Power distribution cabinet fault prediction and self-healing control method and system

Abstract

The present application relates to the field of power distribution system intelligent operation and maintenance, in particular to a power distribution cabinet fault prediction and self-healing control method and system. The method comprises: obtaining multi-physical field sensor raw data, constructing a continuous time sequence and node binding structure; combining with the device topology to construct a weighted graph structure, using a graph convolution network to extract features and generate a node health degree atlas; calculating the abnormal probability based on the health degree and dividing the risk level; further constructing a digital twin model, deducing and optimizing the optimal recovery path; generating control instructions according to the path, executing self-healing actions and recording action logs; finally, generating the logs into a certificate structure and writing into a blockchain to complete the trusted operation and maintenance closed loop. Compared with the prior art, the present application can realize accurate prediction of implicit faults, dynamic path self-healing and tamper-proof chain recording, effectively improving the fault handling efficiency, automation level and system reliability of the power distribution cabinet.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of power distribution systems, and in particular to a method and system for predicting and self-healing faults in power distribution cabinets. Background Technology

[0002] With the development of smart grids and distributed energy systems, distribution cabinets, as key nodes in power distribution systems, face challenges such as high operating loads, rapid equipment aging, and complex external environments, significantly increasing the risk of failure. Traditional fault prediction methods mainly rely on fixed threshold monitoring and rule-based judgment, which are insufficient to handle the complex dynamic patterns implied by multi-source heterogeneous sensor data. Especially when facing complex fault scenarios such as electromagnetic disturbances, thermal runaway, and mechanical vibration coupling, their fault warning accuracy and real-time performance are severely inadequate. Furthermore, existing systems lack comprehensive modeling and extrapolation capabilities for node states when performing self-healing operations, often relying on static configurations and human experience, failing to achieve accurate simulation of fault propagation paths and closed-loop command control. Currently, there is also a lack of traceable and verifiable maintenance behavior documentation mechanisms, making it difficult to meet the requirements of intelligent, automated, and reliable power distribution system operation and management.

[0003] Therefore, there is an urgent need to construct a fault prediction and self-healing control framework that integrates multi-physics field sensing data, equipment topology, and advanced graph neural network algorithms. This framework can achieve accurate quantification and predictive analysis of node states through graph structure modeling, graph convolution feature extraction, and health graph construction. Furthermore, it can complete digital twin simulation, control command generation, and blockchain-based chain-based evidence storage operations, thereby solving core problems in the operation and maintenance of existing power distribution cabinets, such as prediction lag, inaccurate self-healing response, and lack of information traceability. Summary of the Invention

[0004] This invention provides a method and system for fault prediction and self-healing control of power distribution cabinets, which solves the problem of how to construct a weighted graph structure based on multi-physics time series data and equipment topology relationship, extract node state features through graph convolutional networks and complex analysis operators and generate health maps, so as to achieve accurate prediction and quantifiable evaluation of potential fault nodes in power distribution cabinets, and provide a unified state basis for subsequent digital twin inference, self-healing control command generation and chain-based operation and maintenance evidence storage.

[0005] To address the aforementioned technical problems, this invention provides a method for fault prediction and self-healing control of power distribution cabinets, comprising:

[0006] Acquire raw data, construct continuous time series data structures and configure node bindings to generate time series data structures for graph construction;

[0007] The time series data structure is combined with the device topology to construct a graph structure, and graph convolution processing is performed to generate a node health graph.

[0008] The expression for the graph structure is:

[0009]

[0010] in, The diagram shows the structural relationships; A set of nodes; It is an edge set; Let the edge weight matrix be the edge weight matrix.

[0011] The expression for graph convolution processing is:

[0012]

[0013] in, Indicates the final embedded features; This represents the vector in the i-th row of the feature matrix of the L-th layer of the graph convolutional network; This represents the total number of layers in a graph convolutional neural network (GCN) model. Representing vectors Belongs to the dimension The real number space; This represents the number of dimensions of the output features of the Lth layer of the graph convolutional network;

[0014] For the final embedded features Execute the second-order complex Hessian kernel function Laplace gradient operation:

[0015]

[0016] in, For the first Health score of each node; The second-order derivative tensor of the embedded features after complex analysis kernel function; It is the Frobenius norm;

[0017] The expression for the node health graph is:

[0018]

[0019] in, Represents the node health graph; Indicates the first Each device data acquisition unit; This indicates the total number of sensor nodes deployed in the device;

[0020] Based on the node health graph, the probability of anomalies is calculated, clustering and structural coding are performed, and a risk level structure is generated.

[0021] A digital twin structure is constructed using the risk level structure and node health graph, and then deduced and optimized to generate the optimal recovery path structure.

[0022] Generate a set of control instructions based on the optimal recovery path structure, execute self-healing actions and associated records, and generate an action scheduling log structure.

[0023] Obtain the action scheduling log structure, generate a standardized evidence storage structure and write it to the blockchain to form chain-based operation and maintenance evidence storage data.

[0024] Furthermore, the steps of acquiring raw data, constructing a continuous time series data structure and configuring node bindings to generate a time series data structure for graph construction include the following:

[0025] The raw data collected by temperature sensors, electromagnetic sensors and vibration sensors deployed in the power distribution cabinet are obtained, and time-series segmentation and format standardization are performed to obtain preliminary time-series data.

[0026] Missing values ​​were imputed and marginal filtering was performed on the initial time series data to construct a continuous time series data structure;

[0027] The continuous time series data structure is normalized and configured with node binding to generate a time series data structure for graph construction.

[0028] Furthermore, the step of constructing a graph structure by combining the time-series data structure with the device topology, performing graph convolution processing, and generating the graph structure in the node health graph includes:

[0029] Obtain the time-series data structure and construct a graph structure relationship matrix between nodes by combining the topological relationships of device nodes.

[0030] Furthermore, the graph convolution processing includes the following steps:

[0031] The graph structure relation matrix and time series data structure are input into the graph convolutional network model for training, generating multi-layer graph convolutional feature representations.

[0032] Furthermore, the generation of the node health graph includes the following steps:

[0033] The state evaluation and health mapping processes are performed on the convolutional feature representation of multi-layer graphs to generate a node health map.

[0034] Furthermore, the generation of the risk level structure includes the following steps:

[0035] Obtain the node health graph, calculate the probability of failure for each node in the graph, and extract the anomaly probability information;

[0036] Clustering and segmenting of anomaly probability information is performed to identify different risk level intervals;

[0037] The risk level ranges are labeled and coded to generate a risk level structure.

[0038] Furthermore, generating the optimal recovery path structure includes the following steps:

[0039] Obtain the risk level structure and node health graph, and construct a digital twin structure containing operational status parameters;

[0040] Perform fault propagation simulation and path state update on the digital twin structure, and deduce multiple fault recovery path schemes;

[0041] The fault recovery path schemes are scored and optimized to generate the optimal recovery path structure.

[0042] Furthermore, the generation of the action scheduling log structure includes the following steps:

[0043] Obtain the optimal recovery path structure, perform control command mapping processing, and generate a control instruction set;

[0044] The control cabinet control interface is invoked according to the control instruction set to perform self-healing actions such as circuit switching and load transfer;

[0045] The execution results of the self-healing action are associated with the control instruction set and recorded to generate an action scheduling log structure.

[0046] Furthermore, the generation of chained operation and maintenance evidence data includes the following steps:

[0047] Obtain the action scheduling log structure and construct a log data unit containing execution time, command content, and result status;

[0048] The log data units are encrypted, hashed, and signed to generate a standardized evidence storage structure.

[0049] A standardized evidence storage structure is written into the blockchain network to generate chain-based operation and maintenance evidence storage data.

[0050] Furthermore, a power distribution cabinet fault prediction and self-healing control system, applied to any one of the above methods, includes:

[0051] The raw data processing module is used to acquire raw data, construct continuous time series data structures and configure node binding, and generate time series data structures for graph construction.

[0052] The graph structure modeling module is used to combine time series data structures with device topology to construct graph structures, perform graph convolution processing, and generate node health graphs.

[0053] The risk identification module is used to calculate the probability of anomalies based on the node health graph, perform clustering and structural coding, and generate a risk level structure.

[0054] The digital twin simulation module is used to construct a digital twin structure using the risk level structure and node health graph, and to perform simulation and optimization to generate the optimal recovery path structure.

[0055] The control execution module is used to generate a set of control instructions based on the optimal recovery path structure, execute self-healing actions and associated records, and generate an action scheduling log structure.

[0056] The operation and maintenance evidence storage module is used to obtain the action scheduling log structure, generate a standardized evidence storage structure and write it into the blockchain to form chain-like operation and maintenance evidence storage data.

[0057] The key innovations of this invention include:

[0058] (1) A weighted graph construction method integrating multi-physics field sensing data and equipment topology is proposed, and the edge weights are quantized based on the Shannon entropy function, thereby improving the accuracy of state structure modeling of power distribution equipment.

[0059] (2) Introducing a modeling mechanism that combines graph convolutional networks and complex analysis state kernel functions to form a node health score map, transforming the traditional discrete fault detection method into a differentiable and trainable continuous prediction model.

[0060] (3) A closed-loop “identification-recovery-recording” chain is constructed, integrating three major mechanisms: digital twin simulation, self-healing control instruction set generation and blockchain trusted operation and maintenance record, realizing a full-process intelligent control system from state perception to action execution and then to trusted evidence storage.

[0061] The following are its main beneficial effects:

[0062] (1) This invention achieves deep fusion of multi-physics information and graph structure modeling, significantly improving the accuracy and foresight of fault prediction. By normalizing, binding nodes, and mapping topological data from multiple physical fields such as temperature, electromagnetic fields, and vibration, this invention constructs a node-level time-series data structure and introduces a Shannon entropy-enhanced edge weight calculation mechanism to form a dynamically weighted graph structure relation matrix. Compared to the traditional method that relies solely on single-dimensional sensor threshold judgment, this method can capture the potential coupling effects and anomaly propagation trends between nodes, effectively improving the ability to identify fault precursors and enabling fault prediction to move from "static rules" to "structure learning," thereby increasing accuracy.

[0063] (2) A combined modeling method of graph convolutional networks and complex analysis kernel functions is introduced to construct a node health map, realizing quantitative assessment and hierarchical management of node status. This invention innovatively embeds a multi-layer graph convolutional computation process into the graph neural network, extracts cross-node correlation feature vectors, and calculates the state gradient response based on the complex Hessian kernel function to generate a node health score. The health score, as an interval normalized result, reflects the structural stability and potential failure risk of each node. This mechanism breaks the limitation of traditional state assessment methods relying on fixed rules, realizes continuous quantification of fault states, supports more refined risk identification and level classification, and provides a structural basis for subsequent digital twin simulation and recovery path selection.

[0064] (3) Construct a closed-loop system for self-healing control and trusted recording to achieve an automated operation and maintenance process of "fault identification - path optimization - control execution - blockchain evidence storage". Based on the health graph and risk level structure, this invention constructs a digital twin model that integrates state factors and risk structure, performs multiple path deductions and optimizations, and forms the optimal recovery path structure. Subsequently, the system automatically maps and generates a set of control instructions to drive the distribution cabinet to perform self-healing actions such as loop switching and load reconfiguration, and standardizes and records the action results as log data units, performs hash encryption and block signing, and finally writes them into the blockchain to form chain-based operation and maintenance evidence storage data. This mechanism ensures the verifiability and immutability of operation and maintenance execution, and effectively improves the automation level, transparency and compliance assurance capabilities of the power distribution system. Attached Figure Description

[0065] Figure 1 A flowchart illustrating the power distribution cabinet fault prediction and self-healing control method provided in the embodiments of this application; Figure 2 This is a structural block diagram of the power distribution cabinet fault prediction and self-healing control system provided in an embodiment of this application. Detailed Implementation

[0067] Example 1: Refer to Figure 1 This is a flowchart illustrating the power distribution cabinet fault prediction and self-healing control method provided in an embodiment of the present invention. The process may include at least steps S100-S600:

[0068] SS100 generates a time series data structure for graph construction by acquiring raw data, constructing a continuous time series data structure and configuring node binding.

[0069] S200: Combine the time series data structure with the device topology to construct a graph structure, perform graph convolution processing, and generate a node health graph.

[0070] S300: Calculate the anomaly probability based on the node health graph, perform clustering and structural coding, and generate a risk level structure;

[0071] S400: Construct a digital twin structure using the risk level structure and node health graph, and perform deduction and optimization to generate the optimal recovery path structure;

[0072] S500: Generate a control instruction set based on the optimal recovery path structure, execute self-healing actions and associated records, and generate an action scheduling log structure.

[0073] S600: Obtain the action scheduling log structure, generate a standardized evidence storage structure and write it into the blockchain to form chain-based operation and maintenance evidence storage data.

[0074] Step S100 includes at least steps S110-S130:

[0075] S110. Acquire the raw data collected by the temperature sensor, electromagnetic sensor and vibration sensor deployed in the power distribution cabinet, perform time-series segmentation and format standardization processing to obtain preliminary time-series data.

[0076] This step forms the initial foundation for the time series data structure. Through the synchronous acquisition and standardization of data from multiple sensors, a data structure that can be input into subsequent models is generated. Specifically:

[0077] The system deploys multiple physical sensing units within the distribution cabinet, including temperature sensors, electromagnetic sensors, and vibration sensors, to collect multi-physical field state data in real time during the cabinet's operation. The temperature sensors collect temperature rise data at cable interfaces, busbar nodes, and the surface of control modules; the electromagnetic sensors detect changes in the power frequency magnetic field strength near the main busbar and circuit breaker nodes; and the vibration sensors monitor minute vibration changes in the cabinet structure during operation to reflect the operational stability of the power equipment.

[0078] During data acquisition, a synchronous acquisition mechanism is used to ensure consistency in the sampling time and frequency of each type of sensor. The sampling time interval is set by platform configuration parameters, and the sampling frequency can be automatically adjusted according to load dynamics or equipment operating cycle to obtain multidimensional raw signal data with temporal continuity and physical integrity. Each type of raw data is labeled with the source node number, physical field type, sampling timestamp, and raw signal value.

[0079] The system performs time-series segmentation processing on the collected raw data, dividing the continuous data stream obtained according to the sampling period into fixed-length sliding windows with a uniform time step, generating original multi-dimensional time segment segments. The length and step size of the sliding window, together with the synchronization strategy, determine the temporal granularity of the data that can be used for model training.

[0080] After segmentation, format standardization processing needs to be performed. This includes uniformly mapping the data fields output by different sensors to a structured format, such as unifying temperature, electromagnetic field, and vibration data into a four-tuple format of "time-node number-physical type-signal value", unifying units, retaining decimal places and redundancy in precision, and removing data segments that do not conform to the timestamp order or have lost signals.

[0081] The system completes the acquisition, time-series segmentation, and format standardization of multi-source sensor data, generating preliminary time-series data with a unified format, time series labels, and node labels. This data serves as the input data structure required by S120.

[0082] S120. Perform missing value imputation and edge filtering on the preliminary time series data to construct a continuous time series data structure.

[0083] This step, based on the preliminary time series data generated in S110, further processes the sampled data segments that have missing or boundary anomalies, constructs a time series structure with modelable continuity, ensures data integrity and continuity, and meets the requirements for subsequent graph structure construction and modeling.

[0084] Specifically, the system performs missing value identification processing on the preliminary time series data. Missing points are identified using a dual criterion of timestamp interval and physical numerical abrupt changes: if the time interval between adjacent sampling points exceeds twice the set period, or the numerical abrupt change is greater than three times the normal fluctuation standard deviation, then that point or time period is marked as a missing segment. The system automatically builds a missing value index table based on the marking results.

[0085] During the missing value repair phase, a bidirectional interpolation strategy is used to fill in missing values ​​for points that are temporally continuous but numerically missing. For missing time periods, the system prioritizes estimating the possible value using the average of adjacent time periods and a weighted linear extrapolation strategy, while recording the filling type and filling level identifier for subsequent model adjustments. All interpolation operations are performed while preserving the original timestamp structure, avoiding structural changes.

[0086] After missing value imputation, the system proceeds to the edge filtering stage. Considering potential current surges, noise signals, or impulsive interference during the operation of the distribution cabinet, a window smoothing mechanism is introduced to perform edge filtering on each sensor data channel. Specifically, a combination of a weighted median filter and a local extremum eliminater is used to remove high-frequency interference components without altering the true trend. The filter window length is adaptively set based on the sensing data sampling frequency and boundary change rate to ensure that early abnormal signals are not mistakenly filtered out.

[0087] After missing data completion and filtering, the time series data is transformed into a complete, continuous, unbroken, and noise-constrained continuous time series data structure. This data structure serves as the input data for normalization and node binding in S130 and will participate in the subsequent graph model construction process.

[0088] S130. Normalize the continuous time series data structure and configure node binding to generate a time series data structure for graph construction.

[0089] This step, based on the continuous time series data structure constructed in S120, uses normalization transformation and node binding label operations to ultimately generate a data format that can be directly used for graph model construction. This data serves as the sole input source for subsequent graph convolution training.

[0090] Specifically, the system normalizes each type of physical field in the continuous time series data structure. This normalization process employs a scaling method based on the maximum and minimum values ​​of each channel, treating each physical sensing field as an independent channel for normalization transformation. This scales the time series of each channel to a fixed interval, improving the numerical balance and gradient stability of the model across feature dimensions. This operation ensures that temperature, electromagnetic, and vibration fields are input to the subsequent graphical model structure at a uniform numerical scale, avoiding the dominance of a single variable in training.

[0091] After normalization, node binding configuration is performed. This includes binding each time series sample to its acquisition node number, establishing a one-to-one mapping between node labels and time series samples. Node numbers are uniformly assigned during platform deployment and are consistent with the device numbers in the distribution cabinet topology, ensuring that time series data from the same node are accurately mapped to graph structure nodes during graph modeling. This process creates a three-element binding structure in the database: "node number—sensor channel—time period index," which the system subsequently uses to construct the graph structure relationships.

[0092] Furthermore, the system integrates the normalized signal values, node numbers, and channel categories into a single data structure, forming a multi-channel time series data set. Each set contains complete time steps, physical channels, and node binding relationships, providing basic data support for constructing the graph structure relationship matrix in the subsequent step S210.

[0093] Understandably, the node binding configuration result will be used as one of the basic attribute fields for constructing the graph structure relationship in S210. The system completes the edge weight calculation and adjacency matrix initialization in the graph based on the time series vector corresponding to each node and the topological relationship between nodes. Therefore, the data structure completed in S130 not only provides basic input data, but also directly affects the spatial modeling accuracy in the graph convolution training stage.

[0094] Through the implementation steps S110–S130, the system has formed a multi-channel time series data structure with a unified structure, normalized processing, completed node binding, and meeting the input requirements of graph models. This provides a standardized input foundation for constructing node health maps for graph convolutional networks and establishes a data starting point for subsequent risk assessment and fault recovery path deduction, realizing an input closed-loop processing mechanism from physical perception to graph intelligent analysis.

[0095] Step S200 includes at least steps S210-S230:

[0096] S210. Obtain the time series data structure and construct a graph structure relationship matrix between nodes by combining the topological relationship between device nodes.

[0097] Specifically, the system first calls the time-series data structure that has been normalized and configured for node binding in step S130, assuming the node set is:

[0098]

[0099] in, This indicates the total number of sensor nodes deployed in the device. Indicates the first Each device acquisition unit includes three data dimensions: temperature, electromagnetic fields, and vibration. The time series feature vector is represented as:

[0100] formula :

[0101]

[0102] in:

[0103] : No. The fused feature vector of each node;

[0104] :node In time The collected temperature values;

[0105] :node In time The collected electromagnetic signal values;

[0106] :node In time The collected vibration signal values;

[0107] Number of time series steps.

[0108] The system obtains the connectivity matrix between nodes based on the topology configuration. This indicates whether there is a physical connection between any two nodes; further, we introduce a Shannon entropy-enhanced weighting function to define the edge weights:

[0109]

[0110] in:

[0111] : No. The fused feature vector of each node;

[0112] : No. The fused feature vector of each node;

[0113] :node With nodes The edge weights between them;

[0114] Feature splicing operation;

[0115] Shannon entropy function, used to measure the complexity of node features;

[0116] :node and Whether the space is connected (1 indicates connected, 0 indicates disconnected);

[0117] This represents the natural exponential function.

[0118] The graph structure relationship is ultimately represented as a graph. edge set Edge weight matrix .

[0119] S220. Input the graph structure relation matrix and time series data structure into the graph convolutional network model for training to generate multi-layer graph convolutional feature representations.

[0120] After completing the graph structure and its weight matrix Then, the system will Features of each node Input graph convolutional network model Let the input feature matrix be... ,in The initial settings are:

[0121]

[0122] The system constructs the recursive formula for the feature representation of multi-layer graph convolution as follows:

[0123]

[0124] in:

[0125] : Graph convolution Layer input features;

[0126] : Edge weight matrix with self-loops;

[0127] For the angle matrix, ;

[0128] : No. Layer training parameter matrix;

[0129] : Activation function, using ReLU or LeakyReLU.

[0130] Graph convolutional network stack The final embedded features are generated after the layers:

[0131]

[0132] in:

[0133] : This represents the final embedded feature, indicating the feature vector output by the i-th node in the L-th layer of the Graph Convolutional Neural Network (GCN). It is the final representation of the current node after L layers of feature transfer and nonlinear transformation, and has the following characteristics: Dimension;

[0134] : Represents the vector in the i-th row of the feature matrix of the L-th layer of the graph convolutional network;

[0135] : Represents the total number of layers in a graph convolutional neural network (GCN) model;

[0136] : Represents a vector Belongs to the dimension The real number space;

[0137] : Indicates the dimension of the output features of the Lth layer of the graph convolutional network.

[0138] S230. Perform state evaluation and health mapping on the multi-layer graph convolutional feature representation to generate a node health map.

[0139] To achieve quantitative prediction and health assessment of node states in the distribution cabinet, the system further introduces a state gradient operator in the form of complex analysis to the final embedded features. Execute the second-order complex Hessian kernel function Laplace gradient operation:

[0140]

[0141] in:

[0142] : No. The health score of each node is represented by a smaller value, indicating poorer health.

[0143] The second-order derivative tensor of the embedded features after complex analysis kernel function;

[0144] : Frobenius norm.

[0145] Finally, all nodes Health score Integrate and generate a node health graph:

[0146]

[0147] in, This represents the node health graph.

[0148] This map will serve as the core input data source for risk level classification and structural coding in the subsequent S300 module.

[0149] Through the construction of this module, the system achieves high-dimensional graph structure modeling based on multi-physics field data (temperature, electromagnetic, vibration), and combines complex analysis with graph neural networks to overcome the problem that traditional threshold detection methods cannot predict latent faults. This health graph, as a core early warning indicator, can accurately identify potential fault nodes in advance, improve the accuracy of validation set predictions, significantly enhance the interpretability and practicality of fault prediction, and provide solid data support and status evidence for subsequent risk level modeling, recovery path generation, and blockchain operation and maintenance records.

[0150] Step S300 includes at least steps S310-S330:

[0151] S310. Obtain the node health map, calculate the probability of failure for each node in the map, and extract the anomaly probability information.

[0152] Specifically, the system first extracts the status score information of key equipment nodes in all distribution cabinets from the node health map generated by S230. The health map is a set of structured vectors generated by combining time series evolution data and graph convolution spatial propagation mechanism, where each vector corresponds to the health score result of a graph node within a given period window. To achieve the fault prediction function, the system needs to perform standardized mapping and anomaly calculation processing on the above score results.

[0153] The system first traverses the vector information of all graph nodes in the health graph in order of node number. The system then calls the graph structure indexing function to extract the historical rating sequence of each graph node using a sliding window, and calculates the time-related stability features and mutability indicators of that rating sequence. Based on a preset time period aggregation length, the system extracts local statistical features from the health sequence of each node and calculates derived indicators such as rating fluctuation amplitude, gradient mutation degree, and rating decline rate.

[0154] Furthermore, the system inputs the aforementioned derived indicators into the anomaly scoring mapping module. This module uses a discrimination model trained based on historical labeled samples, combined with the node scoring feature distribution within the current period, to estimate the anomaly probability of each graph node. The anomaly probability information is a set of numerical sequences that correspond one-to-one with the graph node numbers, indicating the probability of potential faults occurring in each equipment node of the current distribution cabinet structure within the prediction period.

[0155] To enhance the stability of anomaly identification, the system introduces a dual-iteration mechanism during probability extraction, cross-validating the scoring trends over time and the structural adjacency propagation trends in space. By comparing the consistency of anomaly probabilities across adjacency structures at different scales, the system can eliminate false alarms caused by individual mutations or measurement biases, thereby improving the global consistency and prediction reliability of the anomaly probability results.

[0156] Through this step, the system constructs a stable set of anomaly probability information with structural consistency support based on the node health map generated by S230, providing input basis and distinguishing parameters for subsequent risk level structure modeling.

[0157] S320. Cluster and segment the abnormal probability information to identify different risk level intervals.

[0158] After extracting the anomaly probability information in S310, the system enters the stage of risk level identification based on the probability information. This step aims to extract clustering results with risk interval distribution characteristics based on the distribution features of the anomaly probability, form risk level interval division results, and build a clear hierarchical system for subsequent structural coding.

[0159] Specifically, the system first calculates the probability distribution of all nodes in the current anomaly probability sequence, and then initializes the clustering model parameters based on the set sliding window size and the number of interval boundaries. The system adopts an improved density-driven clustering mechanism, using the anomaly probability value as the main axis of the clustering dimension and the evolutionary stability parameter of node health as an auxiliary constraint feature, to dynamically divide multiple initial clusters of risk categories.

[0160] To ensure structural consistency of the clustering results in the node space, the system introduces a structural consistency weight function during the clustering calculation process. This function uses the difference in outlier probabilities among nodes in the same topological subgraph as a cluster boundary adjustment factor, performing multiple edge corrections on the partitioning boundaries. The system uses a clustering quality evaluation index based on a graph partitioning optimization strategy to filter the partitioning results in each round until the mean probability of nodes within each cluster converges to a certain stable interval.

[0161] After the clustering operations described above are completed, the system performs segmented modeling on each resulting cluster. Segmented modeling includes interval boundary refinement, node sample representativeness extraction, and overlap matching analysis with historical label data. By comparing the clusters with the probability intervals of standard fault samples, the system can identify the risk level intervals corresponding to each cluster and map the cluster results to a predefined set of risk labels (such as "low risk," "medium risk," "high risk," "critical risk," etc.). Each risk level interval will serve as an important basis for risk level structure coding.

[0162] This step outputs a risk level interval structure containing three items: node number, risk level label, and corresponding probability interval, which serves as the basic dataset for constructing the risk level structure.

[0163] S330. Mark and code the risk level ranges to generate a risk level structure.

[0164] After completing the clustering of anomaly probabilities and identification of risk level intervals in S320, the system enters the risk level structure generation stage. This stage involves performing standardized coding and structural organization operations on the existing risk level labels to achieve hierarchical mapping and structural modeling of node failure risks.

[0165] Specifically, the system first invokes the risk level interval structure, indexes and extracts all nodes that have been assigned level labels, and constructs a mapping structure with node number as key and risk label as attribute. During the construction of the mapping relationship, the system extracts the topology of each level node, generating a risk level-topology mapping graph, which is used for level fusion in the subsequent construction of the digital twin.

[0166] Furthermore, the system encodes the constructed risk label structures. The system uses a hybrid coding rule based on hierarchical depth and topological connectivity to assign a unified structured coding prefix to each risk level, followed by a structural suffix related to the node spatial number, thus forming a complete risk level code. Each structure code contains three types of content: risk level label information, node spatial information, and time period label, which can be used to uniquely identify the risk level identity of a specific node within a specific period.

[0167] To enhance the schedulability and traceability of this structure in subsequent modules, the system also associates and binds the aforementioned structure code with the aforementioned anomaly probability information, and records it in the system's level information index table. This table can be called by the digital twin structure in the subsequent S400 to realize risk level perception and path optimization in the fault propagation simulation process.

[0168] This step generates a complete risk level structure, which includes a structure coding index table, a node level mapping table, and a level topology subgraph. This provides level criteria and constraint parameters for the next stage of fault propagation path deduction and recovery path scoring.

[0169] Through the implementation path of the S300 module of this invention patent, the system successfully realized a closed-loop mapping process from node health map to risk level structure. S310 achieved a precise conversion from healthy state to fault probability, S320 completed the structural division of probability distribution into level partitions, and S330 implemented level constraint input in the fault recovery path construction process through the structural encoding of level labels. The entire module execution process ensured that node fault risk was quantifiable, classifiable, and schedulable, providing a controllable foundation and accurate support for subsequent digital twin-driven self-healing strategies, and overall enhancing the predictive maintenance capability and intelligent recovery efficiency of the power distribution cabinet system.

[0170] S410. Obtain the risk level structure and node health graph, and construct a digital twin structure containing operational status parameters.

[0171] The core task of this step is to integrate the risk level structure and node health map generated by the preceding modules, and to construct a digital twin structure based on this. Specifically, the system first calls the risk level structure completed in step S330 to extract the risk level interval labels identified by each node in different time segments. The risk level structure has combined anomaly probability information, clustering partitioning models, and level coding rules to identify the risk evolution trend of each key node in the system. At the same time, the system further calls the node health map generated in step S230 to extract the current health value of each node and its time-evolutionary health trend characteristics.

[0172] Based on the two core inputs mentioned above, the system constructs a digital twin structure. Specifically, each node is considered the smallest unit for twin modeling. Node values ​​in the health graph are mapped to health status weights in the twin model, and level intervals in the risk level structure are mapped to the risk strain coefficient of that twin node. Furthermore, topology numbers and associated constraint information consistent with the node binding configuration in step S130 are introduced, organizing the aforementioned twin nodes into a real-time evolving digital twin graph structure according to the actual distribution cabinet structure. The system simultaneously calls the continuous time-series data structure completed in step S120 to drive the dynamic updates of the twin node state parameters over time.

[0173] Furthermore, to enhance the scalability and dynamic response capability of the twin structure modeling, the system categorizes the state parameters of each twin node into three types: first, initial state parameters generated based on the health graph; second, level weighting factors generated based on the risk level structure; and third, operational status records dynamically updated through continuous time series data. These three types of state parameters serve as the initial values, adjustment factors, and dynamic update functions of the twin node during twin construction, respectively, forming a complete set of state vectors. Ultimately, the system generates a digital twin structure with topological structure, health status, risk level, and dynamic evolution capabilities for subsequent fault propagation and path prediction.

[0174] S420: Perform fault propagation simulation and path state update on the digital twin structure, and deduce multiple fault recovery path schemes.

[0175] After constructing the digital twin structure containing operational status parameters, the system proceeds to this step, formally executing fault propagation simulation and path state update processing based on the twin structure. Specifically, the system first identifies the set of nodes currently at the edge health threshold based on the node health status values ​​in the twin structure, and treats this set as a set of potential fault sources. Relying on the state assessment model in step S230, the system constructs a fault propagation weighted mapping graph by evaluating the probability of node states propagating to neighboring nodes.

[0176] Under this diffusion model, the system performs multiple rounds of simulation. In each round of simulation, the system determines the probability of fault propagation at each node based on historical health trend parameters and dynamically updates the path status in the twin structure, i.e., adjusting the health value and risk level of all nodes on each connected path. The path status update logic comprehensively considers the current node risk level, historical health trend, and topology weight structure, where the topology structure is derived from the node graph structure relationship matrix in step S210.

[0177] Furthermore, during each round of simulation, the system records the fault propagation path chain and the state change sequence of each fault path, forming multiple sets of possible fault evolution paths. For each path set, the system extracts its corresponding path weight set, node health decline trend, node level mutation events, and other features to form a path state record. Finally, this step generates a set of feasible fault recovery path candidate schemes, providing a basis for scoring and optimization in subsequent steps.

[0178] S430. Perform path scoring and optimization on the fault recovery path scheme to generate the optimal recovery path structure.

[0179] After generating multiple sets of fault recovery path schemes, the system enters the path scoring and optimization stage. First, the system performs an overall score on the status of each node in each recovery path scheme. The scoring indicators include, but are not limited to, the following aspects: the number of highest-risk nodes in the path, the number of consecutive risk propagation events in the path, the path length, the frequency of policy intervention in the path, the estimated recovery time of the path, and the path controllability level.

[0180] Specifically, the system invokes the strategy scoring model library to configure and normalize the weights of the aforementioned scoring dimensions, calculating the comprehensive score for each path solution. Simultaneously, the system uses the anomaly probability information identified in step S310 and the risk level structure generated in S330 to assist in filtering path validity. For path solutions with scores falling within the lowest quantile range, the system performs a rejection operation, retaining only the highest-scoring candidate solutions for the next round of selection.

[0181] Within the candidate solution set, the system further introduces control execution constraint parameters (such as load migration reachability, backup loop activation time, control interface availability, etc.) to verify the feasibility of the remaining path solutions and determine the final feasible path set based on actual control conditions. Finally, the system selects the path solution with the highest comprehensive score that meets the feasibility constraints, generating the optimal recovery path structure.

[0182] Through the digital twin structure, fault propagation simulation, and path optimization scoring mechanism constructed in S410 to S430 of this invention, the system achieves the following technical breakthroughs: First, the system no longer relies on static rules or simple redundant logic, but instead simulates real operational evolution based on the twin, enabling dynamic deduction of fault recovery strategies under multi-path, multi-node, and multi-intervention conditions; second, the path scoring mechanism supports multi-objective function evaluation, achieving a comprehensive trade-off between recovery efficiency, execution complexity, and system stability; finally, the generation of the optimal recovery path structure not only possesses dynamic adaptability but also provides the control execution module with a clear instruction structure containing state basis and scoring reasons, greatly enhancing the intelligence and decision-making transparency of the fault self-healing process.

[0183] Step 500 includes at least steps S510-S530:

[0184] S510: Obtain the optimal recovery path structure, perform control command mapping processing, and generate a control instruction set.

[0185] Specifically, in this step, the system first calls the optimal recovery path structure generated in step S430. This path structure clearly indicates the optimal switching scheme, backup path activation strategy, and load migration path topology for each key node in the current distribution cabinet under fault conditions. The optimal recovery path structure not only includes the node mapping relationship for loop reconfiguration, but also the path scoring results generated under the combined influence of time series data, equipment health status, and risk level structure.

[0186] After obtaining the optimal recovery path structure, the system performs control command mapping processing. This mapping process first binds and identifies the target node identifiers in the path structure against the control template library. The control template library is a set of pre-set structured control mapping rules, each rule corresponding to a specific type of operation command, including but not limited to loop disconnection, backup power connection, load migration execution, power polarity adjustment, and bypass activation. Based on the node status identifiers, control targets, execution priorities, and system operating modes in the current path structure, the system generates a set of parameters for the corresponding operation instructions and organizes them into a control instruction set.

[0187] The control instruction set is a structured data collection containing ordered control sequences. Each control instruction unit consists of the following fields: operation object identifier, target state description, execution mode parameters, constraint condition judgment, and rollback strategy identifier. During the control command mapping process, the system automatically assigns and combines the above fields, generating an instruction set structure with real-time execution capabilities.

[0188] Furthermore, the system performs consistency checks on the initially generated control instruction set, including execution logic conflict detection, node concurrent access protection policy verification, priority sorting compliance confirmation, and path structure integrity verification. If control conflicts or policy conflicts are found during the detection process, the system will perform control instruction rollback, merging, or splitting operations according to built-in rules to ensure that the instruction set is executable and secure.

[0189] The processed control instruction set will be transmitted to the control interface module as the standard input parameter set for subsequent self-healing control actions, providing a precise control basis for the execution of step S520.

[0190] S520 calls the control interface of the power distribution cabinet according to the control instruction set to perform self-healing actions such as circuit switching and load transfer.

[0191] In this step, the system first receives the generated control instruction set from step S510 and interacts with the control execution unit in the physical power distribution cabinet through the control interface module. The control interface module is an integrated device control bus system, whose structure includes a protocol adapter unit, an interface parser, a status listener, and a control execution trigger.

[0192] Specifically, the system uses an interface parser to perform protocol-level translation of the control command set, ensuring strict matching of communication format, frame structure, check bits, command type, and other aspects with the actual control unit. The system supports multiple industry standard communication protocols, including but not limited to Modbus, CANopen, and IEC61850, and can load corresponding driver modules according to differences in control unit manufacturers.

[0193] After successful protocol matching, the system triggers the control execution trigger, sending control commands to the target control unit one by one. During execution, the system activates loop switching operations, load transfer commands, and backup power access control in priority order. The loop switching operation involves a series of steps, including powering off the faulty node, activating the backup path, and performing transition current pre-detection. The system will quickly calculate based on historical data of node current and voltage fluctuations in the health graph to confirm the switching window and precisely control the operation of relays or circuit breakers.

[0194] For load transfer operations, the system uses the current load data of high-risk nodes marked in the risk level structure as parameters input to the migration module. Through the load distribution model, the transfer target is rationally allocated to standby nodes with higher health levels. The load migration process involves operations such as power synchronization, minimizing wiring delays, and controlling the safe threshold of transfer current. All processes are completed in real time under the scheduling and monitoring of the control interface module.

[0195] During the above operations, the system's status listener continuously tracks feedback information from each control unit, including command response delay, execution status code, and electrical quantity changes. If the listener detects that a command execution has failed or the preset feedback conditions have not been met, the system will immediately trigger a rollback strategy and execute a predefined fault recovery path to avoid cascading anomalies.

[0196] By executing the above control commands one by one, the system achieves a series of actions such as loop switching, load transfer, and self-healing closure without manual intervention, and ensures the logical closed loop and physical safety of each action node during the execution process, thereby significantly shortening the system fault recovery time and improving the availability of the power distribution system.

[0197] S530. Associate the execution results of the self-healing action with the control instruction set and generate an action scheduling log structure.

[0198] After executing all control commands, the system enters the result recording stage described in this step. Specifically, the system first obtains the execution feedback information corresponding to each command from the control interface module, including but not limited to fields such as the actual execution timestamp, execution status code, voltage and current parameter changes, device response delay, and whether the rollback mechanism was triggered.

[0199] The system structurally associates the above feedback information with the original control command set to generate a complete action scheduling record unit. This record unit is built on the basis of a one-to-one correspondence between "control command → actual execution response" and simultaneously embeds path structure information, node health index snapshots, and corresponding numbers of fault propagation simulations to ensure that the entire scheduling record has traceability, state correlation, and path coherence.

[0200] The system further organizes multiple action scheduling record units into a scheduling log structure according to the control flow sequence. This structure adopts a two-dimensional nested mode of time series and node topology, and the recorded content includes the execution trigger source, control target node, operation type identifier, response feedback data, and abnormal event identifier code. The action scheduling log structure will serve as the basic data source for subsequent chain-style writing by the evidence storage module, and will support historical backtracking, control analysis, and responsibility attribution determination through unique identifiers.

[0201] To ensure the authenticity and integrity of the log structure data, the system introduces a log integrity verification mechanism, which employs strategies such as multi-point hash signature, instruction index backtracking, and chain connection verification to encrypt and verify the log content and implement version control, ensuring the legality and technical consistency of the subsequent S600 evidence storage process.

[0202] Through the implementation of steps S510 to S530, this invention achieves a seamless connection between the output path structure of the fault prediction model and the actual control execution. The generation of the control instruction set ensures high-precision matching and safety of control actions, the dynamic self-healing capability of loop switching and load transfer operations significantly improves the system's fault response efficiency, and the recording and encrypted management of the action scheduling log structure provides a complete data foundation and structural support for subsequent blockchain evidence storage, constituting the key central part of the "prediction-response-evidence storage" technology closed loop of this invention.

[0203] Step S600 includes at least steps S610-S630:

[0204] S610. Obtain the action scheduling log structure and construct a log data unit containing execution time, command content, and result status.

[0205] Specifically, before executing this step, the association record between the self-healing action execution result and the control instruction set in step S530 needs to be completed. The action scheduling log structure is a set of structured information generated in the aforementioned steps. This set records the generation source path, execution node identifier, type and function definition of the called interface, response timestamp, feedback status code, and final execution result for each control instruction. In this step, the system first performs field decomposition processing on the action scheduling log structure, extracting the execution time data, control command content, and operation result status from the structure as basic fields to construct log data units.

[0206] Furthermore, the system invokes the log parsing module to analyze the information of each control action execution record, re-encapsulates it according to the "execution time-command content-result status" triple format, and adds the action execution number, command initiation source identifier, and target device identifier to form a set of log data units with a complete and traceable chain. Each log data unit structurally satisfies the constraints of field integrity and uniqueness, ensuring no duplication or omission in the subsequent encryption hashing and chained evidence storage process.

[0207] Understandably, the execution time field is generated based on the synchronization timestamp mechanism in the control system, ensuring that all execution records are organized in a high-precision time chain; the command content field is encapsulated in the form of structured instruction parameters, facilitating subsequent traceability and auditing; the result status field originates from the action response content returned by the control interface in step S520, covering status identification information such as whether the execution was successful and whether the response delay meets the tolerance range. Through the above operations, the system constructs a set of log data units for standardized evidence storage, providing an input foundation for subsequent data encryption and evidence storage structure generation.

[0208] S620. Perform cryptographic hashing and signing operations on log data units to generate a standardized evidence storage structure.

[0209] The system first calls the hash digest generation module to perform a cryptographic hash calculation operation on each log data unit. The hash algorithm used is a collision-resistant cryptographic hash function; in this embodiment, the SHA-256 algorithm is preferred. The triple log field and auxiliary identification information are digested and compressed to generate a unique and irreversible hash fingerprint value. This hash fingerprint serves as a unique index identifier for subsequent block notarization, ensuring that even minor data changes can be identified.

[0210] After completing the hash calculation, the system further performs a signature operation on the hash fingerprint. Specifically, the system calls the platform key management module to obtain the digital signature key of the current operation and maintenance entity, and uses an asymmetric encryption algorithm (such as RSA or ECC) to encrypt the hash fingerprint to generate a signature structure. The signature structure can be used to provide identity authentication and data consistency verification capabilities in subsequent data verification or accountability audits.

[0211] The system encapsulates log data units, corresponding hash fingerprints, signature structures, and related auxiliary fields (such as node addresses and issuance times) into a standardized evidence storage structure. This standardized evidence storage structure has the following characteristics: field arrangement conforms to predefined specifications, supporting multi-dimensional field indexing and structured retrieval; field values ​​adopt a unified encoding system to ensure cross-platform compatibility; and the structure is complete and the links are clear, meeting the data integrity and compliance requirements of on-chain operations.

[0212] Understandably, the standardized evidence storage structure consists of five main parts: first, the log field, which records the core data of the action execution; second, the cryptographic hash field, which identifies the uniqueness of the data; third, the signature field, which provides legally valid proof of the operation; fourth, the index field, which locates the node's identity and the evidence storage time; and fifth, the extension field, which is used for subsequent version upgrades and cross-platform compatibility support. All of these fields have a clear structure and stable meaning, providing standardized input for the blockchain system to accept and store evidence.

[0213] S630: Write the standardized evidence storage structure into the blockchain network to generate chain-based operation and maintenance evidence storage data.

[0214] After completing the construction of the standardized evidence storage structure, the system enters the evidence storage writing stage. First, the system calls the linking module to connect with the blockchain network nodes deployed in the operation and maintenance management platform. The Hyperledger Fabric platform is preferred as the underlying support to ensure that the characteristics of the consortium blockchain support controlled access, hierarchical permissions, and traceable operations.

[0215] The system encapsulates each standardized evidence storage structure into transaction data and calls the chaincode interface to perform the evidence storage and writing operation. During this process, the system uses the transaction body as input parameters for the smart contract, calls the on-chain contract to execute storage instructions, and synchronously broadcasts the transaction request in the node network. Driven by the consensus mechanism, each consortium node verifies, reaches consensus on, and packages the transaction data to form the final block, which is then written into the chain ledger.

[0216] After the write operation is completed, the system records the transaction ID, block height, and on-chain storage time of the evidence storage transaction, and writes this information back to the log data index table for subsequent evidence verification, fault traceability, and operation and maintenance responsibility tracking. This forms a complete chain-based operation and maintenance evidence storage data structure, which includes the original data of the action execution process, a standardized structure, and blockchain write records for triple data redundancy, ensuring that the fault handling process is traceable, tamper-proof, and has legal evidentiary value.

[0217] After the evidence data is generated, the system also supports building a reconciliation mechanism and evidence audit interface based on contract calls, which facilitates cross-verification of the operation process by the operation and maintenance supervision entity, equipment manufacturers, and third-party certification bodies. The chained operation and maintenance evidence data not only improves the credibility and transparency of power distribution cabinet fault handling, but also serves as a trigger condition for subsequent smart contracts to participate in advanced functional modules such as automatic compensation, liability definition, and operation and maintenance performance.

[0218] By executing the processing steps S610 to S630, after the self-healing control action is completed, the system can generate verifiable and tamper-proof blockchain-based evidence data based on log data, achieving full-link traceability and closed-loop accountability in the operation and maintenance process of the power distribution cabinet. Compared with traditional log archiving methods, this implementation method has higher data integrity assurance capabilities and stronger audit traceability, significantly improving the intelligent management level and operation and maintenance compliance capabilities of the power distribution system. Especially in application scenarios facing multi-entity collaborative operation and maintenance and compliance supervision, the blockchain-based evidence storage mechanism has the technical advantages of being implementable, auditable, and verifiable.

[0219] The key innovations of this invention include:

[0220] (1) A weighted graph construction method integrating multi-physics field sensing data and equipment topology is proposed, and the edge weights are quantized based on the Shannon entropy function, thereby improving the accuracy of state structure modeling of power distribution equipment.

[0221] (2) Introducing a modeling mechanism that combines graph convolutional networks and complex analysis state kernel functions to form a node health score map, transforming the traditional discrete fault detection method into a differentiable and trainable continuous prediction model.

[0222] (3) A closed-loop “identification-recovery-recording” chain is constructed, integrating three major mechanisms: digital twin simulation, self-healing control instruction set generation and blockchain trusted operation and maintenance record, realizing a full-process intelligent control system from state perception to action execution and then to trusted evidence storage.

[0223] The following are its main beneficial effects:

[0224] (1) This invention achieves deep fusion of multi-physics information and graph structure modeling, significantly improving the accuracy and foresight of fault prediction. By normalizing, binding nodes, and mapping topological data from multiple physical fields such as temperature, electromagnetic fields, and vibration, this invention constructs a node-level time-series data structure and introduces a Shannon entropy-enhanced edge weight calculation mechanism to form a dynamically weighted graph structure relation matrix. Compared to the traditional method that relies solely on single-dimensional sensor threshold judgment, this method can capture the potential coupling effects and anomaly propagation trends between nodes, effectively improving the ability to identify fault precursors and enabling fault prediction to move from "static rules" to "structure learning," thereby increasing accuracy.

[0225] (2) A combined modeling method of graph convolutional networks and complex analysis kernel functions is introduced to construct a node health map, realizing quantitative assessment and hierarchical management of node status. This invention innovatively embeds a multi-layer graph convolutional computation process into the graph neural network, extracts cross-node correlation feature vectors, and calculates the state gradient response based on the complex Hessian kernel function to generate a node health score. The health score, as an interval normalized result, reflects the structural stability and potential failure risk of each node. This mechanism breaks the limitation of traditional state assessment methods relying on fixed rules, realizes continuous quantification of fault states, supports more refined risk identification and level classification, and provides a structural basis for subsequent digital twin simulation and recovery path selection.

[0226] (3) Construct a closed-loop system for self-healing control and trusted recording to achieve an automated operation and maintenance process of "fault identification - path optimization - control execution - blockchain evidence storage". Based on the health graph and risk level structure, this invention constructs a digital twin model that integrates state factors and risk structure, performs multiple path deductions and optimizations, and forms the optimal recovery path structure. Subsequently, the system automatically maps and generates a set of control instructions to drive the distribution cabinet to perform self-healing actions such as loop switching and load reconfiguration, and standardizes and records the action results as log data units, performs hash encryption and block signing, and finally writes them into the blockchain to form chain-based operation and maintenance evidence storage data. This mechanism ensures the verifiability and immutability of operation and maintenance execution, and effectively improves the automation level, transparency and compliance assurance capabilities of the power distribution system.

[0227] Example 2: Figure 2 A structural block diagram of a power distribution cabinet fault prediction and self-healing control system according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include:

[0228] The raw data processing module 10 is used to acquire multi-source sensor data deployed in the power distribution cabinet, and perform continuous time series construction and node binding operations on the data to generate a unified time series data structure that meets the requirements of graph modeling. The multi-source data includes raw physical parameter data collected by temperature sensors, electromagnetic sensors, and vibration sensors installed in different key parts of the power distribution cabinet. This module further constructs graph structure input data with integrity, continuity, and node label identification through time series segmentation, format standardization, missing value completion, edge filtering, normalization processing, and device node mapping operations, providing a unified semantic foundation for subsequent graph convolution analysis.

[0229] The graph structure modeling module 20 is used to fuse the time-series data structure with the equipment topology of the distribution cabinet, construct a graph structure relationship matrix, and perform multi-layer graph signal feature extraction based on the Graph Convolutional Network (GCN) model to finally generate a health map of the distribution cabinet nodes. This module first constructs a node adjacency matrix based on the electrical connections, physical deployment structure, and control dependency paths between power distribution components. Then, it inputs this adjacency structure along with the constructed time-series data into the GCN model, performs multi-layer convolutional abstraction operations, and aggregates spatial correlations and temporal evolution trends layer by layer, outputting node-level graph representation vectors. Finally, through a state mapping function, the high-dimensional graph feature vectors are converted into standardized health scores to form an equipment node health map for subsequent risk identification and control decisions.

[0230] The risk identification module 30 is used to perform fault probability analysis and cluster assessment on key nodes in the distribution cabinet based on the node health map, constructing a risk level structure with hierarchical risk identification capabilities. This module first calculates the anomaly probability value of each node, then clusters the anomaly features using probability thresholding and hierarchical clustering algorithms to clarify different risk ranges. Subsequently, the system combines node type, topological location, and anomaly propagation path to assign structured labels to nodes of different risk levels, forming a risk level structure that provides quantitative risk input for the digital twin-driven recovery simulation process.

[0231] The digital twin simulation module 40 is used to construct a virtual mapping model, i.e., a digital twin of the power distribution cabinet, based on the node health map and risk level structure. It then performs diffusion simulation, path scoring, and recovery strategy optimization based on state parameters and fault models, outputting the optimal recovery path structure. This module first maps the current state parameters (health value, anomaly level, connection topology) to the digital twin model structure, simulating fault point propagation, load migration, and control command issuance scenarios, dynamically generating multiple recovery path combinations. Next, it compares the merits of each scheme through a multi-index scoring mechanism (such as path complexity, impact range, and recovery latency), ultimately outputting the recovery path result with the best performance as the basis for control command generation.

[0232] The control execution module 50 is used to generate a control instruction set based on the optimal recovery path structure, and execute self-healing operations such as loop switching and load transfer. Simultaneously, it records the control path, executed actions, and response results to form an action scheduling log structure. This module uses an instruction mapping mechanism to convert each recovery path into an executable combination of low-level instructions, which are then sent to the distribution cabinet controller via a driver interface. During instruction execution, the system collects action response information in real time and encapsulates all instruction content, target object, execution time, execution status, and other fields into log entries, constructing an action scheduling log structure to provide a complete event chain for subsequent chain-based evidence storage.

[0233] The Operation and Maintenance Evidence Preservation Module 60 is used to generate standardized evidence preservation data based on the action scheduling log structure and write it into the blockchain network, forming chain-based operation and maintenance evidence preservation data with tamper-proof characteristics. This module first performs cryptographic hashing on the log data to generate a data digest fingerprint; then it calls the platform key generation module to complete the log signature structure construction; finally, the system encapsulates the log fields, hash digest, and signature result together into a standardized block transaction body, which is broadcast onto the chain through consortium blockchain nodes and written into the distributed ledger system. After this process is completed, the system writes the block index information back to the log index table, supporting subsequent accountability and audit query operations.

[0234] The beneficial effects of this embodiment:

[0235] (1) Significantly improve fault prediction capability: By introducing a graph convolutional network structure and integrating multi-physics sensor data to construct a device health map, potential hidden faults can be identified earlier than traditional alarm mechanisms based on static thresholds, thus improving early warning efficiency. Improve node-level fault prediction and realize the transformation of equipment operation and maintenance from "post-event response" to "proactive prediction" mode.

[0236] (2) Significantly reduce recovery response time: By establishing a digital twin structure and performing multi-path recovery scheme simulation in virtual space, the system of the present invention can generate and issue the optimal recovery command at the moment the fault occurs, shortening the power distribution recovery time in typical scenarios and greatly reducing power outage losses and service interruptions caused by faults.

[0237] (3) Enhance data credibility and compliance assurance capabilities: The system has a built-in blockchain module that automatically generates chain-based operation and maintenance evidence data after the self-healing action is completed. Each control path, execution instruction and recovery result can be recorded on the chain, realizing the immutability and traceability of responsibility throughout the entire operation and maintenance process, and meeting the requirements of government and enterprise units for the compliance and auditability of power operation and maintenance data.

[0238] (4) Supporting multi-party collaboration and regulatory audit: By constructing a standardized evidence storage structure and deploying it on the consortium blockchain platform, the system supports access and cross-verification by multiple parties such as government and enterprise regulators, operation and maintenance units, and equipment manufacturers, providing a structured and legal data foundation for subsequent power service quality supervision, service level assessment and commercial compliance audit.

[0239] (5) It has scalable deployment capability: This system is based on modular design, adapts to multiple types of power distribution cabinet structures, and supports multi-level topology automatic learning capability. It has good scalability and cross-scenario reuse capability, which makes it easy to promote and apply in various power supply and distribution environments such as municipal infrastructure, data centers, and industrial parks.

[0240] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A method for fault prediction and self-healing control of a power distribution cabinet, characterized in that, Includes the following steps: Acquire raw data, construct continuous time series data structures and configure node bindings to generate time series data structures for graph construction; The time series data structure is combined with the device topology to construct a graph structure, and graph convolution processing is performed to generate a node health graph. The expression for the graph structure is: in, The diagram shows the structural relationships; A set of nodes; It is an edge set; Let the edge weight matrix be the edge weight matrix. The expression for graph convolution processing is: in, Indicates the final embedded features; This represents the vector in the i-th row of the feature matrix of the L-th layer of the graph convolutional network; This represents the total number of layers in a graph convolutional neural network (GCN) model. Representing vectors Belongs to the dimension The real number space; This represents the number of dimensions of the output features of the Lth layer of the graph convolutional network; For the final embedded features Execute the second-order complex Hessian kernel function Laplace gradient operation: in, For the first Health score of each node; The second-order derivative tensor of the embedded features after complex analysis kernel function; It is the Frobenius norm; The expression for the node health graph is: in, Represents the node health graph; Indicates the first Each device data acquisition unit; This indicates the total number of sensor nodes deployed in the device; Based on the node health graph, the probability of anomalies is calculated, clustering and structural coding are performed, and a risk level structure is generated. A digital twin structure is constructed using the risk level structure and node health graph, and then deduced and optimized to generate the optimal recovery path structure. Generate a set of control instructions based on the optimal recovery path structure, execute self-healing actions and associated records, and generate an action scheduling log structure. Obtain the action scheduling log structure, generate a standardized evidence storage structure and write it to the blockchain to form chain-based operation and maintenance evidence storage data.

2. The method for fault prediction and self-healing control of power distribution cabinets according to claim 1, characterized in that, The process of acquiring raw data, constructing a continuous time series data structure and configuring node bindings to generate a time series data structure for graph construction includes the following steps: The raw data collected by temperature sensors, electromagnetic sensors and vibration sensors deployed in the power distribution cabinet are obtained, and time-series segmentation and format standardization are performed to obtain preliminary time-series data. Missing values ​​were imputed and marginal filtering was performed on the initial time series data to construct a continuous time series data structure; The continuous time series data structure is normalized and configured with node binding to generate a time series data structure for graph construction.

3. The method for fault prediction and self-healing control of power distribution cabinets according to claim 1, characterized in that, The steps of constructing a graph structure by combining time-series data structures with device topology, performing graph convolution processing, and generating a graph structure in the node health graph include: Obtain the time-series data structure and construct a graph structure relationship matrix between nodes by combining the topological relationships of device nodes.

4. The method for fault prediction and self-healing control of power distribution cabinets according to claim 1, characterized in that, The graph convolution process includes the following steps: The graph structure relation matrix and time series data structure are input into the graph convolutional network model for training, generating multi-layer graph convolutional feature representations.

5. The method for fault prediction and self-healing control of power distribution cabinets according to claim 1, characterized in that, The process of generating the node health map includes the following steps: The state evaluation and health mapping processes are performed on the convolutional feature representation of multi-layer graphs to generate a node health map.

6. The method for fault prediction and self-healing control of power distribution cabinets according to claim 1, characterized in that, The process of generating the risk level structure includes the following steps: Obtain the node health graph, calculate the probability of failure for each node in the graph, and extract the anomaly probability information; Clustering and segmenting of anomaly probability information is performed to identify different risk level intervals; The risk level ranges are labeled and coded to generate a risk level structure.

7. The method for fault prediction and self-healing control of power distribution cabinets according to claim 1, characterized in that, The process of generating the optimal recovery path structure includes the following steps: Obtain the risk level structure and node health graph, and construct a digital twin structure containing operational status parameters; Perform fault propagation simulation and path state update on the digital twin structure, and deduce multiple fault recovery path schemes; The fault recovery path schemes are scored and optimized to generate the optimal recovery path structure.

8. The method for fault prediction and self-healing control of power distribution cabinets according to claim 1, characterized in that, The process of generating the action scheduling log structure includes the following steps: Obtain the optimal recovery path structure, perform control command mapping processing, and generate a control instruction set; The control cabinet control interface is invoked according to the control instruction set to perform self-healing actions such as circuit switching and load transfer; The execution results of the self-healing action are associated with the control instruction set and recorded to generate an action scheduling log structure.

9. The method for fault prediction and self-healing control of power distribution cabinets according to claim 1, characterized in that, The generation of chained operation and maintenance evidence storage data includes the following steps: Obtain the action scheduling log structure and construct a log data unit containing execution time, command content, and result status; The log data units are encrypted, hashed, and signed to generate a standardized evidence storage structure. A standardized evidence storage structure is written into the blockchain network to generate chain-based operation and maintenance evidence storage data.

10. A fault prediction and self-healing control system for a power distribution cabinet, applied to the method described in any one of claims 1-9, characterized in that, include: The raw data processing module is used to acquire raw data, construct continuous time series data structures and configure node binding, and generate time series data structures for graph construction. The graph structure modeling module is used to combine time series data structures with device topology to construct graph structures, perform graph convolution processing, and generate node health graphs. The risk identification module is used to calculate the probability of anomalies based on the node health graph, perform clustering and structural coding, and generate a risk level structure. The digital twin simulation module is used to construct a digital twin structure using the risk level structure and node health graph, and to perform simulation and optimization to generate the optimal recovery path structure. The control execution module is used to generate a set of control instructions based on the optimal recovery path structure, execute self-healing actions and associated records, and generate an action scheduling log structure. The operation and maintenance evidence storage module is used to obtain the action scheduling log structure, generate a standardized evidence storage structure and write it into the blockchain to form chain-like operation and maintenance evidence storage data.

Images