Recent advances in computer vision have provided new solutions for intelligent welding. However, existing vision-based weld seam extraction techniques exhibit limited adaptability to various workpieces in unstructured environments. Therefore, this paper proposes a three-dimensional vision-based method tailored for weld seam extraction and path generation. The proposed method synergizes a deep learning-based point cloud segmentation technique with an improved multi-scale point cloud registration algorithm to reconstruct the complete point cloud model of all weld regions in the workpieces. Subsequently, the welding paths and torch poses are calculated using an optimized multi-plane fitting algorithm integrated with geometry model of weld seam. Experimental validation on four workpieces demonstrates that the proposed method achieves good accuracy and outperforms the existing techniques in terms of efficiency and applicability, offering a robust solution for automated welding of steel structures.

Bridge weigh-in-motion (B-WIM) utilizes bridges as weighing scales to estimate the axle weights of passing vehicles. Conventional B-WIM relies on a static assumption, obtaining axle weights via least-squares fitting between measured and predicted responses. However, the system of equations may be ill-conditioned, resulting in significant errors in the estimated individual axle weights. To address this limitation, this paper proposes a novel Bayesian B-WIM approach, using the maximum posterior probability to determine axle weights. The posterior probability incorporates prior knowledge about the likely axle weights and the likelihood of the bridge response. These are characterized by the mean and covariance of several factors at both system calibration level, represented by influence line ordinates, and at individual vehicle level, represented by the axle weights. If the bridge response is treated as a sequence of independent variables with identical standard deviations for each time scan, the Bayesian solution reduces to the classical Moses B-WIM algorithm. If each axle weight is independent and identically distributed, the Bayesian formulation reduces to the Regularized B-WIM algorithm. The proposed approach is evaluated using both simulated and field test data. Results are compared against the Moses and Regularization algorithms. For all calculations, the Bayesian approach achieves the highest accuracy, with particularly notable improvements in the identification of individual axle weights. In field tests, the mean absolute errors of the first axle weights are reduced by 21.7% compared with the Moses algorithm and 14.0% compared with the Regularization algorithm. The proposed method can effectively mitigate the effects of fluctuations in axle weights, bridge influence lines, and measurement noise on weighing accuracy.

The traditional finite element method (FEM) incurs prohibitive computational costs in the dynamic analysis of large-scale complex structures (e.g., structures with a large number of degrees of freedom (DOFs)), while purely data-driven deep learning approaches are hindered by issues of interpretability, robustness, and data dependency. To overcome these challenges, this study explores a gray-box modeling approach, the Physics-encoded Time Integrator Graph Network (PeTIGN), which integrates physics-based principles with data-driven learning. The central novelty of PeTIGN lies in three synergistic components: (i) an FEM-consistent graph representation that embeds deformation-compatibility constraints directly into the graph neural network with explicit physical priors; (ii) a time-integration scheme that encodes an enhanced Newmark-beta recursion within the message-passing procedure, ensuring accurate and stable time marching; and (iii) a learnable graph encoder-decoder that compresses large-scale systems into a low-dimensional latent space, slashing memory and runtime overheads. The method is validated on a three-dimensional frame structure with 17,400 DOFs subjected to various seismic excitations, where displacement time-history responses are predicted. Across ten ground-motion scenarios, PeTIGN achieves high accuracy with an average normalized mean squared error (NMSE) of 0.00089 and a speedup of approximately two orders of magnitude over the FEM solver.

Automatic recognition of Railway Sleeper Serial Numbers (RSSNs) is essential for traceability, quality management, and lifecycle maintenance of railway infrastructure. In practice, embossed serial numbers on concrete surfaces exhibit extremely low discernibility with minimal height variations (<1 mm) and negligible color differentiation. Traditional Red-Green-Blue-based (RGB-based) image enhancement and Optical Character Recognition (OCR) methods face a fundamental limitation: they cannot directly capture the three-dimensional geometric features distinguishing embossed characters from their surroundings. To address this challenge, this study proposes an integrated framework based on line laser height imaging and position prior-guided detection with three key innovations: (1) a cascaded processing framework leverages geometric height information to overcome RGB-based method limitations; (2) a Dual-stage Adaptive Image Enhancement (DAIE) strategy converts 16 binary-digit (bit) height images into optimized 8-bit visualizations by systematically selecting optimal methods: modulo truncation for global structure and Minimum-Maximum (Min-Max) normalization for local detail enhancement; and (3) a Position Prior-guided Spatial Attention (PPSA) Feature Pyramid Network (FPN) integrates statistically-derived position priors to enhance small target detection. Comprehensive validation on 2234 images demonstrates superior performance: 98.2% F1-score and 99.38% recognition accuracy at 27 Frames Per Second (FPS), achieving 2.4% improvement over state-of-the-art methods. Ablation experiments confirm the individual contributions of the PPSA module (4.0%), the Small Target Enhancement (STE) module (1.4%), and the DAIE strategy (3.08%). Field testing in a prefabricated factory validates industrial applicability, providing a scalable technical framework and valuable reference for low-discernibility embossed industrial character recognition. Code is publicly available at https://github.com/shipeng38/RSSN-recognition.

Cables are crucial components of cable-supported bridges, and cable force measurement is essential for assessing bridge safety. While the frequency method is widely used to estimate cable forces, it faces challenges in determining boundary conditions and bending stiffness, especially for short cables. Therefore, this study proposes a cable force measurement method based on computer vision (CV) and high-order mode shapes. The broad-band phase-based video motion magnification (BPVMM) technique is first used to magnify small vibrations of the cable, and then the line tracking algorithm is employed to identify cable displacement from the magnified video to obtain high-order mode shapes. Additionally, a theoretical formula for calculating the cable force based on the high-order mode shapes is proposed. The proposed method is verified through laboratory cable tests and field tests of cable-stayed bridges. The results indicate that the cable forces calculated using the proposed method align closely with the measured values, with the maximum error within 3%. This method effectively eliminates the effects of boundary conditions and bending stiffness, significantly improves the accuracy of cable force identification. This study provides technical support for cable force identification when boundary conditions and bending stiffness are unknown.

Finite element method suffers from high computational cost and poor generalization ability, especially in cases requiring repeated calculations under multiple loading conditions, while deep learning-based surrogate model could be a highly efficient alternative that enables fast prediction of structural responses across different conditions. Herein, a mechanics-inspired message-passing network (MMPN) is proposed for generalization computation of bending structures. The model consists of mechanics-inspired network and physics-constrained module, where the structural mechanics and equilibrium equations are encoded into the graph neural network through message passing mechanism. The generalization capability of the model for different load cases and structural forms was validated and further confirmed through a realistic case analysis, which achieves high prediction accuracy with load generalization errors below 3 % and markedly outperforms conventional surrogate models of graph neural network. In addition, a series of factor analyses were conducted to elucidate its scalability and application prospects. Results indicate that the model maintains good generalization performance for large-scale structures, and can achieve considerable prediction accuracy with limited training sample size.

Computer vision offers promising solutions for automated welding. However, existing vision-based methods heavily rely on predefined geometric models tailored to specific welding seam conditions. Consequently, automatic welding of steel box girder components remains highly challenging due to the welding seams' complex and variable geometry. Therefore, this paper proposes an automated welding seam extraction and path planning method using 3D vision. The method integrates four key modules: point cloud segmentation using planar filters, seam extraction by local spatial vector analysis, path planning using nearest-neighbor algorithm, and welding posture planning with smoothing strategy. It was deployed on an actual welding robot for testing. The results demonstrate that the method is capable of automatically extracting and planning paths for various types, including straight, curved, straight-curved and spatially intersecting welding seams. Compared to existing methods, it exhibits greater adaptability and higher extraction accuracy, offering a robust solution for automated welding of steel box girders.

Diffusion phenomena are fundamental in physics, engineering, biology, and etc., yet their simulation via traditional numerical methods is computationally intensive for complex systems. While deep learning offers promise, prevailing approaches like physics-informed neural network (PINN) suffers from limited generalizability and error accumulation, especially with sparse data or irregular geometries. Here, we propose a physics-encoded graph neural network (PeMN) that integrates graph message-passing mechanism with discrete differential geometry principles. By explicitly encoding differential operators such as Laplacian and gradient operators into network architecture for spatial discretization, PeMN reduces nonlinearity in function approximation while enforcing strict adherence to physics. Meanwhile, combining with recurrent temporal modeling it achieves global spatiotemporal optimization, avoiding error propagation inherent in numerical iterative methods. Benchmark study demonstrates very significant improvements in accuracy over other models. Notably, PeMN remains robust and enables generalization across various initial conditions. Our framework establishes a new paradigm for embedding PDEs into network architecture.

The construction industry continues to lag behind other sectors in productivity and safety due to its reliance on manual labor. Robotic automation, especially for rebar tying, faces significant perception challenges stemming from repetitive textures and minimal distinctive features of rebar structures. Existing camera pose estimation methods, including deep learning-based and structure-driven approaches, typically fail in these environments. To address these limitations, this paper proposes an autonomous, feature-aware camera pose estimation framework tailored specifically for rebar structures. The framework introduces methods for structure-aware feature extraction, ideal image generation, structural-constraint-based matching, and closed-loop image evaluation. Laboratory and field tests demonstrate that the proposed method consistently achieves a 100% pose estimation success rate within three attempts, significantly outperforming deep learning-based, structure-driven, and manual methods. This research substantially enhances the reliability and efficiency of rebar tying robots, enabling broader and more effective deployment in real-world construction environments.

Despite significant progress in automation within prefabricated construction, rebar tying remains a labor-intensive task, contributing to production delays and ergonomic risks. Although robotic automation offers potential solutions, existing six degrees of freedom (6-DoF) tying pose estimation methods often struggle with variations in rebar dimensions, which can lead to tying inaccuracies and necessitate rework due to structural non-compliance. To address this problem, a dimension-aware 6-DoF tying pose estimation framework (DimenPose) is proposed in this study. DimenPose integrates rebar dimension measurement into the pose estimation process. First, a coarse-to-fine point cloud reconstruction method is developed to integrate multiple viewpoints to establish a comprehensive three-dimensional (3D) representation. Next, a bidirectional density anomaly detection-based automatic rebar segmentation approach is designed for geometric information extraction from dense and repetitive rebar arrangements. Finally, a geometric analysis approach determines dimensional parameters and tying poses based on spatial and geometric information. By enabling dimension-aware robotic perception through point cloud processing, DimenPose offers a perception basis for improving pose estimation accuracy and supporting dimensional compliance assessment in robotic rebar tying, while its generalizability to broader industrial scenarios requires further validation.

Prestressed concrete railway sleepers (PCRS) are critical railway construction components requiring strict dimensional compliance with industry standards. Traditional manual measurement methods are inefficient and labor-intensive, while existing 3D point cloud reconstruction techniques introduce substantial computational overhead incompatible with real-time production requirements. This paper presents a systematic three-stage feature extraction method based on multimodal line laser scanning (MLLS) image data, which processes height and intensity images to achieve real-time detection of 36 dimensional indices for PCRS. An innovative Dynamic Weighted Least Squares with Rapid Convergence (DWLS-RC) algorithm delivers 64% efficiency improvement over traditional methods. A self-developed automated PCRS dimensional measurement system (APDMS) integrating a high-stability detection platform and line laser sensors enables effective industrial implementation. Experimental and factory tests validate accuracy (<0.3 mm), stability (accuracy's standard deviation < 0.1 mm), and efficiency (8.5 s per PCRS), representing over 60-fold speed enhancement with significantly reduced costs compared to existing automated methods. This work demonstrates a cost-effective, accurate, and scalable engineering solution for automated PCRS measurement, effectively bridging the gap between laboratory research and industrial deployment.

Steel shock absorbers have attracted growing attention for their role in seismic energy dissipation in railway bridges. However, the absence of systematic investigations into their adaptation for various bridge configurations, coupled with an incomplete understanding of the underlying mechanisms of critical design parameters-such as stiffness, yield displacement, and initial gap-has constrained their widespread engineering implementation. To address these challenges, a simply supported girder railway bridge was selected as a representative case. A finite element model was developed using OpenSees and validated through large-scale shaking table experiments. Based on this model, the seismic impact of steel shock absorber design parameters was systematically analyzed for bridges with 8 m and 25 m piers respectively, through which optimal parameter ranges and corresponding configuration schemes were identified. The results demonstrate that: (1) There exists an optimal stiffness range for steel shock absorbers. When stiffness falls within 7000-10000 kN/m, seismic damping efficiency in terms of tensile strain reduction in longitudinal reinforcement can exceed 50%. (2) The initial gap must balance energy dissipation with operational smoothness. (3) Recommended design parameters are as follows: for an 8 m pier, stiffness of 7000-10000 kN/m, yield displacement of 7-10 mm, and initial gap of 20-23 mm; for a 25 m pier, stiffness of 6000-8000 kN/m, yield displacement of 14-16 mm, and initial gap of 13-15 mm. These findings provide practical configuration guidelines for implementing steel shock absorbers in railway bridges with typical spans and varying pier heights, serving as a theoretical reference for performancebased seismic design.

Corrosion will accelerate the degradation of materials in cables and thereby compromise the long-term serviceability of bridges. Most existing studies consider the deepest corrosion pit as both the crack initiation point and the fracture location, and employ a single corrosion parameter, such as maximum/average corrosion depth or total corrosion area, for fatigue life prediction. However, these methods exhibit inherent limitations in capturing the complex interactions between the corrosion distribution, fatigue load, and material properties. Therefore, this study proposes a two-stage deep learning model for comprehensive fatigue performance evaluation of corroded steel wires. In the first stage, the 3D scanning technique is used to obtain surface morphological images, and the training dataset is constructed by combining the finite element simulation data. The Fracture Prediction Pix2Pix (FP-Pix2Pix) model is then developed to predict fracture locations. In the second stage, the physics-informed neural network (PINN) is adopted by integrating prior information, including corrosion, fatigue load, material properties, and fracture locations predicted in the first stage, to predict the residual fatigue life (RFL). Experimental results show that the proposed two-stage model outperforms existing models in both fracture location prediction and RFL prediction with errors less than 5 %. The proposed method provides a robust and efficient approach for evaluating the fatigue performance of corroded steel wires in bridge cables.

In practical engineering, high-resolution (HR) imaging devices have become increasingly utilized for capturing structural surface crack images. However, the effectiveness of current deep learning (DL) segmentation models in accurately predicting refined masks for HR crack images is hindered by the discrete sampling methods inherent in traditional DL architectures and the limited computational resources of GPUs. To tackle this issue, this investigation incorporates the point-based rendering methodology originating from computer graphics disciplines into the encoding-decoding framework, introducing an innovative Crack Boundary Point Rendering Network (CBPRN). The CBPRN endeavors to accomplish elaborate delineation of crack visual samples possessing resolutions surpassing 4K. Initially, an edge feature extractor integrated with a super-resolution encoder is devised to guide rendering heads in efficiently focusing computational power on ambiguous edge regions. Subsequently, a rendering-based prediction head is introduced with the function of efficiently sampling rendering points for the training and inference phases, respectively. Furthermore, a tailored composite objective function is deployed to enhance the learning procedure, enabling the architecture to equilibrium substantial disparities in pixel counts among positive and negative instances within crack visual data. Ultimately, to substantiate the practical applicability of the CBPRN, an on-site crack identification investigation was executed on an actual bridge structure located in Changsha utilizing an unmanned aerial vehicle (UAV). The CBPRN demonstrated remarkable effectiveness on 4K-resolution visual samples acquired by the unmanned aerial vehicle, attaining achieving overlap ratio (Intersection over Union, IoU), average boundary precision (mean Boundary Accuracy, mBA), and Dice similarity index metrics of 85.46%, 86.00%, and 92.16%, correspondingly. This outstanding effectiveness improves both the operational security and processing efficiency of unmanned aerial vehicle-assisted crack assessment procedures, offering enhanced flexibility in choosing flight trajectories for the inspection workflow.

Accurate measurement of bridge vibrations under load excitation is often impeded using traditional computer vision-based measurement (CVBM) methods in complex environments, particularly when small amplitudes and low-texture surfaces are involved. This paper proposes a CVBM method that estimates the full-field dynamic displacement and modal parameters of bridge structures by integrating the swin transformer motion magnification (STMM) with the improved recurrent all-pairs field transform (IRAFT) algorithm. The STMM algorithm could capture high-frequency information about structural vibrations while effectively suppressing noise, blurring, and motion artifacts during magnification. The IRAFT algorithm could improve the calculation accuracy of full-field optical flow in low-texture and large-motion scenes. First, vibrational videos of bridge structures are acquired and calibrated, and then the pixel motion in the video is magnified by the STMM algorithm. Subsequently, the IRAFT algorithm is employed to compute full-field optical flow from the magnified video. Finally, pixel motion within regions of interest is converted to a displacement time-history curve through motion normalization and a scale-factor method, from which modal parameters are identified based on fast Fourier transform and covariance-driven stochastic subspace identification. The proposed method was validated on a synthetic truss bridge, a laboratory experiment of a Q235 simply-supported beam, and on-site inspection of a pedestrian overpass. The results show that the utilization of the STMM algorithm is not only effective in acquiring high-frequency information of structural vibrations but also has the advantage of suppressing motion noise and artifacts during magnification. By integrating the STMM with the IRAFT algorithm, the identification performance of the full-field dynamic displacement and modal parameters of structures is significantly improved, with better robustness to illumination changes and partial occlusion.

High-resolution (HR) imaging is crucial for structural defect detection, yet conventional deep learning models struggle to balance edge segmentation precision with computational efficiency in HR crack image analysis. To address these challenges, a HR Crack Image Rendering Segmentation Network (HRCRSN) is proposed. Three customized improvements were made, enabling the HRCRSN to exploit the advantages of edge-aware rendering technique from the field of computer graphics in the precise segmentation of HR crack images. First, a Transformer-based crack localization module with adaptive multi-scale feature fusion (MSFAWFS) enhances pixel-level guidance while preserving micro-crack details. Second, dynamic point sampling prioritizes ambiguous boundaries and sub-millimeter defects via asymmetric density allocation during training/inference. Third, a synthetic augmentation framework recombines crack objects to address data scarcity. Experiments on UAV-acquired datasets achieve state-of-the-art performance (IoU: 85.36%, mBA: 92.07%, DICE: 91.78%). This approach improves inspection efficiency and establishes a new framework for UAV-based civil infrastructure monitoring.