Computer-Aided Civil and Infrastructure Engineering最新文献

Accurate identification of slab conditions is critical for ensuring the seepage safety of concrete face rockfill dams (CFRDs). However, existing methods for monitoring slab damage are limited and inconvenient. There is an urgent need to utilize available monitoring data to rapidly and accurately assess the condition of slab damage, and to implement responsive decision-making and early warning measures to prevent dam failure. To address this issue, this study proposes a deep learning (DL) methodology for the rapid identification of slab damage. A multi-objective optimization algorithm, Non-dominated Sorting Genetic Algorithm ll (NSGA-II), is employed to carry out inversion analysis and obtain the actual permeability coefficients, which are then used as inputs in numerical seepage simulations of dam behavior after slab damage. Based on the simulation results, a DL model is constructed to establish an accurate mapping relationship between the information of slab damage and the corresponding monitoring data. Given that current DL models often fail to explicitly and effectively capture the importance of features across sequences—leading to redundancy, reduced accuracy, and poor interpretability, a hierarchical optimization structure based on multi-head attention mechanisms is proposed. Specifically, two multi-head attention modules controlling input weights are innovatively integrated into both the input and hidden layers of the DL model, forming a dual multi-head attention enhanced (DMAE) architecture. This structure can be embedded within basic DL models for training and prediction. A case study of the cracked Sanbanxi CFRD project shows that the DMAE-Bi-directional Long Short-Term Memory (BiLSTM) model outperforms other DL models in terms of prediction accuracy and robustness, suggesting it is the most suitable for the identification and prediction of slab damage. Furthermore, the visualization of input attention weights reveals that the key factors in identifying slab damage and should be prioritized in future seepage pressure monitoring. This study fills a critical gap in the field of slab damage identification, provides both technical support and theoretical foundations for intelligent diagnosis and interpretability analysis of slab damage in CFRDs.

Multiple sensors are strategically deployed within concrete dams to monitor structural behavior under intricate environmental conditions. The diverse monitoring parameters, spatial configurations, and temporal variations across these sensors often engender performance conflicts. It is different to obtain the comprehensive dam safety status under the intricate relations and behavior conflicts. To solve the problem, the proposed model integrates anomaly detection, trust propagation, consensus measurement, and information fusion for dam safety assessment. The multi-expert variational autoencoder facilitates anomaly score computations. The social trust network delineates spatiotemporal relationships among sensors. The consensus measurement mitigates information conflicts for data integration using the interval-valued fusion strategy. Empirical validation through a case study involving an arch dam underscores the model's efficacy in identifying anomalies. Through the results analysis, the spatial relationships exhibit divergent attributes in response to changes in water levels. It indicates that the spatial relations are necessary factors in the dam safety assessment.

Vehicle-induced fires present a critical risk to cable-supported bridges, where the integrity of cable components is especially vulnerable. However, conventional monitoring solutions face significant limitations: infrared cameras are often economically prohibitive, and standard smoke detectors are ineffective in open bridge environments. To address these challenges, this paper proposes a multi-stage computer vision framework that utilizes existing visual surveillance infrastructure for real-time fire detection and preliminary cable safety assessment. The proposed system integrates a you only look once v11-m model for accurate vehicle detection, a BoT-SORT tracker with re-identification (Re-ID) capabilities to maintain target consistency through visual obstructions such as smoke, and a ResNet-50 classifier for vehicle-centric fire identification. The framework's novelty lies in the demonstrated synergistic operation of these components across various scenarios, particularly under actual fire conditions. The integration of the Re-ID module proves essential for eliminating false alarms by preserving target identity, while the vehicle-centric approach directly associates fire events with specific vehicles and their tracking identifiers. This linkage provides the fundamental basis for real-time safety evaluation of adjacent cables. Consequently, the framework establishes a cost-effective, readily deployable, and scalable solution for bridge monitoring, offering management authorities a practical tool for immediate fire detection and instant structural assessment.

Retrieval-augmented generation (RAG) enabled learning assistants are promising for engineering education, given their capability to supplement domain-specific knowledge and enhance student support. However, it is also a known problem that RAG demands adequate knowledge bases and can experience unreliable retrieval generation alignment. This study proposes a proactive evaluation framework for RAG-based learning assistants, eliminating the need for student feedback in system evaluation. The framework is demonstrated using a Civil Engineering education tool, CivASK. The evaluation framework identifies the deficiencies in CivASK, including database gap, contextual misunderstanding, and incomplete retrievals, based on the performances under simulated student inquiries, automated retrieval ranking, and expert-validated evaluations. Specifically, 742 student queries are analyzed, and 374 test questions are generated for assessment, showing the practical utility of the proposed evaluation framework for real-world education assist development. The application of the proposed framework is transferable to assist other engineering courses as well.

High-resolution (HR) imaging technology is increasingly employed to capture crack images in civil infrastructure, which is vital for ensuring the safety of the bridge inspection process conducted via unmanned aerial vehicles (UAVs). Such applications require the development of advanced algorithms for the segmentation of HR images. Traditional deep learning-based segmentation methods for inferencing HR images consume considerable GPU resources, which prompts the authors to draw inspiration from the cost-effective rendering technique in computer graphics and try to apply this advanced method to the refined segmentation of HR crack images. However, the original rendering method, designed to guide rendering points by the coarse segmentation masks, often inadequately directs rendering points towards the crucial boundary areas of tiny cracks, leading to unclear or missing boundary predictions. To address this, an innovative rendering technique was proposed, utilizing probability maps to precisely direct rendering points towards crack boundaries and tiny-crack branches during inference. This method enhances the accuracy of crack boundary segmentation and reduces the miss rate of tiny crack branches from HR images, all while conserving computational resources. Through model parameter experiments and ablation studies, the optimal model was obtained, and the effectiveness of the improved components was demonstrated. Furthermore, the field test has confirmed that, equipped with the proposed point rendering technique, the UAV is permitted to effectively perform crack inspection within a 3-m distance from the main beam. Compared to traditional low-resolution semantic segmentation methods, the UAV bridge inspection time is significantly reduced by 50% while maintaining the same accuracy.

Currently, real-time assessment of surface damage to bridges is crucial for ensuring infrastructure safety. Unfortunately, existing methods often present a challenge: overly complex computational models are incompatible with systems that have limited resources, while lightweight models struggle to achieve sufficient detection accuracy. This task is further complicated by the diverse nature of bridge damages, such as cracks, exposed reinforcement, and efflorescence, as well as the challenges of data acquisition under varied conditions from sources like unmanned aerial vehicles and specialized datasets. This work presents an efficient framework developed to improve such applications. The Lightweight Feature Enhancement and Triplet Attention Network for Bridge Damage Detection includes: (1) a multi-scale feature learning module, (2) a slim-neck-based optimized feature pyramid integration module, and (3) a triplet attention-based damage detector module; (1) extracts multi-scale representations of bridge surface features, (2) enhances multi-scale feature integration for lightweight computation, while maintaining accuracy, and (3) optimizes the framework with a three-branch structure for cross-latitude interaction, reducing the importance of irrelevant features. Extensive experiments on the MCDS and CODEBRIM datasets demonstrated its advantages: a $��5.6�\%�$ increase in Mean Average Precision, a $��13.6�\%�$ computational load reduction, and a 45 frames per second real-time performance. The model's computational complexity scales linearly with the input instances processed per unit time during inference.

Road networks face increasing disruptions, yet vulnerability assessment methods either oversimplify traffic dynamics or require extensive computational simulations. This research introduces a novel approach integrating traffic simulation, graph theory, and machine learning for efficient and accurate vulnerability assessment. Analysis across numerous disruption scenarios showed that static weighting is inadequate for capturing traffic redistribution effects. In contrast, dynamic weighting aligns strongly with simulation results but was computationally infeasible. To overcome this limitation, a specialized multilayer perceptron artificial neural network (ANN) model was developed with a dual-pathway architecture and a novel automated closure propagation algorithm, separating static network attributes from spatial relationships. This surrogate model generates predictions significantly faster than traffic simulations, enabling comprehensive vulnerability analyses, previously deemed impractical. Testing across diverse disruption scales demonstrated surrogate effectiveness and limitations. This research presents a transferable and scalable methodology for constructing simulation-informed ANN surrogate models, providing practical deployment guidance for informed resilient transportation network planning.

To overcome the limitations of conventional single-factor analysis, this study proposed a framework for investigating interaction effects of influencing factors on the resilient modulus (M_r) of stabilized aggregate base. First, cross-validation was utilized to compare the predictive accuracy and generalization capability of gradient boosting (GB) and random forest (RF) in predicting the M_r. The grid search algorithm was used to optimize hyperparameters. After optimization, the coefficient of determination for GB reached 0.99 on the training set and 0.96 on the test set, while those for RF were 0.98 and 0.94, respectively. The results indicated that GB demonstrated higher predictive accuracy for the M_r. Finally, the importance analysis, univariate sensitivity analysis, and bivariate interaction sensitivity analysis of influencing factors were systematically conducted using partial dependence plots (PDP) and Shapley additive explanations (SHAP). The research results showed that the importance of influencing factors on the M_r decreases in the order of maximum dry density to optimum moisture content ratio, wet–dry cycles (WDC), deviator stress, confining pressure, and ratio of oxide compounds in the cementitious materials. The bivariate interaction sensitivity analysis of the WDC, deviator stress, confining pressure, and ratio of oxide compounds in the cementitious materials did not disrupt their single-variable sensitivity relationships with the M_r. The variation of the WDC would destroy the single variable sensitivity relationship between the optimum moisture content ratio and M_r.

Rapid and accurate estimation of seismic responses in city-scale buildings is critical for post-earthquake loss assessment and pre-event identification of vulnerable buildings. However, conventional numerical simulation methods struggle to balance efficiency and accuracy when applied to large-scale buildings, while existing data-driven methods often rely on single-source datasets, limiting generalizability. Numerical simulation data of varying detail (e.g., floor- and component-based models) and field monitoring data form inherently multi-fidelity datasets, but integrating these heterogeneous sources remains challenging, particularly when different fidelities correspond to different building targets. To address this gap, we propose a multi-fidelity meta-learning algorithm that extends deep learning methods for seismic response prediction, demonstrated on multiple high-rise shear wall buildings. The proposed algorithm enables incremental data learning and model updates and is applied and validated across datasets of varying fidelities, including multiple numerical simulations and field monitoring data. Under small-sample field monitoring scenarios, the proposed method reduces overall prediction errors by 40.4%, compared to the typical transfer learning approach, demonstrating superior learning capabilities in limited-data settings. Additionally, to account for inaccuracies and potential noise in acquired structural information inputs under real-world conditions, the meta-learning model was trained and evaluated with varying levels of noise based on field monitoring data. Results indicate that the proposed meta-learning algorithm exhibits strong robustness when handling noisy inputs.

This paper introduces PRIMoS (Probabilistic Risk matrix Integration with MOnte carlo Simulation), an advanced computational framework that enhances cost overrun risk assessment and uncertainty quantification in infrastructure project management. PRIMoS is an innovative Bayesian Monte Carlo simulation framework integrated with a probabilistic risk matrix, providing comprehensive cost risk analysis. The proposed framework simultaneously addresses both cost uncertainties and time uncertainties, the latter through discount rate assessment, extending beyond traditional cost-focused approaches. PRIMoS employs a novel method to define risk magnitude (RM) levels for all project components, enabling adaptive probability distributions for Monte Carlo inputs. This approach allows for the capture of specific cost-related interdependencies and evolving risk patterns within the financial aspects of the project lifecycle. The framework's efficacy was demonstrated through application to a large infrastructure project, showcasing its ability to provide more accurate and detailed cost overrun forecasts compared to conventional methods. The proposed model improved cost estimation accuracy by predicting an increase in contingencies, thereby reducing the estimation error to less than 5%. PRIMoS offers a powerful tool for proactive risk management and informed decision-making in large-scale infrastructure development.