Abstract Postdisaster inspection of concrete structures demands rapid, reliable assessment tools, yet conventional manual surveys are slow, subjective, and impractical under emergency conditions. Although deep learning has advanced crack detection, most prior work is limited to binary segmentation and neglects other critical surface damages such as spalling. In addition, postdisaster environments present two major challenges: (1) severe class imbalance at the pixel level; and (2) constrained computing resources that hinder deployment of computationally intensive models. The objective of this study is to systematically evaluate U-Net variants for multiclass concrete damage segmentation under class imbalance and computing-resource constraints. This study systematically evaluates four U-Net variants, the original U-Net, dilated U-Net, attention U-Net, and spatial and channel squeeze-and-excitation (SCSE) U-Net, alongside a well-established external benchmark, DeepLabv3 with a ResNet-101 backbone. All models were trained, validated, and tested on an annotated dataset of cracks and spalling and assessed using both quantitative metrics and qualitative inspections of segmentation outputs. Computational efficiency, including training time, inference time, and memory usage, was also monitored to gauge practical feasibility for on-site deployment. Results show that all U-Net variants achieved strong segmentation performance with only marginal differences between them. Attention U-Net delivered the most balanced class-specific performance, while the original U-Net combined balanced accuracy with the lightest computational footprint. Compared to the benchmark, U-Net variants achieved 5.9% faster training, 49.8% faster inference, and overall reduced memory usage, underscoring their suitability for practical deployment in postdisaster environments. This study provides a systematic evaluation of U-Net variants for multiclass concrete damage segmentation. It offers empirical insights into the persistent challenge of class imbalance and emphasizes computational efficiency as a core evaluation dimension. These findings lay the groundwork for future frameworks that link automated damage maps with structural models to accelerate disaster reconnaissance.
Abstract Extracting a building information model (BIM) from point cloud data of an existing structure is referred to as scan-to-BIM. It is crucial for effective renovation planning, progress monitoring, and life-cycle asset management. Terrestrial close-range photogrammetry is a cost-effective data acquisition technique for scan-to-BIM, with view planning being a critical preparatory step. However, traditional view planning often relies on iterative, experience-based camera placements, leading to inefficiencies, potential data gaps, and increased project costs until a satisfactory point cloud is acquired. This study systematically optimizes view planning by assessing the object surface photogrammetric information (OSPI) generated from multiview images before field data acquisition. Based on camera-building geometry, the method categorizes potential camera positions into stereo, transition, and convergent zone at the center (CZC). By calculating an OSPI value for points on the building surface, considering constraints such as uniform resolution and convergence angles, the method enables informed decision-making on camera network design. A minimum OSPI threshold helps determine the optimal number and distribution of camera positions. For instance, to achieve a level of detail 1 (LoD1) BIM for a 90° corner, the study found that two images in the CZC, three in each transition zone, and normal images with 90% overlap provided an accurate point cloud. Similarly, for a more detailed LoD2 BIM, the requirements increased to six images in CZC and five additional normal images in stereo zone. The results of the study were validated on a large-scale building (1,033 images) by developing a BIM model from the point cloud data with an accuracy of 2 cm. Factors influencing cost are discussed. This low-cost, predictive view planning (PVP) approach empowers project managers and BIM coordinators to make strategic decisions on camera placement, minimizing rework, optimizing field time, and ensuring the generation of point clouds fit for scan-to-BIM purposes.
Abstract In tunnel boring machine (TBM) operations, a large volume of feedback data related to the excavation process is collected in real time, which directly reflects the interaction between the TBM and surrounding rock. The key challenge in ensuring efficient and safe tunneling is to accurately extract sensitive features from vast, fluctuating, and complex multidimensional time series data to understand the geological conditions ahead of the cutterhead. In this study, a novel intelligent recognition method for TBM rock mass classification is proposed, based on piecewise aggregate approximation (PAA) and a multichannel one-dimensional convolutional neural network (MC-1D-CNN). The method effectively maps the interaction between the TBM and surrounding rock to the associated rock classes information ahead of the cutterhead. Five tunneling parameters are selected as input features. First, the PAA technique is applied to transform the multidimensional, raw TBM time series data, which exhibit different scales, into fixed-length one-dimensional feature vectors. This process reduces dimensionality and fuses the data along spatial channels. Then, a MC-1D-CNN model is constructed, which takes the multichannel PAA sequence matrix as input for training. Through multilayer nonlinear mappings and dynamic optimization, the model automatically learns the features of univariate time series within each channel, further reducing the dimensionality and extracting high-order features more strongly correlated with the rock classes. Additionally, to enhance the feature extraction ability and generalization capability of the model, a MC-1D dilated CNN (MC-1D-DCNN) is developed by incorporating dilated convolutions. The method is validated using TBM monitoring data from the second section of the Yinchao project, demonstrating its capability to accurately classify the rock classes. Comparative analysis with several ensemble machine learning models based on key tunneling parameter time series statistical features shows that the proposed TBM surrounding rock level intelligent method PAA-MC-1D-CNN has better feature mining ability and higher recognition accuracy and robustness.
Abstract This study introduces a physics-informed deep learning framework for modeling the viscoelastic behavior of asphalt binders using a fractional-order Kelvin–Voigt series model embedded within a physics-informed neural network (PINN). By enforcing Kramers–Kronig constraints and integrating depth-frequency-dependent input features, the proposed fractional-order constitutive PINN (FO-PINN) framework captures broadband dynamic modulus behavior with high accuracy (global R 2 > 0.96 ). A unique material embedding strategy enables the network to generalize across different binder types while preserving material-specific characteristics. Sensitivity analysis reveals that long-term viscosities ( η 2 , η 3 ) and the fractional derivative order ( α ) dominate the rheological response, reflecting the multiscale nature of asphalt microstructure. The model demonstrates superior performance compared to conventional identification methods in terms of accuracy, generalizability, and physical consistency. Furthermore, its capability to decode depth-resolved viscoelastic gradients and support cross-material parameter estimation highlights its potential as a foundational tool for digital twins of pavement systems. These findings promote the adoption of physics-constrained machine learning in civil material informatics and infrastructure resilience assessment.
Abstract The tunneling parameters of tunnel boring machines (TBMs) exhibit distinct spatiotemporal characteristics under different rock mass conditions, particularly their non-Euclidean spatial properties. This paper developed a real-time rock mass identification model based on the spatial-temporal graph convolutional network (STGCN) to simultaneously capture the spatiotemporal characteristics of tunneling parameters. TBM tunneling data and geological data were collected from an 8.41-km-long tunneling section of the YinChaoJiLiao (YCJL) project in China. By matching the data based on tunnel mileage, 2,367 valid samples were obtained. A downsampling strategy was applied to reduce data redundancy and computational cost while maintaining model performance. The optimized model achieved an accuracy of 93.04% and an F 1 -score of 89.75%, and comparative experiments verified its superior performance compared with the mainstream model in this task. A statistical analysis of the adjacency matrix derived from the graph convolutional network (GCN) module was conducted and the influence of input time length on model performance was analyzed to interpret the spatial and temporal characteristics of the tunneling parameters, respectively. Furthermore, the model was validated on a extension section of the same tunnel, confirming its accuracy and efficiency for practical application. The results indicate that the STGCN model can simultaneously capture the complex spatiotemporal characteristics of TBM data and provide accurate real-time rock mass identification results, showing great potential for engineering application in intelligent tunneling.
