Rail transit data processing system adopting high-performance computing power node board card

Abstract

The invention relates to the technical field of rail transit data processing, in particular to a rail transit data processing system adopting a high-performance computing power node board card, which is used for solving the problems that the existing rail transit data processing technology has obvious defects in a complex operation scene, the state prediction depends on a black box model, and the reliability is poor. Explicit modeling of physical constraints such as track curvature, gradient and speed limit is lacked, and real anomaly and model misalignment are difficult to distinguish; structured features and compressed time sequence data are fused through the rail traffic physical credible prediction module, the train operation stage is intelligently recognized, adaptive deep learning, time sequence Bayesian or reinforcement learning sub-models are dynamically called, high-precision state prediction is achieved, meanwhile, physical constraints such as the track curvature, the gradient and the speed limit are combined, and the reliability and reliability of the system are improved. And the causal consistency verification based on the mechanics principle is executed, so that the physical mismatch abnormity is effectively discriminated, and the prediction credibility, the abnormity identification accuracy and the operation and maintenance decision intelligent level are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rail transit data processing, in particular to a rail transit data processing system using high-performance computing node board cards. BACKGROUND

[0002] The current rail transit system needs to process multi-source heterogeneous data in real time in automatic control, intelligent operation and maintenance and video analysis. The traditional industrial control or embedded platform is difficult to support AI reasoning, multi-channel video and high-speed signal fusion tasks. The existing acceleration scheme is mostly deployed in the whole machine, which has high power consumption, large volume, poor heat dissipation and poor electromagnetic compatibility, and is difficult to meet the strict requirements of rail transit in reliability, environmental adaptability and installation space. The system lacks modular high-performance computing units, and the resource scheduling is not flexible. Although the high-performance computing node board card integrates multi-core processors, AI acceleration engines and high-speed interfaces, it has not been effectively integrated into the rail transit data processing system, and there are still technical gaps in master control cooperation, multi-board parallelism, redundancy fault tolerance and vehicle-mounted network docking.

[0003] The patent application with publication number CN111397593A discloses a rail transit equipment positioning and guiding data processing system and method. The rail transit equipment navigation data processing system comprises a geological judgment module and a self-adaptive selection module. The geological judgment module is used to judge the geological type in front according to the input advanced drilling data, and divide it into two types: normal geology and abnormal geology. The self-adaptive selection module is used to select different processing methods for laser navigation data according to different geological types. The beneficial effects of the present application are: input the data of advanced drilling into the processing system as the basis for geological judgment, select appropriate data processing methods according to the geological type, correct the navigation data, and then output the corrected navigation data, so as to realize the closed-loop control between the positioning system and the equipment body, better adapt to various working conditions, and realize high precision and high accuracy of large rail transit equipment tunneling route; However, the existing rail transit data processing technology has obvious deficiencies in complex operation scenarios. The state prediction relies on black box models, lacks explicit modeling of physical constraints such as rail curvature, slope and speed limit, and is difficult to distinguish between real anomalies and model inaccuracies. The task scheduling generally adopts a static strategy, which cannot dynamically adjust the priority of data processing and abnormal response tasks according to the train operation stage and rail risk, and does not adapt to the computing characteristics and real-time load of heterogeneous computing power such as CPU, GPU and FPGA, resulting in delayed response of key tasks or low resource utilization.

[0004] Therefore, the present application proposes a rail transit data processing system using high-performance computing node board cards to solve the above problems. SUMMARY

[0005] The application aims to provide a rail transit data processing system using high-performance computing node board cards, solve the obvious deficiencies of existing rail transit data processing technology in complex operation scenarios, and the state prediction relies on black box models, lacks explicit modeling of physical constraints such as rail curvature, slope and speed limit, and it is difficult to distinguish between real abnormalities and model inaccuracies; the task scheduling generally uses static strategies, cannot dynamically adjust the priority of data processing and abnormal response tasks according to the train operation stage and rail risk, and does not adapt to the computing characteristics and real-time load of heterogeneous computing power such as CPU, GPU and FPGA, resulting in the problems of delayed response of key tasks or low resource utilization.

[0006] The application achieves the above-mentioned purpose through the following technical solutions. A rail transit data processing system using high-performance computing node board cards is applied to a rail transit vehicle-mounted intelligent terminal and includes: A rail transit multi-source data preprocessing module is used to collect multi-source heterogeneous sensing parameters from the train operation process, track structure, operation environment and vehicle-mounted sensing system, preprocess the collected multi-source heterogeneous sensing parameters, and output structured feature vectors and compressed time series data. A rail transit physically credible prediction module is used to call deep learning, time series Bayesian or reinforcement learning prediction sub-models according to the train operation stage based on the structured feature vectors and compressed time series data, perform abnormal physical causal verification using rail curvature, slope and speed limit, predict the coupling evolution trend of abnormalities and rails and generate abnormal warning information. A rail transit heterogeneous computing power scheduling module is used to schedule data processing tasks based on structured feature vectors and compressed time series data and abnormal response tasks based on abnormal warning information, dynamically adjust the task priority using deep reinforcement learning and fuzzy logic fusion methods according to the train operation stage and rail topology prior constraints, and map the tasks to adaptive heterogeneous computing power resources in CPU, GPU or FPGA to perform dynamic scheduling and task migration in the case of computing node failure. A rail transit computing power thermal control module is used to monitor the hot spot area of the board card in real time according to the task load on CPU, GPU and FPGA, dynamically control the local cooling of the cooling execution mechanism, and adaptively degrade the computing performance when the temperature exceeds the preset safety threshold.

[0007] As a preferred embodiment of the application, the process of the rail transit physically credible prediction module based on the state representation of the structured feature vectors and compressed time series data and calling deep learning, time series Bayesian or reinforcement learning prediction sub-models according to the train operation stage includes: The structured feature vector of the current time window and the corresponding compressed time series data are received, from which the traction power, braking force, speed sequence and acceleration sequence are read, if the traction power is greater than zero, the speed is monotonically increasing and the acceleration is continuously positive, it is determined as an acceleration stage, if the traction power and the braking force do not exceed the traction power threshold and the braking force threshold configured in operation respectively, and the range of the speed sequence does not exceed the speed change threshold configured in operation, it is determined as a uniform speed stage, if the braking force is greater than zero, the speed is monotonically decreasing and the acceleration is continuously negative, it is determined as a braking stage, the deep learning prediction sub-model is called in the acceleration stage, the time series Bayesian prediction sub-model is called in the uniform speed stage, and the reinforcement learning prediction sub-model is called in the braking stage, the compressed time series data is input into the called prediction sub-model, and the prediction values of the speed, acceleration, traction power, braking force and axle box vibration at the next time are output.

[0008] As a preferred embodiment of the present application, the process of performing abnormal physical causal verification by the rail transit physical credible prediction module using track curvature, slope and speed limit includes: The track curvature and track slope at the current position of the train are obtained, the acceleration component of gravity along the running direction is calculated according to the track slope, the line speed limit at the current position of the train and the train speed at the current time are obtained, the train speed is compared with the line speed limit, the acceleration sensor output value at the current time is obtained, the traction force, braking force and total mass of the train at the current time are obtained, the net driving force acceleration component obtained by dividing the difference between the traction force and the braking force by the total mass of the train is calculated, the centripetal acceleration component determined by the train speed and the track curvature is calculated, the gravity acceleration component, the centripetal acceleration component and the net driving force acceleration component are added to obtain the theoretical acceleration, the absolute deviation of the acceleration sensor output value and the theoretical acceleration is calculated as the physical causal consistency residual, the physical causal consistency residual is compared with the physical residual threshold configured in operation, if the physical causal consistency residual is greater than the physical residual threshold, or the train speed is greater than the line speed limit, a physical causal mismatch event is output.

[0009] As a preferred embodiment of the present application, the process of predicting the coupling evolution trend of the abnormality and the track and generating abnormal warning information by the rail transit physical credible prediction module includes: The predicted values and actual observed values of the speed, acceleration, traction power, braking force and axle box vibration at the current time are obtained, the errors of each item are calculated respectively, if the five errors in the past three consecutive times all exceed the error threshold configured in operation respectively, the prediction abnormality is marked, the physical causal consistency residual at the current time is obtained, if the physical causal consistency residual exceeds the physical residual threshold configured in operation, a physical causal mismatch event is marked; Obtain track irregularity amplitude, track alignment deviation and track gauge deviation of the current position of the train, respectively combine the respective corresponding hazard weights to calculate three risk components, and obtain the track risk contribution value after fusion, if the track risk contribution value exceeds the track risk threshold configured in operation, and the predicted abnormal or physical cause mismatch event is established, it is determined that there is train and track coupling evolution difference; According to the overall level of prediction error, the physical cause consistency residual and the track risk contribution value, the comprehensive risk score is calculated by combining the respective corresponding weight coefficients, if the comprehensive risk score exceeds the comprehensive risk threshold configured in operation, an abnormal early warning is output, otherwise a normal operation is output.

[0010] As a preferred embodiment of the application, the rail transit heterogeneous computing power scheduling module schedules data processing tasks based on structured feature vectors and compressed time series data and abnormal response tasks based on abnormal early warning information, and dynamically adjusts the task priority according to the train operation stage and the track topology prior constraint, and the process includes: Obtain the task type, if it is an abnormal response task, obtain the emergency level and the maximum allowed response delay, if it is a data processing task, obtain the data processing complexity and the timeliness requirement, obtain the acceleration weight, the uniform speed weight and the braking weight, obtain the track section number, query the track stage risk value corresponding to the track section number and the three weights, if the risk value corresponding to the track section number and the three weights is greater than the maximum allowed risk threshold configured in operation, discard the current scheduling scheme, and output a prior constraint violation error code; Obtain CPU usage, memory remaining capacity, number of tasks to be processed and whether there is an abnormal event, calculate the fuzzy priority through the fuzzy logic rule base, calculate the reinforcement learning priority through the reinforcement learning model, obtain the original reward value and the consistency constraint coefficient, calculate the modified reward value, update the reinforcement learning model parameters using the modified reward value, fuse the fuzzy priority and the reinforcement learning priority, generate the final task priority, and generate the scheduling instruction based on the final task priority.

[0011] As a preferred embodiment of the application, the process that the rail transit heterogeneous computing power scheduling module maps the task to the adaptive heterogeneous computing power resource in CPU, GPU or FPGA includes: Obtain the task type, obtain the recommended hardware set corresponding to the task type, obtain the load upper limit and the CPU load percentage, if the recommended hardware set contains CPU and the CPU load percentage does not exceed the load upper limit, select CPU as the target hardware, if the target hardware is not selected, obtain the GPU load percentage, if the recommended hardware set contains GPU and the GPU load percentage does not exceed the load upper limit, select GPU as the target hardware; If the target hardware is not selected, the FPGA load percentage is obtained, if the recommended hardware set contains the FPGA and the FPGA load percentage does not exceed the load upper limit, the FPGA is selected as the target hardware, if the recommended hardware set is empty or the load of all hardware exceeds the load upper limit, the CPU is selected as the target hardware, and the task is mapped to the target hardware.

[0012] As a preferred embodiment of the present application, the process of the rail transit heterogeneous computing power scheduling module performing dynamic scheduling and task migration in the case of computing node failure includes: The number of the failed computing node is obtained, the first running task on the failed computing node is obtained, the remaining execution time estimation value of the current task is obtained when there is a current running task, the migration exemption threshold is obtained, if the remaining execution time estimation value is less than or equal to the migration exemption threshold, the current task is marked as allowed to be completed locally, otherwise, other computing nodes are checked in ascending order of node number, the target node is selected, and the checking is ended; If the target node has been selected, the current task is rebuilt on the target node, the current task resources on the failed computing node are released, the current task state is updated to migration success, otherwise, the current task is marked as migration failure and allowed to be completed locally, the next running task on the failed computing node is obtained, and the iteration is ended.

[0013] As a preferred embodiment of the present application, the process of the rail transit computing power thermal control module monitoring the hot spot area of the board card in real time and dynamically controlling the local cooling of the cooling execution mechanism according to the task load on the CPU, GPU and FPGA includes: The CPU task load value and the CPU load threshold are obtained, if the CPU task load value is greater than the CPU load threshold, the cooling execution mechanism covering the CPU physical area is started, otherwise, the cooling execution mechanism covering the CPU physical area is stopped, the GPU task load value and the load threshold are obtained, if the GPU task load value is greater than the GPU load threshold, the cooling execution mechanism covering the GPU physical area is started, otherwise, the cooling execution mechanism covering the GPU physical area is stopped; The FPGA task load value and the FPGA load threshold are obtained, if the FPGA task load value is greater than the FPGA load threshold, the cooling execution mechanism covering the FPGA physical area is started, otherwise, the cooling execution mechanism covering the FPGA physical area is stopped.

[0014] As a preferred embodiment of the present application, the process of the rail transit computing power thermal control module adaptively degrading the computing performance when the temperature exceeds the preset safety threshold includes: Obtaining the CPU measured temperature and the CPU safety threshold, if the CPU measured temperature is greater than the CPU safety threshold and the CPU is not currently limited, then performing CPU performance limitation, if the CPU measured temperature is less than or equal to the CPU safety threshold, the CPU recovery delay timing state is completed and the CPU is currently limited, then performing CPU performance recovery; Obtaining the GPU measured temperature and the GPU safety threshold, if the GPU measured temperature is greater than the GPU safety threshold and the GPU is not currently limited, then performing GPU performance limitation, if the GPU measured temperature is less than or equal to the GPU safety threshold, the GPU recovery delay timing state is completed and the GPU is currently limited, then performing GPU performance recovery; Obtaining the FPGA measured temperature and the FPGA safety threshold, if the FPGA measured temperature is greater than the FPGA safety threshold and the FPGA is not currently limited, then performing FPGA performance limitation, if the FPGA measured temperature is less than or equal to the FPGA safety threshold, the FPGA recovery delay timing state is completed and the FPGA is currently limited, then performing FPGA performance recovery.

[0015] Compared with the prior art, the advantages of the present application are: (1) In the present application, the rail transit physical credible prediction module fuses structured features and compressed time series data, intelligently identifies the train running stage and dynamically calls the adapted deep learning, time series Bayesian or reinforcement learning sub-model, realizes high-precision state prediction, simultaneously combines physical constraints such as track curvature, slope and speed limit, performs causal consistency verification based on the principle of mechanics, effectively identifies physical mismatch anomalies, further fuses prediction errors, physical residuals and track irregularity risks, constructs a weighted comprehensive risk scoring mechanism, accurately determines and warns train-track coupling evolution anomalies, and significantly improves the prediction credibility, anomaly identification accuracy and intelligent level of operation and maintenance decision-making; (2) In the present application, the rail transit heterogeneous computing power scheduling module combines the train running stage and the track topology constraint, fuses deep reinforcement learning and fuzzy logic to dynamically adjust task priority, realizes intelligent sorting of data processing and abnormal response tasks under the premise of safety, through the task characteristics and CPU, GPU, FPGA resource matching mechanism, adaptively allocates computing power according to the load state, and supports multi-condition task migration based on remaining execution time, hardware compatibility, load and safety domain when the node fails, balances real-time performance, safety and resource efficiency, and significantly improves the system scheduling intelligence and operation robustness. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The system block diagram of the embodiment one in the present application is shown in the figure; Figure 2 The system block diagram of the embodiment two in the present application is shown in the figure. DETAILED DESCRIPTION

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] Example 1: As Figure 1 As shown, this invention proposes a rail transit data processing system using high-performance computing node cards, applied to rail transit onboard intelligent terminals, comprising: The rail transit multi-source data preprocessing module is used to collect multi-source heterogeneous sensor parameters from the train operation process, track structure, operating environment and on-board sensing system, including train instantaneous speed, train longitudinal acceleration, train lateral acceleration, brake cylinder pressure, track vertical vibration acceleration, track lateral vibration acceleration, rail strain, ambient temperature, relative humidity, electromagnetic interference field strength and video frame sequence. The module preprocesses the collected multi-source heterogeneous sensor parameters. The preprocessing operations include data cleaning, time synchronization, multi-source fusion, noise suppression, feature extraction and intelligent compression. Feature extraction includes deriving traction power and braking force based on the train instantaneous speed and brake cylinder pressure, combined with the preset train dynamics relationship, and forming speed sequence and acceleration sequence, outputting structured feature vectors and compressed time series data. By integrating heterogeneous data from multiple sources, including trains, tracks, environment, and onboard sensing, through a multi-source data preprocessing module, high-precision time synchronization, intelligent cleaning, and noise suppression are used to ensure data quality. Combined with train dynamics models, higher-order features such as traction power and braking force are derived from raw parameters to enhance information dimensionality and interpretability. Intelligent compression and structured output are employed to reduce redundancy while meeting the needs of both real-time and offline applications. This provides high-quality, low-latency data support for intelligent operation and maintenance tasks such as health monitoring, condition assessment, and risk warning, comprehensively improving the safety and intelligence level of the rail transit system.

[0019] The rail transit physical reliability prediction module is used to call deep learning, temporal Bayes or reinforcement learning prediction sub-models based on structured feature vectors and compressed time series data according to the train operation stage, and use track curvature, gradient and speed limit to perform abnormal physical causal verification, predict the coupling evolution trend of anomalies and track and generate anomaly early warning information. The reliable prediction module for rail transit physics, based on the state representation of structured feature vectors and compressed time-series data, and calling deep learning, temporal Bayesian, or reinforcement learning prediction sub-models according to the train operation stage, includes the following process: The structured feature vector and corresponding compressed time series data of the current time window are received, from which the traction power, braking force, speed sequence and acceleration sequence are read, it is judged whether the speed is monotonically increasing or decreasing, and whether the acceleration sign is continuously positive or negative, if the traction power is greater than zero, the speed is monotonically increasing and the acceleration is continuously positive, it is determined that it is in the acceleration stage, if the traction power and the braking force are respectively not more than the operation configured traction power threshold and braking force threshold, and the speed sequence range is not more than the operation configured speed change threshold, it is determined that it is in the uniform speed stage, if the braking force is greater than zero, the speed is monotonically decreasing and the acceleration is continuously negative, it is determined that it is in the braking stage, in the acceleration stage, a deep learning prediction sub-model is called, in the uniform speed stage, a time series Bayesian prediction sub-model is called, and in the braking stage, a reinforcement learning prediction sub-model is called, the compressed time series data is input into the called prediction sub-model, and the prediction values of the speed, acceleration, traction power, braking force and axle box vibration at the next moment are output; The process of performing abnormal physical causal verification by the rail transit physical credible prediction module using track curvature, slope and speed limit includes: The track curvature and track slope at the current position of the train are obtained, the acceleration component of gravity along the running direction is calculated according to the track slope, the line speed limit at the current position of the train and the train speed at the current time are obtained, the train speed is compared with the line speed limit, the acceleration sensor output value at the current time is obtained, the traction force, braking force and total mass of the train at the current time are obtained, the net driving force acceleration component obtained by dividing the difference between the traction force and the braking force by the total mass of the train is calculated, the centripetal acceleration component determined by the train speed and the track curvature is calculated, the gravity acceleration component, the centripetal acceleration component and the net driving force acceleration component are added to obtain the theoretical acceleration, the absolute deviation between the acceleration sensor output value and the theoretical acceleration is calculated as the physical causal consistency residual, and the physical causal consistency residual is calculated by the following formula: , wherein represents the physical causal consistency residual at time t, represents the measured acceleration output by the acceleration sensor at time t, represents the gravity acceleration, which is a fixed value of 9.81, represents the slope angle of the track at the current position of the train, represents the speed of the train at time t, represents the curvature of the track at the current position of the train, represents the force output by the traction system at time t, represents the force output by the braking system at time t, represents the total mass of the train, the physical causal consistency residual is compared with the operation configured physical residual threshold, if the physical causal consistency residual is greater than the physical residual threshold, or the train speed is greater than the line speed limit, a physical causal mismatch event is output; The process of the rail transit physically credible prediction module predicting the coupling evolution trend of the anomaly and the track and generating anomaly warning information includes: Obtain the predicted value and the actual observed value of the speed, acceleration, traction power, braking force and axle box vibration at the current time, calculate the error of each item respectively, if the five errors in the past three consecutive times all exceed the error threshold configured for operation, mark the prediction anomaly, if any state quantity error at the current time exceeds the mutation threshold configured for operation and there is no driver operation instruction record, also mark the prediction anomaly, obtain the physical and causal consistency residual at the current time, if the physical and causal consistency residual exceeds the physical residual threshold configured for operation, mark the physical and causal mismatch event; Obtain the track high-low irregularity amplitude, track deviation and track gauge deviation at the current position of the train, calculate the three risk components respectively combined with the respective corresponding hazard weight, and obtain the track risk contribution value after fusion, if the track risk contribution value exceeds the track risk threshold configured for operation, and the prediction anomaly or the physical and causal mismatch event is established, it is determined that there is a coupling evolution anomaly of the train and the track; According to the overall level of prediction error, the physical and causal consistency residual and the track risk contribution value, calculate the comprehensive risk score combined with the respective corresponding weight coefficient, if the comprehensive risk score exceeds the comprehensive risk critical value configured for operation, output the anomaly warning, otherwise output the normal operation, determine the main cause of the anomaly warning: if the physical and causal mismatch event is established, the main cause is marked as physical and causal mismatch, otherwise, if the track risk contribution value exceeds the track risk threshold configured for operation and the prediction anomaly or the physical and causal mismatch event is established, the main cause is marked as track coupling risk, otherwise, if the prediction anomaly is established, the main cause is marked as prediction model deviation; Through the rail transit physically credible prediction module, the structured feature vector and the compressed time series data are comprehensively fused, the train running stage (acceleration, uniform speed, braking) is intelligently identified, and the deep learning, time series Bayesian or reinforcement learning prediction sub-model is called accordingly to accurately predict the key running parameters at the next time. This module not only realizes high-precision representation and prediction of the train state, but also performs abnormal physical and causal checking by introducing track curvature, slope and speed limit and other actual running environment parameters, effectively ensuring the physical consistency and reliability of the prediction results. In addition, by using the comprehensive risk score mechanism combined with the prediction error, the physical and causal consistency residual and the track risk contribution value, the potential train and track coupling evolution anomaly can be accurately evaluated and warned, greatly improving the safety and operation efficiency of the system. The setting of the physical and causal mismatch event and the track risk threshold makes the anomaly detection more scientific and reasonable, and can accurately mark the main cause of the anomaly, providing clear guidance for subsequent maintenance, and finally realizing more intelligent and efficient monitoring and management of the rail transit system. This multi-level and multi-dimensional data processing and analysis method greatly enhances the safety, stability and intelligence level of rail transit operation.

[0020] The rail transit heterogeneous computing power scheduling module is used for scheduling data processing tasks based on structured feature vectors and compressed time sequence data and abnormal response tasks based on abnormal early warning information, dynamically adjusting task priorities according to train running stages and rail topology prior constraints by using a deep reinforcement learning and fuzzy logic fusion method, and mapping the tasks to adaptive heterogeneous computing power resources in CPUs, GPUs or FPGAs to perform dynamic scheduling and task migration in the event of computing power node failure. The process in which the rail transit heterogeneous computing power scheduling module schedules data processing tasks based on structured feature vectors and compressed time sequence data and abnormal response tasks based on abnormal early warning information and dynamically adjusts task priorities according to train running stages and rail topology prior constraints by using a deep reinforcement learning and fuzzy logic fusion method includes: The type of the task is obtained, if it is an abnormal response task, the emergency level and the maximum allowed response delay are obtained, if it is a data processing task, the data processing complexity and the timeliness requirement are obtained, the acceleration weight, the uniform speed weight and the braking weight are obtained, the rail section number is obtained, the rail stage risk value corresponding to the rail section number and the three weights is queried, if the risk value corresponding to the rail section number and the three weights is greater than the maximum allowed risk threshold configured in operation, the current scheduling scheme is discarded, and an error code for violating the prior constraint is output; The CPU usage, the memory remaining capacity, the number of tasks to be processed and whether there is an abnormal event are obtained, the fuzzy priority is calculated through a fuzzy logic rule base, the reinforcement learning priority is calculated through a reinforcement learning model, the original reward value and the consistency constraint coefficient are obtained, the modified reward value is calculated, and the calculation formula is: wherein represents the current system state, represents the action of assigning the priority, is an original reward integer returned by a task execution monitor, is a priority integer returned by a reinforcement learning subroutine, is a priority integer returned by a fuzzy logic subroutine, is a consistency constraint coefficient configuration item, is a modified reward used for model updating, the reinforcement learning model parameters are updated using the modified reward value, the fuzzy priority and the reinforcement learning priority are fused, the final task priority is generated, and the scheduling instruction is generated based on the final task priority; The process in which the rail transit heterogeneous computing power scheduling module maps the tasks to adaptive heterogeneous computing power resources in CPUs, GPUs or FPGAs includes: Get the task type, get the recommended hardware set corresponding to the task type, get the load limit and CPU load percentage. If the recommended hardware set includes CPU and the CPU load percentage does not exceed the load limit, then select CPU as the target hardware. If no target hardware is selected, get the GPU load percentage. If the recommended hardware set includes GPU and the GPU load percentage does not exceed the load limit, then select GPU as the target hardware. If no target hardware is selected, the FPGA load percentage is obtained. If the recommended hardware set includes an FPGA and the FPGA load percentage does not exceed the load limit, the FPGA is selected as the target hardware. If the recommended hardware set is empty or all hardware in it exceeds the load limit, the CPU is selected as the target hardware, and the task is mapped to the target hardware. The process of dynamic scheduling and task migration in case of computing node failure in the heterogeneous computing power scheduling module of rail transit includes: Obtain the faulty compute node number, obtain the first running task on the faulty compute node, if there is a currently running task, obtain the estimated remaining execution time of the current task, obtain the migration exemption threshold, if the estimated remaining execution time is less than or equal to the migration exemption threshold, mark the current task as allowed to complete locally, otherwise, check other compute nodes in ascending order of node number, for each checked node, obtain its status, if the status is not equal to ready, skip, obtain its hardware type, if the hardware type required by the current task does not match the hardware type of the checked node, skip, obtain its current load rate, if the current load rate is greater than the load limit threshold, skip, obtain its security domain identifier with the current task, if the security domain identifier is inconsistent, skip, select the checked node as the target node, and end the check; If a target node has been selected, rebuild the current task on the target node, release the current task resources on the failed computing node, update the current task status to migration successful; otherwise, mark the current task as migration failed, allow local completion, obtain the next running task on the failed computing node, and end the traversal. By integrating deep reinforcement learning and fuzzy logic into the heterogeneous computing power scheduling module for rail transit, and combining train operation phases and prior constraints of track topology, task priorities are dynamically generated that balance real-time performance, safety, and resource efficiency. Through a hardware-aware task mapping mechanism, CPU, GPU, or FPGA resources are intelligently allocated to optimize computing power utilization while ensuring task timeliness. It also has fault adaptive capabilities, which can efficiently complete the migration of critical tasks or security degradation based on multi-dimensional conditions such as remaining execution time, hardware compatibility, load status, and security domain consistency when computing power nodes fail, ensuring high system availability and business continuity, and significantly improving the robustness, response speed, and intelligent resource scheduling level of the rail transit edge intelligent system.

[0021] Example 2: The technical solution of this embodiment of the invention differs from that of Example 1 in that: like Figure 2 As shown, the rail transit computing power thermal control module is used to monitor the hot spots of the board in real time according to the task load on the CPU, GPU and FPGA, dynamically control the heat dissipation actuator to perform local cooling, and adaptively degrade computing performance when the temperature exceeds the preset safety threshold. The rail transit computing power thermal control module monitors hot spots on the CPU, GPU, and FPGA in real time based on the task load, and dynamically controls the heat dissipation actuators to perform localized cooling. The process includes: Get the CPU task load value and CPU load threshold. If the CPU task load value is greater than the CPU load threshold, start the heat dissipation execution mechanism covering the CPU physical area; otherwise, stop the heat dissipation execution mechanism covering the CPU physical area. Get the GPU task load value and load threshold. If the GPU task load value is greater than the GPU load threshold, start the heat dissipation execution mechanism covering the GPU physical area; otherwise, stop the heat dissipation execution mechanism covering the GPU physical area. Obtain the FPGA task load value and FPGA load threshold. If the FPGA task load value is greater than the FPGA load threshold, start the heat dissipation actuator covering the FPGA physical area; otherwise, stop the heat dissipation actuator covering the FPGA physical area. The process by which the thermal control module for rail transit computing power adaptively degrades computing performance when the temperature exceeds a preset safety threshold includes: Get the actual CPU temperature and CPU safety threshold. If the actual CPU temperature is greater than the CPU safety threshold and the CPU is not currently restricted, then execute CPU performance limiting. Get the actual CPU temperature, CPU safety threshold and CPU recovery delay timer status. If the actual CPU temperature is less than or equal to the CPU safety threshold, the CPU recovery delay timer status is complete and the CPU is currently restricted, then execute CPU performance recovery. Get the actual GPU temperature and GPU safety threshold. If the actual GPU temperature is greater than the GPU safety threshold and the GPU is not currently restricted, then perform GPU performance limiting. Get the actual GPU temperature, GPU safety threshold and GPU recovery delay timer status. If the actual GPU temperature is less than or equal to the GPU safety threshold, the GPU recovery delay timer status is complete and the GPU is currently restricted, then perform GPU performance recovery. Get the measured FPGA temperature and FPGA safety threshold. If the measured FPGA temperature is greater than the FPGA safety threshold and the FPGA is not currently restricted, then perform FPGA performance limiting. Get the measured FPGA temperature, FPGA safety threshold, and FPGA recovery delay timer status. If the measured FPGA temperature is less than or equal to the FPGA safety threshold, the FPGA recovery delay timer status is complete, and the FPGA is currently restricted, then perform FPGA performance recovery. The rail intersection computing power thermal control module can monitor the hot spot area of the board card in real time according to the task load of CPU, GPU and FPGA, start and stop the local heat dissipation execution mechanism covering the physical area of each chip as needed, and realize precise and efficient on-demand cooling; at the same time, when the temperature of any chip exceeds the preset safety threshold, the targeted computing performance degradation is automatically triggered, and after the temperature falls back to the safety range and meets the recovery delay condition, the computing power is recovered in order, thereby maximizing the system computing capacity under the premise of guaranteeing thermal safety, effectively improving the energy efficiency, stability and long-term operation reliability of the rail transit edge computing platform under high load working conditions.

[0022] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change within the technical range disclosed by the present application according to the technical scheme and improvement concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A rail transit data processing system using a high-performance computing node board card, applied to a rail transit vehicle-mounted intelligent terminal, characterized in that, The rail transit system comprises: a rail transit multi-source data preprocessing module, configured to collect multi-source heterogeneous sensing parameters from train operation processes, track structures, operating environments, and vehicle-mounted sensing systems, preprocess the collected multi-source heterogeneous sensing parameters, and output structured feature vectors and compressed time series data; a rail transit physically credible prediction module, configured to call deep learning, time series Bayesian, or reinforcement learning prediction sub-models according to train operation stages based on the structured feature vectors and the compressed time series data, perform abnormal physical causal verification using track curvature, slope, and speed limit, predict the coupling evolution trend of abnormalities and tracks, and generate abnormal warning information; a rail transit heterogeneous computing power scheduling module, configured to schedule data processing tasks based on the structured feature vectors and the compressed time series data, and abnormal response tasks based on the abnormal warning information, dynamically adjust task priorities according to train operation stages and track topology prior constraints using a deep reinforcement learning and fuzzy logic fusion method, and map the tasks to adaptive heterogeneous computing power resources in CPUs, GPUs, or FPGAs to perform dynamic scheduling and task migration in the event of computing power node failure; a rail transit computing power thermal control module, configured to monitor hot spots in the board card in real time according to the task load on the CPUs, GPUs, and FPGAs, dynamically control the local cooling of the heat dissipation execution mechanism, and adaptively degrade the computing performance when the temperature exceeds the preset safety threshold.

2. The rail transit data processing system employing high-performance computing node board cards according to claim 1, characterized in that, The process of calling deep learning, time series Bayesian, or reinforcement learning prediction sub-models according to train operation stages based on the structured feature vectors and the compressed time series data by the rail transit physically credible prediction module comprises: receiving the structured feature vectors and the corresponding compressed time series data of the current time window, reading the traction power, braking force, speed sequence, and acceleration sequence therefrom, determining that it is an acceleration stage if the traction power is greater than zero, the speed is monotonically increasing, and the acceleration is continuously positive, determining that it is a uniform speed stage if the traction power and the braking force are respectively less than the traction power threshold and the braking force threshold configured in the operation, and the range of the speed sequence is less than the speed change threshold configured in the operation, determining that it is a braking stage if the braking force is greater than zero, the speed is monotonically decreasing, and the acceleration is continuously negative, calling the deep learning prediction sub-model in the acceleration stage, calling the time series Bayesian prediction sub-model in the uniform speed stage, calling the reinforcement learning prediction sub-model in the braking stage, inputting the compressed time series data into the called prediction sub-model, and outputting the predicted values of the speed, acceleration, traction power, braking force, and axle box vibration at the next time.

3. The rail transit data processing system with high-performance computing node board cards according to claim 2, characterized in that, The process of performing abnormal physical causal verification using track curvature, slope, and speed limit by the rail transit physically credible prediction module comprises: Obtain the track curvature and track slope of the current position of the train, calculate the acceleration component of gravity along the running direction according to the track slope, obtain the line speed limit of the current position of the train and the current time of the train speed, compare the train speed with the line speed limit, obtain the acceleration sensor output value at the current time, obtain the current time of the traction force, braking force and total mass of the train, calculate the net driving force acceleration component obtained by the difference between the traction force and the braking force divided by the total mass of the train, calculate the centripetal acceleration component determined by the train speed and the track curvature, add the gravity acceleration component, the centripetal acceleration component and the net driving force acceleration component to obtain the theoretical acceleration, calculate the absolute deviation of the acceleration sensor output value and the theoretical acceleration as the physical and causal consistency residual, compare the physical and causal consistency residual with the physical residual threshold configured in operation, if the physical and causal consistency residual is greater than the physical residual threshold, or the train speed is greater than the line speed limit, then output the physical and causal mismatch event.

4. The rail transit data processing system employing high-performance computing node board cards according to claim 3, characterized in that, The process of the rail transit physical credible prediction module predicting the coupling evolution trend of the abnormality and the track and generating abnormal early warning information includes: Obtain the predicted value and actual observed value of the speed, acceleration, traction power, braking force and axle box vibration at the current time, calculate the error of each item respectively, if the five errors in the past three consecutive times exceed the error threshold configured in operation respectively, mark the prediction abnormality, obtain the physical and causal consistency residual at the current time, if the physical and causal consistency residual exceeds the physical residual threshold configured in operation, mark the physical and causal mismatch event; Obtain the track high-low irregularity amplitude, track deviation and track gauge deviation of the current position of the train, calculate the three risk components respectively combined with the respective corresponding hazard weights, and obtain the track risk contribution value after fusion, if the track risk contribution value exceeds the track risk threshold configured in operation, and the prediction abnormality or the physical and causal mismatch event is established, it is determined that there is a coupling evolution of the train and the track; According to the overall level of prediction error, the physical and causal consistency residual and the track risk contribution value, calculate the comprehensive risk score combined with the respective corresponding weight coefficients, if the comprehensive risk score exceeds the comprehensive risk critical value configured in operation, output the abnormal early warning, otherwise output the normal operation.

5. The rail transit data processing system employing high-performance computing node board cards according to claim 1, wherein, The process of the rail transit abnormality structure computing power scheduling module scheduling data processing tasks based on structured feature vectors and compressed time series data and abnormal response tasks based on abnormal early warning information, dynamically adjusting the task priority by using a deep reinforcement learning and fuzzy logic fusion method according to the train running stage and the track topology prior constraint includes: Obtain the task type, if it is an abnormal response task, obtain the emergency level and the maximum allowed response delay, if it is a data processing task, obtain the data processing complexity and the timeliness requirement, obtain the acceleration weight, the uniform speed weight and the braking weight, obtain the track section number, query the track stage risk value corresponding to the track section number and the three weights, if the risk value corresponding to the track section number and the three weights is greater than the maximum allowed risk threshold configured in operation, discard the current scheduling scheme and output the prior constraint violation error code. The CPU usage, the memory remaining capacity, the number of tasks to be processed and whether there is an abnormal event are acquired, the fuzzy priority is calculated through a fuzzy logic rule base, the reinforcement learning priority is calculated through a reinforcement learning model, the original reward value and the consistency constraint coefficient are acquired, the modified reward value is calculated, the modified reward value is used to update the reinforcement learning model parameters, the fuzzy priority and the reinforcement learning priority are fused, and the final task priority is generated, and the scheduling instruction is generated based on the final task priority.

6. The rail transit data processing system employing high-performance computing node board cards according to claim 5, wherein, The process that the rail transit heterogeneous computing scheduling module maps the task to the adaptive heterogeneous computing resource in the CPU, the GPU or the FPGA includes: The task type is acquired, the recommended hardware set corresponding to the task type is acquired, the load upper limit and the CPU load percentage are acquired, if the recommended hardware set contains the CPU and the CPU load percentage does not exceed the load upper limit, the CPU is selected as the target hardware, if the target hardware is not selected, the GPU load percentage is acquired, if the recommended hardware set contains the GPU and the GPU load percentage does not exceed the load upper limit, the GPU is selected as the target hardware; If the target hardware is not selected, the FPGA load percentage is acquired, if the recommended hardware set contains the FPGA and the FPGA load percentage does not exceed the load upper limit, the FPGA is selected as the target hardware, if the recommended hardware set is empty or all hardware loads exceed the load upper limit, the CPU is selected as the target hardware, and the task is mapped to the target hardware.

7. The rail transit data processing system employing high-performance computing node board cards according to claim 6, characterized in that, The process that the rail transit heterogeneous computing scheduling module performs dynamic scheduling and task migration when the computing node fails includes: The fault computing node number is acquired, the first running task on the fault computing node is acquired, when there is a current running task, the remaining execution time estimation value of the current task is acquired, the migration exemption threshold is acquired, if the remaining execution time estimation value is less than or equal to the migration exemption threshold, the current task is marked as allowed to be completed locally, otherwise, other computing nodes are checked in ascending order of node number, the checked node is selected as the target node, and the checking is ended; If the target node has been selected, the current task is rebuilt on the target node, the current task resource on the fault computing node is released, the current task state is updated to migration success, otherwise, the current task is marked as migration failure and allowed to be completed locally, the next running task on the fault computing node is acquired, and the iteration is ended.

8. The rail transit data processing system employing high-performance computing node board cards according to claim 1, wherein, The process that the rail transit computing thermal control module monitors the hot spot area of the board card in real time according to the task load on the CPU, the GPU and the FPGA, and dynamically controls the local cooling of the heat dissipation execution mechanism includes: The CPU task load value and the CPU load threshold are acquired, if the CPU task load value is greater than the CPU load threshold, the heat dissipation execution mechanism covering the CPU physical area is started, otherwise, the heat dissipation execution mechanism covering the CPU physical area is stopped, the GPU task load value and the load threshold are acquired, if the GPU task load value is greater than the GPU load threshold, the heat dissipation execution mechanism covering the GPU physical area is started, otherwise, the heat dissipation execution mechanism covering the GPU physical area is stopped; The FPGA task load value and the FPGA load threshold value are acquired, and if the FPGA task load value is greater than the FPGA load threshold value, a heat dissipation execution mechanism covering the FPGA physical area is started, otherwise the heat dissipation execution mechanism covering the FPGA physical area is stopped.

9. The rail transit data processing system employing high-performance computing node board cards according to claim 8, characterized in that, The process of the rail transit algorithmic thermal control module adaptively degrading the computing performance when the temperature exceeds the preset safety threshold value includes: The CPU measured temperature and the CPU safety threshold value are acquired, and if the CPU measured temperature is greater than the CPU safety threshold value and the CPU is not currently limited, CPU performance limitation is performed, and if the CPU measured temperature is less than or equal to the CPU safety threshold value, the CPU recovery delay timing state is completed, and the CPU is currently limited, CPU performance recovery is performed; The GPU measured temperature and the GPU safety threshold value are acquired, and if the GPU measured temperature is greater than the GPU safety threshold value and the GPU is not currently limited, GPU performance limitation is performed, and if the GPU measured temperature is less than or equal to the GPU safety threshold value, the GPU recovery delay timing state is completed, and the GPU is currently limited, GPU performance recovery is performed; The FPGA measured temperature and the FPGA safety threshold value are acquired, and if the FPGA measured temperature is greater than the FPGA safety threshold value and the FPGA is not currently limited, FPGA performance limitation is performed, and if the FPGA measured temperature is less than or equal to the FPGA safety threshold value, the FPGA recovery delay timing state is completed, and the FPGA is currently limited, FPGA performance recovery is performed.

Images