Fiber-reinforced polymer (FRP) can substitute for steel bars to improve the durability problem of reinforced concrete (RC) beams attributed to corrosion. But A high-precision and interpretable prediction method for the flexural strength of FRP-RC beams has not yet been constructed. This study proposed a genetic algorithm optimized artificial neural network (GA-ANN) model to predict the flexural strength of FRP-RC beams. A database of 166 samples was established to train and validate the model. The input parameters include the FRP reinforcement area, FRP ultimate tensile strength, FRP type, elastic modulus of FRP, concrete compressive strength, beam width, and beam depth. The prediction accuracy and practicability of the GA-ANN model were assessed by comparison with other machine learning (ML) models and design guidelines. A parametric sensitivity analysis was performed based on the proposed model. Finally, the SHapley Additive exPlanation (SHAP) was introduced to investigate the intrinsic mechanisms and the parameter contribution of the ML prediction. The results revealed the GA-ANN model achieves superior prediction performance, with a coefficient of determination (R2) on the validation set of 0.992, which is 1.74% to 6.43% higher than that of other models. Moreover, the trends of flexural strength with the input parameters can be well captured, which is highly consistent with the design guidelines. Interpretability analysis shows that the beam depth and the FRP reinforcement area are the dominant factors affecting flexural strength. This study provides reliable support for the accurate prediction of flexural strength and effective reference for engineering applications.

The corrosion behavior of HRB400 steel in simulated concrete pore solutions (SCPS) containing different chloride concentrations and pH values was systematically investigated through electrochemical measurements, surface characterization, and density functional theory (DFT) calculations. The results demonstrate that the composition and compactness of the passive film are strongly dependent on solution pH. At pH 13.5 and 10.5, the passive film is dominated by Fe2O3, whereas a transition toward a Fe3O4-rich passive layer occurs at pH 9.5, resulting in improved compactness and enhanced corrosion resistance despite the reduced alkalinity. A non-monotonic corrosion dependence on chloride concentration was identified at pH 10.5, where the most severe corrosion occurred at an intermediate Cl concentration of 0.35 M. Electrochemical impedance spectroscopy and MottSchottky analysis revealed that this critical chloride concentration produced the highest defect density within the passive film, thereby facilitating chloride penetration and localized film breakdown. At higher chloride concentrations, the accumulation of Fe3O4 and FeOOH corrosion products partially suppressed corrosion by forming a secondary protective barrier. DFT calculations further demonstrated that Cl exhibits stronger adsorption and greater electronic perturbation toward Fe2O3 than Fe3O4, explaining the inferior corrosion resistance of Fe2O3-dominated passive films. The corrosion mechanism is therefore governed by the synergistic interaction among passive film composition, defect density, chloride activity, and repassivation behavior. These findings provide mechanistic insight into chloride-induced depassivation of reinforcing steel in concrete environments.

Artificial Intelligence (AI) is affecting structural engineering, particularly in data analysis and safety enhancement. In this context, structural health monitoring (SHM) enables condition-based maintenance and extended service life of civil structures. However, vibration-based, data-driven damage detection is often compromised by sensor faults, whose effects on measured signals can closely resemble those induced by structural damage. This similarity can significantly impair damage detection performance, leading to unreliable decisions and inefficient maintenance. To address this challenge, this study proposes a novel AI-powered framework that robustly distinguishes anomalies caused by structural damage from those caused by sensor faults. The framework leverages meta-learning, which is well-suited for data-scarce scenarios typical of real-world SHM deployments. First, an autoencoder trained on healthy-state data extracts damage-sensitive features via reconstruction errors. Second, multiple time-series similarity measures characterize sensor fault-sensitive behaviors. These complementary features are then integrated into a deep neural network classifier. Using model-agnostic meta-learning, the classifier’s initial parameters are optimized to enable rapid adaptation to new monitoring tasks with only a few labeled samples. By transferring prior knowledge, the proposed approach reduces both false positives and false negatives, thereby improving overall detection reliability. It also reduces data and computational requirements, supporting long-term, resource-efficient monitoring. The framework is validated using numerical simulations of a shear-type building and experimental tests on a frame structure, achieving average classification accuracies of 97.1% and 98.7%, respectively.

Accurate modeling of retraction-slip of multiple strands in pretensioned concrete structures is challenging due to stringent meshing requirements. This study proposes a meso-scale numerical model that explicitly incorporates the helical geometry of strands and heterogeneous concrete to investigate retraction-slip. A novel prestress simulation approach is established to reproduce the tensioning-casting-release sequence in a physically consistent manner. A meshing strategy combining global sizing with a non-standard element-growth technique is implemented to ensure numerical stability under narrow spacing and stress localization. The proposed model is validated with experimental results. Parametric analyses indicate that, compared with monostrand configuration, twin strands induce inter-strand interaction, which alters stress redistribution in the cut-end and promotes localized damage at the strand-concrete interface. Increasing the concrete strength from 30 MPa to 50 MPa reduces the transfer length by 59%. Retraction-slip increases overall with decreasing strand spacing and increasing initial prestress. The cut-slip in twin-strand configurations increase by 80% as the eccentricity increases from 30 mm to 50 mm.

Structural health monitoring (SHM) systems generate massive volumes of dynamic response data, posing significant challenges for real-time transmission, storage, and analysis. To address these constraints, this paper proposes a fusion attention-enhanced deep convolutional generative adversarial network (DCGAN) for the compressive sensing (CS) of structural response in SHM. The proposed framework integrates a symmetric deep convolutional generator adapted from the U-Net architecture and a standard convolutional neural network discriminator, which can directly learn optimal sparse representations from raw vibration data, thereby eliminating the need for predefined transformation bases. A fusion attention-enhanced encoder-decoder architecture is designed to preserve multi-scale features under high compression ratios and enhance reconstruction fidelity. Furthermore, inter-layer self-attention and channel attention mechanisms are incorporated to stabilize the adversarial training and improve the quality of signal generation. The proposed method is validated using the Qatar Grandstand benchmark dataset and field monitoring data from a cable-stayed bridge. The results demonstrate that the proposed method effectively stabilizes the convergence process of the loss function. The values of the modal assurance criterion between the reconstructed and original responses consistently exceed 0.99, confirming robust fitting agreement across varying compression ratios.

Hot-cast anchorage is a critical force-transmitting device in cables, which transfers loads through the interfacial bond between steel wires and casting filler. However, the progressive debonding mechanism of anchorage is still not clear due to the invisibility in encapsulated cables. This paper proposes an analytical framework based on a trilinear cohesive zone model (CZM) to accurately reveal the full-range bond-slip behavior and damage evolution at the interface between steel wire and casting filler. The closed-form solutions for load-displacement response, interfacial slip, and shear stress distribution are derived for both long and short bond lengths. A series of pull-out tests are carried out to validate the reliability and precision of the proposed model. The result shows that the proposed model has high fidelity, with predicted ultimate loads deviating from experimental measurements by less than 6.1%. All test specimens exhibited a short-bond-length failure mode. The back-calibrating key control parameters revealed the distributions of the interfacial slip and shear stress throughout the pull-out process. The proposed model provides an effective tool for characterizing the dynamic damage evolution of the bond interface in hot-cast anchorages.

Steel box girder bridges are susceptible to widespread fatigue damage (WFD) under long-term heavy traffic load. Fatigue assessment methods based on single crack initiation and propagation models are inadequate for quantifying structural safety under multi-crack interaction. This study proposes a methodology for probabilistic fatigue life prediction and system reliability assessment of steel box girder bridges with WFD. A stochastic multi-crack initiation and propagation model are developed and validated for finite-width perforated plates with multiple site damage (MSD). A failure criterion considering multi-crack interaction and coalescence is established. Following this, the diaphragm with arc-shaped cutouts is idealized as a finite-width perforated plate, and finite element models are developed to obtain the stress distribution. The crack initiation, crack growth, and fatigue failure life of diaphragms with MSD are predicted and validated against experimental datasets from different test configurations in the open literatures. A fatigue failure probability model for steel box girder bridges with WFD is established using parallel system theory to evaluate the fatigue life distribution and system reliability. The results show that the proposed method effectively captures the evolution from crack initiation to fatigue failure in MSD structures and provides reliable fatigue life predictions. Using the fatigue life of a single diaphragm with MSD to represent the service life of the steel box girder bridge with WFD results in conservative fatigue safety assessments. The proposed system level reliability analysis provides a quantitative basis for maintenance decision-making of steel box girder bridges.

The multiobjective optimization of the maintenance strategy for bridge networks becomes more complex with the increasing environmental problems. A maintenance strategy model of bridge networks considering the coordination of cost, reliability, and carbon emissions is proposed to reduce the environmental impact during the bridge maintenance. The model comprehensively considers the connectivity reliability of bridge networks, reliability of key bridges, and comprehensive cost and further considers the influence of carbon emissions under the time effect. At the same time, the improved nondominated sorting genetic algorithm III (NSGA-III) algorithm is first proposed to optimize the maintenance strategy. To enhance the search capability of the algorithm and avoid local optima, the rotation transform operator and axial transform operator are added in the selection operation stage, while a dynamic cross-mutation mechanism is also introduced. Case study shows that the improved NSGA-III, compared with the original NSGA-III, can obtain nondominated solutions with higher diversity and the stronger convergence. Therefore, the improved NSGA-III can provide a better solution to the maintenance strategy for bridge networks, i.e., obtaining higher connectivity reliability and critical bridge reliability under the condition of lower maintenance costs and CO2 emissions.

In engineering practice, fatigue failure is the primary failure mode of bridge steel, which is a brittle fracture without obvious symptoms. This work characterizes the microstructure of Q370qD steel before and after fatigue testing using a series of characterization techniques, including optical microscopy (OM), scanning electron microscopy (SEM), energy-dispersive X-ray spectroscopy (EDS), and electron back-scatter diffraction (EBSD). The fatigue fracture mechanism of Q370qD steel at different stress levels is further investigated. Results indicate that the fatigue strength of the S-N curve (97.7% guarantee rate) at N=2 & times;106 is 245 MPa for naturally linear fitting and 197 MPa for fixed-slope linear fitting. The initial microstructure is characterized by uniformly distributed Fe-C dendritic eutectic phases with the dendrite spacing of approximately 10.63 mu m. The microstructure analysis near the fracture surface indicates that grain size and the percentage of low-angle grain boundaries significantly affect fatigue performance. The deformation band can drastically reduce the energy needed for crack initiation and provide a low-resistance path for crack propagation. Small-sized grains can hinder dislocation motion and delay crack initiation. In addition, crack initiation is dominant in high-cycle fatigue. Average kernel average misorientation (Ave. KAM) is a statistical representation of the overall dislocation structure and has a weak effect on fatigue performance. Consequently, the correlation between microfactors affecting fatigue performance can be expressed as deformation band > grain size > Ave. KAM. As the stress level increases, the {100} texture evolves from two pole density peaks, which deviate 45 degrees from normal direction (ND) toward transverse direction (TD), to three pole density peaks extending along the loading direction. For the {111} texture, the pole density peaks parallel to ND gradually dismiss, and the other two pole density peaks that deviate 45 degrees from ND to TD are retained. This study provides strong data support for the high-cycle fatigue behavior and fracture mechanism of bridge steel from the perspective of microstructure.

Local fatigue cracks in steel box girders can damage structural integrity under heavy-traffic loads. This study proposes a collaborative local-global safety evaluation method by integrating strain ratio theory with a genetic algorithm-based support vector machine (GA-SVM) model. A strain ratio calculation model based on bridge influence line theory is developed to quantify stiffness changes of bridge after local fatigue crack damage. Wheel load strain responses of each measurement point under three typical fatigue crack damage scenarios are obtained through numerical simulation. The mapping relationship between the strain ratio probability density distribution and the type and location of fatigue cracks is established. A fatigue crack importance coefficient is introduced to quantitatively correlate local crack damage with global structural safety performance. A GA-SVM model is then constructed to detect fatigue crack damage types and evaluate the global safety level. The results indicate that the strain ratio-based crack damage detection method can effectively capture local stiffness variations and accurately identify different types of fatigue crack. The proposed GA-SVM model integrates strain information from multiple measurement points to provide reliable evaluation of structural safety under local crack damage. Based on these findings, an online monitoring framework for safety evaluation of the existing steel bridges is proposed.

This study explores the feasibility of learning-based motion magnification (LMM) for revealing cable motion under micro-vibrations. LMM enhances micro-vibrations without requiring a predefined target frequency band, while producing fewer bilateral artifacts around the cable. Therefore, this paper proposes a novel motion magnification-based method (LMMLT) for rapid measurement of cable vibration displacement under small-amplitude motion. LMMLT combines the LMM algorithm with improved line tracking technology (ILTT). Low image contrast under varying illumination is addressed using an enhanced TA-DeepLabV3plus model, which segments magnified frames and generates binary cable masks. ILTT then extracts subpixel displacements from the binary image sequence. Vibration frequencies are obtained through Fourier analysis of the extracted pixel displacements and used to estimate cable tension. The effectiveness of the LMMLT is verified by laboratory experiments and field tests on a cable-stayed bridge. ILTT reduces the runtime from 2.60 s/frame to 76 ms/frame at 640 & times; 480 pixels. The maximum cable force estimation error of the LMMLT approach is 1.75%. Field tests further demonstrate that the proposed method accurately identifies cable vibration under environmental excitations, with estimated frequencies highly consistent with accelerometer measurements and theoretical values.

An improved sequence segment method is proposed to consider the bond degradation into RC beam deformation failure analysis. A new criterion is developed to distinguish the anchoring behavior of steel rebar and the overall compatible relation within the whole beams is employed to quantify the incompatible strain caused by bond degradation. The accuracy of the proposed model is verified using the published experimental data. The effects of corrosion-induced bond degradation on load-deflection behavior and ultimate load of corroded RC beams are discussed and clarified based on the proposed model. Results show that the proposed model has high accuracy to incorporate the bond degradation and to predict the load-deflection behavior for corroded RC beams. Bond degradation has negligible effects on load-deflection behavior and ultimate load of corroded RC beams as the corrosion loss is below a critical value (3.5%-5.0%). As the corrosion loss exceeds the critical value, bond degradation progressively deteriorates the ultimate load and load-carrying stiffness of RC beams. After then, when the corrosion loss exceeds another critical threshold, however, the bond degradation-induced ultimate load of RC beams stabilizes and no longer changes with further increases of corrosion loss. The load-carrying stiffness exhibits a continuous trend of degrading earlier with increasing corrosion loss.

Crushing waste coral concrete into recycled aggregates to create recycled coral aggregate con-crete(RCAC)contributes to sustainable construction development on offshore islands and reefs.To investigate the impact of recycled coral aggregate on concrete properties,this study performed a comprehensive analysis of the physical properties of recycled coral aggregate and the basic mechanical properties and microstructure of RCAC.The test results indicate that,compared to coral debris,the crushing index of recycled coral aggregate was reduced by 9.4%,while porosity decreased by 33.5%.Furthermore,RCAC retained the early strength char-acteristics of coral concrete,with compressive strength and flexural strength exhibiting a notable increase as the water-cement ratio decreased.Under identical conditions,the compressive strength and flexural strength of RCAC were 12.7%and 2.5%higher than coral concrete's,respectively,with porosity correspondingly reduced from 3.13%to 5.11%.This enhancement could be attributed to the new mortar filling the recycled coral aggre-gate.Scanning electron microscopy(SEM)analysis revealed three distinct interface transition zones within RCAC,with the'new mortar-old mortar'interface identified as the weakest.The above findings provided a reference for the sustainable use of coral concrete in constructing offshore islands.

Evaluation of corrosion’s effect on the development length of prestressing strand is crucial for pretensioned prestressed concrete structures serving in the aggressive environment. This paper presents an analytical model for predicting the development length of corroded prestressing strand, in which the effects of strand corrosion on the transfer length and flexural bond length are considered separately. The effect of corrosion on the initial prestressing transfer length is analyzed by considering the combined effects of initial releasing cracking and subsequent corrosion-induced cracking on the evolution of confinement around the prestressing strand. Thereafter, beyond the transfer stage, the effect of strand corrosion on the additional flexural bond length is confirmed based on the calculation of maximum bond strength for various corrosion degrees, in which the effect of strand rotation along its axis is considered for the pull-out failure mode in a well-confined condition. Comparison of analytical and experimental results shows that this model can reasonably predict the development length of corroded prestressing strand.

Health monitoring data from fiber-reinforced polymer (FRP) reinforced concrete structures contain valuable information related to structural performance. However, the acquisition of long-term monitoring data is often constrained by long testing durations, high costs, and complex influencing variables. To address this challenge, this study proposes an improved deep convolutional generative adversarial network (DCGAN) framework for generating reliable monitoring data and establishing an augmented database. Firstly, long-term static bending tests were conducted on FRP-reinforced concrete beams to obtain initial deflection monitoring data. Secondly, ensemble learning was incorporated into the DCGAN framework to develop an ensemble deep convolutional generative adversarial network (EDCGAN). The proposed generator consists of multiple sub-generators, while the discriminator includes both global and local discriminators to enhance training stability. Thirdly, the images generated by EDCGAN were converted back into deflection time-series data to verify their consistency with the real monitoring data in terms of physical characteristics, thereby establishing a reliable augmented database. Finally, five different deflection prediction models were trained using the initial and augmented databases to evaluate the effectiveness of the proposed model. The results indicate that the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) of images generated by EDCGAN are improved by 16.5% and 4.3%, respectively, while the frechet inception distance (FID) is reduced by 35.8% compared with those of images generated by DCGAN. The generated data exhibit good physical consistency, and the prediction accuracy of the models trained using the augmented database is significantly improved.

Accurate modeling of the interfacial behavior of helical strand is essential for characterizing the transfer of prestress. This study proposes a novel separated modeling strategy that simulates the prestressing process using independent tensioning and release models. The strategy overcomes the incompatibility of conventional prestress transfer models to accommodate the helical structure of strands by decoupling the tensioning and release simulations. The deformed mesh and element stresses are transferred from the tensioning model to the release model to establish the full prestressing process. The transfer bond is generated through the interfacial contact between helical ribs and heterogeneity concrete, without relying on bond elements. The rationality of the proposed model is validated based on existing experiment, and series of parametric studies are conducted. The results indicate that the helical structure of prestressing strands leads to a non-uniform distribution of interfacial stresses and damage during prestress transfer. A shorter strand lay length enhances the effective prestress and reduces the transfer length after release, while simultaneously increasing the risk of concrete splitting at the release end. Increasing the prestress level from 0.4 to 0.75 of the strand tensile strength results in a 1.28-times increase in prestress loss.

A damage-incorporated method for generating reinforcement material (RM) topological configurations in damaged reinforced concrete (RC) structures is introduced in this study. It addresses the issue of computationally expensive nonlinear finite element analysis (FEA) inherent in existing topology optimization methods. First, a dimensionality reduction method for design variables and a damage-incorporated adaptive RM evolution strategy are developed to compress the design space and reduce redundant FEA iterations. Based on this framework, an optimization objective and design sensitivity formulations for strengthening damaged structures are established, and a complete optimization procedure is constructed. Numerical examples demonstrate that the proposed method generates rational RM configurations for damaged RC structures and significantly reduces the computational cost of optimization, with its efficiency advantage becoming particularly evident in large-scale reinforcement problems.

High-fidelity forecasting of structural dynamic responses from sparse sensor networks is critical for ensuring the reliability of structural health monitoring and safety assessments. Existing data forecasting methods, including state-of-the-art Transformer models, have limited ability to simultaneously capture the complex spatial topology of sensor networks and long-range temporal dependencies. To address this challenge, this paper proposes a multi-scale adaptive graph attention network for forecasting structural dynamic responses with high fidelity under data scarcity. First, the framework employs the Fast Fourier Transform to decompose complex time-series signals into distinct, physically meaningful temporal scales. Subsequently, it utilizes an adaptive graph convolutional network to explicitly model the unique spatial correlations between sensors within each identified scale. Finally, a multi-head attention mechanism is applied to capture intricate intra-series temporal patterns and long-range dependencies. The model was validated using acceleration data from the Alamosa Canyon Bridge and demonstrated competitive forecasting performance in long-horizon prediction. More importantly, the study reveals a significant decoupling between linear sensor correlation and predictive utility, finding that sensors with low linear correlation provide critical complementary information for forecasting, while highly correlated sensors often introduce redundancy. In addition, under selected sparse sensing configurations, the model achieves forecasting accuracy comparable to denser sensor settings while reducing computational cost. This study provides empirical support for evaluating accuracy-efficiency trade-offs in sparse-sensor structural response forecasting.

Carbonation is a common and slow process that occurs in cement-based material, resulting in durability degradation of reinforced concrete structures. In this research, the relative velocity change (d v/ v) of both ultrasonic direct waves and coda waves are extracted based on the step-wise stretching method for concrete carbonation monitoring. To this end, two-dimensional mesoscale models are established to investigate the ultrasonic propagation behaviors and assess the effectiveness of direct waves and coda waves on concrete carbonation monitoring. The numerical results indicate that direct waves could evaluate the initial carbonation, but exhibit limited sensitivity for severe carbonation. Conversely, the d v/ v values of coda waves show a linear correlation with carbonation depth in all conditions. Accelerated concrete carbonation experiments are conducted to validate the numerical findings. The experimental results and numerical findings mutually corroborate, validating the effectiveness of velocity changes of coda waves on all-stage concrete carbonation monitoring. This research contributes a novel method for monitoring concrete carbonation and enhances the understanding of ultrasonic wave propagation in concrete.

Ultrasonic guided wave (UGW) technology is an effective method for intelligent monitoring of crack damage in thin-walled plate structures. In this paper, a finite-element (FE) model of UGW propagation in Q345 steel plate with crack damage was established and verified. A visual localization method for crack damage in steel plate based on the time reversal-probability damage imaging (TR-PDI) method is proposed. On this basis, the time-frequency domain characteristics of UGW signal are extracted using continuous wavelet transform. Two deep learning (DL) models, the two-dimensional convolutional neural network (2D-CNN) and DenseNet, are used to quantify the crack length and depth by UGW signal. The robustness and generalization performance of the automatic quantification method of crack damage based on UGW-DL algorithm were verified using three databases. The results show that the TR-PDI method can realize visual localization of crack damage in steel plate. The accuracy of two DL models in quantifying crack length and depth exceeded 92.9% and 99%, respectively, and the DenseNet model outperformed the 2D-CNN model. The fusion database constructed using the experimental results and FE data can significantly improve the robustness and generalization ability of the UGW-DL algorithm in crack diagnosis.