Train intelligent control method and system based on big data analysis

Abstract

The invention discloses a train intelligent control method and system based on big data analysis, and relates to the technical field of train intelligent control, and the method comprises the steps: obtaining train traffic original data through a traffic sensor network, and constructing a traffic feature matrix; constructing a train track topological structure chart, and constructing an anomaly detection model to detect abnormal nodes; combining an attention mechanism and double-layer time modeling to construct a dynamic graph neural network prediction model, and generating a cellular node prediction state; calculating propagation influence intensity of abnormal nodes based on a dynamic weight propagation formula, generating a dynamic abnormal propagation weight matrix, and calculating an abnormal influence index; and constructing a state space and an action space based on the state diagram, optimizing train scheduling by adopting a DQN of a dual-network architecture, and generating an intelligent scheduling strategy. Full-process optimization from data acquisition, anomaly detection, state prediction, anomaly influence evaluation to intelligent scheduling is realized, and the intelligent level, safety and efficiency of train operation are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of train intelligent control technology, and in particular to a train intelligent control method and system based on big data analysis. Background Art

[0002] With the continuous advancement of urbanization and the rapid expansion of public transportation systems, trains, as an important component of urban and intercity transportation, are facing a direct impact on traffic efficiency, public safety, and energy consumption levels through intelligent operation and control. Traditional train dispatching systems, which rely primarily on static dispatch tables and manual intervention, are slow to respond and have poor adaptability when faced with new challenges such as increased train operation density and complex and changing operating conditions. In recent years, the integration of the Internet of Things (IoT) and big data technologies has promoted the refined development of traffic perception and data collection. Rail transit systems have begun to possess real-time perception capabilities for factors such as train operation status, track occupancy, and signal status, thereby driving the evolution of traffic control systems towards a higher level of intelligence.

[0003] Although existing technologies have made initial progress in intelligent train control, they still face several key challenges. First, most existing models are limited to modeling traffic data based on static features or coarse-grained time series, lacking a deep expression of the track network structure and the spatiotemporal evolution of nodes, making it difficult to effectively extract potential anomaly diffusion paths and behavioral logic. Second, current anomaly detection methods are mostly single-stream models (such as detection methods based on statistical features or unidirectional graph convolution), which make it difficult to simultaneously focus on the synergy between node state features and structural context, resulting in insufficient anomaly identification accuracy. Third, traditional scheduling strategy optimization methods lack a dynamic feedback mechanism for the propagation effects of abnormal events, making it impossible to achieve predictive scheduling responses for future situations. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a train intelligent control method and system based on big data analysis to solve the problems of delayed perception of train operation status, insufficient modeling of abnormal event propagation, and weak intelligent optimization capability of scheduling strategy.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a train intelligent control method based on big data analysis, which includes:

[0008] The original train traffic data is obtained through the traffic sensor network, and the spatiotemporal features are extracted and preprocessed by road network division to construct a traffic feature matrix.

[0009] Construct a train track topology diagram and use a two-stream attention embedding network to build an anomaly detection model to detect abnormal events and classify abnormal nodes by probability.

[0010] Combining the attention mechanism with two-layer temporal modeling to build a dynamic graph neural network prediction model to generate the predicted state of cellular nodes;

[0011] The propagation impact strength of abnormal nodes is calculated based on the dynamic weight propagation formula, and a dynamic abnormal propagation weight matrix is generated. A bidirectional nested model is constructed based on Bi-GRU and neighborhood graph convolution to extract spatiotemporal embedding features. The abnormal impact index is calculated by combining the spatiotemporal embedding features and the dynamic abnormal propagation weight matrix.

[0012] Construct state space and action space based on state graph, and use DQN with dual network architecture to optimize train scheduling and generate intelligent scheduling strategy;

[0013] Based on the feedback results of the intelligent scheduling strategy, the deep Q network is regularly updated.

[0014] As a preferred solution of the train intelligent control method based on big data analysis described in the present invention, wherein: the raw train traffic data is obtained through the traffic sensor network, the spatiotemporal features are extracted and preprocessed by road network division, and the traffic feature matrix is constructed, including:

[0015] Deploy a train traffic sensor network to collect train operation status data, track usage status data, and train trajectory data to construct an original data set;

[0016] The original data set is cleaned using the adaptive time window method, and train location, travel speed, trajectory path, passing time, train density, and acceleration fluctuation data are extracted to form a train travel data set.

[0017] Based on the train travel dataset, the road network is divided into hexagonal honeycomb grids. For each time window, the spatial distribution characteristics of train tracks, the time-varying characteristics of train flow, and the changing characteristics of train travel time in the honeycomb unit are extracted. After normalization, the data are aggregated into the corresponding honeycomb unit to construct the traffic feature matrix.

[0018] As a preferred solution of the train intelligent control method based on big data analysis described in the present invention, wherein: the construction of the train track topology structure diagram, the use of the dual-stream attention graph embedding network to build an anomaly detection model to detect abnormal events and divide abnormal nodes by probability, including:

[0019] A train track topology map is constructed based on the traffic feature matrix. Cellular units are set as nodes, and the traffic feature matrix carried by each cell node is used as the node attribute. Edges are added between directly connected cell nodes with continuous track segments.

[0020] Use a two-stream attention graph embedding network to build an anomaly detection model, including an input layer, a two-stream attention mechanism, an embedding layer, and an output layer;

[0021] The input layer receives a set of nodes in the train track topology graph. The input features of each node are divided into spatial feature vectors and temporal feature vectors.

[0022] The spatial feature vector refers to the spatial distribution feature of the train track;

[0023] The time feature vector includes the time-varying feature of train flow and the time-varying feature of train travel;

[0024] Generate an adjacency matrix A(i) for each node i based on the edge connection relationship in the train track topology graph structure;

[0025] The dual-stream attention mechanism calculates the spatial feature attention weight and the temporal feature attention weight for each neighboring node j∈A(i) of node i;

[0026] Generate spatial enhancement features and temporal enhancement features through attention weights;

[0027] The spatial enhancement features and temporal enhancement features are fused to generate the final feature representation F of the node e (i);

[0028] The embedding layer performs graph convolution calculation on the final feature representation to generate the embedded representation of the honeycomb node;

[0029] The output layer uses the final embedding feature H of the node (K) , calculate the abnormal probability of cellular nodes;

[0030] Use the training set to train the model, select the cross entropy loss function to calculate the difference between the classification result and the true label, use the Adam optimizer to perform gradient descent optimization, update the model parameters, and stop iterating and output the model if the model loss no longer decreases significantly during the continuous iteration process;

[0031] Use the trained model to generate the abnormal probability of each node in the train track topology map, set a threshold based on historical experience, and classify a node as an abnormal node when the abnormal probability of the node is greater than the threshold.

[0032] As a preferred solution of the train intelligent control method based on big data analysis described in the present invention, wherein: the dynamic graph neural network prediction model is constructed by combining the attention mechanism with the double-layer time modeling to generate the cellular node prediction state, including:

[0033] Combining the attention mechanism with two-layer temporal modeling, a dynamic graph neural network prediction model is constructed, including an input layer, a graph convolution layer, a two-layer temporal modeling module, and an output layer.

[0034] The input layer inputs the node set and adjacency matrix in the train track topology graph;

[0035] The graph convolution layer aggregates information of the cellular nodes at each moment through the attention mechanism and generates the aggregated node representation H t ;

[0036] The two-layer time modeling module includes local time modeling and global time modeling;

[0037] The local time modeling uses causal convolution to H t Capture the nearest neighbor time dependency and generate local information Z t ;

[0038] The global time modeling adopts the self-attention mechanism to calculate H t Full temporal context representation, generating global information S t ;

[0039] The local information is weightedly fused with the global information to generate the final representation U t ;

[0040] The output layer uses a feedforward neural network with shared parameters to t Decode and predict the time series to generate the next time step prediction state of the train position, speed, acceleration, running direction and train density in each cellular node;

[0041] Use the training set to train the prediction model, update the model parameters, and output the model after training is completed;

[0042] Use the trained model to generate the predicted state of each cell node.

[0043] As a preferred solution of the train intelligent control method based on big data analysis described in the present invention, wherein: the propagation influence intensity of abnormal nodes is calculated based on the dynamic weight propagation formula, a dynamic abnormal propagation weight matrix is generated, a bidirectional nested model is constructed based on Bi-GRU and neighborhood graph convolution to extract spatiotemporal embedding features, and the abnormal impact index is calculated by combining the spatiotemporal embedding features and the dynamic abnormal propagation weight matrix, including:

[0044] Define the dynamic propagation weight formula, calculate the propagation impact intensity of abnormal nodes based on historical data and real-time data, and generate a dynamic abnormal propagation weight matrix;

[0045] Build a bidirectional nested model based on Bi-GRU and neighborhood graph convolution;

[0046] Extracting train running status feature time series X based on spatiotemporal traffic feature matrix i (t);

[0047] Use Bi-GRU to extract the propagation dynamics in the time series and generate propagation state embeddings;

[0048] Based on the propagation state, spatial dependencies are captured through neighborhood graph convolution to generate spatial embeddings;

[0049] Use the training set to train the bidirectional nested model and update the model parameters. If the model loss no longer decreases significantly during the continuous iteration, stop iterating and output the model.

[0050] Combining the dynamic anomaly propagation weight matrix and the trained bidirectional nested model, the anomaly impact indicators of abnormal nodes in the train track topology graph are calculated, including the propagation index, impact range, and efficiency reduction ratio.

[0051] As a preferred solution of the train intelligent control method based on big data analysis described in the present invention, wherein: the state space and action space are constructed based on the state diagram, and the DQN with a dual network architecture is used to optimize train scheduling and generate an intelligent scheduling strategy, including:

[0052] The real-time traffic feature matrix at each moment, the predicted state of each cellular node, the adjacency matrix of the track topology graph, and the abnormal impact index of the abnormal node are represented as a state graph;

[0053] Based on the state diagram, construct the state space S t ;

[0054] Based on the train track topology structure and historical scheduling experience, the action space A is constructed. t ;

[0055] Design a multi-objective optimization model, taking minimizing the propagation index, reducing the impact range, and improving efficiency as optimization goals, and design a reward function;

[0056] A deep Q-network using a dual-network architecture consisting of a master network and a target network is used to find the optimal scheduling strategy through reinforcement learning.

[0057] In the current state, the best action is selected by predicting the Q value through the main network based on the ∈-greedy strategy. The main network parameter is θ main ;

[0058] After executing the action, the reward and next state diagram are obtained;

[0059] The target network calculates the target Q value based on the reward, the next state diagram, and the action selected by the main network, and synchronizes the main network parameters to the target network every k steps;

[0060] The Bellman error is selected as the damage function of the main network, the difference between the predicted Q value and the target Q value is calculated, and the parameters of the main network are updated using the Adam optimizer for gradient descent optimization.

[0061] After training is completed, the main network can give the optimal action based on the current state and generate an intelligent scheduling strategy.

[0062] As a preferred solution of the train intelligent control method based on big data analysis of the present invention, wherein: the feedback result based on the intelligent scheduling strategy performs a timed update operation on the deep Q network, including:

[0063] Schedule execution according to the intelligent scheduling strategy generated by the main network, and collect new status and rewards after execution;

[0064] The current state, optimal action, new state, and reward are stored in the experience buffer pool. An update interval is set. Every update interval, the data generated in the previous update interval is sampled from the experience buffer pool to update the main network weights.

[0065] In a second aspect, the present invention provides a train intelligent control system based on big data analysis, comprising:

[0066] The data acquisition module collects train operation data in real time through distributed traffic sensors and performs unified spatiotemporal mapping and normalization processing to build a standardized traffic feature matrix;

[0067] The anomaly detection module builds a graph structure based on the train track network and uses a two-stream attention embedding model to predict the abnormal probability of node status and classify abnormal nodes;

[0068] The state prediction module combines the attention mechanism with the two-layer temporal modeling mechanism to build a dynamic graph neural network model to predict the future state evolution trend of cellular nodes;

[0069] The impact assessment module calculates the propagation impact of abnormal nodes on the neighborhood status through a dynamic weight propagation formula and a bidirectional nested feature extraction model, and outputs a node-level abnormal impact index;

[0070] The intelligent scheduling decision module constructs the state space and action space based on the traffic state graph, and uses the deep Q network with a dual network structure to optimize the train scheduling strategy and generate real-time scheduling actions;

[0071] The feedback learning module regularly updates the deep Q network strategy based on the scheduling execution feedback, realizing the dynamic self-evolution of the scheduling strategy.

[0072] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the train intelligent control method based on big data analysis as described in the first aspect of the present invention is implemented.

[0073] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the train intelligent control method based on big data analysis as described in the first aspect of the present invention.

[0074] The beneficial effects of the present invention are as follows: by constructing a traffic feature matrix, a structured expression of multi-dimensional traffic sensor data is achieved, providing a unified input for downstream modeling; a graph embedding network with a dual-stream attention mechanism is used to accurately identify abnormal nodes in the rail system, greatly improving the real-time and accuracy of anomaly detection; a dynamic graph neural network model is combined with local and global time modeling mechanisms to accurately predict the future state of nodes; a bidirectional nested model and a propagation weight matrix are jointly modeled to achieve a quantitative assessment of the diffusion path and intensity of abnormal events, thereby improving the responsiveness of scheduling strategies to systemic risks; and by constructing a state-action space and using a dual-network DQN to optimize scheduling behavior, strategy self-learning and optimization of multi-objective scheduling decisions are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0076] Figure 1 This is a flow chart of the train intelligent control method based on big data analysis in Example 1.

[0077] Figure 2 This is a structural diagram of the train intelligent control system based on big data analysis in Example 1. DETAILED DESCRIPTION

[0078] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0079] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0080] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0081] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, provides a train intelligent control method based on big data analysis, comprising the following steps:

[0082] S1. Obtain raw train traffic data through the traffic sensor network, extract spatiotemporal features through road network division and perform preprocessing to construct a traffic feature matrix.

[0083] Specifically, a train traffic sensor network is deployed to collect train operation status data, track usage status data, and train travel trajectory data to construct an original data set;

[0084] The original data set is cleaned using the adaptive time window method, and train location, travel speed, trajectory path, passing time, train density, and acceleration fluctuation data are extracted to form a train travel data set.

[0085] Based on the train travel dataset, the road network is divided into hexagonal honeycomb grids. For each time window, the spatial distribution characteristics of train tracks, the time-varying characteristics of train flow, and the changing characteristics of train travel time in the honeycomb unit are extracted. After normalization, the data are aggregated into the corresponding honeycomb unit to construct the traffic feature matrix.

[0086] By deploying a traffic sensor network and collecting data on train operation status, track usage status, and trajectory, the road network is spatially divided using a hexagonal honeycomb grid. Traffic characteristics at different time periods are aggregated and normalized to construct a unified structured traffic characteristic matrix. Complex data is converted into an easy-to-process structured form, facilitating the input and calculation of subsequent models. This achieves standardized modeling of complex train operation status, effectively improving the fusion and processing efficiency of multi-source traffic data, and providing a unified input basis for subsequent track topology mapping, status modeling, and scheduling control. Ultimately, this approach achieves the beneficial effect of enhancing the structural consistency, spatial expression accuracy, and computational scalability of train operation status data.

[0087] S2. Construct a train track topology diagram and use a two-stream attention embedding network to build an anomaly detection model to detect abnormal events and divide abnormal nodes by probability.

[0088] Specifically, a train track topology map is constructed based on the traffic feature matrix. Cellular units are set as nodes, and the traffic feature matrix carried by each cell node is used as the node attribute. Edges are added between directly connected cell nodes with continuous track segments.

[0089] Use a two-stream attention graph embedding network to build an anomaly detection model, including an input layer, a two-stream attention mechanism, an embedding layer, and an output layer;

[0090] The input layer receives a set of nodes in the train track topology graph. The input features of each node are divided into spatial feature vectors and temporal feature vectors.

[0091] The spatial feature vector refers to the spatial distribution feature of the train track;

[0092] The time feature vector includes the time-varying feature of train flow and the time-varying feature of train travel;

[0093] Generate an adjacency matrix A(i) for each node i based on the edge connection relationship in the train track topology graph structure;

[0094] The dual-stream attention mechanism calculates the spatial feature attention weight and the temporal feature attention weight for each neighboring node j∈A(i) of node i, which can be expressed as:

[0095]

[0096] in, and Represents the spatial feature attention weight and temporal feature attention weight between nodes i and j, respectively, F s (i) and F s (j) distribution represents the spatial feature vector of nodes i and j, F t (i) and F t (j) represents the time feature vector of node i and node j respectively, and MLP represents multi-layer perceptron;

[0097] The spatial enhancement features and temporal enhancement features are generated by attention weights, which are expressed as:

[0098]

[0099] Among them, F sd (i) and F td (i) represent spatial enhancement features and temporal enhancement features respectively;

[0100] The spatial enhancement features and temporal enhancement features are fused to generate the final feature representation F of the node e (i);

[0101] The embedding layer performs graph convolution calculation on the final feature representation to generate the embedded representation of the honeycomb node, which is expressed as:

[0102] H (k+1) =σ(D -1 AH (k) W (k) +b (k) );

[0103] Among them, H (k) represents the k-th layer feature matrix, H (0) =F e (i), A represents the adjacency matrix, D -1 represents the normalization operation of the adjacency matrix, σ represents the nonlinear activation function, W (k) and b (k) Represent the weight and bias of graph convolution respectively;

[0104] The output layer uses the final embedding feature H of the node (K) , calculate the abnormal probability of the cellular node, expressed as:

[0105] P ab (i)=σ(W out H (K) +b out );

[0106] Among them, P ab (i) represents the abnormal probability of node i, σ represents the Sigmoid activation function, W out and b out Represent the weight and bias of the output layer respectively;

[0107] Use the training set to train the model, select the cross entropy loss function to calculate the difference between the classification result and the true label, use the Adam optimizer to perform gradient descent optimization, update the model parameters, and stop iterating and output the model if the model loss no longer decreases significantly during the continuous iteration process;

[0108] Use the trained model to generate the abnormal probability of each node in the train track topology map, set a threshold based on historical experience, and classify a node as an abnormal node when the abnormal probability of the node is greater than the threshold.

[0109] A topology map is constructed based on the traffic feature matrix, and anomaly detection of nodes is performed using a dual-stream attention graph embedding network. By training the model and setting thresholds to divide abnormal nodes, abnormal nodes in the track topology map can be discovered in a timely manner, providing early warning information for subsequent predictions and scheduling, and avoiding potential risks and accidents. The dual-stream attention mechanism processes spatial features and temporal features respectively, and generates enhanced features through attention weights. The enhanced features are fused to generate highly expressive node representation vectors, which can more accurately identify abnormal events, realize the modeling of complex spatiotemporal correlation patterns between nodes, and the dynamic identification of abnormal patterns. Compared with traditional methods, the accuracy of anomaly detection is significantly improved. The model is optimized through the training set so that it can adapt to different track topologies and operating states. It has good generalization ability and can be applied to a variety of scenarios. Rapid detection of abnormal nodes provides timely risk warnings for train scheduling and operation management, which helps to take measures in advance to ensure the safety of train operation.

[0110] S3. Combining the attention mechanism with two-layer temporal modeling, a dynamic graph neural network prediction model is constructed to generate the predicted status of cellular nodes.

[0111] Specifically, we combine the attention mechanism with the two-layer temporal modeling to build a dynamic graph neural network prediction model, which includes an input layer, a graph convolution layer, a two-layer temporal modeling module, and an output layer.

[0112] The input layer inputs the node set and adjacency matrix in the train track topology graph;

[0113] The graph convolution layer aggregates information of the cellular nodes at each moment through the attention mechanism and generates the aggregated node representation H t ;

[0114] The two-layer time modeling module includes local time modeling and global time modeling;

[0115] The local time modeling uses causal convolution to H t Capture the nearest neighbor time dependency and generate local information Z t ;

[0116] The global time modeling adopts the self-attention mechanism to calculate H t Full temporal context representation, generating global information S t ;

[0117] The local information is weightedly fused with the global information to generate the final representation U t ;

[0118] The output layer uses a feedforward neural network with shared parameters to t Decode and predict the time series to generate the next time step prediction state of the train position, speed, acceleration, running direction and train density in each cellular node;

[0119] Use the training set to train the prediction model, update the model parameters, and output the model after training is completed;

[0120] Use the trained model to generate the predicted state of each cell node.

[0121] A dynamic graph neural network prediction model is constructed, including graph convolution layers and double-layer time modeling modules. By integrating time information through local time modeling and global time modeling, the next moment state of the cellular node is predicted, and the train operation state at future moments is predicted, providing forward-looking information for train scheduling, helping to plan and adjust the operation plan in advance. The combination of attention mechanism and double-layer time modeling can fully capture the spatiotemporal dependencies of nodes, and significantly improves the accuracy of prediction compared with single time modeling or spatial modeling methods. The combination of local time modeling and global time modeling takes into account both recent dynamic changes and long-term trends, enabling the model to adapt to different time scales and be suitable for complex train operation scenarios. Finally, a detailed predicted state of each cellular node is generated, including position, speed, acceleration, etc., which provides a comprehensive reference basis for train scheduling, helps to optimize train operation plans and improve operation efficiency.

[0122] S4. Calculate the propagation impact intensity of abnormal nodes based on the dynamic weight propagation formula, generate a dynamic abnormal propagation weight matrix, build a bidirectional nested model based on Bi-GRU and neighborhood graph convolution to extract spatiotemporal embedding features, and calculate the abnormal impact index by combining the spatiotemporal embedding features and the dynamic abnormal propagation weight matrix.

[0123] Specifically, a dynamic propagation weight formula is defined. Based on historical data and real-time data, the propagation impact intensity of abnormal nodes is calculated to generate a dynamic abnormal propagation weight matrix, which is expressed as:

[0124]

[0125] in, represents the abnormal propagation weight between cellular nodes i and j at time t, represents the current real-time traffic flow between cellular nodes i and j, represents the historical maximum traffic flow between cellular nodes i and j, V avg represents the real-time average speed of the train between cellular nodes i and j, V max represents the maximum speed limit between cellular nodes i and j, represents the exponential decay term;

[0126] Build a bidirectional nested model based on Bi-GRU and neighborhood graph convolution;

[0127] Extracting train running status feature time series X based on spatiotemporal traffic feature matrix i (t);

[0128] Use Bi-GRU to extract the propagation dynamics in the time series and generate the propagation state embedding, which is expressed as:

[0129] H i (t) = Bi-GRU(X i (t));

[0130] Among them, Bi-GRU represents a bidirectional gated recurrent unit, H i (t) represents the propagation state embedding of node i at time t;

[0131] Based on the propagation state, the spatial dependency is captured by the neighborhood graph convolution to generate a spatial embedding, which is expressed as:

[0132]

[0133] Among them, Z i (t) represents the spatial embedding of node i at time t, and A(i) represents the set of direct neighbor nodes of node i;

[0134] Use the training set to train the bidirectional nested model and update the model parameters. If the model loss no longer decreases significantly during the continuous iteration, stop iterating and output the model.

[0135] Combining the dynamic anomaly propagation weight matrix and the trained bidirectional nested model, the anomaly impact indicators of the abnormal nodes in the train track topology graph are calculated, including the propagation index, impact range, and efficiency reduction ratio, which are expressed as:

[0136]

[0137] Among them, PI represents the propagation index, R im represents the impact range, ELR represents the efficiency reduction ratio, i represents the abnormal node index, N represents the total number of abnormal nodes currently detected by the anomaly detection model, j represents the neighboring node of node i, ΔR ij represents the connection length of the track segment between cellular nodes i and j, which is extracted from the track topology graph. represents the average traffic between cellular nodes i and j, Represents the current traffic between cellular nodes i and j.

[0138] A dynamic propagation weight formula is defined to calculate the propagation impact intensity of abnormal nodes, a bidirectional nested model is constructed to extract spatiotemporal embedding features, and the weight matrix is combined to calculate the abnormal impact index to quantify the impact of abnormal nodes on the entire track topology. This provides a quantitative basis for subsequent scheduling decisions and helps evaluate the severity and impact range of abnormal events. Through the dynamic abnormal propagation weight matrix and bidirectional nested model, the propagation impact intensity of abnormal nodes can be accurately calculated, and abstract abnormal events can be converted into specific quantitative indicators for easy understanding and evaluation. The combination of Bi-GRU and neighborhood graph convolution can simultaneously capture the dynamic changes of time series and the dependencies of spatial structures, enabling the model to comprehensively consider the propagation characteristics of abnormal events in the spatiotemporal dimensions. Abnormal impact indicators include propagation index, impact range, and efficiency reduction ratio, etc., which provide multi-dimensional quantitative information for scheduling decisions, help to formulate more accurate and effective response measures, and reduce the impact of abnormal events on train operations.

[0139] S5. Construct the state space and action space based on the state graph, and use the DQN with a dual network architecture to optimize train scheduling and generate intelligent scheduling strategies.

[0140] Specifically, the real-time traffic feature matrix at each moment, the predicted state of each cellular node, the adjacency matrix of the track topology graph, and the abnormal impact index of the abnormal node are represented as a state graph;

[0141] Based on the state diagram, construct the state space S t ;

[0142] Based on the train track topology structure and historical scheduling experience, the action space A is constructed. t ;

[0143] Design a multi-objective optimization model, taking minimizing the propagation index, reducing the impact range, and improving efficiency as optimization goals, and design a reward function;

[0144] A deep Q-network using a dual-network architecture consisting of a master network and a target network is used to find the optimal scheduling strategy through reinforcement learning.

[0145] In the current state, the best action is selected by predicting the Q value through the main network based on the ∈-greedy strategy. The main network parameter is θ main ;

[0146] After executing the action, the reward and next state diagram are obtained;

[0147] The target network calculates the target Q value based on the reward, the next state diagram, and the action selected by the main network, and synchronizes the main network parameters to the target network every k steps;

[0148] The Bellman error is selected as the damage function of the main network, the difference between the predicted Q value and the target Q value is calculated, and the parameters of the main network are updated using the Adam optimizer for gradient descent optimization.

[0149] After training is completed, the main network can give the optimal action based on the current state and generate an intelligent scheduling strategy.

[0150] Construct state diagrams, state spaces, and action spaces, design multi-objective optimization models and reward functions, use DQN with a dual-network architecture for reinforcement learning, generate intelligent scheduling strategies, and automatically generate the optimal train scheduling strategy based on the current traffic status and abnormal conditions, realizing intelligent management and optimized scheduling of train operations. Through the reinforcement learning algorithm, the model can automatically learn the optimal scheduling strategy based on the real-time status. Compared with traditional manual scheduling or fixed-rule scheduling, it has higher efficiency and adaptability, and can effectively improve the overall performance of train operation. The designed multi-objective optimization model comprehensively considers multiple objectives such as propagation index, impact range, and efficiency improvement, so that the scheduling strategy can achieve balance and optimization in multiple aspects, avoiding the one-sidedness that may be caused by single-objective optimization. The DQN with a dual-network architecture can update and adjust strategies in real time to adapt to dynamically changing traffic environments and abnormal conditions, ensuring that the scheduling strategy is always in the optimal state, and improving the stability and reliability of train operation.

[0151] S6. Perform a timed update operation on the deep Q network based on the feedback results of the intelligent scheduling strategy.

[0152] Specifically, the scheduling is performed according to the intelligent scheduling strategy generated by the main network, and the new status and rewards are collected after execution;

[0153] The current state, optimal action, new state, and reward are stored in the experience buffer pool. An update interval is set. Every update interval, the data generated in the previous update interval is sampled from the experience buffer pool to update the main network weights.

[0154] According to the execution results of the intelligent scheduling strategy, the relevant data is stored in the experience buffer pool, and data is regularly sampled from the buffer pool to update the main network weights. The parameters of the deep Q network are continuously optimized through the feedback mechanism, so that it can better adapt to the actual operating environment and further improve the quality and performance of the scheduling strategy. Regular updates are performed based on the feedback results, so that the model can continuously learn and adjust, gradually improve the accuracy and adaptability of the scheduling strategy, and achieve continuous optimization of the strategy. By storing and sampling data in the experience buffer pool, excessive fluctuations that may be caused by a single update are avoided, making the model update process more stable, improving the model's convergence speed and stability, and the model can automatically adjust according to feedback information in actual operation, with stronger adaptive capabilities, and can better cope with complex operating environments and dynamically changing traffic needs.

[0155] This embodiment also provides a train intelligent control system based on big data analysis, including:

[0156] The data acquisition module collects train operation data in real time through distributed traffic sensors and performs unified spatiotemporal mapping and normalization processing to build a standardized traffic feature matrix;

[0157] The anomaly detection module builds a graph structure based on the train track network and uses a two-stream attention embedding model to predict the abnormal probability of node status and classify abnormal nodes;

[0158] The state prediction module combines the attention mechanism with the two-layer temporal modeling mechanism to build a dynamic graph neural network model to predict the future state evolution trend of cellular nodes;

[0159] The impact assessment module calculates the propagation impact of abnormal nodes on the neighborhood status through a dynamic weight propagation formula and a bidirectional nested feature extraction model, and outputs a node-level abnormal impact index;

[0160] The intelligent scheduling decision module constructs the state space and action space based on the traffic state graph, and uses the deep Q network with a dual network structure to optimize the train scheduling strategy and generate real-time scheduling actions;

[0161] The feedback learning module regularly updates the deep Q network strategy based on the scheduling execution feedback, realizing the dynamic self-evolution of the scheduling strategy.

[0162] This embodiment also provides a computer device suitable for the case of a train intelligent control method based on big data analysis, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the train intelligent control method based on big data analysis proposed in the above embodiment.

[0163] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0164] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the train intelligent control method based on big data analysis proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, disk or optical disk.

[0165] In summary, the present invention constructs a traffic feature matrix; constructs a topological graph based on the traffic feature matrix, uses a dual-stream attention graph embedding network to detect node anomalies and divide abnormal nodes; constructs a dynamic graph neural network prediction model, fuses time information through local time modeling and global time modeling, and predicts the next moment state of the cellular node; defines a dynamic propagation weight formula and calculates the anomaly impact index; constructs a state graph based on the predicted state and the anomaly impact index, designs a multi-objective optimization model and reward function, uses a dual-network architecture DQN for reinforcement learning, generates an intelligent scheduling strategy, and realizes intelligent management and optimized scheduling of train operation.

[0166] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A train intelligent control method based on big data analysis, characterized by: include, The original train traffic data is obtained through the traffic sensor network, and the spatiotemporal features are extracted and preprocessed by road network division to construct a traffic feature matrix. Construct a train track topology diagram and use a two-stream attention embedding network to build an anomaly detection model to detect abnormal events and classify abnormal nodes by probability. Combining the attention mechanism with two-layer temporal modeling to build a dynamic graph neural network prediction model to generate the predicted state of cellular nodes; The propagation impact strength of abnormal nodes is calculated based on the dynamic weight propagation formula, and a dynamic abnormal propagation weight matrix is generated. A bidirectional nested model is constructed based on Bi-GRU and neighborhood graph convolution to extract spatiotemporal embedding features. The abnormal impact index is calculated by combining the spatiotemporal embedding features and the dynamic abnormal propagation weight matrix. Construct state space and action space based on state graph, and use DQN with dual network architecture to optimize train scheduling and generate intelligent scheduling strategy; Based on the feedback results of the intelligent scheduling strategy, the deep Q network is regularly updated.

2. The train intelligent control method based on big data analysis according to claim 1, characterized in that: The method obtains the original train traffic data through the traffic sensor network, extracts the spatiotemporal features through road network division and performs preprocessing, and constructs the traffic feature matrix, including: Deploy a train traffic sensor network to collect train operation status data, track usage status data, and train trajectory data to construct an original data set; The original data set is cleaned using the adaptive time window method, and train location, travel speed, trajectory path, passing time, train density, and acceleration fluctuation data are extracted to form a train travel data set. Based on the train travel dataset, the road network is divided into hexagonal honeycomb grids. For each time window, the spatial distribution characteristics of train tracks, the time-varying characteristics of train flow, and the changing characteristics of train travel time in the honeycomb unit are extracted. After normalization, the data are aggregated into the corresponding honeycomb unit to construct the traffic feature matrix.

3. The train intelligent control method based on big data analysis according to claim 2, characterized in that: The train track topology diagram is constructed, and the anomaly detection model is constructed using a dual-stream attention graph embedding network to detect abnormal events and divide abnormal nodes by probability, including: A train track topology map is constructed based on the traffic feature matrix. Cellular units are set as nodes, and the traffic feature matrix carried by each cell node is used as the node attribute. Edges are added between directly connected cell nodes with continuous track segments. Use a two-stream attention graph embedding network to build an anomaly detection model, including an input layer, a two-stream attention mechanism, an embedding layer, and an output layer; The input layer receives a set of nodes in the train track topology graph. The input features of each node are divided into spatial feature vectors and temporal feature vectors. The spatial feature vector refers to the spatial distribution feature of the train track; The time feature vector includes the time-varying feature of train flow and the time-varying feature of train travel; Generate an adjacency matrix A(i) for each node i based on the edge connection relationship in the train track topology graph structure; The dual-stream attention mechanism calculates the spatial feature attention weight and the temporal feature attention weight for each neighboring node j∈A(i) of node i; Generate spatial enhancement features and temporal enhancement features through attention weights; The spatial enhancement features and temporal enhancement features are fused to generate the final feature representation F of the node e (i); The embedding layer performs graph convolution calculation on the final feature representation to generate the embedded representation of the honeycomb node; The output layer uses the final embedding feature H of the node (K) , calculate the abnormal probability of cellular nodes; Use the training set to train the model, select the cross entropy loss function to calculate the difference between the classification result and the true label, use the Adam optimizer to perform gradient descent optimization, update the model parameters, and stop iterating and output the model if the model loss no longer decreases significantly during the continuous iteration process; Use the trained model to generate the abnormal probability of each node in the train track topology map, set a threshold based on historical experience, and classify a node as an abnormal node when the abnormal probability of the node is greater than the threshold.

4. The train intelligent control method based on big data analysis according to claim 3, characterized in that: The dynamic graph neural network prediction model is constructed by combining the attention mechanism with the double-layer time modeling to generate the predicted state of the cellular node, including: Combining the attention mechanism with two-layer temporal modeling, a dynamic graph neural network prediction model is constructed, including an input layer, a graph convolution layer, a two-layer temporal modeling module, and an output layer. The input layer inputs the node set and adjacency matrix in the train track topology graph; The graph convolution layer aggregates information of the cellular nodes at each moment through the attention mechanism and generates the aggregated node representation H t ; The two-layer time modeling module includes local time modeling and global time modeling; The local time modeling uses causal convolution to H t Capture the nearest neighbor time dependency and generate local information Z t ; The global time modeling adopts the self-attention mechanism to calculate H t Full temporal context representation, generating global information S t ; The local information is weightedly fused with the global information to generate the final representation U t ; The output layer uses a feedforward neural network with shared parameters to t Decode and predict the time series to generate the next time step prediction state of the train position, speed, acceleration, running direction and train density in each cellular node; Use the training set to train the prediction model, update the model parameters, and output the model after training is completed; Use the trained model to generate the predicted state of each cell node.

5. The train intelligent control method based on big data analysis according to claim 4, characterized in that: The method calculates the propagation influence strength of abnormal nodes based on the dynamic weight propagation formula, generates a dynamic abnormal propagation weight matrix, constructs a bidirectional nested model based on Bi-GRU and neighborhood graph convolution to extract spatiotemporal embedding features, and calculates abnormal impact indicators by combining spatiotemporal embedding features and the dynamic abnormal propagation weight matrix, including: Define the dynamic propagation weight formula, calculate the propagation impact intensity of abnormal nodes based on historical data and real-time data, and generate a dynamic abnormal propagation weight matrix; Build a bidirectional nested model based on Bi-GRU and neighborhood graph convolution; Extracting train running status feature time series X based on spatiotemporal traffic feature matrix i (t); Use Bi-GRU to extract the propagation dynamics in the time series and generate propagation state embeddings; Based on the propagation state, spatial dependencies are captured through neighborhood graph convolution to generate spatial embeddings; Use the training set to train the bidirectional nested model and update the model parameters. If the model loss no longer decreases significantly during the continuous iteration, stop iterating and output the model. Combining the dynamic anomaly propagation weight matrix and the trained bidirectional nested model, the anomaly impact indicators of abnormal nodes in the train track topology graph are calculated, including the propagation index, impact range, and efficiency reduction ratio.

6. The train intelligent control method based on big data analysis according to claim 5, characterized in that: The state space and action space are constructed based on the state graph, and the DQN with dual network architecture is used to optimize train scheduling and generate intelligent scheduling strategies. include, The real-time traffic feature matrix at each moment, the predicted state of each cellular node, the adjacency matrix of the track topology graph, and the abnormal impact index of the abnormal node are represented as a state graph; Based on the state diagram, construct the state space S t ; Based on the train track topology structure and historical scheduling experience, the action space A is constructed. t ; Design a multi-objective optimization model, taking minimizing the propagation index, reducing the impact range, and improving efficiency as optimization goals, and design a reward function; A deep Q-network using a dual-network architecture consisting of a master network and a target network is used to find the optimal scheduling strategy through reinforcement learning. In the current state, the best action is selected by predicting the Q value through the main network based on the ∈-greedy strategy. The main network parameter is θ main ; After executing the action, the reward and next state diagram are obtained; The target network calculates the target Q value based on the reward, the next state diagram, and the action selected by the main network, and synchronizes the main network parameters to the target network every k steps; The Bellman error is selected as the damage function of the main network, the difference between the predicted Q value and the target Q value is calculated, and the parameters of the main network are updated using the Adam optimizer for gradient descent optimization. After training is completed, the main network can give the optimal action based on the current state and generate an intelligent scheduling strategy.

7. The train intelligent control method based on big data analysis according to claim 6, characterized in that: The feedback result based on the intelligent scheduling strategy performs a timed update operation on the deep Q network, including: Schedule execution according to the intelligent scheduling strategy generated by the main network, and collect new status and rewards after execution; The current state, optimal action, new state, and reward are stored in the experience buffer pool. An update interval is set. Every update interval, the data generated in the previous update interval is sampled from the experience buffer pool to update the main network weights.

8. A train intelligent control system based on big data analysis, based on the train intelligent control method based on big data analysis according to any one of claims 1 to 7, characterized in that: include, The data acquisition module collects train operation data in real time through distributed traffic sensors and performs unified spatiotemporal mapping and normalization processing to build a standardized traffic feature matrix; The anomaly detection module builds a graph structure based on the train track network and uses a two-stream attention embedding model to predict the abnormal probability of node status and classify abnormal nodes; The state prediction module combines the attention mechanism with the two-layer temporal modeling mechanism to build a dynamic graph neural network model to predict the future state evolution trend of cellular nodes; The impact assessment module calculates the propagation impact of abnormal nodes on the neighborhood status through a dynamic weight propagation formula and a bidirectional nested feature extraction model, and outputs a node-level abnormal impact index; The intelligent scheduling decision module constructs the state space and action space based on the traffic state graph, and uses the deep Q network with a dual network structure to optimize the train scheduling strategy and generate real-time scheduling actions; The feedback learning module regularly updates the deep Q network strategy based on the scheduling execution feedback, realizing the dynamic self-evolution of the scheduling strategy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the train intelligent control method based on big data analysis described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the train intelligent control method based on big data analysis according to any one of claims 1 to 7 are implemented.

Images