A heavy truck charging health degree evaluation method based on an optimized graph convolution network

Abstract

The application discloses a heavy truck charging health degree evaluation method based on an optimized graph convolution network and particularly relates to the technical field of electric power operation and maintenance; through collecting charging pile multi-source heterogeneous operation data and performing unified time window segmentation, a dynamic multiple graph structure fusing spatial, electrical and behavior coupling relationships is constructed; a node local state evolution trend and a neighbor influence factor are extracted to generate a joint feature vector; the features under multiple time windows are input into a multi-scale attention enhanced graph convolution network to obtain a joint latent health degree representation; a health score is calculated based on the latent representation, a high-risk node is identified, and a latent failure propagation path atlas is constructed; a maintenance priority table is generated by combining a centrality index and a degradation rate to output an optimized inspection path; the application realizes spatiotemporal correlation modeling and intelligent maintenance path planning of the charging pile health state and has the advantages of high accuracy and strong applicability.

Description

Technical Field

[0001] This invention relates to the field of power operation and maintenance technology, specifically to a method for evaluating the charging health of heavy-duty trucks based on optimized graph convolutional networks. Background Technology

[0002] With the rapid popularization of new energy vehicles, heavy-duty truck charging piles have been deployed extensively in cities and along highways as supporting infrastructure. However, due to factors such as long-term operation, electrical aging, environmental corrosion, and uneven use, the health of heavy-duty truck charging piles is showing a deterioration trend, leading to problems such as increased charging failure rates, increased grid impact, and decreased user experience. Existing health assessment methods mostly rely on manual inspections or statistical analysis based on single sensor data, lacking in-depth analysis of the complex relationships between the internal structural features of charging piles, making it difficult to accurately identify early signs of health degradation.

[0003] Especially in urban secondary arterial roads and high-frequency usage areas, due to the high density of heavy-duty truck charging piles and their highly heterogeneous operating states, there are complex spatial and topological relationships between charging piles. Their health status often does not evolve in isolation but exhibits graph structure characteristics influenced by the coupling of neighboring devices. Existing methods ignore this spatial graph structure, resulting in evaluation results that cannot accurately reflect the true distribution of device health status in the overall area, making it difficult to implement preventive maintenance deployment and posing potential safety risks. Summary of the Invention

[0004] The purpose of this invention is to provide a method for evaluating the charging health of heavy-duty trucks based on optimized graph convolutional networks, in order to address the shortcomings in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for evaluating the charging health of heavy-duty trucks based on an optimized graph convolutional network, comprising:

[0006] S100: Acquire multi-source heterogeneous data of all charging piles in the target area, including time-series voltage and current data, internal temperature change trajectory of the module, communication delay, user operation frequency and geographical location information, and perform unified time window segmentation on the data.

[0007] S200: Based on a unified segmented time window, the coupling behavior between charging piles is modeled, and a dynamic multi-graph structure is constructed, in which each subgraph corresponds to a different type of coupling relationship, and the temporal order of the edges in the graph is encoded by a sliding window method.

[0008] S300. For each subgraph in the dynamic multigraph structure, the local state evolution trend of the nodes in the graph and the influence factor of the neighbor graph are fused to obtain the joint feature vector Fi of the nodes.

[0009] S400. Concatenate the joint feature vector Fi according to the time window order to form a sequence feature tensor T, and input it into a multi-scale attention-enhanced graph convolutional network for feature encoding to obtain the joint potential health representation Z.

[0010] S500, calculates the predicted health score H for each charging pile based on Z;

[0011] S600: Based on the predicted health score H, it automatically identifies nodes that exhibit structural vulnerability in multiple heterogeneous graphs and, combined with their topological clustering in space, constructs a potential failure propagation path map P for charging piles.

[0012] S700: Based on the centrality index and degradation rate in the failure propagation path graph P, output a priority-ranked table of high-risk maintenance for charging piles, and generate an optimized inspection path by combining the maintenance radius and resource availability.

[0013] Preferably, the modeling of the coupling behavior between charging piles and the construction of a dynamic multigraph structure includes:

[0014] Based on the unified segmented time window data, three types of initial coupling relationship matrices are constructed according to the spatial distance between charging piles, the correlation of electrical parameters, and the coordinated changes in operating status.

[0015] Generate corresponding subgraphs based on each type of coupling relationship matrix, and construct a graph set containing multiple heterogeneous graphs, where the edges of each subgraph represent a type of coupling.

[0016] For the edge relationships in each subgraph, the evolution trend of edge weights at different times is embedded with timestamps, and the dynamic change characteristics of the edges are preserved.

[0017] The dynamic change characteristics of edges are jointly represented with the characteristics of nodes.

[0018] Preferably, for each subgraph in the dynamic multigraph structure, the local state evolution trend of nodes in the graph and the influence factors of neighboring graphs are integrated, including:

[0019] For each subgraph in the dynamic multigraph structure, the state change sequence of each node under a continuous time window is extracted, and the local state evolution trend vector is extracted using a gated cyclic unit.

[0020] Based on the neighbor set of each node, calculate the weighted average feature representation of its neighbor nodes in the current subgraph;

[0021] The local state evolution trend vector of a node is concatenated with the neighbor graph influence factor, and then nonlinearly fused through a fully connected neural network to obtain the joint feature vector Fi of the node.

[0022] Perform a fusion operation on each type of subgraph to obtain joint feature representations of nodes under multiple coupling relationships.

[0023] Preferably, the step of concatenating the joint feature vector Fi into a sequence feature tensor T according to the time window order, and inputting it into a multi-scale attention-enhanced graph convolutional network for feature encoding includes:

[0024] Arrange the joint feature vectors of nodes under each time window in chronological order to construct a three-dimensional sequence feature tensor, where the first dimension of the tensor represents the time window index, the second dimension represents the node number, and the third dimension represents the node feature dimension.

[0025] Graph convolution operations are performed on sequence feature tensors, and a multi-scale graph convolution structure composed of multiple convolution scales is used to capture local and global topological relationship features between nodes with different adjacency ranges.

[0026] The features of different time windows are weighted and fused, and the attention weight is calculated by the magnitude of node state change and the consistency of adjacency structure.

[0027] The fused features are transformed by an activation function and output as a joint potential health representation Z for each node.

[0028] Preferably, the step of calculating the predicted health score H for each charging pile based on Z includes:

[0029] Perform feature normalization on the joint potential health representation of each node obtained;

[0030] The normalized features are input into a fully connected regression neural network with residual connection structure. The neural network consists of two linear transformation layers and one activation function, and the output is a continuous numerical score.

[0031] During the model training phase, mean squared error is used as the loss function to optimize the deviation between the network output score and the historical actual health label through backpropagation.

[0032] Based on the predicted health score H output by the model and the set state classification threshold, charging piles are classified into healthy, sub-healthy, and abnormal states.

[0033] Preferably, the automatic identification of nodes exhibiting structural vulnerability in multiple heterogeneous graphs based on the predicted health score H includes:

[0034] Based on the obtained predicted health score H, structural vulnerability analysis is performed on nodes in multiple heterogeneous graphs. The vulnerability index is jointly calculated using the sparsity of edge weights between nodes, the lower bound of the health score, and the characteristic volatility. Nodes with vulnerability indices exceeding a preset stability threshold are then selected.

[0035] Extract the adjacency structure of nodes identified as structurally fragile in the spatial coupling subgraph and calculate their topological clustering coefficient in the local region;

[0036] Based on the health score difference between nodes, edge weight propagation probability, and spatial distance, a weighted directed graph is constructed as a potential failure propagation path map, where the direction of the edge represents the direction of degradation propagation and the weight represents the propagation intensity.

[0037] Perform a depth-first traversal of the path graph to identify path chains that trigger cascading degradation.

[0038] Preferably, a weighted directed graph is constructed as a potential failure propagation path map based on the health score difference between nodes, edge weight propagation probability, and spatial distance, including:

[0039] Obtain the health score difference between all node pairs in multiple heterogeneous graphs, and calculate the score decrease between any two adjacent nodes as the initial driving factor for failure propagation.

[0040] The graph structure edge weights between node pairs are normalized, and the propagation probability of the edges is calculated by combining the dynamic change frequency of the edges.

[0041] The distance attenuation factor is calculated based on the geographic spatial distance between node pairs, and the spatial distance is weighted and compressed using a Gaussian function.

[0042] The weighted fusion of score decline rate, edge weight propagation probability and distance decay factor generates the final weight of the edge, and a weighted directed graph is constructed as a potential failure propagation path graph, in which the direction of the edge is from high health node to low health node.

[0043] Preferably, the step of generating an optimized inspection path by combining the maintenance radius and resource availability includes:

[0044] Based on the constructed potential failure propagation path map, the propagation path centrality index of each node is calculated. The propagation path centrality index is obtained by weighted statistics of the frequency of occurrence of the node in all failure paths and the path length.

[0045] Calculate the degradation rate of each node, which is the ratio of the decrease in its health score within a continuous time window to the time interval;

[0046] A priority table for the maintenance of high-risk charging piles is generated by weighting and sorting nodes based on their path centrality index and degradation rate.

[0047] By combining the geographical location of each high-risk node, the maximum service radius of maintenance personnel, and the constraints of currently available resources, the Traveling Salesman Path (TSP) optimization algorithm is used to calculate the shortest inspection path and output a list of optimal inspection sequences.

[0048] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0049] 1. This invention introduces unified temporal modeling of multi-source heterogeneous data to construct a dynamic multi-graph structure, and combines multi-scale graph convolution and temporal attention mechanisms to achieve accurate modeling and trend prediction of charging pile health status. Compared with existing health assessment methods based on static features or single models, this invention can effectively capture the spatiotemporal dependencies and evolutionary patterns of equipment operating status, thereby significantly improving prediction accuracy and early degradation identification capabilities.

[0050] 2. Based on the prediction results, this invention further constructs a potential failure propagation path map and generates a high-risk maintenance priority ranking by combining centrality indicators and degradation rates. By integrating spatial geographical constraints and resource scheduling conditions, it generates the optimal inspection path, realizing an intelligent closed loop from data modeling to maintenance decision-making. This solution not only reduces the risk of systemic failures caused by charging pile degradation but also significantly improves the efficiency and accuracy of operation and maintenance resource allocation, demonstrating promising engineering application prospects. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0052] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] For examples, please refer to Figure 1 As shown in this embodiment, a method for evaluating the charging health of heavy-duty trucks based on optimized graph convolutional networks includes:

[0055] S100: Acquire multi-source heterogeneous data of all charging piles in the target area, including time-series voltage and current data, internal temperature change trajectory of the module, communication delay, user operation frequency and geographical location information, and perform unified time window segmentation on the data.

[0056] The core of step S100 lies in comprehensively collecting multi-source heterogeneous data covering the operating status of charging piles, so as to provide a foundation for subsequent graph structure modeling and graph neural network feature learning. The multi-source heterogeneous data specifically includes:

[0057] Time-series voltage and current data: Collect the voltage and current variation curves of the charging pile in various time periods, which can reflect its power output stability and circuit health status;

[0058] Internal temperature change trajectory of the module: Collect the temperature change trajectory of key components inside the charging pile module to monitor potential problems such as overheating and abnormal heat dissipation;

[0059] Communication delay: Record the communication response delay between the charging pile and the platform, and analyze potential control delays or communication interruption issues;

[0060] User operation frequency: Statistics on the average number of times users start charging per day, operation frequency, and abnormal termination records are used to identify usage frequency and user experience-related characteristics;

[0061] Geographic location information: Collect the latitude and longitude coordinates and installation location (such as underground / ground, wall-mounted / independent bracket, etc.) of each charging pile for calculating spatial adjacency relationships in subsequent graph modeling.

[0062] Considering the issues of inconsistent sampling frequencies and misaligned timestamps among the aforementioned data sources, a unified sliding time window mechanism is adopted to align the time of all data. The specific method is as follows:

[0063] Set a uniform time window length Δt, such as 5 minutes or 10 minutes, to ensure that all sensor data are aggregated into a single vector within this time period;

[0064] For continuous data (such as voltage, current, and temperature), statistical features are extracted using methods such as mean, standard deviation, and extreme values.

[0065] For discrete data (such as operation records, number of communication delay anomalies), perform counting or binarization;

[0066] All feature vectors are stacked sequentially according to the time window to form the state sequence of each charging pile under multiple consecutive time windows.

[0067] S200: Based on a unified time window segmentation, the coupling behavior between charging piles is modeled, and a dynamic multigraph structure is constructed, in which each subgraph corresponds to a different type of coupling relationship, and the temporal sequence of the edges in the graph is encoded by a sliding window method.

[0068] First, based on the charging pile operation data within each time window, the following three types of coupling relationship indicators are calculated to construct three types of initial coupling relationship matrices:

[0069] Spatial distance coupling matrix: Based on the latitude and longitude coordinates of charging piles, the geographical distance between any two charging piles is calculated using the spherical distance formula. When the distance is less than a set threshold D (e.g., 500 meters), they are defined as having a spatial adjacency relationship, and the corresponding element in the matrix is ​​assigned a value of 1; otherwise, it is 0.

[0070] Electrical parameter correlation matrix: Calculate the Pearson correlation coefficient of voltage, current and other parameters for each pair of charging piles within the current time window. If the absolute value of the correlation coefficient exceeds the threshold R (e.g., 0.85), it is considered that there is an electrical behavior coupling relationship, and the matrix elements are assigned the value of the correlation coefficient.

[0071] Operational status co-evolution matrix: Calculate the usage frequency (e.g., number of charging times), fault records, and user operation behavior trends of each pair of charging piles within the time window. Use dynamic time warping algorithm to match their state sequences. If the similarity is higher than the threshold S (e.g., 0.9), it is marked as having a co-evolutionary relationship.

[0072] The three types of matrices mentioned above constitute the initial structural basis for spatial coupling, electrical coupling, and behavioral coupling, respectively. They are all symmetric matrices and are subsequently used to generate the corresponding subgraphs.

[0073] The three types of coupling relationship matrices are mapped sequentially to three independent subgraphs. Each graph has a charging pile as a node. Non-zero elements in the matrix indicate that there are edges connecting the nodes, and the edges are assigned specific numerical weights.

[0074] The three graphs are named the spatial coupling subgraph, the electrical coupling subgraph, and the cooperative behavior subgraph, respectively, forming a heterogeneous graph set. This set preserves the topological differences between charging piles in different dimensions, supports multi-channel graph neural network input, and improves the expressive power of the graph model.

[0075] To capture the changing trends of edge weights across different time windows, a timestamp embedding operation is introduced in each subgraph. The specific steps are as follows:

[0076] Within each time window, the weight values ​​of all edges are recorded to form a continuous temporal edge weight vector; the temporal weight change trend of each edge is differentially processed, and its first derivative is extracted to represent the weight change rate; the original weight sequence and the change rate sequence of each edge are concatenated to form the dynamic representation vector of that edge; this vector is encoded by a one-dimensional convolutional neural network to obtain the embedding features of the edge in the current time window.

[0077] The generated dynamic edge embedding features are jointly represented with the node's state feature vector. Specifically, for each node, its local operating features and the dynamic embedding features of all its adjacent edges are integrated and fused by weighted summation. The weighting factor is normalized according to the edge weight. The resulting joint representation vector is used as the input feature of the graph neural network model and input into the graph convolutional neural network in the next step for spatial-temporal health learning and prediction.

[0078] S300. For each subgraph in the dynamic multigraph structure, the local state evolution trend of the nodes in the graph and the influence factor of the neighboring graph are fused to obtain the joint feature vector Fi of the nodes.

[0079] First, for all nodes in each subgraph, based on a uniformly segmented time window, their operational state vectors within multiple consecutive time segments are extracted. These state vectors include data such as the node's mean voltage, standard deviation of current, extreme temperature values, operational frequency statistics, and mean communication delay. The state vectors from multiple consecutive time windows are then concatenated to form the node's time-series input sequence.

[0080] A gated recurrent unit (GRU) is introduced to model the time-series input sequence, capturing the dynamic trend of node state evolution over time. The GRU is a recurrent neural network with an update gate and a reset gate, controlling the retention of historical states and the updating of the current state. Network weights are optimized end-to-end using training data, and the output is a fixed-length local state evolution trend vector with a dimension of 64.

[0081] In the current subgraph, for any node, traverse its set of adjacent nodes. Based on the dynamic edge embedding features generated in the aforementioned dynamic edge modeling steps, extract the weight of each adjacent edge in the current time window and perform normalization processing. The normalization method uses the Softmax function to map the weights of all adjacent edges so that their sum is 1.

[0082] By using normalized edge weights as weighting coefficients, a weighted average is calculated on the local state evolution trend vectors of all neighboring nodes to obtain the neighbor graph influence factor of the target node. This representation can reflect the potential influence of the neighboring node group on the state evolution of the target node, which is beneficial for modeling the dependencies between local structures.

[0083] The local state evolution trend vector of a node is concatenated with the influence factors of its neighboring graphs according to their dimensions to form an intermediate fused feature vector. This vector serves as the input to a fully connected neural network and is processed through one or more linear transformations and nonlinear activation functions (such as Rectified Linear Units) to generate the final joint feature vector of the nodes.

[0084] The first layer of the fully connected neural network has an input dimension of 128 and an output dimension of 64. The weights are initialized using Xavier uniform initialization, and the network is optimized through error backpropagation during training.

[0085] For each type of subgraph in the dynamic multigraph structure (such as spatially coupled subgraph, electrically coupled subgraph, and cooperative behavior subgraph), the above three steps are performed to generate joint feature representations of nodes under various coupling relationships, named the first joint feature vector, the second joint feature vector, and the third joint feature vector.

[0086] S400. The joint feature vector Fi is concatenated into a sequence feature tensor T according to the time window order, and then input into a multi-scale attention-enhanced graph convolutional network for feature encoding to obtain the joint potential health representation Z.

[0087] The joint feature vectors of nodes obtained in each time window are arranged in chronological order to construct a set of three-dimensional sequence feature tensors. The first dimension of the tensor represents the time window index, the second dimension represents the charging pile node number, and the third dimension represents the node feature dimension.

[0088] If we set the total number of time windows to N, the total number of nodes to M, and the node feature dimension to D, then the final sequence feature tensor will have a dimension of N×M×D. This tensor is used to preserve the joint distribution structure of node features in both time and space dimensions, providing input for subsequent graph convolution operations.

[0089] For each time slice of the aforementioned 3D sequence feature tensor, a graph convolution operation is performed to capture the graph topological dependencies between nodes. A multi-scale graph convolution structure is employed, with different receptive field sizes (i.e., the extent to which adjacent nodes are extended) set in each convolutional layer.

[0090] For example, the first graph convolutional layer only considers first-order adjacent node features, the second layer expands to second-order adjacent nodes, and the third layer expands to third-order adjacent nodes, and the outputs of different layers are concatenated. Each graph convolutional layer uses Chebyshev polynomial expansion for graph frequency domain filtering, with a filter order of 3, and the weight parameters are automatically optimized through end-to-end training.

[0091] To integrate feature differences across different time windows, a temporal attention mechanism is introduced to weight the node features at each time step. The attention weights are jointly determined by the following two factors:

[0092] Node state change amplitude: Calculate the feature difference between the node in the current time window and the previous time window, and find its norm to reflect the degree of state fluctuation;

[0093] Adjacency structure consistency: Calculates the structural similarity of adjacent edges in the current time window, the degree of structural preservation compared to the previous time window, and reflects topological stability.

[0094] The attention weights for each time window are obtained by normalizing and then summing the two factors separately. The feature vectors of all time windows are then summed after weighting to form the fused time series representation.

[0095] The fused node features are input into a nonlinear transformation unit, and a Rectified Linear Unit is used as the activation function to perform an activation transformation on the weighted sum features. The output is the joint potential health representation Z of the nodes. The dimension of this potential representation can be set according to task requirements, such as 64 or 128 dimensions, for subsequent health scoring and risk classification.

[0096] S500 calculates the predicted health score H for each charging station based on Z.

[0097] First, the joint potential health representation vector of each node is normalized to eliminate scale differences between different feature dimensions and ensure the numerical stability of the model input. The normalization method adopts the minimum-maximum normalization approach: subtract the minimum value of each feature dimension in the training data and divide by the difference between the maximum and minimum values, so that all input feature values ​​are mapped to a continuous interval between 0 and 1.

[0098] The normalized vector retains the same dimensional structure as the original latent representation and serves as the input to the neural network scoring module.

[0099] The normalized joint feature vector is input into a fully connected regression neural network with residual connections. This network is used to learn the mapping relationship between the input features and the target health score. The neural network includes the following structure:

[0100] First linear transformation layer: The input dimension is the same as the joint feature vector, the output dimension is set to 64, and the weight matrix is ​​initialized using the Xavier method;

[0101] Activation function layer: The output of the first layer is nonlinearly transformed using a rectified linear unit function;

[0102] The second linear transformation layer has an output dimension of 1 and is used to generate a single continuous numerical health score.

[0103] Residual connection: The original input vector is added directly to the output of the second linear layer through a skip connection, which is used to alleviate gradient vanishing and enhance feature preservation.

[0104] During the model training phase, manually labeled health scores from historical charging pile operation data are selected as the supervised learning objective, and the mean square error between the predicted output score and the true label is used as the loss function for optimization.

[0105] The mean squared error is calculated as follows: for each node, the difference between its predicted score and actual label is squared, and the average is calculated over all node samples. The model uses the backpropagation algorithm to update all trainable parameters in the neural network, with the optimization objective being to make the loss function converge simultaneously on both the training and validation sets.

[0106] Early stopping strategy and learning rate decay mechanism are adopted during training to improve the model's generalization ability.

[0107] The predicted health score H for each node output by the model is a continuous real number, defined as ranging from 0 to 1. To achieve interpretable management of charging pile status, two status classification thresholds are set: the first threshold is set to 0.35, and the second threshold is set to 0.7.

[0108] The health status is classified based on the relationship between the predicted health score H and the aforementioned thresholds: a predicted health score H greater than or equal to 0.7 is considered a healthy state; a predicted health score H greater than or equal to 0.35 and less than 0.7 is considered a sub-healthy state; and a predicted health score H less than 0.35 is considered an abnormal state. This classification result is used to support subsequent risk warnings, inspection scheduling, and equipment replacement decisions.

[0109] S600 automatically identifies nodes exhibiting structural vulnerability in multiple heterogeneous graphs based on the predicted health score H, and constructs a potential failure propagation path map P of charging piles by combining their topological clustering in space.

[0110] First, structural vulnerability analysis is performed on all nodes across multiple heterogeneous graphs (including spatial coupling graphs, electrical coupling graphs, and cooperative behavior graphs). For each node, the following three factors are calculated to comprehensively assess its vulnerability:

[0111] Sparsity of edge weights between nodes: Calculate the edge weight density between the current node and its first-order adjacent nodes. If the edge weights are generally low or unevenly distributed, it indicates that the connection stability is poor.

[0112] Lower bound of health score: Select the minimum value in the predicted health score sequence to reflect the lowest operating state of the node in the time evolution;

[0113] Characteristic volatility: Calculates the standard deviation of the characteristic vector of a node in each time window, used to measure the degree of fluctuation in its operating status.

[0114] The three indicators mentioned above are used to synthesize a vulnerability index for a node through linear weighting, with weights set to 0.3, 0.4, and 0.3, respectively. If the vulnerability index of a node is greater than a preset stability threshold (e.g., 0.7), it is determined to be a structurally vulnerable node and included in the subsequent failure path analysis.

[0115] For identified structurally vulnerable nodes, their adjacency structures are extracted from the spatially coupled subgraph, and a local subgraph centered on that node is constructed. The topological clustering coefficient of this local subgraph is calculated, which is the ratio of the actual number of connected edges existing at the current node to the theoretical maximum number of edges between all its neighboring nodes.

[0116] The clustering coefficient is used to reflect the density of the network in the area where a node is located. The higher the coefficient, the denser the local structure. If a vulnerable node is in a highly clustered area, its failure may be propagated to neighboring devices through multiple paths, amplifying the risk of degradation propagation.

[0117] For all structurally vulnerable nodes and their adjacent nodes, a weighted directed graph is constructed by combining health scores, structural edge weights, and geographical distance information. This involves the following sub-operations:

[0118] Health score difference: Calculate the health score difference between any pair of adjacent nodes, retain only the direction from the node with the higher score to the node with the lower score, and take the difference as the initial driving factor for failure propagation;

[0119] Transmission probability estimation: Based on the historical dynamic change frequency of edges in the aforementioned heterogeneous graph structure, calculate the proportion of non-zero states that occur in multiple time windows, and use this as the transmission probability of the edge.

[0120] Distance decay factor: Gaussian function is applied to compress the straight-line spatial distance between nodes. The function is in the form of an exponential function, and the exponent is the square of the distance divided by the square of the set standard deviation. The standard deviation can be set to 500 meters.

[0121] A weighted linear combination method was used to fuse the score difference, propagation probability, and distance attenuation factor, with fusion weights set to 0.5, 0.3, and 0.2, respectively. Finally, the weight value of each edge was calculated to construct a failure propagation path graph P. The failure propagation path graph P is constructed with structurally vulnerable nodes as the central nodes. The direction of its edges reflects the potential propagation direction of health degradation, and the edge weights reflect the propagation intensity, providing a structural foundation for subsequent path identification.

[0122] In the constructed weighted directed graph, a path traversal operation based on a depth-first search algorithm is performed. Starting from each structurally vulnerable node, a degenerate path that can be reached outwards is recursively searched, and path chains in which the cumulative edge weight exceeds a set path threshold (e.g., 1.5) are retained.

[0123] The identified path chains are potential cascading degradation paths, which can be used for risk warning, priority scheduling of maintenance resources, and regional network structure optimization decisions.

[0124] S700: Based on the centrality index and degradation rate in the failure propagation path graph P, output a priority-ranked table of high-risk maintenance for charging piles, and generate an optimized inspection path by combining the maintenance radius and resource availability.

[0125] First, based on the aforementioned failure propagation path graph P, the propagation path centrality index is calculated for all nodes in the graph. The specific process is as follows:

[0126] In all failed path chain groups, count the frequency of each node appearing in the path;

[0127] Calculate the length of each path and use the reciprocal of the path length as a weighting factor;

[0128] The path centrality index of a node is calculated by weighting the frequency of its occurrence in all paths with the reciprocal of the corresponding path length.

[0129] This centrality index reflects the degree to which a node is central to the propagation of the entire potentially degenerate network; the higher the value, the more likely the node is to become a key transmission hub for degenerate signals.

[0130] Obtain the health score sequence for each node within the most recent consecutive time windows. Calculate the difference between the current score and the initial score in the sequence, and divide it by the window time span to obtain the node's health degradation rate. The calculation formula is: degradation rate equals the rate of score decrease divided by the number of time windows.

[0131] This rate is used to reflect the speed at which a node's recent health condition deteriorates, highlighting high-risk targets with rapidly declining health in the short term.

[0132] A comprehensive risk score is generated by weighting the path centrality index of a node with its health degradation rate. The weight ratio can be set to 0.6 and 0.4, that is, the centrality index has a weight of 0.6 and the degradation rate has a weight of 0.4. All nodes are sorted from high to low according to this score to generate a priority table for the maintenance of high-risk charging piles.

[0133] This priority table clearly reflects which devices not only have significant performance degradation but also have strong structural propagation capabilities, providing a basis for subsequent resource deployment.

[0134] Based on the top K high-risk charging pile nodes in the priority table (the K value can be set according to resource conditions, such as 10 or 20), their geographical location information is extracted, and the following operation and maintenance constraints are considered:

[0135] The maximum service radius of maintenance personnel or equipment (e.g., not exceeding 5 kilometers);

[0136] The maximum number of charging stations that can be covered at any time;

[0137] Constraints on current human resources or inspection time windows.

[0138] Based on the above constraints, the target nodes are constructed as a graph structure with coordinate attributes, and the Traveling Salesman Path (TSP) algorithm is used for path search. The objective function is to minimize the total path length while satisfying the service radius and resource conditions. The algorithm outputs one or more optimal inspection paths, each path containing a sequence of nodes to be visited and a suggested inspection order.

[0139] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for evaluating the charging health of heavy-duty trucks based on optimized graph convolutional networks, characterized in that: include: S100: Acquire multi-source heterogeneous data of all charging piles in the target area, including time-series voltage and current data, internal temperature change trajectory of the module, communication delay, user operation frequency and geographical location information, and perform unified time window segmentation on the data. S200: Based on a unified segmented time window, the coupling behavior between charging piles is modeled, and a dynamic multi-graph structure is constructed, in which each subgraph corresponds to a different type of coupling relationship, and the temporal order of the edges in the graph is encoded by a sliding window method. S300. For each subgraph in the dynamic multigraph structure, fuse the local state evolution trend of the nodes in the graph with the influence factors of the neighbor graph to obtain the joint feature vector Fi of the nodes. S400. Concatenate the joint feature vector Fi according to the time window order to form a sequence feature tensor T, and input it into a multi-scale attention-enhanced graph convolutional network for feature encoding to obtain the joint potential health representation Z. S500, calculates the predicted health score H for each charging pile based on Z; S600: Based on the predicted health score H, it automatically identifies nodes that exhibit structural vulnerability in multiple heterogeneous graphs and, combined with their topological clustering in space, constructs a potential failure propagation path map P for charging piles. S700: Based on the centrality index and degradation rate in the failure propagation path graph P, output a priority-ranked table of high-risk maintenance for charging piles, and generate an optimized inspection path by combining the maintenance radius and resource availability. The centrality index is obtained by weighting the frequency of a node's occurrence in all failed paths with the path length, and the degradation rate is the ratio of the decrease in its health score to the time interval within a continuous time window.

2. The method for evaluating the charging health of heavy-duty trucks based on optimized graph convolutional networks according to claim 1, characterized in that: The modeling of the coupling behavior between charging piles and the construction of a dynamic multigraph structure include: Based on the unified segmented time window data, three types of initial coupling relationship matrices are constructed according to the spatial distance between charging piles, the correlation of electrical parameters, and the coordinated changes in operating status. Generate corresponding subgraphs based on each type of coupling relationship matrix, and construct a graph set containing multiple heterogeneous graphs, where the edges of each subgraph represent a type of coupling. For the edge relationships in each subgraph, the evolution trend of edge weights at different times is embedded with timestamps, and the dynamic change characteristics of the edges are preserved. The dynamic change characteristics of edges are jointly represented with the characteristics of nodes.

3. The method for evaluating the charging health of heavy-duty trucks based on optimized graph convolutional networks according to claim 1, characterized in that: For each subgraph in the dynamic multigraph structure, the local state evolution trend of nodes in the graph and the influence factors of neighboring graphs are integrated, including: For each subgraph in the dynamic multigraph structure, the state change sequence of each node under a continuous time window is extracted, and the local state evolution trend vector is extracted using a gated cyclic unit. Based on the neighbor set of each node, calculate the weighted average feature representation of its neighbor nodes in the current subgraph; The local state evolution trend vector of a node is concatenated with the influence factors of the neighbor graph, and then nonlinearly fused through a fully connected neural network to obtain the joint feature vector Fi of the node. Perform a fusion operation on each type of subgraph to obtain joint feature representations of nodes under multiple coupling relationships.

4. The method for evaluating the charging health of heavy-duty trucks based on optimized graph convolutional networks according to claim 3, characterized in that: The step of concatenating the joint feature vector Fi into a sequence feature tensor T according to the time window order, and inputting it into a multi-scale attention-enhanced graph convolutional network for feature encoding includes: Arrange the joint feature vectors of nodes under each time window in chronological order to construct a three-dimensional sequence feature tensor, where the first dimension of the tensor represents the time window index, the second dimension represents the node number, and the third dimension represents the node feature dimension. Graph convolution operations are performed on sequence feature tensors, and a multi-scale graph convolution structure composed of multiple convolution scales is used to capture local and global topological relationship features between nodes with different adjacency ranges. The features of different time windows are weighted and fused, and the attention weight is calculated by the magnitude of node state change and the consistency of adjacency structure. The fused features are transformed by an activation function and output as a joint potential health representation Z for each node.

5. The method for evaluating the charging health of heavy-duty trucks based on optimized graph convolutional networks according to claim 1, characterized in that: The predicted health score H for each charging station calculated based on Z includes: Perform feature normalization on the joint potential health representation of each node obtained; The normalized features are input into a fully connected regression neural network with residual connection structure. The neural network consists of two linear transformation layers and one activation function, and the output is a continuous numerical score. During the model training phase, mean squared error is used as the loss function to optimize the deviation between the network output score and the historical actual health label through backpropagation. Based on the predicted health score H output by the model and the set state classification threshold, charging piles are classified into healthy, sub-healthy, and abnormal states.

6. The method for evaluating the charging health of heavy-duty trucks based on optimized graph convolutional networks according to claim 1, characterized in that: The method, based on a predicted health score H, automatically identifies nodes exhibiting structural vulnerability across multiple heterogeneous graphs, including: Based on the obtained predicted health score H, structural vulnerability analysis is performed on nodes in multiple heterogeneous graphs. The vulnerability index is jointly calculated using the sparsity of edge weights between nodes, the lower bound of the health score, and the characteristic volatility. Nodes with vulnerability indices exceeding a preset stability threshold are then selected. Extract the adjacency structure of nodes identified as structurally fragile in the spatial coupling subgraph and calculate their topological clustering coefficient in the local region; Based on the health score difference between nodes, edge weight propagation probability, and spatial distance, a weighted directed graph is constructed as a potential failure propagation path map, where the direction of the edge represents the direction of degradation propagation and the weight represents the propagation intensity. A depth-first traversal is performed on the potential failure propagation path graph to identify path chains that trigger cascading degradation.

7. The method for evaluating the charging health of heavy-duty trucks based on optimized graph convolutional networks according to claim 6, characterized in that: Based on the health score differences between nodes, edge weight propagation probabilities, and spatial distances, a weighted directed graph is constructed as a potential failure propagation path map, including: Obtain the health score difference between all node pairs in multiple heterogeneous graphs, and calculate the score decrease between any two adjacent nodes as the initial driving factor for failure propagation. The graph structure edge weights between node pairs are normalized, and the propagation probability of the edges is calculated by combining the dynamic change frequency of the edges. The distance attenuation factor is calculated based on the geographic spatial distance between node pairs, and the spatial distance is weighted and compressed using a Gaussian function. The weighted fusion of score decline rate, edge weight propagation probability and distance decay factor generates the final weight of the edge, and a weighted directed graph is constructed as a potential failure propagation path graph, in which the direction of the edge is from high health node to low health node.

8. The method for evaluating the charging health of heavy-duty trucks based on optimized graph convolutional networks according to claim 1, characterized in that: The method of generating optimized inspection paths by combining maintenance radius and resource availability includes: Based on the constructed potential failure propagation path map, the propagation path centrality index of each node and the degradation rate of each node are calculated. A priority table for the maintenance of high-risk charging piles is generated by weighting and sorting nodes based on their path centrality index and degradation rate. By combining the geographical location of each high-risk node, the maximum service radius of maintenance personnel, and the constraints of currently available resources, the Traveling Salesman Path (TSP) optimization algorithm is used to calculate the shortest inspection path and output a list of optimal inspection sequences.

Images