Automated pavement crack repair offers a promising approach to significantly extend road lifespan and is crucial for intelligent road maintenance. To tackle challenges associated with real-time crack trajectory extraction and substantial sealing errors, the Automated Pavement Crack Sealing Robot (APCSbot) was developed. APCSbot integrates a real-time crack trajectory segmentation network (S2TNet) and a cross-entropy-based adaptive fuzzy control method (CEAFC) for crack sealing repair. The S2TNet incorporates Anchor Ratio IoU Sampling (ARIS) and Balanced Fine-Grained Features (BFGF) to enhance the detector's capability in predicting bounding boxes and segmenting instance binary masks, consequently improving crack trajectory extraction accuracy. The CEAFC method employs cross-entropy optimization iterations to tune controller parameters and constructs fuzzy logic to enhance repair control robustness. Furthermore, an unmanned wheeled robot framework based on four-wheel independent differential drive was established, integrating the crack segmentation network and tracking repair control methods. Extensive experiments conducted on DeepCrack, CFD, and S2T-Crack datasets demonstrate a real-time pavement crack segmentation accuracy of 80.21%. The crack sealing repair process achieves a speed of approximately 0.05 m·s^-1, with an average sealing error for slender cracks of 5.17 mm. The APCSbot showcases its accuracy and robustness in pavement crack sealing repair, thus providing technical support for intelligent road maintenance.

Pavement crack segmentation is a key technology for road maintenance management and is crucial for enhancing road safety and durability. However, due to the complex and diverse pavement scenarios, the irregularity of crack morphology, and background noise interference, implementing high-precision crack segmentation still faces many challenges. To tackle these issues, a pavement crack segmentation method based on the fusion of frequency domain Mamba and recursive gated Transformer is proposed. By deeply fusing frequency domain modeling and spatial feature extraction, it effectively enhances the model's ability to capture crack edge features and global contextual information. Firstly, a Transformer architecture employing a large-kernel recursive attention mechanism is constructed as the backbone of the U-shaped network. By enhancing spatial feature interaction, effective fusion of local details and global contextual features within the module is achieved. Secondly, the Sobel operator and a gating mechanism are integrated into the Feedforward Network (FFN) to dynamically regulate the information flow, significantly improving edge feature extraction capability and noise suppression effect. Finally, a Frequency Domain Mamba enhancement module embedded in the skip connections is proposed. By introducing the State Space Model (SSM) during the frequency domain decomposition of wavelet transform, it aims to enhance the fusion of frequency domain and spatial domain features to compensate for the detail information loss during downsampling. To verify the effectiveness of the proposed method, this paper conducted systematic experiments and comparative analysis on the DeepCrack537, CrackLS315, and YCD datasets. The experimental results demonstrate that this method outperforms current mainstream segmentation methods in both visual effects and quantitative metrics, exhibiting good robustness and practical value. It provides a new solution for the intelligent detection of pavement cracks.

To effectively describe the role of mineral aggregate in the skeleton structure of asphalt mixture, this study introduced complex network theory to characterize and analyze the topological structure of skeleton. Firstly, 11 types of asphalt mixtures were prepared. Then, the information of the mineral aggregates was extracted to constructure an unweighted and undirected skeleton network model by the digital image processing (DIP) technology. Subsequently, complex network theory was introduced to characterize the topological features of the skeleton. The results indicate that the scale-free index α of the skeleton network is much larger than 1, the small-world index S is around 0.8, and the fractal dimension is 1.94 times that of the equivalent ER random network, which indicates that the skeleton network exhibits scale-free, small-word self-similar properties. Its topological structure has significant structural complexity, aligning with the characteristics of complex networks, which confirms that the complex network theory is suitable for characterizing the topology of skeleton. Within the same standard size interval, the number of aggregates initially increases and then decreases with the degree value increasing. Simultaneously, the degree value associated with the peak also gradually increases with an increase in the aggregate size. In AC type of asphalt mixture, the proportion of the degree value contributed from mineral aggregate with size greater than 1.18 mm is about 12%, which is 18.8% and 37.2% less than that of SMA and OGFC types of asphalt mixtures, respectively. The results indicate that the aggregates with larger size have smaller contact distances, and can form a stable skeleton structure for SMA and OGFC types of asphalt mixtures. When the aggregate size is less than 2.36 mm, the average normalized betweenness values of aggregates are all less than 0.01 in different particle size ranges, indicating that these aggregates mainly play the role in filling the skeleton. The average betweenness value of the mineral aggregate within 2.36-9.5 mm slowly increases to 0.05 which indicates that although the mineral aggregate within 2.36-9.5 mm participates in the skeleton composition, it only plays a secondary bearing role. The average betweenness value of the mineral aggregate with size larger than 9.5 mm increases significantly, even reaching up to 0.3, indicating that the mineral aggregate with size larger than 9.5 mm plays a major role in the skeleton composition. Compared to SMA and OGFC types of asphalt mixtures, due to the lower content of coarse aggregates, the betweenness of aggregate particles is larger in AC type asphalt mixtures.

In order to realize the high value and diversified utilization of muddy slag, a new artificial granulation technology that used alkali-activated blast furnace slag (GGBS) was developed to solidify muddy slag and achieve high strength and high water resistance under standard curing conditions. This paper experimentally explored the effects of factors such as alkali activator type, dosage and curing time on the mechanical strength and water resistance of solidified soil particles, and determined Ca(OH)₂ and Na₂SiO₃ as alkali activators and the optimal mixing ratio. On this basis, Fourier transform infrared spectroscopy (FTIR) and scanning electron microscopy (SEM) were further used to microscopically characterize the artificial particles, revealing the strength formation mechanism and evolution law of alkali-activated GGBS solidified soil particles. The study shows that GGBS undergoes hydration reaction under the alkali activation of Na₂SiO₃; when the Na₂SiO₃ dosage reaches more than 10%, the particle strength reaches more than 5 MPa after 7 days of standard curing, and the softening coefficient is higher than 0.75. By adding Ca(OH)₂ to replace part of Na₂SiO₃, the alkali activation effect can be further enhanced, and the optimal mixing ratio is 1∶3. The particle strength is increased by more than 20% compared with the use of Na₂SiO₃ alone. Microscopic experiments show that when alkali-excited GGBS generates hydrated calcium silicate (C—S—H) and hydrated calcium aluminosilicate (C—A—S—H), an inorganic material M_x{—(SiO₂)_zAlO₂—}_n·wH₂O with a high degree of polymerization and a three-dimensional network structure is synthesized, which can fill pores and bond soil particles, significantly enhancing the strength of the soil particle blank. In this paper, artificial aggregate is made by alkali-activated GGBS solidified sludge, which has the characteristics of light weight (1.8-2.0 g·cm^-3), high strength (≥5 MPa), water resistance (softening coefficient＞0.75), low energy consumption (20 ℃ cold curing), green and environmental protection (solid waste utilization rate ≥85%), etc. It can be used to replace natural fillers in traffic roadbed base, backfill of cross-sea bridge pedestals, and protective structures of coastal highways.

Chemical stabilization serves as a principal approach for the resource utilization of shield tunnel spoil by modifying it to meet diverse fill soil requirements. Compared to traditional cement and lime stabilization methods, magnesium oxide (MgO)-based stabilization-carbonation technology exhibits advantages in energy conservation, environmental friendliness, and low-carbon sustainability, rendering it a promising alternative worthy of exploration. In this study, a self-developed accelerated carbonation apparatus was employed to investigate the plasticity and strength evolution of MgO-stabilized-carbonated shield tunnel clayey spoil under varying MgO dosages, curing ages, and carbonation durations, aiming to identify optimal modification conditions for its application as highway subgrade fill. Results demonstrate that MgO stabilization-carbonation treatment of earth pressure shield spoil leads to a slight increase in plastic limit (w_P), accompanied by significant reductions in liquid limit (w_L) and plasticity index (I_P). A complementary relationship exists between curing age and carbonation duration in reducing soil plasticity: appropriately extending carbonation time markedly shortens the curing age required to achieve I_P compliance with specifications. Post-carbonation, the soil transitions from high-liquid-limit clay to low-liquid-limit clay, with unconfined compressive strength (UCS) substantially enhanced, peaking at 4 h of carbonation. During the carbonation process, magnesite (MgCO₃) and hydration products undergo mutual cohesive bonding, effectively filling soil pore spaces to form a compact reticulated structure. However, with prolonged carbonation exceeding 4 hours, partial decomposition of hydration products occurs, accompanied by an increase in pore volume. This microstructural evolution mechanism fundamentally accounts for the persistent reduction in material plasticity and the characteristic strength pattern of initial enhancement followed by subsequent degradation. Under optimal conditions (9% MgO dosage, 7-day curing, and 2 h carbonation), the treated soil exhibits a I_P below 26 and UCS exceeding 800 kPa, satisfying the plasticity and strength criteria for primary highway subgrade applications. This study provides a technical reference for the sustainable treatment and reuse of silty clay waste generated by earth pressure shield tunneling in the Yangtze River Delta region.

Vortex-induced vibration (VIV) of stay cables can lead to fatigue damage, failure of auxiliary structures, increased maintenance costs, and negatively impact the normal operation of bridges. To promote the theoretical understanding of VIV and improve vibration control technologies, this paper provided a comprehensive review of research that had been conducted on VIV of stay cables over the past 30 years. It focused on three key aspects: the mechanism and characteristics of VIV, major influencing factors and their underlying mechanisms, and vibration control measures with their practical effectiveness. Initially, through an analysis of VIV incidents globally and statistical data from 124 cable-stayed bridges with main spans exceeding 500 m, the paper identified a notable shift in research emphasis from international to domestic studies, highlighting persistent future research challenges. Subsequently, the evolution from classical Kármán vortex shedding to complex, large-amplitude multimodal vibrations was detailed from the perspectives of vibration mechanism and influencing factors, emphasizing three crucial parameters: incoming wind characteristics, Reynolds number, and damping ratio. Finally, various vibration control strategies-including aerodynamic measures, auxiliary cables, mechanical dampers, and hybrid systems-were systematically reviewed and analyzed in terms of their structural features, research progress, and practical engineering applications. Additionally, existing research gaps were identified, and future research directions were proposed, providing fresh insights into VIV control for cable-stayed bridges.

The rapid construction of infrastructure will inevitably lead to large-scale maintenance. To address the issues of maintaining the structural performance and enhancing the bearing capacity of long-span bridges, based on a summary of methods for enhancement, as well as closely related technologies such as structural damage simulation, optimization algorithms, and process safety monitoring, this paper focuses on the current status and development trends of improving bridge performance by changing the structural system, systematically reviews the research status and typical engineering applications of enhancement methods such as the transformation of simple-supported into continuous structures, adding support points, cable-stayed composite systems, suspension-composite systems, hanging systems, and beam-arch composite systems, identifies two main types: the method of increasing constraints and the method of additional structures, deeply analyzes the active transformation behavior of the structural system of long-span bridges, as well as the scientific mechanisms behind changes in structural states. The paper outlines the key issues, major challenges, and future development trends related to structural system modification and reinforcement. It highlights that this method involves both benefits and risks, and points out that the compatibility between the old and new systems requires systematic and in-depth research. The establishment of mathematical models for bridge damage remains a shortcoming limiting the research on structural system modification and reinforcement. Issues such as the construction of multi-variable and multi-objective functions and the formulation of optimal solution criteria still need exploration, and breakthroughs in solving algorithms for complex stress processes remain key. With the help of smart devices and digital twins, interactive collaborative design and intelligent construction methods that involve real-time monitoring, analysis, control, and feedback should be further explored.

To thoroughly investigate the fatigue performance of the orthotropic steel-UHPC composite bridge decks with large-size U-ribs, and to address the issues of scale fragmentation and information barriers inherent in traditional full-scale model fatigue tests, a multiscale integrated experimental research method was proposed by taking the interrelationships of performance indicators among the component, subassembly, and structural scales as the starting point. This method adopted a strategy of progressive advancement, integration, and feedback of multiscale performance indicator information, designed multiscale coordinated fatigue tests for components, subassemblies, and structures, established a modular experimental research pathway, and provided effective support for multiscale experimental studies under the same research objective. The results indicate that the proposed multiscale integrated experimental research method, through the information transfer and integration pathways among components, subassemblies, and structures, forms a systematically closed-loop research framework, which can effectively resolve the problem of data fragmentation in multiscale testing. At the component scale, typical fatigue-prone details such as the UHPC material, stud connectors, and steel bridge deck all exhibit performance degradation and fatigue damage accumulation characteristics. These can be quantitatively characterized through mechanical performance degradation models, S-N curves, and damage accumulation criteria, serving as fundamental inputs for upper-scale assessments. At the subassembly scale, segment model tests can elucidate the fatigue damage evolution paths of each segment model, determine their fatigue failure modes, identify the controlling locations of the UHPC layer, stud connectors, and typical fatigue-prone details of the steel bridge deck, and establish mapping relationships between local responses and component performance indicators. This provides experimental basis and theoretical support for design parameter optimization and structural-scale fatigue response analysis. At the structural scale, in-situ monitoring results of the actual bridge show that the strain responses at key controlling locations are stable and the degree of damage is low. The performance indicators demonstrate good applicability and consistency across the component, subassembly, and structural scales.

By following the progressive hierarchy from pure torsion to shear-torsion and then to compression-bending-shear-torsion (CBST), a refined torsional softened membrane model (RTSMM) was proposed for studying the composite mechanical performance of CBST of the concrete box girders in the cable-stayed bridge. Based on the composite mechanical state of girder, the strain gradient generated by torsion, the shear strain generated by shear were combined through a modified constitutive relationship. Using the softened membrane model to solved the target strain values of the superposition zone subjected to initial shear, and the average stress coefficient or stress-strain relationship that changes due to the existence of the initial state was further corrected. Thus, the softened membrane model of the concrete box girder was extended from pure torsion to shear-torsion. The axial pressure plays a pre-stressing role in resisting torsion and shear and was also considered as the initial state before torsion. The bending moment was reflected in the asymmetric axial pressure between the top and bottom plates of the concrete box girder. By combining the box girder and slab components through Bredt's theory, the shear-torsion model was further extended to composite CBST. The correctness of RTSMM was verified by the designed composite torsion test model of the concrete box girder, and it can accurately calculate the torsional stiffness of the girder section under composite stress state, and predict the load-displacement curve of the torsion member throughout the entire process. The results of a real bridge show that the torsional failure first occurs in the weakest plate, which determines and restricts the torsional limit value of the entire section. The combined effect of shear and torsion greatly reduces the ultimate torque value of the section. The shear-torsion coupling effect is significant, with amplification effects on both parties. The torsional failure of box girders under the influence of shear-bending generally occurs in the web plate of the shear-torsion superposition zone and the top/bottom plates of the tension zone by bending. Axial force and prestress can significantly enhance torsional stiffness.

Relying on the engineering background of a large-span suspension bridge with a main span of 928 meters located in a strong typhoon region，a combination of wind tunnel testing and numerical simulation was employed to study the flutter stability of the closed streamlined steel box girder of the large-span suspension bridge，as well as the mechanisms of aerodynamic control measures.Firstly，the dynamic characteristics of the main bridge structure in its completed state were analyzed.Then，wind tunnel tests were conducted on the original design of the closed streamlined girder section and various aerodynamic control measures，obtaining the corresponding flutter critical wind speed and instability characteristics.Finally，a systematic study was carried out using Computational Fluid Dynamics (CFD) to investigate the flutter characteristics of the closed streamlined girder section and the mechanisms of the aerodynamic control measures from two perspectives：aerodynamic forces and flow field characteristics.The results show that the installation of a central stabilizing plate on the closed streamlined steel box girder can significantly enhance the critical flutter wind speed of the bridge while substantially increasing the participation level of vertical vibrations，and the combination of a central stabilizing plate and horizontal splitter plates further enhances the flutter critical wind speed.The critical flutter wind speed of the closed streamlined main deck section based on aerodynamic derivatives method and Fluid-structure-interaction (FSI) agree well with the wind tunnel test results，respectively.The addition of horizontal splitter plates to the closed streamlined steel box girder primarily serves to guide airflow over the bridge deck surface，thereby reducing pressure fluctuations at the leading edge.Concurrently，the installation of a central stabilizer plate facilitates the migration of deck vortices toward the central axis，concentrating pressure fluctuations near the mid-section.This combined configuration effectively mitigates aerodynamic self-excited torsional moments，collectively achieving the objective of enhancing the critical wind speed for bridge flutter onset.

The confinement effect leads to significant differences in the longitudinal shear bearing capacity and force transmission mechanism of PBL in steel plate-constrained UHPC compared to non-steel plate-constrained UHPC. Currently, there is a lack of in-depth research and sufficient understanding regarding this phenomenon. To address this gap, this paper designs two types of seven sets of push-out specimens: the lattice type of steel plate-constrained UHPC and the insertion type of non-steel plate-constrained UHPC. Notably, both types of specimens lack pressure-bearing ends for either perforated or non-perforated steel plates. Through the experiments conducted, we obtained data on the failure modes, ultimate bearing capacity, and load displacement curves for both types of push-out specimens. The analysis focused on the composition of longitudinal shear bearing capacity and the force transmission mechanism of the push-out specimens while exploring how different bearing capacity compositions impact their longitudinal shear performance. The results indicated that, except for the pure bonding specimens, shear failure occurred in the concrete tenon for both specimen types. In the lattice type specimens, the penetrating steel bars experienced pure shear failure, whereas flexural shear failure was observed in the insertion type specimens. In the cell type specimens, the ultimate bearing capacity of PBL increased by 59%, 51%, and 125% respectively when compared to bonded pure tenons, bonded steel bars, and non-bonded PBL specimens. Moreover, the load-relative displacement curve of PBL lattice specimens can be categorized into five stages, indicating ductile failure, while the load-relative displacement curve of PBL insertion type specimens can be divided into four stages, reflecting brittle failure. Under typical engineering conditions, the longitudinal shear bearing capacity of PBL in steel plate-constrained UHPC is 2.05 times greater than that of PBL inserted components. The longitudinal shear bearing capacity in this context is composed of peripheral friction, concrete tenon shear capacity, and penetrating steel bar shear capacity. Based on the experimental results and the shear bearing capacity formulas for PBL proposed by various domestic and foreign scholars, we established a longitudinal shear bearing capacity formula for both PBL and unconstrained components in steel plate-constrained UHPC. The calculated values aligned closely with the experimental values, and this formula can also be used to estimate the longitudinal shear bearing capacity of PBL or unconstrained components in steel plate-constrained concrete.

Accurate cable condition identification is critical for long-span bridge safety and efficient perception in lightweight inspection/monitoring systems. Traditional string theory and empirical methods ignore coupling effects of parameters (e.g., bending stiffness, constraint stiffness), leading to large stay cable identification deviations. These deviations fail to meet lightweight monitoring's high-precision needs and may cause structural internal force redistribution, increasing local component damage risk.This study proposed a multi-parameter intelligent inversion method driven by multi-order frequencies. A vibration/modal equation (integrating inclination angle, sag, bending stiffness, boundary elastic constraints) was established, and its matrix form derived via the finite difference method. A multi-parameter collaborative inversion framework was built: multi-parameter feasible domains were defined using dimensionless characteristic parameters and empirical formulas; an intelligent engine based on the meta-heuristic Grey Wolf-Krill Herd Optimization algorithm was designed; its global convergence and robustness were verified through optimization and generalization analysis. Laboratory vibration tests on rigid short cables and on-site tests on stay cables were conducted. Results show the proposed method achieves high cable force inversion accuracy for both rigid short and medium-long cables, with errors controlled within 5%. Compared with traditional string theory, beam theory, and the dual-frequency method, its accuracy for rigid short cables is significantly improved by 20%-35%. Within the frequency-mapped feasible domain, its bending stiffness inversion accuracy is comparable to the unified modified string theory. Engineering verification indicates the method effectively realizes multi-parameter collaborative inversion for both cable systems, exhibiting excellent engineering applicability in computational efficiency and identification reliability. This study provides key technical support for efficient perception in bridge lightweight inspection and monitoring systems.

In this paper, a new approach for bridge automated operational modal analysis based on the Dirichlet Process Mixture Model (DPMM) that requires only one clustering operation is proposed, addressing the challenges in existing Stochastic Subspace Identification (SSI)-based methods, such as the difficulty in determining thresholds and the cluster count. The proposed method eliminates the need for manual parameter selection or threshold determination. First, the Covariance-driven SSI (SSI-Cov) algorithm is used to generate an initial stability diagram of the bridge. A stabilization diagram cleaning framework with different validation criteria is then applied to remove spurious modal information. The frequency, damping ratio, and mode information after stable diagram cleaning are used as high-dimensional clustering features to construct data samples for the DPMM model. The Dirichlet Process is employed to build an infinite mixture clustering model, and Collapsed Gibbs Sampling (CGS) is used to automatically determine the optimal cluster numbers. The proposed method is applied to the Dowling Hall Footbridge and a large-span suspension bridge, where modal parameters are extracted from measured vibration data. The results demonstrate that the method effectively identifies closely spaced modes of bridge structures, achieving automated identification of bridge modal parameters requiring only one clustering step, while the identified natural frequencies and mode shapes exhibit close agreement with theoretical values.

To reduce impact damage to bridge structures and improve post-impact serviceability, this study explored the use of fiber reinforced polymer (FRP) tubes around bridge piers to enhance shear resistance and confinement. Simultaneously, hybrid reinforcement combining conventional ribbed steel bars and high-strength threaded steel bars was employed to improve self-centering ability. The dynamic impact response and residual load capacity of FRP-confined hybrid reinforced concrete (FHRC) bridge piers were investigated. First, horizontal impact tests using “low-, medium-, and high-velocity” progressive loading revealed that FHRC piers experienced significantly less damage compared to conventional reinforced concrete (RC) piers. Peak displacement was reduced by 24.5%, and residual displacement after three impacts decreased by 53.6%. Then, compression tests on impacted specimens showed that RC piers suffered rapid strength degradation and severe damage, whereas FHRC piers exhibited a 53.8% higher residual axial load capacity, slower strength degradation by 60%, and greater ductility. Finally, a finite element model was developed on the LS-DYNA platform to simulate sequential dynamic impact followed by static compression using a restart analysis. The numerical results aligned well with experimental data, validating the proposed approach. This research provides innovative structural strategies and reliable validation for enhancing the impact resistance and safety of key bridge structures.

Post-earthquake bridge damage assessment constitutes a critical component in maintaining the resilience of transportation networks and urban lifeline systems. While conventional assessment methodologies rely heavily on expert knowledge of structural damage mechanisms, this dependence significantly impedes time-sensitive post-disaster operations. Although artificial intelligence and computer vision demonstrate proficiency in crack recognition and feature extraction, existing approaches lack explicit correlations between visual patterns and structural mechanics. To address these limitations, this study proposes an intelligent framework integrating multilevel physical damage criteria with computer vision and deep learning techniques, enabling automated damage state prediction through visual evidence analysis. The proposed methodology comprises two principal components: ① Extracting bridge pier damage images and physical damage information, including damage feature maps, lateral displacement, and longitudinal rebar stress. The damage feature map is generated by binarizing and skeletonizing the finite element model damage images to simulate the crack formation and development; ② Hybrid damage prediction through multi-level physical criteria fusion. Physical parameters and crack feature maps serve as training data to optimize a Residual Neural Network (ResNet) architecture. The prediction results include the lateral displacement and the longitudinal reinforcement stress, which are combined with the multilevel physical damage criterion to determine the structural damage state. In the prediction results of the test data set, the maximum prediction errors for top displacement and longitudinal stress are 22.38% and 19.4%, respectively, both of which are about 20% of the maximum error with the test data. Accurate damage state classification is achieved across all tested scenarios. This result confirms the robustness of this method in simulating and predicting post-earthquake damage progression in reinforced concrete piers, which provides an important idea for automatically identifying and evaluating the damage state of bridge piers after regional earthquakes

The seismic design of existing bridges in China does not meet current specifications, and some lack seismic design. These bridges are exposed to great seismic risks and have considerable seismic safety concerns. Double-column reinforced concrete (RC) bridge piers, a predominant structural form, urgently require seismic retrofitting. To meet the demand for improved seismic performance and resilience of double-column RC bridge piers, this paper proposed a novel seismic retrofitting technique utilizing additional rocking columns based on the self-centering mechanism. Firstly, analytical models for seismic capacity and displacement of retrofitted bridge piers were developed. Secondly, quasi-static tests were conducted on double-column RC bridge piers with/without additional rocking columns. The cooperative working mechanism between the existing double-column RC bridge piers and additional rocking columns was studied. The seismic damage process, failure modes, and working mechanism of the retrofitted bridge piers were revealed. Finally, finite element models of double-column RC bridge piers with/without additional rocking columns were established based on OpenSees. The influences of key parameters on the seismic performance of retrofitted bridge piers were analyzed, including prestressed tendons and energy dissipation devices of additional rocking columns. Quasi-static test results indicate that the additional rocking column reduces the damage in the plastic hinge region of the existing bridge piers. The retrofitted bridge piers achieve approximately a 71% increase in lateral seismic capacity. When the drift reaches to 5%, about a 50% enhancement in energy dissipation, and a 43.5% reduction in residual displacement compared to existing double-column RC bridge piers. Double-column RC bridge piers with additional rocking columns realize simultaneous improvements of seismic performance and resilience. Finite element analysis shows that design parameters, such as initial prestress, the cross-sectional area of prestressed tendons, and the cross-sectional area of dampers, affect the seismic capacity, energy dissipation, and residual displacement of retrofitted bridge piers. The proposed analytical models accurately predict seismic capacity and displacement of retrofitted bridge piers. This study extends the field of seismic strengthening and retrofitting of existing RC bridge piers, providing a novel method for retrofitting the seismic performance of existing double-column RC bridge piers.

To reveal the influence mechanism of shear deformation on the distortion effect of thin-walled box girders, this study systematically develops an analytical method for box girder distortion considering shear deformation, based on the distribution law of distortion shear flow. First, starting from the geometric equations in elasticity and incorporating the self-balancing condition of cross-sectional stress, a distortion warping displacement mode neglecting the influence of shear deformation is established. Subsequently, based on the stress balance of micro-elements and the self-balancing condition of cross-sectional torque, the warping shear flow distribution function within the cross-section of the box girder is derived. On this basis, by introducing an additional distortion rotation displacement, a reasonable distortion warping displacement function considering gradient variation is constructed, and the corresponding governing differential equations are established. A one-dimensional beam segment finite element formulation for the distortion effect of thin-walled box girders is then developed using the principle of minimum potential energy, with the selection of the Hermite displacement interpolation function. Furthermore, a corresponding beam segment element program system is established. The accuracy and reliability of the method and program system in this paper are verified through experimental values from steel box model beam and PC box girder bridge, as well as numerical solutions obtained from three-dimensional spatial finite element analysis. Numerical results indicate that the calculations presented in this paper align well with experimental values and ANSYS spatial finite element results. For rectangular box girders with uniform wall thickness, shear deformation exhibits a minor impact on their distortion warping stresses and displacements. However, for PC box girder bridges with cantilever slabs, shear deformation not only significantly alters the distribution of distortion warping stresses but also markedly increases the distortion transverse bending moments and distortion angles. Employing the distortion warping displacement function that accounts for shear deformation can effectively improve the accuracy of distortion warping stress calculations. The proposed distortion analysis theory for box girders, which considers the gradient variation of warping displacement, provides a high-precision and practical new method for the refined analysis of box girder distortion effects.

Ultra-High-Performance Concrete (UHPC) has been widely adopted for its exceptional mechanical properties and durability. However, the design-to-prediction cycle is lengthy and experimental costs are high, limiting performance optimization and large-scale engineering deployment. To improve prediction efficiency, this work constructs a machine-learning-based multi-performance prediction framework for rapid prediction and analysis of UHPC properties, including compressive strength (CS), flexural strength (FS), Slump flow (Slf), and porosity (PS). The study first established a dataset spanning CS, FS, Slf, and PS data drawn from published literature and in-house experiments. It included 23, 23, 17, and 20 input variables for CS, FS, Slf, and PS, respectively, with data sizes of 1 059, 537, 129, and 98 samples. Key features were selected using mutual information (MI) scores and correlation analysis. To address small-sample issues, a Conditional Tabular GAN (CTGAN) was used for data augmentation. Additionally, an Isolation Forest (IF) algorithm based on binary search tree principles was used to eliminate outliers, improving data quality and model stability. Building on this, twelve ML models were compared, with hyperparameters tuned via Bayesian optimization combined with cross-validation to identify the optimal predictor. Subsequently, SHAP was employed to reveal the physical contributions of key features to UHPC performance, and the model's generalization ability was validated through new indoor experiments. The results showed that CTGAN-generated data effectively preserve the original multivariate structure, with optimal augmentation ratios of 0.22 for CS, 0.47 for FS, 0.28 for Slf, and 0.25 for PS. The IF algorithm accurately removes outliers, improving data quality. On the four datasets' test sets, XGBoost achieves R² values of 0.964 (CS), 0.920 (FS), 0.946 (Slf), and 0.943 (PS), indicating the best overall performance. MI-score-based feature subset selection significantly enhances model performance, with the best subsets being CS_17, FS_16, Slf_10, and PS_18. Further SHAP analysis revealed the patterns by which different features influence UHPC performance, providing theoretical guidance for material design. Indoor experiments showed that prediction bias is kept within 10%, demonstrating the effectiveness and feasibility of the proposed method for UHPC performance prediction and optimization. The UHPC performance prediction framework presented in this work is expected to shorten the UHPC design cycle and reduce performance-prediction costs.

During the construction of curved sections in super-large cross-section shield tunnels, eccentric jack loads can easily cause joint opening and offset deformation in segment joints. This subsequently induces the degradation of joint waterproofing performance. To address this problem, a complete research framework was established in this study based on a multiple sealing gasket waterproofing system ranging from the “local waterproofing mechanism” to the “global load response”. A fluid-solid coupling analysis model for joint waterproofing performance was set up, and a mechanical analysis model for multi-ring segments under complex construction loads was also established. The study systematically revealed the deformation characteristics of segment joints in super-large shield tunnels caused by eccentric jack loads as well as the degradation pattern of the waterproofing performance of multiple sealing gaskets. The results show that the multiple sealing gasket system exhibits a “gradient barrier and functional synergy” waterproofing mechanism.The waterproofing performance of the three-gasket system is improved by 23.6%~35.3% compared to the double-gasket system, and by 62.5%~76.2% compared to the single-gasket system. The waterproofing performance degradation of the circumferential joint under eccentric loads reaches 21.0%~24.0%. This degradation is most significant when the shield machine adopts an upward attitude. Finally, the degradation level of the waterproofing performance of multiple sealing gaskets in segment joints caused by eccentric jack loads was quantified. A waterproofing safety factor correction method based on the degradation rate of waterproofing performance has been established, and the adjustment values of the corresponding waterproofing safety factors considering the influence of eccentric loads have been clarified. The research can provide theoretical support for the refined design of segment joint waterproofing in shield tunnels with super-large cross-sections under ultra-high water pressure, and provide a scientific basis for shifting the safety factor of joint waterproofing performance from “empirical judgment” to “data-driven”.

Scientifically understanding the seismic response characteristics of tunnels near fault zones is crucial for practical engineering construction. However, current research has neglected the combined effects of active fault dislocation and seismic motion on near-fault tunnels. To address this, this paper proposes a dynamic input method that considers the combined effects of fault dislocation and seismic motion, referred to as “co-seismic and co-dislocation” dynamic input, based on the theoretical relationship between fault dislocation deformation and the distribution ratio of seismic acceleration energy during the fault activation process. A corresponding finite element analysis model of the strata-fault-tunnel structure is established to investigate the dynamic response characteristics of tunnels near fault zones under the coupled effects of seismic motion and fault dislocation. The study clarifies the seismic impact on near-fault tunnel structures, highlighting key factors such as the tunnel's position in hanging or foot walls and the tunnel-fault distance. The results show that, compared to single dislocation loading or single seismic loading, the dynamic response of tunnel structures under the “co-seismic and co-dislocation” effect is more significant. Additionally, the lining on the side closer to the fault exhibits obvious stress concentration, and experiences more complex loading. Due to the hanging wall effect of the fault, the deformation and stress of the tunnel located on the hanging wall are significantly greater than those on the foot wall. As the tunnel-fault distance increases, the deformation and stress of the lining gradually decrease, and the difference in deformation between the tunnels on the hanging and foot walls also gradually diminishes. This study provides a theoretical basis for the seismic design of tunnel structures near active fault zones.

Resilience in shield tunnels refers to their capacity to withstand hazards and rapidly restore functionality. However, a notable gap remains in the quantitative assessment and grading of longitudinal structural resilience. This study proposed a longitudinal resilience evaluation method for shield tunnels under construction disturbance based on “cloud” uncertainty reasoning. Firstly, the uneven settlement coefficient (R) and maximum settlement (S_max) was adopted as performance indicators, and the qualitative rules was established based on the additive model of longitudinal structural damage and repair levels. Secondly, the boundaries of damage and repair levels was quantified according to codes and numerical simulation methods, the qualitative rules were quantified through the forward cloud generator, then the “cloud” reasoning evidence set was established. Subsequently, considering the non-negativity and data preference properties of resilience, a “cloud” generator improved by log-normal distribution was used to quantify resilience levels, and a “cloud” reasoning conclusion set was established. Finally, through uncertainty reasoning, the performance of actual engineering data from the“cloud” reasoning evidence set to the reasoning conclusion set was studied, and the resilience value R_e was calculated. This method was applied to a double-line shield tunnel undercrossing an existing tunnel project in Shanghai. The results show that the resilience values of the existing tunnel induced by tunnel undercrossing disturbances are R_e=0.600 and R_e=0.561, indicating tunnel resilience level 2, which is consistent with the high risk and strong damage caused by large-diameter shield tunnels undercrossing existing tunnels at close proximity in soft soil. Compared with traditional resilience assessment methods, the calculation results of this method differ by approximately 1%, which further verifies the effectiveness of the proposed method. This method achieves resilience evaluation of the longitudinal structure of shield tunnels, effectively integrates qualitative knowledge with quantitative data, and provides decision-making references for shield tunnel resilience improvement.

To systematically investigate the disturbance characteristics of rectangular pipe jacking tunnel construction on sandy strata deformation, a model test platform for earth pressure balanced (EPB) rectangular pipe jacking was developed. This platform consists of a model soil box, a pipe jacking model machine, a grouting system, and a control and monitoring system, enabling precise simulation of key construction processes such as soil cutting, muck discharge, pipe jacking, and lining support. Based on this experimental platform, a model test of rectangular pipe jacking construction in sandy strata was conducted. During the testing process, systematic monitoring was carried out to record stratum deformation and surrounding rock pressure variations induced by the pipe jacking operation, enabling a detailed analysis of the deformation characteristics of the strata and the evolution patterns of surrounding rock pressure caused by rectangular pipe jacking construction. The test results indicate that the surface settlement induced by rectangular pipe jacking tunnel construction exhibits a Gaussian distribution pattern, with significant settlement areas primarily concentrated within 1.0H (where H is the height of the rectangular pipe) on both sides of the tunnel centerline. The surface settlement and settlement trough width are closely correlated with ground loss rate. Specifically, an increase in ground loss rate leads to a significant rise in maximum surface settlement, while concurrently resulting in a corresponding reduction in the settlement trough width. The crown pressure during construction demonstrates an evolution pattern of “initial reduction followed by stabilization,” with the pressure unloading zone mainly distributed within (0.5B+0.25H) on both sides of the tunnel axis (where B is the width of the rectangular pipe). Furthermore, the development of soil settlement demonstrates distinct spatiotemporal evolution characteristics. During the initial construction phase, rapid settlement primarily occurs in the near-field zone (within 0.25H from the crown). As pipe jacking progresses, settlement in the far-field zone (0.75H to 1.25H from the crown) gradually increases. The research findings can provide significant references for ensuring safe construction of rectangular pipe jacking in sandy strata.

This paper proposed Risk Explanation Ability Constructed Technology, an architecture designed to elicit human-like driving risk reasoning capabilities in lightweight pre-trained large language models (LLMs). The method aimed to promote their application in driving risk segment recall and automated analysis of risk causes. This method consisted of a pre-training phase and an iterative optimization phase. In the pre-training phase, a chain-of-thought (CoT) is designed according to the reasoning framework of driving risk, namely risk factor identification, interaction behavior inference, and potential risk determination. A lightweight LLM is then guided using a few-shot learning approach to generate this CoT, enabling it to initially assess driving risk levels and generate corresponding reasoning. In the iterative optimization phase, a guided learning strategy is employed. During the initial optimization stages, a “teacher” model is used to regenerate reasoning for samples with incorrect driving risk level assessments. Correct samples and regenerated samples are collected to conduct supervised fine-tuning. Experiments were conducted using the LLAMA 3-8B model as the base model and Qwen2-72B as the “teacher” model, with 7 000 naturalistic driving segments. The results show that this method improve the risk level assessment accuracy of the lightweight pre-trained model from 0.527 to 0.783. Furthermore, by comparing the similarity between manually constructed risk reasoning and model-generated reasoning, this method improves the ROUGE-L (Recall-Oriented Understudy for Gisting Evaluation-Longest Common Subsequence) metric from 0.517 to 0.616 compared to the baseline model. These results indicate that the proposed method effectively enhances the consistency between the model's reasoning and human risk reasoning. This method provides a feasible approach to automatically analyze the causes of risk, supporting the creation of driver safety profiles and the delivery of targeted safety education.

Route choice modeling in transportation systems faces challenges such as a large number of feasible paths, limited observation data, and high uncertainty in individual preferences. In recent years, the link-based modeling paradigm has avoided the problem of path set construction by modeling route choice as a sequential decision process. Based on different modeling approaches for two core components-link transition utility and link transition probability-several research directions have emerged: knowledge-driven models, like the Recursive Logit (RLogit), offer efficient and analytical calculations but lack accuracy; data-driven models introduce deep neural networks (DNNs) to model probability for higher accuracy but suffer from poor mechanistic explanations and generalization ability, which hinder their application in traffic management; and existing data-knowledge hybrid-driven attempts still face challenges in learning complexity and computational efficiency. To address these issues, this paper proposes a novel data-and-knowledge-driven route choice modeling framework-Deep Robust Recursive Logit (DR²Logit). This method uses DNNs to model link transition utility while maintaining the efficient analytical calculation of link transition probabilities from the traditional RLogit model, thus balancing accuracy, computational efficiency, and mechanistic explanation. To address the robustness challenges of parameter estimation in deep RLogit models, this study proves the necessary and sufficient conditions for the solvability of RLogit, replacing the overly strict conditions used in prior research. This helps precisely determine and expand the feasible parameter domain, and a robust learning algorithm is designed accordingly. The results from virtual and real-world network experiments show that: The robust parameter estimation algorithm based on the sufficient condition reduces the path flow divergence (used as a prediction error metric) by 14% to 98% and 23%, respectively; The DNN-based utility function reduces path flow divergence by 91% to 97% and 71% to 77% compared to the linear utility function; While improving modeling accuracy, DR²Logit maintains high computational efficiency. Experiments on real-world road networks show that parameter estimation takes only 100 to 200 seconds, and predicting 30 000 paths takes just 2.2 seconds.

Driver distraction is one of the major factors leading to road traffic accidents, but existing vision-based deep learning models for distraction detection suffer from significant performance degradation in cross-domain scenarios due to distribution shifts between the source and target domains. To enhance the robustness of distraction detection in real in-vehicle environments, this paper proposes MT-YOLO, an unsupervised domain adaptation-based distraction detection method focusing on driver hand motions. The proposed method adopted YOLOv7 as the backbone network and embedded a lightweight cross-dimensional attention (CDA) module to enhance features in the driver hand regions and suppress background interference; on this basis, a Mean Teacher-based consistency learning framework was constructed, in which class-aware dynamic confidence and IoU thresholds were introduced to generate high-quality pseudo-labels for the target domain; meanwhile, multi-scale instance-level domain classifiers and gradient reversal layers were inserted before the detection head to achieve adversarial alignment of features between the source and target domains. Experimental results show that, on cross-domain object detection tasks, MT-YOLO achieves a 3.7% improvement over the best competing model; on cross-domain distraction detection tasks, introducing CDA alone yields a 24.3% improvement in recognition accuracy, and the performance gap between the full model and a fully supervised model is reduced to 8.8%. At the same time, the proposed framework maintains nearly real-time inference speed on embedded platforms. MT-YOLO thus achieves a favorable balance between cross-domain detection accuracy and deployment efficiency without requiring additional annotations, providing an effective technical solution for building scalable, low-annotation-cost driver distraction monitoring systems.

Max-Pressure (MP) distributed traffic signal control is effective in stabilizing network queue lengths and improving traffic efficiency as well as throughput. However, the traditional MP control method lacks a periodic scheduling mechanism. It activates phases solely based on intersection pressure, which can lead to certain phases being skipped, adversely affecting pedestrian safety and travel experience. Although existing cyclic Max-pressure control ensures all phases are activated, its green time allocation remains primarily static, proportional to pressure, making it difficult to adapt to complex and dynamic urban traffic demands. To address these limitations, this study developed a reinforcement learning-based dynamic Max-pressure traffic signal optimization method (Dynamic Max Pressure Light, DMPLight), aiming to achieve more refined signal control under complex traffic conditions. The proposed method first utilized the Max-pressure control algorithm to dynamically calculate real-time pressure values for each phase, determining the phase activation sequence and ensuring that every phase was activated within the cycle. It then incorporated the Proximal Policy Optimization (PPO) reinforcement learning algorithm to dynamically adjust the green duration for each phase. By continuously acquiring real-time traffic state information from the environment, the method iteratively updated and optimized the control policy until it converged to the optimal strategy and achieved the best traffic performance. Results demonstrated that compared to Fixed-time control, Vehicle-Actuated control, traditional MP control, cyclic MP control, and the reinforcement learning PPO algorithm, the proposed DMPLight method reduced average intersection queue length by 6.58% to 44.7% and average vehicle delay by 10.05% to 66.67% during peak hours. During off-peak hours, it reduced average queue length by 23.86% to 72.91% and average vehicle delay by 14.12% to 73.79%. The proposed dynamic Max-pressure traffic signal optimization method effectively alleviated intersection congestion, significantly reduced vehicle queue lengths and delays, and enhanced overall network throughput. This study provides new insights and technical support for the development of intelligent traffic signal control systems.

To enhance decision-making safety in highway obstacle avoidance and overtaking scenarios, and to address the limitations of existing Deep Reinforcement Learning (DRL) methods, which rely on short-term observations and lack trajectory prediction for surrounding traffic participants. This paper proposes a DRL-based driving decision-making method integrated with trajectory prediction information. First, an interactive trajectory prediction module based on a Spatio-temporal Transformer is constructed, which incorporates a spatial attention mechanism and a temporal convolutional network to extract multi-vehicle interaction features and predict the future trajectories of surrounding vehicles. Combined with these prediction results, a dynamic driving risk field is established to achieve a quantitative evaluation of potential collision risks and long-horizon driving safety. Subsequently, a DRL driving decision-making framework integrated with trajectory prediction is designed. This framework explicitly introduces predicted trajectories into the state space and utilizes long short-term memory networks to extract temporal features for optimizing the Actor-Critic architecture. Concurrently, a safety reward function is constructed based on the dynamic driving risk field, and an interpretable safety constraint mechanism is introduced to further ensure decision-making safety. Finally, experiments are conducted using the CARLA simulation platform and a self-developed Hardware-in-the-Loop testbench. The results demonstrate that the proposed Trust Region Policy Optimization with Trajectory Prediction information (TRPO-P) algorithm improves safety and traffic efficiency by 14.80% and 6.39%, respectively, compared to baseline reinforcement learning algorithms. These findings verify the effectiveness of the proposed method in enhancing vehicle safety and traffic efficiency within complex dynamic highway driving scenarios.

Driver fatigue is one of the leading causes of traffic accidents, with statistics indicating that approximately 20% of accidents worldwide are directly related to driver fatigue, posing a serious threat to road safety. Existing vision-based fatigue detection methods are susceptible to interference from variations in ambient lighting, eye occlusion, and head pose, which can significantly compromise detection accuracy. Although physiological-signal-based methods-such as those utilizing EEG or EMG-offer higher accuracy, they typically require the use of electrodes or other intrusive equipment, leading to discomfort and limiting their practicality in real-world driving environments.This study focuses on drivers within intelligent cockpit environments and addresses two critical challenges in fatigue detection: detection accuracy and intrusiveness. A novel non-intrusive driver fatigue detection method based on steering wheel grip force signals is proposed. Utilizing a flexible thin-film pressure sensor array embedded in the steering wheel, the system enables accurate acquisition of grip force data. Within a driver-in-the-loop simulation platform, EEG and grip force data were collected from 27 participants. EEG power spectral density analysis was employed to establish a fatigue state evaluation benchmark. Combined with subjective rating scales, the grip force data were labeled according to fatigue states.Subsequently, six time-domain features were identified through statistical significance testing as effective indicators for distinguishing between alertness, mild fatigue, and moderate fatigue. On this basis, fatigue detection models were developed and evaluated using one-dimensional convolutional neural networks (1D-CNN) and bidirectional long short-term memory networks (Bi-LSTM), respectively. To further enhance classification performance, a hybrid CNN-BiLSTM model was proposed, integrating the local feature extraction capability of 1D-CNN with the temporal dependency modeling strength of Bi-LSTM.Experimental results demonstrate that the hybrid CNN-BiLSTM model achieved an overall classification accuracy of 96%, outperforming standalone 1D-CNN and Bi-LSTM models by 22% and 5%, respectively. The proposed method enables accurate identification of multiple fatigue states and offers a promising technical solution for driver fatigue detection in intelligent cockpit applications.

The path tracking control faces issues such as the dynamics coupling, external disturbances and multi-actuator cooperative conflicts. The wheel-side dual-motor rear-drive vehicles provide a physical basis for solving the above problems through the cooperative control of the differential steering and active front-wheel steering. Therefore, this work proposes a compound steering control method for the wheel-side rear-drive vehicles based on dynamic game theory. Firstly, based on the Nash game theory, a non-cooperative dynamic game model is established between the wheel-side rear-drive system and active steering system. The coordinated optimization between the two intelligent agents is achieved by solving the globally optimal control solutions. Secondly, within the compound steering game-theoretic control framework, the bounded disturbances such as the road surface friction uncertainty are considered. The finite time-domain robust controller is designed based on the dynamic zero-sum game theory. The optimal anti-disturbance strategy of the system under the disturbance conditions is derived using the minimax optimization method, and the stability of the closed-loop system is proved. The hardware-in-the-loop simulation results demonstrate the effectiveness of the proposed controller. In the test scenario 1, compared with the traditional Nash game controller designed using numerical iterative strategy, the dynamic game-based compound steering controller can reduce the mean values and root mean square values of the lateral offset by 42.98% and 42.99%, respectively. In the test scenario 2, compared with the traditional Nash game controller designed using numerical iterative strategy, the dynamic game-based compound steering controller can reduce the mean values and root mean square values of the lateral offset by 75.50% and 76.13%, respectively. The path tracking accuracy of the wheel-side rear-drive vehicle is significantly improved. By integrating the Nash game-based cooperative optimization strategy with zero-sum game-based robust optimal control, the highly robust compound steering control scheme is developed for wheel-side rear-drive vehicles.

Trackless haulage is the primary transportation mode in underground mines, but it involves high driving risks and costly collision accidents. Before autonomous driving technology is fully deployed underground, adopting human-machine integration driving methods is crucial for gradually enhancing the adaptability of autonomous systems. This paper proposed a human-machine integrated decision-making optimization method for intelligent mining trucks in narrow underground tunnels. The method aimed to solve the problems of joint decision-control optimization and human-machine conflict resolution. First, for autonomous decision-making control of intelligent mining trucks, a driving operation inversion model was constructed. This model directly mapped steering inputs to future swept areas. A unified tunnel-keeping method was proposed by matching real-time tunnel boundary perception with inversion results to guide control decisions. Then, the human-machine integration driving decision system was modeled as a dual potential well system. A harmonic potential well model was introduced and integrated with an impedance interaction mechanism to form a fused potential well. The dual potential well model was derived and analyzed. A parameter initialization method for the potential well was developed based on human driving data. A quantitative human–machine intention conflict model was established, incorporating steering angle deviation and deviation velocity. Based on dynamic-window conflict recognition, a driver-in-the-loop optimization mechanism for machine potential well parameters was established. Finally, human-in-the-loop comparative experiments were carried out on a test platform built using real underground mining scenario data. The results showed that, in narrow curved tunnel scenarios, the proposed dual potential well optimization (DPW-O) method reduced human-machine conflicts, primary steering torque, and time spent in collision risk zones by 40.97%, 21.2%, and 36.90%, respectively. In long straight tunnel scenarios, human-machine conflicts and primary steering torque decreased by 40.82% and 47.56%, respectively. These results verified the feasibility of the DPW-O method and its ability to optimize machine potential well parameters through human impedance feedback under human-machine integration.