ISSN 1673-0291 CN 11-5258/U
The nonlinear and nonstationary characteristics of vibration acceleration time series from train components degrade the predictive accuracy of the Informer model. To address this issue, this study proposes a Selective Kernel Mechanism Informer (SKInformer) model that integrates the Selective Kernel mechanism (SKmechanism) into the Informer architecture for predicting vibration acceleration time series of train components. First, sequence-dimension transformation converts one-dimensional time series into the two-dimensional feature maps required by the SKmechanism, enabling two-dimensional convolution to capture spatial correlations within the time series and enhance feature interactions across time steps. Second, the SKmechanism performs split, fusion, and selection operations on convolutional branches, allowing the network to adjust its receptive-field size dynamically according to the input and thereby improve its adaptability to multiscale features. Third, the SKmechanism is embedded in the Informer encoder. Through dynamic receptive-field selection, the model simultaneously captures short-term abrupt variations and long-term trends in the time series. Finally, the SKmechanism is incorporated into the Informer decoder to adaptively adjust the range of attention over the historical sequence and extract more discriminative local temporal representations. Ablation experiments on a pantograph dataset and comparative experiments on the LRV4313 light-rail vehicle dataset are conducted to validate the proposed method. Compared with the baseline Informer, SKInformer reduces the test-set loss by 34.18% at a learning rate of 0.000 1 in the ablation experiments and achieves a maximum coefficient of determination (R²=0.99) in the comparative experiments. These results demonstrate that the proposed method provides greater accuracy and robustness in predicting acceleration time series for train components.
Wheel-rail profile matching significantly affects the dynamic performance of wheel-rail interaction in fixed crossings. Changes in wheel and rail profiles during service intensify wheel-rail dynamic impacts and impair train running performance. To clarify the effects of wheel-tread wear and changes in the nose-rail profile on wheel-rail dynamic characteristics in the crossing area, this study investigates the nose-rail profile parameters under worn wheel conditions.First, a dynamic evaluation method that accounts for parametric reconstruction of the nose-rail profile is developed based on Non-Uniform Rational B-Splines (NURBS) and vehicle-track coupled dynamics. The wheel-rail dynamic characteristics of a train passing through crossings with different nose-rail profile parameters are then investigated for both standard and worn wheel treads. The relative effects of the wheel and nose-rail profiles are quantified, and the effects of the railhead slope and proportional coefficient on the dynamic responses are determined. Finally, appropriate nose-rail profile parameters are selected by comparing the dynamic responses of all parameter sets. The results show that, for a standard wheel tread, the nose-rail profile parameters significantly affect the wheel-rail interaction forces but have little effect on safety and ride-quality indices. For worn wheel treads, the railhead slope has a greater effect on the peak vertical and lateral wheel-rail forces than the proportional coefficient. Changes in the wheel-tread profile exert greater effects on the wheel-rail interaction forces, derailment coefficient, and wheel-load reduction rate than the nose-rail profile parameters, whereas both factors have little effect on the vertical and lateral carbody accelerations. A railhead slope of 1/20 combined with a proportional coefficient of 1/3 produces the fewest positive deviations in the maximum vertical and lateral wheel-rail forces across the operating stages; this combination is therefore selected as the nose-rail profile suitable for worn wheel treads. These findings provide a reference for the structural design and optimization of crossing areas and may further improve the dynamic performance of trains passing through crossings.
To address the low tracking accuracy, large error fluctuations, and limited disturbance rejection capability of conventional control methods during target-speed tracking in the automatic operation of high-speed trains, a novel speed control algorithm based on Improved Sliding Mode Active Disturbance Rejection Control (ISM-ADRC) is proposed. First, a single-point-mass model of a high-speed train is established. A novel nonlinear function is then introduced to replace the conventional fal function in the design of the extended state observer, thereby mitigating the chattering caused by the nonsmooth behavior of the fal function near the origin. Second, a saturation function is incorporated into an adaptive variable-speed reaching law, which is combined with a nonsingular fast terminal sliding mode surface to derive the sliding mode control law. This approach suppresses the chattering caused by the inherent discontinuous switching characteristics of sliding mode control. The stability of the control system is subsequently demonstrated using Lyapunov stability theory. Finally, simulations are conducted under both disturbance-free and disturbed conditions using the parameters of a CR400BF electric multiple unit. The simulation results show that the speed-tracking error of high-speed train remains within the range of -0.1 to 0.2 km/h, regardless of the presence of external disturbances. Compared with the traditional Proportional Integral Derivative (PID) control, Non-Singular Fast Terminal Sliding Mode Control (NFTSMC) and Sliding Mode Linear Active Disturbance Rejection Control (SM-LADRC), the proposed algorithm exhibits superior disturbance rejection capability and higher tracking accuracy. Its steady-state speed-tracking error is reduced by 95%, 89%, and 49%, respectively, relative to the three comparative methods, demonstrating the effectiveness and feasibility of the proposed approach.
Because the safety and comfort implications of flexible marshalling operations have received limited attention, this study compares a conventional longitudinal dynamics model with a refined model that represents the carbody-bogie-frame-wheelset system. Relevant performance indices are analyzed using both models, and a preliminary comparison with coupling tests on Beijing Metro Line 3 is conducted. Differences in train displacement, acceleration, and jerk predicted by the two models at different coupling speeds are examined. The simulation results show that, at marshalling speeds of 3 and 5 km/h, the refined model predicts train displacements that are 52.9% and 53.1% greater, respectively, and vibration accelerations that are 12.2% and 20.5% greater, respectively, than those predicted by the conventional model. Under different gradients, downhill marshalling produces a train displacement up to 89% greater than that under uphill conditions, indicating a higher risk of unintended rolling and reduced operational safety; however, acceleration decreases by 5%, suggesting a modest improvement in passenger comfort. Coupling tests on a test track show that, at 5 km/h, the refined model agrees more closely with the experimental results, reducing the relative error by approximately 30% compared with the conventional model. This confirms the feasibility of the refined model in simulating actural train-coupling scenarios.
To address the effects of the fluctuating and stochastic nature of ambient wind on the operational safety of high-speed trains under strong-wind conditions, this study investigates the aerodynamic characteristics of high-speed trains in a random wind field. First, a random wind model that more closely represents atmospheric turbulence is developed from field-measured wind-speed data and validated. Next, numerical models of a high-speed train in random and mean wind fields are established. Appropriate physical models, boundary conditions, and solvers are used to capture the development of the flow field around the train and the characteristics of the aerodynamic loads. Finally, case studies are conducted to reveal the evolution of the train-surface boundary layer, underbody flow field, and wake under the two wind conditions and to clarify how the wind-field characteristics affect the aerodynamic loads. The results show that the wind-speed time history generated using harmonic synthesis and a Davenport spectrum corrected through power-spectral fitting effectively represents the fluctuating and stochastic characteristics of random wind. Compared with the mean-wind case, the random-wind case exhibits broadly similar train-surface boundary layers, underbody flow fields, wakes, and large-scale leeward vortices; however, the flow field is substantially less stable, the flow structures fluctuate more strongly, and the small-scale structures around the train exhibit pronounced uncertainty. The surface-pressure distribution varies periodically, and the aerodynamic loads on individual cars fluctuate markedly, with different degrees of fluctuation among the aerodynamic forces and moments. The mean aerodynamic forces increase by less than 5%, whereas their extreme values increase by more than 50%. These findings provide theoretical support for aerodynamic shape optimization and operational safety control of high-speed trains in complex wind environments and can inform train-safety design for extreme conditions in random wind fields.
High-speed railway subgrades in karst-prone areas may experience soil loosening, cumulative deformation, and even soil-cavity collapse under long-term train-induced vibrations, thereby threatening ride comfort and operational safety. To address this problem, a method combining field vibration monitoring with three-dimensional dynamic numerical simulation is proposed to analyze the dynamic response of railway subgrades overlying karst soil cavities. First, field vibration tests are conducted on a typical high-speed railway subgrade section prone to karst collapse. Soil vibration acceleration responses during train passage were recorded, and time-frequency analysis is used to systematically examine the frequency distribution and attenuation characteristics of environmental vibrations. Next, based on multibody dynamics and finite element theory, a refined three-dimensional coupled dynamic vehicle-track-subgrade-foundation model incorporating a soil cavity is developed to analyze the dynamic response of the subgrade under train loading. Finally, multi-case parametric analyses are performed using the numerical model to reveal how soil-cavity burial depth and radius affect vibration propagation and the distribution of dynamic responses within the subgrade. The results show that environmental vibrations are concentrated mainly in the range of 0~35 Hz, with high-frequency components attenuating significantly faster than low-frequency components. The soil cavity scatters and obstructs elastic waves induced by train vibrations. Both the vertical dynamic displacement and vibration acceleration amplitudes are markedly lower near the cavity than farther away, and the attenuation rate of dynamic displacement is higher than that of vibration acceleration, indicating that displacement is more sensitive to the cavity-induced reduction in wave energy. Both amplitudes increase with increasing burial depth and decrease with increasing cavity radius; consequently, shallow, large soil cavities have the greatest influence on the dynamic response of the subgrade. These findings provide a reference for the disaster-prevention and mitigation design of railway subgrades in karst areas.
Accurate settlement prediction is essential for ensuring subgrade construction quality and the long-term safe operation of highways. Aeolian sand has low shear strength and can undergo substantial settlement deformation when used as subgrade fill, thereby compromising pavement structural safety. Taking an aeolian-sand section of National Highway G219 as the study site, this study proposes a high-accuracy PSO-BP-MGM(1,n) model for subgrade settlement prediction. First, settlement data collected using embedded monitoring instruments are used as modeling samples. Next, a hybrid model is constructed by integrating the Multivariable Grey Model (MGM(1,n)) with Particle Swarm Optimization (PSO) and a Back Propagation (BP) neural network. The MGM(1,n) model captures settlement trends, while PSO-BP performs nonlinear correction of the MGM(1,n) prediction residuals, thereby compensating for the limited nonlinear fitting capability of MGM(1,n) and reducing the tendency of BP neural networks to overfit small samples. Finally, the proposed model is applied to settlement prediction for the G219 aeolian-sand subgrade. The results show that the settlement predictions obtained using the PSO-BP-MGM(1,n) model agree most closely with the measured values. Compared with the MGM(1,n) model, the maximum reductions in Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and Sum of Squared Errors (SSE) are 1.341 percentage points, 0.528, and 1.817, respectively; compared with the hyperbolic model, the corresponding maximum reductions are 2.21 percentage points, 0.965, and 5.523. All absolute grey relational grades for the PSO-BP-MGM(1,n) model exceed 0.995, outperforming those of the MGM(1,n) and hyperbolic models. The proposed model accounts for interactions among monitoring points and is therefore suitable for joint multi-point prediction. By further modeling and correcting the MGM(1,n) prediction residuals, PSO-BP effectively identifies systematic components hidden in the residual sequences and improves the overall prediction accuracy. The proposed approach can support subgrade settlement control in aeolian-sand regions and settlement prediction under similar geological conditions.
To address the high susceptibility of tunnel portal construction in Grade VI surrounding rock to water and sand inrush, support structure deformation, and even portal collapse, this study proposes a tunnel entrance construction scheme combining vertical jet grouting piles, double-layer pipe roofing, and the Cross Diaphragm （CRD） method, based on a single-bore bidirectional tunnel project. A numerical comparison model using the single-layer pipe roofing-Center Diaphragm （CD） method is also established. The mechanical parameters of sandy loess are determined through laboratory tests, and construction mechanical simulations of the full tunnel excavation process are conducted together with field monitoring. The results show that reinforcing the portal slope with vertical jet grouting piles arranged in a quincunx pattern, together with double-layer advance pipe roofing to form a composite canopy support system, can substantially improve the overall stability of the portal slope and enhance the pre-support capacity of the surrounding rock ahead of the tunnel face. Compared with the original single-layer pipe roofing-CD method, the load carried by the pipe roofing decreases from 58.54 kN to 27.29 kN, representing a reduction of 53.38%; the temporary support stress decreases to 2.95 MPa; and the extrusion deformation of the tunnel face decreases from 10.47 cm to 3.17 cm, a reduction of 69.72%. These improvements significantly mitigate the risks of surrounding rock instability, water inrush, and collapse. Tunnel excavation induces stress redistribution in the surrounding rock, which can readily lead to a sharp increase in support stress. Following the construction principle of “rapid excavation and rapid support, sectional excavation and ring closure,” the CRD method enables rapid closure of the primary support structure and accelerates the convergence of stress and deformation in the surrounding rock–support system, demonstrating a pronounced effect in controlling deformation in extremely weak rock tunnels. Compared with the original CD method, the surrounding rock stress decreases from 6.95 MPa to 1.55 MPa, the internal force of the steel arch decreases from 93.94 kN to 63.74 kN, and the primary support stress decreases from 5.59 MPa to 1.25 MPa, corresponding to a reduction of 77.64%. The combined support system achieves three-dimensional pre-reinforcement of the surrounding rock through vertical jet grouting piles and double-layer pipe roofing, while the CRD excavation method effectively reduces the exposed sliding range of the tunnel face. Under these combined effects, extrusion deformation of the tunnel face is effectively constrained, thereby reducing the risk of tunnel collapse at the source. Field monitoring data indicate that, under the combined support measures, the peak values of crown settlement, surrounding rock stress, and steel arch internal force are 15.76 mm, 1.58 MPa, and 61.16 kN, respectively, and the overall support system remains safe and stable. The numerical simulation results agree well with the field monitoring results, verifying the reliability of the vertical jet grouting pile-double-layer pipe roofing-CRD tunnel entrance construction technique for shallow-buried tunnels in Grade VI surrounding rock.
To describe the nonlinear, dilative, and strain-softening responses of soil-structure interfaces during shearing, a hypoplastic constitutive model for sand-structure interfaces that accounts for in-plane stresses is developed based on the eight-parameter Gudehus-Bauer hypoplastic model for sand. The proposed model avoids the need to predefine yield surfaces and flow rules and overcomes the difficulty of extending conventional elastoplastic interface models to three-dimensional conditions. First, the three-dimensional stress tensor is reduced to a four-dimensional vector comprising the normal stress, two orthogonal shear stresses, and the in-plane stress, and a corresponding strain-rate vector is constructed. The modified stress and strain-rate components are substituted into the fundamental hypoplastic equation to derive the linear and nonlinear terms for the interface state. The model is then implemented in ABAQUS through a user material subroutine UMAT for finite-element analysis, and its predictive performance is validated against test results for interfaces between dam fill materials and steel plates, Hostun sand of different densities and smooth panels, and Ticino sand and aluminum plates. Finally, the QZ4-D2 test pile of the Beijing Sub-Center Station Integrated Transportation Hub is simulated using an axisymmetric finite-element model with a pile length of 68.3 m and a diameter of 1.2 m. For loads below 8 000 kN, the predicted settlements are in good agreement with the measurements, while at higher loads the proposed model provides more accurate predictions than the Mohr-Coulomb model does. At a load of 16 000 kN, the predicted shaft resistances at 12 depths have a mean absolute error of 9.13 kN and a mean absolute percentage error of 17.2%. The model also reasonably predicts the pile axial force, shaft resistance, and toe resistance, demonstrating its numerical robustness and practical applicability.
With the rapid development of urban rail transit, operational energy consumption has increased rapidly. Improving the utilization of regenerative braking energy has therefore become an important means of reducing energy consumption. Regenerative energy utilization is constrained by the spatiotemporal matching of train operations, and the simultaneous braking of multiple trains can cause traction-network voltage fluctuations. Although wayside energy storage systems can effectively address these problems, most existing charging and discharging strategies are based only on current train-storage states and do not consider how future coordination events may affect current decisions. To improve the foresight of such strategies, this study proposes a two-level optimization method based on Adaptive Particle Swarm Optimization and Model Predictive Control (APSO-MPC) for the cooperative utilization of regenerative braking energy by trains and wayside storage systems. First, a regenerative energy calculation model that accounts for dynamic conversion losses and line-transmission losses is developed. In the inner layer, an adaptive prediction horizon triggered by coordination events is designed; model predictive control is used to predict regenerative energy utilization during future coordination events and optimize the charging and discharging strategy of the energy storage systems in a receding-horizon manner. In the outer layer, adaptive particle swarm optimization searches for the optimal speed profile of each train, using the regenerative energy utilization returned by the inner layer as the fitness function. Experiments based on data from an urban railway line in China show that, compared with APSO combined with a default charging and discharging strategy, the proposed method increases indirect regenerative energy utilization by 16.91 kWh and reduces the number of charging and discharging operations by 28. Model predictive control performs best when the prediction horizon contains three discharge events. The minimum and average computation times per step are 0.001 2 s and 0.027 2 s, respectively, satisfying the requirements for online optimization. The method also demonstrates good recovery capability in response to unexpected timetable adjustments.
Evaluating railway track inspection systems requires a large number of high-quality images of defective fasteners. Because real defective samples are difficult to obtain, image synthesis provides an effective means of addressing data scarcity. Image harmonization is a critical step in this process. However, existing diffusion-based methods do not adequately account for the structural differences between foreground fasteners and background rail surfaces in railway inspection images, and the coupling between noise prediction and image harmonization limits the quality of the generated images. To address these issues, this study proposes a dual-branch rail-fastener image harmonization method based on Stochastic Differential Equations (SDEs). First, a Variational Auto-Encoder (VAE) maps the input image into a latent space, reducing the computational cost of processing high-resolution images while preserving essential structural information. Next, a Content-Matching Attention Module (CMAM) establishes correspondences between foreground fasteners and background regions. This module directs the model toward background fasteners that are structurally similar to the foreground fastener while suppressing interference from irrelevant rail-surface regions, thereby improving visual consistency between the foreground and background. Finally, a dual-branch diffusion framework comprising reconstruction and harmonization branches is constructed. The reconstruction branch learns stable noise prediction and guides the training of the harmonization branch through multiscale feature constraints, alleviating the optimization conflict between the denoising and harmonization objectives and further improving image quality. Experiments on a self-constructed dataset of defective railway fasteners show that the proposed method outperforms state-of-the-art image harmonization methods, reducing Mean Squared Error (MSE) and foreground Mean Squared Error (fMSE) by 14.37% and 8.91%, respectively, while achieving the best Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). Subjective evaluations and downstream fastener-detection experiments further confirm the high realism and visual consistency of the generated images, demonstrating that the proposed method can provide high-quality data for evaluating railway track inspection systems.
Sand-dust weather can severely degrade image quality, thereby reducing the robustness and reliability of computer vision systems. Moreover, the scarcity of paired sand-dust images and publicly available annotated datasets limits the application of deep learning methods. To address these issues, a sand-dust image restoration method integrating an improved network with a staged training strategy is proposed. First, the Horizontal/Vertical-Intensity (HVI) color space is extended by incorporating a dust dimension to form the Horizontal/Vertical-Intensity-Dust (HVID) color space, which accurately represents sand-dust degradation through color-cast correction and color adjustment. Second, the backbone of the Color and Intensity Decoupling Network (CIDNet) is enhanced with Sparse Nested Skip Connections (SNSC) and a Bidirectional Coupled Interaction Module (BCIM), yielding the Bidirectional Coupled Interaction Module Network (BCIMNet). This architecture improves cross-scale feature fusion and joint spatial-channel feature representation. Third, training pairs are constructed from synthetic samples, and a prior-guided staged pseudo-label learning framework is developed. Traditional methods and the enhanced network are used to progressively generate pseudo-labels, enabling domain adaptation from synthetic to real data. Finally, lightweight post-processing based on HVI priors is applied during inference to mitigate potential color bias introduced by the pseudo-labels. The results show that the proposed method significantly outperforms mainstream methods in restoring real sand-dust images. Relative to the baseline Horizontal/Vertical-Intensity Color and Intensity Decoupling Network(HVI-CIDNet) model, the Natural Image Quality Evaluator (NIQE), Perception-based Image Quality Evaluator (PIQE), and Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) scores decrease by 55.8%, 47.9%, and 46.1%, respectively. The proposed method provides effective technical support for image enhancement and high-level vision tasks in sand-dust environments.
Transit-Oriented Development (TOD) has promoted the construction of metro over-track complexes and improved the utilization of urban space. However, metro train-induced loads can propagate into the superstructure, amplifying floor vibration and reducing occupant comfort. Many early over-track complexes are constructed without vibration-isolation bearings at the column bases, while subsequent retrofitting is constrained by structural, spatial, construction, operational, and maintenance requirements, leaving few feasible mitigation options. To address excessive metro-induced vibration in existing buildings, this study examines an existing metro over-track complex with an inclined-member truss transfer story. A three-dimensional finite element model of the soil-tunnel-building system is established, and the modeling approach is validated against field measurements and results reported in the literature. Two post-installed vibration mitigation measures which include steel-spring floating floors and elastic supports are then proposed. Their effectiveness is evaluated using acceleration time-history responses, one-third-octave vibration levels, and relevant vibration-control standards. Dynamic elastoplastic analyses are also performed to investigate the effects of the retrofits on the seismic performance of the existing complex. The results show that vibration levels exceed the prescribed limits on several residential floors of the original structure, demonstrating the need for targeted mitigation. Both steel-spring floating floors and elastic supports effectively reduce the vertical vibration response and improve occupant comfort without compromising the overall seismic performance of the original structure. Considering mitigation effectiveness, economy, constructability, and the spatial constraints of existing buildings, the steel-spring floating floor should preferably be installed on the first floor of the superstructure, with a slab thickness of 100~150 mm. For the elastic-support system, a target natural frequency of 3~5 Hz is recommended. These findings provide a basis for vibration control, retrofit-scheme selection, and integrated vibration and seismic control of existing metro over-track complexes with truss transfer stories.
High-speed vacuum-tube trains can travel at speeds exceeding 1 000 km/h through confined, elongated tubes. To capture the enclosed nature of the long, narrow vacuum-tube environment and the nonstationarity caused by high-speed train operation, a Geometry-Based Stochastic Model (GBSM) incorporating a Fresnel-zone-based cluster identification mechanism is proposed to analyze the train-to-ground wireless channel. First, a geometric framework and a three-dimensional Cartesian coordinate system are established for the confined, elongated tube. Second, the channel is geometrically characterized using the GBSM. Third, valid scattering clusters are identified under Fresnel-zone constraints, and the statistical distributions of the azimuth and elevation angles of arrival are obtained. Finally, the Channel Impulse Response (CIR), Doppler shift, and spatiotemporal correlation functions, including the spatial Cross-Correlation Function (CCF) and temporal Auto-Correlation Function (ACF), are derived, and the coherence characteristics in the time, frequency, and spatial domains are analyzed. Simulation results show that the confined, elongated vacuum-tube structure produces a strong waveguide reflection effect, resulting in extremely high spatiotemporal correlations among different transmitter-receiver links within the tube. The ACF remains above 0.967 1, the CCF exhibits multiple high-correlation peaks, and channel coupling among multiple links is pronounced. These findings provide practical guidance for antenna layout and time-frequency synchronization algorithm design in train-to-ground communication systems for high-speed vacuum-tube trains.
To improve the operational efficiency of human-machine mixed traffic at intersections in a connected traffic environment and mitigate conflicts arising from interactions between the Connected and Automated Vehicle (CAV) and the Connected Human-driven Vehicles (CHV), this study proposes a presignal-based separation control method. First, a CAV-dedicated lane, a secondary stop line, presignals, and a queueing area are installed on each intersection approach. By coordinating the timing of the main signals and presignals, CAVs are given priority to enter the queue upstream of the main stop line, thereby separating and sequencing CAVs and CHVs in space and time. Second, a joint vehicle-trajectory and signal-timing optimization model is formulated to minimize average vehicle delay. The model incorporates the signal cycle, green time, vehicle speed and acceleration, safe car-following distance, lane-changing path, and coordination constraints between the main signals and presignals. Particle Swarm Optimization (PSO) is used to solve the model, and fifth-degree Bézier curves are used to plan CAV lane-changing paths. Finally, MATLAB simulations are conducted under different traffic-demand levels, and the proposed method is compared with fixed-time signal control, a First-In, First-Out (FIFO) strategy, and Phase-Continuous Tandem Intersection (PCTI) control. Under high traffic demand, the proposed method reduces average vehicle delay by 28.27% and 19.4% relative to FIFO and PCTI, respectively, while also reducing high-risk lane-changing states and improving stopping characteristics. Sensitivity analysis shows that CAV penetration, queueing-area length, and communication delay significantly affect control performance. The relationship between CAV penetration and average vehicle delay is nonlinear; the queueing-area length should be adjusted dynamically to traffic demand; and low communication delay is more conducive to realizing the benefits of coordinated CAV control. These findings provide a theoretical and technical basis for refined traffic organization and coordinated intersection control under human-machine mixed traffic conditions.
To address the vulnerability and nonlinear characteristics of mixed traffic flows comprising Connected and Automated Vehicles (CAV) and Human-Driven Vehicles (HDV) under adversarial disturbances in expressway merging areas, this study proposes a heterogeneous cooperative pinning control method based on Risk Potential-Driven Spectral Clustering (SC-Risk). First, from the perspective of Cyber-Physical Systems (CPS) and based on composite risk field theory, a risk-driven dynamic network model of vehicle groups is developed to quantify the asymmetric risk coupling strength between nodes. Second, the conventional spectral clustering algorithm is improved by introducing a maximum risk-potential coverage strategy, which identifies causal nodes with high-risk propagation potential as pinning nodes through feature-space mapping. Finally, a heterogeneous hierarchical control protocol combining Dueling Deep Q-Network (Dueling DQN)-based leading control with Cooperative Adaptive Cruise Control (CACC)-based following control is established. The pinning nodes employ reinforcement learning to make defensive decisions and contain risk sources, while the following nodes adopt CACC to maintain string stability. The results demonstrate that the proposed strategy increases the platoon synchronization rate to 0.877 3 and reduces the speed oscillation index to 3.55. By actively interrupting the upstream propagation of traffic shockwaves, it reduces the spatiotemporal impact area of congestion by 73.3% and decreases the delay index from 2.54 to 1.21. The proposed method therefore enables proactive risk prevention and cooperative optimization of mixed traffic flows in expressway merging areas.
To improve the accuracy of traffic flow prediction for urban road networks, this study proposes a Multi-Factor Fusion Spatiotemporal Graph Convolutional Network (MFSGCN). The model integrates long-term trends, short-term transients, periodic patterns, and time-varying external factors within a unified spatial graph convolution framework. First, MFSGCN employs a Graph Convolutional Network (GCN) to effectively capture spatiotemporal dependencies in the traffic network and model spatial relationships. Second, it incorporates a dynamic adjacency matrix to adaptively learn the time-varying characteristics of the traffic network. Third, it uses a multi-head self-attention mechanism to fuse long-term, short-term, periodic, and time-varying features, thereby effectively capturing traffic flow variations during periods of substantial external influence. Finally, extensive experiments on the METR-LA, PEMS-BAY, and PEMS04 datasets verify the effectiveness and validity of MFSGCN. The results show that MFSGCN outperforms benchmark models, including the Spatial-Temporal Transformer Network (STN), Dual Dynamic Spatial-temporal Graph Convolution Network (DDSGCN), and Domain Adversarial Graph Neural Network (DAGNN), in terms of Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). For the 15-min prediction task, MFSGCN reduces RMSE by up to 1.11; for the 60-min long-horizon prediction task, it reduces RMSE by up to 1.6. These findings can inform urban traffic management and traffic signal optimization.
To meet the need for reliable parking-slot perception in automated parking systems under complex conditions and to address the limitations of existing end-to-end parking-slot detection methods in modeling structural integrity and robustness, this study proposes an end-to-end parking-slot detection method based on sparse keypoint detection and structural matching. The method uses parking-slot keypoints to construct a dual-branch architecture that simultaneously performs keypoint localization and predicts matching relationships. First, Sparse Pairwise Focal Loss (SPF-Loss) and point regression loss are adopted to optimize training in terms of structural constraints and geometric accuracy, respectively, thereby improving the model’s ability to identify ambiguous regions and difficult-to-match samples. Next, the Structure-Aware CutMix (SA-CutMix) data augmentation strategy and the Attention-Guided Interpolation (AGI) module are introduced to enhance structural representation and adaptability to environmental variations, improving detection stability under disturbances such as occlusion, shadows, and low illumination. Finally, extensive experiments are conducted on the public ps2.0 dataset and the self-constructed real-world CRPS-D dataset to evaluate the proposed method. The results show that the method achieves a precision of 99.37% and a recall of 99.47% on ps2.0, comparable to those of mainstream methods. On the more challenging CRPS-D dataset, its Average Precision (AP) is 3.47%~4.56% higher than that of existing methods. Tests in several representative complex scenarios further demonstrate the adaptability and robustness of the method, while ablation experiments confirm the effectiveness of each module. The proposed method provides a viable technical approach to visual parking-slot perception for automated parking systems in complex environments.
