Heavy haul train group operation control sensing method and system

Abstract

The present disclosure relates to the technical field of rail transit operation control, and particularly relates to a heavy-haul train group operation control sensing method and system. Train group operation state information is collected through different sensors; the state information is converted into a state graph, in which nodes represent sensor sampling points, edges represent time difference values between the latest sampling points of the sensors, and an adjacent matrix represents the connection relationship between the edges of the nodes; a graph convolutional neural network is established according to the state graph, which is used to estimate the alarm level and alarm processing time of the train group operation state; the graph convolutional neural network is trained according to historical state graphs, and a trained graph convolutional neural network is obtained; the real-time collected state graph is input into the graph convolutional neural network, and the operation control sensing supervision of the train group operation state is performed. Through technical means, the automatic and holographic sensing and monitoring capability of the heavy-haul railway line train operation safety risk is enhanced, which provides a guarantee for realizing the safe operation of the heavy-haul train group.

Description

Technical Field

[0001] This disclosure relates to the field of rail transit operation control technology, and in particular to a sensing method and system for heavy-load train group operation control. Background Technology

[0002] The existing risk monitoring along heavy-haul railway lines mainly relies on human-based security measures supplemented by technological ones. (1) It relies heavily on personnel monitoring, has a large workload for personnel monitoring, and carries heavy safety responsibilities, making it difficult to cover various complex scenarios on site around the clock; (2) The technical security measures are limited, the technical means are relatively simple and the functions are simple, and the monitoring is mainly carried out in key areas; (3) The existing monitoring information is incomplete and the information integration is low. The monitoring results are managed separately by personnel from different departments. At present, a comprehensive and holographic perception and monitoring system has not yet been formed. (4) The existing risk monitoring is limited to the perception and monitoring level. It is necessary to arrange on-duty personnel to judge the degree of risk of the event based on the monitoring results, and then take manual response measures.

[0003] In summary, existing railway operation environment safety risk monitoring systems offer limited monitoring information and rely on simplistic technical methods. Furthermore, they require on-duty personnel to assess the risk level of events based on monitoring results, necessitating manual responses. For example, natural disaster and foreign object intrusion monitoring systems can only provide real-time monitoring, early warning, and alarms for conditions such as wind, rain, snow, earthquakes, and foreign object intrusion across railway lines. They lack effective monitoring coverage for special natural disasters such as landslides, rockfalls, and debris flows on heavy-haul railways, failing to meet the disaster monitoring needs of heavy-haul railways and effectively ensure operational safety. Summary of the Invention

[0004] To address the aforementioned issues, this disclosure provides a sensing method and system for controlling the operation of heavy-haul train groups.

[0005] A first aspect is a sensing method for controlling the operation of heavy-haul train groups, the method comprising: Train group operation status information is collected using different sensors; The state information is transformed into a state graph, where nodes represent sensor sampling points, edges represent the time difference between the latest sensor sampling points, and the proximity matrix represents the connection relationship between the edges of the nodes. A graph convolutional neural network is built based on the state diagram to predict the alarm level and alarm processing time of the train group operation status. Train a graph convolutional neural network based on the historical state graph to obtain a trained graph convolutional neural network; The real-time collected state graphs are input into a graph convolutional neural network to perform operational control perception and supervision of the train group's operating status.

[0006] Furthermore, the state information is transformed into a state diagram, including: Let G represent the state diagram, and let V be the node of the diagram. V represents the sensor data, including: sampling number, sampling location, sampling time, sensor type, sensor data, and whether an alarm is triggered. The edge E of graph G represents the time difference between the latest sampled data from the sensor; The proximity matrix A of graph G represents the connection relationship between edges between nodes.

[0007] Furthermore, the graph G is (V,E,A), where node V is an N×F matrix, N is the number of sensor nodes, and F is the number of data features of the sensor; edge E is an N×N×P matrix, and P is the number of edge data features.

[0008] Furthermore, the information propagation model of sensor nodes in a graph convolutional neural network is as follows:

[0009] In the formula, This represents the output feature of sensor node i in the (l+1)th layer of the graph convolutional neural network. and This represents the output of sensor node i and sensor node j at layer l. Let R be the activation function, and R be the set of relationships between sensor nodes. Let be the set of sensor nodes with relation r among the one-hop neighbors of sensor node i. In this disclosure, relation r indicates the existence of an association. The value is The regularity constant, Let be the weight parameter matrix of neighboring sensor nodes with relation r in the l-th layer of the network. Let be the weight parameter matrix of sensor node i in the l-th layer of the network.

[0010] Furthermore, the propagation model of edge information in graph convolutional neural networks is as follows:

[0011] In the formula, Let i be the neighboring sensor nodes of sensor node i. This is the Softmax regularization function. and These are the node weight parameter matrix and the edge weight parameter matrix, respectively. For edge feature vectors, Let be the node feature vector, and || be the concatenation operation. This is the output feature vector after aggregating edge feature information.

[0012] Furthermore, a graph convolutional neural network is constructed based on the state diagram, including: Graph Convolutional Neural Networks consist of an L-layer Graph Convolutional Network (GCN) and a multi-layer Perceptron (MLP). The activation functions for the GCN and LSTM layers are ReLU (Rectified Linear Unit) functions, while the activation function for the MLP layers is the tanh activation function. The output of the convolutional neural network is O= , These represent the alarm number, alarm location, alarm time, alarm type, alarm level, and alarm processing time, respectively. The alarm number is arranged in the order of alarms. The alarm location and alarm time are the location of the sensor node that triggered the alarm and the sampling time at the time of the alarm, respectively. The alarm level is the estimated level after processing by the graph convolutional neural network, and the alarm processing time is the time required to process the alarm as estimated by the graph convolutional neural network.

[0013] Furthermore, the loss function for graph convolutional neural networks is as follows:

[0014] In the formula, Weights for alarm types set based on experience or iterative learning.

[0015] Secondly, a heavy-haul train group operation control and sensing system includes: a data acquisition unit, a graph conversion unit, a model building unit, and a sensing and supervision unit; The data acquisition unit is used to collect train group operation status information through different sensors; The graph conversion unit is used to convert state information into a state graph. In the graph, nodes represent sensor sampling points, edges represent the time difference between the latest sampling points of the sensor, and the proximity matrix A represents the connection relationship between the edges of the nodes. The model building unit is used to build a graph convolutional neural network based on the state diagram, which is used to predict the alarm level and alarm processing time of the train group operation status. The model building unit is also used to train a graph convolutional neural network based on the historical state graph to obtain a trained graph convolutional neural network. The perception and supervision unit is used to input the real-time acquired state graph into the graph convolutional neural network to perform operation control perception and supervision of the train group's operating status.

[0016] Thirdly, an electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, which stores computer programs; When a processor executes a computer program stored in memory, it implements the aforementioned method for controlling and sensing the operation of heavy-load train groups.

[0017] Fourthly, a computer-readable storage medium stores a computer program that, when executed by a processor, implements the aforementioned method for sensing and controlling the operation of heavy-load train groups.

[0018] This disclosure has at least the following beneficial effects: This disclosure addresses the complex, all-encompassing environment of heavy-haul train group operations, short-distance train tracking, and the development needs of automated train control systems. It enhances the automated and holographic perception and monitoring capabilities of train operation safety risks along heavy-haul railway lines through technical means. It integrates advanced technologies such as big data mining and digital twins to conduct risk analysis, safety situation simulation, and decision-making based on line perception information, thereby providing a guarantee for the safe operation of heavy-haul train groups.

[0019] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description and the accompanying drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the operation control sensing method according to an embodiment of the present disclosure; Figure 2 This is a schematic diagram of the operation control and sensing system architecture according to an embodiment of the present disclosure; Figure 3 The image shows a schematic diagram of a convolutional neural network structure. Figure 4 This is a schematic diagram of the electronic device structure; Figure 5 This is a schematic diagram of the system hardware structure. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0023] In response to the demands of complex all-domain environments, short-distance train tracking, and the development of automated train control systems, there is an urgent need to enhance the holographic and automated perception capabilities of train safety operations through technological means, comprehensively improve the comprehensiveness of perception content and the level of cognitive inference, develop holographic perception and three-dimensional safety assurance for railway lines, and lead the research and development of active safety protection technologies for advanced global rail transit autonomous operation systems.

[0024] like Figure 1 As shown, a sensing method for controlling the operation of heavy-haul train groups includes: S101 collects train group operation status information through different sensors; S102, convert the state information into a state diagram. In the diagram, nodes represent sensor sampling points, edges represent the time difference between the latest sensor sampling points, and the proximity matrix represents the connection relationship between the edges of the nodes. S103, A graph convolutional neural network is established based on the state diagram to predict the alarm level and alarm processing time of the train group's operating status. S104, Train the graph convolutional neural network based on the historical state graph to obtain the trained graph convolutional neural network; S105 inputs the real-time collected state graph into a graph convolutional neural network to perform operational control perception and supervision of the train group's operating status.

[0025] The specific implementation details are as follows: When processing actual sensing data, the sampling frequencies of different sensing terminals are inconsistent, resulting in asynchronous sensing data. At the same time, when a sensor generates an alarm, information from other related sensors may not be collected synchronously. It is necessary to use historical sensor data to determine the correlation of sensing information and the degree of risk of the event.

[0026] Therefore, for the early warning analysis and processing of multi-source sensing data, this disclosure proposes a graph neural network prediction and classification algorithm for asynchronous spatiotemporal correlation analysis, which can process multi-source asynchronous sensing data and alarm data to determine the degree of event risk.

[0027] The graph G of the sensing data processing model is defined with nodes V, containing sensor data (number, location, sampling time, sensor type, sensor data, alarm status, alarm level, and estimated alarm processing time). The edges of graph G are defined with edges E, containing (distance between sensors, time difference between the latest sensor sampling data). The proximity matrix A of graph G represents the connection relationships between nodes. This disclosure obtains the positional and temporal relationships between data with the closest sensor sampling times by processing edge E, thereby obtaining asynchronous spatiotemporal relationships and correlations between sensor data without requiring sensor synchronization.

[0028] Then graph G is (V,E,A), where the nodes are an N×F matrix, N is the number of sensor nodes, and in this disclosure, F=8 represents the number of sensor data features. E is an N×N×P matrix, and in this disclosure, P=2 represents the number of edge data features. This represents the j-th channel of the feature vector of the i-th sensor node in graph G. The p-th channel of the feature vector representing the edge between the i-th sensor node and the j-th sensor node in graph G.

[0029] In the graph convolutional neural network disclosed herein, the information propagation model of the sensor nodes is as follows:

[0030] In the formula, This represents the output feature of sensor node i in the (l+1)th layer of the graph convolutional neural network. and This represents the output of sensor node i and sensor node j at layer l. Let R be the activation function, and R be the set of relationships between sensor nodes. Let be the set of sensor nodes with relation r among the one-hop neighbors of sensor node i. In this disclosure, relation r indicates the existence of an association. The value is The regularity constant, Let be the weight parameter matrix of neighboring sensor nodes with relation r in the l-th layer of the network. Let be the weight parameter matrix of sensor node i in the l-th layer of the network.

[0031] The propagation model of edge information is as follows:

[0032] In the formula, Representing an edge eigenvectors, Let represent the feature vector of sensor node i at the t-th graph convolutional layer, M represent the information transfer function, and U represent the sensor node information update function. This represents an information aggregation operation. Specifically, one implementation of the above formula can be:

[0033] In the formula, Let i be the neighboring sensor nodes of sensor node i. This is the Softmax regularization function. and These are the node weight parameter matrix and the edge weight parameter matrix, respectively. For edge feature vectors, Let be the node feature vector, and || be the concatenation operation. This is the output feature vector after aggregating edge feature information.

[0034] In this disclosure, an L-layer graph convolutional neural network can be constructed as needed to... Perform L iterations, then you can stack LSTM and MLP layers to... The process is performed to obtain the final output O, which is (alarm number, alarm location, alarm time, alarm type, alarm level, alarm processing time). The alarm number is arranged in the order of alarms, the alarm location and alarm time are the location of the sensor node that triggered the alarm and the sampling time at the time of the alarm, respectively, the alarm level is the predicted level after processing by the graph convolutional neural network, and the alarm processing time is the time required to process the alarm as predicted by the graph convolutional neural network.

[0035] like Figure 2 As shown, a heavy-haul train group operation control and sensing system includes: a data acquisition unit 201, a graph conversion unit 202, a model building unit 203, and a sensing and supervision unit 204. Data acquisition unit 201 is used to collect train group operation status information through different sensors; The graph conversion unit 202 is used to convert state information into a state graph. In the graph, nodes represent sensor sampling points, edges represent the time difference between the latest sampling points of the sensor, and the proximity matrix A represents the connection relationship between the edges of the nodes. The model building unit 203 is used to build a graph convolutional neural network based on the state diagram, which is used to predict the alarm level and alarm processing time of the train group operation status. The model building unit 203 is also used to train a graph convolutional neural network based on the historical state graph to obtain a trained graph convolutional neural network. The perception and supervision unit 204 is used to input the real-time acquired state graph into the graph convolutional neural network to perform operation control perception and supervision of the train group's operating status.

[0036] In specific implementation, such as Figure 3As shown, the graph convolutional neural network in this disclosure includes an L-layer graph convolutional network (GCN), an optional long short-term memory (LSTM) layer, and a multilayer perceptron (MLP) layer. The activation function of the GCN layer and the LSTM layer is the ReLU (Rectified Linear Unit) function, and the activation function of the MLP layer is the tanh activation function.

[0037] The output result O is (alarm number, alarm location, alarm time, alarm type, alarm level, alarm processing time), set as a variable O= Then the loss function is:

[0038] In the formula, Weights for alarm types set based on experience or iterative learning.

[0039] Network training process: Dataset Construction. First, alarm data for a section of railway within a year was collected, including sensor data (number, location, sampling time, sensor type, sensor data, alarm status, alarm level, and estimated alarm processing time). The sensor data was also processed to obtain edge information, including (distance between sensors and time difference between the latest sensor sampling data). Furthermore, based on expert experience, the network output was configured, including (alarm number, alarm location, alarm time, alarm type, alarm level, and alarm processing time). The (alarm number, alarm location, alarm time, and alarm type) were obtained from the actual alarm data from the sensors; the alarm level was set based on expert experience; and the alarm processing time was obtained by subtracting a random number from the system's maximum required processing time. The range of the random number was also set based on expert experience. The dataset was then divided into training, testing, and validation sets.

[0040] The designed network is trained on the training set, and the internal parameters of the model are updated based on the loss function to obtain the trained network model. Then, the network model is tested and validated using the test set and validation set.

[0041] like Figure 4 As shown, this disclosure provides an electronic device, including a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404. Memory 403 stores computer programs; The processor 401 implements the above method when executing a computer program stored in the memory 403.

[0042] This disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0043] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments; or it may exist independently and not assembled into the device / apparatus. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0044] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0045] To enable those skilled in the art to better understand this disclosure, the principles of this disclosure are explained below in conjunction with the accompanying drawings: Existing methods rely on a single means of sensing and monitoring train operation safety risks, and the assessment of safety risks is mainly based on a single source, lacking cross-verification. This results in high false alarm and false negative rates, requiring manual judgment of event risks. In contrast, the holographic sensing system designed in this disclosure integrates and combines multi-source and multi-dimensional sensing information, and through safety strategies, can significantly improve the reliability of risk warnings.

[0046] This disclosure proposes a graph neural network prediction and classification algorithm for asynchronous spatiotemporal correlation analysis. By processing edge feature information, it obtains the positional and temporal relationships between data with the closest sensor sampling times. Without the need for sensor synchronization, it obtains asynchronous spatiotemporal relationships, acquires the correlation between sensor data, and can determine the degree of event risk.

[0047] By employing deep mining technology of sensory data, and constructing an integrated three-dimensional holographic sensing system and platform, multi-source and multi-dimensional sensory information is converged and integrated. Through cross-verification, the technology of simulation of safety status and risk identification of heavy-haul train group operation is realized.

[0048] System Architecture: This disclosure proposes a holographic perception system for the operation control of heavy-haul train groups. The data analysis and decision-making functions of the holographic perception system are all based on a large amount of perception information. Based on the access of massive perception data, the representativeness and breadth of the perception data are enhanced to ensure the accuracy and comprehensiveness of information mining. At the same time, through multi-level, multi-scale, and multi-modal complex data deep mining and digital twin based on the holographic perception network, the system outputs safety situation perception and inference information of each element, thereby improving the anti-interference capability of the perception data.

[0049] The system adopts a two-tier architecture: on-site and holographic perception center. On-site: Composed of a series of sensing and monitoring terminals, it realizes high-quality collection of basic sensing data and interfaces with the holographic sensing center, including infrastructure sensing terminals, operating environment sensing terminals, catenary foreign object monitoring, rail breakage monitoring terminals, equipment monitoring, etc.

[0050] Holographic Sensing Center: Performs data analysis and processing on the sensed data. The data analysis stage provides data support services, including data extraction, data transformation, data inspection, and data storage, and completes data aggregation and convergence. The data processing stage provides services such as train operation risk warning, train operation safety situation simulation, in-depth data mining and digital twin services, multi-dimensional data analysis, data visualization, and data reports, enabling the simulation of the safety situation and risk warning of heavy-load train groups based on sensed data.

[0051] Hardware platform architecture: The holographic perception system hardware platform adopts a secure computer platform, including a power supply unit, a computing unit, an internal network unit, a communication interface unit, a maintenance and management unit, and a storage unit. The system hardware structure is as follows: Figure 5 As shown.

[0052] 1) Secure Computer Platform: The holographic perception system adopts the company's independently developed DS6-80 secure computer platform, providing a two-out-of-two security architecture based on task-level synchronization. The secure computer platform has independently passed SIL4 level certification by the relevant project.

[0053] 2) Power supply unit: Provides stable and reliable DC power supply for other devices in the holographic perception system (the security computing unit itself includes a power supply module). Its input is AC220 and its output is DC24V. It is also equipped with a UPS as a backup power supply device.

[0054] 3) Computation Unit: The computation unit performs logical operations. The holographic perception system contains multiple computation units, each consisting of two symmetrical CPU modules. Each CPU module is a fully functional embedded minimum system, integrating CPU, memory, CF card storage, Ethernet communication, and other functions. The CPUs communicate with each other via Ethernet ports, enabling intra-system dual-CPU data cross-comparison and inter-system data synchronization.

[0055] 4) Internal network unit: It undertakes the switching function of the internal network of the holographic perception system and is implemented by multiple switches.

[0056] 5) Communication interface unit: Used to realize communication between the host unit of the holographic perception system and external devices. It is a non-security related unit.

[0057] 6) Maintenance and Management Unit: This unit contains independent monitoring and maintenance software and hardware. Its main function is to monitor and record the real-time working status of the system and alarm information. It is a non-safety-related unit.

[0058] 7) Storage unit: Provides data storage services. It runs as a relatively independent process, providing storage services for external business, including sensing information and data collected by signal equipment, and also providing access services for log data generated by the holographic sensing system itself.

[0059] The system's main functions include business functions and management functions: (1) Business functions, 1) Sensing data reception, The system can receive monitoring data, alarm data, etc. from the sensing and monitoring terminal; 2) Sensing data filtering, The system can process the information received from the sensing and monitoring terminal, mark abnormal data such as data outside the measurement range or garbled characters as suspicious data, and prompt data abnormality alarms; 3) Display of perceived information, The holographic sensing system can display various monitoring information of the sensing monitoring points in the form of text, graphics and other forms. The specific display content can be dynamically configured according to user needs. 4) Sensor information storage and backup, The holographic sensing system can store and back up the monitoring information uploaded by the sensing and monitoring terminal; 5) Sensor information alarm, When the sensing and monitoring terminal uploads alarm information, the holographic sensing center can issue an alarm prompt and provide the alarm information to the CTC and the on-site data center; 6) The alarm for the sensed information has been deactivated. The holographic sensing system supports manual confirmation to deactivate the alarm; 7) Perceive historical data query, The holographic sensing system can provide real-time sensing and monitoring information and alarm information for a specified time period through charts, curves, etc., and can also provide output and printing functions for query result files; 8) Signal equipment collects and stores data. Holographic sensing systems can store data collected by devices; 9) Processing monitoring data from sensing devices. The holographic sensing system has the function of analyzing and processing the monitoring data of a single sensing terminal, and generating event warnings and alarms.

[0060] 10) Multi-source fusion sensing data mining, The holographic sensing system has the function of matching the monitoring data of different sensing terminals. For sensing data with relevant relationships, it can perform multi-source data fusion, situational analysis, risk prediction and early warning protection.

[0061] (2) Business functions 1) Configuration management, The holographic perception system can provide functions such as alarm parameter configuration, terminal function configuration, and user role configuration; 2) Fault information management, The holographic sensing system can receive fault information from the sensing and monitoring terminal, and classify, statistically analyze, and perform analysis on the faults. 3) Performance management, The holographic sensing system can provide real-time monitoring of server CPU utilization, I / O performance, memory usage, and storage device usage. 4) User authentication management Users should use password authentication when logging into the holographic perception system (specific scenarios need to be listed). 5) Access control, The holographic sensing system can be configured with corresponding software function module permissions and data information item permissions according to different monitoring terminals; 6) Log management, The holographic perception system log should include operation logs (operation traces), quality logs (error logs, alarm logs, abnormal data, and other information records), and login logs (user, time, and other information records for each login to the system); 7) Network status management, The holographic sensing system can monitor the network connection status of interface devices in real time and generate network alarm information in real time when the network connection is abnormal. 8) Time synchronization, The holographic sensing system has a clock synchronization function.

[0062] Data processing, When processing actual sensing data, the sampling frequencies of different sensing terminals are inconsistent, resulting in asynchronous sensing data. At the same time, when a sensor generates an alarm, information from other related sensors may not be collected synchronously. It is necessary to use historical sensor data to determine the correlation of sensing information and the degree of risk of the event.

[0063] Therefore, for the early warning analysis and processing of multi-source sensing data, this disclosure proposes a graph neural network prediction and classification algorithm for asynchronous spatiotemporal correlation analysis, which can process multi-source asynchronous sensing data and alarm data to determine the degree of event risk.

[0064] The graph G of the sensing data processing model is defined with nodes V, containing sensor data (number, location, sampling time, sensor type, sensor data, alarm status, alarm level, and estimated alarm processing time). The edges of graph G are defined with edges E, containing (distance between sensors, time difference between the latest sensor sampling data). The proximity matrix A of graph G represents the connection relationships between nodes. This disclosure obtains the positional and temporal relationships between data with the closest sensor sampling times by processing edge E, thereby obtaining asynchronous spatiotemporal relationships and correlations between sensor data without requiring sensor synchronization.

[0065] Then graph G is (V,E,A), where the nodes are an N×F matrix, N is the number of sensor nodes, and in this disclosure, F=8 represents the number of sensor data features. E is an N×N×P matrix, and in this disclosure, P=2 represents the number of edge data features. This represents the j-th channel of the feature vector of the i-th sensor node in graph G. The p-th channel of the feature vector representing the edge between the i-th sensor node and the j-th sensor node in graph G.

[0066] In the graph convolutional neural network disclosed herein, the information propagation model of the sensor nodes is as follows:

[0067] In the formula, This represents the output feature of sensor node i in the (l+1)th layer of the graph convolutional neural network. and This represents the output of sensor node i and sensor node j at layer l. Let R be the activation function, and R be the set of relationships between sensor nodes. Let be the set of sensor nodes with relation r among the one-hop neighbors of sensor node i. In this disclosure, relation r indicates the existence of an association. The value is The regularity constant, Let be the weight parameter matrix of neighboring sensor nodes with relation r in the l-th layer of the network. Let be the weight parameter matrix of sensor node i in the l-th layer of the network.

[0068] The propagation model of edge information is as follows:

[0069] In the formula, Representing an edge eigenvectors, Let represent the feature vector of sensor node i at the t-th graph convolutional layer, M represent the information transfer function, and U represent the sensor node information update function. This represents an information aggregation operation. Specifically, one implementation of the above formula can be:

[0070] In the formula, Let i be the neighboring sensor nodes of sensor node i. This is the Softmax regularization function. and These are the node weight parameter matrix and the edge weight parameter matrix, respectively. For edge feature vectors, Let be the node feature vector, and || be the concatenation operation. This is the output feature vector after aggregating edge feature information.

[0071] In this disclosure, an L-layer graph convolutional neural network can be constructed as needed to... Perform L iterations, then you can stack LSTM and MLP layers to... The process is performed to obtain the final output O, which is (alarm number, alarm location, alarm time, alarm type, alarm level, alarm processing time). The alarm number is arranged in the order of alarms, the alarm location and alarm time are the location of the sensor node that triggered the alarm and the sampling time at the time of the alarm, respectively, the alarm level is the predicted level after processing by the graph convolutional neural network, and the alarm processing time is the time required to process the alarm as predicted by the graph convolutional neural network.

[0072] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A sensing method for controlling the operation of heavy-haul train groups, characterized in that, The method includes: Train group operation status information is collected using different sensors; The state information is transformed into a state graph, where nodes represent sensor sampling points, edges represent the time difference between the latest sensor sampling points, and the proximity matrix represents the connection relationship between nodes. A graph convolutional neural network is built based on the state diagram to predict the alarm level and alarm processing time of the train group operation status. Train a graph convolutional neural network based on the historical state graph to obtain a trained graph convolutional neural network; The real-time collected status graphs are input into a graph convolutional neural network to perform operational control perception and supervision of the train group's operating status. Transforming state information into a state diagram includes: Let G represent the state diagram, and let V be the node of the diagram. V represents the sensor data, including: sampling number, sampling location, sampling time, sensor type, sensor data, and whether an alarm is triggered. The edge E of graph G represents the time difference between the latest sampled data from the sensor; The proximity matrix A of graph G represents the connection relationship between edges between nodes; Building a graph convolutional neural network based on the state diagram includes: Graph Convolutional Neural Networks consist of an L-layer Graph Convolutional Network (GCN) and a multi-layer Perceptron (MLP). The activation functions for the GCN and LSTM layers are both ReLU, while the activation function for the MLP layer is tanh. The output of the convolutional neural network is O= , These represent the alarm number, alarm location, alarm time, alarm type, alarm level, and alarm processing time, respectively. The alarm number is arranged in the order of alarms. The alarm location and alarm time are the location of the sensor node that triggered the alarm and the sampling time at the time of the alarm, respectively. The alarm level is the estimated level after processing by the graph convolutional neural network, and the alarm processing time is the time required to process the alarm as estimated by the graph convolutional neural network. The loss function F of the graph convolutional neural network is as follows: In the formula, Weights for alarm types set based on experience or iterative learning.

2. The sensing method for heavy-haul train group operation control according to claim 1, characterized in that, Graph G is (V,E,A), where node V is an N×F matrix, N is the number of sensor nodes, and F is the number of data features of the sensor; edge E is an N×N×P matrix, and P is the number of edge data features.

3. The sensing method for heavy-haul train group operation control according to claim 1, characterized in that, The information propagation model of sensor nodes in a graph convolutional neural network is as follows: In the formula, This represents the output feature of sensor node i at the (l+1)th layer of the graph convolutional neural network. and This represents the output of sensor node i and sensor node j at layer l. Let R be the activation function, and R be the set of relationships between sensor nodes. Let r be the set of sensor nodes that are one-hop neighbors of sensor node i, where r indicates that there is an association relationship. The value is The regularity constant, Let be the weight parameter matrix of neighboring sensor nodes with relation r in the l-th layer of the network. Let be the weight parameter matrix of sensor node i in the l-th layer of the network.

4. The sensing method for heavy-haul train group operation control according to claim 1, characterized in that, The propagation model of edge information in graph convolutional neural networks is as follows: In the formula, Let i be the neighboring sensor nodes of sensor node i. This is the Softmax regularization function. and These are the node weight parameter matrix and the edge weight parameter matrix, respectively. For edge feature vectors, Let be the node feature vector, and || be the concatenation operation. This is the output feature vector after aggregating edge feature information.

5. A heavy-haul train group operation control and sensing system, characterized in that, include: Data acquisition unit, graph transformation unit, model building unit, and perception supervision unit; The data acquisition unit is used to collect train group operation status information through different sensors; The graph conversion unit is used to convert state information into a state graph. In the graph, nodes represent sensor sampling points, edges represent the time difference between the latest sampling points of the sensor, and the proximity matrix A represents the connection relationship between the edges of the nodes. The model building unit is used to build a graph convolutional neural network based on the state diagram, which is used to predict the alarm level and alarm processing time of the train group operation status. The model building unit is also used to train a graph convolutional neural network based on the historical state graph to obtain a trained graph convolutional neural network. The perception and supervision unit is used to input the real-time acquired state graph into the graph convolutional neural network to perform operation control perception and supervision of the train group's operating status. Transforming state information into a state diagram includes: Let G represent the state diagram, and let V be the node of the diagram. V represents the sensor data, including: sampling number, sampling location, sampling time, sensor type, sensor data, and whether an alarm is triggered. The edge E of graph G represents the time difference between the latest sampled data from the sensor; The proximity matrix A of graph G represents the connection relationship between edges between nodes; Building a graph convolutional neural network based on the state diagram includes: Graph Convolutional Neural Networks consist of an L-layer Graph Convolutional Network (GCN) and a multi-layer Perceptron (MLP). The activation functions for the GCN and LSTM layers are both ReLU, while the activation function for the MLP layer is tanh. The output of the convolutional neural network is O= , These represent the alarm number, alarm location, alarm time, alarm type, alarm level, and alarm processing time, respectively. The alarm number is arranged in the order of alarms. The alarm location and alarm time are the location of the sensor node that triggered the alarm and the sampling time at the time of the alarm, respectively. The alarm level is the estimated level after processing by the graph convolutional neural network, and the alarm processing time is the time required to process the alarm as estimated by the graph convolutional neural network. The loss function F of the graph convolutional neural network is as follows: In the formula, Weights for alarm types set based on experience or iterative learning.

6. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, which stores computer programs; A processor, when executing a computer program stored in memory, implements the heavy-haul train group operation control sensing method according to any one of claims 1-4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the heavy-haul train group operation control sensing method according to any one of claims 1-4.

Images