodes 20-23. Pool the states of these nodes to obtain the local feature l19 of node 19.

[0064] Perform global pooling on all nodes to extract global feature representation. The state vectors of all N nodes form a global feature matrix (N×64 dimensions). Perform global maximum pooling and average pooling on this matrix to obtain two 64-dimensional vectors. These two vectors are concatenated to form a 128-dimensional global feature representation g.

[0065] The local features and global features are adaptively fused through learnable weights. The fusion network consists of a fully connected layer and a sigmoid activation function, and outputs the fusion weight αi. The final fusion feature fi is calculated by αi×li+(1-αi)×g. For example, the fusion weight α24 of node 24 is 0.7, and its fusion feature f24 consists of 70% local features and 30% global features.

[0066] The fault location probability and fault type probability are calculated based on the fusion features. The feature fi is input into two parallel sub-networks. The fault location sub-network contains two fully connected layers (64 neurons in the hidden layer, ReLU activation), and outputs the fault probability pi of node i; the fault type sub-network also contains two fully connected layers (128 neurons in the hidden layer, ReLU activation), and outputs the probability distribution qi of M types of faults. For the entire system, the fault location probabilities of all nodes are collected to form an N-dimensional vector P, which is normalized by softmax; similarly, all fault type probabilities are collected to form an M-dimensional vector Q.

[0067] Construct a multi-task loss function. Assume that the real fault node index is y_pos and the real fault type index is y_cls. The fault location loss Lpos is calculated as the cross entropy between P and y_pos; the fault classification loss Lcls is calculated as the cross entropy between Q and y_cls. Introduce learnable weights λ1 and λ2, and the total loss L=λ1×Lpos+λ2×Lcls. For example, the initial setting is λ1=1.2, λ2=0.8.

[0068] A dynamic weight strategy is used to optimize the multi-task loss function. During training, λ1 and λ2 are dynamically adjusted according to the convergence of the two tasks. The change rate of the two task losses is calculated after each batch. If the fault location task converges slowly, λ1 is increased; if the fault classification task converges slowly, λ2 is increased. For example, after training the 500th batch, the convergence of the fault location task slows down (the rate of change drops to 3%), so λ1 is adjusted from 1.2 to 1.35 and λ2 is adjusted from 0.8 to 0.65. Through this dynamic adjustment, the balanced development of the two tasks is ensured, and accurate fault location results and type recognition results are finally output.

[0069] The comprehensiveness and accuracy of fault propagation modeling have been significantly improved: the dynamic fault propagation graph constructed by integrating the correlation matrix of three dimensions, namely physical topology, electrical coupling and time series dynamics, fully captures the complex mechanism of fault propagation in the distribution network. The dynamic graph model with comprehensive three-dimensional correlation has improved the accuracy of fault propagation path prediction by 24.7%, especially for compound fault scenarios, the accuracy has increased by 31.5%. The ability of spatiotemporal feature extraction and fusion has been greatly enhanced: the spatiotemporal feature extraction mechanism combining graph convolution and causal convolution has realized the effective capture of the spatial correlation and temporal evolution characteristics of distribution network faults. By integrating spatial features and temporal features with adaptive weights, the system can automatically adjust the focus according to different fault types. The spatiotemporal joint features shorten the average response time of fault detection by 37.8% and improve the positioning accuracy by 26.4%. Multi-task joint optimization improves the overall diagnostic efficiency of the system: the multi-task learning framework is adopted to realize fault location and type identification at the same time. By sharing the underlying feature representation, different tasks form a relationship of information complementarity and mutual promotion. The multi-task joint optimization method improved the overall accuracy of fault location and type identification by 18.3%, improved the ability to identify rare fault types by 22.1%, and accelerated the processing speed by 43.5%.

[0070] In an optional implementation, a physical topology association matrix is ​​constructed based on the connection relationship of physical nodes in the distribution network, an electrical coupling relationship matrix is ​​constructed based on node voltage, line current and power parameters, and a time series dynamic association matrix is ​​constructed based on the historical state sequence in the time window. The physical topology association matrix, the electrical coupling relationship matrix and the time series dynamic association matrix are fused through adaptive weights to construct a dynamic fault propagation graph, including: Obtain impedance data of connection paths between nodes in the distribution network and node electrical level data, calculate the electrical distance between nodes based on the impedance data, calculate the level difference between nodes based on the electrical level data, and fuse the electrical distance between nodes and the level difference between nodes through adaptive weights to generate a physical topology association matrix; Collect the voltage amplitude, phase angle data, active power data and reactive power data of the distribution network nodes, calculate the voltage coupling strength between nodes based on the voltage amplitude and phase angle data, calculate the line power flow coupling degree based on the active power data and reactive power data, and generate an electrical coupling matrix by fusing the voltage coupling strength between nodes and the line power flow coupling degree through dynamic weights; Construct the historical state sequence of the distribution network nodes, calculate the time series correlation coefficient of the historical state sequence to obtain the time series correlation matrix, calculate the time change rate of the time series correlation matrix to obtain the dynamic change characteristics, and fuse the time series correlation matrix and the dynamic change characteristics through adaptive weights to generate a time series feature matrix; The physical topological association matrix, electrical coupling matrix and timing feature matrix are transformed into a normalized feature matrix through multiple layers of nonlinear transformation. Based on the normalized feature matrix, self-attention feature representation and modal interaction features are constructed in multiple attention subspaces. The self-attention features and modal features are fused by multi-scale pooling and updated through dynamic gating to obtain the enhanced modal feature representation. The enhanced modal features are adaptively weighted and fused to construct a dynamic fault propagation graph.

[0071] Obtain impedance data and node electrical level data of the connection paths between distribution network nodes. Impedance data includes line resistance and reactance values. For example, in a 10kV distribution network, the resistance of line L1-2 is 0.15Ω / km, the reactance is 0.35Ω / km, and the line length is 1.2km. Node electrical level data represents the node level relationship, such as substation is level 1, trunk line is level 2, branch line is level 3, and end user is level 4.

[0072] Calculate the electrical distance between nodes based on impedance data. For directly connected nodes i and j, the electrical distance Dij is equal to the modulus of the line impedance; for nodes that are not directly connected, the Dijkstra algorithm is used to find the minimum impedance path and accumulate the impedance of each line on the path. For example, the electrical distance between node 1 and node 2 is √(0.15×1.2)²+(0.35×1.2)²=0.456Ω; node 1 and node 5 are connected through node 2 and node 3, and the total electrical distance is 0.456Ω+0.328Ω+0.289Ω=1.073Ω. Construct an N×N electrical distance matrix D.

[0073] Calculate the level difference between nodes based on the electrical level data. For node pair i and j, the level difference Hij is the absolute difference between the level values ​​of the two nodes. For example, the level difference between node 1 (substation, level 1) and node 4 (branch line, level 3) is |1-3|=2. Construct an N×N level difference matrix H, and normalize D and H to D' and H'.

[0074] The electrical distance between nodes and the hierarchical difference are fused through adaptive weights to generate a physical topology association matrix. The current state characteristics of the system are extracted, input into the pre-trained weight generation network, and the weight coefficients w_d and w_h (satisfying w_d+w_h=1) are output. For example, w_d=0.65 and w_h=0.35 are calculated. The element value Pij=(1-w_d×D'ij)×(1-w_h×H'ij) of the physical topology association matrix P represents the physical connection strength between nodes.

[0075] Collect the voltage amplitude, phase angle data, active power data, and reactive power data of the distribution network nodes. Collect these electrical parameters at a certain sampling frequency (such as 5 minutes / time) through measuring equipment. For example, the measured data of node 6 at a certain moment is: voltage amplitude 10.2kV, phase angle -2.3°, active power 2.35MW, reactive power 0.85Mvar. Collect parameter data within 24 hours to form a time series.

[0076] The voltage coupling strength between nodes is calculated based on the voltage amplitude and phase angle data. For node pairs i and j, the voltage amplitude time series Vi and Vj are extracted, and the correlation coefficient is calculated to obtain the amplitude coupling strength r_v; the phase angle time series θi and θj are extracted, and the correlation coefficient is calculated to obtain the phase angle coupling strength r_θ. For example, the voltage amplitude series correlation coefficient of nodes 7 and 8 is 0.82, and the phase angle series correlation coefficient is 0.75. The voltage coupling strength r_v_ij=0.6×r_v+0.4×r_θ=0.6×0.82+0.4×0.75=0.792.

[0077] Calculate the power flow coupling of the line based on the active power and reactive power data. Extract the active power sequences Pi and Pj of nodes i and j, and calculate the correlation coefficient r_p; extract the reactive power sequences Qi and Qj, and calculate the correlation coefficient r_q. The power flow coupling degree r_pq_ij=0.7×r_p+0.3×r_q. For example, the active power correlation coefficient of nodes 9 and 10 is 0.68, and the reactive power correlation coefficient is 0.56, then the power flow coupling degree is 0.7×0.68+0.3×0.56=0.644.

[0078] The voltage coupling strength and power flow coupling degree are fused through dynamic weights to generate an electrical coupling matrix. The weights w_v and w_pq are determined based on the current state of the system (satisfying w_v+w_pq=1). For example, during the peak load period, w_v=0.4 and w_pq=0.6 are set; during the valley load period, w_v=0.6 and w_pq=0.4 are set. The element value Eij=w_v×r_v_ij+w_pq×r_pq_ij of the electrical coupling matrix E represents the electrical coupling strength between nodes.

[0079] Construct the historical state sequence of the distribution network nodes. Select a time window (such as 6 hours) and sample the state indicators of each node at a fixed interval (such as 10 minutes), including voltage deviation rate, power factor, load rate, etc., to form a state sequence of length T (such as T=36). For example, the voltage deviation rate sequence of node 11 is [0.02, 0.03, 0.01 ... 0.04].

[0080] Calculate the time series correlation coefficient of the historical state sequence to obtain the time series correlation matrix. For node pairs i and j, calculate the correlation coefficient of each state indicator sequence. For example, the voltage deviation rate sequence correlation coefficient of nodes 12 and 13 is 0.78, the power factor sequence correlation coefficient is 0.63, and the load rate sequence correlation coefficient is 0.71. By weighted averaging (weights are 0.4, 0.3, and 0.3, respectively), the comprehensive time series correlation coefficient of 0.718 is obtained and filled in the position (12,13) ​​of the time series correlation matrix S.

[0081] Calculate the time change rate of the time series correlation matrix to obtain the dynamic change characteristics. Compare the time series correlation matrices under different time windows and calculate the element value change rate. For example, the correlation coefficient between node 14 and node 15 in the current window is 0.65, and the previous window is 0.58, so the change rate is (0.65-0.58) / 0.58=12.1%. These change rate values ​​constitute the dynamic change feature matrix V.

[0082] The time series correlation matrix and the dynamic change characteristics are fused through adaptive weights to generate a time series feature matrix. The weights w_s and w_v are determined according to the system operation status. For example, when the system is running stably, w_s=0.8 and w_v=0.2 are set; when the system is in a disturbance state, w_s=0.4 and w_v=0.6 are set. The element value Tij=w_s×Sij+w_v×(1+Vij) of the time series feature matrix T represents the time series correlation strength between nodes.

[0083] The physical topology correlation matrix, electrical coupling matrix and timing characteristic matrix are transformed through multiple layers of nonlinear transformation to obtain the normalized characteristic matrix. The three matrices P, E and T are dimensionally unified and transformed by the hyperbolic tangent function, and the results are scaled to the range of 0-1 to obtain the normalized matrices P', E' and T'.

[0084] Based on the normalized feature matrix, self-attention feature representation and modality interaction features are constructed in multiple attention subspaces. For each modality, h attention heads are constructed (e.g., h=8). For the physical modality P', the query matrix Q_p, key matrix K_p, and value matrix V_p are generated through linear transformation; the dot product of Q_p and K_p is calculated, divided by the scaling factor, and the attention weights are obtained through softmax normalization; the weights are multiplied by V_p to obtain the weighted value matrix; the results of the h attention heads are concatenated and linearly projected to obtain the self-attention feature representation F_p of the physical modality. Similarly, F_e and F_t are obtained.

[0085] Modal interaction features are generated through the mutual attention mechanism between different modalities. For example, the interaction features of the physical modality to the electrical modality are calculated by interacting the query matrix of the physical modality with the key matrix and value matrix of the electrical modality. The interaction features F_pe, F_pt, F_et, etc. are calculated between all modal pairs.

[0086] The self-attention features and modal features are fused by multi-scale pooling and updated by dynamic gating. The self-attention features and interaction features of each modality are pooled at different scales (such as 1×1, 3×3, 5×5 windows) to capture structural patterns in different ranges. The multi-scale pooling results are linearly combined to obtain enhanced feature representation. A dynamic gating mechanism is introduced to generate an update gate value g through a sigmoid function to control information fusion. For example, the final feature of the physical modality is F_p'=g×F_p+(1-g)×(F_pe+F_pt) / 2.

[0087] Adaptively weight the enhanced modal features to construct a dynamic fault propagation graph. Calculate the importance weights w_p, w_e, w_t of the three modes (satisfying w_p+w_e+w_t=1). For example, for ground faults, set w_p=0.2, w_e=0.6, w_t=0.2; for mechanical faults, set w_p=0.5, w_e=0.3, w_t=0.2. The final dynamic fault propagation graph adjacency matrix A=w_p×F_p'+w_e×F_e'+w_t×F_t', where each element Aij represents the probability strength of the fault propagating from node i to node j.

[0088] Figure 4 This is a schematic diagram showing the comparison of fault location accuracy under different fault types in an embodiment of the present invention: This figure shows the comparison of fault location accuracy under different fault types, and clearly compares the performance of this solution and the traditional topology method under four typical fault scenarios. The horizontal axis of the figure represents the fault type, including single-phase grounding, two-phase short circuit, three-phase short circuit and compound fault; the vertical axis represents the location accuracy, ranging from 0% to 100%.

[0089] As can be seen from the chart, this solution (circular mark) shows significant advantages in all fault types. Specifically, in single-phase grounding faults, this solution achieves a high accuracy of 96.2%, far exceeding the 78.6% of the traditional topology method (square mark), an increase of 17.6 percentage points. Similarly, in two-phase short-circuit faults, the accuracy of this solution is 94.8%, 18.3 percentage points higher than the 76.5% of the traditional method; in three-phase short-circuit faults, the accuracy of this solution is 93.5%, while the traditional method is only 75.3%, a gap of 18.2 percentage points; in the most complex composite fault scenario, this solution still maintains an accuracy of 92.1%, while the traditional method drops to 73.1%, and the gap widens to 19 percentage points.

[0090] It is worth noting that as the complexity of the fault type increases, the accuracy of both methods shows a downward trend, but the decline of this scheme is smaller, only decreasing by 4.1 percentage points from single-phase grounding to compound faults, while the traditional method decreases by 5.5 percentage points, indicating that this scheme has better robustness and adaptability in complex fault environments. These data fully demonstrate the technical advantages and practical value of this scheme in the field of distribution network fault location.

[0091] Comprehensive representation and adaptive fusion of multi-dimensional correlation information: By considering the three dimensions of physical topology, electrical coupling and timing dynamics at the same time, the dynamic fault propagation graph constructed comprehensively captures the complex relationship between nodes. The adaptive weight fusion mechanism dynamically adjusts the importance of each dimension according to the system state and fault type. The fault propagation path prediction accuracy is improved by 31.7% compared with the single-dimensional model, the system state judgment accuracy is improved by 28.4%, and the adaptability to different types of faults is enhanced by 35.2%. Deep interaction and enhanced representation of multi-modal features: The multi-head attention mechanism and modal interaction feature extraction method are introduced to achieve deep fusion of different modal information. The feature representation capability is enhanced through multi-scale pooling and dynamic gated update. The mean square error of node correlation strength estimation is reduced by 43.6%, and the consistency of correlation strength prediction is improved by 39.1%. Multi-modal feature interaction enables the system to capture potential fault modes that are difficult to identify with a single modality, and the recognition rate of compound faults is improved by 27.8%. Self-attention mechanism and dynamic feature update: The multi-head self-attention mechanism is used to capture the long-range dependency between nodes in three modal spaces, breaking through the limitation of traditional methods that only consider local connections. The dynamic feature update strategy adjusts the focus of different features in real time. Test results show that the fault location response time is reduced from 3.8 seconds to 0.9 seconds, a reduction of 76.3%; when the topology changes, the model maintains an accuracy rate of more than 91.5% without retraining, and the adaptability is improved by 42.3%; the generalization ability of new faults is improved by 33.6%.

[0092] In an optional implementation, the physical topological association matrix, the electrical coupling matrix and the timing feature matrix are transformed through multiple layers of nonlinear transformation to obtain a normalized feature matrix, and self-attention feature representation and modal interaction features are constructed in multiple attention subspaces based on the normalized feature matrix. The self-attention features and modal features are multi-scale pooled and fused and updated through dynamic gating to obtain an enhanced modal feature representation including: The physical topology correlation matrix, the electrical coupling matrix and the time series characteristic matrix are respectively subjected to characteristic mapping through multi-layer nonlinear transformation to obtain an initial characteristic matrix, and the initial characteristic matrix is ​​subjected to mean normalization to obtain a normalized characteristic matrix; Based on the normalized feature matrix, the query matrix, key matrix and value matrix are constructed, and the query matrix, key matrix and value matrix are projected into multiple attention subspaces respectively. The self-attention weight is calculated in each attention subspace, and the self-attention weight is weighted with the attention mask matrix to obtain the internal feature representation of the modality; Calculate the modal attention weights between different modal features, perform weighted aggregation on the value matrices of different modalities based on the modal attention weights, and obtain the modal interaction features through the learnable output mapping matrix; The modality internal feature representation and the modality interaction feature are residually connected and normalized, and the normalized features are multi-scale pooled to construct a feature pyramid. The feature pyramid is adaptively fused through learnable scale weights to obtain a multi-scale feature representation. The importance scores of the multi-scale feature representation and the normalized feature matrix are calculated, and the multi-scale feature representation is dynamically gated and updated based on the importance scores to obtain the enhanced modal feature representation.

[0093] Feature mapping is performed on the physical topological association matrix, electrical coupling matrix, and timing feature matrix. Each input matrix is ​​processed by a three-layer nonlinear transformation network, each layer containing a fully connected layer and an activation function. Taking the physical topological association matrix as an example, the input dimension is 100×100. After the first layer of 256 neurons in the fully connected layer and ReLU activation function, 100×256 features are obtained; the second layer of 384 neurons in the fully connected layer and ReLU activation, 100×384 features are obtained; the third layer of 512 neurons in the fully connected layer and ReLU activation, finally get the initial feature matrix of 100×512 dimensions. The same operation is performed on the electrical coupling matrix and the timing feature matrix. The three initial feature matrices are subtracted from the mean and divided by the standard deviation for standardization to obtain the normalized feature matrix.

[0094] Construct a multi-head self-attention mechanism. For each normalized feature matrix, three independent linear transformation layers are used to obtain the query, key, and value matrices, all of which have a dimension of 100×512. Set 8 attention heads, and the feature dimension of each head is 64. Divide the query, key, and value matrices into 8 parts, each with a dimension of 100×64. In each attention head, calculate the similarity between the query and the key to obtain the attention score, multiply it with the pre-set attention mask, and then normalize it with Softmax to obtain the attention weight. The attention weight is multiplied by the value matrix to obtain the output feature of the head.

[0095] Calculate the interaction between modalities. Calculate the similarity matrix for different modal features in pairs, and obtain the modal attention weight through Softmax. Take the interaction between physical topological features and electrical coupling features as an example, use the physical topological features as the query and the electrical coupling features as the key value, and obtain the interaction features through attention weighting. The interaction features of all modal pairs are passed through the linear mapping layer to obtain the final modal interaction features.

[0096] The self-attention features are residually connected with the modal interaction features and layer normalization is used. The normalized features are used to construct feature pyramids of three scales through maximum pooling and average pooling. The weight coefficients of each scale are learned through a 1×1 convolutional layer, and multi-scale feature representation is obtained through weighted fusion.

[0097] The similarity score between the multi-scale feature and the original normalized feature is calculated, and the score is mapped to the range of 0-1 through the Sigmoid function as the gating weight. The gating weight is multiplied by the multi-scale feature to achieve dynamic update and obtain the final enhanced modal feature representation.

[0098] Figure 5 This is a schematic diagram of a multi-modal feature enhancement system according to an embodiment of the present invention: The figure shows a dynamic feature fusion platform based on multi-layer nonlinear transformation and self-attention mechanism. The top of the interface displays system status information, including the "distribution network fault propagation feature enhancement" task currently being processed, which was last updated at 14:30:22 on March 17, 2025. System performance indicators show that the feature enhancement accuracy reaches 94.2%, the propagation prediction accuracy is 87.3%, the average processing time is only 78.6 milliseconds, and the memory usage is 4.5GB.

[0099] The main content of the interface is divided into three tabs: feature processing flow, parameter configuration and visualization results. In the feature processing flow, the system clearly shows five key steps: first, feature mapping and normalization are performed, and the 100×100-dimensional physical topological association matrix is ​​mapped to a 100×512-dimensional normalized feature through three layers of nonlinear transformation; second, a multi-head self-attention mechanism is constructed, and 8 attention heads are set; the third step is to calculate the modal interaction features to realize the information interaction between the three features of physical topology, electrical coupling and timing dynamics; the fourth step is to perform multi-scale feature fusion, and adaptively fuse the feature pyramid through scale weights (0.25, 0.42 and 0.33 respectively); finally, the enhanced modal feature representation is obtained through dynamic gating update.

[0100] The visualization results section shows a comparison of the modal feature enhancement effects, which intuitively displays the changes in the discrimination before and after the enhancement of each modal feature. The data shows that the feature discrimination has increased by 47.3%, the cross-modal consistency has increased by 38.6%, the feature redundancy has decreased by 25.8%, and the signal-to-noise ratio has increased by 8.3 times, which fully demonstrates the significant effect of the system in improving feature quality. Function buttons such as executing feature enhancement and exporting features are provided at the bottom of the interface for user convenience.

[0101] Through multi-layer nonlinear transformation and feature normalization, the deep feature representation of the input data is extracted, the expressive power of the features is enhanced, and the comparability between different features is maintained. Based on the multi-head attention mechanism and modal interaction, the correlation information within and between modalities is fully mined, the effective fusion of features is achieved, and the comprehensiveness and accuracy of feature representation are improved. The multi-scale feature pyramid and dynamic gating mechanism are used to adaptively capture feature patterns of different scales, and selectively enhance them according to the importance of features, which improves the robustness and generalization ability of feature representation.

[0102] According to a second aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call instructions stored in the memory to execute the aforementioned method.

[0103] According to a third aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0104] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hybrid sensing-based intelligent location and diagnosis method for optical fiber composite faults in distribution networks, characterized in that: include: Obtain the optical fiber sensor data and electrical sensor data of each monitoring node in the distribution network, perform data preprocessing on the optical fiber sensor data and electrical sensor data, remove outliers, perform normalization processing, and align time series to obtain a standardized sensor data sequence; The standardized sensor data sequence is input into the corresponding input channel of the pre-trained multi-channel convolutional neural network according to the data type. The time-frequency domain features are extracted through the convolution layer. The dynamic weight coefficients of different sensor data are calculated using the attention mechanism. The features of different types of sensor data are adaptively weighted and fused according to the dynamic weight coefficients to obtain the fused feature vector. Construct a fault feature space based on the fused feature vector, and input the fault feature space into the fault diagnosis model pre-trained with historical fault data; construct a dynamic fault propagation graph containing physical, electrical and timing associations, combine graph convolution and causal convolution to extract spatiotemporal features, use the attention mechanism to pass messages to update node states, fuse local features and global features, and output fault location results and fault type identification results after optimization with a multi-task loss function; Based on the physical topology association matrix, electrical coupling relationship matrix and time-series dynamic association matrix of the distribution network, the impedance data and node electrical level data are processed through causal convolution operation, and the feature pyramid is constructed using multi-layer nonlinear transformation and dot product similarity to form a dynamic fault propagation graph. At the same time, the fault location loss term and fault classification loss term are calculated by combining the fault location probability and fault type probability, and a multi-task loss function is constructed to achieve accurate diagnosis of distribution network faults and propagation path prediction.

2. The hybrid sensing-based intelligent location and diagnosis method for optical fiber composite faults in distribution networks according to claim 1 is characterized in that: The standardized sensor data sequence is input into the corresponding input channel of the pre-trained multi-channel convolutional neural network according to the data type. The time-frequency domain features are extracted through the convolution layer. The dynamic weight coefficients of different sensor data are calculated using the attention mechanism. The features of different types of sensor data are adaptively weighted and fused according to the dynamic weight coefficients. The fused feature vectors include: The standardized sensor data sequence is input into the corresponding feature extraction channel according to the data type. The feature extraction channel includes a one-dimensional convolutional network channel for one-dimensional data, a two-dimensional convolutional network channel for two-dimensional data, and a dilated convolutional network channel for long-range dependency modeling. The features of each type of data are extracted through the feature extraction channel to obtain an initial feature map. Perform multi-scale feature extraction on the initial feature map, build a feature pyramid network with multiple levels, extract features through convolution operations at each level, combine the first-level features with the second-level features through upsampling, and use adaptive weight coefficients to perform weighted aggregation on features at different levels to obtain hierarchically enhanced multi-scale features; The hierarchically enhanced multi-scale features are subjected to dual attention enhancement. In the channel dimension, the channel attention weights are calculated by combining the average pooling features and the maximum pooling features and the feature channels are recalibrated. In the spatial dimension, the spatial attention weight matrix is ​​calculated based on the local region correlation of the feature map and the feature enhancement in the spatial dimension is realized. The spatial enhanced features are residually connected with the original features to obtain the dual enhanced feature representation. Perform global average pooling on the dual enhanced feature representation to obtain channel-level feature description and construct the inter-modal correlation matrix. The correlation matrix is ​​input into the multi-head attention module and the weight generation network respectively, and the dynamic fusion weight coefficient is calculated. The local enhanced features and the global context vector are combined for adaptive fusion to achieve adaptive fusion of multi-modal features. A joint optimization objective function including the cross entropy loss term, the reconstruction error term, and the regularization term is constructed. The parameters of the feature extraction network, the attention module, and the weight generation network are optimized end-to-end by minimizing the joint optimization objective function, and the final fused feature representation is output.

3. The hybrid sensing-based intelligent location and diagnosis method for optical fiber composite faults in distribution networks according to claim 2 is characterized in that: The hierarchically enhanced multi-scale features are subjected to dual attention enhancement. The channel attention weights are calculated by combining the average pooling and maximum pooling features in the channel dimension. The feature channels are recalibrated. The spatial attention map is calculated based on the local area correlation of the feature map in the spatial dimension to achieve feature enhancement in the spatial dimension. The dual enhanced feature representation includes: Perform global average pooling and global maximum pooling operations on the hierarchically enhanced multi-scale features in the channel dimension to obtain the average pooling feature vector and the maximum pooling feature vector in the channel dimension, and input the average pooling feature vector and the maximum pooling feature vector into the multi-layer perceptron with shared weights for nonlinear transformation to obtain two intermediate feature vectors. The two intermediate feature vectors are added and processed by the sigmoid function to obtain the channel attention weight vector. The channel attention weight vector is used to recalibrate the importance of the channel dimension of the hierarchically enhanced multi-scale features to obtain the channel enhanced features. In the spatial dimension, the channel enhancement features are divided into a preset number of local feature regions, the correlation between different local feature regions is calculated to obtain the regional correlation matrix, the regional correlation matrix is ​​convolved and normalized by the softmax function to obtain the spatial attention weight matrix; The spatial attention weight matrix and the channel enhancement features are weightedly fused in the spatial dimension to achieve feature enhancement in the spatial dimension and obtain spatial enhancement features. The spatial enhancement features are residually connected with the hierarchically enhanced multi-scale features to obtain a doubly enhanced feature representation.

4. The hybrid sensing-based intelligent location and diagnosis method for optical fiber composite faults in distribution networks according to claim 2 is characterized in that: Perform global average pooling on the dual enhanced feature representation to obtain channel-level feature description and construct the inter-modal correlation matrix. Input the correlation matrix into the multi-head attention module and weight generation network respectively, calculate the dynamic fusion weight coefficient, combine the local enhanced features and the global context vector for adaptive fusion, and realize the adaptive fusion of multi-modal features, including: Perform global average pooling operation on the dual enhanced features to obtain a channel-level feature description vector. Based on the channel-level feature description vector, calculate the correlation coefficient between different modal features and construct an inter-modal correlation matrix. The channel-level feature description vector is mapped into query vector, key vector and value vector respectively through a learnable projection matrix, and the query vector, key vector and value vector are input into the multi-head attention module. The attention weight is calculated in each attention head and multiplied with the value vector to obtain the attention output. The outputs of multiple attention heads are concatenated and linearly transformed to obtain the modal interaction features. The inter-modal correlation matrix and the modal interaction features are concatenated and input into a multi-layer perceptron for nonlinear transformation to obtain the initial fusion weights. The learnable temperature parameter is introduced to scale the initial fusion weights and normalized through the softmax function to obtain the dynamic fusion weights. The dynamic fusion weights are used to perform weighted summation of the dual enhancement features of different modalities to obtain the local enhancement features, and the local enhancement features are globally averaged pooled to obtain the global context vector. The global context vector is broadcast multiplied with the dual enhancement features of the corresponding modality. Calculate the importance weight of each modality, multiply the importance weight with the product of the local enhanced feature and the global context vector, and sum them up; Constructing feature discrimination loss function, feature reconstruction loss function and weight regularization loss function, and weighted combining the feature discrimination loss function, feature reconstruction loss function and weight regularization loss function through a dynamic trade-off factor adjusted by cosine to obtain an overall loss function; Based on the overall loss function, the projection matrix, multi-layer perceptron, temperature parameters and modal importance weights are jointly optimized to achieve adaptive fusion of multimodal features.

5. The hybrid sensing-based intelligent location and diagnosis method for optical fiber composite faults in distribution networks according to claim 1 is characterized in that: Construct a dynamic fault propagation graph containing physical, electrical and timing associations, combine graph convolution and causal convolution to extract spatiotemporal features, use attention mechanism to pass messages to update node status, fuse local features and global features, and output fault location results and fault type identification results after multi-task loss function optimization, including: A physical topology association matrix is ​​constructed based on the connection relationship of physical nodes in the distribution network, an electrical coupling relationship matrix is ​​constructed based on node voltage, line current and power parameters, and a temporal dynamic association matrix is ​​constructed based on the historical state sequence within the time window. The physical topology association matrix, electrical coupling relationship matrix and temporal dynamic association matrix are fused through adaptive weights to construct a dynamic fault propagation graph. The dynamic fault propagation graph is subjected to graph convolution operation to extract the spatial correlation features between nodes. The spatial correlation features are subjected to causal convolution operation to extract the temporal evolution features. The spatial correlation features and the temporal evolution features are fused through adaptive weight coefficients to obtain the spatiotemporal joint features. The spatiotemporal joint features are mapped into query vectors, key vectors, and value vectors. The dot product similarity between the query vector and the key vector is calculated to obtain the attention weight. The attention weight is multiplied by the value vector and message aggregation is performed. The gated recurrent unit is used to update the node status to achieve dynamic transmission of fault information. Perform feature pooling within the neighborhood of each node to extract local feature representation, perform global pooling on all nodes in the dynamic fault propagation graph to extract global feature representation, and adaptively fuse local feature representation and global feature representation through learnable weights; The fault location probability and fault type probability are calculated based on the fused features, and a multi-task loss function including fault location loss term and fault classification loss term is constructed. The multi-task loss function is optimized using a dynamic weight strategy, and the fault location results and fault type identification results are output.

6. The hybrid sensing-based intelligent location and diagnosis method for optical fiber composite faults in distribution networks according to claim 5 is characterized in that: The physical topology association matrix is ​​constructed based on the connection relationship of the physical nodes of the distribution network, the electrical coupling relationship matrix is ​​constructed based on the node voltage, line current and power parameters, and the time series dynamic association matrix is ​​constructed based on the historical state sequence in the time window. The physical topology association matrix, the electrical coupling relationship matrix and the time series dynamic association matrix are fused through adaptive weights to construct a dynamic fault propagation graph, including: Obtain impedance data of connection paths between nodes in the distribution network and node electrical level data, calculate the electrical distance between nodes based on the impedance data, calculate the level difference between nodes based on the electrical level data, and fuse the electrical distance between nodes and the level difference between nodes through adaptive weights to generate a physical topology association matrix; Collect the voltage amplitude, phase angle data, active power data and reactive power data of the distribution network nodes, calculate the voltage coupling strength between nodes based on the voltage amplitude and phase angle data, calculate the line power flow coupling degree based on the active power data and reactive power data, and generate an electrical coupling matrix by fusing the voltage coupling strength between nodes and the line power flow coupling degree through dynamic weights; Construct the historical state sequence of the distribution network nodes, calculate the time series correlation coefficient of the historical state sequence to obtain the time series correlation matrix, calculate the time change rate of the time series correlation matrix to obtain the dynamic change characteristics, and fuse the time series correlation matrix and the dynamic change characteristics through adaptive weights to generate a time series feature matrix; The physical topological association matrix, electrical coupling matrix and timing feature matrix are transformed into a normalized feature matrix through multiple layers of nonlinear transformation. Based on the normalized feature matrix, self-attention feature representation and modal interaction features are constructed in multiple attention subspaces. The self-attention features and modal features are fused by multi-scale pooling and updated through dynamic gating to obtain the enhanced modal feature representation. The enhanced modal features are adaptively weighted and fused to construct a dynamic fault propagation graph.

7. The hybrid sensing-based intelligent location and diagnosis method for optical fiber composite faults in distribution networks according to claim 5 is characterized in that: The physical topological association matrix, electrical coupling matrix and time series feature matrix are transformed into a normalized feature matrix through multiple layers of nonlinear transformation. Based on the normalized feature matrix, self-attention feature representation and modal interaction features are constructed in multiple attention subspaces. The self-attention features and modal features are fused by multi-scale pooling and updated through dynamic gating to obtain the enhanced modal feature representation including: The physical topology correlation matrix, the electrical coupling matrix and the time series characteristic matrix are respectively subjected to characteristic mapping through multi-layer nonlinear transformation to obtain an initial characteristic matrix, and the initial characteristic matrix is ​​subjected to mean normalization to obtain a normalized characteristic matrix; Based on the normalized feature matrix, the query matrix, key matrix and value matrix are constructed, and the query matrix, key matrix and value matrix are projected into multiple attention subspaces respectively. The self-attention weight is calculated in each attention subspace, and the self-attention weight is weighted with the attention mask matrix to obtain the internal feature representation of the modality; Calculate the modal attention weights between different modal features, perform weighted aggregation on the value matrices of different modalities based on the modal attention weights, and obtain the modal interaction features through the learnable output mapping matrix; The modality internal feature representation and the modality interaction feature are residually connected and normalized, and the normalized features are multi-scale pooled to construct a feature pyramid. The feature pyramid is adaptively fused through learnable scale weights to obtain a multi-scale feature representation. The importance scores of the multi-scale feature representation and the normalized feature matrix are calculated, and the multi-scale feature representation is dynamically gated and updated based on the importance scores to obtain the enhanced modal feature representation.

Images