Intelligent fault diagnosis method and system for power distribution terminal equipment based on Internet of Things

Abstract

The invention discloses a power distribution terminal equipment fault intelligent diagnosis method and system based on the Internet of Things. The method comprises the following steps: S1, collecting multiple items of operation data of a power distribution terminal to construct a time sequence sample; s2, extracting power disturbance characteristics based on a sliding window, generating a behavior coupling matrix and a topological connection matrix, and calculating node redundancy; s3, fusing the two types of relationships to construct a dynamic graph; s4, inputting a graph diffusion network, fusing a structure and a behavior diffusion result, and generating a node state feature vector; s5, inputting a multi-label classification model to identify the fault type and probability; s6, generating a response instruction in combination with the fault type and the redundancy; s7, response is executed, the relation matrix and the classification model are updated, and the diagnosis process is optimized in a closed-loop mode. According to the invention, accurate fault identification and quick response of the power distribution terminal are realized, and the intelligent operation and maintenance level of a power supply system is improved.

Description

Technical Field

[0001] The present invention relates to the field of power distribution automation technology, and in particular to an intelligent diagnosis method and system for power distribution terminal equipment faults based on the Internet of Things. Background Art

[0002] As the intelligence level of distribution networks increases, distribution terminal equipment plays an increasingly critical role in the system. These terminals are primarily used to collect power parameters, detect operating status, and execute control commands, serving as a vital bridge between substations and higher-level systems. In recent years, distribution terminals have generally acquired the ability to collect multi-dimensional data such as voltage, current, and power. They connect to IoT platforms via wireless or wired communication, enabling basic information reporting and remote management. However, in actual operation, relying on the terminals themselves or single-point analysis methods to identify faults is difficult to meet the current system's requirements for diagnostic accuracy and response efficiency.

[0003] Traditional methods often rely on manually set thresholds, such as determining whether voltage is excessively high or power is fluctuating abnormally. These methods are usable in some stable scenarios, but are prone to false positives or missed positives in distribution systems with widely distributed terminals and frequently changing statuses. Furthermore, most existing methods only analyze data from a single terminal and lack an understanding of the relationships between multiple terminals. This makes it difficult to effectively assess the impact relationships between nodes, especially in ring networks or with backup paths.

[0004] Another common problem is the disconnect between diagnosis and control. Existing systems typically only issue alarms when an anomaly is detected, requiring manual confirmation and operational decision-making for subsequent processing. This is inefficient and susceptible to human interference. While some existing research has attempted to improve judgment capabilities by introducing graph neural networks or other algorithmic approaches, these methods often overlook the dynamic coupling between the physical wiring relationships and operational behavior of distribution terminals, resulting in limited diagnostic effectiveness and difficulty in practical application.

[0005] Especially when multiple types of faults occur simultaneously, traditional classification methods fail to adequately consider the dependencies between labels, which can easily lead to inaccurate identification results. Furthermore, existing control strategies are mostly fixed and lack mechanisms for adjusting them based on real-time information such as network topology and node redundancy. The execution layer also lacks closed-loop feedback, which affects overall control effectiveness.

[0006] Given these challenges, there is an urgent need for a new diagnostic method that comprehensively considers the physical connections and operational status of distribution terminals. This requires not only improved identification accuracy but also the integration of diagnostic results with control strategies, achieving a truly closed-loop "identification-response-optimization" approach. Therefore, providing an IoT-based intelligent diagnosis method and system for distribution terminal equipment faults is a pressing challenge for those skilled in the art. Summary of the Invention

[0007] One purpose of the present invention is to propose an intelligent diagnosis method and system for distribution terminal equipment faults based on the Internet of Things. The present invention integrates dynamic graph modeling, multi-scale graph diffusion reasoning and multi-label classification, and describes in detail the joint modeling and intelligent response mechanism of the physical connection relationship and disturbance behavior coupling relationship between distribution terminals. It has the advantages of high fault identification accuracy, strong response control closed-loop, and strong adaptability to dynamic topology structures.

[0008] According to an embodiment of the present invention, a method for intelligent diagnosis of power distribution terminal equipment faults based on the Internet of Things includes the following steps: S1. Collect the voltage, current, active power, reactive power, frequency, status code and sampling timestamp of the power distribution terminal to build operation data; S2. Perform sliding window processing on the operating data to extract node power change trends, calculate the power disturbance synchronization between any node pairs, generate a behavior coupling relationship matrix, generate a topology connection relationship matrix based on the terminal physical connection information, calculate the number of connection redundant paths, and obtain the redundancy index of each node; S3. Fusion of the topological connection relationship matrix and the behavioral coupling relationship matrix to construct a dynamic graph containing structural edges and behavioral edges; S4. Input the dynamic graph into the dual-scale graph diffusion inference network, perform local diffusion based on structural edges and global diffusion based on behavioral edges respectively, and fuse the diffusion paths through the gating mechanism to output the node state feature vector; S5. Input the node status feature vector into a trainable multi-label classification model to identify the fault type and probability and output the classification label; S6. Based on the fault type and redundancy index, generate a response instruction including a control action and a scheduling target; S7. Execute the response instructions, record the control results, update the behavior coupling relationship matrix and the parameters of the multi-label classification model, and complete the diagnostic closed-loop optimization.

[0009] Optionally, the S2 specifically includes: S21. Active power time series of each distribution terminal node Perform sliding window processing and set the window step size to , calculate the nodes in each time window Power change rate ,in, Indicates the first Sampling moments; S22, for any two distribution terminal nodes and , extract the power change rate sequence within the same time window and ,in, Indicates sampling At each moment, calculate the power disturbance synchronization , define the behavioral coupling relationship matrix No. Item is ; S23. Extract the direct connection status between nodes based on the physical connection relationship between nodes and construct a topological connection relationship matrix , topological connection relationship matrix is a symmetric adjacency matrix representing the structural connectivity elements of the distribution network and is defined as follows: ; S24, based on the topological connection relationship matrix , use the depth-first search algorithm to enumerate each node All reachable paths, count the number of reachable redundant paths , and calculate the node redundancy index ,in, Indicates the maximum number of redundant paths among all nodes in the current power distribution network. For measuring nodes The relative connection redundancy level of the node. A higher redundancy index means that the node has more alternative paths for carrying out failover tasks.

[0010] Optionally, the depth-first search algorithm in S24 is used to enumerate each node The process of finding all reachable paths includes: taking each distribution terminal as the starting node, building a connection relationship graph based on the topological connection relationship matrix, and using the depth-first search algorithm to traverse all paths connected to the node. During the traversal process, the path length is limited to not more than the maximum allowed number of hops set in the system, and any path containing repeated nodes or forming a closed loop is skipped. Each path that meets the requirements of no duplication, no loop, and can reach other nodes within the limited length range is counted as a valid path. The total number of valid paths that can be enumerated starting from the distribution terminal is the number of reachable paths for the node.

[0011] Optionally, the process of constructing a dynamic graph in S3 includes: using the topological connection relationship matrix as the connection basis of the structural edge, establishing the structural edge between the distribution terminals, using the behavioral coupling relationship matrix as the weight reference, establishing the behavioral edge that reflects the synchronization relationship of the power disturbance between the nodes, marking the attribute categories of the structural edge and the behavioral edge respectively, and integrating them into a heterogeneous graph structure. In the heterogeneous graph, each node represents a distribution terminal, and the edge type is divided into physical connection edge and behavioral coupling edge, where the structural edge indicates that there is an actual wiring connection between the nodes, and the behavioral edge indicates that the power disturbance between the nodes is synchronous.

[0012] Optionally, the S4 specifically includes: S41, using the structural edge corresponding to the topological connection relationship matrix as the local diffusion path input channel, performing local graph state propagation based on the adjacent node set of each node, and generating a structural edge diffusion output; S42, using the behavior edge corresponding to the behavior coupling relationship matrix as the global diffusion path input channel, propagating the behavior information between non-adjacent nodes based on the synchronization weights between the nodes, and generating the behavior edge diffusion output; S43, generate node state vectors for the structural edge diffusion output and the behavioral edge diffusion output respectively, and record them as structural state vectors respectively. and behavior state vector , the structural state vector and behavior state vector In the input gating mechanism, by calculating the structural state vector and behavior state vector The correlation distribution, self-information and mutual information feature density in the vector space are adaptively generated by the gating mechanism to generate a fusion weight coefficient vector, dynamically controlling the fusion ratio of the structural diffusion output and the behavioral diffusion output; S44, use the fusion weight coefficient vector output by the gating mechanism to perform weighted fusion on the diffusion results of the structural edge and the behavioral edge to generate the fused node state feature vector .

[0013] Optionally, the S5 specifically includes: the process of inputting the node state feature vector into the multi-label classification model to identify the fault type in the S5 includes: inputting the node state feature vector generated by the gating mechanism into the multi-label classification model to identify the fault type Input into a trainable multi-label classification model, which includes an input layer, a feature compression layer, a label association modeling layer and an output layer. In the feature compression layer, the node state feature vector Perform dimension compression to obtain compressed feature vector , in the label association modeling layer, a label co-occurrence matrix based on historical sample statistics is constructed, multi-label feature mapping is performed based on the label co-occurrence matrix, and a multi-label embedding vector is generated. , in the output layer, the multi-label embedding vector Execute the activation operation and output the fault type and probability score corresponding to each node. The fault types include power failure, overvoltage abnormality, undervoltage abnormality, current imbalance, reactive power abnormality, communication abnormality and status code abnormality.

[0014] Optionally, S6 specifically includes: retrieving the corresponding control action according to the fault type output by the multi-label classification model, adjusting the control action in combination with the redundancy index of the node, determining the scheduling target parameters according to the fault type and the position of the node in the topology, and generating a response instruction including the control action and the scheduling target.

[0015] According to an embodiment of the present invention, an intelligent fault diagnosis system for power distribution terminal equipment based on the Internet of Things includes the following modules: The operation data acquisition module is used to collect the voltage, current, active power, reactive power, frequency, status code and sampling timestamp of the distribution terminal to construct the operation data; A relationship matrix building module is used to generate a behavior coupling relationship matrix and a topology connection relationship matrix based on operation data; A redundancy calculation module is used to calculate the number of redundant paths of each node based on the topological connection relationship matrix; A dynamic graph construction module is used to fuse structural edges and behavioral edges to generate a heterogeneous graph structure representing the status of distribution terminals; Graph diffusion fusion module, which performs local diffusion on the structural side and global diffusion on the behavioral side, and outputs the fused state vector through a gating mechanism; Multi-label classification module, used to input the fused state vector into the multi-label classification model to identify the fault type and probability score; Response strategy generation module, used to generate control actions and scheduling target parameters based on fault type and redundancy index; The closed-loop update module is used to execute response instructions and feedback control results, and update the parameters of the behavior coupling relationship and the multi-label classification model.

[0016] The beneficial effects of the present invention are: (1) The present invention constructs a dynamic graph model that integrates the topological connection relationship and the power disturbance synchronization relationship, combines the dual-scale graph diffusion mechanism with the gated fusion strategy, extracts the structural and behavioral correlation characteristics between distribution terminals, and improves the fault location accuracy under complex topological structures, especially in concurrent multi-fault scenarios, with good recognition robustness.

[0017] (2) The present invention introduces a multi-label classification model to identify specific fault types, combines it with the redundancy index of the distribution terminal, generates structured response instructions from the preset strategy library, clarifies the control action type and scheduling target, and enables the diagnosis results to directly drive the control behavior. It constructs a full-process closed-loop mechanism from identification, response to model update, and solves the problem of "disconnection between identification and response" in traditional solutions.

[0018] (3) The present invention supports dynamic adjustment of graph structure and classification model parameters during operation, and realizes online retraining and self-optimization by executing feedback to update the behavior coupling relationship and classification model weights. The system has the ability to quickly adapt to changes in wiring methods and fluctuations in electrical status, which can help build an adaptive intelligent distribution terminal management solution for the Internet of Things environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is an overall flow chart of an intelligent fault diagnosis method for power distribution terminal equipment based on the Internet of Things proposed by the present invention; Figure 2 This is a module connection diagram of an IoT-based intelligent fault diagnosis system for power distribution terminal equipment proposed by the present invention. DETAILED DESCRIPTION

[0020] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0021] refer to Figure 1 and Figure 2 , an intelligent diagnosis method for power distribution terminal equipment fault based on the Internet of Things, comprising the following steps: S1. Collect the voltage, current, active power, reactive power, frequency, status code and sampling timestamp of the power distribution terminal to build operation data; S2. Perform sliding window processing on the operating data to extract node power change trends, calculate the power disturbance synchronization between any node pairs, generate a behavior coupling relationship matrix, generate a topology connection relationship matrix based on the terminal physical connection information, calculate the number of connection redundant paths, and obtain the redundancy index of each node; S3. Fusion of the topological connection relationship matrix and the behavioral coupling relationship matrix to construct a dynamic graph containing structural edges and behavioral edges; S4. Input the dynamic graph into the dual-scale graph diffusion inference network, perform local diffusion based on structural edges and global diffusion based on behavioral edges respectively, and fuse the diffusion paths through the gating mechanism to output the node state feature vector; S5. Input the node status feature vector into a trainable multi-label classification model to identify the fault type and probability and output the classification label; S6. Based on the fault type and redundancy index, generate a response instruction including a control action and a scheduling target; S7. Execute the response instructions, record the control results, update the behavior coupling relationship matrix and the parameters of the multi-label classification model, and complete the diagnostic closed-loop optimization.

[0022] This invention achieves end-to-end automated fault diagnosis for distribution terminal equipment by constructing a complete process from distribution terminal operational data collection, coupling relationship identification, graph modeling, graph diffusion fusion, fault identification, to response strategy feedback. Compared to existing technical approaches that rely on manual threshold setting and single-point analysis, this invention significantly improves diagnostic accuracy and system adaptability through a dynamic data-driven approach that integrates structural behavior. It is particularly suitable for real-world operating environments with complex topologies and frequent power disturbances, and offers greater practicality and robustness.

[0023] In this embodiment, S2 specifically includes: S21. Active power time series of each distribution terminal node Perform sliding window processing and set the window step size to , calculate the nodes in each time window Power change rate : ; in, Indicates the first Sampling time, a total of a moment; S22, for any two distribution terminal nodes and , extract the power change rate sequence within the same time window and , calculate the power disturbance synchronization , define the behavioral coupling relationship matrix No. Item is : ; in, and Represents nodes respectively With node The average value of the power change rate sequence in the current time window, the behavioral coupling relationship matrix Characterizes the consistency of disturbance trends between different nodes. When it is close to 1, it means that there is a highly synchronized power disturbance trend between the two nodes; The present invention introduces a formula for calculating the degree of synchronization of disturbances between nodes based on the rate of change of power disturbances, which can characterize the dynamic behavior similarity between any pair of distribution terminal nodes within a sliding time window. This formula constructs a behavioral coupling relationship matrix with more physical behavioral significance by modeling the correlation of active power change trends. Compared with traditional association judgment methods based on static numerical values ​​or fixed thresholds, this formula has stronger dynamic adaptability and abnormal sensitivity, especially in scenarios such as fault propagation, fluctuation interference and disturbance coordination. It can more accurately capture the potential influence chain between nodes, provide high-quality edge weight support for subsequent dynamic graph modeling and graph diffusion reasoning, and significantly improve the accuracy of fault identification and causal analysis capabilities.

[0024] S23. Extract the direct connection status between nodes based on the physical connection relationship between nodes and construct a topological connection relationship matrix , topological connection relationship matrix is a symmetric adjacency matrix representing the structural connectivity elements of the distribution network and is defined as follows: ; S24, based on the topological connection relationship matrix , use the depth-first search algorithm to enumerate each node All reachable paths, count the number of reachable redundant paths , and calculate the node redundancy index ,in, Indicates the maximum number of redundant paths among all nodes in the current power distribution network. For measuring nodes The relative connection redundancy level of the node. A higher redundancy index means that the node has more alternative paths for carrying out failover tasks.

[0025] This invention introduces a mechanism for calculating disturbance synchronization and constructing a relationship matrix, enabling quantitative modeling of inter-terminal operational behavior. Compared to traditional methods based solely on physical connections or static numerical judgments, this invention dynamically identifies the degree of behavioral correlation between distribution terminals based on power variation trends. This can reflect the characteristics of multi-terminal coordinated response or interference propagation, offering significant advantages in identifying potential fault propagation paths and building a more realistic terminal relationship network.

[0026] In this embodiment, the depth-first search algorithm used in S24 is used to enumerate each node The process of finding all reachable paths includes: taking each distribution terminal as the starting node, building a connection relationship graph based on the topological connection relationship matrix, and using the depth-first search algorithm to traverse all paths connected to the node. During the traversal process, the path length is limited to not more than the maximum allowed number of hops set in the system, and any path containing repeated nodes or forming a closed loop is skipped. Each path that meets the requirements of no duplication, no loop, and can reach other nodes within the limited length range is counted as a valid path. The total number of valid paths that can be enumerated starting from the distribution terminal is the number of reachable paths for the node.

[0027] This paper proposes a depth-first search method based on a topological connectivity matrix to accurately calculate the number of redundant paths for distribution terminal nodes, thereby obtaining a node redundancy index. This index can be used to assist in generating more feasible and safe response strategies. This addresses the problem of existing response mechanisms ignoring topological redundancy and lacking control and fault tolerance, providing decision support for strategy optimization and high-reliability power supply.

[0028] In this embodiment, the process of constructing a dynamic graph in S3 includes: using the topological connection relationship matrix as the connection basis of the structural edge, establishing the structural edge between the distribution terminals, using the behavioral coupling relationship matrix as the weight reference, establishing the behavioral edge that reflects the synchronization relationship of the power disturbance between the nodes, marking the attribute categories of the structural edge and the behavioral edge respectively, and integrating them into a heterogeneous graph structure. In the heterogeneous graph, each node represents a distribution terminal, and the edge type is divided into physical connection edge and behavioral coupling edge. The structural edge indicates that there is an actual wiring connection between the nodes, and the behavioral edge indicates that the power disturbance between the nodes is synchronous.

[0029] This paper constructs a heterogeneous dynamic graph model by fusing structural and behavioral edges, effectively integrating the physical connection information between distribution terminals and the power disturbance coupling relationship. Compared with methods that only use static wiring diagrams or single feature graphs, this graph modeling approach can express multi-type and multi-granular node interactions, possesses dynamic update capabilities, and provides stronger structural expression capabilities, providing a high-quality input foundation for graph diffusion reasoning.

[0030] In this embodiment, the S4 specifically includes: S41, using the structural edge corresponding to the topological connection relationship matrix as the local diffusion path input channel, performing local graph state propagation based on the adjacent node set of each node, and generating a structural edge diffusion output; S42, using the behavior edge corresponding to the behavior coupling relationship matrix as the global diffusion path input channel, propagating the behavior information between non-adjacent nodes based on the synchronization weights between the nodes, and generating the behavior edge diffusion output; S43, generate node state vectors for the structural edge diffusion output and the behavioral edge diffusion output respectively, and record them as structural state vectors respectively. and behavior state vector , the structural state vector and behavior state vector In the input gating mechanism, by calculating the structural state vector and behavior state vector The correlation distribution, self-information and mutual information feature density in the vector space are adaptively generated by the gating mechanism to generate the fusion weight coefficient vector, dynamically controlling the fusion ratio of the structure diffusion output and the behavior diffusion output. The specific process includes: first calculating the structure state vector and behavior state vector The correlation distribution in the vector space is used to measure its directional consistency; secondly, the self-information of the two vectors is calculated to measure the independent feature density carried by each; then the mutual information feature density between the two is calculated to evaluate the degree of feature overlap; finally, the correlation distribution, self-information and mutual information feature density are used as input features of the gating function, and a trainable gating network is used to generate a fusion weight coefficient vector to perform the gating on the structural state vector. and behavior state vector Perform element-by-element weighted fusion to achieve dynamic proportional control of the output of structural and behavioral diffusion paths.

[0031] S44, use the fusion weight coefficient vector output by the gating mechanism to perform weighted fusion on the diffusion results of the structural edge and the behavioral edge to generate the fused node state feature vector .

[0032] This paper proposes a dual-scale graph diffusion inference network and a gated fusion mechanism. This propagates state information along both structural and behavioral edge paths, with the gating mechanism adaptively calculating fusion weights to achieve dynamic regulation of information fusion and feature extraction. Compared to traditional graph neural networks, which rely on fixed diffusion paths or average aggregation, this approach can flexibly adjust propagation bias in different scenarios, effectively improving the graph model's ability to represent complex power disturbance characteristics.

[0033] In this embodiment, the S5 specifically includes: the process of inputting the node state feature vector into the multi-label classification model to identify the fault type in the S5 includes: inputting the node state feature vector generated by the gating mechanism into the multi-label classification model to identify the fault type Input into a trainable multi-label classification model, which includes an input layer, a feature compression layer, a label association modeling layer and an output layer. In the feature compression layer, the node state feature vector Perform dimension compression to obtain compressed feature vector , in the label association modeling layer, a label co-occurrence matrix based on historical sample statistics is constructed, multi-label feature mapping is performed based on the label co-occurrence matrix, and a multi-label embedding vector is generated. , in the output layer, the multi-label embedding vector Execute the activation operation and output the fault type and probability score corresponding to each node. The fault types include power failure, overvoltage abnormality, undervoltage abnormality, current imbalance, reactive power abnormality, communication abnormality and status code abnormality.

[0034] This paper designs a trainable multi-label classification model that uses the fused node state vector as input to identify multiple fault labels, including power outages, overvoltage, undervoltage, current imbalance, reactive power anomalies, communication anomalies, and status code anomalies. It supports probabilistic output and modeling of inter-label associations. This model overcomes the limitations of traditional classification methods, which cannot handle label co-occurrence or ignore multiple class combinations. It improves the accuracy and generalization of fault identification, providing reliable support for intelligent diagnosis in complex scenarios.

[0035] In this embodiment, the S6 specifically includes: retrieving the corresponding control action according to the fault type output by the multi-label classification model, adjusting the control action in combination with the redundancy index of the node, determining the scheduling target parameters according to the fault type and the position of the node in the topology, and generating a response instruction including the control action and the scheduling target. The control action includes instructions such as cutting off power supply, reducing voltage operation, switching to a backup channel, remotely restarting the communication module, and issuing a manual inspection request; the scheduling target parameters are used to limit the operating goals and execution conditions that the control action should achieve, including key parameters such as the target voltage value, current threshold, action duration, target switching node number, and backup path activation flag.

[0036] Based on fault type labels and node redundancy indicators, this invention matches and adjusts response instructions containing control actions and scheduling targets from a policy library, achieving an automatic transition from diagnostic results to execution control. Compared with traditional approaches that rely on manual judgment and fixed control instructions, this invention achieves a closed-loop diagnosis-response linkage, improving response efficiency and policy adaptability, and enhancing the intelligent level of terminal-level automatic control.

[0037] According to an embodiment of the present invention, an intelligent fault diagnosis system for power distribution terminal equipment based on the Internet of Things includes the following modules: The operation data acquisition module is used to collect the voltage, current, active power, reactive power, frequency, status code and sampling timestamp of the distribution terminal to construct the operation data; A relationship matrix building module is used to generate a behavior coupling relationship matrix and a topology connection relationship matrix based on operation data; A redundancy calculation module is used to calculate the number of redundant paths of each node based on the topological connection relationship matrix; A dynamic graph construction module is used to fuse structural edges and behavioral edges to generate a heterogeneous graph structure representing the status of distribution terminals; Graph diffusion fusion module, which performs local diffusion on the structural side and global diffusion on the behavioral side, and outputs the fused state vector through a gating mechanism; Multi-label classification module, used to input the fused state vector into the multi-label classification model to identify the fault type and probability score; Response strategy generation module, used to generate control actions and scheduling target parameters based on fault type and redundancy index; The closed-loop update module is used to execute response instructions and feedback control results, and update the parameters of the behavior coupling relationship and the multi-label classification model.

[0038] Through modular design, this invention divides the entire diagnostic system into multiple functional modules, including acquisition, modeling, diffusion, identification, strategy, and feedback. Each module has clear functions and data interfaces, facilitating system integration and iterative functional upgrades. Compared to traditional centralized processing structures, this system architecture offers greater scalability and engineering deployment value, making it suitable for intelligent diagnostic deployment scenarios for large-scale power distribution terminal equipment.

[0039] Embodiment 1: To verify the feasibility of the present invention, it was applied to a high-voltage substation. The distribution system covers over 100 distribution terminals, supporting residential, commercial, and small industrial loads. This substation suffers from a complex topology, diverse equipment types, and frequent electrical disturbances. Traditional operation and maintenance methods, which primarily rely on manual inspections and simple threshold alarms, are not only inefficient in diagnosis and prone to frequent false alarms, but also unable to address the cascading faults caused by inter-terminal disturbance transmission. This severely restricts the stability and recovery speed of the distribution system.

[0040] The proposed IoT-based intelligent fault diagnosis method for power distribution terminal equipment was deployed in the substation's data center. The operating data acquisition module collects voltage, current, active / reactive power, frequency, status code, and other information from all terminals on a one-minute cycle. This information is pre-processed by the edge computing node before being uploaded. The system then performs a sliding window analysis on the operating data, calculating the power disturbance change rate and synchronization between terminals, generating a behavioral coupling relationship matrix. Simultaneously, the system automatically extracts physical connection relationships based on the wiring diagram, constructs a topological connection relationship matrix, and counts the number of redundant paths for each terminal.

[0041] The graph modeling module integrates the aforementioned structural and behavioral relationships to construct a dynamic graph structure. This is then fed into a dual-scale graph diffusion inference network, which performs local diffusion based on structural edges and global diffusion based on behavioral edges. The gating mechanism dynamically generates fusion weights based on the correlation, self-information, and mutual information density between each node's structural state and behavioral state vector, enabling differential dynamic control and obtaining highly expressive node state vectors. The multi-label classification module identifies fault types in the state vector, including power outages, undervoltage anomalies, and reactive power anomalies, and outputs the probability of each type of fault. Finally, the policy module generates response instructions based on the fault type and node redundancy, enabling precise control and resource scheduling. The response results are recorded for self-updating model parameters.

[0042] Six typical terminals were selected from the actual operating data after the system was deployed for comparative analysis. As shown in the table, these terminals have different degrees of voltage fluctuation and power disturbance characteristics, the disturbance synchronization ranges from 0.33 to 0.91, and the structural redundancy is between 1 and 3. Before the model of the present invention was deployed, the average diagnostic accuracy of the system was 74.6%, and the response time was about 16 seconds. After the deployment of the present invention, the overall diagnostic accuracy was improved to 87.9%, the false alarm rate was significantly reduced, and the average response control time was shortened to less than 10 seconds. In particular, the improvement effect of nodes with high disturbance synchronization and complex path structure (such as terminals 103 and 106) was more significant, which fully verified the effectiveness and robustness of the present invention in processing power terminal fault diagnosis with high coupling complexity and fast state changes.

[0043] Table 1: Data sheet for intelligent diagnosis and verification of power distribution terminal faults

[0044] The "Distribution Terminal Fault Intelligent Diagnosis Verification Data Table" demonstrates significant advantages in actual distribution terminal applications. The table compares the performance of six typical terminal nodes before and after deploying the fault diagnosis system across multiple dimensions, including disturbance behavior, structural characteristics, diagnostic effectiveness, and response efficiency.

[0045] First, judging by the voltage fluctuation frequency and power disturbance amplitude, terminal 103 exhibited the highest power disturbance value of 1.01kW and a voltage fluctuation frequency of 7 times per day, indicating that its operating state fluctuates violently, making it a high-disturbance node. In contrast, terminal 102 exhibited relatively stable disturbance characteristics, with a power disturbance amplitude of only 0.22kW and the lowest fluctuation frequency of only 1 time per day. These differences were accurately quantified as different disturbance synchronization indicators in the behavioral coupling calculation. For example, the synchronization degrees of terminals 101 and 104 were as high as 0.91 and 0.87, respectively, indicating that they have a strong cooperative response relationship with other nodes in the network and are susceptible to the influence of neighboring disturbances.

[0046] Secondly, considering the number of redundant paths, terminals 103 and 106 both have three, indicating they are highly fault-tolerant. However, terminal 102 only has a single access path, resulting in low structural redundancy and limited reconfiguration potential after a failure. In the response strategy generation of the present invention, this redundancy metric is used to assist in scheduling path optimization and improve system resilience.

[0047] From the perspective of diagnostic performance, before initial deployment, the fault diagnosis accuracy of each terminal generally ranged from 65% to 82%, with terminal 103 having the lowest accuracy, at only 64.3%. After deploying the present invention, the accuracy significantly improved, with terminal 103 rising to 84.0%, an increase of nearly 20 percentage points, validating the superior characterization capabilities of the fused dynamic graph and dual-scale diffusion model under complex disturbances. The accuracy of other nodes, such as 101 and 104, also increased by more than 9%, with the overall average accuracy increasing by more than 13%.

[0048] In terms of response efficiency, the optimization method reduces the average response time from approximately 16 seconds to under 10 seconds. For example, terminal 105 took 16.4 seconds before optimization, but this was reduced to 10.2 seconds after optimization. Terminal 102 responded in just 8.7 seconds after optimization, demonstrating that the system maintains good decision-making efficiency despite its low-redundancy architecture.

[0049] Overall, the table truly reflects the robustness and high performance of the present invention under multi-type node conditions. It not only improves the accuracy and refinement of fault identification, but also achieves a significant improvement in response execution speed, and is suitable for intelligent diagnosis and adaptive control scenarios of large-scale power distribution terminals.

[0050] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An intelligent fault diagnosis method for power distribution terminal equipment based on the Internet of Things, characterized in that: The steps include: S1. Collect the voltage, current, active power, reactive power, frequency, status code and sampling timestamp of the power distribution terminal to build operation data; S2. Perform sliding window processing on the operating data to extract node power change trends, calculate the power disturbance synchronization between any node pairs, generate a behavior coupling relationship matrix, generate a topology connection relationship matrix based on the terminal physical connection information, calculate the number of connection redundant paths, and obtain the redundancy index of each node; S3. Fusion of the topological connection relationship matrix and the behavioral coupling relationship matrix to construct a dynamic graph containing structural edges and behavioral edges; S4. Input the dynamic graph into the dual-scale graph diffusion inference network, perform local diffusion based on structural edges and global diffusion based on behavioral edges respectively, and fuse the diffusion paths through the gating mechanism to output the node state feature vector; S5. Input the node status feature vector into a trainable multi-label classification model to identify the fault type and probability and output the classification label; S6. Based on the fault type and redundancy index, generate a response instruction including a control action and a scheduling target; S7. Execute the response instructions, record the control results, update the behavior coupling relationship matrix and the parameters of the multi-label classification model, and complete the diagnostic closed-loop optimization.

2. The method for intelligent fault diagnosis of power distribution terminal equipment based on the Internet of Things according to claim 1, characterized in that: The S2 specifically includes: S21. Active power time series of each distribution terminal node Perform sliding window processing and set the window step size to , calculate the nodes in each time window Power change rate ,in, Indicates the first Sampling moments; S22, for any two distribution terminal nodes and , extract the power change rate sequence within the same time window and ,in, Indicates sampling At each moment, calculate the power disturbance synchronization , define the behavioral coupling relationship matrix No. Item is ; S23. Extract the direct connection status between nodes based on the physical connection relationship between nodes and construct a topological connection relationship matrix , topological connection relationship matrix is a symmetric adjacency matrix representing the structural connectivity elements of the distribution network and is defined as follows: ； S24, based on the topological connection relationship matrix , use the depth-first search algorithm to enumerate each node All reachable paths, count the number of reachable redundant paths , and calculate the node redundancy index ,in, Indicates the maximum number of redundant paths among all nodes in the current power distribution network. For measuring nodes The relative connection redundancy level of the node. A higher redundancy index means that the node has more alternative paths for carrying out failover tasks.

3. The method for intelligent fault diagnosis of power distribution terminal equipment based on the Internet of Things according to claim 2, characterized in that: The depth-first search algorithm used in S24 enumerates each node The process of finding all reachable paths includes: taking each distribution terminal as the starting node, building a connection relationship graph based on the topological connection relationship matrix, and using the depth-first search algorithm to traverse all paths connected to the node. During the traversal process, the path length is limited to not more than the maximum allowed number of hops set in the system, and any path containing repeated nodes or forming a closed loop is skipped. Each path that meets the requirements of no duplication, no loop, and can reach other nodes within the limited length range is counted as a valid path. The total number of valid paths that can be enumerated starting from the distribution terminal is the number of reachable paths for the node.

4. The method for intelligent fault diagnosis of power distribution terminal equipment based on the Internet of Things according to claim 3, characterized in that: The process of constructing the dynamic graph in S3 includes: using the topological connection relationship matrix as the connection basis of the structural edge to establish the structural edge between the distribution terminals, using the behavioral coupling relationship matrix as the weight reference to establish the behavioral edge that reflects the synchronization relationship of the power disturbance between the nodes, and after marking the attribute categories of the structural edges and the behavioral edges respectively, integrating them into a heterogeneous graph structure. In the heterogeneous graph, each node represents a distribution terminal, and the edge types are divided into physical connection edges and behavioral coupling edges. The structural edge indicates that there is an actual wiring connection between the nodes, and the behavioral edge indicates that the power disturbance between the nodes is synchronized.

5. The method for intelligent fault diagnosis of power distribution terminal equipment based on the Internet of Things according to claim 4, characterized in that: The S4 specifically includes: S41, using the structural edge corresponding to the topological connection relationship matrix as the local diffusion path input channel, performing local graph state propagation based on the adjacent node set of each node, and generating a structural edge diffusion output; S42, using the behavior edge corresponding to the behavior coupling relationship matrix as the global diffusion path input channel, propagating the behavior information between non-adjacent nodes based on the synchronization weights between the nodes, and generating the behavior edge diffusion output; S43, generate node state vectors for the structural edge diffusion output and the behavioral edge diffusion output respectively, and record them as structural state vectors respectively. and behavior state vector , the structural state vector and behavior state vector In the input gating mechanism, by calculating the structural state vector and behavior state vector The correlation distribution, self-information and mutual information feature density in the vector space are adaptively generated by the gating mechanism to generate a fusion weight coefficient vector, dynamically controlling the fusion ratio of the structural diffusion output and the behavioral diffusion output; S44, use the fusion weight coefficient vector output by the gating mechanism to perform weighted fusion on the diffusion results of the structural edge and the behavioral edge to generate the fused node state feature vector .

6. The method for intelligent fault diagnosis of power distribution terminal equipment based on the Internet of Things according to claim 5, characterized in that: The S5 specifically includes: the process of inputting the node state feature vector into the multi-label classification model to identify the fault type in the S5 includes: Input into a trainable multi-label classification model, which includes an input layer, a feature compression layer, a label association modeling layer and an output layer. In the feature compression layer, the node state feature vector Perform dimension compression to obtain compressed feature vector , in the label association modeling layer, a label co-occurrence matrix based on historical sample statistics is constructed, multi-label feature mapping is performed based on the label co-occurrence matrix, and a multi-label embedding vector is generated. , in the output layer, the multi-label embedding vector Execute the activation operation and output the fault type and probability score corresponding to each node. The fault types include power failure, overvoltage abnormality, undervoltage abnormality, current imbalance, reactive power abnormality, communication abnormality and status code abnormality.

7. The method for intelligent fault diagnosis of power distribution terminal equipment based on the Internet of Things according to claim 6, characterized in that: The S6 specifically includes: retrieving the corresponding control action according to the fault type output by the multi-label classification model, adjusting the control action in combination with the redundancy index of the node, determining the scheduling target parameters according to the fault type and the position of the node in the topology, and generating a response instruction including the control action and the scheduling target.

8. An intelligent fault diagnosis system for power distribution terminal equipment based on the Internet of Things, applied to an intelligent fault diagnosis method for power distribution terminal equipment based on the Internet of Things according to any one of claims 1 to 7, characterized in that: Includes the following modules: The operation data acquisition module is used to collect the voltage, current, active power, reactive power, frequency, status code and sampling timestamp of the distribution terminal to construct the operation data; A relationship matrix building module is used to generate a behavior coupling relationship matrix and a topology connection relationship matrix based on operation data; A redundancy calculation module is used to calculate the number of redundant paths of each node based on the topological connection relationship matrix; A dynamic graph construction module is used to fuse structural edges and behavioral edges to generate a heterogeneous graph structure representing the status of distribution terminals; Graph diffusion fusion module, which performs local diffusion on the structural side and global diffusion on the behavioral side, and outputs the fused state vector through a gating mechanism; Multi-label classification module, used to input the fused state vector into the multi-label classification model to identify the fault type and probability score; Response strategy generation module, used to generate control actions and scheduling target parameters based on fault type and redundancy index; The closed-loop update module is used to execute response instructions and feedback control results, and update the parameters of the behavior coupling relationship and the multi-label classification model.

Images