Real-time fault-tolerant decision-making system for multi-source sensing data of rail transit equipment

Abstract

The invention relates to the field of rail transit equipment, and discloses a rail transit equipment multi-source sensing data real-time fault-tolerant decision-making system, which comprises a multi-source sensing module used for collecting rail environment and locomotive state data through a 3D laser radar, a binocular camera and a Beidou positioning terminal, including obstacle, turnout, signal lamp and locomotive position information; the data fusion processing module is used for carrying out preprocessing, redundancy check and feature extraction on the multi-source sensing data to generate a fusion sensing result; and the fault-tolerant decision module is used for judging the safety state of the track based on the fusion sensing result, triggering an alarm / braking instruction and executing self-checking logic when the sensor is abnormal. Through the combination of double groups of redundant sensors and a three-level progressive fault-tolerant strategy, closed-loop control from data acquisition to safety decision is formed, the system robustness under complex working conditions is remarkably improved, and an omnibearing safety protection barrier is constructed for rail transit equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of rail transit equipment, and in particular to a real-time fault-tolerant decision-making system for multi-source perception data of rail transit equipment. Background Art

[0002] As rail transit equipment develops towards intelligent and unmanned operations, scenarios such as automated port terminals, trunk railway transportation, and urban rail transit have placed stringent requirements on driving safety and the reliability of environmental perception. On complex track paths, such as curves and switch areas, and in harsh environments such as dust, rain, snow, and strong light, locomotives need to accurately detect obstacles, identify signal light status, and obtain global position information in real time to avoid collision risks and ensure the coordinated operation of trackside equipment. Traditional single-sensor detection solutions are difficult to cover all scenarios, while multi-sensor collaborative systems face problems such as low data fusion efficiency and insufficient fault handling capabilities. There is an urgent need for a safety decision-making system that can achieve efficient collaboration of multi-source data and has hierarchical fault tolerance to meet the needs of high-precision and high-reliability driving safety.

[0003] In existing technologies, rail transit environment perception systems either rely on a single sensor, resulting in prominent detection blind spots and single equipment failures that easily lead to monitoring interruptions. Alternatively, although multiple sensors are used, there are problems such as insufficient data registration accuracy and simple conflict handling mechanisms. This makes it difficult to integrate the advantages of multi-source data and the reliability of decision-making basis is low. In addition, there is a general lack of hierarchical fault-tolerant strategies for single sensor anomalies, dual-source data conflicts and system-level failures. In extreme scenarios, it is impossible to respond quickly and trigger safety instructions. The overall detection reliability and fault handling capabilities are insufficient, making it difficult to meet the stringent requirements of intelligent rail transit equipment for high-precision, high-reliability safety decisions. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a real-time fault-tolerant decision-making system for multi-source perception data of rail transit equipment, which solves the problems of the existing rail transit environment perception system, such as prominent single-sensor detection blind spots, insufficient multi-sensor fusion accuracy, and weak conflict handling capabilities. In addition, the system lacks a hierarchical fault-tolerant strategy, resulting in low detection reliability and delayed fault response under complex working conditions, making it difficult to meet the safety decision-making needs of intelligent rail transit.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: The real-time fault-tolerant decision-making system for multi-source perception data of rail transit equipment includes: A multi-source perception module, which uses 3D lidar, binocular cameras, and Beidou positioning terminals to collect track environment and locomotive status data, including information on obstacles, switches, signals, and locomotive positions; The data fusion processing module is used to pre-process, perform redundancy check and feature extraction on multi-source perception data to generate fusion perception results; A fault-tolerant decision-making module is used to determine the track safety status based on the fused perception results, trigger alarms / brake commands, and execute self-check logic when a sensor anomaly occurs; Real-time communication module, used to interact with the locomotive control system through the CAN bus to transmit instructions and status information; Visualization module, used to display locomotive position, sensor status and safety warning information.

[0006] Preferably, the multi-source perception module includes a 3D laser radar unit, a binocular camera unit and a Beidou positioning unit. The 3D laser radar unit is used to detect obstacles in front of the track and the position of switches through a point cloud algorithm and output point cloud processing results. The binocular camera unit is used to identify obstacles, switches and signal light status through a deep neural network and output image analysis results. The Beidou positioning unit is used to obtain the locomotive's global position information in real time and upload it to the visualization module.

[0007] Preferably, the 3D lidar unit uses a pass filter, a statistical filter, and a voxel filtering algorithm to perform noise reduction and regional screening on the point cloud data, and combines the RANSAC algorithm to perform track plane segmentation and switch detection, outputting obstacle distance and track geometric features. The binocular camera unit realizes dynamic obstacle recognition and signal light language analysis through camera calibration and deep learning models.

[0008] Preferably, the data fusion processing module includes a redundant check unit and a feature fusion unit. The redundant check unit is used to compare the detection results of the 3D laser radar and the binocular camera. If obstacles exist in both sets of data, an alarm is immediately triggered. If a single set of data is abnormal for 5 consecutive frames, an early warning is triggered. The feature fusion unit is used to integrate the laser point cloud features and the image target detection results to generate a comprehensive track safety status assessment.

[0009] Preferably, the fault-tolerant decision module includes a self-check trigger unit and a dynamic decision unit. The self-check trigger unit is used to monitor abnormal data caused by sensor occlusion or extreme weather. If the occlusion area exceeds a threshold or the abnormality lasts for more than 10 frames, the system failure is determined and a request for manual intervention is reported. The dynamic decision unit is used to determine whether the switch position is correct and whether the traffic light allows passage based on the fusion perception results. If any condition is not met, a braking command is triggered.

[0010] Preferably, the fault-tolerant decision module implements a three-level fault-tolerant strategy, which is in descending order of fault severity: The first-level strategy targets single-sensor anomalies. When the point cloud density drops by more than 30% or the image grayscale is abnormal, the anomaly is marked and the opposite sensor is triggered to take over. The switching time does not exceed 100ms. The secondary strategy, for conflicting detection results from two sensors, selects a data source with a confidence level exceeding 95% based on DS evidence theory and simultaneously initiates a manual review request. The third-level strategy targets system-level failures. When both sets of sensors fail, emergency braking is triggered immediately, with a response time of no more than 200ms and reporting to the entire system fault log.

[0011] Preferably, the real-time communication module interacts with the locomotive control system via a standard CAN bus protocol, the output fields include obstacle status, switch status, signal light status and system self-test fault codes, and the input fields include locomotive travel direction, speed and position information.

[0012] Preferably, the CAN bus output frame data field includes obstacle distance, sensor status and heartbeat signal, the heartbeat signal is used to monitor the online status of the sensor, and the input frame parses the train operation instruction and synchronizes it to the multi-source perception module to adjust the detection direction.

[0013] Preferably, the visualization module integrates Beidou positioning data through the port management platform, displays the locomotive position trajectory in real time, and distinguishes between safety status, warning status and fault status through color coding.

[0014] Preferably, the system uses the SD System robot operating system to implement data interaction and task scheduling between modules, including lidar drive, binocular camera processing, decision logic and communication interface function packages. Each function package implements data interaction through a standardized interface and supports dynamic configuration of detection parameters, including obstacle warning distance, switch calibration tolerance and self-test trigger threshold.

[0015] The present invention provides a real-time fault-tolerant decision-making system for multi-source perception data of rail transit equipment. It has the following beneficial effects: 1. The present invention achieves full-dimensional perception of the track environment through the multi-source coordinated deployment of 3D lidar, binocular camera and Beidou positioning, effectively covering key information such as obstacle detection, switch status identification, signal light analysis and global positioning. The dual sets of redundant sensors are combined with a three-level progressive fault-tolerant strategy to provide layered protection against single sensor anomalies, dual-source data conflicts and system-level failures. When a single sensor anomaly occurs, it automatically switches to the opposite device to maintain detection continuity. In the event of a dual-source conflict, high-confidence data is screened based on evidence theory. In the event of a system-level failure, emergency braking is quickly triggered, forming a closed-loop control from data acquisition to safety decision-making, significantly improving the system robustness under complex working conditions, and building a comprehensive safety protection barrier for rail transit equipment.

[0016] 2. The data fusion processing module of this invention deeply integrates laser point cloud and visual image features, generating comprehensive safety assessment results through spatiotemporal registration and redundancy verification, effectively resolving multi-sensor data conflicts and providing more reliable environmental perception information for the decision-making module. The real-time communication module achieves efficient data exchange based on the industrial-grade CAN bus, dynamically adapts to the locomotive's operating status to adjust detection strategies, and transmits control commands to the actuator in real time, ensuring precise triggering and rapid response to safety commands such as braking and deceleration. This design overcomes the detection blind spots and decision-making lag limitations of traditional single-sensor systems, achieving efficient "perception-fusion-control" collaboration and improving the real-time and accuracy of driving safety control.

[0017] 3. The visualization module of the present invention integrates Beidou positioning and multi-source detection results, and displays the locomotive position, operating status and warning information in real time through an intuitive human-machine interface, supporting operators to quickly identify anomalies and intervene. The historical data tracing function realizes the correlated playback of multi-source data at the time of the fault, significantly shortening the fault location and troubleshooting time. The SD System robot operating system adopts a modular design. Each functional package interacts through a standardized interface, supports dynamic configuration of detection parameters and independent maintenance. The failure of a single module does not affect the overall operation of the system, significantly improving the system's scalability and fault resistance. This feature meets the long-term stable operation requirements of high-load scenarios such as ports and railways, and provides efficient and convenient technical support for intelligent operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a diagram of the architecture of the present invention; Figure 2 This is a diagram of the multi-source perception module architecture of the present invention; Figure 3 This is a diagram of the fault-tolerant decision module architecture of the present invention; Figure 4 This is a diagram showing the relationship between the modules of the present invention; Figure 5 Schematic diagram of the system software structure of the present invention; Figure 6 is a system logic diagram of the present invention; Figure 7 A schematic diagram of the installation of the locomotive forward viewing device of the present invention; Figure 8 This is a simplified installation diagram of the locomotive rearview device of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solution of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] Please see the attached Figure 1 -Attached Figure 8 The embodiment of the present invention provides a real-time fault-tolerant decision-making system for multi-source perception data of rail transit equipment, comprising: A multi-source perception module, which uses 3D lidar, binocular cameras, and Beidou positioning terminals to collect track environment and locomotive status data, including information on obstacles, switches, signals, and locomotive positions; The data fusion processing module is used to preprocess, perform redundancy check and feature extraction on multi-source perception data to generate fusion perception results; A fault-tolerant decision-making module is used to determine the track safety status based on the fused perception results, trigger alarms / brake commands, and execute self-check logic when a sensor anomaly occurs; Real-time communication module, used to interact with the locomotive control system through the CAN bus to transmit instructions and status information; Visualization module, used to display locomotive position, sensor status and safety warning information.

[0021] Please see the attached Figure 1 , Attachment Figure 2 and attached Figure 4 The multi-source perception module includes a 3D lidar unit, a binocular camera unit and a Beidou positioning unit. The 3D lidar unit is used to detect obstacles in front of the track and the position of switches through a point cloud algorithm and output point cloud processing results. The binocular camera unit is used to identify obstacles, switches and signal light status through a deep neural network and output image analysis results. The Beidou positioning unit is used to obtain the locomotive's global position information in real time and upload it to the visualization module.

[0022] The 3D lidar unit uses pass-through filters, statistical filters, and voxel filtering algorithms to reduce noise and perform regional screening on point cloud data. It combines the RANSAC algorithm for track plane segmentation and switch detection, and outputs obstacle distances and track geometric features. The binocular camera unit uses camera calibration and deep learning models to achieve dynamic obstacle recognition and signal light language analysis.

[0023] Specifically, the multi-source perception module builds full-dimensional environmental perception capabilities through the deployment of dual sets of redundant sensors and multimodal fusion technology.

[0024] The 3D LiDAR unit, with two high-resolution LiDAR sensors integrated at the front and rear of the locomotive, utilizes a three-stage noise reduction algorithm (through filtering to remove ground clutter, statistical filtering to eliminate outliers, and voxel filtering to compress point cloud data) combined with the RANSAC algorithm for track plane segmentation and turnout detection. This design achieves high-precision obstacle distance detection within a range of 0.5 to 150 meters, with an error of ±5 cm, and precise turnout position identification with an error of ≤10 cm. This eliminates blind spots within 2 meters in front and behind the locomotive, ensuring comprehensive monitoring in complex scenarios such as curves and turnouts, and providing precise geometric feature data for trackside equipment status verification.

[0025] The binocular camera unit, with a pair of stereo cameras deployed at the front and rear of the train, utilizes precise calibration of external parameter errors to ≤0.5px and time synchronization technology to ≤1ms. Combined with the YOLOv8 deep neural network, it enables dynamic obstacle recognition and signal light interpretation. The depth information provided by the binocular vision system accurately distinguishes pedestrians, vehicles, and trackside equipment. Even in rainy and foggy environments with visibility less than 10m or strong sunlight greater than 10,000 lux, it can reliably identify dynamic targets within a range of 1.5 to 50m, providing the decision-making module with rich semantic information such as target type and direction of movement.

[0026] The BeiDou positioning terminal utilizes the BeiDou-3 RTK high-precision positioning solution with an accuracy of ±0.1m. Through hard synchronization of PPS signals with sensor timestamps, it converts global coordinates into the locomotive coordinate system with an error of ≤2cm. It records the locomotive's position trajectory in real time and supports 72 hours of historical data backtracking. This function provides a unified spatiotemporal reference for multi-source data, enabling precise correlation of location, image, and point cloud, providing reliable geographic information support for remote scheduling and fault analysis.

[0027] Please see the attached Figure 1 , Attachment Figure 2 and attached Figure 4 The data fusion processing module includes a redundant check unit and a feature fusion unit. The redundant check unit is used to compare the detection results of the 3D laser radar and the binocular camera. If there are obstacles in both sets of data, an alarm is triggered immediately. If a single set of data is abnormal for 5 consecutive frames, an early warning is triggered. The feature fusion unit is used to integrate the laser point cloud features and image target detection results to generate a comprehensive track safety status assessment.

[0028] Specifically, the data fusion processing module improves data reliability and the quality of decision-making basis through a double verification mechanism.

[0029] The detection results of the lidar and binocular camera are compared in real time. When both confirm the existence of an obstacle, the highest priority alarm is triggered within 50ms, ensuring that the risk of missed detection in complex environments is reduced to an extremely low level. If a single group of sensors shows an abnormality for five consecutive frames, such as a sudden drop in point cloud density or blurred images, the system automatically switches to the opposite sensor to take over the detection task. The switching process is completed within 100ms, maintaining 100% detection coverage, effectively avoiding monitoring interruptions caused by single device failures, and ensuring detection continuity under extreme working conditions.

[0030] Using a spatiotemporal registration algorithm, the laser point cloud coordinates and camera pixel coordinates are aligned to the locomotive coordinate system with an error of ≤5cm. Combined with the target confidence output by a deep neural network, a comprehensive track safety status assessment is generated, encompassing location, type, and risk level. This technology effectively integrates the advantages of multiple sensors, providing comprehensive and reliable input to the decision-making module. Even when sensor data conflicts slightly, accurate environmental assessment results can be output, ensuring the scientific nature of subsequent safety decisions.

[0031] Please see the attached Figure 1 , Attachment Figure 2 and attached Figure 4 The fault-tolerant decision-making module includes a self-check trigger unit and a dynamic decision-making unit. The self-check trigger unit is used to monitor abnormal data caused by sensor occlusion or extreme weather. If the occlusion area exceeds the threshold or the abnormality lasts for more than 10 frames, the system fault is determined and a manual intervention request is reported. The dynamic decision-making unit is used to determine whether the switch position is correct and whether the traffic light allows passage based on the fusion perception results. If any condition is not met, the braking command is triggered.

[0032] The fault-tolerance decision module implements a three-level fault-tolerance strategy, which is ranked from low to high in terms of fault severity: The first-level strategy targets single-sensor anomalies. When the point cloud density drops by more than 30% or the image grayscale is abnormal, the anomaly is marked and the opposite sensor is triggered to take over. The switching time does not exceed 100ms. The secondary strategy, for conflicting detection results from two sensors, selects a data source with a confidence level exceeding 95% based on DS evidence theory and simultaneously initiates a manual review request. The third-level strategy targets system-level failures. When both sets of sensors fail, emergency braking is triggered immediately, with a response time of no more than 200ms and reporting to the entire system fault log.

[0033] Specifically, the fault-tolerant decision-making module builds a comprehensive security protection system through a three-level progressive strategy. In the first-level strategy, when a single sensor group experiences data anomalies due to factors such as occlusion or dust, such as a sudden drop in LiDAR point cloud density or abnormal camera image grayscale values, the system immediately flags the anomalous sensor and triggers the opposing sensor to take over detection. This process is completed within 100ms, ensuring that the detection process is not affected by temporary interference. Even in harsh environments such as port dust and snow-covered railways, stable detection performance is maintained, providing continuous environmental monitoring for locomotive operations.

[0034] In the secondary strategy, when the LiDAR and camera detection results conflict for three consecutive frames, the system calculates the confidence level of each data source based on DS evidence theory, prioritizes the data with a higher confidence level (>95%), and simultaneously sends a review request to the manual monitoring platform. This mechanism effectively resolves the issue of inconsistent multi-sensor data, ensuring accurate decision-making in complex scenarios, such as those caused by reflective objects, and providing a reliable basis for generating safety instructions.

[0035] The third-level strategy triggers an emergency braking command if both sensor groups fail or a critical decision module experiences an anomaly. The system outputs the braking signal within 200 milliseconds, ensuring the train stops quickly in extreme failure scenarios, with the braking distance error controlled within 1 meter. The system also automatically records raw data from the 10 seconds preceding the failure and reports it to the operations and maintenance platform, providing a complete information chain for post-fault analysis and minimizing the risk of safety incidents.

[0036] Based on the fused perception results, the dynamic decision-making unit determines the switch position and signal light status in real time. When any condition does not meet the safety requirements, it immediately triggers the corresponding control command warning, deceleration or braking to ensure that the locomotive operation is in line with the status of the trackside equipment, avoiding safety hazards caused by signal misjudgment or switch abnormalities from the source.

[0037] Based on the fused perception results, the dynamic decision-making unit determines the switch position and signal light status in real time. When any condition does not meet the safety requirements, it immediately triggers the corresponding control command (warning, deceleration or braking) to ensure that the locomotive operation complies with the status of the trackside equipment, avoiding safety hazards caused by signal misjudgment or switch abnormalities from the source.

[0038] Please see the attached Figure 1 , Attachment Figure 4 and attached Figure 5 ,The real-time communication module interacts with the locomotive control system through the ,standard CAN bus protocol. The output fields include obstacle status, switch status, signal light status and ,system self-test fault codes. The input fields include the locomotive ,travel direction, speed and position information.

[0039] The CAN bus output frame data field contains obstacle distance, sensor status and heartbeat signal. The heartbeat signal is used to monitor the online status of the sensor. The input frame parses the train operation instructions and synchronizes them to the multi-source perception module to adjust the detection direction.

[0040] Specifically, the real-time communication module realizes efficient and reliable data interaction through the industrial-grade CAN bus. The output signal transmits the obstacle status distance, position, whether the switch status position is correct, signal light status light language analysis results and system self-test fault codes to the locomotive control system in real time, providing accurate environmental data for the braking system, supporting its real-time calculation of key parameters such as braking pressure and distance, and ensuring the accuracy and timeliness of braking action.

[0041] The system receives input signals, receives operating parameters such as the locomotive's direction, speed, and position, and dynamically adjusts sensor detection strategies. For example, it increases the detection frequency of the rear sensor when reversing, and focuses on side obstacle monitoring when turning, so that system resource allocation matches the current operating conditions, reduces invalid data processing, and improves overall operating efficiency.

[0042] The CAN bus uses AES-128 encryption technology and priority management mechanism to ensure the security and real-time performance of data transmission. Even in strong electromagnetic interference environments, it can still guarantee the accurate interaction of control instructions and status information, providing stable communication support for the collaborative work of various system modules.

[0043] Please see the attached Figure 1 and attached Figure 4 ,The visualization module integrates Beidou positioning data through the port ,management platform, and displays the locomotive position trajectory in real time, and ,differentiates between safety status, warning status and fault status through color ,coding.

[0044] Specifically, the visualization module improves operation and maintenance efficiency and decision-making transparency through an intuitive human-computer interaction interface and powerful data traceability functions. It presents the locomotive position trajectory with an accuracy of ±0.1m in real time on the port management platform, and clearly distinguishes the operating status through color coding of green, yellow and red. Operators can identify abnormal situations within 3 seconds, significantly shortening emergency response time and ensuring timely response measures.

[0045] It stores 72 hours of historical data and supports the associated playback of multi-source data point clouds, images, and positioning information. When a fault occurs, operation and maintenance personnel can quickly locate the fault timestamp and simultaneously view the sensor data and device status at that moment, shortening the fault location time from 15 minutes in traditional solutions to less than 2 minutes, significantly improving troubleshooting efficiency.

[0046] It provides a friendly human-machine interface, supporting operators to dynamically adjust detection parameters according to actual working conditions, such as shortening the obstacle warning distance in bad weather and temporarily modifying the turnout calibration tolerance. Configuration instructions take effect in real time via the CAN bus, enabling the system to flexibly adapt to different operating scenarios and enhance overall environmental adaptability.

[0047] Please see the attached Figure 1 , Attachment Figure 4 and attached Figure 5 The system uses the SD System robot operating system to realize data interaction and task scheduling between modules, including lidar drive, binocular camera processing, decision logic and communication interface function packages. Each function package realizes data interaction through a standardized interface and supports dynamic configuration of detection parameters, including obstacle warning distance, switch calibration tolerance and self-test trigger threshold.

[0048] Specifically, the system achieves efficient module collaboration and task scheduling through the independently developed SD System robot operating system, and adopts a priority preemption algorithm to ensure that key tasks such as emergency braking and fault handling are executed first. The end-to-end delay from data collection to decision output is controlled within 200ms, meeting the strict real-time requirements of rail transit and providing underlying guarantees for the rapid generation and execution of safety instructions.

[0049] Through a high-precision time synchronization protocol with an error of ≤1ms, the spatiotemporal alignment of data from lidar, cameras, and Beidou terminals is achieved, avoiding detection errors caused by time deviations, such as the risk of distance misjudgment caused by asynchrony, ensuring the accuracy of multi-source data fusion and providing a reliable time benchmark for subsequent decision-making.

[0050] The system is divided into functional packages, including lidar driver, binocular camera processing, decision logic, and communication interfaces. Each module exchanges data through standardized interfaces, supporting independent deployment and maintenance. In the event of a single module failure, the system automatically isolates the faulty unit and activates redundancy to ensure overall functionality is not affected. This significantly improves system reliability, achieving a mean time between failures (MTBF) exceeding 10,000 hours, meeting the long-term stable operation requirements of high-load scenarios such as ports and railways.

[0051] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A real-time fault-tolerant decision-making system for multi-source perception data of rail transit equipment, characterized by: include: A multi-source perception module, which uses 3D lidar, binocular cameras, and Beidou positioning terminals to collect track environment and locomotive status data, including information on obstacles, switches, signals, and locomotive positions; The data fusion processing module is used to pre-process, perform redundancy check and feature extraction on multi-source perception data to generate fusion perception results; A fault-tolerant decision-making module is used to determine the track safety status based on the fused perception results, trigger alarms / brake commands, and execute self-check logic when a sensor anomaly occurs; Real-time communication module, used to interact with the locomotive control system through the CAN bus to transmit instructions and status information; Visualization module, used to display locomotive position, sensor status and safety warning information.

2. The rail transit equipment multi-source perception data real-time fault-tolerant decision-making system according to claim 1 is characterized in that: The multi-source perception module includes a 3D laser radar unit, a binocular camera unit and a Beidou positioning unit. The 3D laser radar unit is used to detect obstacles in front of the track and the position of switches through a point cloud algorithm and output point cloud processing results. The binocular camera unit is used to identify obstacles, switches and signal light status through a deep neural network and output image analysis results. The Beidou positioning unit is used to obtain the locomotive's global position information in real time and upload it to the visualization module.

3. The real-time fault-tolerant decision-making system for multi-source perception data of rail transit equipment according to claim 2 is characterized in that: The 3D lidar unit uses a pass filter, a statistical filter, and a voxel filter algorithm to reduce noise and perform regional screening on point cloud data. It combines the RANSAC algorithm to perform track plane segmentation and switch detection, and outputs obstacle distance and track geometric features. The binocular camera unit uses camera calibration and a deep learning model to achieve dynamic obstacle recognition and signal light language analysis.

4. The rail transit equipment multi-source perception data real-time fault-tolerant decision-making system according to claim 1 is characterized in that: The data fusion processing module includes a redundancy check unit and a feature fusion unit. The redundancy check unit is used to compare the detection results of the 3D laser radar and the binocular camera. If obstacles exist in both sets of data, an alarm is immediately triggered. If a single set of data is abnormal for five consecutive frames, an early warning is triggered. The feature fusion unit is used to integrate the laser point cloud features and image target detection results to generate a comprehensive track safety status assessment.

5. The rail transit equipment multi-source perception data real-time fault-tolerant decision-making system according to claim 1 is characterized in that: The fault-tolerant decision-making module includes a self-check trigger unit and a dynamic decision-making unit. The self-check trigger unit is used to monitor abnormal data caused by sensor occlusion or extreme weather. If the occlusion area exceeds a threshold or the abnormality lasts for more than 10 frames, the system failure is determined and a manual intervention request is reported. The dynamic decision-making unit is used to determine whether the switch position is correct and whether the traffic light allows passage based on the fusion perception results. If any condition is not met, a braking command is triggered.

6. The rail transit equipment multi-source perception data real-time fault-tolerant decision-making system according to claim 5 is characterized in that: The fault-tolerance decision module implements a three-level fault-tolerance strategy, which is ranked from low to high in terms of fault severity: The first-level strategy targets single-sensor anomalies. When the point cloud density drops by more than 30% or the image grayscale is abnormal, the anomaly is marked and the opposite sensor is triggered to take over. The switching time does not exceed 100ms. The secondary strategy, for conflicting detection results from two sensors, selects a data source with a confidence level exceeding 95% based on DS evidence theory and simultaneously initiates a manual review request. The third-level strategy targets system-level failures. When both sets of sensors fail, emergency braking is triggered immediately, with a response time of no more than 200ms and reporting to the entire system fault log.

7. The rail transit equipment multi-source perception data real-time fault-tolerant decision-making system according to claim 1 is characterized in that: The real-time communication module interacts with the locomotive control system through the standard CAN bus protocol. The output fields include obstacle status, switch status, signal light status and system self-test fault codes, and the input fields include locomotive direction, speed and position information.

8. The rail transit equipment multi-source perception data real-time fault-tolerant decision-making system according to claim 7 is characterized in that: The CAN bus output frame data domain contains obstacle distance, sensor status and heartbeat signal. The heartbeat signal is used to monitor the online status of the sensor. The input frame parses the train operation instruction and synchronizes it to the multi-source perception module to adjust the detection direction.

9. The rail transit equipment multi-source perception data real-time fault-tolerant decision-making system according to claim 1 is characterized in that: The visualization module integrates Beidou positioning data through the port management platform, displays the locomotive position trajectory in real time, and distinguishes between safety status, warning status and fault status through color coding.

10. The real-time fault-tolerant decision-making system for multi-source perception data of rail transit equipment according to any one of claims 1 to 9, characterized in that: The system uses the SD System robot operating system to implement data exchange and task scheduling between modules. It includes lidar drive, binocular camera processing, decision logic, and communication interface function packages. Each function package implements data exchange through a standardized interface and supports dynamic configuration of detection parameters, including obstacle warning distance, switch calibration tolerance, and self-test trigger threshold.

Images