Rail transit dispatching and operation intelligent decision and exercise system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHAANXI TRANSPORTATION VOCATIONAL & TECH COLLEGE
- Filing Date
- 2026-03-19
- Publication Date
- 2026-06-16
Smart Images
Figure CN122222291A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of auxiliary decision-making technology for emergency response to sudden incidents in rail transit, specifically involving a smart decision-making and drill system for rail transit dispatching and maintenance. Background Technology
[0002] With the rapid development of urban rail transit network operations, train operating density is constantly increasing, passenger flow is continuously growing, and equipment systems are becoming increasingly complex. When emergencies such as train fires, switch malfunctions, or signal interruptions occur, improper handling can lead to train delays and passenger congestion, or even operational paralysis and casualties. Traditional emergency response mainly relies on the manual experience of personnel in various positions, telephone communication, and paper-based plans, which have the following prominent problems:
[0003] Information silos and delayed collaboration: Poor information transmission among dispatching, operation, and maintenance personnel, and the lack of a unified data sharing platform, lead to delayed emergency command and decision-making and difficulties in cross-departmental collaboration.
[0004] Decision-making relies on human experience: When faced with complex and ever-changing fault scenarios, dispatchers need to quickly consult a large number of regulations and cases, and make the best decisions under great pressure. The risk of human error is high, and it is difficult to guarantee the scientific nature and timeliness of the handling plan.
[0005] The training and drills are not effective: existing training often uses theoretical explanations or simple simulations, lacking a highly realistic virtual drill environment, making it difficult for trainees to truly master the practical skills of multi-position collaborative handling.
[0006] Train operation adjustments are difficult: Unexpected events often cause timetable disruptions, requiring dispatchers to manually adjust train operation plans. This involves complex calculations, is time-consuming, and makes it difficult to guarantee the optimality of the adjustment plan.
[0007] Existing technologies include several auxiliary decision-making systems for rail transit emergency response. For example, Chinese patent "An Emergency Command System and Method for Urban Rail Transit" (CN112561284A) proposes receiving alarm information and automatically and intelligently matching emergency plans. However, it focuses on auxiliary decision-making for a single position and fails to achieve AI-based intelligent decision-making and multi-position collaborative process generation. Another example is "An Emergency Drill Management Platform for Urban Rail Transit" (CN117022409A), which mainly focuses on the digital management of emergency drill processes and lacks the use of intelligent algorithms to adjust the core operation plan of rail transit. Therefore, there is an urgent need for an intelligent decision-making and drill system that can deeply integrate AI large-scale models, simulation modeling, and intelligent algorithms to comprehensively improve the intelligence level and collaborative efficiency of rail transit emergency response. Summary of the Invention
[0008] The purpose of this invention is to provide a smart decision-making and drill system for rail transit dispatching and maintenance, in order to solve the problems of low emergency response efficiency, poor coordination, reliance on human experience for decision-making, and lack of intelligent auxiliary means in the existing technology, and to achieve rapid, accurate, and visualized emergency response decision support and drill training.
[0009] The technical solution adopted in this invention is a smart decision-making and exercise system for rail transit dispatching and inspection, including a problem extraction module, a knowledge base construction module, a thinking and problem-solving module, a train operation adjustment module, a human-computer interaction interface module, and a solution output module.
[0010] The invention is further characterized in that, The question extraction module, based on the AI large-scale model Qwen3, performs semantic understanding and keyword extraction on the user-input description of sudden event faults or query questions. The Qwen3 model is based on the Transformer decoder architecture, and its core structure includes an embedding layer, a multi-layer decoder module, and an output layer. The embedding layer converts the input text into a vector representation; each decoder module contains a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer captures the dependencies between different words in the text, and the feedforward neural network layer performs non-linear transformations; the layer normalization module RMSnorm stabilizes the training process; the output layer generates a probability distribution on the vocabulary through linear transformations and outputs classification results containing key information such as fault type, location of occurrence, and scope of impact.
[0011] The knowledge base construction module is used to store and classify typical rail transit fault cases, emergency response regulations, industry standards, and basic data including BIM models of operating environments and equipment such as stations and lines, and equipment parameters. It also uses Anylogic simulation software to perform virtual simulation modeling of typical fault scenarios to build a knowledge base that includes a fault case library, a regulation and standard library, and a simulation model library. The problem-solving module employs Retrieval Enhancement Generative Planning (RAG) technology. Based on key information output from the problem extraction module, it performs multi-source searches in the knowledge base construction module, matching relevant emergency failure cases, regulatory standards, and virtual simulation models. Simultaneously, it utilizes AI large-scale model technology to deeply analyze the search results. This deep analysis process includes: the large model performing similarity comparisons and extracting handling experience from the retrieved failure cases; compliance screening and applicability assessment of regulations; scenario matching and key parameter extraction from the simulation model; and combining the above information with the real-time context input by the user to generate a preliminary emergency response plan framework. This framework covers the division of responsibilities, key operational steps, collaborative interaction nodes, and safety precautions for dispatching, operations, and maintenance positions. Then, it optimizes the solution based on the user's actual problem, ultimately generating a decision-making-supporting emergency response plan, specifically including emergency response key points for dispatching, operations, and maintenance positions, collaborative response flowcharts, and virtual simulation videos.
[0012] The train operation adjustment module, when a sudden fault affects train operation, constructs a train operation adjustment optimization model based on real-time timetable data, train positions, and track occupancy, with the objective of minimizing total delay time and passenger waiting time. This model is a typical N-Phard problem, i.e., a bi-objective mixed-integer linear programming problem. The specific calculation process is as follows: First, data including train times, section travel time, station dwell time, train sequence, and track occupancy status from the current timetable are collected, along with the affected sections and their expected duration. Then, the optimization model is constructed, with decision variables including the adjustment amount of departure time for each train, the adjustment amount of dwell time, the section operation level, and the necessary adjustments. The system considers the train's choice between turning back or stopping at intermediate stations. Constraints include: train tracking interval time constraints, section running time constraints, station stopping time constraints, and line occupancy uniqueness constraints. The objective function is to minimize the weighted sum of total delay time and passenger waiting time. An improved Grey Wolf optimization algorithm is used: each solution is represented as a vector of train adjustments; the population is initialized; the fitness value of each individual (i.e., the objective function value) is calculated; the Grey Wolf positions are updated iteratively; mutation operations are introduced to avoid local optima; and the optimal solution is output after the termination condition is met. Finally, an adjusted train timetable is generated based on the optimal solution, and the train operation diagram is redrawn to provide decision support for dispatchers.
[0013] The solution output module evaluates and analyzes the emergency response solutions output by the thinking and solving module. By comparing the effects of historical cases, simulation results, and regulatory compliance checks, it outputs optimized auxiliary decision-making solutions and presents them in a graphical way, including a flowchart of collaborative handling by dispatch, operation, and maintenance personnel, key energy points, and safety precautions. It can also play corresponding simulation videos to intuitively demonstrate the handling process.
[0014] The human-computer interaction interface module adopts a web-based application platform, integrates the above modules, provides users with a unified input interface, supports fault query, solution generation, and operation diagram adjustment functions, and displays the results in the form of visual flowcharts and simulation videos.
[0015] In the knowledge base construction module, the fault case library comes from historical accident reports, operation and maintenance records and theoretical knowledge, and is structured; the rules and standards library covers relevant standards at the national, industry and enterprise levels; the simulation model library uses tools such as Anylogic and BIM to build detailed models of equipment in typical station, section and vehicle environments, which can dynamically simulate passenger flow evacuation, train operation and equipment maintenance under different fault scenarios.
[0016] The graphical flowchart generated by the solution output module clearly defines the task sequence and interaction relationship of each position in scheduling, operation, and maintenance, and supports exporting as an image or PDF; at the same time, it can generate corresponding simulation videos to intuitively demonstrate the handling process.
[0017] The train operation adjustment module employs various intelligent algorithms to solve a multi-objective optimization problem. The decision variables are the arrival and departure times, stop times, and interval running order of each train. Constraints include tracking intervals, interval running times, and station capacity. The objective function is to minimize the weighted total delay time and passenger waiting time. The solution process is as follows: First, the multi-objective problem is transformed into a single objective using a weighted summation method, with weights dynamically adjusted according to operational needs. Then, an improved Grey Wolf optimization algorithm or Particle Swarm Optimization algorithm is used for the solution. The algorithm flow includes initializing the population, calculating fitness, updating individual and global optima, and iterating until convergence. In each iteration, the adjustment amount for each train is calculated based on the current solution, and it is checked whether all constraints are satisfied. If not, corrections are made. Finally, the adjustment scheme that satisfies the constraints and optimizes the objective function is output, i.e., the new train timetable.
[0018] The beneficial effects of this invention are as follows: Improved emergency response efficiency: By rapidly retrieving knowledge from the knowledge base using an AI-powered large-scale model and generating preliminary response plans, emergency decision-making time is significantly shortened from minutes to seconds. Enhanced multi-position collaboration: The system outputs a collaborative flowchart involving dispatching, operation, and inspection, clearly defining the responsibilities and connections of each position, reducing communication errors, and improving overall collaboration capabilities. Visualized and rehearsable decision-making: Complex processes are transformed into graphical flowcharts and accompanied by simulation videos, enabling trainees and frontline personnel to intuitively understand the response process and improving training effectiveness. Optimized train operation adjustments: Intelligent algorithms automatically generate adjusted operation schedules, reducing the impact of faults on overall line operations and minimizing passenger delays. Compliance with industry standards: The system has a built-in library of regulations and standards, ensuring that the generated response plans comply with current regulations and avoiding the risk of violations. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the overall modules of the intelligent decision-making and drill system for rail transit dispatching and inspection of the present invention; Figure 2 This is a schematic diagram of the human-computer interaction module in this invention; Figure 3 This is an example of generating a train fire emergency response flowchart in an embodiment of the present invention; Figure 4 This is an example of generating a simulated video of a train fire emergency response in an embodiment of the present invention; Figure 5 This is an example of the original train timetable before the fault in an embodiment of the present invention; Figure 6 This is an example of adjusting the train timetable after a fault in an embodiment of the present invention. Detailed Implementation
[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.
[0021] Example 1 This invention relates to a smart decision-making and drill system for rail transit dispatching and maintenance, such as... Figure 1 As shown, it includes a problem extraction module, a knowledge base construction module, a problem-solving module, a train operation adjustment module, a human-computer interaction interface module, and a solution output module.
[0022] The question extraction module, based on the AI large-scale model Qwen3, performs semantic understanding and keyword extraction on the user-input description of sudden event faults or query questions. The Qwen3 model is based on the Transformer decoder architecture, and its core structure includes an embedding layer, a multi-layer decoder module, and an output layer. The embedding layer converts the input text into a vector representation; each decoder module contains a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer captures the dependencies between different words in the text, and the feedforward neural network layer performs non-linear transformations; the layer normalization module RMSnorm stabilizes the training process; the output layer generates a probability distribution on the vocabulary through linear transformations and outputs classification results containing key information such as fault type, location of occurrence, and scope of impact.
[0023] The knowledge base construction module is used to store and classify typical rail transit fault cases, emergency response regulations, industry standards, and basic data including BIM models of operating environments and equipment such as stations and lines, as well as equipment parameters. It also uses Anylogic simulation software to perform virtual simulation modeling of typical fault scenarios, such as fires in train sections, fires on station platforms, turnout faults at turnaround stations, and faults in onboard signaling equipment, to build a knowledge base that includes a fault case library, a regulation and standard library, and a simulation model library. The problem-solving module employs Retrieval Enhancement Generative Intelligence (RAG) technology. Based on key information output from the problem extraction module, it performs multi-source searches in the knowledge base construction module, matching relevant emergency failure cases, regulatory standards, and virtual simulation models. Simultaneously, it utilizes AI large-scale model technology to deeply analyze the search results. This deep analysis process includes: the large model comparing the similarity of retrieved failure cases and extracting handling experience; screening regulations for compliance and applicability; matching simulation models to scenarios and extracting key parameters; and combining this information with the user's real-time input context (such as failure location, impact range, and current passenger flow). Through reasoning, it generates a preliminary emergency response plan framework, covering the responsibilities, key operational steps, collaborative interaction nodes, and safety precautions for dispatching, operations, and maintenance positions. Then, it optimizes the solution based on the user's actual problem, ultimately generating a decision-making-supporting emergency response plan, specifically including emergency response points for dispatching, operations, and maintenance positions, collaborative response flowcharts, and virtual simulation videos.
[0024] The train operation adjustment module, when a sudden fault affects train operation, constructs a train operation adjustment optimization model based on real-time timetable data, train positions, and track occupancy, with the objective of minimizing total delay time and passenger waiting time. This model is a typical N-Phard problem, i.e., a bi-objective mixed-integer linear programming problem. The specific calculation process is as follows: First, data including train times, section travel time, station dwell time, train sequence, and track occupancy status from the current timetable are collected, as well as the affected sections and expected duration of the fault. Then, the optimization model is constructed, with decision variables including the adjustment amount of departure time for each train, the adjustment amount of dwell time, the section operation level (affecting travel time), and, if necessary, the choice of train turnaround or suspension at intermediate stations. Constraints include: train tracking interval time constraint (the departure interval between preceding and following trains is not less than the minimum tracking interval), section travel time constraint (train passing through...). The constraints include the shortest and longest allowed time between sections, station dwell time constraints (minimum dwell time required for passengers to board and alight), and line occupancy uniqueness constraints (only one train can occupy the same section or station at the same time). The objective function is to minimize the weighted total delay time (the sum of the delay times of all trains at each station) and the passenger waiting time (the extra waiting time for passengers due to train delays, estimated based on passenger flow data). An improved gray wolf optimization algorithm is used to solve the problem: each solution is represented as a vector of train adjustment amounts. The population is initialized, the fitness value of each individual (i.e., the objective function value) is calculated, and the gray wolf positions are iteratively updated (alpha, beta, delta wolf guidance). Mutation operations are introduced to avoid local optima. The optimal solution is output after the termination condition is met. Finally, an adjusted train timetable is generated based on the optimal solution, and the train operation diagram is redrawn to provide decision support for dispatchers.
[0025] The solution output module evaluates and analyzes the emergency response plans output by the problem-solving module. By comparing historical case results, simulation results, and compliance checks with regulations and standards, it outputs optimized decision support plans, presented graphically. This includes flowcharts of collaborative responses among dispatch, operations, and maintenance personnel, key performance indicators, and safety precautions. Figure 3 As shown; simultaneously, corresponding simulation videos can be played to intuitively demonstrate the handling process, such as... Figure 4 As shown; like Figure 2 As shown, the human-computer interaction interface module adopts a web-based application platform, integrates the above modules, provides users with a unified input interface, supports fault query, solution generation, and operation diagram adjustment functions, and displays the results in the form of visual flowcharts and simulation videos.
[0026] In the knowledge base construction module, the fault case library comes from historical accident reports, operation and maintenance records and theoretical knowledge, and is structured; the rules and standards library covers relevant standards at the national, industry and enterprise levels; the simulation model library uses tools such as Anylogic and BIM to build detailed models of equipment in typical station, section and vehicle environments, which can dynamically simulate passenger flow evacuation, train operation and equipment maintenance under different fault scenarios.
[0027] The graphical flowchart generated by the solution output module clearly defines the task sequence and interaction relationship of each position in scheduling, operation, and maintenance, and supports exporting as an image or PDF; at the same time, it can generate corresponding simulation videos to intuitively demonstrate the handling process.
[0028] The train operation adjustment module employs various intelligent algorithms to solve a multi-objective optimization problem. The decision variables are the arrival and departure times, stop times, and interval running order of each train. Constraints include tracking intervals, interval running times, and station capacity. The objective function is to minimize the weighted total delay time and passenger waiting time. The solution process is as follows: First, the multi-objective problem is transformed into a single objective using a weighted summation method, with weights dynamically adjusted according to operational needs. Then, an improved Grey Wolf optimization algorithm or Particle Swarm Optimization algorithm is used for the solution. The algorithm flow includes initializing the population, calculating fitness, updating individual and global optima, and iterating until convergence. In each iteration, the adjustment amount for each train is calculated based on the current solution, and it is checked whether all constraints are satisfied. If not, corrections are made. Finally, the adjustment scheme that satisfies the constraints and optimizes the objective function is output, i.e., the new train timetable.
[0029] Example 2 The present invention relates to a smart decision-making and exercise system for rail transit dispatching and maintenance, comprising a problem extraction module, a knowledge base construction module, a problem-solving module, a train operation adjustment module, a human-computer interaction interface module, and a solution output module.
[0030] The question extraction module, based on the AI large-scale model Qwen3, performs semantic understanding and keyword extraction on the user-input description of sudden event faults or query questions. The Qwen3 model is based on the Transformer decoder architecture, and its core structure includes an embedding layer, a multi-layer decoder module, and an output layer. The embedding layer converts the input text into a vector representation; each decoder module contains a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer captures the dependencies between different words in the text, and the feedforward neural network layer performs non-linear transformations; the layer normalization module RMSnorm stabilizes the training process; the output layer generates a probability distribution on the vocabulary through linear transformations and outputs classification results containing key information such as fault type, location of occurrence, and scope of impact.
[0031] Example 3 This invention relates to a smart decision-making and drill system for rail transit dispatching and maintenance, such as... Figure 1 As shown, it includes a problem extraction module, a knowledge base construction module, a problem-solving module, a train operation adjustment module, a human-computer interaction interface module, and a solution output module.
[0032] The question extraction module, based on the AI large-scale model Qwen3, performs semantic understanding and keyword extraction on the user-input description of sudden event faults or query questions. The Qwen3 model is based on the Transformer decoder architecture, and its core structure includes an embedding layer, a multi-layer decoder module, and an output layer. The embedding layer converts the input text into a vector representation; each decoder module contains a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer captures the dependencies between different words in the text, and the feedforward neural network layer performs non-linear transformations; the layer normalization module RMSnorm stabilizes the training process; the output layer generates a probability distribution on the vocabulary through linear transformations and outputs classification results containing key information such as fault type, location of occurrence, and scope of impact.
[0033] The knowledge base construction module is used to store and classify typical rail transit fault cases, emergency response regulations, industry standards, and basic data including BIM models of operating environments and equipment such as stations and lines, as well as equipment parameters. It also uses Anylogic simulation software to perform virtual simulation modeling of typical fault scenarios, such as fires in train sections, fires on station platforms, turnout faults at turnaround stations, and faults in onboard signaling equipment, to build a knowledge base that includes a fault case library, a regulation and standard library, and a simulation model library. Example 4 This invention relates to a smart decision-making and drill system for rail transit dispatching and maintenance, such as... Figure 1 As shown, it includes a problem extraction module, a knowledge base construction module, a problem-solving module, a train operation adjustment module, a human-computer interaction interface module, and a solution output module.
[0034] The question extraction module, based on the AI large-scale model Qwen3, performs semantic understanding and keyword extraction on the user-input description of sudden event faults or query questions. The Qwen3 model is based on the Transformer decoder architecture, and its core structure includes an embedding layer, a multi-layer decoder module, and an output layer. The embedding layer converts the input text into a vector representation; each decoder module contains a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer captures the dependencies between different words in the text, and the feedforward neural network layer performs non-linear transformations; the layer normalization module RMSnorm stabilizes the training process; the output layer generates a probability distribution on the vocabulary through linear transformations and outputs classification results containing key information such as fault type, location of occurrence, and scope of impact.
[0035] The knowledge base construction module is used to store and classify typical rail transit fault cases, emergency response regulations, industry standards, and basic data including BIM models of operating environments and equipment such as stations and lines, as well as equipment parameters. It also uses Anylogic simulation software to perform virtual simulation modeling of typical fault scenarios, such as fires in train sections, fires on station platforms, turnout faults at turnaround stations, and faults in onboard signaling equipment, to build a knowledge base that includes a fault case library, a regulation and standard library, and a simulation model library. The problem-solving module employs Retrieval Enhancement Generative Intelligence (RAG) technology. Based on key information output from the problem extraction module, it performs multi-source searches in the knowledge base construction module, matching relevant emergency failure cases, regulatory standards, and virtual simulation models. Simultaneously, it utilizes AI large-scale model technology to deeply analyze the search results. This deep analysis process includes: the large model comparing the similarity of retrieved failure cases and extracting handling experience; screening regulations for compliance and applicability; matching simulation models to scenarios and extracting key parameters; and combining this information with the user's real-time input context (such as failure location, impact range, and current passenger flow). Through reasoning, it generates a preliminary emergency response plan framework, covering the responsibilities, key operational steps, collaborative interaction nodes, and safety precautions for dispatching, operations, and maintenance positions. Then, it optimizes the solution based on the user's actual problem, ultimately generating a decision-making-supporting emergency response plan, specifically including emergency response points for dispatching, operations, and maintenance positions, collaborative response flowcharts, and virtual simulation videos.
[0036] Example 5 This invention relates to a smart decision-making and drill system for rail transit dispatching and maintenance, such as... Figure 1 As shown, it includes a problem extraction module, a knowledge base construction module, a problem-solving module, a train operation adjustment module, a human-computer interaction interface module, and a solution output module.
[0037] The question extraction module, based on the AI large-scale model Qwen3, performs semantic understanding and keyword extraction on the user-input description of sudden event faults or query questions. The Qwen3 model is based on the Transformer decoder architecture, and its core structure includes an embedding layer, a multi-layer decoder module, and an output layer. The embedding layer converts the input text into a vector representation; each decoder module contains a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer captures the dependencies between different words in the text, and the feedforward neural network layer performs non-linear transformations; the layer normalization module RMSnorm stabilizes the training process; the output layer generates a probability distribution on the vocabulary through linear transformations and outputs classification results containing key information such as fault type, location of occurrence, and scope of impact.
[0038] The knowledge base construction module is used to store and classify typical rail transit fault cases, emergency response regulations, industry standards, and basic data including BIM models of operating environments and equipment such as stations and lines, as well as equipment parameters. It also uses Anylogic simulation software to perform virtual simulation modeling of typical fault scenarios, such as fires in train sections, fires on station platforms, turnout faults at turnaround stations, and faults in onboard signaling equipment, to build a knowledge base that includes a fault case library, a regulation and standard library, and a simulation model library. The problem-solving module employs Retrieval Enhancement Generative Intelligence (RAG) technology. Based on key information output from the problem extraction module, it performs multi-source searches in the knowledge base construction module, matching relevant emergency failure cases, regulatory standards, and virtual simulation models. Simultaneously, it utilizes AI large-scale model technology to deeply analyze the search results. This deep analysis process includes: the large model comparing the similarity of retrieved failure cases and extracting handling experience; screening regulations for compliance and applicability; matching simulation models to scenarios and extracting key parameters; and combining this information with the user's real-time input context (such as failure location, impact range, and current passenger flow). Through reasoning, it generates a preliminary emergency response plan framework, covering the responsibilities, key operational steps, collaborative interaction nodes, and safety precautions for dispatching, operations, and maintenance positions. Then, it optimizes the solution based on the user's actual problem, ultimately generating a decision-making-supporting emergency response plan, specifically including emergency response points for dispatching, operations, and maintenance positions, collaborative response flowcharts, and virtual simulation videos.
[0039] The train operation adjustment module, when a sudden fault affects train operation, constructs a train operation adjustment optimization model based on real-time timetable data, train positions, and track occupancy, with the objective of minimizing total delay time and passenger waiting time. This model is a typical N-Phard problem, i.e., a bi-objective mixed-integer linear programming problem. The specific calculation process is as follows: First, data including train times, section travel time, station dwell time, train sequence, and track occupancy status from the current timetable are collected, as well as the affected sections and expected duration of the fault. Then, the optimization model is constructed, with decision variables including the adjustment amount of departure time for each train, the adjustment amount of dwell time, the section operation level (affecting travel time), and, if necessary, the choice of train turnaround or suspension at intermediate stations. Constraints include: train tracking interval time constraint (the departure interval between preceding and following trains is not less than the minimum tracking interval), section travel time constraint (train passing through...). The constraints include the shortest and longest allowed time between sections, station dwell time constraints (minimum dwell time required for passengers to board and alight), and line occupancy uniqueness constraints (only one train can occupy the same section or station at the same time). The objective function is to minimize the weighted total delay time (the sum of the delay times of all trains at each station) and the passenger waiting time (the extra waiting time for passengers due to train delays, estimated based on passenger flow data). An improved gray wolf optimization algorithm is used to solve the problem: each solution is represented as a vector of train adjustment amounts. The population is initialized, the fitness value of each individual (i.e., the objective function value) is calculated, and the gray wolf positions are iteratively updated (alpha, beta, delta wolf guidance). Mutation operations are introduced to avoid local optima. The optimal solution is output after the termination condition is met. Finally, an adjusted train timetable is generated based on the optimal solution, and the train operation diagram is redrawn to provide decision support for dispatchers.
[0040] Example 6 This invention relates to a smart decision-making and drill system for rail transit dispatching and maintenance, such as... Figure 1 As shown, it includes a problem extraction module, a knowledge base construction module, a problem-solving module, a train operation adjustment module, a human-computer interaction interface module, and a solution output module.
[0041] The question extraction module, based on the AI large-scale model Qwen3, performs semantic understanding and keyword extraction on the user-input description of sudden event faults or query questions. The Qwen3 model is based on the Transformer decoder architecture, and its core structure includes an embedding layer, a multi-layer decoder module, and an output layer. The embedding layer converts the input text into a vector representation; each decoder module contains a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer captures the dependencies between different words in the text, and the feedforward neural network layer performs non-linear transformations; the layer normalization module RMSnorm stabilizes the training process; the output layer generates a probability distribution on the vocabulary through linear transformations and outputs classification results containing key information such as fault type, location of occurrence, and scope of impact.
[0042] The knowledge base construction module is used to store and classify typical rail transit fault cases, emergency response regulations, industry standards, and basic data including BIM models of operating environments and equipment such as stations and lines, as well as equipment parameters. It also uses Anylogic simulation software to perform virtual simulation modeling of typical fault scenarios, such as fires in train sections, fires on station platforms, turnout faults at turnaround stations, and faults in onboard signaling equipment, to build a knowledge base that includes a fault case library, a regulation and standard library, and a simulation model library. The problem-solving module employs Retrieval Enhancement Generative Intelligence (RAG) technology. Based on key information output from the problem extraction module, it performs multi-source searches in the knowledge base construction module, matching relevant emergency failure cases, regulatory standards, and virtual simulation models. Simultaneously, it utilizes AI large-scale model technology to deeply analyze the search results. This deep analysis process includes: the large model comparing the similarity of retrieved failure cases and extracting handling experience; screening regulations for compliance and applicability; matching simulation models to scenarios and extracting key parameters; and combining this information with the user's real-time input context (such as failure location, impact range, and current passenger flow). Through reasoning, it generates a preliminary emergency response plan framework, covering the responsibilities, key operational steps, collaborative interaction nodes, and safety precautions for dispatching, operations, and maintenance positions. Then, it optimizes the solution based on the user's actual problem, ultimately generating a decision-making-supporting emergency response plan, specifically including emergency response points for dispatching, operations, and maintenance positions, collaborative response flowcharts, and virtual simulation videos.
[0043] The train operation adjustment module, when a sudden fault affects train operation, constructs a train operation adjustment optimization model based on real-time timetable data, train positions, and track occupancy, with the objective of minimizing total delay time and passenger waiting time. This model is a typical N-Phard problem, i.e., a bi-objective mixed-integer linear programming problem. The specific calculation process is as follows: First, data including train times, section travel time, station dwell time, train sequence, and track occupancy status from the current timetable are collected, as well as the affected sections and expected duration of the fault. Then, the optimization model is constructed, with decision variables including the adjustment amount of departure time for each train, the adjustment amount of dwell time, the section operation level (affecting travel time), and, if necessary, the choice of train turnaround or suspension at intermediate stations. Constraints include: train tracking interval time constraint (the departure interval between preceding and following trains is not less than the minimum tracking interval), section travel time constraint (train passing through...). The constraints include the shortest and longest allowed time between sections, station dwell time constraints (minimum dwell time required for passengers to board and alight), and line occupancy uniqueness constraints (only one train can occupy the same section or station at the same time). The objective function is to minimize the weighted total delay time (the sum of the delay times of all trains at each station) and the passenger waiting time (the extra waiting time for passengers due to train delays, estimated based on passenger flow data). An improved gray wolf optimization algorithm is used to solve the problem: each solution is represented as a vector of train adjustment amounts. The population is initialized, the fitness value of each individual (i.e., the objective function value) is calculated, and the gray wolf positions are iteratively updated (alpha, beta, delta wolf guidance). Mutation operations are introduced to avoid local optima. The optimal solution is output after the termination condition is met. Finally, an adjusted train timetable is generated based on the optimal solution, and the train operation diagram is redrawn to provide decision support for dispatchers.
[0044] The solution output module evaluates and analyzes the emergency response plans output by the problem-solving module. By comparing historical case results, simulation results, and compliance checks with regulations and standards, it outputs optimized decision support plans, presented graphically. This includes flowcharts of collaborative responses among dispatch, operations, and maintenance personnel, key performance indicators, and safety precautions. Figure 3 As shown; simultaneously, corresponding simulation videos can be played to intuitively demonstrate the handling process, such as... Figure 4 As shown; like Figure 2 As shown, the human-computer interaction interface module adopts a web-based application platform, integrates the above modules, provides users with a unified input interface, supports fault query, solution generation, and operation diagram adjustment functions, and displays the results in the form of visual flowcharts and simulation videos.
[0045] Example 7 Fault Case Database: This database collects typical accident reports from the past decade in the rail transit sector (such as train fires and switch malfunctions), extracting key fields such as fault phenomena, causes, handling processes, and lessons learned to form structured case entries. Each case entry is tagged (fault type, line, equipment, consequences, etc.) for easy retrieval.
[0046] The regulations and standards library compiles the national "Urban Rail Transit Operation Organization and Management Measures", the industry standard "Urban Rail Transit Operation Emergency Drill Management Measures", and the corporate regulations of various local subway companies, and digitizes the texts and annotates key clauses.
[0047] Simulation Model Library: Utilizing Anylogic simulation software, 3D models of typical stations (such as transfer stations and terminal stations) are constructed, including facilities such as platforms, concourses, turnstiles, escalators, and equipment models such as trains and signaling systems. For typical faults (such as train fires and switch malfunctions), simulation models with multiple pre-set handling scenarios are provided to simulate passenger evacuation and train delays under different handling measures. Simultaneously, BIM models are imported for detailed display of station structure and equipment layout.
[0048] Problem extraction module: Users input a fault description via a web interface, such as: "A fire occurred on train 012 between stations A and B." This module uses a large AI model (such as Qwen3) to perform semantic parsing and extract key information: fault type = "train fire", train number = "012", section = "station A - station B", and potential impact = "section operation, passenger safety". The extracted results are then passed to the solution module.
[0049] Thinking and Solving Module: Using RAG technology, based on fault type keywords, such as "train fire", a search is performed in the knowledge base construction module to retrieve relevant cases, regulations, and two simulation models (such as "platform fire evacuation model" and "section fire rescue model").
[0050] Then, by combining the search results with regulations and case studies, preliminary suggestions for collaborative handling by the three parties involved in "transportation, dispatch, and inspection" are generated. See below: Dispatching duties include: remotely stopping subsequent trains, initiating smoke extraction in the section, notifying the station ahead to prepare for evacuation, and adjusting the train schedule.
[0051] Operations positions (stations): Activate fire emergency plans, open turnstiles, maintain access control in open mode, broadcast announcements to guide passengers, and organize station staff to wear fire-fighting equipment.
[0052] Maintenance position: Rush to the scene and inspect the damage to the vehicle.
[0053] Solution output module: This module evaluates and optimizes the output solutions; compares them with successful experiences from similar historical cases to check the completeness of the recommendations; uses the corresponding scenarios in the simulation model library to conduct rapid simulations and evaluate the expected effects of different handling steps (such as evacuation time and train delays); and checks the regulatory standards library to ensure that the recommendations do not violate any regulations.
[0054] The final output is the optimized decision support solution, presented in the form of a flowchart (see...). Figure 3 At the same time, a corresponding 3D simulation video is generated to intuitively demonstrate the treatment process.
[0055] Train operation adjustment module: Combination Figure 5 , Figure 6 , Figure 5 This is an example of the original train timetable before the fault in an embodiment of the present invention. Figure 6 This is an example of adjusting the train timetable after a fault in this embodiment of the invention. This module is automatically triggered if a fault causes train delays or line interruptions. Inputs include the fault type, fault location, current timetable data (train timetable, position, speed), line occupancy, and estimated fault duration. A mixed-integer linear programming model is constructed with the goal of minimizing total delay time and passenger waiting time. A smart algorithm is used to solve the problem, resulting in a new train timetable, including adjustments to train stop times, changes in section operation levels, and the reversal or cancellation of some trains. The algorithm provides optimization results within one minute and displays a comparison between the original and new timetables (see...). Figure 4 The dispatcher can adopt or make minor adjustments.
[0056] Human-computer interaction interface: The main functional areas of the interface include: fault input entry (supports voice input), knowledge base management, intelligent decision triggering, and learning progress.
[0057] Theoretical feasibility analysis: The core of this invention lies in the deep integration of the understanding and reasoning capabilities of a large AI model, the structured information of a knowledge base, the predictive capabilities of simulation modeling, and the computational power of optimization algorithms. The large AI model can understand complex fault scenarios described in natural language and utilize the knowledge base to provide factual evidence; the simulation model can verify the feasibility of the solution; and the optimization algorithm can quantify the decision-making objectives. This collaborative architecture of "knowledge + data + model" ensures the scientific rigor, compliance, and operability of the generated solutions, aligning with the development direction of intelligent rail transit.
[0058] Train fire emergency response: The user entered: "A fire broke out on train 012 between stations A and B, and there was thick smoke in the carriages."
[0059] System processing: Problems identified: fire, 012, interval AB.
[0060] Search and consideration: Retrieve 3 similar cases from the "Train Fire Case Database" and Article X of the "Urban Rail Transit Operation Organization and Management Measures" for fire handling, and call up the "Intersection Fire Evacuation Simulation Model".
[0061] In-depth thinking: Generate preliminary suggestions (dispatch and stop subsequent trains, start smoke extraction; prepare for evacuation at the station; maintenance personnel rush to the scene).
[0062] Optimization module: Simulation results show that if the train forces its way into the station, evacuation is safer than evacuation within the station area. It is recommended that passengers be cleared after the train enters the station.
[0063] Output: Generates a flowchart showing the sequence of dispatch instructions, station actions, and maintenance actions, along with a simulation video.
Claims
1. A smart decision-making and drill system for rail transit dispatching and maintenance, characterized in that: It includes a problem extraction module, a knowledge base construction module, a problem-solving module, a train operation adjustment module, a human-computer interaction interface module, and a solution output module.
2. The intelligent decision-making and drill system for rail transit dispatching and maintenance according to claim 1, characterized in that, The question extraction module, based on the AI model Qwen3, performs semantic understanding and keyword extraction on the user-input description of sudden event faults or query questions. The Qwen3 model is based on the Transformer decoder architecture, and its core structure includes an embedding layer, a multi-layer decoder module, and an output layer. The embedding layer converts the input text into a vector representation; each decoder module contains a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer captures the dependencies between different words in the text, and the feedforward neural network layer performs non-linear transformations; the layer normalization module RMSnorm stabilizes the training process; the output layer generates a probability distribution on the vocabulary through linear transformations and outputs classification results containing key information such as fault type, location of occurrence, and scope of impact.
3. The intelligent decision-making and drill system for rail transit dispatching and maintenance according to claim 1, characterized in that, The knowledge base construction module is used to store and classify typical rail transit fault cases, emergency response regulations, industry standards, and basic data including BIM models of operating environments and equipment such as stations and lines, as well as equipment parameters. It also uses Anylogic simulation software to perform virtual simulation modeling of typical fault scenarios, and constructs a knowledge base that includes a fault case library, a regulation and standard library, and a simulation model library.
4. The intelligent decision-making and drill system for rail transit dispatching and maintenance according to claim 1, characterized in that, The problem-solving module employs Retrieval Enhancement Generative Intelligence (RAG) technology. Based on key information output by the problem extraction module, it performs multi-source searches in the knowledge base construction module, matching relevant emergency failure cases, regulatory standards, and virtual simulation models. Simultaneously, it utilizes AI large-scale model technology to deeply analyze the search results. This deep analysis process includes: the large model performing similarity comparisons and extracting handling experience from the retrieved failure cases; compliance screening and applicability judgment of regulatory clauses; scenario matching and key parameter extraction from the simulation model; and, by integrating the above information with the real-time context input by the user, generating a preliminary emergency response plan framework through reasoning. This framework covers the division of responsibilities, key operational steps, collaborative interaction nodes, and safety precautions for dispatching, operations, and maintenance positions. Then, it optimizes the solution based on the user's actual problem, ultimately generating an emergency response plan to support decision-making. This plan includes key emergency response points for dispatching, operations, and maintenance positions, a collaborative response flowchart, and a virtual simulation video.
5. The intelligent decision-making and drill system for rail transit dispatching and maintenance according to claim 1, characterized in that, The train operation adjustment module, when a sudden fault affects train operation, constructs a train operation adjustment optimization model based on real-time operation diagram data, train positions, and track occupancy status, with the goal of minimizing total delay time and passenger waiting time. This model is a typical N-Phard problem, namely a bi-objective mixed integer linear programming problem. The specific calculation process is as follows: First, data including train times, interval travel time, station dwell time, train sequence, and track occupancy status in the current operation diagram are collected, as well as the affected section and the expected duration of the fault. Then, an optimization model is constructed, with decision variables including the adjustment amount of departure time for each train, the adjustment amount of stop time, the section operation level, and the choice of train turning back or stopping at intermediate stations when necessary; constraints include: train tracking interval time constraint, section operation time constraint, station stop time constraint, and line occupancy uniqueness constraint. The objective function is to minimize the weighted sum of total delay time and passenger waiting time. An improved Grey Wolf optimization algorithm is used to solve this problem: each solution is represented as a vector of train adjustment amounts. The population is initialized, the fitness value of each individual (i.e., the objective function value) is calculated, the Grey Wolf positions are updated iteratively, and mutation operations are introduced to avoid local optima. The optimal solution is output after the termination condition is met. Finally, an adjusted train timetable is generated based on the optimal solution, and the train operation diagram is redrawn to provide decision support for dispatchers.
6. The intelligent decision-making and drill system for rail transit dispatching and maintenance according to claim 1, characterized in that, The solution output module evaluates and analyzes the emergency response plan output by the thinking and solving module. By comparing the effects of historical cases, simulation results, and regulatory compliance checks, it outputs an optimized auxiliary decision-making plan, which is presented in a graphical manner. This includes a flowchart of the collaborative handling process among dispatching, operation, and maintenance personnel, key energy points, and safety precautions. It can also play corresponding simulation videos to intuitively demonstrate the handling process.
7. The intelligent decision-making and drill system for rail transit dispatching and maintenance according to any one of claims 1 to 6, characterized in that, The human-computer interaction interface module adopts a web-based application platform, integrates the above modules, provides users with a unified input interface, supports fault query, solution generation, and operation diagram adjustment functions, and displays the results in the form of visual flowcharts and simulation videos.
8. The intelligent decision-making and drill system for rail transit dispatching and maintenance according to claim 2, characterized in that, In the knowledge base construction module, the fault case library is derived from historical accident reports, operation and maintenance records and theoretical knowledge, and is structured; the rules and standards library covers relevant standards at the national, industry and enterprise levels; the simulation model library uses tools such as Anylogic and BIM to build detailed models of equipment in typical station, section and vehicle environments, which can dynamically simulate passenger flow evacuation, train operation and equipment maintenance under different fault scenarios.
9. The intelligent decision-making and drill system for rail transit dispatching and maintenance according to claim 3, characterized in that, The graphical flowchart generated by the solution output module clearly defines the task sequence and interaction relationship of each position in scheduling, operation, and maintenance, and supports exporting as an image or PDF; at the same time, it can generate corresponding simulation videos to intuitively show the handling process.
10. The intelligent decision-making and drill system for rail transit dispatching and maintenance according to claim 4, characterized in that, The train operation adjustment module employs various intelligent algorithms to solve a multi-objective optimization problem. The decision variables are the arrival and departure times, stop times, and interval running order of each train. Constraints include tracking intervals, interval running times, and station capacity. The objective function is to minimize the weighted total delay time and passenger waiting time. The solution process is as follows: First, the multi-objective problem is transformed into a single objective using a weighted summation method, with weights dynamically adjusted according to operational needs. Then, an improved Grey Wolf optimization algorithm or Particle Swarm Optimization algorithm is used for the solution. The algorithm flow includes initializing the population, calculating fitness, updating individual and global optima, and iterating until convergence. In each iteration, the adjustment amount for each train is calculated based on the current solution, and it is checked whether all constraints are met. If not, corrections are made. Finally, the adjustment scheme that satisfies the constraints and has the optimal objective function is output, i.e., the new train timetable.