35kV collection line fault diagnosis method based on cloud edge collaboration

Abstract

The invention relates to a 35kV collection line fault diagnosis method based on cloud edge cooperation, and the method comprises the steps: fusing multi-modal features, and constructing local graph nodes; hybrid coding compression is carried out based on the topology sensing edge weight and then the edge weight is uploaded to the cloud; aggregating the plurality of local graphs into a global graph; fusing the signal timing characteristics and the line topology to generate enhanced features; high-dimensional features are extracted through the dynamic graph convolutional network; and eliminating time sequence deviation through multi-node time sequence calibration, and executing distributed weighted optimization to calculate node contribution weight to obtain a final fault point. Through adaptive feature extraction and topology perception compression, optimized graph node data is generated, and collaborative optimization of high resolution, robustness and low bandwidth occupation is realized; high-precision fault identification and positioning are realized through a time-space mixed attention mechanism and a dynamic graph convolutional network; and in combination with the global graph and the wave head time, the fault point location of the line is accurately determined through multi-node time sequence calibration and weighted optimization.

Description

Technical Field

[0001] This invention relates to the field of transmission line fault location technology, specifically to a fault diagnosis method for 35kV collection lines based on cloud-edge collaboration. Background Technology

[0002] In recent years, the number of abnormal states in actual operation of new energy power plants has been 3-4 times per year, and the time for handling each power outage is about 3-5 days. Based on 3 abnormal power outages and 3 days for handling each abnormal state, the direct economic loss reduced by the implementation of the project in the current year is: 328,000 kWh (average daily power generation) * 3 days * 0.86 yuan = 846,240 yuan.

[0003] Traditional fault diagnosis algorithms exhibit severe inadequacy in the complex electromagnetic environment of wind farms. The electromagnetic environment of wind farm collector lines is filled with high-frequency harmonic noise (ranging from kHz to MHz) generated by wind turbine converters and photovoltaic inverters, forming a strong noise background. Meanwhile, key transient characteristic signals (such as partial discharge and traveling wave fronts) characterizing early insulation degradation or transient faults are usually very weak. Traditional algorithms struggle to effectively separate and extract these weak fault features from such a complex noise background, resulting in a low signal-to-noise ratio for fault identification and persistently high rates of false positives and false negatives.

[0004] Wind farm collector lines connect dozens of dispersed wind turbine generators, typically employing a complex tree-like or ring-like topology with multiple branches. Traditional fault location methods (such as traveling wave location methods) are usually based on a simple two-end measurement model, relying primarily on the time difference between the arrival of the fault signal at two measurement points. When the line has three or more branches, the propagation path of the fault traveling wave signal becomes highly variable, resulting in complex reflections, refractions, and superpositions at each branch point. This algorithm based on a simple two-end model cannot effectively distinguish between direct waves and multiple reflected waves, making it difficult to accurately determine which specific branch the fault point is located on. This leads to location failures or a sharp increase in error (often exceeding several hundred meters), resulting in low reliability of the location results. Furthermore, traditional methods treat the line as a static geometric structure, failing to effectively consider the dynamic changes in signal propagation paths caused by line galloping and topology switching (such as switching operations) during actual operation, further weakening the accuracy and reliability of the location.

[0005] When faced with the unique characteristics of new energy power plants, such as strong electromagnetic noise, power fluctuations, and complex line topologies, the shortcomings of existing technologies, such as rigid diagnostic models and single location methods, are fully exposed. They are difficult to achieve early and accurate warning of faults and meter-level accurate location, and cannot meet the needs of modern wind farms for efficient and intelligent operation and maintenance. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a fault diagnosis method for 35kV collection lines based on cloud-edge collaboration, so as to overcome the shortcomings of the prior art.

[0007] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A fault diagnosis method for a 35kV aggregation line based on cloud-edge collaboration, comprising the following steps: Step S01: Dynamically acquire high-frequency transient signals, generate optimized time spectrum using dynamic wavelet packet decomposition; fuse multimodal features to construct local graph nodes; perform hybrid encoding compression based on topology-aware edge weights and upload to the cloud; Step S02: The cloud receives compressed data and decompresses it to restore the local map; multiple local maps are aggregated into a global map and feature standardization is performed; signal timing characteristics and line topology are fused to generate enhanced features; high-dimensional features are extracted through a dynamic graph convolutional network, and network parameters are dynamically updated to adapt to topology changes; fault classification results and initial location results are output. Step S03: Based on the fault probability, initial location distance, global topology map and wavefront time, eliminate timing deviation through multi-node timing calibration, and perform distributed weighted optimization to calculate the node contribution weights to obtain the final fault point coordinates; Step S04: Based on the final fault point coordinates and fault probability, and combined with the global graph, perform priority assessment and dynamic feedback.

[0008] The beneficial effects of this invention are as follows: This method utilizes a distributed data acquisition and preprocessing approach at the edge, with the core being the "Dynamic Wavelet Packet Decomposition and Graph Node Adaptive Generation" (DWPD-GNAG) mechanism. This mechanism upgrades the distributed monitoring equipment (deployed on each line's towers or branch points, equipped with wideband current sensors, APN communication modules, GPS synchronization modules, and embedded computing units) of the 35kV collection lines at new energy power plants into intelligent sensing nodes. This not only efficiently acquires high-frequency transient electrical signals (sampling rate ≥ 1MHz, frequency band 10kHz-10MHz), but also generates optimized graph node data through adaptive feature extraction and topology-aware compression for further analysis in the cloud. DWPD-GNAG is designed for the complex electromagnetic environment of new energy power plants (such as high-frequency inverter noise and signal reflection caused by multi-branch topology), overcoming the limitations of traditional preprocessing (such as fixed wavelet transform and single-mode feature extraction), and achieving coordinated optimization of high resolution, robustness, and low bandwidth occupancy.

[0009] This method utilizes a cloud-based Hybrid Attention Deep Model (HADM) to process complex fault signals in the 35kV collection lines of renewable energy power plants. HADM receives the generated compressed data and, based on local graphs and fused features, achieves high-precision fault identification and localization through a temporal-spatial hybrid attention mechanism and a dynamic graph convolutional network (GNN). Addressing the non-stationary signal characteristics of renewable energy power plants (such as high-frequency noise from inverters and signal reflections caused by multi-branch topologies) and the need for multi-node collaboration, HADM overcomes the limitations of traditional deep learning models (such as CNNs and RNNs), significantly improving fault classification accuracy (>95%) and localization error (<10m).

[0010] This method utilizes a cloud-based distributed traveling wave localization mechanism. Based on fault probability and initial location distance, combined with a global map and wavefront time, it accurately determines the location of fault points (such as partial discharge and arc discharge) in the 35kV collection lines of new energy power plants through multi-node time-series calibration and weighted optimization. This mechanism addresses the complex topology of new energy power plants (such as multiple branch lines and dynamic switching states) and non-stationary interference (such as high-frequency noise from inverters), overcoming the limitations of traditional traveling wave localization methods (such as single-end ranging and double-end ranging). It achieves a localization error of <5m and improves robustness by approximately 40%.

[0011] Based on the above technical solution, the present invention can be further improved as follows.

[0012] Furthermore, step S01 specifically includes: High-frequency transient signals are collected by distributed monitoring devices to generate the original signal matrix. For the signal matrix Calculate the frequency domain entropy matrix ; Frequency domain entropy matrix Generate dynamic scale factors by combining noise spectrum Based on scale factor Perform adaptive decomposition to obtain wavelet packet coefficients. ; Multi-scale time-spectrum matrices are generated and optimized through mutual information filtering. The optimized time-frequency spectrum matrix is ​​obtained. Perform multimodal fusion, extract key features of traveling waves, and generate a fused feature vector. ; Constructing a local graph Estimating edge weights based on dynamic edge weights Compressed data is obtained after topology-sensing compression. Compress the data Uploaded to the cloud.

[0013] Furthermore, step S01 specifically includes the following steps: Step S11: Collect high-frequency transient signals through distributed monitoring devices to generate the original signal matrix. : ; in, The original signal matrix, , , These are current, voltage, and temperature signal vectors, respectively. Step S12: For the signal matrix The formula for calculating the frequency domain entropy matrix is ​​as follows: ; in, The frequency domain entropy matrix, For frequency band index, frequency band Energy distribution, For frequency variables; Frequency domain entropy matrix The dynamic scaling factor is generated by combining the noise spectrum, as shown in the following formula: ; in, Scale factor; To prevent the denominator from being zero; typical NSR range is 0.1-0.5; According to the scaling factor Select the decomposition level and mother wavelet, based on the spectral correlation coefficient. The decomposition formula is: ; in, These are wavelet packet coefficients. It is a current signal. For the conjugate of the mother wavelet, For time, For translation parameters, As a scale parameter, focus on the 1-5MHz frequency band; By filtering with mutual information, a multi-scale time-spectrum matrix is ​​generated: ; in, The time-frequency spectrum matrix, For scale indexing, For time indexing; Optimization of the time spectrum matrix : ; in, The optimized time-frequency spectrum matrix, For mutual information, For frequency components; Step S13: Apply the Hilbert transform The analytic signal is generated using the following formula: ; in, To analyze the signal, The result of the Hilbert transform; Extract the following parameters: Wave head timing: , accuracy <10ns; Energy gradient: ,in The signal energy reflects the intensity of the disturbance. Phase shift: Capture signal phase abrupt changes; Multimodal fusion generates fused feature vectors : ; in, To fuse feature vectors, , , This is the weight matrix. For voltage phase shift, This is the temperature correction vector; weight matrix Through optimization, the formula is as follows: ; in, This is the weight matrix. Signal matrix The covariance matrix, For regularization parameters, It is the Frobenius norm; Step S14: Generate a local undirected graph Node characteristics: ; in, For nodes feature, For nodes The fused feature vector, The wavehead time is given; the edge features are: ; in, For the edge feature, The interval length is... For interference weights; Estimating edge weights using dynamic edge weights: ; in, For border rights, , , For coefficients, Due to time difference, To store the impedance, The covariance of the fusion features; Prioritize compressing high-weight edges, using hybrid encoding. The compression formula is as follows: ; in, To compress the data, Compression ratio; .

[0014] Furthermore, step S11 also includes: For signal matrix Perform signal complexity evaluation: ; in, This represents the normalized probability density. Joint entropy combined with real-time environmental noise spectrum generates triggering factors: ; in, The trigger factor is NSR, which is the noise-to-signal ratio. when Exceeding the dynamic threshold When triggered, it switches to high-frequency sampling mode, and under normal circumstances, it maintains low-frequency sampling.

[0015] Furthermore, step S02 specifically includes: Receive compressed data in the cloud Local images are restored through reverse encoding. Aggregate the local graphs of multiple edge devices into a global graph. ; Based on global graph Calculate the time attention weight vector Spatial attention weight vector The two are fused to generate hybrid attention features. ; Hybrid attention features Input dynamic graph convolutional network, convolutional features Classification probabilities are generated through fully connected layers. Predicting the initial location of the fault point using a regression model. .

[0016] Furthermore, step S02 specifically includes the following steps: Step S21: Receive compressed data from the cloud Local images are restored through reverse encoding. , among which node features Edge features and border rights ; partial diagrams of multiple edge devices Aggregate into a global graph : ; in, The set of features for all nodes. It is the set of all edge features; Through initial standardization: ; in, For standardized node features, , for The mean and standard deviation; Step S22: For the standardized node features and wave head time The time attention weights are calculated using the following formula: ; in, For nodes Temporal attention vector, , , For queries, key, and value vectors; , , For learnable weight matrix, The dimension of the key vector. This is the normalization function; Spatial Attention: Based on Global Graph Calculate the spatial attention weights: ; in, For nodes For nodes Spatial attention weights, , , , , , For learnable weight matrix, For border rights, ; Fuse temporal attention vectors and spatial attention weights to generate enhanced features: ; in, For nodes Hybrid attention features , These are the weighting coefficients. For nodes The set of neighbors; Step S23: Move the node Hybrid attention features Input dynamic graph convolutional network: ; in, For the first Layer nodes The convolutional features, It is the ReLU activation function. For the first Layer weight matrix, For bias, For border rights; Border rights Updated every minute: ; in, For a moment The right to the border, For update rate; This represents the increment of edge weights; Convolutional features Generating classification probabilities through fully connected layers: ; in, For nodes The probability of failure, , These are the parameters for the fully connected layer; Location is determined by predicting the distance to the fault point using a regression model: ; in, For initial location, it represents the distance from the fault point to the node. The distance; , For regression parameters; Step S24: HADM uses a joint loss function to optimize classification and localization: ; in, For cross-entropy loss, Mean square error, , These are the weighting coefficients.

[0017] Furthermore, step S03 specifically includes: Based on the edge weights between nodes and line impedance compute nodes and Propagation delay Perform timing calibration to obtain the calibrated wavefront time. ; Based on calibrated wavehead time Combined with failure probability and initial positioning Perform distributed weighted localization optimization to determine the coordinates of the fault point. ; Calculate the localization contribution weight for each node based on the calibrated wavefront time. and initial positioning Optimize the global coordinates of the fault point ; Introducing edge weights Correct global coordinates of the fault point The final coordinates of the fault point are obtained. .

[0018] Furthermore, step S03 specifically includes the following steps: Step S31: Based on the edge weights between nodes and line impedance compute nodes and Propagation delay: ; in, To delay the spread, The distance between nodes. For the speed of travel wave propagation, For interference weights, To store the impedance, Reference impedance; Calibrate wavehead time: ; in, For the calibrated wavehead time, For nodes The neighborhood group, For border rights; Step S32: Calculate the positioning contribution weight of each node: ; in, For nodes Contribution weight, This represents the probability of failure. For convolutional features, , for The mean and standard deviation; Based on calibrated wavehead time and initial positioning Optimize the coordinates of the fault point: ; in, The coordinates of the fault point, For nodes The known coordinates, For the traveling wave velocity, For reference only; Introducing edge weights Corrected location results: ; in, The corrected coordinates of the fault point. For projection onto the edge Position offset.

[0019] Furthermore, step S03 also includes the following steps: Step S33: Update node contribution weights in each round: ; in, For the first Weighting For the first Wheel positioning results The standard deviation of the distance is used; after 3-5 iterations, the positioning error is converged to <5m. Integrating results from multiple rounds: ; in, The coordinates of the final fault point. For the first Total weight of rounds.

[0020] Furthermore, step S04 specifically includes: Based on the inspection results, the node weights are updated using the following formula: ; in, For the updated weights, For the original weights, For feedback gain, This is a fault confirmation indication function; after updating, step S03 is re-executed to optimize subsequent inspections. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a diagram showing the device deployment structure of the present invention. Detailed Implementation

[0022] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0023] like Figures 1-2 As shown in Example 1, a fault diagnosis method for a 35kV collection line based on cloud-edge collaboration includes the following steps: Step S01: Dynamically acquire high-frequency transient signals, generate optimized time spectrum using dynamic wavelet packet decomposition; fuse multimodal features to construct local graph nodes; perform hybrid encoding compression based on topology-aware edge weights and upload to the cloud; Step S02: The cloud receives compressed data and decompresses it to restore the local map; multiple local maps are aggregated into a global map and feature standardization is performed; signal timing characteristics and line topology are fused to generate enhanced features; high-dimensional features are extracted through a dynamic graph convolutional network, and network parameters are dynamically updated to adapt to topology changes; fault classification results and initial location results are output. Step S03: Based on the fault probability, initial location distance, global topology map and wavefront time, eliminate timing deviation through multi-node timing calibration, and perform distributed weighted optimization to calculate the node contribution weights to obtain the final fault point coordinates; Step S04: Based on the final fault point coordinates and fault probability, and combined with the global graph, perform priority assessment and dynamic feedback.

[0024] This method employs a distributed data acquisition and preprocessing approach at the edge, with the core mechanism being the "Dynamic Wavelet Packet Decomposition and Graph Node Adaptive Generation" (DWPD-GNAG) mechanism. This mechanism upgrades the distributed monitoring equipment (deployed on each line's towers or branch points, equipped with wideband current sensors, APN communication modules, GPS synchronization modules, and embedded computing units) of the 35kV collection lines at renewable energy power plants into intelligent sensing nodes. This not only efficiently acquires high-frequency transient electrical signals (sampling rate ≥1MHz, frequency band 10kHz-10MHz), but also generates optimized graph node data through adaptive feature extraction and topology-aware compression for further analysis in the cloud. DWPD-GNAG is designed for the complex electromagnetic environment of renewable energy power plants (such as high-frequency inverter noise and signal reflections caused by multi-branch topologies), overcoming the limitations of traditional preprocessing methods (such as fixed wavelet transform and single-mode feature extraction), achieving a synergistic optimization of high resolution, robustness, and low bandwidth occupancy.

[0025] This method utilizes a cloud-based Hybrid Attention Deep Model (HADM) to process complex fault signals in the 35kV collection lines of renewable energy power plants. HADM receives the generated compressed data and, based on local graphs and fused features, achieves high-precision fault identification and localization through a temporal-spatial hybrid attention mechanism and a dynamic graph convolutional network (GNN). Addressing the non-stationary signal characteristics of renewable energy power plants (such as high-frequency noise from inverters and signal reflections caused by multi-branch topologies) and the need for multi-node collaboration, HADM overcomes the limitations of traditional deep learning models (such as CNNs and RNNs), significantly improving fault classification accuracy (>95%) and localization error (<10m).

[0026] This method utilizes a cloud-based distributed traveling wave localization mechanism. Based on fault probability and initial location distance, combined with a global map and wavefront time, it accurately determines the location of fault points (such as partial discharge and arc discharge) in the 35kV collection lines of new energy power plants through multi-node time-series calibration and weighted optimization. This mechanism addresses the complex topology of new energy power plants (such as multiple branch lines and dynamic switching states) and non-stationary interference (such as high-frequency noise from inverters), overcoming the limitations of traditional traveling wave localization methods (such as single-end ranging and double-end ranging). It achieves a localization error of <5m and improves robustness by approximately 40%.

[0027] Example 2 is a further improvement based on Example 1, and its details are as follows: Step S01 specifically includes: High-frequency transient signals are collected by distributed monitoring devices to generate the original signal matrix. For the signal matrix Calculate the frequency domain entropy matrix ; Frequency domain entropy matrix Generate dynamic scale factors by combining noise spectrum Based on scale factor Perform adaptive decomposition to obtain wavelet packet coefficients. ; Multi-scale time-spectrum matrices are generated and optimized through mutual information filtering. The optimized time-frequency spectrum matrix is ​​obtained. Perform multimodal fusion, extract key features of traveling waves, and generate a fused feature vector. ; Constructing a local graph Estimating edge weights based on dynamic edge weights Compressed data is obtained after topology-sensing compression. Compress the data Uploaded to the cloud.

[0028] Example 3 is a further improvement based on Example 2, and its details are as follows: Step S01 specifically includes the following steps: Step S11: Data acquisition is optimized for high-frequency transient signals (such as fault discharges and electromagnetic pulses caused by tripping) of the 35kV collection line; the wideband current sensor adopts a non-invasive design to capture signals in the 10kHz-10MHz frequency band, with a sampling rate ≥1MHz, ensuring the integrity of wavefront steepness and frequency transitions. The GPS synchronization module provides high-precision timestamps (error <10ns) and supports multi-node signal timing alignment. The equipment integrates multi-modal sensors (current, voltage, temperature); high-frequency transient signals are acquired through distributed monitoring equipment to generate the original signal matrix. : ; in, The original signal matrix (typical dimension 1024×3 samples / 10ms window). , , These are current, voltage, and temperature signal vectors, respectively. Step S12: Data acquisition proceeds to the DWPD stage, which is a groundbreaking innovation of this method. DWPD surpasses traditional fixed-scale wavelet transform by generating high-resolution features through multi-scale entropy analysis and dynamic spectrum reconstruction. The specific steps are as follows: Multiscale entropy analysis: for signal matrices The formula for calculating the frequency domain entropy matrix is ​​as follows: ; in, The frequency domain entropy matrix, For frequency band index, frequency band Energy distribution (calculated via FFT) For frequency variables; Frequency domain entropy matrix The dynamic scaling factor is generated by combining the noise spectrum, as shown in the following formula: ; in, Scale factor; To prevent the denominator from being zero; typical NSR range is 0.1-0.5; Adaptive decomposition: based on scale factor Choose the decomposition level (typically 4-8) and mother wavelet (e.g., db4 or sym4), based on the spectral correlation coefficient. The decomposition formula is: ; in, These are wavelet packet coefficients. It is a current signal. For the conjugate of the mother wavelet, For time, For translation parameters, As a scale parameter, focus on the 1-5MHz frequency band; Feature reconstruction: A multi-scale time-spectral matrix is ​​generated through mutual information filtering (threshold MI=0.1). ; in, The time-frequency spectrum matrix, For scale indexing, For time indexing (time resolution 0.1μs, frequency resolution 10kHz). Optimization of the time spectrum matrix : ; in, The optimized time-frequency spectrum matrix, For mutual information, For frequency components (based on frequency variables) ); Step S13: Optimized time-spectrum matrix Entering the multimodal fusion stage, key features of the traveling wave are extracted; Hilbert transform is applied. The analytic signal is generated using the following formula: ; in, To analyze the signal, The result of the Hilbert transform; Extract the following parameters: Wave head timing: , accuracy <10ns; Energy gradient: ,in The signal energy reflects the intensity of the disturbance. Phase shift: Capture signal phase abrupt changes; Multimodal fusion generates fused feature vectors (Dimension 256): ; in, To fuse feature vectors (dimension 256). , , This is the weight matrix (typical values ​​0.6, 0.3, 0.1). Voltage phase shift (based on voltage signal) Hilbert transform). Temperature correction vector (based on) (Normalization) weight matrix Through optimization, the formula is as follows: ; in, This is the weight matrix. Signal matrix The covariance matrix, For regularization parameters, It is the Frobenius norm; After fusion, based on the energy-phase joint threshold ( , Perturbation screening is performed; Step S14: Local Graph Construction: Generate Local Undirected Graph Node characteristics: ; in, For nodes feature, For nodes The fused feature vector, The wavehead time is given; the edge features are: ; in, For the edge feature, The interval length is... Interference weights (based on noise spectrum) (Estimation) Estimating edge weights using dynamic edge weights: ; in, For border rights, , , The coefficients are obtained through regression optimization. Due to time difference, To store the impedance, The covariance of the fusion features; Topology-aware compression: Prioritize compressing high-weight edges (sorted) It employs hybrid coding (Huffman + differential), and the compression formula is: ; in, To compress the data, For compression ratio, dynamically adjusted from 8 to 16 bits; Compression ratio 20:1-30:1, data packet <500B.

[0029] Example 4 is a further improvement on Example 3, and its details are as follows: Step S11 also includes: For signal matrix Perform signal complexity evaluation: ; in, The normalized probability density (estimated by Gaussian kernel density); The triggering factor is generated by combining joint entropy with the real-time ambient noise spectrum (estimated via Fast Fourier Transform, frequency band 10kHz-10MHz): ; in, The triggering factor is NSR, which is the noise-to-signal ratio, typically... ; Adaptive triggering: when Exceeding the dynamic threshold hour,( , (Based on historical mean and standard deviation, updated via a 1-hour sliding window), triggering a switch to high-frequency sampling mode (1MHz), while maintaining low-frequency sampling (10kHz) under normal conditions, reducing power consumption to below 0.8W. This mechanism is optimized for intermittent interference in new energy power plants (such as photovoltaic inverter pulses and wind power grounding fluctuations), reducing energy consumption by 70% compared to traditional continuous high-frequency acquisition, with a disturbance capture rate >99%. Device installation as Figure 2 As shown, the deployment structure of the equipment on the pole tower includes an integrated layout of current sensors, Beidou modules, and APN communication units, which optimizes signal acquisition and data transmission efficiency.

[0030] Example 5 is a further improvement based on Example 1, and its details are as follows: Step S02 specifically includes: Receive compressed data in the cloud Local images are restored through reverse encoding. ; partial diagrams of multiple edge devices Aggregate into a global graph ; Based on global graph Calculate the time attention weight vector Spatial attention weight vector The two are fused to generate hybrid attention features. ; Hybrid attention features Input dynamic graph convolutional network, convolutional features Classification probabilities are generated through fully connected layers. Predicting the initial location of the fault point using a regression model. .

[0031] Example 6 is a further improvement on Example 5, and its details are as follows: Step S02 specifically includes the following steps: Step S21: Receive compressed data from the cloud The local image is reconstructed through inverse encoding (Huffman + differential decoding). , among which node features Edge features and border rights To adapt to global analysis in the cloud, HADM (Hybrid Attention Deep Learning Model) integrates local graphs from multiple edge devices. Aggregate into a global graph : ; in, The set of features for all nodes. It is a set of all edge features, generated based on topological connection rules; Global Graph After reconstruction, the nodes fusion feature vector (Dimension 256, including time-frequency features, phase shift, and temperature correction) After initial normalization: ; in, For standardized node features, , for The mean and standard deviation (calculated based on all nodes) are calculated. This step ensures consistent feature scale and improves model training stability. Step S22: The core innovation of HADM lies in its time-space hybrid attention mechanism, which combines the time dimension (signal timing characteristics) and the spatial dimension (line topology) to optimize feature extraction; the specific steps are as follows: Temporal attention: targeting standardized node features and wave head time The time attention weights are calculated using the following formula: ; in, For nodes Temporal attention vector, , , For queries, key, and value vectors; , , It is a learnable weight matrix (dimension 256×64). The dimension of the key vector. The function is a normalized function; temporal attention is focused on wavefront time. The dynamic changes of fault signals are captured to determine their temporal correlation. Spatial Attention: Based on Global Graph Calculate the spatial attention weights: ; in, For nodes For nodes Spatial attention weights, , , , , , For learnable weight matrix, For border rights, Spatial attention utilizes line topology (edge ​​weights) Enhance the correlation between nodes to adapt to multi-branch topologies; Hybrid attention fusion: fusing temporal attention vectors and spatial attention weights to generate enhanced features. ; in, For nodes Hybrid attention features , These are the weighting coefficients (optimized through grid search). For nodes The neighbor set; hybrid attention combined with temporal and topological information significantly improves the accuracy of fault feature extraction; Step S23: Move the node Hybrid attention features Input the dynamic graph convolutional network to further extract high-dimensional features and perform fault classification and localization: ; in, For the first Layer nodes The convolutional features, It is the ReLU activation function. For the first Layer weight matrix, For bias, Edge weights are used; graph convolution utilizes topological structure to aggregate neighbor information, enhancing feature representation; Dynamic updates: To adapt to the dynamic topology of new energy power stations (such as line switching), edge weights... Updated every minute: ; in, For a moment The right to the border, For update rate; The edge weight increment is based on the attention feature covariance; Fault Classification and Localization: Convolutional Features Generating classification probabilities through fully connected layers: ; in, For nodes The probability of failure (categories include partial discharge, arc discharge, and normal). , These are the parameters for the fully connected layer; Location is determined by predicting the distance to the fault point using a regression model: ; in, For initial location, it represents the distance from the fault point to the node. The distance; , These are regression parameters; classification accuracy > 95%, localization error < 10m; Step S24: HADM uses a joint loss function to optimize classification and localization: ; in, For cross-entropy loss, Mean square error, , These are the weighting coefficients.

[0032] Example 7 is a further improvement based on Example 1, and its details are as follows: Step S03 specifically includes: Based on the edge weights between nodes and line impedance compute nodes and Propagation delay Perform timing calibration to obtain the calibrated wavefront time. ; Based on calibrated wavehead time Combined with failure probability and initial positioning Perform distributed weighted localization optimization to determine the coordinates of the fault point. ; Calculate the localization contribution weight for each node based on the calibrated wavefront time. and initial positioning Optimize the global coordinates of the fault point ; Introducing edge weights Correct global coordinates of the fault point The final coordinates of the fault point are obtained. .

[0033] Example 8 is a further improvement on Example 7, and its details are as follows: Step S03 specifically includes the following steps: Step S31: Distributed traveling wave positioning utilizes wavefront time (Accuracy <10ns, based on Hilbert transform) and global graph To eliminate timing discrepancies among multiple nodes (such as GPS synchronization errors and signal propagation delays), a distributed timing calibration algorithm is employed. Temporal bias estimation: based on inter-node edge weights and line impedance compute nodes and Propagation delay: ; in, To delay the spread, This represents the distance between nodes (edge ​​feature), i.e., the interval length. This is the speed of travel wave propagation (near the speed of light). For interference weights, To store the impedance, Reference impedance; Timing calibration: Calibrate wavefront time: ; in, For the calibrated wavehead time, For nodes The neighborhood group, The edge weights are used; the calibration formula uses topological weighted averaging to eliminate multi-node timing bias, with an error of <5ns; Step S32: Based on the calibrated wavefront time Combined with failure probability and initial positioning Perform distributed weighted localization optimization to determine the global coordinates of the fault point. The specific steps are as follows: Node contribution weight: Calculate the location contribution weight of each node. ; in, For nodes Contribution weight, This represents the probability of failure. For convolutional features, , for The mean and standard deviation (based on all nodes), typically The weighting combines fault probability and feature significance, prioritizing nodes with high confidence. Weighted localization optimization: based on calibrated wavefront time and initial positioning Optimize the coordinates of the fault point: ; in, The coordinates of the fault point (a two-dimensional or one-dimensional location on the line, i.e., the location of the fault point in the global line diagram). For nodes The known coordinates (based on GPS positioning). For the traveling wave velocity, The reference time (earliest wavehead time) is used; optimization is achieved by weighted least squares method, taking into account multi-node topological constraints; Topology constraint correction: To adapt to multi-branch topologies, edge weights are introduced. Corrected location results: ; in, The corrected coordinates of the fault point. For projection onto the edge Position offset, based on topological distance The corrected formula ensures that the fault point is located on the physical line topology.

[0034] Example 9 is a further improvement based on Example 1, and its details are as follows: Step S03 also includes the following steps: Step S33: To improve localization robustness, multi-round iterative fusion is adopted: Iterative optimization: Update node contribution weights in each round: ; in, For the first Weighting For the first Wheel positioning results The standard deviation of the distance is used; after 3-5 iterations, the positioning error is converged to <5m. Fusion Output: Fusion of multi-round results: ; in, The coordinates of the final fault point. For the first Overall weighting. Integration improves positioning stability and adapts to dynamic interference.

[0035] Example 10 is a further improvement based on Example 1, and its details are as follows: Step S04 is as follows: Based on the inspection results, the node weights are updated using the following formula: ; in, For the updated weights, For the original weights, For feedback gain, This is a fault confirmation indication function (1 for confirmation, 0 for non-confirmation); after updating, step S03 is re-executed to optimize subsequent inspections.

[0036] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A fault diagnosis method for a 35kV collection line based on cloud-edge collaboration, characterized in that, Includes the following steps: Step S01: Dynamically acquire high-frequency transient signals, generate optimized time spectrum using dynamic wavelet packet decomposition; fuse multimodal features to construct local graph nodes; perform hybrid encoding compression based on topology-aware edge weights and upload to the cloud; Step S02: The cloud receives compressed data and decompresses it to restore the local map; multiple local maps are aggregated into a global map and feature standardization is performed; signal timing characteristics and line topology are fused to generate enhanced features; high-dimensional features are extracted through a dynamic graph convolutional network, and network parameters are dynamically updated to adapt to topology changes; Output the fault classification results and initial location results; Step S03: Based on the fault probability, initial location distance, global topology map and wavefront time, eliminate timing deviation through multi-node timing calibration, and perform distributed weighted optimization to calculate the node contribution weights to obtain the final fault point coordinates; Step S04: Based on the final fault point coordinates and fault probability, and combined with the global graph, perform priority assessment and dynamic feedback.

2. The method for fault diagnosis of a 35kV collection line based on cloud-edge collaboration according to claim 1, characterized in that, Step S01 specifically includes: High-frequency transient signals are collected by distributed monitoring devices to generate the original signal matrix. ; for the signal matrix Calculate the frequency domain entropy matrix ; The frequency domain entropy matrix Generate dynamic scale factors by combining noise spectrum Based on the scale factor Perform adaptive decomposition to obtain wavelet packet coefficients. ; Multi-scale time-spectrum matrices are generated and optimized through mutual information filtering. The optimized time-frequency spectrum matrix is ​​obtained. Perform multimodal fusion, extract key features of traveling waves, and generate a fused feature vector. ; Constructing a local graph Estimating edge weights based on dynamic edge weights Compressed data is obtained after topology-sensing compression. The compressed data Uploaded to the cloud.

3. The method for fault diagnosis of a 35kV collection line based on cloud-edge collaboration according to claim 2, characterized in that, Step S01 specifically includes the following steps: Step S11: Collect high-frequency transient signals through distributed monitoring devices to generate the original signal matrix. : ； in, The original signal matrix, , , These are the current, voltage, and temperature signal vectors, respectively; Step S12: For the signal matrix The formula for calculating the frequency domain entropy matrix is ​​as follows: ； in, The frequency domain entropy matrix, For frequency band index, frequency band Energy distribution, For frequency variables; The frequency domain entropy matrix The dynamic scaling factor is generated by combining the noise spectrum, as shown in the following formula: ； in, Scale factor; To prevent the denominator from being zero; typical NSR range is 0.1-0.5; According to the scale factor Select the decomposition level and mother wavelet, based on the spectral correlation coefficient. The decomposition formula is: ； in, These are wavelet packet coefficients. It is a current signal. For the conjugate of the mother wavelet, For time, For translation parameters, As a scale parameter, focus on the 1-5MHz frequency band; By filtering with mutual information, a multi-scale time-spectrum matrix is ​​generated: ； in, The time-frequency spectrum matrix, For scale indexing, For time indexing; Optimize the time-spectrum matrix : ； in, The optimized time-frequency spectrum matrix, For mutual information, For frequency components; Step S13: Apply the Hilbert transform The analytic signal is generated using the following formula: ； in, To analyze the signal, The result of the Hilbert transform; Extract the following parameters: Wave head timing: , accuracy <10ns; Energy gradient: ,in The signal energy reflects the intensity of the disturbance; Phase shift: Capture signal phase abrupt changes; Multimodal fusion generates fused feature vectors : ； in, To fuse feature vectors, , , This is the weight matrix. For voltage phase shift, This is the temperature correction vector; weight matrix Through optimization, the formula is as follows: ； in, This is the weight matrix. Signal matrix The covariance matrix, For regularization parameters, It is the Frobenius norm; Step S14: Generate a local undirected graph Node characteristics: ； in, For nodes feature, For nodes The fused feature vector, The wavehead time is given; the edge features are: ； in, For the edge feature, The interval length is... For interference weights; Estimating edge weights using dynamic edge weights: ； in, For border rights, , , For coefficients, Due to time difference, To store the impedance, The covariance of the fusion features; Prioritize compressing high-weight edges, using hybrid encoding. The compression formula is as follows: ； in, To compress the data, Compression ratio; .

4. The method for fault diagnosis of a 35kV collection line based on cloud-edge collaboration according to claim 3, characterized in that, Step S11 further includes: For the signal matrix Perform signal complexity evaluation: ； in, This represents the normalized probability density. Joint entropy combined with real-time environmental noise spectrum generates triggering factors: ； in, The trigger factor is NSR, which is the noise-to-signal ratio. when Exceeding the dynamic threshold When triggered, it switches to high-frequency sampling mode, and under normal circumstances, it maintains low-frequency sampling.

5. The method for fault diagnosis of a 35kV collection line based on cloud-edge collaboration according to claim 1, characterized in that, Step S02 specifically includes: The cloud receives the compressed data Local images are restored through reverse encoding. ; partial diagrams of multiple edge devices Aggregate into a global graph ; Based on global graph Calculate the time attention weight vector Spatial attention weight vector The two are fused to generate hybrid attention features. ; The hybrid attention features Input dynamic graph convolutional network, convolutional features Classification probabilities are generated through fully connected layers. Predicting the initial location of the fault point using a regression model. .

6. The method for fault diagnosis of a 35kV collection line based on cloud-edge collaboration according to claim 5, characterized in that, Step S02 specifically includes the following steps: Step S21: The cloud receives the compressed data. Local images are restored through reverse encoding. , among which node features Edge features and border rights ; partial diagrams of multiple edge devices Aggregate into a global graph : ； in, The set of features for all nodes. It is the set of all edge features; Through initial standardization: ； in, For standardized node features, , for The mean and standard deviation; Step S22: For the standardized node features and wave head time The time attention weights are calculated using the following formula: ； in, For nodes Temporal attention vector, , , For queries, key, and value vectors; , , For learnable weight matrix, The dimension of the key vector. This is the normalization function; Spatial Attention: Based on Global Graph Calculate the spatial attention weights: ； in, For nodes For nodes Spatial attention weights, , , , , , For learnable weight matrix, For border rights, ; Fuse temporal attention vectors and spatial attention weights to generate enhanced features: ； in, For nodes Hybrid attention features , These are the weighting coefficients. For nodes The set of neighbors; Step S23: Move the node Hybrid attention features Input dynamic graph convolutional network: ； in, For the first Layer nodes The convolutional features, It is the ReLU activation function. For the first Layer weight matrix, For bias, For border rights; Border rights Updated every minute: ； in, For a moment The right to the border, For update rate; This represents the increment of edge weights; Convolutional features Generating classification probabilities through fully connected layers: ； in, For nodes The probability of failure, , These are the parameters for the fully connected layer; Location is determined by predicting the distance to the fault point using a regression model: ； in, For initial location, it represents the distance from the fault point to the node. The distance; , For regression parameters; Step S24: HADM uses a joint loss function to optimize classification and localization: ； in, For cross-entropy loss, Mean square error, , These are the weighting coefficients.

7. The method for fault diagnosis of a 35kV collection line based on cloud-edge collaboration according to claim 1, characterized in that, Step S03 specifically includes: Based on the edge weights between nodes and line impedance compute nodes and Propagation delay Perform timing calibration to obtain the calibrated wavefront time. ; Based on calibrated wavehead time Combined with failure probability and initial positioning Perform distributed weighted localization optimization to determine the coordinates of the fault point. ; Calculate the positioning contribution weight for each node based on the calibrated wavefront time. and the initial positioning Optimize the global coordinates of the fault point ; Introducing edge weights Correct the global coordinates of the fault point The final coordinates of the fault point are obtained. .

8. The method for fault diagnosis of a 35kV collection line based on cloud-edge collaboration according to claim 7, characterized in that, Step S03 specifically includes the following steps: Step S31: Based on the edge weights between nodes and line impedance compute nodes and Propagation delay: ； in, To delay the spread, The distance between nodes. For the speed of travel wave propagation, For interference weights, To store the impedance, Reference impedance; Calibrate wavehead time: ； in, For the calibrated wavehead time, For nodes The neighborhood group, For border rights; Step S32: Calculate the positioning contribution weight of each node: ； in, For nodes Contribution weight, This represents the probability of failure. For convolutional features, , for The mean and standard deviation; Based on calibrated wavehead time and initial positioning Optimize the coordinates of the fault point: ； in, The coordinates of the fault point, For nodes The known coordinates, For the traveling wave velocity, For reference only; Introducing edge weights Corrected location results: ； in, The corrected coordinates of the fault point. For projection onto the edge Position offset.

9. A method for fault diagnosis of a 35kV collection line based on cloud-edge collaboration as described in claim 8, characterized in that, Step S03 further includes the following steps: Step S33: Update node contribution weights in each round: ； in, For the first Weighting For the first Wheel positioning results The standard deviation of the distance is used; after 3-5 iterations, the positioning error is converged to <5m. Integrating results from multiple rounds: ； in, The coordinates of the final fault point. For the first Total weight of rounds.

10. A method for fault diagnosis of a 35kV collection line based on cloud-edge collaboration according to claim 1, characterized in that, Step S04 specifically involves: Based on the inspection results, the node weights are updated using the following formula: ； in, For the updated weights, For the original weights, For feedback gain, This is a fault confirmation indication function; after updating, step S03 is re-executed to optimize subsequent inspections.

Images