Railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network

Abstract

The invention belongs to the technical field of railway dispatching management, and provides a railway dispatching safety risk assessment method based on causal inference and a heterogeneous graph neural network in order to solve the technical problem that risk levels cannot be accurately identified and quantified due to the fact that an existing machine learning method cannot cope with complex and changeable railway dispatching environments. Comprising the following steps: S1, constructing a railway dispatching risk index system; s2, constructing a risk event causal graph; s3, calculating a causal effect weight; s4, constructing a heterogeneous graph neural network model based on causal inference; s5, constructing multi-source heterogeneous data; and S6, inputting the multi-source heterogeneous data into the heterogeneous graph neural network model based on causal inference to obtain the risk level of each key risk variable. According to the method, causal inference and the heterogeneous graph neural network are combined, modeling analysis is carried out on complex space-time interaction and node causal relationships in a railway dispatching system, and the accuracy of risk identification and the precision of an evaluation model are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of railway dispatching management, and in particular relates to a railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network. Background Art

[0002] As an important part of the modern transportation system, the safety of the railway dispatching system not only directly affects the safety of passengers and goods, but also plays a key role in the efficient operation of the entire railway network. With the continuous expansion of the scale of the railway transportation network, the complexity of the dispatching system is also increasing, resulting in increasing challenges in railway dispatching safety management.

[0003] Against the background of the expansion of the railway network, the risk sources and uncertainties involved in the dispatching process are increasing, and the interaction between key entities such as stations, trains and signals is becoming more complex and has significant temporal and spatial interactions. Existing railway dispatching safety management methods mainly rely on preset rules and expert experience, which are effective in relatively simple railway dispatching systems in the early days. However, with the increasing complexity of modern railway dispatching scenarios and the occurrence of emergencies, traditional methods are gradually showing their limitations. Preset rules lack flexibility and are difficult to cope with dynamically changing risk environments. Expert experience is limited by individual knowledge and is difficult to fully identify and cover potential risk factors in the dispatching system, resulting in a decrease in dispatching efficiency and system security.

[0004] In recent years, the rapid development of artificial intelligence and big data technologies has promoted the application of machine learning in the field of railway dispatching, identifying potential risks in the system through data-driven methods. Existing machine learning methods still have certain limitations when dealing with the interactive relationships between multi-level and multi-type nodes in railway dispatching systems. The heterogeneity of railway dispatching data is prominent, and different types of nodes have different characteristics and interaction patterns. Traditional machine learning models are difficult to fully capture these characteristics. In addition, the causal relationship between nodes is affected by the interaction of multiple factors. The existing models are insufficient in establishing and explaining causal relationships, making it difficult to provide reliable decision support for railway dispatching safety management.

[0005] Therefore, there is an urgent need for a railway dispatching safety risk assessment method that can cope with complex and changeable dispatching environments, accurately identify and quantify risk levels, and have good interpretability. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network (Heterogeneous Graph Neural Network, HetGNN), which combines causal inference and heterogeneous graph neural network to enhance the risk identification and assessment capabilities in the dispatching system, optimize dispatching decisions, and improve the operating efficiency and safety of the system.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network includes the following steps:

[0009] S1. Based on the historical risk events in the railway dispatching system, a railway dispatching risk indicator system is constructed and multiple key risk variables are obtained;

[0010] S2. construct a risk event cause-effect diagram based on the multiple key risk variables obtained in step S1;

[0011] S3, based on the risk event causal diagram obtained in step S2, calculating the causal effect weight;

[0012] S4. According to the causal effect weights obtained in step S3, a heterogeneous graph neural network model based on causal inference is constructed by introducing the causal effect weights into the graph convolutional network module and the graph attention network module respectively;

[0013] S5. Collect data according to the multiple key risk variables obtained in step S1, pre-process the data to obtain multi-source heterogeneous data, and represent the multi-source heterogeneous data as a feature vector;

[0014] S6. Input the multi-source heterogeneous data feature vector obtained in step S5 into the heterogeneous graph neural network model based on causal inference obtained in step S4 to obtain the risk level of each key risk variable and node.

[0015] Preferably, step S1 comprises the following steps:

[0016] S11. Based on historical risk events in the railway dispatching system, screen and classify different risk factors that affect dispatching safety and obtain multiple key risk factors;

[0017] S12. Constructing a railway dispatching risk indicator system based on the multiple key risk factors obtained in step S11;

[0018] S13, refining the multiple key risk factors in the railway scheduling risk indicator system obtained in step S12 into secondary indicators to obtain multiple key risk variables.

[0019] Preferably, step S2 comprises the following steps:

[0020] S21, inputting the multiple key risk variables obtained in step S1 and the causal relationship between the multiple key risk variables into the heterogeneous graph neural network, and extracting and obtaining an embedding result;

[0021] S22, according to the embedding result obtained in step S21, identify the key risk variables with high influence in the heterogeneous graph neural network through the degree centrality of the nodes, obtain the important key risk variables, and generate the important key risk variable influence path;

[0022] S23. Based on the important key risk variable impact paths generated in step S22, a risk event cause-and-effect diagram is constructed.

[0023] Preferably, in step S22, when the degree centrality of a node is greater than 0.7, the node is determined to be a key risk variable with high influence in the heterogeneous graph neural network.

[0024] Preferably, in step S3, the causal effect weight is calculated by combining the risk event causal graph obtained in step S2 with a causal inference algorithm, and the calculation formula of the causal effect weight is:

[0025] τ=Ε[Y j |do(X i =x)]-Ε[Y j |do(X i = x′)]

[0026] In the formula, τ is the causal effect weight, x and x′ are both intervention variables, do(X i = x) and do(X i = x′) are all do operations, Ε[Y j |do(X i = x)] means that when the intervention variable is x, the outcome variable Y j The expected value of Ε[Y j |do(X i = x′)] means that when the intervention variable is x′, the outcome variable Y j expected value.

[0027] Preferably, in step S4, the constructed heterogeneous graph neural network model based on causal inference includes:

[0028] The input module is used to generate a feature matrix based on the input multi-source heterogeneous data feature vector, and transmit the generated feature matrix to the graph convolution network module and the graph attention network module;

[0029] The graph convolutional network module is used to traverse the neighbor nodes according to the spatial neighborhood structure of the station node in the generated feature matrix, aggregate the neighbor node information through the graph convolution operation, update the features and neighborhood information of the station nodes layer by layer, and obtain the updated spatiotemporal features and neighborhood information of the station nodes;

[0030] The graph attention network module is used to adaptively learn the relationship between train nodes and signal nodes through the attention mechanism according to the dynamic interaction characteristics of train nodes and signal nodes in the generated feature matrix, adjust the attention weights of different neighbor nodes, and obtain the updated spatiotemporal characteristics and neighborhood information between train nodes and signal nodes;

[0031] A causal inference module, which is used to adjust the aggregation mode of neighbor node information by the causal effect weight obtained in step S3 when the graph convolution network module aggregates neighbor node information, and to adjust the influence of neighbor nodes on the current node by the causal effect weight obtained in step S3 when the graph attention network module adjusts the attention weights of different neighbor nodes;

[0032] The risk assessment module is used to input the updated spatiotemporal characteristics and neighborhood information of the station nodes and the updated spatiotemporal characteristics and neighborhood information between the train nodes and the signal nodes into the multi-layer perceptron to obtain the risk level of each node, that is, the risk level of each key risk variable.

[0033] Preferably, the multilayer perceptron consists of a first fully connected layer, a second fully connected layer and an output layer;

[0034] The input dimension of the first fully connected layer is the node embedding dimension, the output dimension is 128, and the activation function is ReLU. The input dimension of the second fully connected layer is 128, the output dimension is 64, the activation function is ReLU, the input dimension of the output layer is 64, the output dimension is the number of risk levels, and the activation function is Softmax.

[0035] Preferably, in step S5, the multi-source heterogeneous data includes station node data and data between train nodes and signal nodes.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. This invention combines causal inference with heterogeneous graph neural networks to model and analyze complex spatiotemporal interactions and node causal relationships in railway dispatching systems, thereby improving the accuracy of risk identification and the precision of evaluation models;

[0038] 2. The present invention explains and describes the causal chain of risk events in the system by constructing a causal relationship diagram, and deeply understands the causes of risks, so that dispatchers can formulate targeted dispatch measures according to risk levels and optimize dispatch responses;

[0039] 3. This invention provides a new perspective for railway dispatching safety risk management. The introduction of causal learning improves the explanatory and generalization capabilities of heterogeneous graph neural network models. It can not only better explain the causes and impacts of risks, but also has strong generalization capabilities, providing reference and reference for applications in other complex networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flow chart of the railway dispatching safety risk assessment method proposed by the present invention;

[0041] Figure 2 It is a railway dispatching safety risk indicator system in an embodiment of the present invention;

[0042] Figure 3 A risk event cause-and-effect diagram in an embodiment of the present invention;

[0043] Figure 4 The railway network in the embodiment of the present invention;

[0044] Figure 5 It is a comparison chart of risk scoring curves of different methods;

[0045] Figure 6 Heat map comparison of different methods; (a) complete model; (b) removing the causal inference component; (c) removing the heterogeneous graph neural network; (d) graph convolutional neural network model; (e) graph attention network model. DETAILED DESCRIPTION

[0046] The following is a description of the embodiments of the present invention. Figures 1 to 6 , the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] like Figure 1 As shown, an embodiment of the present invention provides a railway scheduling safety risk assessment method based on causal inference and heterogeneous graph neural network, which specifically includes the following steps:

[0048] S1. Based on the three dimensions of personnel, technology and environment, and based on historical risk events in the railway dispatching system, a railway dispatching risk indicator system is constructed, and multiple key risk variables are obtained, which specifically includes the following steps:

[0049] S11. Based on historical risk events in the railway dispatching system, historical dispatching data, equipment failure records and personnel operation behavior analysis, different risk factors affecting dispatching safety are screened and classified to obtain multiple key risk factors, which are: train risk, station risk, signal risk, overall system risk, train-station interaction risk, train-signal interaction risk; and the definition and description of high-speed railway risk events are shown in Table 1;

[0050] Table 1 Classification and definition of high-speed railway risk events

[0051]

[0052] S12. According to the multiple key risk factors obtained in step S11, a railway dispatching risk index system is constructed. The railway dispatching risk index system is represented by a matrix as W=[W1, W2, ..., W n ], where W i Indicates different risk events;

[0053] S13, refining the multiple key risk factors in the railway dispatching risk indicator system obtained in step S12 into secondary indicators respectively, and obtaining multiple key risk variables, such as Figure 2 As shown in Figure 1, multiple key risk variables include train derailment W1, train collision W2, train delay W3, station congestion W4, station safety accident W5, signal failure W6, signal display error W7, signal misjudgment W8, signal response delay W9, station entry safety W1, signal failure W2, signal response delay W3, signal failure W4, signal failure W5, signal failure W6, signal failure W7, signal failure W8, signal response delay W9, station entry safety W10, signal failure W11, signal failure W12, signal failure W13, signal failure W14, signal failure W15, signal failure W16, signal failure W17, signal failure W18, signal failure W19, signal failure W20, signal failure W21, signal failure W22, signal failure W23, signal failure W24, signal failure W25, signal failure W26 10 , Outbound Security 11 , Dispatching system failure 12 , Comprehensive operational risk 13 .

[0054] S2. Analyze the causes, impact paths and consequences of risk events, and construct a risk event causal diagram based on the multiple key risk variables obtained in step S1, G c =(W c ,E c ), Wc is a node set, representing different key risk variables, E c is an edge set, which represents the causal relationship between different key risk variables; it specifically includes the following steps:

[0055] S21, inputting the multiple key risk variables (i.e., node features) obtained in step S1 and the causal relationship between the multiple key risk variables (i.e., edge features) into the heterogeneous graph neural network, and extracting and obtaining an embedding result;

[0056] S22. According to the embedding result obtained in step S21, the key risk variables with high influence in the heterogeneous graph neural network are identified through the degree centrality of the node. Specifically, when the degree centrality of the node is greater than 0.7, the node is determined to be a key risk variable with high influence in the heterogeneous graph neural network, that is, an important key risk variable is obtained, and an important key risk variable influence path is generated. The calculation formula of the degree centrality of the node is:

[0057]

[0058] Among them, D i is the degree centrality of node i, k i is the number of edges connected to node i, and N is the maximum number of edges that may be connected to node i.

[0059] S23. Based on the important key risk variable impact path generated in step S22, a risk event cause-effect diagram is constructed, such as Figure 3 As shown, Figure 3 The nodes in the risk indicator system represent key risk variables, such as train status, signal status, and environmental variables, and the edges represent the causal relationship between these variables.

[0060] S3, based on the risk event causal diagram obtained in step S2, calculating the causal effect weight;

[0061] Based on the risk event causal graph obtained in step S2, the present invention uses a causal inference algorithm, uses front-door criteria and back-door criteria to perform causal inference, calculates the causal effect weight τ, evaluates the impact of each key risk variable on the overall risk of the system, identifies the root cause of the risk event through the causal path, and when the path causal weight exceeds a preset threshold, uses its starting node as a potential cause node;

[0062] The causal inference algorithm can analyze the causal relationship between different risk events under the condition of intervention, and quantify the impact of the risk source on the result. The causal inference algorithm uses the front-door criteria and the back-door criteria to perform causal inference and evaluate the impact of intervention on the risk result. The causal inference algorithm is a form of prior knowledge, and the results of causal problems can be obtained by probability analysis. For causal inference, it is necessary to first establish a causal model, that is, a risk event causal graph, that is, it is necessary to first assume how this model works and what event causes what event consequences. Once the risk event causal graph is drawn, the effect of the intervention can be simulated by destroying the causal network, that is, deleting all nodes that cause the do operation and the corresponding edges, and finally calculating the causal effect weight through the formula. The calculation formula for the causal effect weight is:

[0063] τ=Ε[Y j |do(X i=x)]-Ε[Y j |do(X i = x′)]

[0064] In the formula, τ is the causal effect weight, x and x′ are both intervention variables, do(X i = x) and do(X i = x′) are all do operations, Ε[Y j |do(X i = x)] means that when the intervention variable is x, the outcome variable Y j The expected value of Ε[Y j |do(X i = x′)] means that when the intervention variable is x′, the outcome variable Y j expected value.

[0065] The description of the causal inference algorithm is shown in Table 2 below.

[0066] Table 2 Causal inference algorithm calculation flow chart

[0067]

[0068] S4. In view of the differences in structure, source and characteristics of railway data, as well as the complexity of spatiotemporal interactions between nodes and edges, the causal effect weights obtained in step S3 are introduced into the graph convolutional network module and the graph attention network module respectively to construct a heterogeneous graph neural network model G based on causal inference. h =(V h ,E h ,X h ), Vh is a node set, representing physical nodes such as stations and trains, E h is an edge set, representing the connection relationship between nodes, X h Taking the feature vector as the model input, the nodes with stronger spatiotemporal correlation and causal relationship are identified and analyzed through causal inference, and the transmission path of risk events from the root cause node to the result node is traced based on the causal path, which improves the accuracy of risk assessment of the railway dispatching system and the explanatory power of the model;

[0069] In an embodiment of the present invention, the constructed heterogeneous graph neural network model based on causal inference includes:

[0070] The input module is used to generate a feature matrix based on the input multi-source heterogeneous data feature vector, and send the generated feature matrix to the graph convolutional network module and the graph attention network module; the specific multi-source heterogeneous data includes train operation data, equipment status data, weather data, etc., and independent feature embedding layers are designed for different types of nodes and edges to generate feature matrices

[0071] The graph convolutional network module is used to traverse the neighbor nodes according to the spatial neighborhood structure of the station node in the generated feature matrix, aggregate the neighbor node information through graph convolution operations, update the features and neighborhood information of the station nodes layer by layer, and obtain the updated spatiotemporal features and neighborhood information of the nodes; specifically, the features of the direct neighbor nodes of the station node are weighted and summed through normalization operations, and the nonlinear activation function ReLU is used to process the weighted features, extract spatial structure information and enhance the ability to represent the relationship between station nodes.

[0072] The graph attention network module is used to adaptively learn the relationship between train nodes and signal nodes through the attention mechanism according to the dynamic interaction characteristics of train nodes and signal nodes in the generated feature matrix, adjust the attention weights of different neighbor nodes, obtain the updated spatiotemporal characteristics and neighborhood information of station nodes, and highlight the important nodes of spatiotemporal interaction; specifically, by introducing the causal effect weight τ into the attention mechanism, the weight distribution of neighbor nodes is dynamically adjusted, and the nodes with relatively important spatiotemporal interaction between trains and signal machines are highlighted;

[0073] A causal inference module, which is used to adjust the aggregation mode of neighbor node information by the causal effect weight obtained in step S3 when the graph convolution network module aggregates neighbor node information, and to adjust the influence of neighbor nodes on the current node by the causal effect weight obtained in step S3 when the graph attention network module adjusts the attention weights of different neighbor nodes;

[0074] The risk assessment module is used to input the updated spatiotemporal characteristics and neighborhood information of the station nodes and the updated spatiotemporal characteristics and neighborhood information between the train nodes and the signal nodes into the multi-layer perceptron to obtain the risk level of each node, that is, the risk level of each key risk variable.

[0075] In the embodiment of the present invention, the multilayer perceptron is composed of a first fully connected layer, a second fully connected layer and an output layer;

[0076] The input dimension of the first fully connected layer is the node embedding dimension, the output dimension is 128, and the activation function is ReLU. The input dimension of the second fully connected layer is 128, the output dimension is 64, and the activation function is ReLU. The input dimension of the output layer is 64, the output dimension is the number of risk levels, and the activation function is Softmax. Based on the output of the graph convolutional network module and the graph attention network module, according to the probability distribution of the risk level of the node, the probability distribution of the risk level corresponding to each node is output by using the Softmax function, the risk level of the node is determined according to the probability distribution, and the risk threshold is set to conduct risk assessment on the node.

[0077] The following is a further description of the design principle of the heterogeneous graph neural network model based on causal inference:

[0078] Station nodes have relatively stable physical locations and structural characteristics in the railway network. In order to effectively aggregate the information of neighboring nodes, the graph convolutional network module is used to design the convolution kernel. The feature information of the station node is provided by its direct neighboring nodes. These features are normalized, weighted and summed, and processed by a nonlinear activation function σ to generate a new node embedding representation. The causal effect weight τ is used to adjust the aggregation method of neighbor node information. When aggregating information, the model determines which neighbor nodes have a greater influence on the information transmission of the current node, that is:

[0079]

[0080] in, represents the embedding of station node i at the kth layer, represents the neighbor set of node i, τ represents the causal effect weight, c ij is the normalization constant, W (k) is the weight matrix of the kth layer, and σ is the activation function.

[0081] According to the dynamic interaction characteristics of train and signal nodes, the spatiotemporal convolution kernel is designed using the graph attention network module. In the graph attention network module, the information transmission between nodes depends on the attention weight α ij , which is used to reflect the importance of node i to neighbor node j. The attention mechanism further integrates the causal weight τ, and adjusts the influence of neighbor nodes on the current node according to the causal effect weight between nodes. This ensures that the model can not only aggregate information based on structured graph relationships, but also capture the causal effects between nodes, that is:

[0082]

[0083] Among them, α ij is the attention weight between node i and node j, that is:

[0084]

[0085] Among them, a T is a learnable parameter in the attention mechanism, LeakyReLU is the activation function, and || represents the concatenation operation.

[0086] In the operation process of each layer, the node receives information from its neighbor nodes and updates its embedded representation. After multiple layers of message passing and aggregation operations, the node gradually accumulates and refines the structured information from the neighborhood to form a more expressive node representation. The update formula of node i is:

[0087]

[0088] Aggregate represents an aggregation function, which can be used to calculate the average value.

[0089] The output layer of the model maps the node representation to a specific risk level space through a fully connected layer, and uses a multi-layer perceptron (MLP) to complete the classification of risk levels. MLP consists of multiple fully connected layers, where the first layer has an input dimension of node embedding, an output dimension of 128, and an activation function of ReLU. The second layer has an input dimension of 128, an output dimension of 64, and an activation function of ReLU. The input dimension of the output layer is 64, the output dimension is the number of risk levels, and the activation function is Softmax. The node embedding representation that introduces causal relationships is input into the multi-layer perceptron to predict the risk level, that is:

[0090] z i =Softmax(W3·ReLU(W2·ReLU(W1h i +b1)+b2)+b3)

[0091] Among them, z i is the risk score of the category corresponding to node i, W1, W2, W3 are learnable parameter matrices, and b1, b2, b3 are bias terms.

[0092] Through multiple layers of linear transformation and nonlinear activation functions, the model maps the embedded representation to different risk level spaces. The specific risk level division is shown in Table 3 below, and the threshold of the risk score is set to less than 0.25 for low risk, greater than 0.65 for high risk, and medium risk to 0.25-0.65.

[0093] Table 3 Risk level classification

[0094]

[0095] S5. Collect data based on the multiple key risk variables obtained in step S1, preprocess the data to obtain multi-source heterogeneous data, and represent the multi-source heterogeneous data as a feature vector; the multi-source heterogeneous data includes station node data and data between train nodes and signal nodes.

[0096] S6. Input the multi-source heterogeneous data feature vector obtained in step S5 into the heterogeneous graph neural network model based on causal inference obtained in step S4 to obtain the risk level of each key risk variable.

[0097] The performance of the railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network provided by the embodiment of the present invention is studied below.

[0098] 1. Experimental design

[0099] A survey was conducted on 10 stations along the railway line. The survey results are as follows: Figure 4 As shown, the actual operation data in the railway network is obtained. The present invention adopts Python programming language and Pytorch deep learning framework. In terms of hyperparameter setting, the learning rate of the model is set to 0.001, the optimizer uses Adam, the batch size is 64, and the number of training rounds is 100. Dropout regularization technology is used to prevent overfitting, the Dropout rate is 0.5, and L2 regularization is used to constrain the weight.

[0100] 2. Results comparison

[0101] (1) The present invention is compared with three classic machine learning algorithms (LR, SVM, RF) and two mainstream deep learning methods: graph neural network model (GNN) and graph attention network model (GAT). The comparative experimental results are shown in Table 4 below.

[0102] Table 4 Comparison results of different methods

[0103]

[0104] From the results in Table 4, it can be seen that the model provided by the present invention is significantly superior to other benchmark models in all indicators, and can make full use of the complex information of multiple types of nodes and edges in heterogeneous graphs, especially in railway scheduling, effectively capturing the interactions between nodes, and providing a more comprehensive perspective for risk identification. However, the performance of traditional machine learning models (such as LR, SVM, RF) is limited because they fail to utilize this structured information. In addition, the present invention combines causal learning, which greatly improves the interpretability and prediction accuracy of the model by identifying the causal relationship between data. In contrast, although GNN and GAT can process graph data, they lack causal inference capabilities and perform poorly in processing complex relationships.

[0105] (2) The contribution of key components in the causal inference-based heterogeneous graph neural network model of the present invention to the overall performance is evaluated through ablation experiments. Three configurations are designed: a complete model, removing the causal learning component, and removing the heterogeneous graph neural network component to evaluate the impact of different components on the performance of the model of the present invention, as shown in Table 5 below.

[0106] Table 5 Experimental results of different configurations

[0107]

[0108] According to the results in Table 5, it can be seen that after removing the causal learning component, the accuracy and F1-score of the model decrease, and the model cannot effectively distinguish complex relationships, resulting in performance degradation. This shows the importance of using causal relationships in railway dispatching systems to improve the accuracy and interpretability of model risk identification. There are not only structured connections between nodes in the railway dispatching system, such as the relationship between trains and stations, but also temporal and logical causal relationships, such as the impact of train delays on subsequent train operations. Through causal learning, the model can effectively identify and utilize these causal relationships and improve its ability to handle complex risk scenarios.

[0109] The removal of the heterogeneous graph neural network component leads to a significant drop in model performance, reflecting that the graph structure information of railway scheduling is crucial to the evaluation of the model. The railway network has a natural graph structure, and the connections between nodes reflect the actual physical location and operation logic. The heterogeneous graph neural network integrates the relationship information between nodes, allowing the model to more comprehensively understand the risk propagation path in the system. After removal, the model can only rely on the local information of the nodes and lacks the support of global structured information, so the performance is weak. It shows that the complex network structure in the railway scheduling system needs to be effectively modeled by heterogeneous graph neural networks to achieve accurate risk identification.

[0110] The complete model enhances the model's interpretation and generalization capabilities by combining causal inference and heterogeneous graph neural networks. In the task of railway dispatching safety risk identification, the synergy of the two is particularly important in the highly complex spatiotemporal dependencies of the railway system.

[0111] (3) Input the preprocessed data into the model, and generate node risk score curves of different algorithms by deeply analyzing the characteristics of each node. Figure 5 As shown in the figure, the risk score threshold is set to be less than 0.25 for low risk, greater than 0.65 for high risk, and medium risk between 0.25 and 0.65. Figure 5 It can be seen that the model (complete model) provided by the present invention can make full use of input information, especially in capturing the characteristics of high-risk nodes. The risk scoring accuracy of the ablation experiment of the model after removing the causal inference component further decreased, indicating that graph structure information and causal explanation are crucial to the prediction of the model. GCN and GAT showed insufficient scoring ability and were not sensitive to high-risk nodes, indicating that the capabilities of these methods are limited by model design and it is difficult to fully characterize the risk characteristics of nodes.

[0112] like Figure 6As shown, the model (complete model) provided by the present invention shows clearer clustering. After combining the causal information, the node correlation is strong, and the risk propagation path and mutual influence are more obvious. After removing the causal inference component, the performance is somewhat weakened, and the model cannot fully understand the potential causal relationship in the data. After removing the heterogeneous graph neural network component, more low-correlation areas appear in the figure, and the performance of GCN and GAT is also slightly insufficient.

[0113] It can be seen from the description of the above embodiments that the present invention is more comprehensive and systematic than the prior art. In the process of risk identification and assessment, the priority order of high-risk events is clarified through the hierarchical management of risk events, the allocation and scheduling of resources are optimized, and the security of the system is improved. In addition, the risk assessment ideas proposed by the present invention are clear, and management measures can be formulated for high-risk events to enhance the actual effect of railway dispatching risk management.

[0114] In summary, the railway transportation network risk assessment method based on causal learning and heterogeneous graph neural network proposed in the embodiment of the present invention can evaluate the impact of various risk events in the line on the safe operation of the railway network based on historical accident and fault data, and realize the system-level assessment of operational risks. The present invention helps the dispatching department to formulate risk management plans for key lines and stations based on the assessment results, and provides a scientific basis for safety management in the railway dispatching system through quantitative risk assessment, helping dispatchers to identify high-risk nodes, optimize resource allocation and formulate management measures during the dispatching process.

[0115] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network, characterized in that: The following steps are involved: S1. Based on the historical risk events in the railway dispatching system, a railway dispatching risk indicator system is constructed and multiple key risk variables are obtained; S2. construct a risk event cause-effect diagram based on the multiple key risk variables obtained in step S1; S3, based on the risk event causal diagram obtained in step S2, calculating the causal effect weight; S4. According to the causal effect weights obtained in step S3, a heterogeneous graph neural network model based on causal inference is constructed by introducing the causal effect weights into the graph convolutional network module and the graph attention network module respectively; S5. Collect data according to the multiple key risk variables obtained in step S1, pre-process the data to obtain multi-source heterogeneous data, and represent the multi-source heterogeneous data as a feature vector; S6. Input the multi-source heterogeneous data feature vector obtained in step S5 into the heterogeneous graph neural network model based on causal inference obtained in step S4 to obtain the risk level of each key risk variable and node.

2. The railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network according to claim 1 is characterized in that: Step S1 includes the following steps: S11. Based on historical risk events in the railway dispatching system, screen and classify different risk factors that affect dispatching safety and obtain multiple key risk factors; S12, constructing a railway dispatching risk indicator system according to the multiple key risk factors obtained in step S11; S13, refining the multiple key risk factors in the railway scheduling risk indicator system obtained in step S12 into secondary indicators to obtain multiple key risk variables.

3. The railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network according to claim 1 is characterized in that: Step S2 includes the following steps: S21, inputting the multiple key risk variables obtained in step S1 and the causal relationship between the multiple key risk variables into the heterogeneous graph neural network, and extracting and obtaining an embedding result; S22, according to the embedding result obtained in step S21, identify the key risk variables with high influence in the heterogeneous graph neural network through the degree centrality of the nodes, obtain the important key risk variables, and generate the important key risk variable influence path; S23. Based on the important key risk variable impact paths generated in step S22, a risk event cause-and-effect diagram is constructed.

4. The railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network according to claim 3 is characterized in that: In step S22, when the degree centrality of a node is greater than 0.7, the node is determined to be a key risk variable with high influence in the heterogeneous graph neural network.

5. The railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network according to claim 1 is characterized in that: In step S3, the causal effect weight is calculated by combining the risk event causal graph obtained in step S2 with the causal inference algorithm. The calculation formula of the causal effect weight is: τ=Ε[Y j |do(X i =x)]-Ε[Y j |do(X i =x′)] In the formula, τ is the causal effect weight, x and x′ are both intervention variables, do(X i = x) and do(X i = x′) are all do operations, Ε[Y j |do(X i = x)] means that when the intervention variable is x, the outcome variable Y j The expected value of Ε[Y j |do(X i = x′)] means that when the intervention variable is x′, the outcome variable Y j expected value.

6. The railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network according to claim 1 is characterized in that: In step S4, the constructed heterogeneous graph neural network model based on causal inference includes: The input module is used to generate a feature matrix based on the input multi-source heterogeneous data feature vector, and transmit the generated feature matrix to the graph convolution network module and the graph attention network module; The graph convolutional network module is used to traverse the neighbor nodes according to the spatial neighborhood structure of the station node in the generated feature matrix, aggregate the neighbor node information through the graph convolution operation, update the features and neighborhood information of the station nodes layer by layer, and obtain the updated spatiotemporal features and neighborhood information of the station nodes; The graph attention network module is used to adaptively learn the relationship between train nodes and signal nodes through the attention mechanism according to the dynamic interaction characteristics of train nodes and signal nodes in the generated feature matrix, adjust the attention weights of different neighbor nodes, and obtain the updated spatiotemporal characteristics and neighborhood information between train nodes and signal nodes; A causal inference module, which is used to adjust the aggregation mode of neighbor node information by the causal effect weight obtained in step S3 when the graph convolution network module aggregates neighbor node information, and to adjust the influence of neighbor nodes on the current node by the causal effect weight obtained in step S3 when the graph attention network module adjusts the attention weights of different neighbor nodes; The risk assessment module is used to input the updated spatiotemporal characteristics and neighborhood information of the station nodes and the updated spatiotemporal characteristics and neighborhood information between the train nodes and the signal nodes into the multi-layer perceptron to obtain the risk level of each node, that is, the risk level of each key risk variable.

7. The railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network according to claim 6 is characterized in that: The multilayer perceptron consists of the first fully connected layer, the second fully connected layer, and the output layer; The input dimension of the first fully connected layer is the node embedding dimension, the output dimension is 128, and the activation function is ReLU. The input dimension of the second fully connected layer is 128, the output dimension is 64, the activation function is ReLU, the input dimension of the output layer is 64, the output dimension is the number of risk levels, and the activation function is Softmax.

8. The railway dispatching safety risk assessment method based on causal inference and heterogeneous graph neural network according to claim 6 is characterized in that: In step S5, the multi-source heterogeneous data includes station node data and data between train nodes and signal nodes.

Images