Intelligent plug-in box monitoring method and system based on UL authentication standard

Abstract

The application discloses a smart plug-in box monitoring method and system based on UL authentication standards, and relates to the technical field of plug-in box monitoring. The method comprises the following steps: deploying K cross-modal sensing arrays in the smart plug-in box in combination with the safety constraints of UL authentication standards; connecting the K cross-modal sensing arrays to K domain control nodes and constructing a dynamic weighted topology graph of the K domain control nodes; training a heterogeneous fault propagation model of the dynamic weighted topology graph based on a graph neural network to obtain a graph node state reasoning engine; uploading K multi-modal sensing flow data to the K domain control nodes to drive the graph node state reasoning engine to perform cross-domain fault correlation reasoning and output a fault correlation topology state; and performing high-response predictive maintenance decision on the smart plug-in box. The technical problems of lagging plug-in box fault hidden danger identification and untimely predictive maintenance response in the prior art are solved, early warning and active maintenance of plug-in box fault risks are realized, and thus the operation safety and operation and maintenance efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of plug-in box monitoring technology, specifically to an intelligent plug-in box monitoring method and system based on UL certification standards. Background Technology

[0002] As a crucial power distribution unit in low-voltage power distribution systems, plug-in junction boxes are widely used in industrial plants, data centers, commercial complexes, and rail transit systems to facilitate branch power intake and load connection for busbar systems. With the diversification and increasing density of electrical loads, the internal electrical connection structure of plug-in junction boxes is becoming increasingly complex. During long-term operation, they are susceptible to potential faults due to factors such as overload, electric arcing, poor contact, insulation aging, and changes in ambient temperature and humidity. To meet electrical safety requirements, the industry typically regulates the structural design and safety performance of plug-in junction boxes according to UL certification standards. However, existing technologies primarily focus on structural safety design and factory testing, relying mainly on single-point current, voltage, or temperature sensors for status monitoring during operation, lacking the ability to collaboratively analyze multiple fault-related domains. Under complex electrical topology coupling relationships, local anomalies often propagate between different functional areas through electromagnetic, thermal effects, or load fluctuations. Traditional monitoring methods struggle to identify cross-domain fault correlation characteristics in a timely manner, leading to delayed fault identification, untimely predictive maintenance responses, and a high risk of localized overheating, insulation breakdown, or even electrical fires. Summary of the Invention

[0003] This application provides a monitoring method and system for intelligent plug-in boxes based on UL certification standards, which solves the technical problems of delayed identification of potential plug-in box faults and untimely response to predictive maintenance in the prior art.

[0004] A first aspect of this application provides a monitoring method for a smart junction box based on UL certification standards, the method comprising: Combining the safety constraints of UL certification standards, K cross-modal sensor arrays are deployed in K pre-divided fault association domains of the intelligent plug-in box. Edge computing nodes connect the K cross-modal sensor arrays to K domain control nodes, and a dynamic weighted topology graph of the K domain control nodes is constructed based on the electrical topology coupling relationships of the K fault association domains. K sets of spatiotemporal feature vectors within the domains are extracted from historical multi-dimensional operational fault records of the K fault association domains as training data. A heterogeneous fault propagation model for the dynamic weighted topology graph is trained based on a graph neural network to obtain a graph node state inference engine. The K cross-modal sensor arrays monitor and upload K multi-modal sensor stream data of the K fault association domains to the K domain control nodes, driving the graph node state inference engine to perform cross-domain fault association inference and output the fault association topology state. High-response predictive maintenance decisions are made for the intelligent plug-in box based on the fault association topology state.

[0005] A second aspect of this application provides an intelligent plug-in box monitoring system based on UL certification standards, the system comprising: Sensor Array Deployment Module: Combining the safety constraints of UL certification standards, K cross-modal sensor arrays are deployed in K pre-divided fault association domains of the smart plug-in box; Topology Construction Module: Edge computing nodes are used to connect the K cross-modal sensor arrays to K domain control nodes, and a dynamic weighted topology graph of the K domain control nodes is constructed based on the electrical topology coupling relationship of the K fault association domains; Model Training Module: K sets of spatiotemporal feature vectors within the domains are extracted from historical multi-dimensional operational fault records of the K fault association domains as training data, and a heterogeneous fault propagation model of the dynamic weighted topology graph is trained based on graph neural networks to obtain a graph node state inference engine; Fault Inference Module: The K cross-modal sensor arrays monitor and upload K multi-modal sensor stream data of the K fault association domains to the K domain control nodes, driving the graph node state inference engine to perform cross-domain fault association inference and output the fault association topology state; Maintenance Decision Module: Based on the fault association topology state, a high-response predictive maintenance decision is made for the smart plug-in box.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, based on the safety constraints of UL certification standards, K cross-modal sensor arrays are deployed in K pre-divided fault association domains within the intelligent plug-in box. Next, edge computing nodes connect the K cross-modal sensor arrays to K domain control nodes, and a dynamically weighted topology graph is constructed for the K domain control nodes based on the electrical topology coupling relationships of the K fault association domains. Then, K sets of spatiotemporal feature vectors within the domains are extracted from historical multi-dimensional operational fault records as training data for the K fault association domains. A heterogeneous fault propagation model for the dynamically weighted topology graph is trained using a graph neural network, resulting in a graph node state inference engine. Finally, the K cross-modal sensor arrays monitor and upload K multi-modal sensor stream data from the K fault association domains to the K domain control nodes, driving the graph node state inference engine to perform cross-domain fault association inference and output the fault association topology state. Based on the fault association topology state, highly responsive predictive maintenance decisions are made for the intelligent plug-in box. This technology solves the technical problems of delayed identification of potential faults in plug-in boxes and untimely response to predictive maintenance in existing technologies. It enables early warning and proactive maintenance of plug-in box fault risks, thereby improving operational safety and maintenance efficiency. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic diagram of the monitoring method for a smart plug-in box based on UL certification standards provided in this application embodiment; Figure 2 A schematic diagram of the intelligent plug-in box monitoring system based on UL certification standards provided in this application embodiment.

[0009] Figure labeling: 11 Sensor array deployment module, 12 Topology construction module, 13 Model training module, 14 Fault reasoning module, 15 Maintenance decision module. Detailed Implementation

[0010] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0011] Example 1, as Figure 1 As shown, this application provides a monitoring method for intelligent plug-in boxes based on UL certification standards, wherein the method includes: Based on the safety constraints of UL certification standards, K cross-modal sensor arrays are deployed in K pre-divided fault association domains of the smart junction box.

[0012] Furthermore, in accordance with the safety constraints of UL certification standards, K cross-modal sensor arrays are deployed in the K pre-divided fault association domains of the smart junction box, including: By aggregating historical multi-dimensional operational fault records, the smart plug-in box space is divided into K fault association domains; based on the K sets of aggregated fault risk characteristics of the K fault association domains, and combined with the multi-dimensional safety threshold constraints of the UL certification standard, the K cross-modal sensor arrays are deployed in the K fault association domains.

[0013] Preferably, historical multi-dimensional operational fault records of the intelligent plug-in box under different operating conditions are summarized. These records include at least current, voltage, power factor, temperature rise data, contact resistance change data, insulation resistance data, partial discharge data, and ambient temperature and humidity data. The data undergoes time alignment, anomaly cleaning, and standardization. Based on the fault location, fault type, and propagation path, spatial clustering analysis and risk contribution assessment are performed on the historical fault samples. The fault occurrence frequency, average fault severity level, and cross-domain propagation impact factor of each spatial unit are calculated. According to a preset risk classification rule, the internal space of the intelligent plug-in box is divided into K fault association domains, ensuring relative consistency of risk characteristics within each domain and significant differences between domains. Subsequently, K sets of aggregated fault risk characteristics are extracted for each of the K fault association domains. Features include electrical overload risk index, heat accumulation risk index, insulation degradation risk index, and fault propagation sensitivity coefficient. Based on this, corresponding clauses of the UL certification standard are retrieved to extract multi-dimensional safety threshold constraints such as upper limit of temperature rise, electrical clearance and creepage distance requirements, rated current carrying capacity margin, and insulation withstand voltage level. These multi-dimensional safety threshold constraints are matched and analyzed with the aggregated fault risk characteristics of each fault-related domain to determine the key physical quantities and safety margin ranges to be monitored in each fault-related domain. According to the matching results, corresponding cross-modal sensor arrays are deployed in each fault-related domain. These arrays include electrical parameter acquisition units, thermal state monitoring units, and insulation state monitoring units. The number of sensors, measurement range, and sampling frequency are determined based on the risk weight of each domain, thereby achieving precise multi-modal sensing deployment for risk zones while meeting the safety constraints of the UL certification standard.

[0014] Edge computing nodes are used to connect the K cross-modal sensor arrays to K domain control nodes, and a dynamic weighted topology graph of the K domain control nodes is constructed based on the electrical topology coupling relationship of the K fault-related domains.

[0015] First, a corresponding edge computing node is deployed within each fault-related domain. This edge computing node establishes a data communication link with the cross-modal sensor array of its domain via an industrial bus or Ethernet interface. It performs local caching, time synchronization, and data preprocessing on the collected electrical, thermal, and insulation parameters, including filtering and noise reduction, outlier removal, preliminary feature extraction, and UL safety threshold compliance verification. Then, each edge computing node is logically mapped to a domain control node and assigned a unique node identifier, forming an initial topology with K domain control nodes as vertices. Based on this, the busbar topology configuration library and the electrical schematic diagram of the plug-in box are retrieved to extract the electrical connection paths, current-carrying directions, loop coupling relationships, and physical adjacency relationships between the K fault-related domains, constructing an inter-domain coupling matrix. Furthermore, by combining the propagation paths and frequencies of different faults between domains in historical multi-dimensional operational fault records, the electrical coupling strength coefficient, thermal conduction correlation coefficient, and historical fault propagation frequency coefficient between each domain control node are calculated. The above multi-dimensional correlation indicators are normalized and fused according to preset weights to form a three-dimensional correlation weight vector. Based on the spatial critical distance and electrical connectivity judgment conditions, effective connection edges are selected, and the three-dimensional correlation weight vector is mapped to edge weight parameters to construct a dynamic weighted topology graph containing a set of nodes, a set of edges, and a set of edge weights. During operation, edge computing nodes update the inter-domain correlation indicators based on real-time multimodal sensor flow data and dynamically adjust the edge weight parameters, thereby realizing the construction of a dynamic weighted topology graph that can reflect the changes in electrical topology coupling strength and fault propagation trends.

[0016] Furthermore, based on the electrical topology coupling relationships of the K fault-related domains, a dynamic weighted topology graph of the K domain control nodes is constructed, including: The interactive busbar topology configuration library obtains the K electrical interconnection parameters and K physical spatial relationships of the K fault association domains in the intelligent plug-in box; extracts K time-series fault records of the K fault association domains from the historical multi-dimensional operational fault records; calculates multi-dimensional association indicators among the K domain control nodes based on the K electrical interconnection parameters, K physical spatial relationships, and K time-series fault records, and constructs a third-order dynamic weight matrix, wherein the multi-dimensional association indicators cover electrical coupling strength, thermal conduction association degree, and historical fault conduction frequency; and constructs the dynamic weighted topology graph of the K domain control nodes based on the third-order dynamic weight matrix, using spatial critical distance and topology connectivity threshold as topology connectivity determination conditions.

[0017] First, data interaction is performed with the busbar topology configuration library through the interface program to read the connection structure information of each branch circuit inside the intelligent plug-in box. K electrical interconnection parameters and K physical spatial relationships corresponding to K fault association domains are extracted. The electrical interconnection parameters include circuit number, current carrying direction, rated current, equivalent impedance, and common node connection relationship. The physical spatial relationships include inter-domain spatial distance, installation height difference, and adjacent arrangement relationship. Subsequently, K time-series fault records corresponding to the K fault association domains are extracted from historical multi-dimensional operational fault records in chronological order. These time-series fault records include fault occurrence timestamp, fault type code, fault duration, and propagation path marker. Based on this, a multi-dimensional inter-domain correlation index calculation model is constructed, based on the electrical interconnection parameters... The electrical coupling strength between domain control nodes is calculated, the inter-domain thermal conduction correlation is calculated based on the physical spatial relationship, and the historical fault propagation frequency between domains is statistically analyzed based on the time-series fault records. These indicators are then normalized and weighted to form a third-order dynamic weight matrix, where the first order represents the electrical coupling weight, the second order represents the thermal conduction weight, and the third order represents the historical fault propagation weight. Finally, using a preset spatial critical distance and topological connectivity threshold as connectivity criteria, redundant connections that do not meet the connectivity conditions are pruned, and nodes that meet the conditions are assigned corresponding edge weights based on the third-order dynamic weight matrix. This constructs a dynamic weighted topology graph of K domain control nodes, including a set of nodes, a set of edges, and dynamic weight parameters, thus providing a structured graph model foundation for subsequent cross-domain fault propagation reasoning.

[0018] Furthermore, using spatial critical distance and topological connectivity threshold as topological connectivity determination conditions, the dynamically weighted topological graph of the K domain control nodes is constructed based on the third-order dynamic weight matrix, including: Redundant connection topology pruning of the K domain control nodes is performed with the spatial critical distance as a constraint to obtain the basic topology skeleton; after verifying the electrical connectivity of the basic topology skeleton according to the topology connectivity threshold, the topology edge weights are dynamically allocated using the third-order dynamic weight matrix, and the dynamically weighted topology graph is output.

[0019] Preferably, the three-dimensional spatial coordinates of K domain control nodes within the intelligent plug-in box are obtained. The spatial distance matrix between nodes is calculated based on the Euclidean distance between any two domain control nodes, and a spatial critical distance threshold is set. When the spatial distance between nodes exceeds the critical distance threshold, it is determined that the node pair has no possibility of direct physical coupling. The corresponding connection edge is marked as a redundant edge, and topology pruning is performed to delete connection relationships that lack a basis for spatial coupling, resulting in a basic topology skeleton that retains only potentially physically related edges. Subsequently, based on the basic topology skeleton, the bus electrical circuit connection relationship and conductive path information are read, and a topology connectivity threshold is set, including a minimum electrical coupling strength threshold or a minimum historical fault conduction frequency threshold. Each candidate connection edge is then subjected to electrical... For connectivity verification, if the corresponding electrical coupling strength or historical conduction frequency is lower than the topology connectivity threshold, the connecting edge is further removed to ensure that the topology structure meets the actual electrical connectivity constraints. After completing the connectivity screening, the electrical coupling weight, heat conduction weight, and historical fault conduction weight parameters in the third-order dynamic weight matrix are called for the retained connecting edges. The weighted calculation is performed according to the preset fusion rules to generate a comprehensive weight value for each connecting edge, and this comprehensive weight value is assigned as the edge weight to the topology structure. During system operation, the parameters of the third-order dynamic weight matrix are dynamically corrected according to the real-time updated multimodal sensing data, and the edge weight values ​​are updated synchronously, thereby outputting a dynamically weighted topology map that reflects changes in spatial constraints, electrical connectivity relationships, and fault propagation intensity.

[0020] Based on the K fault association domains, K sets of spatiotemporal feature vectors within the domains are extracted from historical multi-dimensional operational fault records as training data. The heterogeneous fault propagation model of the dynamically weighted topology graph is trained based on the graph neural network to obtain the graph node state inference engine.

[0021] Furthermore, based on the K fault-related domains, K sets of spatiotemporal feature vectors within the domains are extracted from historical multi-dimensional operational fault records as training data. A heterogeneous fault propagation model of the dynamically weighted topology graph is trained using a graph neural network to obtain a graph node state inference engine, including: The historical multi-dimensional operational fault records are decomposed using UL standard violation event labeling to obtain multiple fault event fragments. Each fault event fragment includes a violation event type code, a fault type code, and multimodal temporal sensing data. Spatiotemporal feature extraction and UL diagnosis are performed on the multimodal temporal sensing data in the multiple fault event fragments to obtain multiple basic feature vectors and multiple historical fault confidence vectors. The multiple basic feature vectors, multiple historical fault confidence vectors, and violation event type codes and fault type codes in the multiple fault event fragments are concatenated to obtain multiple labeled feature vectors. Using the K fault association domains as spatial units, the spatial aggregation of the multiple labeled feature vectors is performed according to the position coordinates of the multiple fault event fragments to obtain the K sets of domain-specific spatiotemporal feature vectors. The K sets of domain-specific spatiotemporal feature vectors are used as training data to train the heterogeneous fault propagation model of the dynamically weighted topology graph based on a graph neural network to obtain the graph node state inference engine.

[0022] First, historical multi-dimensional operational fault records are decomposed using UL standard violation event labeling. Continuous operational data is segmented according to violation trigger time windows, identifying and labeling various violations such as excessive temperature rise, current overload, insulation breakdown, or abnormal creepage distance, forming multiple fault event segments. Each fault event segment contains a violation event type code, a fault type code, and multimodal time-series sensor data within the corresponding time interval. Subsequently, the multimodal time-series sensor data undergoes time alignment, filtering and denoising, and sliding window segmentation to extract statistical features, trend features, and abrupt change features, constructing multiple basic feature vectors. These are then combined with the UL safety threshold model to perform compliance diagnostic calculations on each time segment, generating multiple historical fault confidence vectors characterizing the severity and evolution trend of violations. Based on this, the basic feature vectors and historical fault... The obstacle confidence vector is concatenated and fused with the corresponding violation event type code and fault type code to form multiple labeled feature vectors with supervised label information. Using the K fault association domains as spatial units, based on the physical location coordinates and domain division results corresponding to each fault event fragment, the multiple labeled feature vectors are spatially aggregated and weighted to obtain K sets of intra-domain spatiotemporal feature vectors reflecting the temporal evolution characteristics of each domain. Finally, the K sets of intra-domain spatiotemporal feature vectors are mapped to the node feature space of the dynamically weighted topology graph to construct a graph neural network training sample set. The edge weights of the dynamically weighted topology graph are used as information propagation coefficients to perform multiple rounds of graph convolution and parameter iterative updates until the loss function converges. The trained heterogeneous fault propagation model is output and encapsulated as the graph node state inference engine for real-time cross-domain fault association inference.

[0023] Furthermore, using the K sets of spatiotemporal feature vectors within the domain as training data, a heterogeneous fault propagation model of the dynamically weighted topology graph is trained based on a graph neural network to obtain the graph node state inference engine, including: A three-channel heterogeneous graph neural network model is constructed based on the dynamically weighted topology graph. In this model, the feature input layer of each node within a domain includes parallel electrical propagation channels, thermal propagation channels, and insulation propagation channels. The K sets of spatiotemporal feature vectors within the domain are decomposed into K sets of sample feature subgroups based on the physical propagation mechanism. Each sample feature subgroup includes an electrical correlation feature group, a thermal correlation feature group, and a fault conduction feature group. During the cross-modal fusion training of the electrical propagation channels, thermal propagation channels, and insulation propagation channels in the three-channel heterogeneous graph neural network model using the K sets of sample feature subgroups to construct a training sample set, an UL safety margin constraint is introduced to suppress overfitting, resulting in a trained heterogeneous graph neural network model. Knowledge distillation is performed on the heterogeneous graph neural network model to generate the graph node state inference engine.

[0024] First, a three-channel heterogeneous graph neural network model is constructed based on the node set and edge weight structure of a dynamically weighted topological graph. Each domain-controlled node is used as the graph node input unit. Parallel electrical propagation channels, thermal propagation channels, and insulation propagation channels are set in the node feature input layer. The electrical propagation channel is used to model the electrical coupling propagation characteristics caused by changes in current, voltage, and impedance. The thermal propagation channel is used to model the thermal conduction characteristics caused by temperature rise gradient and thermal diffusion trend. The insulation propagation channel is used to model the fault conduction characteristics caused by insulation degradation and leakage characteristics. Subsequently, based on the physical propagation mechanism, the K groups of spatiotemporal feature vectors within the domain are decoupled. The dimensions are decomposed according to electrical correlation characteristics, thermal correlation characteristics, and fault conduction characteristics to obtain K groups of sample feature subgroups. Each group of sample feature subgroups is input to the corresponding propagation channel. The model forms parallel feature representations. During the training phase, a training sample set is constructed using the K sets of sample feature subgroups. Cross-node information propagation and cross-modal fusion are achieved through graph convolution operations combined with the edge weight parameters in the dynamically weighted topology graph. In the backpropagation update process, a UL safety margin constraint term is introduced, and the deviation between the node prediction result and the UL safety threshold is added to the loss function as a regularization factor to suppress overfitting and enhance the model's sensitivity to safety boundaries, resulting in a trained heterogeneous graph neural network model. Finally, knowledge distillation is performed on the trained heterogeneous graph neural network model. By constructing a lightweight student model to learn the node representation distribution and fault probability output distribution of the teacher model, model compression and inference acceleration are achieved, generating a graph node state inference engine that can be deployed on domain control nodes or edge computing nodes.

[0025] The K cross-modal sensor arrays monitor and upload K multimodal sensor stream data of the K fault association domains to the K domain control nodes, drive the graph node state inference engine to perform cross-domain fault association inference, and output the fault association topology state.

[0026] Furthermore, the K cross-modal sensor arrays monitor and upload K multimodal sensor stream data of the K fault association domains to the K domain control nodes, driving the graph node state inference engine to perform cross-domain fault association inference and outputting the fault association topology state, including: The K cross-modal sensor arrays monitor the K multimodal sensor stream data of the K fault association domains in real time; the K multimodal sensor stream data are uploaded to the K domain control nodes through K virtual channels for local UL safety threshold compliance diagnosis and spatiotemporal feature extraction, generating K local fault confidence vectors and K real-time feature vectors; the K local fault confidence vectors and K real-time feature vectors are input into the graph node state inference engine for cross-domain fault association inference, and the fault association topology state is output.

[0027] First, the K cross-modal sensor arrays are deployed in corresponding K fault-related domains, respectively. Electrical parameter data, thermal state data, and insulation state data are collected in real time according to a preset sampling frequency, forming K multimodal sensor streams. Timestamp synchronization and cache management are performed on each data stream. Then, a corresponding virtual data transmission channel is established for each fault-related domain. The K multimodal sensor streams are uploaded to the corresponding K domain control nodes through these K virtual channels. At the domain control nodes, local UL safety threshold compliance diagnostics are performed, comparing parameters such as current, temperature rise, and insulation resistance with UL standard limits item by item to generate K local fault confidence vectors. The uploaded multimodal sensor stream data is segmented by sliding window, and statistical features, trend features, and normalized to form K real-time feature vectors consistent with the basic feature structure in the graph neural network modeling stage. Finally, the K local fault confidence vectors and K real-time feature vectors are loaded as node input features into the corresponding nodes of the dynamic weighted topology graph. The graph node state inference engine is called to perform graph structure forward propagation calculation. Cross-domain fault association inference is completed through cross-node information transmission and attention weighted fusion mechanism. The fault association topology state is output, including node-level fault probability distribution, cross-domain fault influence intensity matrix, and UL violation event code set.

[0028] Furthermore, the K local fault confidence vectors and K real-time feature vectors are input into the graph node state inference engine for cross-domain fault association inference, and the fault association topology state is output, including: The K real-time feature vectors are decomposed based on the physical propagation mechanism to obtain K sets of real-time feature subgroups. A three-channel activation weight matrix is ​​generated based on the K local fault confidence vectors. After dynamically configuring the three-channel inference intensity of the K domain control nodes in the graph node state inference engine, the K sets of real-time feature subgroups are input for three-channel parallel propagation inference. The three-channel outputs are then fused through an attention weighting mechanism to obtain K node fused feature vectors. The graph node state inference engine iterates the fault propagation of the K node fused feature vectors using the dynamically weighted topology graph as the transmission structure, and outputs the fault-associated topology state.

[0029] First, based on the electrical coupling propagation mechanism, thermal diffusion propagation mechanism, and insulation degradation conduction mechanism, the K real-time feature vectors are decoupled and decomposed according to the physical attribute dimension, and electrical correlation feature subgroups, thermal correlation feature subgroups, and insulation fault conduction feature subgroups are extracted respectively to form K groups of real-time feature subgroups. Then, the current risk intensity level of each domain is calculated based on the K local fault confidence vectors, and a corresponding three-channel activation weight matrix is ​​generated through a normalized mapping function. The three channels correspond to the electrical propagation channel, thermal propagation channel, and insulation propagation channel, respectively. The activation weights are used to dynamically adjust the inference intensity of each domain control node in the graph node state inference engine on different propagation channels. After completing the inference intensity configuration, the K groups of real-time feature subgroups are... Each feature subgroup is input into its corresponding propagation channel to perform a three-channel parallel graph convolutional propagation calculation. Cross-domain information transmission is achieved through the edge weights of the dynamically weighted topology graph, and multiple rounds of adjacent node feature aggregation are completed within each channel. Subsequently, an attention weighting mechanism is used to fuse the output results of the three channels, and K node fusion feature vectors are generated based on the importance coefficients of each channel output. Finally, in the graph node state inference engine, the dynamically weighted topology graph is used as the transmission structure to perform fault propagation iterative calculation on the K node fusion feature vectors until the preset propagation rounds or state convergence conditions are reached. The output includes the fault-related topology state, which includes the node-level fault probability distribution, the cross-domain fault influence intensity matrix, and the UL violation event encoding set.

[0030] Furthermore, the fault-related topology state includes a node-level fault probability distribution, a cross-domain fault impact intensity matrix, and a UL violation event code set.

[0031] Specifically, based on the K node fusion feature vectors output by the graph node state inference engine, the probability values ​​of each domain control node under a preset fault category set are calculated through a fully connected mapping layer and a Softmax normalization function, forming a node-level fault probability distribution to characterize the probability of different types of faults occurring in each fault-related domain at the current moment. Based on the information transmission strength between nodes during graph propagation and the edge weight parameters in the dynamically weighted topology graph, combined with the change amplitude of the node fusion feature vectors, the influence coefficient between node pairs is calculated, constructing a K×K dimension cross-domain fault influence strength matrix, where the matrix elements represent the influence strength of the source domain fault state on the target domain state evolution. Finally, fault categories exceeding a preset UL safety threshold in the node-level fault probability distribution are judged as violations and mapped to the corresponding UL standard clause codes, summarizing to form a UL violation event code set, used to identify the UL safety clause violation types that may be triggered under the current system state, thus constituting a complete fault-related topology state output result.

[0032] The intelligent plug-in box makes highly responsive predictive maintenance decisions based on the fault-related topology status.

[0033] First, the node-level fault probability distribution, cross-domain fault impact intensity matrix, and UL violation event code set in the fault-related topology are read. The fault probability of each domain control node is compared with a preset risk classification threshold, and classified into normal, warning, and high-risk levels. Second, the key propagation source nodes and affected nodes are identified by combining the cross-domain fault impact intensity matrix, and the fault propagation trend index and potential propagation time window are calculated to determine the priority order of handling. Subsequently, a preset maintenance strategy library is matched according to the UL violation event code set. The maintenance strategy library includes load current limiting strategy, loop isolation strategy, temperature rise mitigation strategy, and insulation detection strategy. The system implements verification strategies and on-site maintenance work order generation rules. Strategies are selected and optimized based on risk level and diffusion trend index to generate predictive maintenance decision schemes for each fault-related domain. After generating the maintenance decision scheme, control commands are sent to the corresponding domain control nodes via the edge control interface. These commands include dynamically adjusting load distribution, triggering alarm prompts, executing automatic circuit breaker protection, or scheduling manual maintenance. Simultaneously, real-time feedback on the operational status after execution is collected, updating the fault probability distribution and impact intensity matrix. This enables closed-loop evaluation of maintenance effectiveness and adaptive strategy correction, thereby completing a highly responsive predictive maintenance decision-making process oriented towards UL safety constraints.

[0034] In summary, the embodiments of this application have at least the following technical effects: First, based on the safety constraints of UL certification standards, K cross-modal sensor arrays are deployed in K pre-divided fault association domains within the intelligent plug-in box. Next, edge computing nodes connect the K cross-modal sensor arrays to K domain control nodes, and a dynamically weighted topology graph is constructed for the K domain control nodes based on the electrical topology coupling relationships of the K fault association domains. Then, K sets of spatiotemporal feature vectors within the domains are extracted from historical multi-dimensional operational fault records as training data for the K fault association domains. A heterogeneous fault propagation model for the dynamically weighted topology graph is trained using a graph neural network, resulting in a graph node state inference engine. Finally, the K cross-modal sensor arrays monitor and upload K multi-modal sensor stream data from the K fault association domains to the K domain control nodes, driving the graph node state inference engine to perform cross-domain fault association inference and output the fault association topology state. Based on the fault association topology state, highly responsive predictive maintenance decisions are made for the intelligent plug-in box. This technology solves the technical problems of delayed identification of potential faults in plug-in boxes and untimely response to predictive maintenance in existing technologies. It enables early warning and proactive maintenance of plug-in box fault risks, thereby improving operational safety and maintenance efficiency.

[0035] Example 2 is based on the same inventive concept as the UL-certified smart plug-in box monitoring method in the previous examples, such as... Figure 2 As shown, this application provides an intelligent plug-in box monitoring system based on UL certification standards, wherein the system includes: Sensor array deployment module 11: Combining the safety constraints of UL certification standards, deploy K cross-modal sensor arrays in K pre-divided fault association domains of the smart plug-in box; Topology construction module 12: Connect the K cross-modal sensor arrays to K domain control nodes using edge computing nodes, and construct a dynamic weighted topology graph of the K domain control nodes based on the electrical topology coupling relationship of the K fault association domains; Model training module 13: Extract K sets of domain-specific spatiotemporal feature vectors from historical multi-dimensional operational fault records of the K fault association domains as training data, and train a heterogeneous fault propagation model of the dynamic weighted topology graph based on graph neural networks to obtain a graph node state inference engine; Fault inference module 14: Monitor and upload K multimodal sensor stream data of the K fault association domains to the K domain control nodes, drive the graph node state inference engine to perform cross-domain fault association inference, and output the fault association topology state; Maintenance decision module 15: Make high-response predictive maintenance decisions for the smart plug-in box based on the fault association topology state.

[0036] Furthermore, the model training module 13 is used to perform the following methods: The historical multi-dimensional operational fault records are decomposed using UL standard violation event labeling to obtain multiple fault event fragments. Each fault event fragment includes a violation event type code, a fault type code, and multimodal temporal sensing data. Spatiotemporal feature extraction and UL diagnosis are performed on the multimodal temporal sensing data in the multiple fault event fragments to obtain multiple basic feature vectors and multiple historical fault confidence vectors. The multiple basic feature vectors, multiple historical fault confidence vectors, and violation event type codes and fault type codes in the multiple fault event fragments are concatenated to obtain multiple labeled feature vectors. Using the K fault association domains as spatial units, the spatial aggregation of the multiple labeled feature vectors is performed according to the position coordinates of the multiple fault event fragments to obtain the K sets of domain-specific spatiotemporal feature vectors. The K sets of domain-specific spatiotemporal feature vectors are used as training data to train the heterogeneous fault propagation model of the dynamically weighted topology graph based on a graph neural network to obtain the graph node state inference engine.

[0037] Furthermore, the model training module 13 is used to perform the following methods: A three-channel heterogeneous graph neural network model is constructed based on the dynamically weighted topology graph. In this model, the feature input layer of each node within a domain includes parallel electrical propagation channels, thermal propagation channels, and insulation propagation channels. The K sets of spatiotemporal feature vectors within the domain are decomposed into K sets of sample feature subgroups based on the physical propagation mechanism. Each sample feature subgroup includes an electrical correlation feature group, a thermal correlation feature group, and a fault conduction feature group. During the cross-modal fusion training of the electrical propagation channels, thermal propagation channels, and insulation propagation channels in the three-channel heterogeneous graph neural network model using the K sets of sample feature subgroups to construct a training sample set, an UL safety margin constraint is introduced to suppress overfitting, resulting in a trained heterogeneous graph neural network model. Knowledge distillation is performed on the heterogeneous graph neural network model to generate the graph node state inference engine.

[0038] Furthermore, the fault reasoning module 14 is used to perform the following method: The K cross-modal sensor arrays monitor the K multimodal sensor stream data of the K fault association domains in real time; the K multimodal sensor stream data are uploaded to the K domain control nodes through K virtual channels for local UL safety threshold compliance diagnosis and spatiotemporal feature extraction, generating K local fault confidence vectors and K real-time feature vectors; the K local fault confidence vectors and K real-time feature vectors are input into the graph node state inference engine for cross-domain fault association inference, and the fault association topology state is output.

[0039] Furthermore, the fault reasoning module 14 is used to perform the following method: The K real-time feature vectors are decomposed based on the physical propagation mechanism to obtain K sets of real-time feature subgroups. A three-channel activation weight matrix is ​​generated based on the K local fault confidence vectors. After dynamically configuring the three-channel inference intensity of the K domain control nodes in the graph node state inference engine, the K sets of real-time feature subgroups are input for three-channel parallel propagation inference. The three-channel outputs are then fused through an attention weighting mechanism to obtain K node fused feature vectors. The graph node state inference engine iterates the fault propagation of the K node fused feature vectors using the dynamically weighted topology graph as the transmission structure, and outputs the fault-associated topology state.

[0040] Furthermore, the sensor array deployment module 11 is used to perform the following methods: By aggregating historical multi-dimensional operational fault records, the smart plug-in box space is divided into K fault association domains; based on the K sets of aggregated fault risk characteristics of the K fault association domains, and combined with the multi-dimensional safety threshold constraints of the UL certification standard, the K cross-modal sensor arrays are deployed in the K fault association domains.

[0041] Furthermore, the topology building module 12 is used to perform the following methods: The interactive busbar topology configuration library obtains the K electrical interconnection parameters and K physical spatial relationships of the K fault association domains in the intelligent plug-in box; extracts K time-series fault records of the K fault association domains from the historical multi-dimensional operational fault records; calculates multi-dimensional association indicators among the K domain control nodes based on the K electrical interconnection parameters, K physical spatial relationships, and K time-series fault records, and constructs a third-order dynamic weight matrix, wherein the multi-dimensional association indicators cover electrical coupling strength, thermal conduction association degree, and historical fault conduction frequency; and constructs the dynamic weighted topology graph of the K domain control nodes based on the third-order dynamic weight matrix, using spatial critical distance and topology connectivity threshold as topology connectivity determination conditions.

[0042] Furthermore, the topology building module 12 is used to perform the following methods: Redundant connection topology pruning of the K domain control nodes is performed with the spatial critical distance as a constraint to obtain the basic topology skeleton; after verifying the electrical connectivity of the basic topology skeleton according to the topology connectivity threshold, the topology edge weights are dynamically allocated using the third-order dynamic weight matrix, and the dynamically weighted topology graph is output.

[0043] Furthermore, the fault reasoning module 14 is used to perform the following method: The fault-related topology state includes node-level fault probability distribution, cross-domain fault impact intensity matrix, and UL violation event code set.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A monitoring method for intelligent plug-in boxes based on UL certification standards, characterized in that, The method includes: Combining the safety constraints of UL certification standards, K cross-modal sensor arrays are deployed in K pre-divided fault association domains of the smart junction box; Edge computing nodes are used to connect the K cross-modal sensor arrays to K domain control nodes, and a dynamic weighted topology graph of the K domain control nodes is constructed based on the electrical topology coupling relationship of the K fault-related domains. Based on the K fault association domains, K sets of spatiotemporal feature vectors within the domains are extracted from historical multi-dimensional operational fault records as training data. The heterogeneous fault propagation model of the dynamically weighted topology graph is trained based on the graph neural network to obtain the graph node state inference engine. The K cross-modal sensor arrays monitor and upload K multimodal sensor stream data of the K fault association domains to the K domain control nodes, drive the graph node state inference engine to perform cross-domain fault association inference, and output the fault association topology state; The intelligent plug-in box makes highly responsive predictive maintenance decisions based on the fault-related topology status.

2. The intelligent plug-in box monitoring method based on UL certification standards as described in claim 1, characterized in that, Based on the K fault association domains, K sets of spatiotemporal feature vectors within the domains are extracted from historical multi-dimensional operational fault records as training data. A heterogeneous fault propagation model of the dynamically weighted topology graph is trained using a graph neural network to obtain a graph node state inference engine. The method includes: The historical multi-dimensional operational fault records are decomposed using UL standard violation event marking to obtain multiple fault event fragments. Each fault event fragment includes a violation event type code, a fault type code, and multimodal time-series sensor data. Spatiotemporal feature extraction and UL diagnosis are performed on the multimodal time-series sensing data in the multiple fault event segments to obtain multiple basic feature vectors and multiple historical fault confidence vectors; By concatenating the multiple basic feature vectors, multiple historical fault confidence vectors, and the violation event type codes and fault type codes from multiple fault event segments, multiple labeled feature vectors are obtained. Using the K fault-related domains as spatial units, spatial aggregation of the multiple labeled feature vectors is performed based on the position coordinates of the multiple fault event fragments to obtain the K sets of spatiotemporal feature vectors within the domains; Using the K sets of spatiotemporal feature vectors within the domain as training data, the heterogeneous fault propagation model of the dynamically weighted topology graph is trained based on a graph neural network to obtain the graph node state inference engine.

3. The intelligent plug-in box monitoring method based on UL certification standards as described in claim 2, characterized in that, Using the K sets of spatiotemporal feature vectors within the domain as training data, a heterogeneous fault propagation model of the dynamically weighted topology graph is trained based on a graph neural network to obtain the graph node state inference engine. The method includes: A three-channel heterogeneous graph neural network model is constructed based on the dynamic weighted topology graph, wherein the feature input layer of each node in the three-channel heterogeneous graph neural network model includes parallel electrical propagation channels, thermal propagation channels and insulation propagation channels; Based on the physical propagation mechanism, the K groups of spatiotemporal feature vectors within the domain are decomposed into K groups of sample feature subgroups, wherein each sample feature subgroup includes an electrical correlation feature group, a thermal correlation feature group, and a fault conduction feature group; In the process of constructing a training sample set using the K groups of sample feature subgroups to perform cross-modal fusion training of the electrical propagation channel, thermal propagation channel and insulation propagation channel in the three-channel heterogeneous graph neural network model, UL safety margin constraints are introduced to suppress overfitting, resulting in a trained heterogeneous graph neural network model. Knowledge distillation is performed on the heterogeneous graph neural network model to generate the graph node state inference engine.

4. The intelligent plug-in box monitoring method based on UL certification standards as described in claim 1, characterized in that, The K cross-modal sensor arrays monitor and upload K multimodal sensor stream data of the K fault association domains to the K domain control nodes, drive the graph node state inference engine to perform cross-domain fault association inference, and output the fault association topology state. The method includes: The K cross-modal sensor arrays monitor the K multimodal sensor stream data of the K fault-related domains in real time; The K multimodal sensor stream data are uploaded to the K domain control nodes through K virtual channels for local UL safety threshold compliance diagnosis and spatiotemporal feature extraction, generating K local fault confidence vectors and K real-time feature vectors. The K local fault confidence vectors and K real-time feature vectors are input into the graph node state inference engine to perform cross-domain fault association inference and output the fault association topology state.

5. The intelligent plug-in box monitoring method based on UL certification standards as described in claim 4, characterized in that, The method involves inputting the K local fault confidence vectors and K real-time feature vectors into the graph node state inference engine to perform cross-domain fault association inference and outputting the fault association topology state. Based on the physical propagation mechanism, the K real-time feature vectors are decomposed to obtain K groups of real-time feature subgroups; Based on the K local fault confidence vectors, a three-channel activation weight matrix is ​​generated. After dynamically configuring the three-channel inference intensity of the K domain control nodes in the graph node state inference engine, the K sets of real-time feature subgroups are input for three-channel parallel propagation inference. The three-channel outputs are then fused through an attention weighting mechanism to obtain K node fused feature vectors. The graph node state inference engine uses the dynamically weighted topology graph as the transmission structure to perform fault propagation iteration of the fused feature vectors of the K nodes, and outputs the fault-associated topology state.

6. The intelligent plug-in box monitoring method based on UL certification standard as described in claim 1, characterized in that, Combining the safety constraints of UL certification standards, K cross-modal sensor arrays are deployed in K pre-divided fault association domains of the smart junction box. The method includes: By aggregating historical multi-dimensional operational fault records, the smart plug-in box space is divided into K fault-related domains; Based on the K sets of aggregated fault risk characteristics of the K fault association domains, and combined with the multidimensional safety threshold constraints of the UL certification standard, the K cross-modal sensor arrays are deployed in the K fault association domains.

7. The intelligent plug-in box monitoring method based on UL certification standard as described in claim 1, characterized in that, Based on the electrical topology coupling relationships of the K fault-related domains, a dynamic weighted topology graph of the K domain control nodes is constructed, the method comprising: The interactive busbar topology configuration library obtains the K electrical interconnection parameters and K physical spatial relationships of the K fault association domains in the intelligent plug-in box; Extract K time-series fault records from the K fault association domains from the historical multi-dimensional operational fault records; Based on the K electrical interconnection parameters, K physical spatial relationships and K time-series fault records, multi-dimensional correlation indicators among the K domain control nodes are calculated, and a third-order dynamic weight matrix is ​​constructed. The multi-dimensional correlation indicators cover electrical coupling strength, thermal conduction correlation degree and historical fault conduction frequency. Using spatial critical distance and topological connectivity threshold as topological connectivity determination conditions, a dynamic weighted topological graph of the K domain control nodes is constructed based on the third-order dynamic weight matrix.

8. The intelligent plug-in box monitoring method based on UL certification standard as described in claim 7, characterized in that, Using spatial critical distance and topological connectivity threshold as topological connectivity determination conditions, and constructing the dynamically weighted topological graph of the K domain control nodes based on the third-order dynamic weight matrix, the method includes: Redundant connection topology pruning of the K domain control nodes is performed with the spatial critical distance as a constraint to obtain the basic topology skeleton; After verifying the electrical connectivity of the basic topology skeleton based on the topology connectivity threshold, the third-order dynamic weight matrix is ​​applied to dynamically allocate the topology edge weights, and the dynamically weighted topology graph is output.

9. The intelligent plug-in box monitoring method based on UL certification standard as described in claim 5, characterized in that, The fault-related topology state includes node-level fault probability distribution, cross-domain fault impact intensity matrix, and UL violation event code set.

10. A smart plug-in box monitoring system based on UL certification standards, characterized in that, For implementing the intelligent plug-in box monitoring method based on UL certification standards according to any one of claims 1-9, the system comprises: Sensor array deployment module: Combining the safety constraints of UL certification standards, deploy K cross-modal sensor arrays in K pre-divided fault association domains of the smart plug-in box; Topology construction module: The K cross-modal sensor arrays are connected to the K domain control nodes using edge computing nodes, and a dynamic weighted topology graph of the K domain control nodes is constructed based on the electrical topology coupling relationship of the K fault-related domains. Model training module: Based on the K fault association domains, extract K sets of spatiotemporal feature vectors within the domains from historical multi-dimensional operational fault records as training data, and train the heterogeneous fault propagation model of the dynamic weighted topology graph based on graph neural network to obtain the graph node state inference engine; Fault reasoning module: The K cross-modal sensor arrays monitor and upload K multimodal sensor stream data of the K fault association domains to the K domain control nodes, drive the graph node state reasoning engine to perform cross-domain fault association reasoning, and output the fault association topology state; Maintenance decision module: Makes highly responsive predictive maintenance decisions for the smart plug-in box based on the fault-related topology status.

Images