Circuit hidden danger diagnosis method and system based on multi-dimensional data and edge calculation

Abstract

The invention provides a circuit hidden danger diagnosis method and system based on multi-dimensional data and edge calculation, and the method comprises the steps: carrying out the real-time collection of multi-source data from current monitoring, temperature monitoring, vibration monitoring and partial discharge monitoring through a multi-mode spatial-temporal feature fusion engine; performing space-time alignment on multi-source data in a multi-modal space-time feature fusion engine, and dynamically calibrating the relationship between a current abrupt change point and a temperature hot spot so as to improve the accuracy of fault detection; a hierarchical computing capsule network is adopted for data processing, the hierarchical computing capsule network automatically loads a corresponding computing model according to the detected abnormal mode, and model parameters are adjusted in a self-adaptive mode according to the condition of computing resources; a fault deduction system driven by a knowledge graph is used, a federal evolution learning mechanism is combined, possible faults and hidden danger propagation paths in a circuit are predicted, and corresponding suggestions and measures are provided, so that early warning and preventive maintenance of circuit hidden dangers are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of circuit safety detection, and particularly relates to a circuit hidden danger diagnosis method and system based on multi-dimensional data and edge computing. BACKGROUND

[0002] In the field of power systems, especially in grid operation and maintenance and industrial automation control, early diagnosis of circuit hidden dangers is crucial for ensuring the safe operation and efficient management of power systems. Existing circuit hidden danger diagnosis techniques mainly rely on single modal data monitoring, such as current monitoring or temperature monitoring, and lack comprehensive analysis and utilization of multi-modal signals. In addition, due to the complexity and dynamic changes of circuit systems, traditional fault diagnosis methods based on single points or single models often fail to accurately predict possible faults and hidden danger propagation paths in circuits, especially in scenarios where current mutations and temperature changes are interrelated, the accuracy and timeliness of diagnosis are limited by the following factors: 1) Lack of multi-modal signal fusion, existing technologies mostly collect current, temperature and other signals separately, and fail to effectively fuse multi-source data (such as vibration, partial discharge) for comprehensive analysis, which limits the accuracy and comprehensiveness of fault detection.

[0003] 2) Inefficient use of computing resources, traditional circuit hidden danger diagnosis systems lack effective resource management and scheduling mechanisms at the edge computing level, and use the same set of computing models and parameters in both low-risk and high-risk scenarios, which not only consumes excessive computing resources, but also may cause processing delays, especially on resource-constrained edge devices.

[0004] 3) Limitations of fault prediction, current fault diagnosis systems focus more on emergency handling after faults occur rather than early warning before faults occur.

[0005] 4) Model updating and adaptability, traditional circuit hidden danger diagnosis models have long update cycles and weak adaptability to new scenarios, especially when the power grid structure changes or new power equipment is introduced, model updates often lag behind, affecting the accuracy of diagnosis.

[0006] Therefore, the present application designs a circuit hidden danger diagnosis method and system based on multi-dimensional data and edge computing. SUMMARY

[0007] Therefore, the present application aims to overcome the defects in the prior art that affect the efficiency of circuit hidden danger monitoring and diagnosis, and to provide a circuit hidden danger diagnosis method and system based on multi-dimensional data and edge computing.

[0008] To solve the above problems, the present application provides a circuit hidden danger diagnosis method and system based on multi-dimensional data and edge computing, which comprises: S1: using a multi-modal spatio-temporal feature fusion engine, collecting multi-source data from current monitoring, temperature monitoring, vibration monitoring and partial discharge monitoring in real time; S2: through a spatio-temporal cross attention mechanism, aligning the multi-source data in the multi-modal spatio-temporal feature fusion engine, dynamically calibrating the relationship between current mutation points and temperature hot spots to improve the accuracy of fault detection; S3: on the edge of the heterogeneous computing gateway, using a hierarchical computing capsule network for data processing, the hierarchical computing capsule network automatically loads the corresponding computing model according to the detected abnormal mode, and adaptively adjusts the model parameters according to the computing resource situation, in order to achieve the purpose of high efficiency and energy saving; S4: using a knowledge graph driven fault deduction system combined with a federal evolutionary learning mechanism, predicting possible faults and hidden danger propagation paths in the circuit, and providing corresponding measures, so as to realize early warning and preventive maintenance of circuit hidden dangers.

[0009] Preferably, the real-time collection process of multi-source data by the multi-modal spatio-temporal feature fusion engine in step S1 includes: A heterogeneous sensor network is formed by deploying high-precision current sensing devices, distributed temperature monitoring arrays, micro-electro-mechanical system vibration sensors and ultra-high frequency partial discharge detection modules at key nodes of the circuit; Each sensor node realizes parallel collection of multi-dimensional data based on a timestamp synchronization mechanism. The collected data is preprocessed through time-frequency domain conversion, noise suppression and feature enhancement, and then input to the feature alignment layer of the spatio-temporal cross attention mechanism; The alignment layer dynamically calibrates the sampling frequency deviation and spatial position offset of different modal data by learning the spatial correlation weight and time sequence dependence relationship of each monitoring point in the circuit topology structure. Then, through the learnable modal fusion gate unit, the feature contribution degree of each type of monitoring data is adaptively allocated according to the running state of the circuit, and finally a fusion data matrix containing transient changes and steady-state features is generated; Wherein, the weight parameters of the modal fusion gate unit are iteratively optimized based on historical fault case data through the federal learning framework.

[0010] Preferably, the spatio-temporal cross attention mechanism in step S2 includes a time synchronization calibration layer, a spatial correlation modeling unit and a dynamic weight distributor. The time synchronization calibration layer unifies multi-source heterogeneous data to the same time scale through sliding time window and interpolation algorithm; the spatial correlation modeling unit constructs a monitoring point spatial adjacency matrix based on the circuit topology structure to quantify the spatial coupling degree of different modal data; the dynamic weight distributor calculates the spatio-temporal response correlation between current mutation points and temperature hot spots through a three-dimensional attention tensor; The dimensions of the three-dimensional attention tensor correspond to time offset, spatial distance, and feature similarity, respectively. The attention weight matrix generated by the corresponding tensor can dynamically enhance the feature mapping relationship between abnormal current fluctuations and the corresponding temperature rise in the region, and suppress non-correlated noise signals, thereby achieving accurate alignment of multi-source data in the spatiotemporal dimensions and enhancement of key features.

[0011] Preferably, the hierarchical computing capsule network in step S3 is deployed on an edge heterogeneous computing gateway integrating CPU, FPGA and NPU, and the hierarchical computing capsule network includes: a feature parsing layer, a pattern matching layer and a dynamic execution layer; The feature parsing layer performs hierarchical feature extraction on multimodal fusion data and generates abnormal feature vectors. The pattern matching layer performs similarity retrieval through a pre-built abnormal pattern feature map library to determine the appropriate computing model. The dynamic execution layer adaptively adjusts the number of network layer activations, feature channel dimensions, and iterative calculation accuracy based on the computing load, memory usage, and energy consumption thresholds fed back by the gateway's real-time resource monitoring module, through model parameter pruning algorithms and computing unit scheduling strategies. In low-resource scenarios, only basic capsule units are enabled for rapid screening, while in high-resource scenarios, enhanced capsule clusters are activated to achieve in-depth analysis.

[0012] Preferably, in step S4: a knowledge graph-driven fault inference system package is used, combined with the fault prediction sub-models trained locally at each edge node in the federated evolutionary learning mechanism, to aggregate and generate globally optimized model parameters under a privacy protection protocol. The federated evolutionary learning mechanism iteratively optimizes the weights of sub-models through a non-dominated sorting genetic algorithm and dynamically adjusts the fault propagation probability matrix. This enables the system to complete the localization of potential fault points, the sorting of multi-path propagation probabilities, and the generation of preventive maintenance plans that include component replacement priorities and operation sequences in a short period of time. The aggregation process of global model parameters uses homomorphic encryption technology to ensure data security.

[0013] Preferably, in step S4, the knowledge graph-driven fault inference system realizes circuit hazard prediction through a multi-level knowledge graph architecture, wherein the circuit topology layer contains the component connection relationship and physical parameter ontology, the fault rule layer stores the association rules based on historical cases and expert experience, and the failure mode layer records the feature vectors and propagation thresholds of typical faults. By using a graph neural network to perform entity linking and path search on the abnormal features output from the edge, and combining the knowledge reasoning engine to calculate the association confidence of each potential fault node, the distributed reasoning sub-model based on the non-dominated sorting genetic algorithm in the federated evolutionary learning mechanism is called. Under the protection of the homomorphic encryption protocol, the local inference results of each edge node are aggregated to generate hierarchical early warning information containing the probability of fault occurrence, multi-path propagation timing and impact range. Based on the circuit criticality score, a preventive maintenance plan containing component replacement priority and operation steps is automatically generated.

[0014] This invention also provides a circuit hazard diagnosis system based on multidimensional data and edge computing, including the circuit hazard diagnosis method based on multidimensional data and edge computing described in the preceding claim, comprising: The multimodal spatiotemporal feature fusion engine module is used to collect and fuse multi-source data from current, temperature, vibration and partial discharge monitoring in real time; The spatiotemporal cross-attention module, connected to the multimodal spatiotemporal feature fusion engine module, is used to perform spatiotemporal alignment of multi-source data and dynamically calibrate the relationship between current abrupt changes and temperature hotspots. The edge heterogeneous computing gateway is configured with a hierarchical computing capsule network, which is used to automatically load the corresponding computing model and adaptively adjust the model parameters according to the detected abnormal patterns. The knowledge graph-driven fault prediction system communicates with the edge heterogeneous computing gateway and combines a federated evolutionary learning mechanism to predict circuit faults and potential hazard propagation paths, and outputs preventive maintenance suggestions.

[0015] Preferably, the multimodal spatiotemporal feature fusion engine module includes: A heterogeneous sensing network, comprising: a current sensing device, a distributed temperature monitoring array, a MEMS vibration sensor, and an ultra-high frequency partial discharge detection module; The time-frequency domain preprocessing unit is used to perform time-frequency domain transformation, noise suppression, and feature enhancement on the acquired data; The spatiotemporal feature alignment layer achieves parallel acquisition of multi-dimensional data based on a timestamp synchronization mechanism and outputs a fused data matrix; The modality fusion gating unit iteratively optimizes weight parameters through a federated learning framework to adaptively allocate the feature contribution of each type of monitoring data.

[0016] Preferably, the spatiotemporal cross-attention module includes: The time synchronization calibration layer unifies the time scale of multi-source heterogeneous data through a sliding time window and interpolation algorithm; The spatial correlation modeling unit constructs a spatial adjacency matrix of monitoring points based on the circuit topology to quantify the spatial coupling degree; The dynamic weight allocator uses a three-dimensional attention tensor to calculate the spatiotemporal response correlation between current abrupt changes and temperature hotspots, and outputs an attention weight matrix. The edge heterogeneous computing gateway integrates CPU, FPGA and NPU; The hierarchical computing capsule network includes: Feature parsing layer, used for hierarchical extraction of abnormal feature vectors; The pattern matching layer uses a similarity search based on an anomaly pattern feature map library to determine the computational model. The dynamic execution layer adaptively adjusts the number of network layer activations, feature channel dimensions, and iterative calculation accuracy based on the computing load, memory usage, and energy consumption thresholds fed back by the real-time resource monitoring module.

[0017] Preferably, the knowledge graph-driven fault deduction system includes: The federated evolutionary learning module uses a non-dominated sorting genetic algorithm to aggregate fault prediction sub-models of each edge node under a privacy protection protocol, and dynamically adjusts the fault propagation probability matrix. A globally optimized parameter aggregator uses homomorphic encryption to ensure data security. A preventative maintenance plan generator is used to output maintenance plans that include component replacement priorities, operation sequences, and scope of impact. The multi-level knowledge graph architecture includes: a circuit topology layer that stores the connection relationships of components and the ontology of physical parameters; a fault rule layer that stores the association rules of historical cases and expert experience; and a failure mode layer that records typical fault feature vectors and propagation thresholds. The graph neural network inference engine is used to perform entity linking and path search on abnormal features, and combines the distributed inference sub-model output by the federated evolutionary learning module to generate hierarchical early warning information.

[0018] The circuit defect diagnosis method and system based on multidimensional data and edge computing provided by this invention have the following beneficial effects: 1. This invention constructs a complete multi-source data collaborative analysis system through a multimodal spatiotemporal feature fusion engine and a spatiotemporal cross-attention mechanism. It integrates four core monitoring data types: current, temperature, vibration, and partial discharge. Parallel acquisition is achieved through heterogeneous sensor networks, and data quality is ensured through preprocessing such as time-frequency domain conversion and noise suppression. The STCA mechanism can solve the spatiotemporal offset problem of multimodal data by time synchronization calibration, spatial correlation modeling, and three-dimensional attention tensor calculation, thereby enhancing key fault features and suppressing non-correlated noise. The modal fusion gating unit optimizes weights iteratively based on federated learning, which can adaptively allocate the feature contribution of each modality data according to the circuit operating state, generating a fusion data matrix containing transient / steady-state features. This avoids the limitations of one-sided diagnosis based on single data, and the fault detection accuracy is significantly improved compared with traditional methods.

[0019] 2. This invention also constructs a dynamic resource scheduling system through edge heterogeneous computing gateways and hierarchical computing capsule networks. The gateway integrates multi-architecture computing units such as CPU, FPGA, and NPU. HCCN achieves full-process resource adaptation through feature parsing layer, pattern matching layer, and dynamic execution layer. In low-risk scenarios, only basic capsule units are activated, which can save up to 95% of computing resources and meet the needs of rapid screening. In high-risk scenarios, enhanced capsule clusters are automatically activated to ensure the accuracy and real-time performance of deep analysis. The dynamic execution layer combines computing load, memory usage, and energy consumption thresholds to adjust the number of network activations, feature channel dimensions, and computing accuracy in real time, solving the problem of one-size-fits-all resource waste and taking into account the processing efficiency and energy-saving needs of edge devices. It is especially suitable for resource-constrained industrial sites. 3. This invention also achieves proactive prevention and control of potential hazards through a knowledge graph-driven fault deduction system and federated evolutionary learning; the multi-level knowledge graph integrates component physical parameters, historical fault cases, expert experience, and failure thresholds to provide complete knowledge support for fault deduction; KG-Forecaster combined with graph neural network reasoning can locate potential fault points, sort multi-path propagation probabilities, predict the scope and timing of impact, and generate early warning information including fault occurrence probability, propagation path, and impact level; the system not only outputs early warning results, but can also automatically generate preventive maintenance plans based on circuit criticality scores, including component replacement priorities, operation sequences, and maintenance steps, upgrading traditional post-fault maintenance to pre-fault prevention and control, significantly reducing downtime maintenance costs and the risk of accidents. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0021] like Figure 1 As shown, this invention provides a circuit defect diagnosis method based on multidimensional data and edge computing, which includes the following steps: S1: Use a multimodal spatiotemporal feature fusion engine to collect multi-source data from current monitoring, temperature monitoring, vibration monitoring and partial discharge monitoring in real time; S2: Through the Spatiotemporal Cross Attention (STCA) mechanism, the multi-source data is spatiotemporally aligned within the multimodal spatiotemporal feature fusion engine to dynamically calibrate the relationship between current abrupt change points and temperature hotspots, thereby improving the accuracy of fault detection. S3: On the heterogeneous computing gateway at the edge, a hierarchical computing capsule network (HCCN) is used for data processing. The hierarchical computing capsule network automatically loads the corresponding computing model according to the detected abnormal patterns and adaptively adjusts the model parameters according to the computing resource situation to achieve the purpose of high efficiency and energy saving. S4: Utilize a knowledge graph-driven fault prediction (KG-Forecaster) system, combined with a federated evolutionary learning mechanism, to predict the propagation paths of potential faults and hidden dangers in circuits, and provide corresponding suggested measures, thereby achieving early warning and preventive maintenance of circuit hidden dangers.

[0022] In some implementations, the real-time acquisition process of multi-source data by the multimodal spatiotemporal feature fusion engine in step S1 includes: A heterogeneous sensing network is formed by deploying high-precision current sensing devices, distributed temperature monitoring arrays, microelectromechanical system vibration sensors, and ultra-high frequency partial discharge detection modules at key circuit nodes. Each sensor node achieves parallel acquisition of multi-dimensional data based on a timestamp synchronization mechanism. The acquired data is preprocessed by time-frequency domain conversion, noise suppression, and feature enhancement, and then input to the feature alignment layer of the spatiotemporal cross-attention mechanism. The alignment layer learns the spatial correlation weights and time series dependencies of each monitoring point in the circuit topology, dynamically calibrates the sampling frequency deviation and spatial position offset of different modal data, and then the learnable modal fusion gating unit adaptively allocates the feature contribution of each type of monitoring data according to the circuit operating state, finally generating a fusion data matrix containing transient changes and steady-state features. The weight parameters of the modal fusion gating unit are iteratively optimized based on historical fault case data using a federated learning framework.

[0023] Specifically, the system comprises multiple sensor nodes, each responsible for monitoring specific types of signals, such as current, temperature, vibration, and partial discharge. These include, but are not limited to: a quantum galvanometer for high-frequency transient current monitoring; a graphene thermal imaging camera for high-resolution temperature distribution acquisition; a MEMS laser vibrometer for precise displacement and vibration measurement; and an AI-UHF antenna for capturing partial discharge signals. Each sensor node is equipped with a high-speed data acquisition card, enabling continuous signal capture during operation and immediate uploading to edge computing devices. The system uses a spatiotemporal cross-attention mechanism (STCA) to align multi-source data spatiotemporally, and a learnable modal weight adjustment module dynamically determines the importance and correlation of different modal data, thereby effectively fusing current, temperature, vibration, and partial discharge monitoring data. In data processing and preprocessing, current monitoring data undergoes Fourier transform to extract frequency domain features, vibration monitoring data is denoised and feature extracted using wavelet transform, while temperature and partial discharge monitoring data are initially interpreted directly through a deep convolutional network.

[0024] In some implementations, the spatiotemporal cross-attention mechanism (STCA) in step S2 includes: a time synchronization calibration layer, a spatial correlation modeling unit, and a dynamic weight allocator; The time synchronization calibration layer unifies multi-source heterogeneous data to the same time scale through a sliding time window and interpolation algorithm; the spatial correlation modeling unit constructs a spatial adjacency matrix of monitoring points based on the circuit topology to quantify the spatial coupling degree of different modal data; and the dynamic weight allocator calculates the spatiotemporal response correlation between current abrupt changes and temperature hotspots through a three-dimensional attention tensor. The dimensions of the three-dimensional attention tensor correspond to time offset, spatial distance, and feature similarity, respectively. The attention weight matrix generated by the corresponding tensor can dynamically enhance the feature mapping relationship between abnormal current fluctuations and the corresponding temperature rise in the region, and suppress non-correlated noise signals, thereby achieving accurate alignment of multi-source data in the spatiotemporal dimensions and enhancement of key features.

[0025] Specifically, the Spatiotemporal Cross-Attention (STCA) mechanism is a data processing technique used to capture and quantify the spatiotemporal dependencies between different modalities of data. STCA dynamically adjusts the attention distribution by analyzing the time series and spatial distribution of multi-source data to highlight corresponding temporally and spatially related feature points. Before fusing multi-source data, the STCA mechanism performs data preprocessing to unify signals of different frequencies and sampling rates onto the same time base, ensuring the alignment of current, temperature, vibration, and partial discharge data on the time axis.

[0026] Specifically, the STCA mechanism identifies abrupt changes in current data, i.e., moments when a fault may occur, and compares and calibrates these with hotspots in temperature data. By calculating the temporal delay and spatial location differences between current abrupt changes and temperature hotspots, STCA dynamically adjusts its attention window, focusing on the temperature change region caused by the corresponding current abrupt change, thereby accurately locating potential circuit fault points. After STCA calibration, the fusion strategy within the multimodal spatiotemporal feature fusion engine further incorporates learnable modal weights to perform a weighted average of multi-source data to generate the final fused features. The fusion process not only considers the value of the data itself but also adjusts it according to the spatiotemporal consistency of the data to ensure the accuracy of fault detection.

[0027] Specifically, the STCA mechanism first analyzes the current data sequence and marks all current abrupt changes. Then, it scans the temperature data to find the temperature hotspots that are spatially and temporally closest to any current abrupt change. The STCA mechanism calculates the correlation score between two corresponding feature points and dynamically adjusts its attention distribution based on this score to enhance the influence of highly correlated regions. Finally, the adjusted multi-source data is fed into a fusion strategy to generate fused features, which are then output to the fault detection model for analysis. Through the spatiotemporal alignment and dynamic calibration of the STCA mechanism, key information from multi-source data is effectively combined, significantly improving the accuracy of circuit fault detection. Especially in scenarios where current abrupt changes and temperature increases occur simultaneously, the STCA mechanism can identify fault signs more quickly and accurately.

[0028] In some implementations, the hierarchical computing capsule network in step S3 is deployed on an edge heterogeneous computing gateway integrating a CPU, FPGA, and NPU. The hierarchical computing capsule network includes a feature parsing layer, a pattern matching layer, and a dynamic execution layer. The feature parsing layer performs hierarchical feature extraction on multimodal fusion data and generates abnormal feature vectors. The pattern matching layer performs similarity retrieval through a pre-built abnormal pattern feature map library to determine the appropriate computing model. The dynamic execution layer adaptively adjusts the number of network layer activations, feature channel dimensions, and iterative calculation accuracy based on the computing load, memory usage, and energy consumption thresholds fed back by the gateway's real-time resource monitoring module, through model parameter pruning algorithms and computing unit scheduling strategies. In low-resource scenarios, only basic capsule units are enabled for rapid screening, while in high-resource scenarios, enhanced capsule clusters are activated to achieve in-depth analysis.

[0029] Specifically, the edge heterogeneous computing gateway used in this application integrates high-performance computing units with various processor architectures, including but not limited to microcontroller units (MCUs), programmable gate arrays (FPGAs), and neural network processors (NPUs), to provide flexible and efficient data processing capabilities. The hierarchical computing capsule network (HCCN) can automatically select and load the most suitable computing model for processing based on the abnormal patterns of the received multimodal spatiotemporal feature data. HCCN enables model parameters to be activated or deactivated in stages on computing resources at different levels, thereby achieving fine-grained management of computing resources and energy consumption. At the same time, in HCCN, once the data preprocessor identifies an abnormal pattern in current, temperature, vibration, or partial discharge, the edge computing gateway will quickly initiate the model search and loading process. The system will retrieve the computing model that matches the abnormal pattern from the pre-stored model library and dynamically load it onto the computing resources specified in HCCN through the network, and immediately begin processing.

[0030] HCCN can adjust the parameters and computing strategies of the computing model in real time based on the current resource status of the edge computing device, including the available capacity of CPU, memory, GPU, or NPU. For low-risk or simple scenarios, it only activates the model parameters of the smallest unit in HCCN, significantly reducing computing requirements and energy consumption. For complex or high-risk scenarios, it automatically expands computing resources and loads more model parameters to improve the accuracy and response speed of fault detection. HCCN's dynamic adjustment feature ensures broad hardware compatibility from different levels of microprocessors to dedicated neural network processing units (NPUs), achieving efficient utilization of computing resources and energy consumption optimization. In low-risk scenarios, HCCN can save up to 95% of computing resources by activating only a small number of model parameters, while in high-risk scenarios, it ensures the real-time performance and accuracy of fault detection by quickly loading and configuring the complete model.

[0031] Specifically, the HCCN mechanism running on the edge heterogeneous computing gateway first analyzes the fused feature vector to identify whether there are abnormal patterns. If so, HCCN will select an appropriate computing model from the model library and load it according to the nature of the abnormal pattern. Then, based on the real-time resource status of the edge device, it dynamically adjusts the active parameters of the computing model to optimize computing efficiency and energy consumption. Finally, HCCN sends the processed data or fault diagnosis results to the cloud or other terminal devices.

[0032] In some implementations, in step S4: the Knowledge Graph-Driven Fault Inference (KG-Forecaster) system package is used, combined with the fault prediction sub-models trained locally at each edge node in the federated evolutionary learning mechanism, to aggregate and generate globally optimized model parameters under a privacy protection protocol. The federated evolutionary learning mechanism iteratively optimizes the weights of sub-models through the non-dominated sorting genetic algorithm (NSGA-II) and dynamically adjusts the fault propagation probability matrix, enabling the system to complete the localization of potential fault points, the sorting of multi-path propagation probabilities, and the generation of preventive maintenance schemes that include component replacement priorities and operation sequences in a short time. The aggregation process of global model parameters adopts homomorphic encryption technology to ensure data security.

[0033] Specifically, this method is based on a pre-constructed circuit knowledge graph, which contains multi-level information such as circuit topology, physical attributes, historical fault records, and the mutual influence relationships between components. Through the knowledge graph, the system can understand the connection methods and mechanisms between various components in the circuit, laying a theoretical foundation for fault prediction and hazard propagation path analysis. The KG-Forecaster system uses the aforementioned knowledge graph, combined with abnormal data obtained from a multimodal spatiotemporal feature fusion engine, to execute a fault inference algorithm. The algorithm predicts circuit faults that may be caused by abnormal data by traversing and evaluating paths in the knowledge graph, while analyzing the possibility and path of fault propagation.

[0034] To continuously optimize the fault prediction accuracy of the KG-Forecaster system, this method employs a federated evolutionary learning mechanism (a distributed machine learning framework). Multiple edge computing devices co-evolve the model without sharing raw data. Each device updates its fault prediction model locally and periodically aggregates model differences via a secure communication protocol, enabling iterative upgrades of the global model and enhancing the system's robustness and generalization capabilities. It can not only predict faults in individual devices or components but also anticipate fault propagation paths and assess their impact on the entire circuit. Once a potential fault or hazard propagation is predicted, the system instantly generates suggested measures, such as guiding maintenance personnel to conduct preventative checks, adjusting load distribution to reduce stress, or replacing expected failure components in advance, thereby achieving early warning and preventative maintenance of circuit hazards.

[0035] Specifically, the process begins with edge computing to collect and fuse multi-source data to identify anomalous patterns. These patterns are then input into the KG-Forecaster system to initiate a fault prediction process. The system analyzes the source and potential impact of these anomalous patterns using a knowledge graph, while a federated evolutionary learning mechanism continuously optimizes the model in the background to improve prediction accuracy. Once the prediction is complete, the KG-Forecaster system outputs a detailed fault prediction report, including the fault type, possible location, estimated fault time, and recommended preventative measures. This helps the maintenance team develop effective maintenance plans. By combining the KG-Forecaster system with the federated evolutionary learning mechanism, this method significantly reduces circuit maintenance costs and improves system reliability. It can successfully predict and address potential faults in circuits in advance, avoiding high downtime repair costs and reducing the risk of accidents caused by faults.

[0036] In some implementations, the knowledge graph-driven fault prediction (KG-Forecaster) system in step S4 uses a multi-level knowledge graph architecture to predict circuit hazards. The circuit topology layer contains the component connection relationships and physical parameter ontology, the fault rule layer stores association rules based on historical cases and expert experience, and the failure mode layer records the feature vectors and propagation thresholds of typical faults. By using a graph neural network to perform entity linking and path search on the abnormal features output from the edge, and combining the knowledge reasoning engine to calculate the association confidence of each potential fault node, the distributed reasoning sub-model based on the non-dominated sorting genetic algorithm in the federated evolutionary learning mechanism is called. Under the protection of the homomorphic encryption protocol, the local inference results of each edge node are aggregated to generate hierarchical early warning information containing the probability of fault occurrence, multi-path propagation timing and impact range. Based on the circuit criticality score, a preventive maintenance plan containing component replacement priority and operation steps is automatically generated.

[0037] Specifically, knowledge graph construction collects detailed information on all components in a circuit system, including physical characteristics, electrical parameters, inter-component connections, behavioral patterns under normal operating conditions, and historical fault records. It determines the dependencies and influence relationships between circuit components, such as the correlation between current abrupt changes and temperature hotspots, or the relationship between mechanical vibration and capacitor solder joint stability. Using graph databases or similar tools, it constructs a knowledge graph, treating components and relationships as nodes and edges to form a multi-layered network structure. The graph contains at least three levels: topology, physical properties, and failure modes.

[0038] In the data fusion and anomaly detection process, multimodal data acquisition uses intelligent sensor arrays to collect multi-source data such as current, temperature, vibration, and partial discharge; spatiotemporal feature fusion uses a spatiotemporal cross-attention mechanism (STCA) to perform spatiotemporal alignment and feature extraction on multi-source data to obtain fused abnormal data patterns; edge data processing uses a hierarchical computing capsule network (HCCN) on the edge computing gateway for data preprocessing and preliminary analysis. Among them, fault pattern matching is as follows: pattern recognition matches the abnormal data patterns processed at the edge with known fault patterns in the knowledge graph to identify possible fault types and their associated components; dynamic model loading is as follows: based on the identified fault type, HCCN dynamically loads the corresponding deep model for further analysis to improve the accuracy and response speed of fault detection. Among them, knowledge graphs are used for inference. Fault propagation path analysis starts from the fault node matched by abnormal data patterns in the knowledge graph and uses the relation edges and failure mode layers in the graph to infer the possible propagation path and scope of impact of the fault. Confidence calculation: Confidence is calculated for each inferred fault propagation path to assess its probability of actual occurrence.

[0039] In the federated evolutionary learning optimization, the model iterates through the federated evolutionary learning mechanism of each edge computing gateway, sharing the optimized model parameters without directly exchanging data, thereby continuously evolving and optimizing the KG-Forecaster system and improving the accuracy and efficiency of fault prediction. Data privacy protection utilizes the federated learning mechanism to ensure data privacy security. Data from each site does not leave its local area, and only the model update part is transmitted, which complies with privacy protection regulations.

[0040] The system combines fault prediction results with confidence assessments to generate a comprehensive report in the output prediction results and recommended measures. The report includes specific fault predictions, the possible scope of impact, and recommended preventive measures, such as disconnecting related circuits to prevent further damage and checking downstream connectors to locate the problem. The system will promptly convey the early warning information and recommended measures to maintenance personnel so that they can take action in advance.

[0041] The entire process is a closed loop, starting with data acquisition, followed by data fusion, fault identification, knowledge graph deduction, and model optimization, ultimately outputting prediction results and suggested measures to achieve early warning and preventive maintenance of circuit hazards. Through continuous learning and optimization, the prediction capability of the KG-Forecaster system will gradually improve, better adapting to complex and ever-changing circuit environments.

[0042] This invention also provides a circuit hazard diagnosis system based on multidimensional data and edge computing, employing any of the circuit hazard diagnosis methods based on multidimensional data and edge computing described in this invention, comprising: The multimodal spatiotemporal feature fusion engine module is used to collect and fuse multi-source data from current, temperature, vibration and partial discharge monitoring in real time; The spatiotemporal cross-attention module, connected to the multimodal spatiotemporal feature fusion engine module, is used to perform spatiotemporal alignment of multi-source data and dynamically calibrate the relationship between current abrupt changes and temperature hotspots. The edge heterogeneous computing gateway is configured with a hierarchical computing capsule network, which is used to automatically load the corresponding computing model and adaptively adjust the model parameters according to the detected abnormal patterns. The knowledge graph-driven fault prediction system communicates with the edge heterogeneous computing gateway and combines a federated evolutionary learning mechanism to predict circuit faults and potential hazard propagation paths, and outputs preventive maintenance suggestions.

[0043] Specifically, such as Figure 2 As shown, the system deployment includes a heterogeneous sensor network at key nodes of the circuit, comprising current sensing devices, a distributed temperature monitoring array, a MEMS vibration sensor, and an ultra-high frequency partial discharge detection module, ensuring comprehensive acquisition of multi-source data from the circuit. An edge-end heterogeneous computing gateway is built, integrating CPU, FPGA, and NPU, and deploying a hierarchical computing capsule network. A knowledge graph-driven fault inference system is constructed, including a multi-level knowledge graph architecture, a graph neural network inference engine, and a federated evolutionary learning module. During data acquisition and preprocessing, a multi-modal spatiotemporal feature fusion engine module acquires multi-source data such as current, temperature, vibration, and partial discharge in real time through the heterogeneous sensor network. The time-frequency domain preprocessing unit processes the acquired data... The data undergoes time-frequency domain transformation, noise suppression, and feature enhancement to improve data quality. The preprocessed data is then spatiotemporally aligned. The spatiotemporal feature alignment layer uses a timestamp synchronization mechanism to achieve parallel acquisition of multi-dimensional data, and the processed data is input into the spatiotemporal cross-attention module. The time synchronization calibration layer of the spatiotemporal cross-attention module unifies the time scale of multi-source heterogeneous data through a sliding time window and interpolation algorithm. The spatial correlation modeling unit constructs a spatial adjacency matrix of monitoring points based on the circuit topology to quantify spatial coupling. The dynamic weight allocator uses a three-dimensional attention tensor to calculate the spatiotemporal response correlation between current abrupt changes and temperature hotspots, and outputs an attention weight matrix to achieve spatiotemporal alignment of multi-source data.

[0044] Specifically, the system performs fault prediction and preventive maintenance. The graph neural network inference engine of the knowledge graph-driven fault inference system performs entity linking and path search on the abnormal features output from the edge end, and calculates the association confidence of each potential fault node in combination with the knowledge inference engine. The federated evolutionary learning module uses a non-dominated sorting genetic algorithm to aggregate the fault prediction sub-models of each edge node under the privacy protection protocol, and dynamically adjusts the fault propagation probability matrix. The global optimization parameter aggregator uses homomorphic encryption technology to ensure data security and aggregates the global optimization model parameters. The preventive maintenance plan generator generates hierarchical early warning information containing fault occurrence probability, multi-path propagation time sequence and impact range based on the calculation results, and automatically generates a preventive maintenance plan containing component replacement priority and operation steps based on the circuit criticality score.

[0045] In some embodiments, the multimodal spatiotemporal feature fusion engine module includes: A heterogeneous sensing network, comprising: a current sensing device, a distributed temperature monitoring array, a MEMS vibration sensor, and an ultra-high frequency partial discharge detection module; The time-frequency domain preprocessing unit is used to perform time-frequency domain transformation, noise suppression, and feature enhancement on the acquired data; The spatiotemporal feature alignment layer achieves parallel acquisition of multi-dimensional data based on a timestamp synchronization mechanism and outputs a fused data matrix; The modality fusion gating unit iteratively optimizes weight parameters through a federated learning framework to adaptively allocate the feature contribution of each type of monitoring data.

[0046] Specifically, the multimodal spatiotemporal feature fusion engine integrates four heterogeneous sensor networks: current, temperature, vibration, and partial discharge. Through time-frequency domain preprocessing (conversion, noise reduction, and feature enhancement) and a time-spatial feature alignment layer (timestamp synchronization + fusion matrix output), it ensures the consistency of multi-dimensional data in the spatiotemporal dimensions, forming a fusion data matrix that includes transient / steady-state features, thereby improving the comprehensiveness and real-time performance of state perception. The modal fusion gating unit dynamically optimizes weights based on a federated learning framework, adaptively allocating the contribution of each modal data feature, enhancing the representation capability of key features (such as the correlation between current abrupt changes and temperature hotspots), and providing a highly robust data foundation for subsequent analysis.

[0047] In some implementations, the spatiotemporal cross-attention module includes: The time synchronization calibration layer unifies the time scale of multi-source heterogeneous data through a sliding time window and interpolation algorithm; The spatial correlation modeling unit constructs a spatial adjacency matrix of monitoring points based on the circuit topology to quantify the spatial coupling degree; The dynamic weight allocator uses a three-dimensional attention tensor to calculate the spatiotemporal response correlation between current abrupt changes and temperature hotspots, and outputs an attention weight matrix. The edge heterogeneous computing gateway integrates CPU, FPGA and NPU; The hierarchical computing capsule network includes: Feature parsing layer, used for hierarchical extraction of abnormal feature vectors; The pattern matching layer uses a similarity search based on an anomaly pattern feature map library to determine the computational model. The dynamic execution layer adaptively adjusts the number of network layer activations, feature channel dimensions, and iterative calculation accuracy based on the computing load, memory usage, and energy consumption thresholds fed back by the real-time resource monitoring module.

[0048] Specifically, the spatiotemporal cross-attention module achieves precise alignment of multi-source data in three dimensions: time offset, spatial distance, and feature similarity through a time synchronization calibration layer (sliding window + interpolation algorithm), a spatial correlation modeling unit (circuit topology adjacency matrix), and a dynamic weight allocator (three-dimensional attention tensor). This dynamically enhances the spatiotemporal response correlation between abnormal current fluctuations and sudden temperature rises, suppresses non-correlated noise, and strengthens key fault characteristics.

[0049] Specifically, the edge heterogeneous computing gateway integrates CPU / FPGA / NPU and deploys a hierarchical computing capsule network: the feature parsing layer extracts hierarchical abnormal feature vectors, the pattern matching layer retrieves the appropriate model through the abnormal pattern graph library, and the dynamic execution layer adaptively adjusts the number of network layer activations, feature channel dimensions and computing accuracy according to the real-time computing power / memory / energy consumption status, enabling rapid screening in low-resource scenarios and in-depth analysis in high-resource scenarios, thus achieving efficient and energy-saving allocation of computing resources.

[0050] In some implementations, the knowledge graph-driven fault deduction system includes: The federated evolutionary learning module uses a non-dominated sorting genetic algorithm to aggregate fault prediction sub-models of each edge node under a privacy protection protocol, and dynamically adjusts the fault propagation probability matrix. A globally optimized parameter aggregator uses homomorphic encryption to ensure data security. A preventative maintenance plan generator is used to output maintenance plans that include component replacement priorities, operation sequences, and scope of impact. The multi-level knowledge graph architecture includes: a circuit topology layer that stores the connection relationships of components and the ontology of physical parameters; a fault rule layer that stores the association rules of historical cases and expert experience; and a failure mode layer that records typical fault feature vectors and propagation thresholds. The graph neural network inference engine is used to perform entity linking and path search on abnormal features, and combines the distributed inference sub-model output by the federated evolutionary learning module to generate hierarchical early warning information.

[0051] Specifically, the knowledge graph-driven fault inference system combines a federated evolutionary learning mechanism (non-dominated sorting genetic algorithm to optimize sub-model weights + homomorphic encrypted aggregation) with multi-level knowledge graphs (topology layer / rule layer / failure mode layer) and graph neural network reasoning to achieve potential fault location, multi-path propagation probability ranking and impact range prediction. Finally, it generates hierarchical early warning information and preventive maintenance plans that include component replacement priorities and operation sequences, supporting early warning and precise maintenance of circuit hazards.

[0052] This application constructs a complete multi-source data collaborative analysis system using a multi-modal spatiotemporal feature fusion engine and a spatiotemporal cross-attention mechanism (STCA). It integrates four core monitoring data types: current, temperature, vibration, and partial discharge. Parallel acquisition is achieved through heterogeneous sensor networks, and data quality is ensured through preprocessing such as time-frequency domain conversion and noise suppression. The STCA mechanism solves the spatiotemporal offset problem of multimodal data by using time synchronization calibration (unifying the time scale), spatial correlation modeling (quantifying the circuit topology coupling), and three-dimensional attention tensor calculation (dynamically correlated current mutations and temperature hotspots). It strengthens key fault features (such as the mapping relationship between current anomalies and sudden temperature rises in the region) while suppressing non-correlated noise. The modal fusion gating unit optimizes weights iteratively based on federated learning, which can adaptively allocate the feature contribution of each modality data according to the circuit operating state, generating a fusion data matrix containing transient / steady-state features. This avoids the limitations of one-sided diagnosis based on single data, and the fault detection accuracy is significantly improved compared with traditional methods.

[0053] This application constructs a dynamic resource scheduling system through an edge heterogeneous computing gateway and a hierarchical computing capsule network (HCCN). The gateway integrates multi-architecture computing units such as CPU, FPGA, and NPU. The HCCN achieves end-to-end resource adaptation through a feature parsing layer (extracting anomaly vectors), a pattern matching layer (adapting the computing model), and a dynamic execution layer (adjusting computing parameters). In low-risk scenarios, only the basic capsule unit is activated, which can save up to 95% of computing resources and meet the needs of rapid screening. In high-risk scenarios, the enhanced capsule cluster is automatically activated to ensure the accuracy and real-time performance of deep analysis. The dynamic execution layer combines computing load, memory usage, and energy consumption thresholds to adjust the number of network activations, feature channel dimensions, and computing accuracy in real time, solving the problem of one-size-fits-all resource waste and taking into account the processing efficiency and energy-saving needs of edge devices, which is especially suitable for resource-constrained industrial sites.

[0054] This application achieves proactive prevention and control of potential hazards through a knowledge graph-driven fault prediction system and federated evolutionary learning. The multi-level knowledge graph integrates component physical parameters, historical fault cases, expert experience, and failure thresholds to provide complete knowledge support for fault prediction. KG-Forecaster, combined with graph neural network reasoning, can locate potential fault points, rank multi-path propagation probabilities, predict the scope and timing of impact, and generate early warning information including fault occurrence probability, propagation path, and impact level. The system not only outputs early warning results but also automatically generates preventive maintenance plans based on circuit criticality scores, including component replacement priorities, operation sequences, and maintenance steps. This upgrades traditional post-fault repair to pre-fault prevention, significantly reducing downtime maintenance costs and accident risks (such as avoiding production interruptions or safety accidents caused by circuit faults).

[0055] This application constructs an adaptive iterative system through federated evolutionary learning and a dynamic model loading mechanism. Federated evolutionary learning supports real-time feedback of local model optimization results from each edge node, and the global model continuously iterates through encrypted aggregation without waiting for centralized updates, enabling rapid adaptation to new scenarios such as power grid structure changes and the access of new equipment. The HCCN's pattern matching layer can dynamically load adapted models from a pre-built model library based on real-time detected abnormal patterns, avoiding diagnostic lag caused by model solidification. The weight parameters of the modal fusion gating unit and the fault rule library of KG-Forecaster are continuously optimized through federated learning and historical cases, ensuring that the model is always synchronized with the dynamic changes of the circuit system, maintaining high diagnostic accuracy over a long period of time, and solving the problems of slow updates and poor adaptation of traditional models.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention. The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the protection scope of the present invention.

Claims

1. A circuit hazard diagnosis method based on multidimensional data and edge computing, characterized in that, Includes the following steps: S1: Use a multimodal spatiotemporal feature fusion engine to collect multi-source data from current monitoring, temperature monitoring, vibration monitoring and partial discharge monitoring in real time; S2: Through the spatiotemporal cross-attention mechanism, the multi-source data is spatiotemporally aligned within the multimodal spatiotemporal feature fusion engine, and the relationship between current abrupt change points and temperature hotspots is dynamically calibrated to improve the accuracy of fault detection. S3: On the heterogeneous computing gateway at the edge, a hierarchical computing capsule network is used for data processing. The hierarchical computing capsule network automatically loads the corresponding computing model according to the detected abnormal patterns and adaptively adjusts the model parameters according to the computing resource situation to achieve the purpose of high efficiency and energy saving. S4: Utilizing a knowledge graph-driven fault prediction system, combined with a federated evolutionary learning mechanism, it predicts the propagation paths of potential faults and hidden dangers in circuits and provides corresponding suggested measures, thereby achieving early warning and preventive maintenance of circuit hidden dangers.

2. The circuit hazard diagnosis method based on multidimensional data and edge computing according to claim 1, characterized in that: The real-time acquisition process of multi-source data by the multimodal spatiotemporal feature fusion engine in step S1 includes: A heterogeneous sensing network is formed by deploying high-precision current sensing devices, distributed temperature monitoring arrays, microelectromechanical system vibration sensors, and ultra-high frequency partial discharge detection modules at key circuit nodes. Each sensor node achieves parallel acquisition of multi-dimensional data based on a timestamp synchronization mechanism. The acquired data is preprocessed by time-frequency domain conversion, noise suppression, and feature enhancement, and then input to the feature alignment layer of the spatiotemporal cross-attention mechanism. The alignment layer learns the spatial correlation weights and time series dependencies of each monitoring point in the circuit topology, dynamically calibrates the sampling frequency deviation and spatial position offset of different modal data, and then the learnable modal fusion gating unit adaptively allocates the feature contribution of each type of monitoring data according to the circuit operating state, finally generating a fusion data matrix containing transient changes and steady-state features. The weight parameters of the modal fusion gating unit are iteratively optimized based on historical fault case data using a federated learning framework.

3. The circuit hazard diagnosis method based on multidimensional data and edge computing according to claim 1, characterized in that: The spatiotemporal cross-attention mechanism in step S2 includes: a time synchronization calibration layer, a spatial correlation modeling unit, and a dynamic weight allocator. The time synchronization calibration layer unifies multi-source heterogeneous data to the same time scale through a sliding time window and interpolation algorithm; the spatial correlation modeling unit constructs a spatial adjacency matrix of monitoring points based on the circuit topology to quantify the spatial coupling degree of different modal data; and the dynamic weight allocator calculates the spatiotemporal response correlation between current abrupt changes and temperature hotspots through a three-dimensional attention tensor. The dimensions of the three-dimensional attention tensor correspond to time offset, spatial distance, and feature similarity, respectively. The attention weight matrix generated by the corresponding tensor can dynamically enhance the feature mapping relationship between abnormal current fluctuations and the corresponding temperature rise in the region, and suppress non-correlated noise signals, thereby achieving accurate alignment of multi-source data in the spatiotemporal dimensions and enhancement of key features.

4. The circuit hazard diagnosis method based on multidimensional data and edge computing according to claim 1, characterized in that: The hierarchical computing capsule network in step S3 is deployed on an edge heterogeneous computing gateway that integrates CPU, FPGA and NPU. The hierarchical computing capsule network includes: a feature parsing layer, a pattern matching layer and a dynamic execution layer. The feature parsing layer performs hierarchical feature extraction on multimodal fusion data and generates abnormal feature vectors. The pattern matching layer performs similarity retrieval through a pre-built abnormal pattern feature map library to determine the appropriate computing model. The dynamic execution layer adaptively adjusts the number of network layer activations, feature channel dimensions, and iterative calculation accuracy based on the computing load, memory usage, and energy consumption thresholds fed back by the gateway's real-time resource monitoring module, through model parameter pruning algorithms and computing unit scheduling strategies. In low-resource scenarios, only basic capsule units are enabled for rapid screening, while in high-resource scenarios, enhanced capsule clusters are activated to achieve in-depth analysis.

5. The circuit hazard diagnosis method based on multidimensional data and edge computing according to claim 1, characterized in that: In step S4: the knowledge graph-driven fault inference system package is used, combined with the fault prediction sub-models trained locally at each edge node in the federated evolutionary learning mechanism, and aggregated to generate global optimized model parameters under the privacy protection protocol. The federated evolutionary learning mechanism iteratively optimizes the weights of sub-models through a non-dominated sorting genetic algorithm and dynamically adjusts the fault propagation probability matrix. This enables the system to complete the localization of potential fault points, the sorting of multi-path propagation probabilities, and the generation of preventive maintenance plans that include component replacement priorities and operation sequences in a short period of time. The aggregation process of global model parameters uses homomorphic encryption technology to ensure data security.

6. The circuit defect diagnosis method based on multidimensional data and edge computing according to claim 5, characterized in that: In step S4, the knowledge graph-driven fault inference system realizes circuit hazard prediction through a multi-level knowledge graph architecture. The circuit topology layer contains the component connection relationship and physical parameter ontology, the fault rule layer stores the association rules based on historical cases and expert experience, and the failure mode layer records the feature vectors and propagation thresholds of typical faults. By using a graph neural network to perform entity linking and path search on the abnormal features output from the edge, and combining the knowledge reasoning engine to calculate the association confidence of each potential fault node, the distributed reasoning sub-model based on the non-dominated sorting genetic algorithm in the federated evolutionary learning mechanism is called. Under the protection of the homomorphic encryption protocol, the local inference results of each edge node are aggregated to generate hierarchical early warning information containing the probability of fault occurrence, multi-path propagation timing and impact range. Based on the circuit criticality score, a preventive maintenance plan containing component replacement priority and operation steps is automatically generated.

7. A circuit hazard diagnosis system based on multidimensional data and edge computing, characterized in that, The circuit defect diagnosis method based on multidimensional data and edge computing, as described in any one of claims 1-6, includes: The multimodal spatiotemporal feature fusion engine module is used to collect and fuse multi-source data from current, temperature, vibration and partial discharge monitoring in real time; The spatiotemporal cross-attention module, connected to the multimodal spatiotemporal feature fusion engine module, is used to perform spatiotemporal alignment of multi-source data and dynamically calibrate the relationship between current abrupt changes and temperature hotspots. The edge heterogeneous computing gateway is configured with a hierarchical computing capsule network, which is used to automatically load the corresponding computing model and adaptively adjust the model parameters according to the detected abnormal patterns. The knowledge graph-driven fault prediction system communicates with the edge heterogeneous computing gateway and combines a federated evolutionary learning mechanism to predict circuit faults and potential hazard propagation paths, and outputs preventive maintenance suggestions.

8. The circuit hazard diagnosis system based on multidimensional data and edge computing according to claim 7, characterized in that: The multimodal spatiotemporal feature fusion engine module includes: A heterogeneous sensing network, comprising: a current sensing device, a distributed temperature monitoring array, a MEMS vibration sensor, and an ultra-high frequency partial discharge detection module; The time-frequency domain preprocessing unit is used to perform time-frequency domain transformation, noise suppression, and feature enhancement on the acquired data; The spatiotemporal feature alignment layer achieves parallel acquisition of multi-dimensional data based on a timestamp synchronization mechanism and outputs a fused data matrix; The modality fusion gating unit iteratively optimizes weight parameters through a federated learning framework to adaptively allocate the feature contribution of each type of monitoring data.

9. The circuit malfunction diagnosis system based on multidimensional data and edge computing according to claim 7, characterized in that: The spatiotemporal cross-attention module includes: The time synchronization calibration layer unifies the time scale of multi-source heterogeneous data through a sliding time window and interpolation algorithm; The spatial correlation modeling unit constructs a spatial adjacency matrix of monitoring points based on the circuit topology to quantify the spatial coupling degree; The dynamic weight allocator uses a three-dimensional attention tensor to calculate the spatiotemporal response correlation between current abrupt changes and temperature hotspots, and outputs an attention weight matrix. The edge heterogeneous computing gateway integrates CPU, FPGA and NPU; The hierarchical computing capsule network includes: Feature parsing layer, used for hierarchical extraction of abnormal feature vectors; The pattern matching layer uses a similarity search based on an anomaly pattern feature map library to determine the computational model. The dynamic execution layer adaptively adjusts the number of network layer activations, feature channel dimensions, and iterative calculation accuracy based on the computing load, memory usage, and energy consumption thresholds fed back by the real-time resource monitoring module.

10. The circuit malfunction diagnosis system based on multidimensional data and edge computing according to claim 7, characterized in that: The knowledge graph-driven fault deduction system includes: The federated evolutionary learning module uses a non-dominated sorting genetic algorithm to aggregate fault prediction sub-models of each edge node under a privacy protection protocol, and dynamically adjusts the fault propagation probability matrix. A globally optimized parameter aggregator uses homomorphic encryption to ensure data security. A preventative maintenance plan generator is used to output maintenance plans that include component replacement priorities, operation sequences, and scope of impact. The multi-level knowledge graph architecture includes: a circuit topology layer that stores the connection relationships of components and the ontology of physical parameters; a fault rule layer that stores the association rules of historical cases and expert experience; and a failure mode layer that records typical fault feature vectors and propagation thresholds. The graph neural network inference engine is used to perform entity linking and path search on abnormal features, and combines the distributed inference sub-model output by the federated evolutionary learning module to generate hierarchical early warning information.

Images