Low-voltage standardized cabinet and contact abnormity tracing method thereof

Abstract

The invention provides a low-voltage standardized cabinet and a contact abnormity tracing method thereof, and the method comprises the steps: firstly obtaining the contact state and environment change characteristics of a low-voltage standardized cabinet contact; performing time sequence alignment on multi-sensor data, extracting a core feature sequence, inputting the core feature sequence into a constructed graph neural anomaly detection model, tracking and identifying local feature disturbance and a diffusion path thereof caused by contact anomaly based on a graph attention network and by utilizing a node state propagation mechanism, and obtaining a graph neural anomaly detection result; obtaining a final state embedding vector of each node at time t, and inputting the final state embedding vector into an abnormal category identification module to obtain prediction category probability output; and when the probability deviation exceeds a dynamic threshold value, triggering local anomaly detection, automatically backtracking sensing data in an anomaly time period, extracting corresponding features to construct a multi-modal anomaly description vector, inputting the multi-modal anomaly description vector into an abnormal state detection module, and outputting an anomaly type. According to the invention, high-precision monitoring, multi-dimensional modeling and safe traceable management of the connection state of the contacts in the cabinet are realized.

Description

Technical Field

[0001] This invention belongs to the field of low-voltage standard cabinet anomaly detection technology based on deep learning, and particularly relates to a low-voltage standard cabinet and its contact anomaly tracing method. Background Technology

[0002] With the continuous expansion of urban power grids and the acceleration of energy transformation, low-voltage distribution systems are undergoing a profound transformation from the traditional passive maintenance model of "operation-fault-maintenance" to an intelligent maintenance model of "state perception-prediction and early warning-proactive operation and maintenance." Against this backdrop, as a crucial carrier in the distribution network for executing switching, current switching, and distribution control, the operational reliability of low-voltage switchgear directly affects the continuity, security, and stability of power supply to end users of the entire power system. In recent years, numerous actual operation and maintenance cases and accident statistics have shown that the contact connection points inside low-voltage switchgear are prone to abnormal phenomena such as poor contact, wear and oxidation, heat accumulation, and arc discharge during long-term operation. These seemingly minor deterioration behaviors often become key causes of electrical faults, equipment burnout, and even fire and explosion accidents. Therefore, how to achieve early identification, early warning, location, and traceability of subtle abnormalities in the operating status of low-voltage switchgear has become a core technical challenge in ensuring the safety of intelligent distribution systems.

[0003] Traditional low-voltage switchgear maintenance primarily relies on manual periodic inspections and regular infrared thermography and partial discharge detection. These methods suffer from drawbacks such as long detection cycles, high response delays, and heavy reliance on experience. In most scenarios, maintenance personnel struggle to detect the slow increase in contact resistance during the initial deterioration stages of equipment, and also find it difficult to capture nonlinear characteristics such as minute hot spot changes and abnormal sound patterns. Once the contact resistance reaches a critical value, it can trigger a rapid temperature rise, insulation aging, and even arcing breakdown within a short period, leading to widespread tripping or damage to the power distribution system. Furthermore, existing monitoring methods generally employ a single signal path, resulting in limited information dimensions and making it difficult to accurately determine the source and evolution trend of anomalies, severely impacting the accuracy of fault location and the timeliness of maintenance response.

[0004] With the continuous maturation of technologies such as the Internet of Things, edge computing, and artificial intelligence, the intelligence level of power systems is constantly improving, providing a technological foundation for achieving "refined perception, intelligent diagnosis, and reliable management" of power distribution equipment status. However, in existing research and engineering applications, several key issues still urgently need to be addressed: First, insufficient perception dimensions make it difficult to comprehensively characterize contact degradation features; second, limited modeling capabilities lack methods for depicting anomaly propagation paths in electrical topologies; third, weak edge intelligence, with models relying on centralized deployment, resulting in limited real-time performance and privacy protection capabilities; and fourth, a lack of unified and reliable data management and anomaly tracing mechanisms. These problems make existing systems inadequate in dealing with nonlinear anomalies, tracing sudden events, and intelligent linkage decision-making in complex operating environments, necessitating a systematic technical solution. Summary of the Invention

[0005] To address the aforementioned issues, this invention utilizes multiple types of sensors, including temperature, gas, micro-vibration, infrared thermal imaging, and acoustic signature sensors, deployed within a standardized cabinet to achieve three-dimensional perception of the contact operation status and the surrounding electrical environment, overcoming the dimensional limitations of single monitoring methods. Simultaneously, it introduces a graph neural network modeling method to abstract the physical structure and operational logic within the low-voltage cabinet into a graph structure, constructing topological relationships between contacts, current paths, and measurement points to accurately depict the disturbance propagation process of abnormal sources. Through the deployment of edge intelligent nodes, local model operation and anomaly identification are supported, improving response time and data security. Furthermore, combined with a federated learning mechanism, joint training of multi-cabinet models is achieved, enhancing the system's generalization ability while protecting data privacy. Finally, a trusted traceability chain is constructed using blockchain technology, enabling verifiable recording of equipment status, model version, and abnormal events, improving the reliability of subsequent fault analysis and responsibility determination.

[0006] This invention proposes a low-voltage standardized switchgear and a method for tracing abnormal contacts, comprising the following processes: S1, based on multiple sensors deployed inside the cabinet, acquires the contact state and environmental change characteristics of the contacts during the "closing-current-disconnection" process of the low-voltage standardized cabinet, including temperature changes, fine structural vibrations, local hot spot images, characteristic gas concentrations released by local arc discharge of contacts, and acoustic signature analysis characteristics. S2 performs time-series alignment on multi-sensor data; a sparse autoencoder is used to reduce the dimensionality of the data and extract the core feature sequence. S3, input the core feature sequence into the constructed graph neural network anomaly detection model. The graph neural network anomaly detection model is constructed with the monitoring site as the node and the physical connection, electrical coupling, structural resonance path and signal propagation path as the edge set. The multimodal features in the core feature sequence are mapped to the graph nodes. Based on the graph attention network, the node state propagation mechanism is used to track and identify the local feature perturbation caused by the contact anomaly and its diffusion path, and obtain the final state embedding vector of each node at time t. S4, embed the final state into the vector input anomaly category recognition module to obtain the first... i The predicted class probability output for each node. ;when When the deviation exceeds the dynamic threshold, local anomaly detection is triggered; the sensing data within the abnormal period is automatically traced back to extract voiceprint features, thermal infrared image features, and microseismic features to construct a multimodal anomaly description vector; S5 inputs the multimodal anomaly description vector into the anomaly state detection module and outputs the anomaly type.

[0007] Preferably, the specific acquisition process of S1 includes: Thermal sensors collect temperature data from the contact area. The sensor is mounted around the metal base of the moving and stationary contacts; The micro-vibration sensor is mounted on the metal base plate and contact fixing assembly of the distribution cabinet to capture subtle structural vibrations and obtain data. ; A multi-channel MOS gas sensor array was used to detect partial discharge products and obtain gas concentration vector data. ; A non-contact infrared array thermal imager is used, installed inside the cabinet door, facing the contacts and their conductive path area, to capture local hotspot images. ; The acoustic signature sensor is equipped with a linear array MEMS microphone array to acquire short-term audio energy mutations caused by abnormal processes during the "closing-current flow" phase, thereby obtaining the original audio frame vector. .

[0008] Preferably, the specific process of S2 is as follows: For data and exist[ Within the interval, a sliding window averaging process was performed to obtain the mean-processed micro-vibration data. and the audio frame data after mean processing. ; For data Data alignment was performed using the nearest-value hold method to obtain aligned temperature data. and aligned gas concentration data ; hotspot images The data is flattened to obtain the flattened hotspot vector data. Based on this data dimension, a unified feature input is obtained by concatenating the data. ; The system will unify feature vectors The data is fed into a sparse autoencoder for dimensionality compression and representation modeling to obtain... Low-dimensional feature representation after time-point encoding ,and ; Total sampling time; Finally, the compressed core feature sequence of all consecutive frames is: .

[0009] Preferably, the neural abnormality detection model in S3 is specifically as follows: Graph Structure Definition and Node Mapping: The physical entities in the low-voltage switchgear are abstracted into a graph according to their structure and function. ; set of nodes in the graph , The total number of nodes equals the number of all modeled key locations in the system; each node Represents an independent physical functional unit, corresponding to a specific component in a low-voltage switchgear; edge set , Represents a node and There is a physical / electrical relationship between them; Feature embedding and node initialization: Based on the core feature sequence, it is assigned and mapped to a graph structure. The initial state vector of the corresponding node in the graph is used to realize the structure-aware data binding process; each node in the graph The features are initialized from the associated core feature sequences. Then, the core feature sequences of each node are fused into a vector of uniform dimension, and linear mapping and concatenation are performed to reduce the dimensionality, resulting in the node... Integrated features The final node features constitute the initial feature matrix H of the input layer of the graph neural network, which is used as the input layer of the model to complete the feature binding process. Edge weight modeling and adjacency matrix construction: The edge weights between nodes represent the strength of physical or informational coupling relationships, using a weighted adjacency matrix. This represents the connection relationships and propagation weights between nodes. Then, the node is calculated based on spatial distance, connection type, and energy coupling degree. With nodes Connection strength Finally, the connection strength between each pair of nodes is obtained, which together form a weighted adjacency matrix. .

[0010] Preferably, in step S3, a multi-layer graph attention network (GAT) mechanism is used to model the interaction propagation and adaptive fusion between nodes. The initial feature matrix H of each node serves as the input feature representation of the graph neural network, and the adjacency matrix A provides the connection topology and propagation weights between nodes. Through the edge connection relationships and weight information in the graph structure, information is guided to propagate from the abnormal source node to adjacent nodes, capturing the diffusion path of the multimodal disturbance signal induced by abnormal contact resistance in the system. Specifically, this includes: The initial feature matrix H of the nodes in the input layer is input into the GAT layer, and the output is... To achieve local perturbation perception and neighborhood propagation, new features are generated after each node interacts with its neighbors. LeakyReLU is used as the activation function. Next The input is fed into the second-layer attention map convolution, using LeakyReLU as the activation function to achieve wider propagation and deeper feature extraction; the GAT structure is compressed into a two-layer attention propagation structure, with parameters denoted as... ; After propagation through an L-layer GAT model, the final state embedding vector of each node at time t is obtained. .

[0011] Preferably, the anomaly category identification module in S4 embeds a vector of the final state of each node at time t. Using a fully connected layer and a softmax function as input, fault prediction is performed to obtain the first... i The predicted class probability output for each node. The predicted probability distribution includes normal, poor contact, arc discharge, and mechanical loosening results; Within each detection cycle, record the actual detection data label of the current node. , and model output results Compare and calculate the absolute deviation vector. ,like The L1 norm exceeds the dynamic threshold When an anomaly is detected, an anomaly detection response is triggered, initiating a data backtracking mechanism for the anomaly period. The backtracking time window is set to [time value missing]. ,in, The time window size is defined; within this interval, voiceprint features, thermal infrared image features, and microseismic features are collected to construct a multimodal anomaly description vector.

[0012] Preferably, the abnormal state detection module in S5 takes the multimodal abnormal description vector as input and outputs the final first... iAbnormal event classification results for each node It includes normal conditions, poor contact, arc discharge, and mechanical loosening; the specific structure of the abnormal state detection module is as follows: Input layer: The input consists of a multimodal description vector that integrates resistance, temperature, vibration, and acoustic signature, as well as the prediction results from the graph neural network; Feature normalization layer: Normalizes the sub-features of each modality in the multimodal description vector to maintain the consistency of feature scale; Anomaly Feature Embedding Layer: A linear layer is used to reduce the dimensionality of the multimodal description vector and map it to a unified representation space to obtain the embedded representation vector; The graph neural network output fusion layer concatenates the output of the graph neural network model with the embedded representation vector. This concatenation is then fed into the fusion classifier to calculate the final classification probability. Classifier: The final fault type, i.e., the classification result of the abnormal event, is obtained by using the extreme value function argmax. And output it.

[0013] Preferably, an integrated edge intelligent node is integrated into each standardized cabinet, combining the graph neural network anomaly detection model, anomaly category identification module, and anomaly state detection module into a system model. This model is deployed in a lightweight form on the edge computing node of each low-voltage cabinet for local operation. The FedAvg federated learning algorithm is used to coordinate multiple low-voltage cabinet edge nodes for distributed modeling. The specific process is as follows: First, federated average model aggregation is performed; in each federated round r, the set of participating edge nodes is K, and the k-th node obtains the model parameters through local training. Then the central server will have parameters for all nodes. Perform weighted aggregation to obtain the updated global model parameters. ; Secondly, a homomorphic encryption communication mechanism is used; to prevent model parameters from being leaked during communication, the system introduces an additive homomorphic encryption function. It satisfies the additivity property with respect to the input variables; if each of the two nodes has an error term. and Then the summation of their error functions is equivalent to calculating the sum of their errors all at once. That is, the sum of the local error function values ​​of node i and node j is equivalent to the global error function value calculated by adding their error parameters together. Edge nodes locally... Encrypted uploads are performed, allowing the central node to complete aggregation calculations without decryption, thus ensuring parameter security.

[0014] Preferably, an early warning strategy is formed by cluster analysis of the anomaly description vectors of different cabinets. The specific process is as follows: After each round of federated training, the central node will collect the anomaly description vector sets uploaded by the edge nodes of each cabinet. : As input, cluster analysis is performed to uncover spatial or temporal patterns of anomaly propagation. The K-means clustering algorithm is used, and the clustering results are used to construct a regional early warning model. Based on the anomaly density and frequency of occurrence of each cluster, the system outputs three warning levels: red, orange, and blue, driving the background scheduling response strategy. The number of clusters is set to K=3, corresponding to the three warning levels: red, orange, and blue, respectively. A red warning level indicates anomaly activity and density, suggesting immediate on-site inspection. An orange warning level indicates the anomaly may spread, requiring planned inspections. A blue warning level indicates a minor anomaly or peripheral disturbance, requiring continuous monitoring by the background system.

[0015] Preferably, nodes based on S5 output anomaly types include node number, fault category, anomaly occurrence time, anomaly event description vector, multimodal fusion features, model version summary, and resistance detection value. These are packaged into an on-chain data structure, and data blocks are recorded across all network nodes using distributed ledger technology. In this process, a data digest is generated using a hash function and uploaded to the blockchain to form a block structure. ,in, This hash value is used to build the chain-like dependency relationship between each block in the blockchain, ensuring that on-chain data is immutable and traceable throughout the entire chain. Represents a random number. This represents the Merkle root of all data packets in the current block; When the system detects an alarm event or receives a maintenance instruction request, the blockchain module automatically backtracks historical chain data, extracts the abnormal event chain, verifies the consistency between the model version and the data, and calculates the credibility index.

[0016] Should The structure is automatically submitted to the consortium blockchain network via a smart contract mechanism, and confirmed through consensus by multiple trusted nodes. Once consensus is reached, This will be formally written into the distributed ledger, forming a time-sequential chain of anomaly event records. Subsequently, when operations and maintenance personnel trace anomalies and analyze the reliability of the model's source tracing, they can use this information... The structure allows for rapid location of event information and model prediction basis, ensuring data integrity and verifiability.

[0017] Compared with the prior art, the present invention has the following innovative features: (1) Breaking through the limitations of traditional single sensing dimension at the perception level, a multi-modal fusion mechanism of heat, sound, air, vibration and electricity is constructed, which can effectively capture a variety of abnormal modes such as contact deterioration, electric arc discharge and mechanical loosening, greatly improving detection sensitivity and discrimination accuracy.

[0018] (2) By introducing graph neural networks to perform graph topology modeling of the cabinet structure and combining attention mechanism to learn abnormal feature propagation, it is possible to trace the source of abnormal root causes and characterize the diffusion path, breaking through the limitation of existing shallow models that can only perform abnormal discrimination.

[0019] (3) The system implements a deployment architecture that combines edge computing and federated learning, enabling the model to have local prediction capabilities and distributed collaborative evolution capabilities, improving the system's real-time performance and ensuring privacy and security.

[0020] (4) The data on-chain and abnormal event recording mechanism based on blockchain can build a data chain that is traceable throughout the entire life cycle, providing objective and reliable data basis for subsequent accident reproduction and regulatory audit.

[0021] (5) A refined mapping logic from multimodal sensing features to physical failure mechanism was constructed. The system can accurately distinguish whether the abnormal contact resistance is caused by process defects due to improper bolt tightening torque or structural deformation due to insufficient conductor mechanical strength. This solves the problem of difficulty in identifying the root cause of heat generation and loosening in traditional operation and maintenance. It provides operation and maintenance personnel with a definite basis for maintenance decision-making, namely, to perform targeted operations such as removing the oxide layer on the contact surface, adjusting the torque or replacing the deformed busbar. This effectively avoids the risk of ineffective maintenance or fault expansion caused by blindly tightening screws. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall implementation logic of the present invention.

[0023] Figure 2 This is a flowchart of the anomaly tracing and detection process of the present invention.

[0024] Figure 3 This is a simulated attention weight heatmap for an embodiment of the present invention.

[0025] Figure 4 This is a two-dimensional visualization of the embedding space of abnormal nodes in an embodiment of the present invention. Detailed Implementation

[0026] This invention provides a low-voltage standardized switchgear and a method for tracing abnormal contacts, as follows: Figure 1 As shown. Its main process is as follows: Multimodal signal acquisition and fusion: Accurately acquire contact state and environmental change characteristics of low-voltage standardized switchgear during operation. The system deploys multiple sensors to obtain temperature changes, subtle structural vibrations, characteristic gas molecules released by local arc discharge at the contacts, and acoustic signature analysis features. A timestamp unification mechanism is used to align the multi-sensor data in time sequence; a sparse autoencoder is employed to reduce the dimensionality of the data and extract core features; and a multimodal operating characteristic trajectory of the standardized switchgear during the "closing-current-disconnection" process is constructed.

[0027] Data-driven anomaly identification and detection: A graph neural network (GNN) anomaly detection model based on a two-layer graph attention mechanism for low-voltage standard switchgear is designed. This model utilizes graph neural network technology to achieve structured modeling and evolutionary analysis of contact resistance anomalies in low-voltage switchgear. Important nodes within the switchgear are abstracted as nodes in a graph. An electrical structure diagram of the low-voltage switchgear is constructed based on physical connections and electrical relationships. The model incorporates a graph attention network (GAT) to map multimodal features to graph nodes. A node state propagation mechanism is used to track and identify local feature disturbances and their propagation paths caused by contact (contact resistance) anomalies. The GNN model output is compared with actual monitoring data. When the prediction deviation exceeds a dynamic threshold, local anomaly detection is triggered. Sensing data within the anomaly period is automatically backtracked, and acoustic signature features, thermal infrared image features, and microseismic features are extracted to construct a multimodal anomaly description vector. The anomaly type (poor contact, arc discharge, mechanical loosening) is determined and uploaded to the blockchain record.

[0028] Trusted training of nodes in edge collaboration: The system integrates a unified edge intelligent node within each standardized cabinet, combining a graph neural network anomaly detection model, anomaly category recognition module, and anomaly state detection module into a system model. This model is deployed in a lightweight form on the edge computing node of each low-voltage cabinet for local operation. The FedAvg federated learning algorithm is employed to coordinate distributed modeling across multiple low-voltage cabinet edge nodes. Homomorphic encryption and differential privacy mechanisms are used to ensure the privacy and security of the model. The system generates early warning strategies by clustering and analyzing the anomaly description vectors of different cabinets.

[0029] Blockchain anomaly tracing mechanism: Key data is recorded on the blockchain, including node number, fault type, time of anomaly occurrence, anomaly event description vector, multimodal fusion features, model version summary, and resistance detection value. This real-time on-chain recording forms an immutable historical record chain. When a fault occurs or a higher-level maintenance request is made, the blockchain record traces the anomaly occurrence process, verifying the reliability of model predictions and tracing the source. Simultaneously, a smart contract triggers a voting mechanism, with multi-node consensus determining whether to initiate automatic alarms or on-site maintenance operations.

[0030] The invention will be further described below with reference to specific embodiments.

[0031] I. Multimodal signal acquisition and fusion This invention designs a multi-sensor collaborative acquisition system for low-voltage switchgear operation, focusing on the local physical response characteristics caused by changes in contact resistance. This system serves as the foundational support layer for intelligent identification, focusing on the accurate acquisition and effective fusion of five types of signals—thermal, electrical, acoustic, gaseous, and vibration—throughout the entire process of contact closing, conduction, and disconnection, forming a high-dimensional state description vector that can be used for subsequent model analysis.

[0032] Multi-sensor deployment and raw data acquisition: Five types of sensor modules are deployed in each low-voltage standardized cabinet to collect physical signals from different dimensions. Each type of sensor... The observed values ​​are defined as follows: Thermal sensors collect temperature data from the contact area. It employs a high-precision thermistor with a response time of less than 250ms. The sensor is mounted around the metal base of the moving and stationary contacts, and the sampling frequency is set to 1Hz.

[0033] The micro-vibration sensor uses a high-sensitivity MEMS sensor with a sampling frequency set to 100Hz. It is mounted on the metal base plate and contact fixing assembly of the distribution cabinet to capture the vibration of minute structures and the release of shock waves, obtaining the data as follows: .

[0034] A gas sensor array, employing a multi-channel MOS gas sensor array, is used to detect partial discharge products. The sampling frequency is 0.2Hz. The gas concentration vector is defined as follows: .

[0035] Infrared thermal imaging module. Employing a non-contact infrared array thermal imager, installed inside the cabinet door, facing the contacts and their conductive paths, it captures localized hotspot images in real time. .

[0036] The acoustic signature sensor is equipped with a linear array MEMS microphone group with a bandwidth of 20Hz-20kHz and a sampling frequency set to 16kHz. It is used to analyze short-term audio energy fluctuations caused by abnormal processes such as contact instability, breakdown, and arc discharge during the "closing-current flow" period. The original audio frame vector is... .

[0037] Time Alignment and Multimodal Fusion: Due to significant differences in the sampling periods of various sensors, the system introduces a unified master time reference synchronization mechanism during deployment, achieving data alignment through local edge processing units. The master clock uses... To align the window, set all sensors to output data at that time point.

[0038] For high-frequency data and exist[ Within the interval, a sliding window averaging process was performed to obtain the mean-processed micro-vibration data. and the audio frame data after mean processing.

[0039] For low-frequency data Data alignment was performed using the nearest-value hold method to obtain aligned temperature data. and aligned gas concentration data ; hotspot images The data is flattened to obtain the flattened hotspot vector data. Based on this data dimension, a unified feature input is obtained by concatenating the data. ; The system will unify feature vectors The data is fed into a sparse autoencoder for dimensionality compression and representation modeling to obtain... Low-dimensional feature representation after time-point encoding ,and ; This represents the total sampling time.

[0040] Finally, the compressed feature sequence of all consecutive frames is This is used to support subsequent graph neural network structure modeling and anomaly comparison.

[0041] II. Data-driven anomaly backtracking identification and detection After completing time alignment and feature extraction of multi-source data, this invention employs graph neural networks (GNNs) to model and embed features of the internal structure of the low-voltage switchgear and the sensing data. The overall process is as follows: Figure 2 As shown.

[0042] Graph Structure Definition and Node Mapping: The physical entities in the low-voltage switchgear are abstracted into a graph according to their structure and function. The set of nodes in the graph , The total number of nodes equals the number of all modeled key locations in the system. Each node This represents an independent physical functional unit, corresponding to a specific component in a low-voltage switchgear, with a clearly defined location and functional attributes (including contacts, cable and current paths, temperature measuring points, partial discharge monitoring areas, and vibration and acoustic signature acquisition points). Edge set , Represents a node and There are physical / electrical relationships between them, representing physical connections, electrical couplings, structural resonance paths, and signal propagation paths.

[0043] Feature embedding and node initialization: Based on the extracted fused feature vector (compressed feature sequence) for each second. Assign and map it to a graph structure The initial state vector of the corresponding node in the graph is used to implement a structure-aware data binding process. Specifically, each node in the graph... The feature initialization comes from the associated multimodal feature set, i.e. Then, the multimodal feature sets of each node are fused into a vector of uniform dimension, and linear mapping and concatenation are performed to reduce the dimensionality, resulting in the node... Integrated features (represents a node) (Integration characteristics at time t). The final obtained node features constitute the initial feature matrix H of the input layer of the graph neural network, which serves as the input layer of the model to complete the feature binding process. This process achieves high-dimensional feature alignment from the multi-source perceptual fusion features F in S1 to the node space of the graph structure G.

[0044] Edge weight modeling and adjacency matrix construction: The edge weights between nodes represent the strength of physical or informational coupling relationships, using a weighted adjacency matrix. This represents the connection relationships and propagation weights between nodes. Then, the node is calculated based on spatial distance, connection type, and energy coupling degree. With nodes Connection strength Finally, the connection strength between each pair of nodes is obtained, which together form a weighted adjacency matrix. .

[0045] Node Propagation Modeling Based on Graph Attention Mechanism: Based on the obtained initial node feature matrix H and adjacency matrix A, this invention designs a multi-layer graph attention network (GAT) mechanism to model the interaction propagation and adaptive fusion between nodes. The initial node feature matrix H serves as the input feature representation of the graph neural network, and the adjacency matrix A provides the connection topology and propagation weights between nodes. Through the edge connections and weight information in the graph structure, information is guided to propagate from the abnormal source node to adjacent nodes, capturing the diffusion path of multimodal disturbance signals induced by abnormal contact resistance in the system. Specifically: The initial feature matrix H of the nodes in the input layer is input into the GAT layer, and the output is... To achieve local perturbation perception and neighborhood propagation, new features are generated after each node interacts with its neighbors. LeakyReLU is used as the activation function.

[0046] Next The input is fed into the second-layer attention map convolution, using LeakyReLU as the activation function to achieve wider propagation and deeper feature extraction. In this invention, the GAT structure is compressed into a structure containing only two attention propagation layers. The parameters are denoted as... .

[0047] This mechanism enables information to propagate not only based on connectivity but also to adaptively identify which neighboring nodes are more important. In particular, when "contact resistance anomalies" gradually cause a series of characteristic disturbances such as temperature rise, acoustic signature changes, and enhanced partial discharge, GAT can track their propagation path and learn the evolution path of the anomalous links.

[0048] State Evolution and Anomaly Focusing: After propagation through an L-layer GAT model, the final state embedding vector of each node at time t is obtained. These states will serve as input to the anomaly category identification module, that is... As input, fault prediction is performed using a fully connected layer and a softmax function to obtain the... i The predicted class probability output for each node. (Including the predicted probability distribution of results for normal, poor contact, arc discharge, and mechanical loosening).

[0049] Furthermore, to achieve a localized anomaly detection mechanism, this invention introduces a prediction bias comparison strategy. Within each detection cycle, the system records the actual detection data label of the current node. , and model output results Compare and calculate the absolute deviation vector. ,like The L1 norm exceeds the dynamic threshold When this occurs, an anomaly detection response is triggered. This includes the threshold function. The system adaptively updates based on its operating status to suppress false alarm rates under different operating conditions.

[0050] Once an anomaly detection is triggered, this system automatically initiates the anomaly period-aware data backtracking mechanism, with the backtracking time window set to [value missing]. ,in, The time window size; within this interval, the following three types of key modal change information are collected: 1) Voiceprint characteristics, collecting node voiceprint signals Extracting signals with abrupt changes in spectral structure and non-steady-state spikes; 2) Thermal infrared image features, analysis of nodal thermal images Morphological evolution and temperature rise trend of the mid-heat spot region; 3) Microseismic characteristics, extracting vibration sensing data Modeling the loosening of minute structures by short-term energy changes.

[0051] Next, the three types of modal features mentioned above are concatenated to construct an anomaly multimodal description vector: .

[0052] The above vector The input is fed into the anomaly detection module, where anomalies are categorized, and the final output is given. i Abnormal event classification results for each node (Including normal conditions, poor contact, arc discharge, and mechanical loosening). The abnormal state detection module structure is as follows: Input layer: The input consists of a multimodal description vector that integrates resistance, temperature, vibration, and acoustic signature, as well as the prediction results from the graph neural network; Feature normalization layer: Normalizes the sub-features of each modality in the multimodal description vector to maintain the consistency of feature scale; Anomaly Feature Embedding Layer: A linear layer is used to reduce the dimensionality of the multimodal description vector and map it to a unified representation space to obtain the embedded representation vector; The graph neural network output fusion layer concatenates the output of the graph neural network model with the embedded representation vector. This concatenation is then fed into the fusion classifier to calculate the final classification probability. Classifier: The final fault type, i.e., the classification result of the abnormal event, is obtained by using the extreme value function argmax. And output it.

[0053] Specifically, the abnormal event classification results output by the abnormal state detection module have clear physical fault mechanism indications. For the poor contact category, the system focuses on tracing contact resistance degradation caused by abnormal bolt tightening torque. Insufficient bolt tightening torque reduces the effective contact area between the moving and stationary contacts, leading to increased contact resistance and localized heating. Excessive bolt tightening torque causes crushing deformation of the copper-aluminum busbar conductors, resulting in microscopic separation of the actual contact surface. Furthermore, this category also covers electrochemical corrosion caused by rough lap surfaces, oil or paint contamination, and the galvanic effect resulting from direct copper-aluminum connections. For the mechanical loosening category, the system focuses on structural deformation caused by insufficient mechanical strength. This deformation includes bending deformation of aluminum busbars due to low tensile strength under short-circuit electrodynamic forces, and sagging displacement due to excessive installation support spacing under its own weight. These deformations directly lead to reduced phase-to-phase distance and flashover short circuits, or cause physical loosening of connection points and accelerated heating and aging, ultimately leading to insulation failure or even fire and explosion accidents.

[0054] Ultimately, the abnormal event record of this node, including node number, fault category, time of occurrence of abnormality, abnormal event description vector, multimodal fusion features, model version summary and resistance detection value, is packaged into an on-chain data structure and written into the blockchain network through a preset smart contract mechanism to ensure the integrity and traceability of the event for subsequent operation and maintenance and status analysis. III. Trusted Training Design Process for Edge Collaboration Nodes To achieve distributed intelligent evolution capabilities in low-voltage switchgear systems, this invention proposes a reliable training scheme for the system model based on edge computing and federated learning mechanisms. The system model integrates a graph neural network anomaly detection model, an anomaly category identification module, and an anomaly state detection module, constructing a multi-node collaborative perception and adaptive model update system. The input to this part comes from the output anomaly description vector. Node prediction results Exception types This serves as the input for edge training and collaborative optimization. The system model is deployed to each standardized low-voltage switchgear, and locally collected data is used as input information. The resulting anomaly description vector, node prediction results, and anomaly type are the three sets of numerical values ​​used as output.

[0055] Edge computing node deployment and local intelligent processing. The system deploys lightweight edge intelligent nodes within each standardized low-voltage cabinet, based on the output anomaly description vector. Node prediction results Exception types When a prediction bias exceeding a dynamic threshold is detected. If this is the case, it indicates that the current node may be abnormal, triggering the data backtracking and description vector construction process. The anomaly description vector is then... Combined with the currently detected anomaly type labels Composition of structured exception data packets .in The location information of the abnormal source node is obtained through three-dimensional structural modeling of the low-voltage switchgear, reflecting its relative layout position.

[0056] Furthermore, through hash functions With timestamp signature Perform digest encryption to obtain encrypted data. By writing to the local blockchain node, the immutability and reliable traceability of abnormal information records are ensured.

[0057] Finally, local edge nodes utilize recent data samples Fine-tuning system model parameters The model parameters are updated using the cross-entropy loss function and stochastic gradient descent. After training is complete, the nodes are ready to participate in the next round of federated collaborative training.

[0058] Federated Learning Collaboration Mechanism Design: To achieve collaborative model evolution and knowledge sharing across low-voltage cabinets, while ensuring user privacy and data security, this invention employs an improved FedAvg algorithm combined with a privacy protection mechanism to construct a complete distributed training process. Specifically, it includes the following steps: First, federated averaging model aggregation is performed. In each federated round r, the set of participating edge nodes is K, and the k-th node obtains the model parameters through local training. Then the central server will have parameters for all nodes. Perform weighted aggregation to obtain the updated global model parameters. .

[0059] Secondly, a homomorphic encryption communication mechanism is implemented. To prevent model parameters from being leaked during communication, the system introduces an additive homomorphic encryption function. It satisfies the additivity property with respect to input variables.

[0060] Specifically, if the two nodes each have an error term and Then, the summation of their error functions is equivalent to calculating the sum of their errors all at once. That is, the sum of the local error function values ​​of node i and node j is equivalent to the global error function value calculated by adding their error parameters together. Edge nodes locally... Encrypted uploads are performed, allowing the central node to complete aggregation calculations without decryption, thus ensuring parameter security.

[0061] Finally, differential privacy perturbation. Furthermore, the system introduces a differential privacy strategy to prevent malicious nodes from reconstructing the training samples of other nodes. Each edge node adds a standard deviation to the model weights before uploading. Gaussian noise (used to control the strength of privacy protection).

[0062] Multi-node anomaly description clustering and early warning strategy output: After each round of federated training, the central node will output the anomaly description vector sets uploaded by the edge nodes of each cabinet. : As input, cluster analysis is performed to uncover spatial or temporal patterns of anomaly propagation. The K-means clustering algorithm is employed, and the clustering results are used to construct a regional early warning model. Based on the anomaly density and frequency of occurrence of each cluster, the system outputs three warning levels: red, orange, and blue, driving the background scheduling response strategy. This invention sets the number of clusters to K=3, corresponding to the three warning levels: red, orange, and blue. A red warning level indicates active and dense anomalies, suggesting immediate on-site maintenance; an orange warning level indicates the anomaly may spread, requiring planned inspections; and a blue warning level indicates minor anomalies or peripheral disturbances, requiring continuous monitoring by the background system.

[0063] The output includes locally optimized system model parameters. That is, the model runs on each edge node i, and after several rounds of local training and updates, it outputs the local model parameters of the current node; anomaly description vector set. That is, each edge node fuses and encodes the perceived multimodal data through the system model, and outputs the abnormal event description vector of its current period. Node prediction results .

[0064] IV. Blockchain Anomaly Tracing Mechanism To achieve reliable traceability and data source tracking of abnormal contact resistance in low-voltage switchgear throughout the entire process, the system uses the federated learning results of edge nodes as input, constructs an immutable historical event chain through blockchain technology, and integrates abnormal event descriptions, model status, and original monitoring data to support intelligent traceability and operation and maintenance decisions at the equipment level.

[0065] Based on the anomaly prediction results updated by federated learning for each edge node, the anomaly description vector is... Node prediction results Encrypted data Pack key information into data structures ,in, For data timestamps, This indicates a data signature signed by the edge node's private key, used for authentication.

[0066] Ultimately Submitted to the blockchain backbone, the data blocks are recorded across all nodes in the network using distributed ledger technology. In the process, a data digest is generated using a hash function and uploaded to the blockchain to form a block structure. ,in, The hash value of the previous block is used to build the chain dependency relationship of each block in the blockchain, ensuring that the data on the chain is immutable and traceable throughout the entire chain. Represents a random number. This represents the Merkle root of all data packets in the current block. This mechanism ensures that critical anomaly prediction data is verifiable and tamper-proof.

[0067] Furthermore, when the system detects an alarm event or receives a maintenance instruction request, the blockchain module automatically backtracks historical chain data, extracts the abnormal event chain, verifies model version and data consistency, and calculates a credibility index.

[0068] in, The sim function represents the credible traceability of an anomalous event at event t, and is a multimodal description function for the similarity between vectors. This indicates the difference in model version hashes. and These are adjustable weighting coefficients. If If the system deems the event prediction unreliable, it will trigger a consensus review of the on-chain smart contract.

[0069] Specifically, the smart contract triggers a coordinated warning mechanism. The system deploys the smart contract. This is used to automatically determine on-chain whether a collaborative alarm or on-site handling should be triggered. Its logical function is:

[0070] in, A predefined set of fault categories. This indicates that an on-chain voting mechanism has been initiated. This is represented as a trust threshold. When the triggering condition is met, the contract will mobilize multiple trusted nodes to participate in voting. Based on the result, it will trigger fault handling operations: automatically issuing alarm broadcasts, scheduling manual maintenance instructions, and pushing equipment shutdown tasks.

[0071] This module fully leverages the decentralized, immutable, and traceable advantages of blockchain to construct a closed-loop data system covering the entire lifecycle of low-voltage switchgear contact resistance anomalies. By storing prediction results, multimodal anomaly characteristics, and model states on the blockchain, and combining this with smart contracts to implement automatic backtracking and collaborative alarm mechanisms, the reliability, security, and automation of the system's anomaly response are enhanced, supporting the construction of a future-oriented intelligent operation and maintenance platform for equipment.

[0072] V. Experimental Analysis like Figure 3 As shown, the distribution of attention weights among nodes in the GAT model when identifying different contact anomaly types reflects its focusing mechanism on the structured graph. Each cell (i,j) represents the attention weight of the i-th node on the j-th node. Nodes with higher weights indicate that they are more sensitive to specific anomaly propagation paths in the GAT model.

[0073] like Figure 4 The figure shows the distribution of node embedding vectors extracted based on a graph attention network (GAT) in a two-dimensional space. By reducing the dimensionality of the high-dimensional embedding vectors of all nodes, the clustering trends of nodes of different anomaly types in the embedding space can be visualized. Nodes are color-coded according to their anomaly type: blue (circle): poor contact; red (triangle): arc discharge; green (square): mechanical loosening. It can be observed from the figure that different categories of anomaly nodes exhibit a clear clustering distribution in the two-dimensional embedding space, with good discriminative boundaries between categories. This demonstrates that the constructed graph neural network model can effectively capture the complex relationships between nodes in terms of electrical topology, sensory feature evolution, etc., and achieves deep representation and feature focusing of node states through layer-by-layer attention propagation.

[0074] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0075] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A low-voltage standardized switchgear and a method for tracing abnormal contacts, characterized in that, The process includes the following: S1, based on multiple sensors deployed inside the cabinet, acquires the contact state and environmental change characteristics of the contacts during the "closing-current-disconnection" process of the low-voltage standardized cabinet, including temperature changes, fine structural vibrations, local hot spot images, characteristic gas concentrations released by local arc discharge of contacts, and acoustic signature analysis characteristics. S2 performs time-series alignment on multi-sensor data; a sparse autoencoder is used to reduce the dimensionality of the data and extract the core feature sequence. S3, input the core feature sequence into the constructed graph neural network anomaly detection model. The graph neural network anomaly detection model is constructed with the monitoring site as the node and the physical connection, electrical coupling, structural resonance path and signal propagation path as the edge set. The multimodal features in the core feature sequence are mapped to the graph nodes. Based on the graph attention network, the node state propagation mechanism is used to track and identify the local feature perturbation caused by the contact anomaly and its diffusion path, and obtain the final state embedding vector of each node at time t. S4, embed the final state into the vector input anomaly category recognition module to obtain the first... i The predicted class probability output for each node. ;when When the deviation exceeds the dynamic threshold, local anomaly detection is triggered; the sensing data within the abnormal period is automatically traced back to extract voiceprint features, thermal infrared image features, and microseismic features to construct a multimodal anomaly description vector; S5 inputs the multimodal anomaly description vector into the anomaly state detection module and outputs the anomaly type.

2. The low-voltage standardized switchgear and its contact anomaly tracing method as described in claim 1, characterized in that: The specific data acquisition process of S1 includes: Thermal sensors collect temperature data from the contact area. The sensor is mounted around the metal base of the moving and stationary contacts; The micro-vibration sensor is mounted on the metal base plate and contact fixing assembly of the distribution cabinet to capture subtle structural vibrations and obtain data. ; A multi-channel MOS gas sensor array was used to detect partial discharge products and obtain gas concentration vector data. ; A non-contact infrared array thermal imager is used, installed inside the cabinet door, facing the contacts and their conductive path area, to capture local hotspot images. ; The acoustic signature sensor is configured with a linear array of MEMS microphones to acquire short-term audio energy fluctuations caused by abnormal processes during the "closing-current flow" phase, thereby obtaining the original audio frame vector. .

3. The low-voltage standardized switchgear and its contact anomaly tracing method as described in claim 2, characterized in that: The specific process of S2 is as follows: For data and exist[ Within the interval, a sliding window averaging process was performed to obtain the mean-processed micro-vibration data. and the audio frame data after mean processing. ; For data Data alignment was performed using the nearest-value hold method to obtain aligned temperature data. and aligned gas concentration data ; hotspot images The data is flattened to obtain the flattened hotspot vector data. Based on this data dimension, a unified feature input is obtained by concatenating the data. ; The system will unify feature vectors The data is fed into a sparse autoencoder for dimensionality compression and representation modeling to obtain... Low-dimensional feature representation after time-point encoding ,and ; Total sampling time; Finally, the compressed core feature sequence of all consecutive frames is: .

4. The low-voltage standardized switchgear and its contact anomaly tracing method as described in claim 1, characterized in that: The neural abnormality detection model in S3 is specifically as follows: Graph Structure Definition and Node Mapping: The physical entities in the low-voltage switchgear are abstracted into a graph according to their structure and function. ; set of nodes in the graph , The total number of nodes equals the number of all modeled key locations in the system; each node Represents an independent physical functional unit, corresponding to a specific component in a low-voltage switchgear; edge set , Represents a node and There is a physical / electrical relationship between them; Feature embedding and node initialization: Based on the core feature sequence, it is assigned and mapped to a graph structure. The initial state vector of the corresponding node is used to realize the structure-aware data binding process; Each node in the diagram The features are initialized from the associated core feature sequences. Then, the core feature sequences of each node are fused into a vector of uniform dimension, and linear mapping and concatenation are performed to reduce the dimensionality, resulting in the node... Integrated features The final node features constitute the initial feature matrix H of the input layer of the graph neural network, which is used as the input layer of the model to complete the feature binding process. Edge weight modeling and adjacency matrix construction: The edge weights between nodes represent the strength of physical or informational coupling relationships, using a weighted adjacency matrix. This represents the connection relationships and propagation weights between nodes. Then, the node is calculated based on spatial distance, connection type, and energy coupling degree. With nodes Connection strength Finally, the connection strength between each pair of nodes is obtained, which together form a weighted adjacency matrix. .

5. The low-voltage standardized switchgear and its contact anomaly tracing method as described in claim 4, characterized in that: In S3, a multi-layer graph attention network (GAT) mechanism is used to model the interaction propagation and adaptive fusion between nodes. The initial feature matrix H of each node serves as the input feature representation of the graph neural network, and the adjacency matrix A provides the connection topology and propagation weights between nodes. Through the edge connections and weight information in the graph structure, information is guided to propagate from the abnormal source node to adjacent nodes, capturing the diffusion path of multimodal disturbance signals induced by abnormal contact resistance in the system. Specifically, this includes: The initial feature matrix H of the nodes in the input layer is input into the GAT layer, and the output is... To achieve local perturbation perception and neighborhood propagation, new features are generated after each node interacts with its neighbors. LeakyReLU is used as the activation function. Next The input is fed into the second-layer attention map convolution, using LeakyReLU as the activation function to achieve wider propagation and deeper feature extraction; the GAT structure is compressed into a two-layer attention propagation structure, with parameters denoted as... ; After propagation through an L-layer GAT model, the final state embedding vector of each node at time t is obtained. .

6. The low-voltage standardized switchgear and its contact anomaly tracing method as described in claim 5, characterized in that: The anomaly category identification module in S4 embeds the final state vector of each node at time t. Using a fully connected layer and a softmax function as input, fault prediction is performed to obtain the first... i The predicted class probability output for each node. The predicted probability distribution includes normal, poor contact, arc discharge, and mechanical loosening results; Within each detection cycle, record the actual detection data label of the current node. , and model output results Compare and calculate the absolute deviation vector. ,like The L1 norm exceeds the dynamic threshold When an anomaly is detected, an anomaly detection response is triggered, initiating a data backtracking mechanism for the anomaly period. The backtracking time window is set to [time value missing]. ,in, The time window size is defined; within this interval, voiceprint features, thermal infrared image features, and microseismic features are collected to construct a multimodal anomaly description vector.

7. The low-voltage standardized switchgear and its contact anomaly tracing method as described in claim 1, characterized in that: The abnormal state detection module in S5 takes the multimodal abnormal description vector as input and outputs the final... i Abnormal event classification results for each node It includes normal conditions, poor contact, arc discharge, and mechanical loosening; the specific structure of the abnormal state detection module is as follows: Input layer: The input consists of a multimodal description vector that integrates resistance, temperature, vibration, and acoustic signature, as well as the prediction results from the graph neural network; Feature normalization layer: Normalizes the sub-features of each modality in the multimodal description vector to maintain the consistency of feature scale; Anomaly Feature Embedding Layer: A linear layer is used to reduce the dimensionality of the multimodal description vector and map it to a unified representation space to obtain the embedded representation vector; Graph Neural Network Output Fusion Layer: Concatenates the output of the graph neural network model with the embedded representation vector; inputs it into the fusion classifier to calculate the final classification probability; Classifier: The final fault type, i.e., the classification result of the abnormal event, is obtained by using the extreme value function argmax. And output it.

8. The low-voltage standardized switchgear and its contact anomaly tracing method as described in claim 1, characterized in that: An integrated edge intelligent node is incorporated into each standardized cabinet, integrating the graph neural network anomaly detection model, anomaly category recognition module, and anomaly state detection module into a system model. This model is deployed in a lightweight form on the edge computing node of each low-voltage cabinet for local operation. The FedAvg federated learning algorithm is used to coordinate multiple low-voltage cabinet edge nodes for distributed modeling. The specific process is as follows: First, federated average model aggregation is performed; in each federated round r, the set of participating edge nodes is K, and the k-th node obtains the model parameters through local training. Then the central server will have parameters for all nodes. Perform weighted aggregation to obtain the updated global model parameters. ; Secondly, a homomorphic encryption communication mechanism is used; to prevent model parameters from being leaked during communication, the system introduces an additive homomorphic encryption function. It satisfies the additivity property with respect to the input variables; if each of the two nodes has an error term. and Then the summation of their error functions is equivalent to calculating the sum of their errors all at once. That is, the sum of the local error function values ​​of node i and node j is equivalent to the global error function value calculated by adding their error parameters together. Edge nodes locally... Encrypted uploads are performed, allowing the central node to complete aggregation calculations without decryption, thus ensuring parameter security.

9. The low-voltage standardized switchgear and its contact anomaly tracing method as described in claim 8, characterized in that: An early warning strategy is formed by clustering analysis of the anomaly description vectors of different cabinets. The specific process is as follows: After each round of federated training, the central node will collect the anomaly description vector sets uploaded by the edge nodes of each cabinet. : As input, cluster analysis is performed to uncover the spatial or temporal patterns of anomaly propagation; the K-means clustering algorithm is used, and the clustering results are used to construct a regional early warning model; the system outputs three levels of early warning (red, orange, and blue) based on the anomaly density and occurrence frequency of each cluster, driving the background scheduling response strategy. The number of clusters is set to K=3, corresponding to the three warning levels of red, orange, and blue, respectively. When the warning level is red, it indicates that the activity is abnormal and dense, and it is recommended to immediately dispatch on-site maintenance. When the warning level is orange, it indicates that the abnormality may spread, and planned inspections need to be arranged. When the warning level is blue, it indicates that the abnormality is slight or a peripheral disturbance, and the background will continue to monitor it.

10. The low-voltage standardized switchgear and its contact anomaly tracing method as described in claim 1, characterized in that: Based on nodes that output anomaly types in S5, anomaly event records are obtained, including node number, fault category, anomaly occurrence time, anomaly event description vector, multimodal fusion features, model version summary, and resistance detection value. These are then packaged into on-chain data structures and recorded in data blocks across all network nodes using distributed ledger technology. In this process, a data digest is generated using a hash function and uploaded to the blockchain to form a block structure. ,in, This hash value is used to build the chain-like dependency relationship between each block in the blockchain, ensuring that on-chain data is immutable and traceable throughout the entire chain. Represents a random number. This represents the Merkle root of all data packets in the current block; When the system detects an alarm event or receives a maintenance instruction request, the blockchain module automatically backtracks the historical chain data, extracts the abnormal event chain and verifies the consistency between the model version and the data, and calculates the credibility index. Should The structure is automatically submitted to the consortium blockchain network via a smart contract mechanism, and confirmed through consensus by multiple trusted nodes. Once consensus is reached, This will be formally written into the distributed ledger, forming a time-sequential chain of exception event records; when operations and maintenance personnel trace anomalies and analyze the reliability of the source tracing model, they can use... The structure quickly locates event information and provides a basis for model prediction.

Images