ISSN 1673-0291 CN 11-5258/U
To meet the stringent requirements for low latency, high reliability, and intelligence posed by emerging applications such as autonomous driving and the Industrial Internet, this study examines the limitations of traditional network architectures in heterogeneous resource integration, service adaptation, and intelligent decision-making. A smart computing integration network architecture based on a “three-layer, three-domain” framework is proposed, and its development prospects are discussed in relation to the deep integration of heterogeneous networks, computing resource scheduling, and network-native intelligence. The research results demonstrate that the proposed smart computing integration network architecture enables a paradigm shift from “passive connection” to “active service” through key technologies such as unified resource representation, computing-network demand analysis, and agile resource scheduling, thereby achieving comprehensive coordination between computing and networking. The research findings provide theoretical references and technical pathways for the construction of intelligent, efficient, and reliable emerging network infrastructures.
China has built and operated the world’s largest high-speed railway and urban rail networks. With the continuous expansion of network scale and the growing complexity of operating environments, existing automated train operation control systems face diverse challenges in improving efficiency and adaptability, and are increasingly unable to meet the requirements of safe and efficient operations under high-density traffic and dynamic conditions. Autonomous operation control for railway transportation, which integrates perception, low-latency and high-reliability communication, autonomous train control, and intelligent scheduling, is expected to enable safe and efficient train operations in complex environments, thereby facilitating the intelligent development of rail transit. This paper provides a comprehensive review of autonomous operation control technologies for railway transportation. It first summarizes domestic and international research progress and practical applications in this field, clarifies the connotation of autonomy, compares the concepts of autonomy in maritime, road, and railway transportation, and systematically outlines the key technology framework composed of integrated perception, low-latency and high-reliability communication, autonomous train control in complex environments, and intelligent train scheduling. The paper then analyzes the application prospects of these key technologies in the rail transit domain and discusses practical implementations through representative case studies. Finally, it identifies the bottlenecks and challenges in real-world deployment and explores future development directions. This paper aims to provide systematic reference and support for both theoretical research and engineering practice in autonomous operation control of railway transportation.
With the continuous evolution of intelligent railway systems, novel railway applications based on spatiotemporal information from the BeiDou Navigation Satellite System (BDS) have garnered significant attention. However, the lack of dedicated testing systems tailored to specific railway applications makes it difficult to meet the full lifecycle requirements of BDS-based railway systems. To address this problem, this paper focuses on exploring the requirements and development potential for specialized testing and evaluation of BDS-based railway applications. It investigates how to leverage effective mechanisms to build dedicated testing instruments and system environments that address the unique characteristics of the railway industry, thereby meeting the full lifecycle needs of BDS-based railway systems. First, existing relevant work is surveyed and summarized, outlining the development needs and vision for dedicated testing of BDS-based positioning in railway. Second, a comprehensive overall architecture for a dedicated railway BDS positioning testing system is established, centered on the concept of "zero-on-site testing". This architecture aims to adopt a modular approach and inter-connected collaborative mechanisms to forge a new pathway that balances conventional satellite positioning testing paradigms with the application characteristics of the railway industry. Third, this paper uses the train positioning unit of a specific type of train control system as a test subject example, detailing the setup of the dedicated test system and the implementation of testing. Finally, this paper analyzes and summarizes the research challenges that may arise in providing comprehensive and reliable testing and evaluation for the progressively developing applications of BDS in railway, offering corresponding approaches. The research results indicate that constructing a dedicated railway BDS positioning testing system can effectively enhance testing coverage, flexibility, adaptability, and compatibility, thereby providing strong support for the deep application of BDS in the railway industry.
With the continuous expansion and technological upgrading of China’s high-speed railway systems, the electromagnetic complexity of the operating environment has significantly increased. In particular, the frequent appearance of strong external electromagnetic interference sources has posed unprecedented challenges to the stable operation of the system. The traditional electromagnetic compatibility (EMC) framework, which focuses primarily on improving device immunity, shows evident limitations when dealing with random and strong external electromagnetic interference, such as delayed response and poor adaptability. These shortcomings make it difficult to meet the current high demands for system functional stability, mission continuity, and rapid recovery. To address these challenges, this paper systematically reviews the current research progress in the field of EMC for high-speed railways in China, with a particular focus on key technical aspects such as interference source modeling, coupling path identification, disturbance immunity evaluation of critical equipment, and system-level protection strategies. On this basis, a new research framework for electromagnetic safety is proposed, with train control capability retention as the core objective. This framework aims to shift from "passive immunity" to "intelligent defense and autonomous recovery," and it includes essential components such as intelligent interference identification and perception, multi-physical-field coupled network modeling, electromagnetic interference risk propagation mechanisms, active protection technologies, and experimental validation methods. Research findings indicate that electromagnetic safety, as an extension of the traditional EMC system, represents a key pathway to ensuring the high safety and high reliability of high-speed railway systems under complex electromagnetic environments. The outcomes of this study can provide theoretical foundations and technical references for the future design and engineering implementation of electromagnetic protection in high-speed railway systems.
Due to constantly interacting targets, rapidly shifting environments, and the inherent heterogeneity of multi-sensor data, dynamic traffic scenarios impose stringent demands on the perceptual robustness and decision reliability of intelligent systems. Multi-modal learning emerges as a critical solution to overcome bottlenecks in dynamic scenario understanding by fusing heterogeneous modalities. This paper offers a systematic review on multimodal robust learning for dynamic traffic scenarios. First, we clarify the definition of multimodal dynamic traffic scenarios, and analyzes the types and dynamic characteristics of multi-source modalities (e.g., optical, radio frequency, and acoustic). Next, we lay out the fundamental principles of multi-modal learning, paying particular attention to key techniques that enhance robustness across data-level processing, model architectures, and training strategies. Furthermore, we delve into core challenges currently confronting the field and future research directions. The review highlights current challenging of data imperfections, model limitations, and the absence of evaluation benchmarks, and chart future directions toward three aspects: technological innovation, technology integration, and collaborative industry efforts. Our aim is to provide a systematic reference for both theoretical research and practical deployment of multimodal robust learning in dynamic traffic scenarios, facilitate the evolution of intelligent transportation systems from being “available in limited scenarios” to achieving “reliability across all conditions,” and provide critical technical support for the widespread adoption of intelligent transportation solutions.
Visual intelligence, as a core branch of AI, seeks to endow machines with human-like capabilities for visual understanding and interaction. Since the breakthrough of deep learning in computer vision in 2012, the field has undergone four progressive stages of evolution. The first stage, represented by AlexNet, VGGNet and ResNet, leveraged large annotated datasets such as ImageNet to achieve remarkable success in closed-domain tasks (e.g., image classification and object detection), but its dependence on labeled data highlighted inherent limitations. The second stage witnessed the rise of self-supervised learning, with models such as MoCo, DION and MAE learning powerful visual representations from massive unlabeled data through contrastive, distillation, and masked reconstruction methods. The third stage marked a shift toward multimodal intelligence, where models like CLIP and GPT-4V integrated vision and language, enabling open-vocabulary understanding and advancing toward fine-grained, intent-driven reasoning. The current frontier is world models, exemplified by Sora, which aim not only to perceive and describe but also to simulate and predict the physical world, paving the way for embodied intelligence capable of interacting with reality. The current frontier is world models, exemplified by Sora, which aim not only to perceive and describe but also to simulate and predict the physical world, paving the way for embodied intelligence capable of interacting with reality. This fundamental transformation from discriminative understanding to generative simulation of the world marks a new consensus: generative modeling is the new deep learning. This survey follows this developmental trajectory, analyzing the core ideas, representative models, and methodological paradigms at each stage, while highlighting ongoing challenges in robustness, reasoning, and generalization. This survey follows this developmental trajectory, analyzing the core ideas, representative models, and methodological paradigms at each stage, while highlighting ongoing challenges in robustness, reasoning, and generalization.
This study addresses lane-change obstacle avoidance for autonomous vehicles under sudden road hazards and proposes SafeLC-DelayDDPG, a vehicle control algorithm based on Deep Reinforcement Learning (DRL). The task is formulated as a Markov Decision Process (MDP), and a structured hybrid state space is constructed by integrating local observations, lane-level semantic information, and the ego vehicle’s global states to enhance environmental perception and risk sensitivity. The action space consists of continuous front-wheel steering angle and longitudinal acceleration. The reward function is centered on a two-dimensional time-to-collision (2D-TTC) metric, balancing safety, efficiency, comfort, and traffic-rule compliance, and employs a TTC-conditioned dynamic weighting mechanism that prioritizes safety under high risk and efficiency under low risk. Furthermore, delayed policy updates and target policy smoothing are introduced, and the Critic network loss is refined to mitigate the training instability and Q-value overestimation issues inherent in Deep Deterministic Policy Gradient (DDPG). The proposed method is validated through traffic simulations across diverse scenarios. Experimental results show that, compared with multiple baseline algorithms, SafeLC-DelayDDPG achieves superior safety and efficiency: during training, the first-attempt and consecutive obstacle-avoidance success rates improve by up to 17.9% and 60.5%, respectively; the safety metric by up to 7.6%; and the average speed by up to 2.1%. In cross-scenario tests, the first-attempt and consecutive success rates improve by up to 13.3% and 44.1%, the safety metric by up to 9.8%, and the average speed by up to 0.6%.
To investigate the decision-making mechanisms of stranded passengers’ travel mode choice during unexpected metro service disruptions and to clarify the influence of key situational factors, this study constructs a baseline Multinomial Logit model incorporating mode-specific attributes, based on scenario simulations and Stated Preference (SP) surveys. A stepwise utility-correction approach is applied to quantify both the impact intensity and patterns of influence exerted by factors such as expected recovery time, weather conditions, travel time and purpose, remaining trip distance, crowd behavior, and the availability of alternative routes. Model improvements are evaluated using the Likelihood Ratio Test (LRT), Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and prediction error rates. The research results indicate that passenger decision-making demonstrates strong scenario dependence and a pronounced risk-averse tendency. The effects of situational factors vary considerably, with expected recovery time having the greatest impact (LR=183.98, prediction error reduced by 1.94%), followed by weather conditions (LR=102.63, error reduced by 1.16%). Trip purpose and remaining journey distance also show notable influence, while crowd behavior and alternative-route availability do not exhibit significant effects. Preferences shifts across public transport modes display similar patterns of change. however, waiting on-site, though frequently chosen as a passive option, maintains a considerable selection rate but declines significantly when systemic and situational risks overlap. Sensitivity to scenario factors differs markedly across modes, with those offering higher certainty and lower risk proving more attractive in disruption scenarios.
The transformation and upgrading of the railway transportation production system represent a core issue and critical challenge currently faced by railway transportation enterprises in their transition toward modern logistics enterprises. To support the construction of a modern logistics-oriented railway transportation production system, the study first analyzes key aspects including top-level design, theoretical architecture, critical technologies, and implementation evaluation indicators. Subsequently, the key scientific and technological bottlenecks that need to be overcome in building the new system are discussed, proposing a key technological architecture oriented to modern logistics. Finally, based on this architecture, a simulation prototype system for railway production organization capable of optimization, evaluation, and simulation is designed to validate the feasibility and effectiveness of the proposed theoretical framework and technical pathways. The research demonstrates that this study offers certain theoretical reference value and practical significance for enhancing the operational efficiency and service quality of railway logistics, as well as promoting the transformation and development of transportation production. Additionally, the developed prototype system can provide a reusable development paradigm and decision-support tool for subsequent system optimization and engineering applications.
For the problem of energy-efficient operation in urban rail transit, existing studies can be grouped into two main strands: Classical methods and technology-driven approaches. The classical strand reviews optimization of train speed profiles, train timetables, and their joint optimization, showing that these methods have developed into mature modeling and solution frameworks capable of reducing traction energy consumption and peak power while enhancing the utilization of regenerative braking energy under safety and service constraints. The second strand summarizes technology-driven progress and discusses the enabling effects and challenges of emerging technologies across three domains: Power supply, control, and operations. On the power side, efforts focus on advancing “source-grid-storage-train” integration, coordinating reversible converters, energy storage, and renewable sources to achieve load shifting and peak suppression. On the control side, autonomous operation and virtual coupling shift the problem toward a dynamic perspective under multi-train coordination. On the operations side, cross-line through services, multi-route patterns, and flexible train formations upgrade capacity supply and reduce inefficient traction. The findings suggest that classical methods already provide well-established frameworks for safe and effective energy saving, while new technology-driven approaches move beyond train-level offline optimization and progressively drive research toward system-level coordination that spans power supply, control, and operations. Future research should place greater emphasis on developing unified benchmark and evaluation systems, while also advancing multi-level coordination and integrated optimization across diverse domains.
Current research on optimizing and controlling traction energy consumption in urban rail transit has largely focused on localized improvements within single systems, making it difficult to coordinate inter-system coupling and achieve global energy efficiency. To address this problem, this study proposes a multi-system collaborative optimization method that integrates power supply, vehicles, and operations. By analyzing the circulation paths of traction energy consumption and the inter-dependencies among decision variables, a four-layer global optimization framework is developed. The first layer optimizes single-train driving curves according to specified interval running times, identifying energy-minimal trajectories under constraints such as punctuality, comfort, and speed limits. The second layer optimizes inter-station running times by adjusting train travel durations between adjacent stations, thereby minimizing overall line energy consumption while ensuring that the total turnaround time equals the given value. The third layer focuses on optimizing multi-train energy-efficient timetables. By adjusting departure intervals and dwell times, this layer reduces total traction energy consumption while balancing passengers’ average waiting times through multi-train coordination. The fourth layer addresses threshold optimization of energy storage and inverter devices. The results of the first three layers of the train scheduling and control optimization process are used to inform the adoption of a grid-voltage-based control strategy. Simulation studies are undertaken in actual line operations. Simulation results demonstrate that the proposed global optimization reduces system traction energy consumption by 13.24%, while increasing passengers’ average waiting time by only 12 seconds. These results verify the effectiveness of the proposed coupled energy-saving optimization method and provide a theoretical foundation for the comprehensive optimization and control of traction energy consumption in urban rail transit.
To address the challenges of precise governance arising from the spatiotemporal heterogeneity and diverse patterns of pollutant and carbon emissions from road mobile sources in China, this study develops a machine learning-based analytical framework for emission characteristics, driving factors, and mitigation pathways. First, a transport-carbon-environment dataset is constructed by integrating multi-source data to analyze the spatiotemporal distribution characteristics of emissions. Second, a knowledge-constrained non-negative matrix factorization model is developed to identify distinct emission patterns and their underlying drivers. Finally, a multi-pattern integrated multiple linear regression model is applied to evaluate effective pathways for pollution and carbon reduction. The results indicate that in recent years, particulate matter emissions from road mobile sources have shown a decreasing trend, while carbon dioxide emissions continue to increase, with emissions of NO x, CO, and VOCs remaining substantial. Three primary emission patterns are identified across China: a “CO2-PM Reduction Pattern,” distributed across border and coastal regions and driven mainly by road passenger and freight transport activities and vehicle mileage (37.2%); a “Pollutant Reduction Pattern,” found in municipalities like Beijing and Tianjin and provinces such as Guangdong, mainly influenced by the Gross Transport Product (30.1%); and a “CO2-NO x Pollution Pattern,” located in the central and western regions, affected by multiple factors including railway development and its electrification, as well as the use of light-duty transport vehicles (23.6%). Pathway analysis indicates that accelerating economic development in the transport sector has the most significant impact on reduction, potentially reducing CO2 and NO x by up to 6.7% and 4.5%, respectively. Other significant measures include restructuring the transport energy system, reducing road freight and vehicle mileage, and promoting the shift from road to rail alongside railway electrification. The proposed framework provides a quantitative basis and strategic guidance for coordinated pollution and carbon reduction from road mobile sources in the context of China’s dual-carbon goals of carbon peaking and carbon neutrality.
To address the limitations of traditional manual inspection of rail transit infrastructure, such as low efficiency, safety risks, and the dependence of existing rail-mounted detection equipment on maintenance time gaps, which leads to blind spots and limited coverage, this study develops an integrated “End-Edge-Cloud-Surveillance” framework for autonomous unmanned aerial vehicle (UAV)-based intelligent inspection in rail transit. At the “End” layer, multi-source perception combining visible light, infrared, and LiDAR, together with visual-inertial state estimation, enables autonomous perception and task-level navigation. At the “Edge” layer, beyond-visual-line-of-sight (BVLOS) communication and secure, efficient data transmission mechanisms are established, alongside lightweight onboard inference for real-time defect and risk detection. At the “Cloud” and “Surveillance” layers, cross-scenario and multi-target inspection applications are conducted with global data analytics, while a low-altitude surveillance system integrating cooperative and non-cooperative surveillance is established to ensure regulatory compliance and operational safety throughout the entire process. The results demonstrate that this work systematically identifies the unique challenges and characteristics of the rail transit domain and, for the first time, unifies UAV-based rail transit inspection within a full-chain “End-Edge-Cloud-Surveillance” framework. This provides a generalizable reference framework for the future deployment of autonomous UAVs in rail infrastructure inspection.
Most existing quality evaluation methods for high-speed railway ballastless track construction are established based on traditional construction processes. As ballastless track construction progressively shifts towards intelligent construction, the existing evaluation methods can no longer adequately meet the requirements for quality evaluation of intelligent ballastless track construction.Taking the construction of the CRTS Ⅲ slab ballastless track as an example, this paper analyzes the key quality factors of intelligent construction processes, constructs an evaluation index system, and establishes a comprehensive quality evaluation model for intelligent ballastless track construction based on the Analytic Hierarchy Process (AHP) and the matter-element extension evaluation theory. Field tests are carried out on an intelligent construction test section of a high-speed railway for verification.The results show that the model established in this paper can objectively evaluate the quality of intelligent ballastless track construction. Compared with the traditional process, the guided intelligent construction process achieves a 32% improvement in the excellent rate. For the intelligent construction of the CRTS Ⅲ slab ballastless track, the construction quality of the track slab is the most important, followed by the quality of the self-compacting concrete construction, while the construction quality of the base slab has the least impact.The research results can provide a basis for evaluating the quality of intelligent construction of CRTS Ⅲ slab ballastless tracks for high-speed railways and offer a reference for the quality management of ballastless track construction in China’s high-speed railways.
To investigate the variation patterns of the temperature field and thermal stress in seamless CRTS Ⅲ slab tracks on simply supported beam bridges under high-temperature conditions, an indirect heat-stress coupling analysis method is employed. A finite element model of the coupled structural system of the heat-ballastless track is established to analyze the distribution and evolution of the temperature field, as well as the longitudinal force, stress, and displacement distribution between track structure layers under high temperatures. The results indicate that the temperature of each structural layer of the ballastless track exhibits a wave-like variation with ambient temperature. Over the course of a day, the maximum and minimum temperatures, 53.1 ℃ and 26.4 ℃, respectively, occur at the top of the track slab. The amplitude of the temperature time-history curves decreases with increasing vertical depth of the track structure, and the peak temperature shows a time lag. In summer high-temperature conditions, the temperature gradient of the track slab approaches zero around 11:00 and 21:00; the positive gradient peaks at 85.5 ℃/m at 15:00, while the negative gradient peaks at -43.8 ℃/m at 03:00. As the depth of the track structure increases, the temperature gradient gradually decreases. Under single-day high-temperature conditions, the rail longitudinal force and track structure displacement reach their maximum around 18:00 every day, representing the most unfavorable state, while the longitudinal stress in the track slab and self-compacting concrete layer peaks between 14:00 and 16:00. These findings provide a t theoretical reference for monitoring and maintenance of track structures in high temperature regions during summer.
In response to the challenge of extracting information from railway signal system station yard diagrams, this study proposes a primitive detection model, YOLO11-AT, based on an improved YOLO11. By constructing a detection model that integrates object detection and keypoint detection, it achieves automatic extraction of primitives and keypoints. First, an Attentional Scale Sequence Fusion (ASF) module is incorporated into the neck network to fuse multi-scale features, thereby enhancing detection performance for small targets. Second, a Task-Aligned Dynamic Detection Head (TADDH) is implemented in the head network, which improves feature interaction between classification and localization tasks through task alignment, reduces feature conflicts, and increases detection accuracy for densely distributed targets. Finally, Slicing Aided Hyper Inference (SAHI) is applied to improve detection accuracy on high-resolution station yard images. Experiments are conducted on a constructed dataset containing multi-style station yard diagrams to validate the proposed method. The results show that, compared with YOLO11s-pose, the proposed YOLO11-AT improves precision, recall, mAP0.5, and mAP0.5-kp by 9%, 2.2%, 4.2%, and 3.2%, respectively, while reducing the number of parameters by 4.3%. Compared with existing mainstream detection models, YOLO11-AT achieves a better balance between detection accuracy and efficiency. The results indicate that the proposed method is adaptable to various styles of station yard diagrams and provide a feasible solution for automated information extraction from station yard drawings.
To address the issues of low detection accuracy, slow detection speed, and frequent missed or false detections in railway track foreign object intrusion detection, this study proposes a lightweight railway track foreign object intrusion detection algorithm based on shallow feature fusion (YOLO-LSF). First, building on the YOLOv8n feature extraction network, the C2f module is improved based on GhostConv to construct the C2f_Ghost module, thereby reducing both the parameter count and computational cost of the model. Second, the MLCA attention mechanism is introduced at the end of the backbone network to enhance the feature representation of the target area and optimize the feature extraction efficiency of the model. Third, deformable convolution DCNv2 is employed to replace some ordinary convolutions in the C2f module of YOLOv8n, constructing the C2f_DCNv2 module and further strengthening the model’s feature extraction capacity. Finally, shallow feature information from the backbone network is integrated into the neck network, effectivelt mitigating detail loss caused by multiple convolution operations and enhancing the model's ability to detect distant foreign objects (small targets). Experimental results show that on a self- constructed railway track foreign object intrusion detection dataset, compared with the original YOLOv8n algorithm, the YOLO-LSF algorithm achieves an improvement of 5.2% in average precision, 3.37% in FPS, a reduction of 20.1% in the number of parameters, and a decrease of 22.2% in computational complexity. These results verify that the proposed algorithm significantly enhances detection accuracy and speed in complex environments while reducing the likelihood of missed and false detections.
