An edge-computing-based power distribution network fault detection method, system, device and medium

Abstract

The application discloses a power distribution network fault detection method and system based on edge computing, equipment and medium, mainly relates to the technical field of fault detection, in order to solve the problem that the existing scheme has low implicit fault recognition rate, and the diagnosis model generally adopts static topology structure, which is difficult to support the precise diagnosis demand of complex power distribution network. Including: taking the multi-dimensional operation characteristic data output at each sampling time as the node attribute, constructing the edge based on the electrical quantity time series similarity and the topological correlation degree, and generating the local operation state atlas of the power distribution network; input the atlas into the lightweight graph attention network, complete the aggregation coding of node features and global topological features, and output the graph neural network coding features; the graph neural network coding is mapped to the unified embedding space through the pre-trained contrast learning encoder, and the current operation state embedding vector is generated; the cosine similarity of the vector and the embedding vector of each fault mode in the fault sample library is calculated, and the fault type and the matching similarity are output.

Description

Technical Field

[0001] This application relates to the field of fault detection technology, and in particular to a method, system, device and medium for fault detection in power distribution networks based on edge computing. Background Technology

[0002] Currently, fault detection in power distribution networks generally relies on a centralized cloud architecture, requiring fault data to be uploaded to a main station for centralized analysis via communication links. This model is limited by network transmission latency (typically 8–15 seconds), communication link stability, and bandwidth constraints. This results in high-frequency electrical transient characteristics (such as traveling waves and harmonic abrupt changes) being downsampled or lost during the upload process, failing to meet the real-time requirements of millisecond-level response in power distribution networks. Especially in remote or poorly communicated areas, intermittent data uploads further exacerbate fault identification delays, causing fault isolation and self-healing operations to be delayed by more than 30 seconds, severely impacting power supply reliability.

[0003] Existing methods often rely on single electrical quantity thresholds for judgment (such as current amplitude and zero-sequence current), resulting in low identification rates for latent faults such as high-resistance grounding and combined open-circuit and grounding faults. Furthermore, their diagnostic models generally employ static topologies, failing to dynamically adapt to changes in operating modes such as distributed power source switching and network reconfiguration. This leads to frequent misjudgments of fault propagation paths and a persistently high false alarm rate, making it difficult to support the accurate diagnostic needs of complex distribution networks. Summary of the Invention

[0004] This application provides a method, system, device, and medium for distribution network fault detection based on edge computing, in order to solve the problems that existing solutions rely on a single electrical quantity threshold for judgment, resulting in low identification rate of hidden faults such as high-resistance grounding and open-circuit-grounding composite faults, and that their diagnostic models generally adopt static topology structures, which are difficult to support the accurate diagnostic needs of complex distribution networks.

[0005] Firstly, this application provides a method for fault detection in a distribution network based on edge computing, the method comprising: Edge computing nodes collect multi-dimensional operational data from power distribution network terminals in real time. This multi-dimensional operational data includes voltage and current time-series waveforms, switch action sequences, equipment temperature and harmonic parameters. Combined with the terminal's historical normal operation dataset, electrical quantity baselines and equipment status baselines are constructed respectively. The standardized deviation of the current operational data from the normal baseline is output. At the same time, electrical feature sub-vectors and equipment status feature sub-vectors of the current sampling period are extracted, and the cross-layer consistency score of the two types of features is calculated. When the standardized deviation and cross-layer consistency score meet the preset trigger conditions, the N consecutive sampling times within the current detection window are taken as nodes, and the multi-dimensional operation feature data output at each sampling time is taken as node attributes. Edges are constructed based on the electrical quantity temporal similarity and topological correlation between nodes to generate a local operation status map of the distribution network. The map is then input into a lightweight graph attention network to complete the aggregation encoding of node features and global topological features, and output the graph neural network encoded features. The graph neural network encoding is mapped to a unified embedding space by a pre-trained contrastive learning encoder to generate the current running state embedding vector; the cosine similarity between this vector and the embedding vectors of various fault modes in the fault sample library is calculated, and the fault type and matching similarity are output.

[0006] In one implementation of this application, the edge computing node collects multi-dimensional operational data from the distribution network terminal in real time; combining the terminal's historical normal operation dataset, it constructs an electrical quantity baseline and an equipment status baseline, and outputs the standardized deviation of the current operational data relative to the normal baseline, specifically including: Based on the voltage and current timing waveforms and switching action sequences in the historical normal operation dataset of the terminal, timing alignment and amplitude normalization preprocessing are performed. Using a sliding time window statistical method, the mean, standard deviation, and upper and lower limits of the 95% confidence interval of preset key electrical quantity characteristics are extracted window by window from the voltage and current time-series waveforms and switching action sequences to form a baseline parameter set of electrical quantities. At the same time, a time-series correlation matrix is ​​established for each electrical quantity to record the coupling relationship between multiple electrical quantities under normal operating conditions. An electrical quantity baseline model containing the normal fluctuation range of each preset key electrical quantity characteristic and the coupling relationship matrix is ​​obtained. Using the historical sampling timestamps corresponding to the electrical quantity baseline model as indexes, equipment temperature and harmonic parameters are synchronously extracted from the historical normal operation dataset and hierarchically clustered according to seasonal periods and load levels. For the equipment temperature data under each type of operating condition, the kernel density estimation method is used to fit the probability distribution curve to determine the normal temperature range and temperature rise rate threshold. The normal upper limit values ​​of each harmonic content rate and total harmonic distortion rate are calculated to construct the equipment status baseline parameter set. The equipment status baseline model containing temperature thresholds and harmonic limits is obtained. The multidimensional operating data of the current sampling period are input into the electrical quantity baseline model and the equipment status baseline model respectively, and the standardized deviation of the current value relative to the baseline interval is calculated dimension by dimension.

[0007] In one implementation of this application, the electrical feature sub-vector and the equipment state feature sub-vector of the current sampling period are extracted, and the cross-layer consistency score of the two types of features is calculated, specifically including: For the voltage and current timing waveforms and switching action sequences of the current sampling period, core electrical features such as voltage RMS deviation, current transient peak value, switching action interval, and fundamental phase difference are extracted through time-domain statistical analysis and fast Fourier transform. After standardizing and normalizing each feature, they are spliced ​​together in a preset dimension order to generate fixed-dimensional electrical feature sub-vectors. Based on the electrical feature sub-vectors and the preset electrical analysis rules, the corresponding first fault analysis vector is determined; Using a unified timestamp identifier as the alignment benchmark, the equipment temperature and harmonic parameter data within the same sampling period are extracted, and equipment status features such as temperature gradient, total harmonic distortion rate, and characteristic subharmonic content rate are extracted. The numerical mapping is completed using the normalization rule consistent with the electrical feature subvector, and the equipment status feature subvector is generated by splicing according to the corresponding dimensions. Based on the equipment status feature sub-vector and the preset equipment analysis rules, the corresponding second fault analysis vector is determined; The first fault analysis vector and the second fault analysis vector are input into the cross-layer consistency calculation unit to obtain a cross-layer consistency score with a value range of 0 to 1.

[0008] In one implementation of this application, the standardized deviation and cross-layer consistency score satisfy a preset trigger condition, specifically including: A pre-defined hierarchical threshold system for standardized deviations is established, including single-dimensional fault thresholds and multi-dimensional early warning thresholds. Each standardized deviation is matched with the corresponding threshold. When the deviation of the core electrical quantity dimension exceeds the single-dimensional fault threshold, or when the deviation of at least two secondary dimensions simultaneously exceeds the multi-dimensional early warning threshold, a preliminary deviation trigger flag is generated. The preliminary judgment result is jointly verified with the cross-layer consistency score. If there is a preliminary deviation from the trigger flag and the cross-layer consistency score is lower than the preset consistency threshold, the preset trigger condition is determined to be met; otherwise, the preset trigger condition is not met.

[0009] In one implementation of this application, N consecutive sampling times within the current detection window are used as nodes, and the multi-dimensional operational feature data output at each sampling time is used as node attributes. Edges are constructed based on the temporal similarity of electrical quantities and topological correlation between nodes to generate a local operational status map of the distribution network, specifically including: N consecutive sampling times within the current detection window are extracted as graph nodes. The standardized deviation, electrical feature sub-vector, and equipment status feature sub-vector corresponding to each sampling time are concatenated to generate a node attribute vector with a unified dimension. All nodes are sequentially numbered according to the sampling time sequence, and a set of node attributes with time sequence index is output. Using the set of node attributes as input, the cosine similarity of electrical feature vectors between nodes is calculated for each pair to obtain the temporal similarity weight; at the same time, the preset topology connection matrix of the distribution network terminal is retrieved to assign topology correlation weight to the node pairs with direct electrical connections; the two types of weights are weighted and merged into a comprehensive edge weight, and node pairs with weights higher than a preset threshold are selected to form an effective edge set, and the weighted edge connection relationship set is output. The set of node attributes and the set of edge connections are input into the graph generation unit. The nodes at the sampling time are used as vertices and the effective edges are used as undirected connecting edges. The graph structure is assembled according to temporal association and topological constraints to generate a local operation status graph of the distribution network that contains multi-dimensional node features and edge weight information.

[0010] In one implementation of this application, a graph neural network encoding is mapped to a unified embedding space using a pre-trained contrastive learning encoder to generate a current running state embedding vector; the cosine similarity between this vector and the embedding vectors of various fault modes in the fault sample library is calculated, and the fault type and matching similarity are output, specifically including: The graph neural network encoded features output by the lightweight graph attention network are input into the pre-trained contrastive learning encoder. The encoder performs nonlinear transformation and dimensionality compression on the high-dimensional encoded features through two fully connected perceptron layers, mapping them to a unified embedding space with fixed dimensions. The mapped feature vector is then subjected to L2 normalization to generate a current running state embedding vector with normalized modulus, which is then output to the similarity matching module. Using the current running state embedding vector as input, retrieve the embedding vectors of various fault modes pre-generated by the isomorphic encoder in the fault sample library; calculate the cosine similarity between the current embedding vector and each fault mode embedding vector for each class, obtain the similarity value for each type of fault, sort them in descending order of similarity to generate a fault matching sorting list, and output a matching result set containing fault type identifiers and corresponding similarity values. Input the fault matching sorting list into the judgment unit and compare the highest similarity value with the preset similarity threshold. If the highest similarity value is not lower than the preset threshold, output the corresponding fault type and matching similarity as the final detection result. If the highest similarity value is lower than the preset threshold, it is determined that the known fault mode is not matched, and the unknown anomaly identifier and the current highest similarity value are output.

[0011] Secondly, this application provides a power distribution network fault detection system based on edge computing, the system comprising: The computing module is used for edge computing nodes to collect multi-dimensional operating data of distribution network terminals in real time. The multi-dimensional operating data includes voltage and current time-series waveforms, switch action sequences, equipment temperature and harmonic parameters. Combined with the terminal's historical normal operation dataset, electrical quantity baselines and equipment status baselines are constructed respectively. The standardized deviation of the current operating data from the normal baseline is output. At the same time, the electrical feature sub-vectors and equipment status feature sub-vectors of the current sampling period are extracted, and the cross-layer consistency score of the two types of features is calculated. The output module is used to generate a local operating status map of the distribution network when the standardized deviation and cross-layer consistency score meet the preset trigger conditions. It takes N consecutive sampling times within the current detection window as nodes and the multi-dimensional operating feature data output at each sampling time as node attributes. It constructs edges based on the time series similarity of electrical quantities and topological correlation between nodes. The map is then input into a lightweight graph attention network to complete the aggregation encoding of node features and global topological features, and outputs the graph neural network encoded features. The matching module is used to map the graph neural network encoding to a unified embedding space through a pre-trained contrastive learning encoder, generating the current running state embedding vector; it calculates the cosine similarity between this vector and the embedding vectors of various fault modes in the fault sample library, and outputs the fault type and matching similarity.

[0012] In one implementation of this application, the computing module includes a computing unit. Used for timing alignment and amplitude normalization preprocessing of voltage and current timing waveforms and switching action sequences based on the historical normal operation dataset of the terminal; Using a sliding time window statistical method, the mean, standard deviation, and upper and lower limits of the 95% confidence interval of preset key electrical quantity characteristics are extracted window by window from the voltage and current time-series waveforms and switching action sequences to form a baseline parameter set of electrical quantities. At the same time, a time-series correlation matrix is ​​established for each electrical quantity to record the coupling relationship between multiple electrical quantities under normal operating conditions. An electrical quantity baseline model containing the normal fluctuation range of each preset key electrical quantity characteristic and the coupling relationship matrix is ​​obtained. Using the historical sampling timestamps corresponding to the electrical quantity baseline model as indexes, equipment temperature and harmonic parameters are synchronously extracted from the historical normal operation dataset and hierarchically clustered according to seasonal periods and load levels. For the equipment temperature data under each type of operating condition, the kernel density estimation method is used to fit the probability distribution curve to determine the normal temperature range and temperature rise rate threshold. The normal upper limit values ​​of each harmonic content rate and total harmonic distortion rate are calculated to construct the equipment status baseline parameter set. The equipment status baseline model containing temperature thresholds and harmonic limits is obtained. The multidimensional operating data of the current sampling period are input into the electrical quantity baseline model and the equipment status baseline model respectively, and the standardized deviation of the current value relative to the baseline interval is calculated dimension by dimension.

[0013] Thirdly, this application provides a power distribution network fault detection device based on edge computing, the device comprising: processor; And a memory containing executable code, which, when executed, causes the processor to perform a power distribution network fault detection method based on edge computing, as described above.

[0014] Fourthly, this application provides a non-volatile computer storage medium storing computer instructions thereon, which, when executed, implement a power distribution network fault detection method based on edge computing as described above.

[0015] As can be seen from the above technical solutions, this application has the following advantages: This solution overcomes the limitations of traditional single electrical quantity threshold judgment by acquiring multi-dimensional operational data such as voltage and current timing waveforms, switching action sequences, equipment temperature, and harmonic parameters in real time through edge computing nodes, and constructing electrical quantity baselines and equipment status baselines respectively. The calculation of standardized deviations effectively captures multi-parameter coordinated anomalies, while cross-layer consistency scores reveal the inherent correlation between electrical characteristics and equipment status, enabling the systematic identification of latent faults such as high-resistance grounding and open-circuit-grounding composite faults even without obvious amplitude changes. This mechanism does not rely on the absolute magnitude of the fault current, but rather on the abnormal coordinated patterns of multi-dimensional features for discrimination, improving the detection sensitivity and coverage of low-amplitude, atypical faults, and solving the core pain point of existing methods that are "invisible and unrecognizable" for latent faults.

[0016] Using continuous sampling times as nodes and multidimensional features as attributes, a local operating state graph is dynamically constructed based on the temporal similarity of electrical quantities and the correlation with the actual network topology, thus overcoming the inherent defects of static topology structures. When distributed power source switching, line reconfiguration, or load transfer causes changes in the network structure, the edge weights and node connections of the graph can be reconstructed in real time, ensuring that the modeling of fault propagation paths remains consistent with the current operating conditions. The lightweight graph attention network further aggregates local node features with the global topology, so that the inference of the fault impact range no longer depends on a preset fixed topology model, thereby reducing misjudgments of fault propagation paths caused by mismatched network structures.

[0017] By using a pre-trained contrastive learning encoder, the encoded features output by the graph neural network are mapped to a unified embedding space and matched with cosine similarity of the embedding vectors of various fault modes labeled in the fault sample library. This scheme achieves end-to-end semantic recognition of fault types. This mechanism does not rely on manually set classification rules or thresholds, but instead learns the distribution structure of fault modes in a high-dimensional space to automatically distinguish similar but fundamentally different fault types (such as transient flashover and continuous high-resistance grounding), suppressing false alarms caused by feature confusion. Simultaneously, the generalization capability of the embedding space allows the system to stably adapt to newly emerging fault modes without frequent model retraining, reducing the false alarm rate and improving the interpretability and engineering reliability of diagnostic results. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a power distribution network fault detection method based on edge computing provided in an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the internal structure of a power distribution network fault detection system based on edge computing, provided in an embodiment of this application.

[0021] Figure 3 This is a schematic diagram of the internal structure of a power distribution network fault detection device based on edge computing, provided in an embodiment of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of this disclosure and do not imply that this disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely used to explain the technical principles of this disclosure and are not intended to limit the scope of protection of this disclosure. Based on the preferred embodiments provided by this disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of this disclosure.

[0024] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0025] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0026] The embodiment provides a distribution network fault detection method based on edge computing, such as Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps: Step 110: Edge computing nodes collect multi-dimensional operation data of distribution network terminals in real time; the multi-dimensional operation data includes voltage and current time-series waveforms, switch action sequences, equipment temperature and harmonic parameters; combined with the terminal's historical normal operation dataset, electrical quantity baseline and equipment status baseline are constructed respectively, and the standardized deviation of the current operation data relative to the normal baseline is output. At the same time, the electrical feature sub-vector and equipment status feature sub-vector of the current sampling period are extracted, and the cross-layer consistency score of the two types of features is calculated.

[0027] In some embodiments, edge computing nodes collect multi-dimensional operational data from distribution network terminals in real time; combining this data with historical normal operation datasets of the terminals, electrical quantity baselines and equipment status baselines are constructed respectively, and the standardized deviation of the current operational data from the normal baselines is output, specifically including: Based on the voltage and current timing waveforms and switching action sequences in the historical normal operation dataset of the terminal, timing alignment and amplitude normalization preprocessing are performed. Using a sliding time window statistical method, the mean, standard deviation, and upper and lower limits of the 95% confidence interval of preset key electrical quantity characteristics are extracted window by window from the voltage and current time-series waveforms and switching action sequences to form a baseline parameter set of electrical quantities. At the same time, a time-series correlation matrix is ​​established for each electrical quantity to record the coupling relationship between multiple electrical quantities under normal operating conditions. An electrical quantity baseline model containing the normal fluctuation range of each preset key electrical quantity characteristic and the coupling relationship matrix is ​​obtained. Using the historical sampling timestamps corresponding to the electrical quantity baseline model as indexes, equipment temperature and harmonic parameters are synchronously extracted from the historical normal operation dataset and hierarchically clustered according to seasonal periods and load levels. For the equipment temperature data under each type of operating condition, the kernel density estimation method is used to fit the probability distribution curve to determine the normal temperature range and temperature rise rate threshold. The normal upper limit values ​​of each harmonic content rate and total harmonic distortion rate are calculated to construct the equipment status baseline parameter set. The equipment status baseline model containing temperature thresholds and harmonic limits is obtained. The multidimensional operating data of the current sampling period are input into the electrical quantity baseline model and the equipment status baseline model respectively, and the standardized deviation of the current value relative to the baseline interval is calculated dimension by dimension.

[0028] It should be further explained that edge computing nodes, by constructing electrical quantity baselines and equipment status baselines, enable localized comparison of multi-dimensional operational data of distribution network terminals. This can effectively identify the coupling relationship between abnormal electrical parameters and deviations in equipment status, reducing the risk of misjudgment caused by fluctuations in a single parameter. The dimensional calculation of standardized deviations provides quantifiable and traceable criteria for judging abnormal signals, improving the reliability of early fault detection. The introduction of cross-layer feature consistency scores helps to distinguish between actual equipment degradation and external interference. For example, when the temperature is abnormal but the electrical characteristics are normal, the impact of sudden load changes can be ruled out, enhancing the physical interpretability of diagnostic results, reducing the frequency of invalid alarms, and supporting maintenance personnel in accurately locating the root cause of problems.

[0029] In some embodiments, the electrical feature sub-vector and the device state feature sub-vector of the current sampling period are extracted, and the cross-layer consistency score of the two types of features is calculated, specifically including: For the voltage and current timing waveforms and switching action sequences of the current sampling period, core electrical features such as voltage RMS deviation, current transient peak value, switching action interval, and fundamental phase difference are extracted through time-domain statistical analysis and fast Fourier transform. After standardizing and normalizing each feature, they are spliced ​​together in a preset dimension order to generate fixed-dimensional electrical feature sub-vectors. Based on the electrical feature sub-vectors and the preset electrical analysis rules, the corresponding first fault analysis vector is determined; Using a unified timestamp identifier as the alignment benchmark, the equipment temperature and harmonic parameter data within the same sampling period are extracted, and equipment status features such as temperature gradient, total harmonic distortion rate, and characteristic subharmonic content rate are extracted. The numerical mapping is completed using the normalization rule consistent with the electrical feature subvector, and the equipment status feature subvector is generated by splicing according to the corresponding dimensions. Based on the equipment status feature sub-vector and the preset equipment analysis rules, the corresponding second fault analysis vector is determined; The first fault analysis vector and the second fault analysis vector are input into the cross-layer consistency calculation unit to obtain a cross-layer consistency score with a value range of 0 to 1.

[0030] Step 120: When the standardized deviation and cross-layer consistency score meet the preset trigger conditions, take the N consecutive sampling times within the current detection window as nodes, and the multi-dimensional operation feature data output at each sampling time as node attributes. Construct edges based on the electrical quantity temporal similarity and topological correlation between nodes to generate a local operation status map of the distribution network. Input the map into a lightweight graph attention network to complete the aggregation encoding of node features and global topological features, and output the graph neural network encoded features.

[0031] In some embodiments, the standardized deviation and cross-layer consistency score meet preset triggering conditions, specifically including: A pre-defined hierarchical threshold system for standardized deviations is established, including single-dimensional fault thresholds and multi-dimensional early warning thresholds. Each standardized deviation is matched with the corresponding threshold. When the deviation of the core electrical quantity dimension exceeds the single-dimensional fault threshold, or when the deviation of at least two secondary dimensions simultaneously exceeds the multi-dimensional early warning threshold, a preliminary deviation trigger flag is generated. The preliminary judgment result is jointly verified with the cross-layer consistency score. If there is a preliminary deviation from the trigger flag and the cross-layer consistency score is lower than the preset consistency threshold, the preset trigger condition is determined to be met; otherwise, the preset trigger condition is not met.

[0032] In some embodiments, N consecutive sampling times within the current detection window are used as nodes, and the multidimensional operational feature data output at each sampling time is used as node attributes. Edges are constructed based on the temporal similarity of electrical quantities and topological correlation between nodes to generate a local operational status map of the distribution network. Specifically, this includes: N consecutive sampling times within the current detection window are extracted as graph nodes. The standardized deviation, electrical feature sub-vector, and equipment status feature sub-vector corresponding to each sampling time are concatenated to generate a node attribute vector with a unified dimension. All nodes are sequentially numbered according to the sampling time sequence, and a set of node attributes with time sequence index is output. Using the set of node attributes as input, the cosine similarity of electrical feature vectors between nodes is calculated for each pair to obtain the temporal similarity weight; at the same time, the preset topology connection matrix of the distribution network terminal is retrieved to assign topology correlation weight to the node pairs with direct electrical connections; the two types of weights are weighted and merged into a comprehensive edge weight, and node pairs with weights higher than a preset threshold are selected to form an effective edge set, and the weighted edge connection relationship set is output. The set of node attributes and the set of edge connections are input into the graph generation unit. The nodes at the sampling time are used as vertices and the effective edges are used as undirected connecting edges. The graph structure is assembled according to temporal association and topological constraints to generate a local operation status graph of the distribution network that contains multi-dimensional node features and edge weight information.

[0033] It should be further explained that when the standardized deviation and cross-layer consistency score meet the triggering conditions, the constructed local operating status map of the distribution network integrates the temporal dynamic characteristics and physical topology relationships into structured graph data, enabling node attributes to carry multi-dimensional state information and edge weights to reflect electrical associations and temporal evolution characteristics. The lightweight graph attention network can effectively aggregate the features of neighboring nodes and the global topology under limited computing resources. The output encoded features retain the local anomaly propagation path and coupling mode, improve the structural sensitivity of fault mode recognition, avoid misjudgments caused by isolated node deviations, and provide physically meaningful low-dimensional representation inputs for subsequent diagnostic modules, reducing the cloud transmission and processing burden.

[0034] Step 130: Map the graph neural network encoding to a unified embedding space using a pre-trained contrastive learning encoder to generate the current running state embedding vector; calculate the cosine similarity between this vector and the embedding vectors of various fault modes in the fault sample library, and output the fault type and matching similarity.

[0035] In some embodiments, a pre-trained contrastive learning encoder maps the graph neural network encoding to a unified embedding space to generate a current running state embedding vector; the cosine similarity between this vector and the embedding vectors of various fault modes in the fault sample library is calculated, and the fault type and matching similarity are output, specifically including: The graph neural network encoded features output by the lightweight graph attention network are input into the pre-trained contrastive learning encoder. The encoder performs nonlinear transformation and dimensionality compression on the high-dimensional encoded features through two fully connected perceptron layers, mapping them to a unified embedding space with fixed dimensions. The mapped feature vector is then subjected to L2 normalization to generate a current running state embedding vector with normalized modulus, which is then output to the similarity matching module. Using the current running state embedding vector as input, retrieve the embedding vectors of various fault modes pre-generated by the isomorphic encoder in the fault sample library; calculate the cosine similarity between the current embedding vector and each fault mode embedding vector for each class, obtain the similarity value for each type of fault, sort them in descending order of similarity to generate a fault matching sorting list, and output a matching result set containing fault type identifiers and corresponding similarity values. Input the fault matching sorting list into the judgment unit and compare the highest similarity value with the preset similarity threshold. If the highest similarity value is not lower than the preset threshold, output the corresponding fault type and matching similarity as the final detection result. If the highest similarity value is lower than the preset threshold, it is determined that the known fault mode is not matched, and the unknown anomaly identifier and the current highest similarity value are output.

[0036] It should be further explained that by mapping the graph neural network encoding to a unified embedding space through a pre-trained contrastive learning encoder, the comparability of features of different fault modes under the same semantic dimension is achieved, giving the cosine similarity calculation a physically meaningful matching basis; L2 normalization ensures the consistency of vector scale and avoids similarity deviation caused by differences in feature amplitude; the pre-stored embedding vectors in the fault sample library are derived from historical real fault records, and the matching process directly reflects the structural similarity between the current state and known fault modes, improving the certainty of fault type identification; when the threshold is not reached, an unknown anomaly label is output to avoid misclassification and provide a clear direction for subsequent manual review.

[0037] In addition, this application Figure 2 This application provides an edge computing-based power distribution network fault detection system as an embodiment. Figure 2 As shown, the system provided in this application embodiment mainly includes: The calculation module 210 is used for edge computing nodes to collect multi-dimensional operating data of distribution network terminals in real time. The multi-dimensional operating data includes voltage and current time-series waveforms, switch action sequences, equipment temperature and harmonic parameters. Combined with the terminal's historical normal operation dataset, electrical quantity baselines and equipment status baselines are constructed respectively. The standardized deviation of the current operating data from the normal baseline is output. At the same time, the electrical feature sub-vectors and equipment status feature sub-vectors of the current sampling period are extracted, and the cross-layer consistency scores of the two types of features are calculated.

[0038] The computing module 210 includes computing units. Used for timing alignment and amplitude normalization preprocessing of voltage and current timing waveforms and switching action sequences based on the historical normal operation dataset of the terminal; Using a sliding time window statistical method, the mean, standard deviation, and upper and lower limits of the 95% confidence interval of preset key electrical quantity characteristics are extracted window by window from the voltage and current time-series waveforms and switching action sequences to form a baseline parameter set of electrical quantities. At the same time, a time-series correlation matrix is ​​established for each electrical quantity to record the coupling relationship between multiple electrical quantities under normal operating conditions. An electrical quantity baseline model containing the normal fluctuation range of each preset key electrical quantity characteristic and the coupling relationship matrix is ​​obtained. Using the historical sampling timestamps corresponding to the electrical quantity baseline model as indexes, equipment temperature and harmonic parameters are synchronously extracted from the historical normal operation dataset and hierarchically clustered according to seasonal periods and load levels. For the equipment temperature data under each type of operating condition, the kernel density estimation method is used to fit the probability distribution curve to determine the normal temperature range and temperature rise rate threshold. The normal upper limit values ​​of each harmonic content rate and total harmonic distortion rate are calculated to construct the equipment status baseline parameter set. The equipment status baseline model containing temperature thresholds and harmonic limits is obtained. The multidimensional operating data of the current sampling period are input into the electrical quantity baseline model and the equipment status baseline model respectively, and the standardized deviation of the current value relative to the baseline interval is calculated dimension by dimension.

[0039] The output module 220 is used to generate a local operating status map of the distribution network when the standardized deviation and cross-layer consistency score meet the preset trigger conditions. It takes N consecutive sampling times within the current detection window as nodes and the multi-dimensional operating feature data output at each sampling time as node attributes. It constructs edges based on the time series similarity of electrical quantities and topological correlation between nodes. The map is then input into a lightweight graph attention network to complete the aggregation encoding of node features and global topological features, and outputs graph neural network encoded features.

[0040] The matching module 230 is used to map the graph neural network encoding to a unified embedding space through a pre-trained contrastive learning encoder to generate the current running state embedding vector; calculate the cosine similarity between this vector and the embedding vectors of various fault modes in the fault sample library, and output the fault type and matching similarity.

[0041] The above are method embodiments of this application. Based on the same inventive concept, this application also provides a power distribution network fault detection device based on edge computing. Figure 3 As shown, the device includes: a processor; and a memory storing executable code thereon, which, when executed, causes the processor to perform a power distribution network fault detection method based on edge computing as described in the above embodiments.

[0042] Specifically, the server collects multi-dimensional operation data of the distribution network terminal in real time through edge computing nodes. The multi-dimensional operation data includes voltage and current time-series waveforms, switch action sequences, equipment temperature and harmonic parameters. Combined with the terminal's historical normal operation dataset, electrical quantity baselines and equipment status baselines are constructed respectively. The standardized deviation of the current operation data from the normal baseline is output. At the same time, the electrical feature sub-vector and equipment status feature sub-vector of the current sampling period are extracted, and the cross-layer consistency score of the two types of features is calculated. When the standardized deviation and cross-layer consistency score meet the preset trigger conditions, the distribution network local operation status map is generated by taking N consecutive sampling times within the current detection window as nodes and the multi-dimensional operation feature data output at each sampling time as node attributes, constructing edges based on the electrical quantity temporal similarity and topological correlation between nodes; the map is input into a lightweight graph attention network to complete the aggregation encoding of node features and global topological features, and outputting graph neural network encoded features; the graph neural network encoding is mapped to a unified embedding space through a pre-trained contrastive learning encoder to generate the current operation status embedding vector; the cosine similarity between this vector and the embedding vectors of various fault modes in the fault sample library is calculated, and the fault type and matching similarity are output.

[0043] In addition, this application embodiment also provides a non-volatile computer storage medium storing executable instructions, which, when executed, implement the above-described edge computing-based power distribution network fault detection method.

[0044] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for fault detection in a distribution network based on edge computing, characterized in that, The method includes: Edge computing nodes collect multi-dimensional operational data from power distribution network terminals in real time. This multi-dimensional operational data includes voltage and current time-series waveforms, switch action sequences, equipment temperature and harmonic parameters. Combined with the terminal's historical normal operation dataset, electrical quantity baselines and equipment status baselines are constructed respectively. The standardized deviation of the current operational data from the normal baseline is output. At the same time, electrical feature sub-vectors and equipment status feature sub-vectors of the current sampling period are extracted, and the cross-layer consistency score of the two types of features is calculated. When the standardized deviation and cross-layer consistency score meet the preset trigger conditions, the N consecutive sampling times within the current detection window are taken as nodes, and the multi-dimensional operation feature data output at each sampling time is taken as node attributes. Edges are constructed based on the electrical quantity temporal similarity and topological correlation between nodes to generate a local operation status map of the distribution network. The map is then input into a lightweight graph attention network to complete the aggregation encoding of node features and global topological features, and output the graph neural network encoded features. The graph neural network encoding is mapped to a unified embedding space by a pre-trained contrastive learning encoder to generate the current running state embedding vector; the cosine similarity between this vector and the embedding vectors of various fault modes in the fault sample library is calculated, and the fault type and matching similarity are output.

2. The distribution network fault detection method based on edge computing according to claim 1, characterized in that, Edge computing nodes collect multi-dimensional operational data from power distribution network terminals in real time. By combining the historical normal operation dataset of the terminal, electrical quantity baselines and equipment status baselines are constructed respectively. The standardized deviation of the current operating data from the normal baseline is output, specifically including: Based on the voltage and current timing waveforms and switching action sequences in the historical normal operation dataset of the terminal, timing alignment and amplitude normalization preprocessing are performed. Using a sliding time window statistical method, the mean, standard deviation, and upper and lower limits of the 95% confidence interval of preset key electrical quantity characteristics are extracted window by window from the voltage and current time-series waveforms and switching action sequences to form a baseline parameter set of electrical quantities. At the same time, a time-series correlation matrix is ​​established for each electrical quantity to record the coupling relationship between multiple electrical quantities under normal operating conditions. An electrical quantity baseline model containing the normal fluctuation range of each preset key electrical quantity characteristic and the coupling relationship matrix is ​​obtained. Using the historical sampling timestamps corresponding to the electrical quantity baseline model as indexes, equipment temperature and harmonic parameters are synchronously extracted from the historical normal operation dataset and hierarchically clustered according to seasonal periods and load levels. For the equipment temperature data under each type of operating condition, the kernel density estimation method is used to fit the probability distribution curve to determine the normal temperature range and temperature rise rate threshold. The normal upper limit values ​​of each harmonic content rate and total harmonic distortion rate are calculated to construct the equipment status baseline parameter set. The equipment status baseline model containing temperature thresholds and harmonic limits is obtained. The multidimensional operating data of the current sampling period are input into the electrical quantity baseline model and the equipment status baseline model respectively, and the standardized deviation of the current value relative to the baseline interval is calculated dimension by dimension.

3. The distribution network fault detection method based on edge computing according to claim 1, characterized in that, Extract the electrical feature sub-vector and equipment status feature sub-vector for the current sampling period, and calculate the cross-layer consistency score between the two types of features, specifically including: For the voltage and current timing waveforms and switching action sequences of the current sampling period, core electrical features such as voltage RMS deviation, current transient peak value, switching action interval, and fundamental phase difference are extracted through time-domain statistical analysis and fast Fourier transform. After standardizing and normalizing each feature, they are spliced ​​together in a preset dimension order to generate fixed-dimensional electrical feature sub-vectors. Based on the electrical feature sub-vectors and the preset electrical analysis rules, the corresponding first fault analysis vector is determined; Using a unified timestamp identifier as the alignment benchmark, the equipment temperature and harmonic parameter data within the same sampling period are extracted, and equipment status features such as temperature gradient, total harmonic distortion rate, and characteristic subharmonic content rate are extracted. The numerical mapping is completed using the normalization rule consistent with the electrical feature subvector, and the equipment status feature subvector is generated by splicing according to the corresponding dimensions. Based on the equipment status feature sub-vector and the preset equipment analysis rules, the corresponding second fault analysis vector is determined; The first fault analysis vector and the second fault analysis vector are input into the cross-layer consistency calculation unit to obtain a cross-layer consistency score with a value range of 0 to 1.

4. The distribution network fault detection method based on edge computing according to claim 1, characterized in that, The standardized deviation and cross-layer consistency score meet the preset trigger conditions, specifically including: A pre-defined hierarchical threshold system for standardized deviations is established, including single-dimensional fault thresholds and multi-dimensional early warning thresholds. Each standardized deviation is matched with the corresponding threshold. When the deviation of the core electrical quantity dimension exceeds the single-dimensional fault threshold, or when the deviation of at least two secondary dimensions simultaneously exceeds the multi-dimensional early warning threshold, a preliminary deviation trigger flag is generated. The preliminary judgment result is jointly verified with the cross-layer consistency score. If there is a preliminary deviation from the trigger flag and the cross-layer consistency score is lower than the preset consistency threshold, the preset trigger condition is determined to be met; otherwise, the preset trigger condition is not met.

5. The distribution network fault detection method based on edge computing according to claim 1, characterized in that, Using N consecutive sampling times within the current detection window as nodes, and the multi-dimensional operational feature data output at each sampling time as node attributes, edges are constructed based on the temporal similarity of electrical quantities and topological correlation between nodes to generate a local operational status map of the distribution network, specifically including: N consecutive sampling times within the current detection window are extracted as graph nodes. The standardized deviation, electrical feature sub-vector, and equipment status feature sub-vector corresponding to each sampling time are concatenated to generate a node attribute vector with a unified dimension. All nodes are sequentially numbered according to the sampling time sequence, and a set of node attributes with time sequence index is output. Using the set of node attributes as input, the cosine similarity of electrical feature vectors between nodes is calculated for each pair to obtain the temporal similarity weight; at the same time, the preset topology connection matrix of the distribution network terminal is retrieved to assign topology correlation weight to the node pairs with direct electrical connections; the two types of weights are weighted and merged into a comprehensive edge weight, and node pairs with weights higher than a preset threshold are selected to form an effective edge set, and the weighted edge connection relationship set is output. The set of node attributes and the set of edge connections are input into the graph generation unit. The nodes at the sampling time are used as vertices and the effective edges are used as undirected connecting edges. The graph structure is assembled according to temporal association and topological constraints to generate a local operation status graph of the distribution network that contains multi-dimensional node features and edge weight information.

6. The distribution network fault detection method based on edge computing according to claim 1, characterized in that, The graph neural network encoding is mapped to a unified embedding space through a pre-trained contrastive learning encoder to generate the current running state embedding vector; Calculate the cosine similarity between this vector and the embedding vectors of various fault modes in the fault sample library, and output the fault type and matching similarity, specifically including: The graph neural network encoded features output by the lightweight graph attention network are input into the pre-trained contrastive learning encoder. The encoder performs nonlinear transformation and dimensionality compression on the high-dimensional encoded features through two fully connected perceptron layers, mapping them to a unified embedding space with fixed dimensions. The mapped feature vector is then subjected to L2 normalization to generate a current running state embedding vector with normalized modulus, which is then output to the similarity matching module. Using the current running state embedding vector as input, retrieve the embedding vectors of various fault modes pre-generated by the isomorphic encoder in the fault sample library; calculate the cosine similarity between the current embedding vector and each fault mode embedding vector for each class, obtain the similarity value for each type of fault, sort them in descending order of similarity to generate a fault matching sorting list, and output a matching result set containing fault type identifiers and corresponding similarity values. Input the fault matching sorting list into the judgment unit and compare the highest similarity value with the preset similarity threshold. If the highest similarity value is not lower than the preset threshold, output the corresponding fault type and matching similarity as the final detection result. If the highest similarity value is lower than the preset threshold, it is determined that the known fault mode is not matched, and the unknown anomaly identifier and the current highest similarity value are output.

7. A power distribution network fault detection system based on edge computing, characterized in that, The system includes: The computing module is used for edge computing nodes to collect multi-dimensional operating data of distribution network terminals in real time. The multi-dimensional operating data includes voltage and current time-series waveforms, switch action sequences, equipment temperature and harmonic parameters. Combined with the terminal's historical normal operation dataset, electrical quantity baselines and equipment status baselines are constructed respectively. The standardized deviation of the current operating data from the normal baseline is output. At the same time, the electrical feature sub-vectors and equipment status feature sub-vectors of the current sampling period are extracted, and the cross-layer consistency score of the two types of features is calculated. The output module is used to generate a local operating status map of the distribution network when the standardized deviation and cross-layer consistency score meet the preset trigger conditions. It takes N consecutive sampling times within the current detection window as nodes and the multi-dimensional operating feature data output at each sampling time as node attributes. It constructs edges based on the time series similarity of electrical quantities and topological correlation between nodes. The map is then input into a lightweight graph attention network to complete the aggregation encoding of node features and global topological features, and outputs the graph neural network encoded features. The matching module is used to map the graph neural network encoding to a unified embedding space through a pre-trained contrastive learning encoder, generating the current running state embedding vector; it calculates the cosine similarity between this vector and the embedding vectors of various fault modes in the fault sample library, and outputs the fault type and matching similarity.

8. The edge computing-based power distribution network fault detection system according to claim 7, characterized in that, The computing module includes computing units. Used for timing alignment and amplitude normalization preprocessing of voltage and current timing waveforms and switching action sequences based on the historical normal operation dataset of the terminal; Using a sliding time window statistical method, the mean, standard deviation, and upper and lower limits of the 95% confidence interval of preset key electrical quantity characteristics are extracted window by window from the voltage and current time-series waveforms and switching action sequences to form a baseline parameter set of electrical quantities. At the same time, a time-series correlation matrix is ​​established for each electrical quantity to record the coupling relationship between multiple electrical quantities under normal operating conditions. An electrical quantity baseline model containing the normal fluctuation range of each preset key electrical quantity characteristic and the coupling relationship matrix is ​​obtained. Using the historical sampling timestamps corresponding to the electrical quantity baseline model as indexes, equipment temperature and harmonic parameters are synchronously extracted from the historical normal operation dataset and hierarchically clustered according to seasonal periods and load levels; For the equipment temperature data under each type of operating condition, the probability distribution curve is fitted using the kernel density estimation method to determine the normal temperature range and the temperature rise rate threshold; the normal upper limit values ​​of each harmonic content rate and total harmonic distortion rate are calculated to construct the equipment state baseline parameter set; and an equipment state baseline model containing temperature thresholds and harmonic limits is obtained. The multidimensional operating data of the current sampling period are input into the electrical quantity baseline model and the equipment status baseline model respectively, and the standardized deviation of the current value relative to the baseline interval is calculated dimension by dimension.

9. A power distribution network fault detection device based on edge computing, characterized in that, The device includes: processor; And a memory having executable code stored thereon, which, when executed, causes the processor to perform a power distribution network fault detection method based on edge computing as described in any one of claims 1-6.

10. A non-volatile computer storage medium, characterized in that, It stores computer instructions, which, when executed, implement a power distribution network fault detection method based on edge computing as described in any one of claims 1-6.

Images