When addressing construction safety hazards in mega hydropower projects (MHPs), rectification measure formulation depends heavily on specialized experience, with manual formulation being prone to inconsistency, non-standardization, and poor practicality. Driven by this limitation, this paper proposes a Cascaded Hybrid Retrieval-Augmented Generation (CH-RAG) method that enhances large language models (LLMs) for generating hazard rectification measures by integrating a tailored Contextualized Late Interaction over BERT (ColBERT). Specifically, first, AdapColBERT adaptively optimizes ColBERT through adaptive representations and training protocols to capture critical hazard elements and boost the precision of rectification-basis ranking. Second, a Cascaded Hybrid Retrieval (CH-Retrieval) mechanism for multi-source engineering texts is built, combining keyword-based, sparse, and AdapColBERT dense semantic retrieval in a cascaded manner to achieve efficient, high-precision recall. Furthermore, a domain knowledge enhancement mechanism is established by integrating the CH-RAG system and fine-tuning the LLM with QLoRA to strengthen understanding of the hazard context and the generation of rectification measures. Extensive experiments indicate that CH-Retrieval with AdapColBERT effectively improves retrieval accuracy while keeping single-query latency under one second. Both CH-RAG and QLoRA greatly enhance the quality of rectification measures, outperforming general-purpose LLMs. Finally, an intelligent MHP hazard rectification workflow is developed around CH-RAG, enabling closed-loop safety hazard management.

Machine Learning (ML)-driven approaches have been employed to replace computationally intensive seismic simulations of hydraulic engineering structures. For the complex seismic responses of arch dams, constructing a metamodel that captures the nonlinear relationship between ground motion inputs and structural response outputs using a limited set of numerical simulations can significantly reduce the computational cost. However, conventional deterministic predictions and fragility analyses fail to account for the high aleatory and epistemic uncertainties inherent in the seismic response of arch dams. To this end, this paper proposes an efficient fragility analysis method for arch dams that integrates probabilistic ML algorithms with the traditional Incremental Dynamic Analysis. By constructing a Natural Gradient Boosting (NGBoost) metamodel for the arch dam dynamic response, not only can the predicted mean value of each response sample be obtained, but also its conditional probability distribution. Superimpose the simulation data with the response distribution predicted by NGBoost, and the binary parameters of the fragility function are estimated, thereby generating both the fragility curve and the uncertain fragility interval of arch dam. Additionally, representative Ground Motion Records (GMRs) for the arch dam are selected using the Partitioning Around Medoids (PAM) unsupervised clustering technique, determining the minimum subset proportion that effectively represents the whole GMR dataset. The effectiveness of the proposed method is validated in a super-high arch dam. The 40% GMR proportion is found to adequately reproduce the fragility curves of the whole dataset, with the reference curve falling within the derived uncertainty interval, achieving a 56.8% reduction in computational cost. The 60% GMR proportion ensured fragility curves with balanced accuracy and effectiveness, exhibiting maximum mean differences of 0.058 and maximum standard deviation differences of 0.031 from reference curves across all damage levels, while reducing computational cost by 39.7%. Comparative results demonstrate the superiority of NGBoost and PAM over existing deterministic metamodels and GMRs selection techniques, respectively. The efficient fragility analysis method proposed in this study ultimately enables the direct characterization of uncertainties in arch dam seismic responses.

Autonomous mobile robots (AMRs) have demonstrated significant potential for addressing a variety of real-world applications, ranging from logistics and inspections to construction and service tasks. Although modern AMRs are equipped with advanced mobility and task execution capabilities, their deploymen t remains largely confined to repetitive and pre-defined processes with limited flexibility. The challenge lies in enabling robots to dynamically interpret missions, manage context-dependent variables, and autonomously activate their capabilities in complex and unstructured environments. This study presents Auto-Mission, a solution-oriented intelligent control methodology that bridges user-defined missions and autonomous execution for mobile robots. Unlike the capability-oriented research, which advances specific robotic technologies, Auto-Mission addresses the practical challenge for enabling end users to define and deploy autonomous missions without close human supervision. This method systematically integrates existing capabilities, navigation, localization, and task execution into a coherent framework that is platform-independent, with the platform adaptation required only at the mission execution layer. Auto-Mission is built on a location-based map framework, where a mission is defined as a list of destinations coupled with specific tasks to be performed at each location. By leveraging spatial and accessibility data from the map, Auto-Mission computes optimal navigation paths, simulates mission workflows, and manages the real-time execution of tasks. Unlike rigid pre-recorded automation schemes, Auto-Mission introduces adaptability by dynamically integrating robot states, environmental contexts, and mission-specific goals. The methodology is designed to be universally applicable across diverse AMR platforms, with the customization required only at the mission execution layer for robot-specific commands and operational data handling. This study details the architecture of Auto-Mission, its integration with mobile robotic systems, and its potential to significantly enhance the deployment flexibility of AMRs in practical applications.

Reliable liquefaction prediction is crucial for the prevention and control of seismic hazards in earth-rock dams. However, the existing studies focus only on the accuracy of seismic liquefaction prediction and neglect the catastrophic consequences caused by underestimated liquefaction risk levels. Moreover, conventional liquefaction predictions are mostly conducted on a single-algorithm Machine Learning (ML)-based model, which can only achieve accurate classification in few specific regions, greatly limiting the applicability of liquefaction prediction models. This study thereby presents a Risk-Averse Hybrid Ensemble Learning (RAHEL) model for seismic liquefaction prediction, which is capable of classifying both Liquefaction States (LS) and Liquefaction Risk Grades (LRG) simultaneously. Specifically, a novel Performance-Weighted Voting (PWV) ensemble strategy is designed to capture complicated mappings from various influencing factors to LS or LRG. The GridSearchCV optimization algorithm simultaneously finds the optimal values of all hyperparameters of heterogeneous ML base-classifiers in the Voting ensemble to generate the most effective classification framework. Further considering the adverse effect of underestimated LRG, a cost function with the penalty mechanism is defined to guide the search direction of GridSearchCV, thus forcing the underestimated risk level to shift to an overestimation. RAHEL integrates the advantages of hybrid and Voting ensemble schemes by combining hyperparameter optimization with risk-averse adaptation with efficient ML models. The effectiveness and generalization ability of the proposed RAHEL is verified comprehensively with reference to four case studies of field data that cover different types of site conditions and levels of liquefaction risk, including three publicly available historical datasets and one real earth-rock dams engineering dataset. The classification performance of RAHEL is compared to that of other single ML models, conventional Voting ensemble, and hybrid models. The analytical results show that RAHEL is the most reliable approach, which can maintain a high prediction accuracy and reduce the underestimation rate of LRG, thus achieving adaptive avoidance of seismic liquefaction risks. Furthermore, the discussion on the category weight schemes confirms the reasonableness of RAHEL, and the analysis based on SHapley Additive exPlanations (SHAP) shows the contribution of each liquefaction factor to LS or LRG. This work develops a highly promising tool that greatly outperforms currently available methods to assist dam engineers in seismic liquefaction forecasting and early-warning.

Rock core logging is critical for evaluating rock mass quality, yet traditional manual measurement of rock quality designation (RQD) indicator from boreholes remains highly labor-intensive. To address the limitations of insufficient detection accuracy and generalizability in mainstream deep learning methods, we propose MG-FSAM4Seg, a vision foundation model adapted for rock core instance segmentation, combined with a fine-grained RQD analytics method. MG-FSAM4Seg adopts an enhanced multi-scale feature learning design that seamlessly integrates the Segment Anything Model (SAM) encoder with additional decoders, where the encoder is further fine-tuned via low-rank adaptation. We evaluate our model on two core imagery datasets, including the self-constructed Core1000 and the independent cross-project Core-G47, which exhibit considerable lithological and resolution differences, benchmarked against 15 advanced segmentation methods. On the Core1000 test set, MG-FSAM4Seg achieves competitive performance in both qualitative and quantitative evaluations, reaching an average precision of 88.07

Although routine inspection of gallery cracking is critical to ensuring dam structural safety, poor lighting in galleries reduces the contrast between cracks and their surroundings, blurring the boundaries and making visual detection challenging. Notably, most deep learning-based methods focus on well-lit cracks, with performance degrading sharply in low-light conditions. Driven by this limitation, this paper introduces an integrated framework, LL-CrackSeg, which includes two major components. The first one is an unsupervised Retinex decomposition-based enhancement network (URDE-Net) to enhance low-light crack images captured in underlit galleries while suppressing extraneous noise by designing unsupervised loss functions and decomposing illumination information. The second component is a customized multi-level feature aggregation network (MLFA-Net) for robust crack segmentation, which leverages a bidirectional path aggregation module and a dual-domain attention block to strengthen the semantic representation of latent features within brightened cracks. Additionally, we create a new benchmark dataset (GalleryLL-1240) for method validation, comprising four test subsets with distinct levels of darkness. Experimental results demonstrate that the proposed framework outperforms state-of-the-art methods, highlighting great potential for low-light inspection tasks.

Accurate modeling of structural behavior is pivotal to dam health monitoring. However, most existing models rely heavily on sufficient monitoring data, failing to address the practical challenge of observational insufficiency. Traditional task-specific modeling approaches further restrict the widespread application of predictive models. Therefore, this study proposes a synthetic pretrained tabular transformer network (SP-TabTNet) to capture the inherent evolutionary patterns in a generalized manner, particularly from small-sample datasets. Specifically, SP-TabTNet first integrates a structural causal model to formalize dam behavior via directed acyclic graphs and structural equations, thereby synthesizing extensive tabular data embedded with physical causal priorities. This design addresses data scarcity while ensuring the cross-scenario physical consistency. Subsequently, a synthetic pretrained sliding in-context learning paradigm is derived to align the feature distribution of target data with pretrained universal evolutionary representation. In this way, it enables gradient-free dynamic adaptation, which facilitates the modeling of nonlinear dam structural behavior using minimal prompts during inference. Moreover, an attention-enhanced transformer is introduced as the prior-data fitted network for generalizable tabular learning, which adopts the dual-attention mechanism to simulate both temporal dynamics and feature dependencies. The proposed network is validated on both dam displacement and seepage datasets collected from real-world dam projects through systematic comparison analyses. Results show that SP-TabTNet retains high modeling accuracy (R2 > 90%) even under ultra-small-sample conditions (using only 10% training data), and outperforms state-of-the-art baseline models across multiple metrics. Extended experiments demonstrate that SP-TabTNet achieves superior cross-point and cross-item generalization performance for structural behavior prediction, offering a generalizable solution for dam health monitoring under small-sample regimes.

Seismic Intensity Measures (IMs) exhibit strong correlations with earthquake hazards and structural responses, and their rational selection underpins fragility modeling by characterizing the mapping relationship between ground motions and structural seismic behaviors. To optimize the hybrid structure-dependent and structure-independent IM sets within arch dam fragility assessment, this paper proposes a Reinforcement Learning Optimization Algorithm (RLOA) for optimal selection of vector-valued IMs. The specific improvements of RLOA to data-driven task adaptation are as follows: (i) Guided by RL principles, a novel adjustment strategy for modification factors is devised to maintain population diversity. (ii) Historical population information is incorporated into the transfer operator to strengthen the global search capability of RLOA. (iii) A feedback operator integrating current and historical feedback terms is constructed to accelerate the convergence speed of RLOA. Furthermore, an optimization objective function is formulated to screen optimal vector-valued IMs from a candidate set of 40 IMs, utilizing dataset derived from 500 seismic dynamic simulations of a high arch dam. Results show that the proposed RLOA outperforms baseline, greedy, and other metaheuristic methods in terms of search efficiency and solution quality for vector-valued IM selection. In the fragility functions of the arch dam, structure-dependent IMs are preferentially selected with relatively stable performance, whereas time-related and dimensionless IMs capture latent interactive characteristics within vector-valued IMs. The optimal vector-valued IM varies with metamodels and structural demand parameters. Specifically, the optimal IM for XGBoost-predicted joint opening contains eight elements with an R2 of 0.882, while that for the damage volume ratio includes seven elements with an R2 of 0.878. The proposed method enhances vector-valued IM selection for arch dam fragility modeling via RL mechanisms, providing a reproducible framework for data-driven seismic performance assessment.

Low-strain reflection testing is widely used for the rapid screening of pile integrity, but its interpretation still relies heavily on manual judgment. This study proposes a dual representation learning framework for classifying the integrity of building foundation piles from low-strain testing records. A dataset containing 1139 piles from engineering projects was established and divided into four integrity classes. Each record was represented in two complementary forms: structured features extracted from engineering parameters and waveform characteristics, and a redrawn waveform image generated from coordinate point data. Support vector machine (SVM), random forest (RF), and convolutional neural network (CNN) models were used as single modality baselines, and their performance was compared with that of a multimodal neural network (MNN) trained on paired structured and image inputs. The multimodal model achieved the highest overall accuracy on the main evaluation subset, reaching 84.65%, whereas the random forest achieved the best Macro-Recall and Macro-F1. This result suggests that multimodal fusion mainly improved overall robustness rather than consistently enhancing performance across all classes. Clearly intact piles and severely defective piles were easier to identify, whereas Class II remained the most difficult category because of its borderline signal characteristics. In the supplementary external validation set, the same ranking of model performance was observed, and the multimodal model achieved an accuracy of 85%. These results indicate that the proposed framework has strong potential for computer-assisted screening of building foundation piles.

Accurate classification of surrounding rock is vital for ensuring the safety and efficiency of TBM operations. To address the limitations of existing methods, which often overlook ascending-stage dynamics, encounter difficulties in multi-source data fusion, and lack interpretability, this study proposes a Cross-Attention Transformer with XGBoost (CA-Trans-XGBoost). The study uses Section IV of the Yinsong Water Diversion Project as a case study, collecting and organizing 802 days of operational data and cutter replacement records. A Transformer encoder was applied to extract dynamic features from the ascending stage, while an MLP modeled structured features from the stable stage. A cross-attention mechanism was introduced to enhance feature interaction, and fused features were further processed with XGBoost for classification and feature importance analysis. Results show that CA-Trans-XGBoost achieves the best performance among six models, with an Accuracy, Precision, Recall, and Macro-F1 of 95.0 %, 93.6 %, 92.6 %, and 92.9 %, respectively. The model shows clear advantages in identifying minority classes II and V. Further analysis confirms that a 30-s ascending stage is the optimal temporal window. The proposed method balances predictive accuracy and interpretability, providing support for intelligent TBM excavation and parameter optimization.

The incorporation of conductive nanomaterials into cementitious matrices to construct functional percolation networks paves the way for sustainable and energy-resilient infrastructure. Accurately characterizing the three-dimensional dispersion state and connectivity of these nanomaterials is prerequisite for optimizing energy conversion and storage efficiencies. However, due to the inherent opacity, heterogeneity, and high hardness of the cementitious matrix, existing microstructural characterizations remain largely confined to two-dimensional planes. This restriction makes it challenging to authentically resolve the 3D spatial distribution of the carbon phase and hydration products at the mesoscale. To overcome this limitation, this study proposes a 3D chemical imaging framework integrating Broad Ion Beam milling with Serial Section Raman Tomography to achieve high-precision volumetric reconstruction. Applying this methodology to a modified cement paste incorporating 4.5 wt.% nano-carbon black (NCB) not only demonstrates the feasibility of 3D visual reconstruction within an opaque matrix but also intuitively reveals the NCB distribution and its spatial nucleation mechanisms. Furthermore, the intrinsic topological parameters of the conductive network are successfully quantified, including tortuosity, fractal dimension, and average coordination number. Utilizing these parameters, the macroscopic conductivity of the composite was accurately predicted at 0.038 S/m, theoretically elucidating the correlation between the formation of the 3D conductive network and the enhanced energy-storage capacity in cementitious materials. This work provides a novel characterization tool for the 3D mesoscopic analysis of complex multiphase systems and corrects stereological errors inherent in planar evaluations, laying a geometric foundation for establishing quantitative structure-property relationships correlating micro-3D architecture with macroscopic energy transport and electrochemical storage.

Reliable displacement prediction is crucial for dam safety monitoring and early-warning. Longer monitoring sequences are usually more informative for modeling the behavioral dynamics of dam structure. Most existing data-driven models, however, struggle with handling extremely long sequences, as they mainly prioritize the latest data information. For dam operators, the interpretability deficit of such black-box methods further undermines their operational credibility in real-world applications. In this paper, we propose an interpretable deep learning model, termed Bahdanau attention-based sequence-to-sequence single-gate recurrent unit (abbreviated as BAS2S-SgRU), which aims to address the long-term displacement prediction problem. BAS2S-SgRU consists of three main components, namely SgRU, S2S encoder-decoder module, and attention mechanism. Specifically, the designed SgRU features only one gate, which is a simplified gating structure to simulate dynamic displacement behavior by incorporating the temporal dependency in both environmental factors and structural responses. The S2S architecture, composed of an encoder and a decoder, excels in modeling sequential monitoring data with varying lengths, and BA further endows it with the ability to capture implicit long-term dependencies, alleviating the error propagation in multi-step prediction. Moreover, the attention mechanism enhances the physical interpretability of BAS2S-SgRU via attention weights, explicitly linking modeling results with context information. The effectiveness of the proposed model is verified using the monitoring data from a real concrete dam project, with a set of comparisons made among our model and other advanced methods. The results show that BAS2S-SgRU achieves the impressive RMSE, MAE, and R2 values of 0.104mm, 0.082mm, and 98.22% on the dataset while training for only 3.14s/epoch, consistently outperforming all baseline models in displacement prediction. The approach developed in this work holds significant potential for structural health monitoring of concrete dams.

Accurate dam displacement prediction is vital for optimizing maintenance and ensuring structural safety. Nevertheless, current models often struggle to effectively capture the complex relationships between structural responses and environmental variables, alongside the interactions between temporal dynamics and multivariate data, resulting in suboptimal predictive accuracy. Therefore, we propose a dual-branch interactive fusion network (DBIFN) for dam displacement prediction using parallel temporal representation and gated cross-attention. The dual-branch architecture, which parallelly integrates the enhanced Transformer (eTransformer) and long short-term memory (LSTM), is designed to optimize feature extraction and interaction modeling across multiple dimensions. Specifically, eTransformer is dedicated to extracting features from targeted displacement sequences, while LSTM effectively processes auxiliary environmental dynamics, enabling a comprehensive analysis of underlying patterns within monitoring data. To fully fuse the interpreted temporal features from dual-branch outputs, we introduce a new cross-attention module to utilize the multi-dimensional gated attention unit to efficiently encode them into semantic representations, followed by a Kolmogorov-Arnold network mapping for further representation enhancement. The effectiveness of the proposed model is validated using real-world monitoring datasets collected from a concrete dam project, with experiments conducted across multiple monitoring points. Results demonstrate that DBIFN achieves superior prediction accuracy compared to both single-branch and conventional baseline models. Across all monitoring points, the proposed model can effectively capture temporal variations, attaining an average coefficient of determination of over 0.95 on the test set and outperforming comparative models in most metrics. Furthermore, statistical significance testing confirms the reliability and reproducibility of the results, while computational efficiency is maintained within inference time constraints. These findings offer valuable insights into the practical application of DBIFN-based monitoring models and support informed decision-making.

The harsh climatic conditions in alpine regions often pose challenges to the construction of concrete face rockfill dams. The purpose of this study is to investigate the enhancement of crack resistance in hydraulic concrete for face slabs by systematically comparing internal admixtures including nano-calcium carbonate (NCC), silica fume (SF), silica fume-fly ash composite (SF-FA), magnesium oxide (MgO), and anti-seepage crack-resistant material (ACM). A two-phase experimental approach was adopted: Phase I optimized a reference mix (water-to-binder ratio of 0.35-0.36, fly ash content of 20-30%) for the target project, while Phase II systematically evaluated the workability, mechanical properties, deformation, and durability of the modified concrete. Experimental findings show that SF significantly enhanced mechanical strength, increasing the 28-day compressive strength to 50.1 MPa, while NCC and ACM effectively reduced the 90-day drying shrinkage strain to approximately 300×10−6 , compared to 380×10−6 for the reference mix. To address inconsistencies in traditional single-index evaluations, an intelligent assessment framework based on the Entropy Weight Method (EWM) was developed, integrating multi-dimensional performance data into a Comprehensive Crack Resistance Index (CCRI). The EWM-based analysis ranked SF, NCC, SF-FA, and MgO as the top four performers based on CCRI. However, considering fresh concrete workability constraints, NCC-concrete, with a slump of 70 mm and the highest overall stability, was identified as the optimal material solution. This datadriven approach provides a validated strategy for enhancing the crack resistance and long-term durability of CFRD face slabs in dry, windy alpine environments.

The timely hazard inspection during arch dam construction is vital for accident prevention and project safety. The construction-site hazard reports generally contain valuable clues for identifying potential risks, enabling dam managers to assess engineering hazards promptly. However, high domain-specific text-labeling costs and lack of a comprehensive analytical process impede their conversion into effective hazard inspection tools. Therefore, this paper proposes a momentum contrast-based semi-supervised hazard text classification model, namely MoCo-ASFormer. Comprising an asynchronous symmetric Transformer and a contrastive learning module, it enables accurate classification of construction safety hazard categories and levels using minimal labeled data. Initially, a momentum-updated asynchronous symmetric Transformer is constructed, and the Memory Bank is refined to ensure more consistent and comprehensive negative-sample collection. A contrastive learning module is then used to optimize the Transformer structure. This module effectively exploits inter-class relationships to amplify the differences among dissimilar samples, thereby significantly boosting the classification performance and facilitating adaptation to multi-attribute hazard classification tasks. We further extract potential hazard risk inspection clues by integrating the latent Dirichlet allocation with Apriori algorithm, and an intelligent inspection workflow for rule-based critical hazard information is implemented. For method validation, over 26,000 hazard records are collected from super-high arch dam construction sites. Results indicate that with 50% fewer labeled samples, MoCo-ASFormer achieves comparable performance to competing models in identifying hazard categories and levels, which rely on fully labeled sample sets. The rule-based recommendation workflow empowers managers to efficiently acquire hazard inspection clues, which offers a new solution for intelligent project safety management.

The dynamic response analysis of hydraulic tunnels in high-intensity seismic regions is often oversimplified in current research. Long tunnels crossing multiple faults, coupled with spatially nonuniform seismic excitation, pose significant hazards. This study introduces a comprehensive modeling methodology for tunnels intersecting multiple fault zones, incorporating multidimensional, multipoint random seismic excitation with intercomponent correlations. The results indicate that the tunnel response evolves progressively with increasing seismic intensity. For peak ground acceleration (PGA) <= 0.1g, displacement responses exhibit approximately linear proportionality to input motion, and no significant plastic deformation is observed. At PGA almost equal to 0.2g, localized plastic zones initiate, and displacement amplification increases markedly (up to 2.34 at Section S6), indicating the onset of nonlinear behavior. For higher intensities (PGA >= 0.3g), plastic zones expand substantially, and at 0.4g PGA, the maximum displacement reaches 12.33 cm with a plastic damage index of 5.07. Vertical displacements at the tunnel crown and invert are most sensitive to fault location, with Section S6 exhibiting the largest response. The plastic zone evolves from circular to elliptical, aligned with the fault dip angle, indicating fault-induced anisotropic stress redistribution. Increasing the number of fault intersections amplifies the peak displacement at the Fault-crossing section S6 by up to 49.30% (z-direction, PGA = 0.2g), while the plastic damage index rises by as much as 87.63%. These findings highlight the need to explicitly consider fault networks and intensity-dependent nonlinear effects in the seismic design of hydraulic tunnels.

This study examines the influence of wind intensity on mixture behavior, forming quality, mechanical performance, hydration characteristics and interfacial transition zone (ITZ) of roller-compacted concrete (RCC). Meanwhile, the formation mechanism of its spatial variability is revealed through the probabilistic statistical analysis of compressive constitutive parameters. The results show that wind disturbances cause surface RCC to present increased VC values, more surface pitting, inhibited hydration and pozzolanic reactions, and reduced microhardness and compressive strength. Especially, force-9 wind increases 123.7% in porosity and 174.8% in thickness of the ITZ. With enhancing wind force, AFt tends to convert to carbonoaluminates and the reaction degree of C3A and C4AF decreases by up to 10%, exhibiting higher wind sensitivity. Statistically, the spatial variability is transmitted across scales and amplified along the thickness direction under high wind-force level, ultimately increasing the coefficients of variation of elastic modulus and peak strain by 4.8 and 2.7 times, respectively.

Accurate displacement prediction is critical for the operation and maintenance of concrete dams. Existing methods overemphasize the global pattern of displacement behavior while neglecting the temporal association in monitoring sequences. Therefore, we propose a Bayesian-augmented Gaussian process (BaGP) integrated with double clustering analysis for temporal-aware dam displacement prediction. First, a double clustering approach is employed to thoroughly explore environmental patterns and further consider temporal association in displacement sequence evolution. Specifically, a deep clustering algorithm is introduced for the first time to extract high-quality data patterns within multidimensional nonlinear environment data. Subsequently, temporal-aware similarity among displacement sequences is quantified using dynamic time warping, and an improved K-means clustering algorithm is employed to identify temporally similar displacement phases within each environment cluster, enabling cluster-wise division. A nonparametric BaGP model is developed by augmenting standard GPs through Markov chain Monte Carlo simulation, Bayesian evidence evaluation, and model selection, to comprehensively address model structural flexibility and parameter adaptability across diverse data patterns. Each cluster's data is sequentially fed into the BaGP model, with ensemble strategies used to generate integrated, cluster-wise predictions. Validation on real-world dam monitoring datasets demonstrates that our method achieves an average R of 0.973, outperforming models that ignore temporal association. Two additional cases also confirm its generalizability, thus providing a novel tool for structural health monitoring.

The operational parameters of cutter suction dredger (CSD) are frequently set based on empirical experience, leading to low productivity and high energy consumption. To address this issue, the present study proposes a Bayesian multi-objective optimization (MOO) framework integrating a multi-scale adaptive graph convolutional network (MAGCN) model and a multi-objective tree-structured Parzen estimator (MOTPE) algorithm. This framework aims to predict and optimize construction productivity and energy consumption performance while providing well-informed support for intelligent decision-making control. First, a hybrid feature selection method combining maximum information coefficient and Pearson correlation coefficient is applied to identify 6 key control parameters from 256 operational parameters. Second, an MAGCN model with multi-scale graph representation is proposed to establish the nonlinear mapping among geological conditions, operational parameters, construction productivity, and energy consumption. Third, the MOO framework with its mathematical formulation is constructed, and the Pareto optimal front for productivity and energy consumption is derived using the MOTPE-based Bayesian optimization algorithm. The optimal trade-off solution and corresponding operational parameters are determined by a weighted sum method. The proposed MOO framework is validated using operational data from the Tian Jing Hao CSD in the Pinglu Canal dredging project in China. The results show that the MAGCN model achieves high prediction accuracy for real-time productivity and energy consumption, with R2 values of 0.942 and 0.979, respectively. The MOTPE-driven optimization reduces standard deviations of operational parameters by 21.24% - 97.90%, enhancing operational stability. The framework improves productivity by 3.79% and reduces energy consumption by 3.53%. This work provides real-time decision-making support for optimizing CSD operations.

Ground penetrating radar(GPR)image interpretation is a critical technique for the geotechnical engineering exploration and subsurface structure detection.Its accuracy and professionalism directly influence the scientificity and safety of engineering decisions.The current GPR interpretation algorithms predominantly rely on the visual deep neural networks,which can only provide rudimentary classification results but is lack of professional geological interpretation capabilities.In this study,we constructed a comprehensive dataset for the GPR images and professional text interpretations,proposed a novel GPR image interpretation method based on the multimodal large models,and achieved a technological leap from traditional detection to professional interpretation by fine-tuning the Florence-2 visual language model with low-rank adaptation(LoRA).The proposed method can be used to conduct multi-dimensional professional analyses of GPR images,including waveform analysis,generation of prediction results,and discrimination of wallrock category.Experimental results demonstrate that a BERTScore of 0.8641 has been obtained by using the fine-tuned model to process the test dataset.Especially,the accuracy rate for discriminating wallrock category is as high as 0.9291.This method serves a novel technical pathway for the intelligent professional interpretation of GPR images.