Computer-Aided Civil and Infrastructure Engineering

Fire safety is crucial for ensuring the stability of building structures, yet evaluating whether a structure meets fire safety requirements is challenging. Fires can originate at any point within a structure, and simulating every potential fire scenario is both expensive and time‐consuming. To address this challenge, we propose the concept of the most fire‐sensitive point (MFSP) and an efficient machine learning framework for its identification. The MFSP is defined as the location at which a fire, if initiated, would cause the most severe detrimental impact on the building's stability, effectively representing the worst‐case fire scenario. In our framework, a graph neural network serves as an efficient and differentiable agent for conventional finite element analysis simulators by predicting the maximum interstory drift ratio under fire, which then guides the training and evaluation of the MFSP predictor. Additionally, we enhance our framework with a novel edge update mechanism and a transfer learning‐based training scheme. Evaluations on a large‐scale simulation dataset demonstrate the good performance of the proposed framework in identifying the MFSP, offering a transformative tool for optimizing fire safety assessments in structural design. All developed datasets and codes are open‐sourced online.

Autonomous robots have the potential to significantly improve the operational efficiency of multi‐robot systems (MRSs) under environment uncertainties. Achieving robust performance in these settings requires effective task allocation and adaptive path‐finding. However, conventional model‐based frameworks often rely on centralized control or global information, making them impractical when communication is intermittent or maps are unavailable. Although recent studies have shown that reinforcement learning (RL)‐based frameworks offer improved performance, problems related to synchronization and adaptability in diverse environments remain unresolved. To address these problems, this study proposes the “RL‐based Task‐Allocation and Path‐Finding under Uncertainty (RL‐TAPU)” framework. This framework incorporates an Action‐Selective Double‐Q‐Learning (ASDQ) algorithm for real‐time task allocation and a Context‐Aware Meta‐Q‐Learning (CA‐MQL) algorithm for adaptive path‐finding. Unlike previous RL‐based frameworks, RL‐TAPU is designed to operate without global maps, uses only local state information, and functions reliably under intermittent and low‐bandwidth communication conditions. The task allocator communicates only minimal information, and the path‐finding component adapts to new environments without the need for complete environmental data. Experimental results show that the RL‐TAPU framework achieves better adaptability and works more efficiently with a shorter total execution time than competitors.

The rise of exascale supercomputing has motivated an increase in high‐fidelity computational fluid dynamics (CFD) simulations. The detail in these simulations, often involving shape‐dependent, time‐variant flow domains and low‐speed, complex, turbulent flows, is essential for fueling innovations in fields like wind, civil, automotive, or aerospace engineering. However, the massive amount of data these simulations produce can overwhelm storage systems and negatively affect conventional data management and postprocessing workflows, including iterative procedures such as design space exploration, optimization, and uncertainty quantification. This study proposes a novel sampling method harnessing the signed distance function (SDF) concept: SDF‐biased flow importance sampling (BiFIS) and implicit compression based on implicit neural network representations for transforming large‐size, shape‐dependent flow fields into reduced‐size shape‐agnostic images. Designed to alleviate the above‐mentioned problems, our approach achieves near‐lossless compression ratios of approximately 17000$\hskip.001pt 17000$:1$\hskip.001pt 1$, reducing the size of a bridge aerodynamics forced‐vibration simulation from roughly 600GB$600 \,\mathrm{GB}$ to about 36MB$36 \,\mathrm{MB}$ while maintaining low reproduction errors, in most cases below 0.5%$0.5 \,\%$, which is unachievable with other sampling approaches. Our approach also allows for real‐time analysis and visualization of these massive simulations and does not involve decompression preprocessing steps that yield full simulation data again. Given that image sampling is a fundamental step for any image‐based flow field prediction model, the proposed BiFIS method can significantly improve the accuracy and efficiency of such models, helping any application that relies on precise flow field predictions. The BiFIS code is available on GitHub.

The dimensional inspection of reinforcing bars at construction sites prior to concrete pouring is essential to ensure structural integrity. However, this process has traditionally relied on manual tape measurements, which are labor‐intensive, unsafe, and prone to human error. To address these limitations, this study introduces a novel method for remotely inspecting the spacings and lengths of reinforcing bars using a single oblique photograph and a new perspective correction technique. This method transforms oblique images into vertical images using constraints based on four specific vectors that must be perpendicular to one another, thereby eliminating the need for the four‐point correspondence required by existing methods. This improvement enhances the practicality of the proposed method. A validation experiment conducted at an apartment construction site yielded a mean absolute error of 5.12 mm in measuring the spacing and length of reinforcing bars, demonstrating field‐level accuracy in compliance with American Concrete Institute 117 and the Gagemaker's Rule from US military standard 120.

Predicting changes in track irregularity after tamping is important for assisting maintenance decisions and improving construction efficiency. To date, most prediction methods lack consideration for the uncertainties related to tamping effects. To fill this gap, a multi‐module prediction interval system composed of feature selection, interval scaling, and intelligent predictor has been constructed. The feature selection module integrates the processes of relevance, redundancy, complementarity, and weighting. The interval scaling module assigns scaling factors to each point in a data‐driven manner, offering great flexibility. Research found that the composite model has significant advantages over traditional models, improving the interval coverage probability by 5.83%–40.62%. It can accurately predict the track relative smoothness after tamping, with the R² between the measurement and the prediction of the 60 m mid‐chord offset reaching 0.95. This model can serve as a reliable and feasible tool for predicting the static irregularity of ballasted tracks after tamping.

Railway track transitions are prone to uneven settlements and track geometry degradation. Traditional monitoring methods are limited in coverage, which highlights the need for novel solutions. This study proposes a method that systematically integrates the high spatial resolution of Persistent Scatterer Interferometric Synthetic Aperture Radar (PS‐InSAR) with the broader coverage of Small Baseline Subset (SBAS). A correction method for abnormal InSAR time series is developed, considering both consecutive phase unwrapping errors as well as outlier displacements. Model parameters are optimized through Monte Carlo analysis embedded with grid search. The proposed PS‐SBAS InSAR processing method is applied to generate the track longitudinal profile of a railway transition section and is compared with track inspection data. The results show: (1) the hybrid PS‐SBAS approach provides higher resolution and robustness for tracking long‐term differential settlement along railway tracks. (2) There is a strong correlation between track longitudinal level and the InSAR‐derived profile in the bridge approaches with high differential settlement rates. (3) InSAR can serve as a complementary method to traditional inspections, capturing the progression of differential settlement and enhancing the understanding of long‐term settlement patterns and their impact on track performance.

Recent advances in deep learning have enabled automated ground‐penetrating radar (GPR) image analysis, particularly through two‐stage models like mask region‐based convolutional neural (Mask R‐CNN) and single‐stage models like you only look once (YOLO), which are two mainstream approaches for object detection and segmentation tasks. Despite their potential, the limited comparative analysis of these methods obscures the optimal model selection for practical field applications in tunnel lining inspection. This study addresses this gap by evaluating the performance of Mask R‐CNN and YOLOv8 for tunnel lining detection and segmentation in GPR images. Both models are trained using the labeled GPR image datasets for tunnel lining and evaluate their prediction accuracy and consistency based on the intersection over union (IoU) metric. The results show that Mask R‐CNN with ResNeXt backbone achieves superior segmentation accuracy with an average IoU of 0.973, while YOLOv8 attains an IoU of 0.894 with higher variability in prediction accuracy and occasional failures in detection. However, YOLOv8 offers faster processing times in terms of training and inference. It appears Mask R‐CNN still excels in accuracy in tunnel GPR lining detection, although recent advancements of the YOLOs often outperform the accuracy of the Mask R‐CNN in a few specific tasks. We also show that ResNeXt‐enhanced Mask R‐CNN further improves the accuracy of the traditional ResNet‐based Mask R‐CNN. The research finding offers useful insights into the trade‐offs between the accuracy, consistency, and computational efficiency of the two mainstream models for the tunnel lining identification task in GPR images. The finding is expected to offer guidance for the future selection and development of optimal deep learning‐based inspection models for practical field applications.

Excavator safety and efficiency are crucial for construction progress. Monitoring their 3D poses is vital but often hampered by resource and accuracy issues with traditional methods. This paper presents a method to reconstruct the 3D poses of excavators using a cost‐effective monocular camera while considering physical constraints. The approach involves two steps: deep learning to identify 2D key points, followed by using excavator kinematic models, coordinate transformation, and camera projection relationships to reconstruct 3D poses with optimization. Experimental results show the method achieves a mean joint position error of 428.58 mm and a mean cylinder length error of 5.12%, outperforming alternative methods. This method can be employed cost‐effectively for safety monitoring and productivity management of excavators on construction sites.

Signal noise removal is an indispensable and critical procedure in obtaining clean pavement texture data for reliable pavement evaluation and management. Nevertheless, the presently established denoising approaches to pavement texture data still rely on traditional techniques that have long struggled with removing noise accurately and consistently. This paper innovatively initiates a one‐dimensional (1D) residual denoising network (R1DNet) to achieve the noise removal of 3D pavement texture data. R1DNet is proposed as a 1D architectural encoder–decoder that considers the unique characteristics of 3D texture data from 3D laser imaging technology. The encoder extracts diverse profile features of input noisy texture data through two favorably developed 1D modular structures: a cascade deep convolutional module and a parallel multi‐scale attention module. The decoder gradually parses the extracted profile features and estimates noise, with which the clean texture data are obtained based on a simple subtraction operation. The architecture of R1DNet is determined to be optimal in both accuracy and efficiency, using a customized performance‐balancing evaluation function. For model development in a supervised manner, a systematic labeling method is specifically developed, which can build the baseline clean texture data from real 0.1 mm noisy 3D texture data. The experimental results show that the proposed R1DNet can effectively eliminate noise and produce clean texture data closely matching the baseline, presenting significant improvements in accuracy and consistency, compared to the traditional denoising methods.

Failure of buried pipelines can result in serious impacts, such as explosions, environmental contamination, and economic losses. Early detection and location of unexpected events is crucial to prevent such events. However, conventional monitoring methods exhibit limited generalization performance under varying environmental and operational conditions. Furthermore, the cross‐correlation‐based time difference of arrival approach, which is widely used for source localization, also lacks the capability to identify anomalous events. This study introduces what is termed as the two‐stream global fusion classifier (TSGFC), a novel multitask deep‐learning model designed to early detection and location of unexpected events in buried pipelines, even under previously unseen conditions. TSGFC combines spatial and temporal features from accelerometer data using a global fusion mechanism, and uniquely performs both event classification and source localization through a unified multitask framework. To ensure generalization across diverse environments, we employed a unique data acquisition strategy that was specifically designed to evaluate the model's performance under domain shift by using training data from controlled experiments and test data from real‐world excavation activities conducted on a completely different pipeline. Our results confirm that TSGFC can identify unexpected excavation activity with 95.45% accuracy and minimal false alarms, even when evaluated on data collected from a completely different buried pipeline under real‐world excavation scenarios unseen during training.

While semi‐autonomous drones are increasingly used for road infrastructure inspection, their insufficient ability to independently handle complex scenarios beyond initial job planning hinders their full potential. To address this, the paper proposes a human–drone collaborative inspection approach leveraging flexible surface electromyography (sEMG) for conveying inspectors' speech guidance to intelligent drones. Specifically, this paper contributes a new data set, sEMG Commands for Piloting Drones (sCPD), and an sEMG‐based Cross‐subject Classification Network (sXCNet), for both command keyword recognition and inspector identification. sXCNet acquires the desired functions and performance through a synergetic effort of sEMG signal processing, spatial‐temporal‐frequency deep feature extraction, and multitasking‐enabled cross‐subject representation learning. The cross‐subject design permits deploying one unified model across all authorized inspectors, eliminating the need for subject‐dependent models tailored to individual users. sXCNet achieves notable classification accuracies of 98.1% on the sCPD data set and 86.1% on the public Ninapro db1 data set, demonstrating strong potential for advancing sEMG‐enabled human–drone collaboration in road infrastructure inspection.

Ground‐penetrating radar (GPR) offers nondestructive subsurface imaging but suffers from a trade‐off between frequency and penetration depth: High frequencies yield better resolution with limited depth, while low frequencies penetrate deeper with reduced detail. This paper introduces a novel frequency enhancement method for GPR images using domain‐adaptive generative adversarial networks. The proposed end‐to‐end framework integrates a Domain Adaptation Module (DAM) and a Frequency Enhancement Module (FEM) to address frequency‐resolution trade‐offs and domain discrepancies. The DAM aligns simulated and real low‐frequency GPR data, enabling effective frequency enhancement by the FEM. Due to inherent differences in signal characteristics between simulated and real‐world GPR data, directly applying models trained on simulated data to real‐world scenarios often results in performance degradation and loss of physical consistency, making domain adaptation essential for bridging this gap. By reducing domain discrepancies and ensuring feature consistency, the framework generates high‐frequency GPR images with enhanced clarity and detail. Extensive experiments show that the method significantly improves image quality, target detection, and localization accuracy, outperforming state‐of‐the‐art approaches and demonstrating strong potential for subsurface imaging applications.

Traffic congestion has substantially induced significant mobility and energy inefficiency. Many research challenges are identified in traffic signal control and management associated with artificial intelligence (AI)‐based models. For example, developing AI‐driven dynamic traffic system models that accurately capture high‐resolution traffic attributes and formulate robust control algorithms for traffic signal optimization is difficult. Additionally, uncertainties in traffic system modeling and control processes can further complicate traffic signal system controllability. To partially address these challenges, this study presents a novel, hybrid neural network model enhanced with a probability density function kernel shaping technique to formulate traffic system dynamics better and improve comprehensive traffic network modeling and control. The numerical experimental tests were conducted, and the results demonstrate that the proposed control approach outperforms the baseline control strategies and reduces overall average delays by 11.64% on average. By leveraging the capabilities of this innovative model, this study aims to address major challenges related to traffic congestion and energy inefficiency toward more effective and adaptable AI‐based traffic control systems.

Although estimating the International Roughness Index (IRI) is crucial, previous studies have faced challenges in addressing IRI prediction for asphalt concrete (AC) overlays on Portland cement concrete (PCC) pavements. This study introduces machine learning to predict the IRI of AC overlays on PCC pavements, focusing on incorporating pre‐overlay treatments to reflect their composite characteristics. These treatments are categorized into concrete pavement restoration (CPR) and fracturing methods. The developed models outperformed conventional approaches by effectively capturing the impact of these pre‐overlay treatments, as evidenced by the distinct differences in their contributions to IRI predictions between the CPR and fracturing methods. Additionally, the types and occurrences of pavement distresses varied depending on the pre‐overlay treatments applied. When separate IRI prediction models were developed for each treatment group, they demonstrated improved performance, compared to the original model that combined all treatments. This demonstrates the significance of individualized modeling based on specific pre‐overlay treatment types.

Near‐fault pulse‐like ground motions can cause severe damage to long‐period engineering structures. A rapid and accurate identification method is essential for seismic design. Deep learning offers a solution by framing pulse‐like motion identification as an image classification task. However, the application of deep learning models faces multiple challenges from data and models for pulse‐like motion classification. This study focuses on suitable input images and model architecture optimization through a comprehensive strategy. The diverse datasets are realized by transferring the original time history into Morlet wavelet time‐frequency diagram, anomaly‐marked velocity time history, Fourier amplitude spectrum and its smoothed diagram, and pixel fusion diagrams. Two types of deep learning models are constructed in the image classification task for these datasets. A convolutional neural network (CNN) is enhanced by integrating the self‐attention mechanism (SAM) to concentrate on local image features. Additionally, a seismic parameter layer is added to this enhanced model to reduce reliance on input data features. Visual Transformers, including Vision Transformer (ViT) and Swin Transformer (SwinT), are adopted in this task as well. The results of the enhanced CNN demonstrate that TF outperforms other images with higher classification accuracy and convergence speed, and dual‐input image presents inferior performance. The accuracy of all input datasets under the constraint of a single‐parameter moment magnitude (Mw) is higher than that under the constraint of rupture distance (Rrup). The accuracy under the two‐parameter constraint of Mw and Rrup is higher than that of the single parameter constraint for all input datasets, in which the accuracy from TF is the highest, and that from dual‐input data is improved. The performance of SwinT is similar to CNN+SAM and better than ViT for single‐input images, in which TF presents the highest accuracy. For dual‐input images, ViT is better than SwinT, and both of them are better than CNN+SAM. In a resource‐limited environment, the enhanced CNN with single‐input TF is the best strategy, and the physical constraint of Mw and Rrup is more effective, especially for the dual‐input images.

The assembly and production of sandwich panels for prefabricated components is crucial for the safety of modular construction. Although computer vision has been widely applied in production quality and safety monitoring, the large‐scale differences among components and numerous background interference factors in sandwich panel prefabricated components pose substantial challenges. Therefore, maintaining the model recognition accuracy remains a big challenge in practical circumstances. This paper presents an instance segmentation model, namely adaptive feature expansion and fusion (AFFS). The proposed model includes a dynamic feature aggregation mechanism and possesses a flattened network architecture, enabling efficient feature processing and precise instance segmentation. Moreover, AFFS supports rapid adaptation to newly added data or component categories by updating only the feature extraction layers. Comprehensive experimental evaluations demonstrate that the proposed AFFS achieves outstanding recognition accuracy (mAP50 reaching 95.8% and mAPmin reaching 99.9%), significantly outperforming several state‐of‐the‐art instance segmentation networks, including You Only Look Once (YOLO), Segmenting Objects by Locations v2 (SOLOv2), and Cascade Mask Region‐based Convolutional Neural Network (Cascade Mask R‐CNN).

The number and proportion of aging infrastructures are increasing, thereby necessitating accurate inspection to ensure safety and structural stability. While computer vision and deep learning have been widely applied to concrete cracks, domain shift issues often result in the poor performance of pretrained models at new sites. To address this, a self‐supervised domain adaptation method using generative artificial intelligence based on inpainting is proposed. This approach generates site‐specific crack images and labels by fine‐tuning Stable Diffusion model with DreamBooth. The resulting data set is then used to train a crack detection neural network using self‐supervised learning. Evaluations across two target domain data sets and eight models show average F1‐score improvements of 25.82% and 17.83%. A comprehensive tunnel ceiling field test further demonstrates the effectiveness of the method. By enhancing real‐world crack detection capabilities, this approach supports better structural safety management.

Safe, efficient, and comfortable autonomous driving is essential for high‐quality transport service in an open road environment. However, most existing driving strategy learning approaches for autonomous driving struggle with varying driving environments, only working properly under certain scenarios. Therefore, this study proposes a novel hierarchical continual reinforcement learning (RL) framework to abstract various driving patterns as skills and support driving strategy adaptation based on vehicle‐cloud collaboration. The proposed framework leverages skill abstracting in the cloud to learn driving skills from massive demonstrations and store them as deep RL models, mitigating catastrophic forgetting and data imbalance for driving strategy adaptation. Connected autonomous vehicles’ (CAVs) driving strategies are sent to the cloud and continually updated by integrating abstracted driving skills and interactions with parallel environments in the cloud. Then, CAVs receive updated driving strategies from the cloud to interact with the real‐time environment. In the experiment, high‐fidelity and stochastic environments are created using real‐world pavement and traffic data. Experimental results showcase the proposed hierarchical continual RL framework exhibits a 34.04% reduction in potentially hazardous events and a 9.04% improvement in vertical comfort, compared to a classical RL baseline, demonstrating superior driving performance and strong generalization capabilities in varying driving environments. Overall, the proposed framework reinvigorates streaming driving data, prevailing motion planning models, and cloud computation resources for life‐long driving strategy learning.

As a fundamental driving behavior, the accurate modeling of car‐following (CF) dynamics is essential for improving traffic flow and advancing autonomous driving technologies. Due to the stochastic nature of CF behaviors, the CF model parameters often exhibit heterogeneity (multimodal trends), distribution uncertainty, and parameter correlations. Most studies have examined correlations among CF model parameters, assuming deterministic marginal distributions, and investigated heterogeneity through driving behavior indicators. However, distribution uncertainty and multimodal trends in CF model parameter characteristics remain insufficiently explored. To address this challenge, this study proposes a driving style–based Bayesian model averaging Copula (DS‐BMAC) framework that simultaneously accounts for heterogeneity, distribution uncertainty, and parameter correlations in CF behavior modeling. Using the intelligent driver model (IDM) as a representative example, its parameters are calibrated using CF trajectory data extracted from the Waymo open motion data set. Based on these calibrated IDM parameters, a multivariate Gaussian mixture model is employed to categorize three distinct driving styles, capturing heterogeneity. Subsequently, a Bayesian model average Copula approach is applied to address distribution uncertainty and parameter correlations. Deterministic and multivehicle ring road simulations were conducted to assess the effectiveness of the proposed DS‐BMAC framework. The results demonstrate that the DS‐BMAC framework provides a precise characterization of CF model parameters and effectively reproduces microscopic CF behaviors compared to other approaches. Additionally, the DS‐BMAC framework offers a realistic representation of traffic flow dynamics. The research findings are valuable for understanding mixed traffic flow dynamics and for developing CF decision‐making models for autonomous vehicles and advanced driver‐assistance systems.

Computer vision‐based ship detection using extensively labeled images is crucial for visual maritime surveillance. However, such data collection is labor‐intensive and time‐demanding, which hinders the practical application of newly built ship inspection systems. Additionally, well‐trained detectors are usually deployed on resource‐constrained edge devices, highlighting the lowered complexity of deep neural networks. This study proposes a simulation‐to‐reality (Sim2Real) domain adaptation framework that alleviates the annotation burden and improves ship detection efficiency by a lightweight adaptive detector. Specifically, a proxy virtual environment is established to generate synthetic images. An automated annotation method is introduced for data labeling, creating a large‐scale synthetic ship detection dataset termed SSDShips. The dataset comprises 4800 images, 23,317 annotated instances, six ship categories, and various scenarios. A novel multi‐level fusion lightweight (MFL) network is developed based on the you only look once version 8 (YOLOv8) framework, referred to as MFL‐YOLOv8. MFL‐YOLOv8 is pre‐trained on the SSDShips and fine‐tuned using both realistic and pseudo‐realistic data through a hybrid transfer learning strategy to minimize cross‐domain discrepancies. The results show that MFL‐YOLOv8 reduces model parameters by 20.5% and giga floating‐point operations per second by 66.0%, while improving detection performance, compared to the vanilla YOLOv8. Sim2Real adaptation boosts the model generalization in practical situations, reaching mean average precision mAP@0.5 and mAP@0.5:0.95 scores of 98.8% and 81.8%, respectively. It also shrinks the size of real‐world labeling by 66.4%, achieving superior detection effectiveness and efficiency, compared to existing ship detection methods within the specific domain. Deployed on the NVIDIA Jetson Orin Nano, the proposed method demonstrates reliable performance in edge‐oriented ship detection. The SSDShips dataset is available at https://github.com/congliaoxueCV/SSDShips .

The prediction of freezing depth and thawing time of unbound pavement layers in cold regions is a critical task in pavement design and management. This study developed long short‐term memory (LSTM)‐based encoder–decoder models to accurately predict freezing depth and thawing time, with air temperature as the sole input variable. The models, which aim to offer a 14‐day advance prediction of the thawing time for effective pavement management, utilized data from the Long‐Term Pavement Performance program's database, provided by the Federal Highway Administration in United States. This database contains extensive records on air temperature and freezing states. The LSTM models were trained using data collected from four regions in North America with severely cold winters (Quebec, Minnesota, Ontario, and Maine) and subsequently validated using data from both severely cold (South Dakota and Vermont) and mild (Idaho and Wyoming) winter regions. During the validation phase, the models demonstrated strong performance in the severely cold regions, with predicted freezing depths deviating from the measured values by only 0.05 to 0.20 m and thawing date predictions differing by just 1 to 3 days. However, in the mild winter regions, the models showed less accuracy, with freezing depth differences ranging from 0.10 to 0.40 m and thawing date delays of 3–6 days. Compared to existing analytical and empirical models, the LSTM prediction models developed in this study provide enhanced convenience while maintaining a satisfactory level of accuracy.

Accurate long‐term deformation prediction is essential to ensure the structural security and ongoing stability of large water‐related concrete structures like ultra‐high arch dams. Traditional statistical regression and shallow machine learning approaches, due to their algorithmic constraints, often fail to comprehensively capture the complex temporal and spatial dependencies inherent in high‐dimensional prototypical monitoring data, thereby limiting their predictive accuracy and robustness. To address these challenges, this study proposes a multi‐point deformation forecasting model that incorporates both spatial and temporal correlations between environmental factors and deformation, utilizing advanced deep learning (DL) techniques. Specifically, we employ a Transformer‐based convolutional long short‐term memory (ConvLSTM) model to capture the spatiotemporal dependencies across numerous temperature and deformation monitoring sequences. Furthermore, the multi‐objective bayesian optimization algorithm is utilized to ascertain the optimal model architecture and hyperparameters, concurrently maximizing the regression coefficient and minimizing the root mean square error (RMSE). The effectiveness of the proposed DL‐based model for high‐arch dam deformation prediction is validated using data from multiple monitoring points of ultra‐high arch dams. Experimental results demonstrate that the TransformerConvLSTM method significantly outperforms other models at five monitoring points. Quantitatively, it consistently achieves lower RMSE and high correlation coefficient values, indicating its superior ability to provide accurate predictions with minimal error.

Real-time prediction of subgrade settlement caused by shield tunneling is crucial in engineering applications. However, data-driven methods are prone to overfitting, while physical methods rely on certain assumptions, making it difficult to select satisfactory parameters. Although there are currently physics-data-driven methods, they typically require extensive iterative calculations with physical models, which makes them unavailable for real-time prediction. This paper introduces a lightweight physics-data-driven method for predicting subgrade settlement caused by shield tunneling. The core concept involves using a single calculation of the physical model to provide a weak constraint. A deep learning network is then designed to capture spatiotemporal correlations based on ConvLSTM. By iteratively incorporating real-time data, the learning of physical constraints is further enhanced. This method combines the predictive power of data-driven method with the reasonable constraints of physical laws, validated a good performance in a practical project. The results demonstrate that this method meets real-time prediction requirements in engineering, achieving an coefficient of determination of 0.980, a root mean square error of 0.22 mm, and a mean absolute error of 0.15 mm. Furthermore, it outperforms both physical and data-driven models and demonstrates good generalization performance. This study provides effective guidance for engineering practices.

Infrastructure facilities require proper maintenance, including diagnosing structural durability and determining appropriate repair methods. Structural analysis is widely used to assess structural conditions, necessitating three‐dimensional (3D) data that accurately reflect the locations of deterioration. Therefore, we investigate a method to generate 3D data of manholes from single‐shot panoramic inspection images, focusing on accurately mapping the condition of wall surfaces, the primary targets of manhole inspections, including their deterioration such as rebar corrosion. However, some areas of the wall are occasionally occluded by internal objects such as cables or adjacent walls, making accurate layout estimation and 3D reconstruction difficult. To address this issue, we propose a model that incorporates 3D data generated from design drawings as additional input, along with a post‐processing method to adjust the estimated layout based on geometry in the drawings. The evaluation results show that our method improves 3D IoU accuracy, especially under occluded conditions. Moreover, with a mapping error of 14.3 cm in the reconstructed 3D data, our approach demonstrates its practical potential for use in the structural analysis of most manholes.

Three‐dimensional (3D) digital design is extensively adopted in the architecture, engineering, consulting, operations, and maintenance (AECOM) industry to enhance collaboration among stakeholders. Although recommendation systems are commonly employed to facilitate purchasing in e‐commerce websites, none involves recommending online products to users from 3D building design models due to dimensional and stylistic discrepancies. This study proposes a vision‐based adaptive cross‐domain online product recommendation method, VacRed, for 3D building design models. First, a cross‐domain approach is proposed to transform design models into e‐commerce images, addressing discrepancies in dimension and style between them. Second, an adaptive mechanism is introduced to solve the issue of image quality instability caused by variations in generator weights during the training process of generative models. Third, a cross‐domain product recommendation system is developed based on deep learning to recommend the top k relevant online products for a given building design product. Finally, experiments were conducted to ascertain the effectiveness of the VacRed method. The experimental results of this method demonstrate its excellent performance, achieving a precision rate ( PR ) of 87.20% and a mean average precision of 83.65%. This study effectively connects two main stages in the AECOM industry, design and purchasing, and two large communities, design and e‐commerce.