To alleviate the mismatch between flight supply and passenger demand, this paper proposes an optimization model for flight supply-demand alignment that incorporates passenger choice behavior. Based on 519 valid choice experiment observations, a Multinomial Logit (MNL) model is employed to characterize the heterogeneous preferences of business and leisure passengers regarding departure time, ticket price, aircraft type, and flight durations. These preferences are translated into differentiated flight demands at the time-slot level to accurately reflect the market’s behavioral responses to flight supply. First, considering multiple constraints such as flight schedule, aircraft assignment, and aircraft maintenance, a bi-objective optimization model is constructed. This model aims to optimize flight schedule allocations and aircraft assignments to minimize both the deviation in market demand fulfillment and operational costs. Second, a hybrid Greedy-Genetic Algorithm (GGA) is adopted to solve the model. By combining the global search capability of a Genetic Algorithm (GA) with the local repair and optimization abilities of a greedy strategy, this approach ensures optimization efficiency and high solution quality. Finally, a case study based on a one-week flight schedule from Airline A is conducted. The results indicate that the optimization scheme significantly improves the match between flight supply and demand, reducing the deviation in market demand fulfillment by 34.3% and lowering operational costs by 12.1%. Furthermore, flight resource utilization is enhanced, and the flight plan’s capacity to accommodate passenger demand is notably improved, particularly during high-demand periods. By comprehensively accounting for the relationships among flight schedules, aircraft types, and passenger demand, the optimization model ensures both the economic viability and operational feasibility of the flight plan. Overall, the proposed optimization model, which integrates passenger choice behavior, holds substantial application value for enhancing the utilization efficiency of flight resources and market responsiveness.

To address the frequent disruptions observed in pilot scheduling, this study proposes a pilot scheduling optimization model incorporating reserve factors. First, under the constraints of CCAR-121 and with the objective of minimizing total scheduling cost, a Min-Max approach is introduced, and a fairness objective is incorporated via ε-constraints to formulate a multi-objective optimization model. Second, a Logic-based Benders Decomposition (LBBD) algorithm is developed for solution. The master problem assigns pilots to flight pairings, while the subproblems verify whether feasible reserve pilot assignments satisfying operational requirements can be obtained based on the master solution. Finally, under controlled cost conditions, the fairness performance of the Min-Max objective is compared with that of a duty-period standard deviation minimization objective. The results indicate that as the planning horizon and workload increase, post hoc reserve assignment based on a given schedule is prone to producing disrupted schedules. The LBBD algorithm effectively identifies and eliminate such infeasible solutions by generating feasible cuts. Compared with direct solving Mixed Integer Programming(MIP) model using Gurobi, the LBBD approach reduces problem complexity and yields superior assignment solutions within the same computational time. In addition, compared with the duty-period standard deviation objective, the Min-Max fairness objective reduces the maximum duty-period deviation by 31.14%, thereby improving the fairness lower bound of the scheduling plan. Decision-makers can select scheduling solutions that balance cost and fairness based on the Pareto frontier according to operational requirements.

To improve the adaptability of high-speed train operation plans to daily passenger flows, this paper proposes an optimization method for high-speed train operation planning that considers the availability of unreserved tickets. Recognizing that not all passengers are willing to accept unreserved tickets, it is first necessary to identify the target demographic for this ticket type. A Logit model is established to analyze passenger acceptance of unreserved tickets when reserved tickets are sold out. Simultaneously, Origin-Destination (OD) passengers are categorized into two groups: those who exclusively accept reserved tickets and those who are willing to accept unreserved tickets. Passenger flow is allocated separately for these two groups during the subsequent solution process. Second, an optimization model is constructed with the dual objectives of minimizing passengers’ generalized travel time and railway operational costs. This is subject to constraints ensuring that the quantities of both ticket types meet diverse travel demands and that the ticket allo-cation in each section complies with train capacity limits. To account for generalized travel time, a time penalty coefficient is applied to the travel time of passengers purchasing unreserved tickets; this penalization aims to minimize unreserved ticket purchases, thereby safeguarding passenger comfort while satisfying the load factor requirement. Given the multitude of variables and the large scale of the decision-making model, a hybrid approach combining a Genetic Algorithm (GA) and the Gurobi solver is developed. The Gurobi solver is utilized to separately allocate passenger flow for reserved and unreserved ticket holders, simultaneously determining the ticket allocation scheme and the objective function values. These results then serve as fitness values for the subsequent GA operations. Finally, a case study based on the Beijing-Shanghai high-speed railway is conducted. The results demonstrate that, compared to operation plans that do not offer unreserved tickets, the proposed plan reduces total passenger travel time by 3.15% and railway operation costs by 8.71%. Furthermore, the ticket allocation scheme aligns closely with passenger demand, verifying the effectiveness of the proposed model and algorithm.

To address the difficulty of optimizing capacity in the collaborative scheduling of express and local trains for urban rail transit under the constraint of fixed express timetables, this paper proposes a coordinated optimization method for train timetabling. Based on the characteristics of publicly available express timetables and the spatiotemporal dynamics of passenger demand in actual operations, a mixed-integer programming model is developed. The model aims to minimize total train operating mileage, subject to the constraints of satisfying passenger travel demand. First, passenger travel characteristics are portrayed using sectional passenger demand, and operational constraints, such as train arrival and departure headways, station dwell times, and turnaround operations, are es-tablished. Second, the correlation between local train timetables and rolling stock circulation plans is formulated to achieve the integrated modeling of the timetable and rolling stock circulation. Third, given the large number of variables and the strong coupling of constraints in the model, an engineering application algorithm named Fast Solution Algorithm for Engineering-Stop Skip (FSAE-SS) is designed. By leveraging the structural characteristics of the model, the constraints are decomposed, and an iterative process is employed to achieve a joint solution for both the timetable and the rolling stock circulation plan. Finally, an empirical analysis is conducted using Line 10 of an urban rail transit system in Southwest China to verify the practical effectiveness of the model and algorithm. The results demonstrate that, compared to a traditional genetic algorithm, the proposed method reduces total operating mileage by 5.6% and maintains stable solution performance under large-scale constraints. Furthermore, compared to the actual operational timetable, the total operating mileage is reduced by 8.1%, achieving a better match between transport supply and passenger demand. The constructed collaborative optimization model successfully coordinates the local train operation plan with rolling stock circulation under fixed express timetable constraints. This provides an implementation pathway for capacity optimization that ensures timetable stability, as well as methodological support for cross-line collaborative scheduling under complex rail net-work conditions.

To alleviate the misallocation of capacity resources caused by the mismatch between passenger demand fluctuations and bus headways, and to balance service quality and operating costs of the transit system, this paper proposes an intelligent scheduling method for fixed-route bus timetables that considers the spatiotemporal distribution characteristics of passenger flow. First, bus passenger flow data are utilized to extract information on passengers’ boarding and alighting stations. A Gaussian mixture model is then employed to describe the spatiotemporal distribution of the passenger flow and fit the distribution of passenger arrivals at stations. Second, by comprehensively considering passenger travel time costs and transit enterprise operating costs, an integrated optimization model for bus timetabling is constructed based on a non-uniform departure time strategy, with the objective of minimizing the total system cost. Key indicators, such as the arrival passenger volume and waiting time cost at each station along the route, are calculated using the fitting results of the spatiotemporal passenger flow distribution. Third, tailored to the model’s characteristics, a Genetic Algorithm (GA) is utilized as the underlying framework, integrating two temporal-difference-based reinforcement learning methods, Q-learning and State-Action-Reward-State-Action (SARSA), to enable the adaptive adjustment of key algorithm parameters and enhance solution quality. Finally, a single route within the fixed-route transit network of Beijing is used as a case study to validate the proposed scheduling procedure. The results indicate that the improved GA incorporating reinforcement learning outperforms the traditional GA in minimizing the total system cost. For the single-route scenario, when the maximum number of departures is set to 10 and 15, the non-uniform scheduling strategy reduces passenger waiting costs by approximately 26.7% and 14.5%, respectively, and decreases total system costs by approximately 25.5% and 13.1%, respectively, compared to a fixed-headway strategy. Furthermore, the proposed scheduling strategy effectively balances the load factor, aligns well with passenger flow fluctuation trends, and enhances overall operational efficiency.

To address the spatiotemporal mismatch between energy supply and demand caused by the spatiotemporal heterogeneity of charging loads in dual-sided expressway service areas, this study proposes a coordinated scheduling method for multi-source energy flows based on direct current interconnection technology. Utilizing actual traffic flow, charging load, and equipment configuration data from the Hancunhe Service Area on the Jingkun Expressway, this study first analyzes the spatiotemporal distribution of traffic flow and charging demand, as well as their coupling characteristics. Second, it constructs a bidirectional energy mutual-support channel between the dual-sided service areas and establishes an hourly optimization model on a daily timescale. The model aims to minimize carbon emissions, maximize economic benefits, and minimize power grid fluctuations, and is solved using the Gurobi solver. Finally, three scheduling schemes, independent operation, Photovoltaic-Storage-Charging (PSC) self-consistency, and Photovoltaic-Storage-Direct Current-charging (PSDC) integration, are established for comparative simulation. Furthermore, a sensitivity analysis is conducted concerning temporal characteristics and weather conditions. The results demonstrate a significant synchronous fluctuation between traffic flow and charging demand on a daily timescale, with spatial load discrepancies between the dual-sided areas becoming notably more pronounced during holidays and other high-travel-intensity scenarios. Compared to the independent operation and PSC self-consistency schemes, the PSDC integration scheme achieves the best comprehensive performance. Specifically, daily carbon emissions are reduced by 71.2% and 41.8%, economic benefits are increased by 17.1% and 7.6%, grid power fluctuations are decreased by 94.6% and 66.7%, and comprehensive economic costs are lowered by 23.4% and 9.0%, respectively. Furthermore, this scheme exhibits strong adaptability under various temporal charging load distributions and photovoltaic output conditions. It performs optimally in typical scenarios characterized by significant spatial load discrepancies, such as peak outbound holiday travel days, peak inbound holiday travel days, and Sundays, where the cross-area mutual-support energy reaches 438.5 kWh, 236.9 kWh, and 166.3 kWh, respectively. These findings provide a valuable reference for the operational scheduling and low-carbon optimization of integrated PSC systems in dual-sided expressway service areas.

To address the instability and difficulty in effectively extracting characteristic differences among passenger groups when applying traditional clustering algorithms to high-speed rail passenger market segmentation, this study proposes a K-Means-Adaptive Learning Particle Swarm Optimization (KM-ALPSO) algorithm based on the Halton sequence. This algorithm is applied to segment the high-speed rail passenger market using ticket data from the Beijing-Shanghai high-speed railway. First, considering passenger age, advance booking time, and departure period as feature variables, the Affinity Propagation (AP) algorithm is adopted to identify representative sample points within the dataset. Second, an initial particle swarm is generated based on the Halton sequence, and the KM-ALPSO algorithm is used to cluster the representative sample points. A comparative analysis is then conducted against the classic K-Means algorithm and the K-Means-Particle Swarm Optimization (KM-PSO) algorithm. The Silhouette Coefficient (SC), Davies-Bouldin (DB) index, and Calinski-Harabasz (CH) index are selected to evaluate the clustering performance and determine the optimal number of clusters. Finally, the characteristic differences among various passenger groups are analyzed, and the Frequent Pattern-Growth (FP-Growth) algorithm is employed to extract strong association rules. The results indicate that preprocessing with the AP algorithm reduces the runtime of the Halton sequence-based KM-ALPSO algorithm to 23.26% of its original duration while maintaining evaluation metrics comparable to those obtained without preprocessing. Furthermore, initializing particle swarm positions with the Halton sequence enhances the global search capability of the KM-ALPSO algorithm. In the analysis of ticket data, the Halton sequence-based KM-ALPSO algorithm achieves an SC of 0.332, a DB of 0.933, and a CH of 708.5, demonstrating that a cluster number of five yields optimal performance and outperforms the baseline algorithms. The five identified passenger groups exhibit significant differences in age, advance booking time, and departure period, revealing distinct travel planning tendencies and departure time preferences.

To address the lack of reliable support for parameter calibration of utility functions in urban rail transit passenger flow assignment models and the insufficient consideration of heterogeneity in passenger travel behavior, this study proposes a parameter calibration method for the travel utility function based on passenger flow estimation. First, feasible routes are identified by integrating the network structure and train timetables, and passenger flow path inference is conducted using Auto Fare Collection (AFC) data. Then, the factors influencing route choice in urban rail transit are analyzed, and a generalized cost function incorporating travel time, number of intermediate stations, and number of transfers is constructed. An optimization model is developed to calibrate utility function parameters, and the Logit model is adopted for passenger flow assignment. Finally, multiple sets of parameters are calibrated using both observed survey data and estimated passenger flow data, and the assignment performance under different data sources and calibration strategies is compared. The applicability of the proposed method under different passenger flow characteristics is also evaluated. The research results show that parameter calibration based on AFC-derived passenger flow estimation yields significantly higher assignment accuracy than calibration based on survey data. Grouping Oorigin-Destination (OD) pairs by travel duration for parameter calibration effectively reduces discrepancies in bidirectional section flows and total network transfer volume. In the case study, the average discrepancy rates of bidirectional section passenger flows of the target lines decrease from 3.79% and 2.70% to 3.65% and 1.86%, respectively, while the discrepancy rate of total network transfer volume is reduced from 5.59% to 5.16%, indicating improved model accuracy. The proposed calibration method demonstrates strong accuracy and stability under different passenger flow patterns and path selection rules on both weekdays and holidays, making it adaptable to diverse urban rail transit travel scenarios.

To address the lack of differentiated analysis across varying travel distances in existing intermodal travel research, this study investigates the intrinsic mechanisms underlying intercity intermodal travel mode choice behavior. First, based on an analysis of traveler choice behavior, a nationally representative hub is selected for the design of a survey questionnaire. Second, from three dimensions, individual traveler attributes, travel characteristics, and latent psychological variables, an integrated SEM-MNL model is constructed by combining a Structural Equation Model (SEM) containing latent psychological variables with a Multinomial Logit (MNL) model for travel mode choice. The results are then compared with those of a traditional MNL model that excludes latent variables. Finally, based on the parameter regression results of the hybrid choice model, the decision-making differences across short-, medium-, and long-distance trips are systematically analyzed. The results indicate that, compared to the traditional MNL model, the SEM-MNL model exhibits superi goodness-of-fit and predictive accuracy, with improvements of 19.1%, 23.4%, and 10.6% for short-, medium-, and long-distance trips, respectively, thereby demonstrating strong practical applicability. Intercity travel choice behavior is jointly influenced by both observable and latent variables. Factors such as gender, education, occupation, income, travel purpose, and travel frequency exhibit significant effects, while latent psychological variables, including economy, safety, convenience, reliability, and comfort, also hold crucial explanatory power. Intermodal travel primarily attracts cost-sensitive occupational groups (e.g., government employees and self-employed individuals), for whom economy and safety constitute the core attractions, although insufficient convenience and comfort remain the primary constraints. As travel distance increases, travelers’ demands for the comfort of intermodal travel rise significantly, whereas their focus on reliability somewhat diminishes. These findings provide a theoretical basis for understanding the complex decision-making behaviors involved in intercity travel and offer an empirical reference for optimizing intermodal travel services and formulating differentiated transportation policies.

To investigate the relationship between the built environment and the carbon emission reduction benefits of bike-sharing, a Geographically Weighted (GW) machine learning model is proposed to analyze the nonlinearity and spatially heterogeneous impacts of the built environment. Based on bike-sharing order data from Shanghai and Xi’an, and integrating life-cycle carbon emission accounting, a “5D” built environment indicator system is established. The GW analysis is incorporated into the eXtreme Gradient Boosting (XGBoost) model, utilizing a bi-square kernel function to calculate spatial weights, thereby constructing the GW-XGBoost model. Localized weighted training is then utilized to capture the nonlinearity and spatial heterogeneity of the relationships among variables, facilitating an exploration of the nonlinear mechanisms by which the built environment influences the carbon emission reduction benefits of bike-sharing. The results indicate that the average carbon emission reduction per kilometer of bike-sharing in Shanghai and Xi’an is 51.09 g and 44.85 g, respectively. The relative importance of built environment variables differs between the two cities, and key variables exhibit nonlinear threshold effects. For instance, when population density is below the threshold of 40 000 people/km², it exerts a positive impact in both cities. However, when the density exceeds this threshold, the impact becomes insignificant. Furthermore, spatial heterogeneity analysis reveals significant spatial variations in the impacts of built environment variables. In Shanghai, for example, the marginal carbon emission reduction benefit of population density peak is highest in the central urban area, with SHapley Additive exPlanations (SHAP) values ranging from 10.01 to 47.52. Conversely, the effect turns negative in suburban areas, with SHAP values ranging from -80.25 to -50.01. Specifically, for every increase in population density of 10 000 people/km², carbon emission reduction can increase by up to 47.52 kg in the central urban area, while it can decrease by up to 80.25 kg in suburban areas. These findings provide a theoretical basis for formulating low-carbon development policies for bike-sharing systems.

To address the limitations of existing spatial morphology prediction studies, which predominantly focus on large scales, coarse classifications, and two-dimensional perspectives, making it difficult to capture the heterogeneous characteristics of land use and three-dimensional spatial morphology at the Traffic Analysis Zone (TAZ) level, this study proposes a TAZ-level spatial morphology prediction model based on a weighted Euclidean distance similarity model. First, based on the proportions of fine-grained land-use types and Floor Area Ratios (FAR), TAZs are categorized into 15 types, including high-density residential, medium-density residential, low-density residential, and industrial areas. Subsequently, a nested Logit model is developed to analyze the impact of population density, employment density, and various types of accessibility (calculated based on a multimodal transportation system) on TAZ categorization. Finally, the parameter results from the nested Logit model are applied as variable weights within the Euclidean distance similarity model, forming a weighted Euclidean distance similarity model to predict the three-dimensional spatial morphology of TAZs. The results indicate that the overall goodness-of-fit ρ ² of the nested Logit model is 0.46. The mean absolute errors for the proportions of fine-grained land-use types and FARs predicted by the weighted Euclidean distance similarity model are 0.062 and 0.592, respectively, demonstrating higher prediction accuracy than the traditional unweighted model. Furthermore, the influencing factors vary significantly across different TAZ categories. The formation of high-density residential areas relies more on service-oriented economies and population agglomeration effects than on industrial activities; medium-density residential areas reflect the characteristics of industry-city integration; the formation of low-density residential areas primarily depends on population distribution.

To address the challenges of calculating the carrying capacity and evaluating the capacity utilization of railway hub district stations, this study proposes a multi-agent-based simulation method for capacity calculation. First, the equipment functions and layouts, types of trains handled, and technical operation workflows at the hub district station are analyzed. Second, by combining multi-agent simulation technology with the technical operations of the station, three types of agents, trains, locomotives, and railcars, are designed. Their behavioral processes are analyzed, and corresponding decision-making modules are developed to establish a multi-agent simulation model for station operations. Third, a method for constructing oversaturated train flow conditions based on the station’s existing train structure is proposed. The generated train flow is input into the simulation model to determine the maximum number of arriving and departing trains under saturated conditions, thereby obtaining the carrying capacity of the station. Finally, a real-world case study is conducted to validate the effectiveness of the proposed simulation method, and various operational scenarios are simulated and analyzed. The research results indicate that the carrying capacity of the case station is 80 trains, and the existing capacity utilization rate is 82.5%, with a relative error of 3.6% compared to the traditional utilization rate method. Under current technical equipment conditions, increasing the proportion of non-reorganized transit freight trains or the number of handling tracks can effectively enhance the carrying capacity of the hub district station. The proposed simulation method effectively evaluate hub district stations’ carrying capacity, providing theoretical support for optimizing station operation organization and improving transportation efficiency.

This study investigates critical node identification in urban rail transit networks to address the limitations of existing methods, which often overlook community structure and differences in macro-level importance, thereby failing to comprehensively characterize the roles of nodes in both overall network organization and local connectivity. First, a critical node identification model integrating community structure and multi-scale node centrality is constructed. An improved Louvain algorithm based on consensus clustering is adopted to partition the rail transit network into communities, which are then abstracted as super-nodes. Subsequently, a weighted PageRank algorithm is used to quantify the macro-level importance of these communities. Second, at the node level, two micro-level centrality indicators, namely node strength and weighted betweenness, are selected. Grey relational analysis is applied to determine the indicator weights, and the macro-level community importance is introduced to revise the initial node importance, thereby establishing a comprehensive node importance evaluation model. Finally, an empirical analysis is conducted using the Beijing rail transit network as a case study. Deliberate attack simulations are performed to verify the model’s identification efficacy and the rationality of the node rankings. The results indicate that the Beijing rail transit network exhibits significant community structure characteristics, with varying macro-level importance across different communities. The identified critical nodes are primarily concentrated at hubs featuring multi-line intersections and cross-regional connections, displaying distinct spatial agglomeration. These nodes play a vital role in maintaining network connectivity, cross-regional links, and overall network stability. Under deliberate attacks simulated according to the model’s identification results, network performance degrades more rapidly. This demonstrates that the proposed model can effectively identify nodes with a critical impact on network stability, providing a novel analytical perspective and methodological reference for critical node identification, resilience enhancement, and the operational management of urban rail transit networks.

To address traffic congestion in the merging area of urban expressway on-ramps, this paper takes the Zhongshan North Road on-ramp of the Xuzhou North Third Ring Expressway as an empirical case study. A modular Deep Reinforcement Learning (DRL) framework integrating a Squeeze-and-Excitation Network (SENet) attention mechanism is proposed to optimize ramp signal control, formulating this dynamic and complex control problem as a Markov Decision Process (MDP). A comprehensive state space is designed, encompassing the grid-based traffic flow distribution characteristics of the merging area, signal phase states, temporal features, and the real-time traffic volume of Connected and Automated Vehicles (CAVs). Furthermore, a multi-objective reward function is constructed to balance real-time microscopic feedback with long-term system objectives. To ensure the simulation environment closely reflects actual traffic scenarios, an orthogonal regression method is adopted to calibrate the traffic flow model during the construction of the road network in the Simulation of Urban MObility (SUMO) microscopic traffic simulation platform. Consequently, the discrepancies between the simulated lane-changing ratios at the upstream, mid-section, and downstream segments of the merging area are controlled within ±10%. compared with field monitoring data. During the training phase, a Double Deep Q-Network (DDQN) is employed alongside prioritized experience replay and a progressive traffic flow learning strategy to enhance learning efficiency and stability. Experimental results indicate that, compared to traditional fixed-time signal control, the proposed method reduces the average delay rate by 40.5% and shortens the queue length by 29.5% during peak hours, outperforming the baseline DDQN model. Furthermore, as the CAV penetration rate increases from 20% to 80%, the average delay rate decreases by an additional 44.7%, demonstrating the model’s excellent congestion mitigation capabilities and strong adaptability in mixed traffic environment.

Existing vehicle trajectory optimization methods based on deep reinforcement learning struggle to achieve the dynamic merging and splitting of platoons in connected, mixed-traffic environments. Simultaneously, conventional signal timing optimization, which operates at the cycle or phase level, is hindered by response lag. The decoupled control between these two elements restricts further improvements in intersection efficiency. To address the necessity for the collaborative optimization of flexible vehicle trajectory control and real-time signal timing, this paper proposes a joint optimization method for platoon trajectories and signal timing based on Proximal Policy Optimization (PPO), termed FT-SIWO (Fleet Trajectory and Signal timing based on Waiting Offsetting). First, the vehicle trajectory optimization problem is formulated as a markov decision process. A state space integrating local vehicle dynamics and global traffic state variables, an action space encompassing car-following and multiple acceleration modes, and a multi-dimensional reward function considering factors such as driving comfort and safety are designed. This formulation enables the agent to comprehensively perceive the environment and learn efficient policies. Subsequently, the PPO-clip algorithm is employed for policy updates to enhance both training stability and sample efficiency. Finally, a real-time signal timing optimization mechanism based on Waiting Offsetting (WO) is introduced. By dynamically calculating and balancing the “time loss” of vehicles in the green phase and the “time surplus” of vehicles in the red phase, this mechanism enables the dynamic adjustment of the green light duration at a second-level granularity, thereby achieving deep coordination with the vehicle trajectory optimization. Simulation results on a joint SUMO and Matlab platform demonstrate that, compared to a non-optimized baseline scheme, FT-SIWO reduces the average vehicle delay at intersections by 28%-32% and the average waiting time by 30%-35%. Additionally, it decreases the total emissions of various pollutants (CO₂, CO, HC, NO _x, and PM _x ) by 34%-38% and reduces fuel consumption by 19%. Furthermore, when compared to state-of-the-art distributed control or vehicle-infrastructure cooperation schemes, FT-SIWO achieves an additional 6%-10% improvement in control effectiveness. By enabling intelligent platoon formation and a second-level linkage with signal timing, this method effectively enhances green time utilization and the overall operational efficiency of traffic flow. It provides a viable collaborative optimization solution for smart urban traffic management systems to handle mixed traffic flows, alleviate congestion, and reduce emissions.

To address the susceptibility of existing railway track foreign object intrusion detection algorithms to complex environments, which leads to low detection accuracy and high computational overhead, this paper proposes a lightweight detection method named Railway Lightweighting-YOLO (RL-YOLO). Based on an improved YOLOv12n architecture, this method is designed to facilitate lightweight deployment in complex environments. First, the High Performance GPU NetV2(HGNetV2) is introduced to replace the original backbone network, which is further optimized by integrating DWConv and GhostConv. This integration achieves dual-dimensional collaborative optimization, specifically, the decoupling of spatial convolution from channel fusion, and the generation of cost-effective “ghost” features alongside core features. This approach significantly reduces the number of model parameters and floating-point operations while enhancing the integrity of feature representation. Second, Lightweight Asymmetric Dual-Head (LADH) is adopted to further streamline the model for lightweight deployment. Third, an Adaptive Fine-Grained Channel Attention (AFGCA) mechanism is incorporated to improve the model’s capacity to interpret complex scenes. Finally, the Wise Focaler ShapeIoU (WF-ShapeIoU) loss function is employed to compel the model to focus more effectively on complex samples, thereby improving its localization capability. Experimental results demonstrate that the proposed method achieves a mean average precision (mAP@0.5) of 89.1% in complex environments, representing a 2.6% improvement over the baseline. The processing speed reaches 159 frames per second, an increase of 21 frames per second. Furthermore, the number of model parameters is reduced by 41.7%,GFLOPs is decreased by 34.5%. These results verify that the method successfully meets the demands for both real-time performance and lightweight deployment.

To address the low prediction accuracy of the Track Quality Index (TQI) in heavy-haul railways, this study proposes an analysis method based on machine learning. First, a raw dataset comprising 303 381 track inspection and maintenance records collected between 2015 and 2023 from 2 906 sections of a heavy-haul railway operated by a major energy enterprise is utilized. Following data cleaning, the dataset was categorized into four track section types based on geometric characteristics: neither gradients nor curves, curves without gradients, gradients without curves, and both gradients and curves. Second, the data for each category were processed into sequential datasets spanning three consecutive months without maintenance interventions. Sequences exhibiting an excessively large range in the TQI were then eliminated using anomaly detection methods. Third, using the TQI and its seven components (left-rail longitudinal level, right-rail longitudinal level, left-rail alignment, right-rail alignment, gauge, cross level, and track twist in triangle form) from the first two months of each sequence as inputs, six machine learning models, such as Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), and linear regression, are leveraged to perform rolling predictions of the TQI for the third month. Finally, the predictive performance of the different models across various section types is comparatively analyzed. The results indicate that when using the six machine learning models to predict the TQI across the four section types, the coefficient of determination R² between the measured and predicted values exceeds 0.94 for all models, suggesting highly accurate and reliable prediction performance. Among the six models, the RNN demonstrates particularly outstanding performance, achieving an average R² of 0.967 5 with a standard deviation of ±0.000 8 across various test sets. Specifically, for the four representative track sections, the prediction R² values are 0.990 1, 0.974 3, 0.988 3, and 0.988 9, respectively. Compared to the other five models, the RNN yields more stable prediction results and exhibits greater adaptability to diverse scenarios. This study demonstrates that machine learning methods effectively meet the demands of TQI prediction in heavy-haul railways, providing a valuable reference for establishing scientific tamping maintenance schedules and advancing intelligent track maintenance practices.

To address the irregularity characteristics of track spectra at different speeds, this study proposes a method for comprehensively evaluating track quality by combining the track irregularity quantile spectrum with the Track Quality Index (TQI). First, based on measured track irregularity data from four high-speed railway lines in East China, the improved Welch method is employed to calculate the power spectral density of track irregularities. The structural characteristics of the track irregularity power spectra at speeds of 200, 250, 300, and 350 km/h are then analyzed. By comparing these with classic domestic and international reference spectra, the structural differences across various speed grades are discussed. Second, the Chinese standard spectrum formula is used to fit the track irregularity data at the four speeds, and the adaptability of the fitted spectral parameters to the current standard spectral parameters is evaluated. Finally, the relationship between different individual standard deviations and the track quantile spectra is analyzed. By calculating the areas of the track quantile spectra at each level, the correlation between the track quantile spectra and the TQI is established. The results indicate that the variation trends of the track spectra at all speed levels are consistent with the current normative standard spectra; however, the spectral parameters exhibit certain discrepancies. Specifically, for vertical irregularities at wavelengths above 10 m and alignment irregularities at wavelengths above 22 m, the spectral lines are significantly elevated compared to the standard spectra, accompanied by substantial changes in the spectral parameters. This suggests that after years of high-speed railway operation, the standard spectral parameters require corresponding revision and refinement. The percentile spectra are divided into four intervals: 0-30%, 30%-70%, 70%-90%, and 90%- 100%, corresponding to excellent, good, fair, and warning track conditions, respectively. Based on actual line conditions, reasonable quantile spectra can be constructed and combined with the TQI to facilitate the comprehensive quality management of high-speed railway tracks.