Techno Press
Tp_Editing System.E (TES.E)
Login Search
You logged in as...

sss
 
CONTENTS
Volume 38, Number 3, September 2026
 

Abstract
To address the challenges of complex background interference, low signal-to-noise ratio of subtle features, and constrained computational resources of embedded devices in automated bridge defect detection, this study proposes YOLO-FSI, a lightweight detection model. First, to minimize the computational overhead, a backbone architecture centered on a multilevel FHGNet was designed, incorporating a focal modulation mechanism. Second, to mitigate the misinterpretation of small-scale targets, such as slippage and cracks, caused by environmental textures, a slim-neck feature fusion structure was developed. This structure optimizes the gradient flow distribution through group shuffle convolution and channel reshuffling techniques, enhancing the fidelity of the multi-scale feature fusion while reducing the parameter redundancy. The Inner-SIoU loss function is then introduced to impose constraints on the bounding boxes via auxiliary scaling factors, effectively accelerating model convergence and reducing localization uncertainty. Finally, experimental evaluations conducted on a dataset comprising 4,707 images of steel and concrete bridges demonstrated that the model parameters, volume, and computational overhead were reduced by 40%, 36%, and 26%, respectively, satisfying the real-time requirements for bridge inspection.

Key Words
bridge defect detection; feature enhancement; lightweight network; YOLO-FSI

Address
Key Laboratory of Urban Rail Transit Intelligent Operation and Maintenance Technology & Equipment of Zhejiang Province, Zhejiang Normal University, Jinhua 321004, China.

Abstract
This study proposes a novel monocular vision and geometry-deep learning fusion depth estimation framework for real-time bolt-loosening monitoring in steel structures. Firstly, bolt images are captured using a single monocular camera, and the bolt regions are automatically localized through YOLOv8 deep learning detection. Secondly, region focused MiDaS model predicts relative depth, and pinhole camera model with bolt group geometric constraints recovers the bolt group spatial configuration. The fusion of the relative depth and bolt spatial positions enables accurate reconstruction of the 3D point cloud of individual bolts. Thirdly, advanced 3D processing techniques—including representative orientation alignment and best-fitting rectangle computation—are applied to align the reconstructed bolt groups into a unified coordination. Finally, bolt-loosening monitoring is performed by estimating bolt head rotation angles and loosening lengths from the aligned 3D point clouds. Experimental validation under diverse loosening conditions and lighting environments demonstrates that the proposed framework achieves stable, accurate, and hardware-efficient performance.

Key Words
3D monocular vision, 3D point cloud, bolt-loosening monitoring, geometry–deep learning fusion, MiDas deep learning, real-time monitoring, YOLO

Address
(1) Ngoc-Lan Pham:
Faculty of Civil Engineering, Industrial University of Ho Chi Minh City, 12 Nguyen Van Bao, Hanh Thong, Ho Chi Minh City, Viet Nam;
(2) Duc-Duy Ho:
Faculty of Civil Engineering, Ho Chi Minh City University of Technology, VNU-HCM, Ho Chi Minh City, Vietnam;
(3) Thanh-Canh Huynh:
Institute of Research and Development, Duy Tan University, Danang 550000, Vietnam;
(4) Thanh-Canh Huynh:
Faculty of Civil Engineering, Duy Tan University, Danang 550000, Vietnam;
(5) Jeong-Tae Kim:
Department of Ocean Engineering, Pukyong National University, 45 Yongso-ro, Nam-gu, Busan 48513, South Korea.

Abstract
To address the inherent limitations of variational nonlinear chirp mode decomposition (VNCMD), specifically its dependency on preset modal numbers and initial instantaneous frequencies (IFs) as well as its sensitivity to noise, this paper proposes a novel data-driven adaptive variational nonlinear chirp mode decomposition (DDAVNCMD) method. Unlike existing approaches that require manual parameter specification, the DDAVNCMD introduces three key innovations: (i) an autonomous modal number determination strategy based on the modal energy proportion, which eliminates the need for presetting the number of components; (ii) a derivative normalization-based algorithm for automatic initial IF estimation, which overcomes the initialization dependency of conventional VNCMD; and (iii) an iterative time-varying filtering mechanism integrated within the decomposition framework to suppress noise interference. The proposed method is validated through single-component and multi-component analytic signals, structural numerical simulations, as well as cable and cantilever structural experiments. The estimated IF results are compared with those obtained from synchroextracting transform (SET), adaptive chirp mode decomposition (ACMD), and VNCMD. Both numerical and experimental results demonstrate that the DDAVNCMD achieves superior accuracy and robustness in time-frequency analysis and IF estimation.

Key Words
data-driven adaptive variational nonlinear chirp mode decomposition; derivative normalizationbased algorithm; instantaneous frequencies; iterative time-varying filter; variational nonlinear chirp mode decomposition

Address
(1) Ping-Ping Yuan, Chen Yang, Zhou-Jie Zhao, Guo-Bao Zhou:
School of Civil Engineering and Architecture, Jiangsu University of Science and Technology, Zhenjiang, 212100, Jiangsu, China;
(2) Wei-Xin Ren:
College of Civil and Transportation Engineering, Shenzhen University, Shenzhen, 518060, Guangdong, China;
(3) Ning Xu:
China MCC20 Group Corp. Ltd., Shanghai, 201999, China.

Abstract
Probabilistic prediction of structural responses is essential for the condition assessment and maintenance of cable-stayed bridges under varying environmental conditions. Although various data-driven approaches, including Gaussian process regression (GPR), have been widely used for this purpose because of their capability to provide both predictions and associated uncertainties, most existing approaches construct independent prediction models for individual cables without explicitly considering correlations among multiple response variables. However, cable systems in cable-stayed bridges behave as integrated structural systems, in which adjacent cable responses are correlated because they are influenced by shared environmental and operational conditions. To address this limitation, this study proposes a probabilistic prediction framework for multiple structural responses using multi-input multioutput Gaussian process regression (MIMO GPR). The proposed framework enables the simultaneous prediction of multiple cable tension forces within a unified model while accounting for correlations among the responses. Environmental variables, including temperature and wind conditions, are used as input features, and a feature selection method based on the Z-index is extended to a multi-output setting to determine a common set of relevant inputs for the correlated responses. The predictive performance of the proposed framework is evaluated using two examples. First, a synthetic dataset generated from a finite element (FE) model is used to numerically validate the proposed method. Second, measurement data obtained from the Second Jin-do Bridge, an actual cable-stayed bridge in the Republic of Korea, are used to compare its performance with that of a conventional multi-input single-output GPR model. To improve the methodological reproducibility of the study, an appendix providing implementation and data-availability details has been added, and the FE model and source code for the proposed framework have been shared through public repositories. The results from the two examples show that the proposed method generally improves prediction accuracy while substantially reducing computational cost. In particular, the proposed approach reduces the computational time by approximately seven times compared with the conventional approach. These demonstrate that jointly considering correlations among structural responses and consistently selecting input features can enhance the accuracy and computational efficiency of probabilistic prediction for cable systems in cable-stayed bridges.

Key Words
cable-stayed bridges; measurement data; multi-input multi-output Gaussian process regression; probabilistic prediction

Address
(1) Minsun Kim:
Track & Civil Infrastructure Division, Korea Railroad Research Institute, 176, Cheoldobangmulgwan-ro,16105, Uiwang-si, Gyeonggi-do, Republic of Korea;
(2) Kyoung-Chan Lee:
Department of Civil and Railroad Engineering, Pai Chai University, Baejaero 15540, 35345, Daejeon, Republic of Korea;
(3) Lee-Sak An:
Department of Smart Infrastructure Engineering, Kongju National University, Cheonan-daero 1223-24, 31080, Cheonan-si, Chungcheongnam-do, Republic of Korea;
(4) Jaebeom Lee:
Division of Physical Metrology, Korea Research Institute of Standards and Science, Gajeong-ro 267, 34113, Daejeon, Republic of Korea;
(5) Jaebeom Lee:
Precision Measurement Major, University of Science and Technology, Gajeong-ro 217, 34113, Daejeon, Republic of Korea;
(6) Young-Joo Lee:
Department of Civil, Urban, Earth, and Environmental Engineering, Ulsan National Institute of Science and Technology (UNIST), UNIST-gil 50, 44919, Ulsan, Republic of Korea.

Abstract
Wave-dispersion analysis of nanomodified cementitious plates requires a consistent link between material state, shear-deformable wave mechanics, data-driven representation, and digital structural information. This study develops an artificial-intelligence-enhanced computational framework for hybrid nano-silica (NS)- and graphenenanoplatelet (GNP)-modified concrete plates with explicit nanoparticle agglomeration descriptors. The nanoscale contribution is represented by a bounded, efficiency-weighted effective-property parameterization, and the structuralscale dispersion problem is solved with a reproducible first-order shear-deformable Mindlin plate formulation that retains transverse shear and rotary inertia. A 6000-point synthetic database is generated from the stated mechanics model. Because flexural phase velocity at fixed dimensionless wavenumber kh is scale-invariant in the homogeneous Mindlin model, thickness is used for dimensional frequency calculation but is not supplied as an independent phasevelocity predictor; finite-plate slenderness is likewise excluded. Multilayer perceptron (MLP), random-forest, and gradient-boosting regressors are trained with a fixed random split and explicitly reported hyperparameters. On the independent test set, gradient boosting achieved R2 = 0.9988 and RMSE = 19.5 m/s, while the MLP achieved R2 = 0.9987 and RMSE = 20.2 m/s. For the reference configuration at kh = 3, increasing the agglomerated-region parameter alpha from 0.05 to 0.55 reduced the predicted phase velocity from approximately 2006 to 1939 m/s. A software-neutral BIM-linked property schema is finally defined to associate plate identity, geometry, nanomaterial descriptors, analysis settings, effective properties, and model outputs. The reported numerical values are computational results of the stated analytical model rather than experimental measurements; the AI and BIM layers are presented as reproducible workflow components for extension to higher-fidelity calibrated models.

Key Words
artificial intelligence; BIM; graphene nanoplatelets; hybrid nanocomposite; nanoparticle agglomeration; nano-silica; smart concrete; structural modeling; wave dispersion

Address
(1) Yajie Geng, Siqi Zhang:
Civil Engineering and Transportation Engineering, Yellow River Conservancy Technical University,
Kaifeng 475004, Henan, China;
(2) H. Elhosiny Ali:
Central Labs, King Khalid University, AlQura'a, P.O. Box 960, Abha, Saudi Arabia;
(3) Rania M. Ghoniem:
Department of Information Technology, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, P.O. Box 84428, Riyadh 11671, Saudi Arabia;
(4) Ayoub Alsarhan:
Department of Data Science and Artificial Intelligence, Hourani Center for Applied Scientific Research, Al-Ahliyya Amman University, Amman, Jordan;
(5) Firas Daghistani:
Department of Civil Engineering, College of Engineering, University of Business and Technology, Jeddah 21361, Saudi Arabia;
(6) José Escorcia-Gutierrez:
Department of Computational Science and Electronics, Universidad de la Costa, CUC, Barranquilla 080002, Colombia.

Techno-Press: Publishers of international journals and conference proceedings.       Copyright © 2026 Techno-Press ALL RIGHTS RESERVED.
P.O. Box 33, Yuseong, Daejeon 34186 Korea.
General Inquiries: info@techno-press.com / Journal Administration: admin@techno-press.com