In situ monitoring of soil water content using actively heated optical frequency domain reflectometry offers high spatiotemporal resolution. However, the accurate retrieval of soil water content from complex thermal response data remains challenging. Conventional theoretical models, limited by idealized physical assumptions, often fail to characterize heat transfer in multiphase porous media. Similarly, purely data-driven inversion methods often lack physical consistency and robustness. To address these issues, this study developed a physics-informed neural network framework validated through laboratory experiments using sand columns. The framework integrates macroscopic physical principles from asymptotic heat transfer analysis into a correlation-based loss function to mitigate measurement noise sensitivity. Specifically, the model incorporates a fixed thermal contact resistance parameter to decouple sand intrinsic properties from interface effects. This enables a submodel to determine the thermal conductivity-water content (7 -w) relationship without prior calibration. Experimental validation shows that the model achieves high accuracy for soil water content (R2 = 0.92, MAE = 0.0125 cm3 & sdot;cm-3) and improved robustness compared to benchmark models. The identified 7 -w function was validated against independent, ground-truth measurements of the sand's thermal conductivity, confirming it captures the correct physical trend. This work provides a reliable approach for the distributed retrieval of water content in coarse-grained media with high physical consistency and interpretability.

Pipeline infrastructure traversing mountainous terrain faces catastrophic threats from rainfall-induced landslides, yet current risk assessments often fail to account for the complex spatial variability of soil properties, particularly rotated anisotropy resulting from geological stratification. This study establishes a comprehensive GPU-accelerated stochastic framework (GCRCEL-OLHS) to quantify the reliability of pipelines under large-deformation landslide impacts. To accurately simulate the disaster evolution, a coupled seepage-stress model incorporating a pressure-sensitive bulk modulus and a modified Mohr-Coulomb criterion was developed to capture the hydraulic-mechanical response, while a strain-softening constitutive relationship was implemented to drive the transition from localized instability to large-scale run-out. A parallelized covariance matrix decomposition strategy, integrated with the Karhunen–Loève expansion, is developed to efficiently simulate random fields with rotated anisotropy. Benchmark tests demonstrate that the GPU-based approach reduces random field generation time from hours to seconds. Parametric studies reveal that the rotational angle of soil variability dictates the failure mechanism: orientations parallel to the slope inclination facilitate translational failure modes, resulting in significantly larger impact forces and pipeline deformations compared to rotational failures. Reliability analysis indicates that conventional deterministic approaches may underestimate pipeline failure probability by neglecting the strain-softening behavior and directional connectivity of weak soil zones. This work establishes a robust high-performance computing pathway for advanced geotechnical risk assessment.

This study addresses the challenge of using normalized contact parameters in discrete element modeling, which limits the accurate representation of spatial variability's impact on the mechanical properties of rockfill materials. A random discrete element modeling approach that accounts for spatial variability is proposed. First, a microscopic parameter interpolation method is introduced, and a Karhunen-Lo & egrave;ve expansion-based, three-dimensional log-normal random discrete element model is developed. This approach establishes an efficient process from random field generation to contact stiffness assignment. A random discrete element model for rockfill material under triaxial compression is constructed. The study analyzes stress-strain curves under different confining pressures and random field parameters, and examines mesoscopic fracture characteristics and contact force distributions. These analyses reveal how spatial variability in contact stiffness regulates macro- and meso-mechanical behaviors. The results show that the random model outperforms traditional models, aligning better with practical engineering applications. The proposed method overcomes the limitations of the homogeneity assumption in conventional models, providing anew approach for studying the mechanical properties of rockfill materials and assessing their engineering safety. (c) 2026 Published by Elsevier B.V. on behalf of The Society of Powder Technology Japan. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

Shield tunneling poses a significant challenge in tunnel engineering by inducing ground settlement, presenting considerable risks to urban infrastructure safety, with potential regional systemic failure leading to. The mechanisms by which shield construction parameters influence soil displacement and surface settlement are currently understood to a limited extent. To address these gaps, this study introduces a coupled continuous-discontinuous model for shield machines. Soil layers at critical excavation points were represented as discontinuous media, while continuous media modeling was utilized for other sections. The benefits of discrete element modeling for large deformations in rock and soil were combined with the computational accuracy and efficiency of the finite difference method by this method. Building on this foundation, the study undertook a comprehensive analysis of settlement evolution patterns at the excavation face during tunneling. The influence patterns of key construction parameters, such as shield advance rate and cutterhead rotational speed, on the deformation characteristics of the soil at the excavation face and and their impacts on ground surface settlement were identified in this study. The distribution of force chains at the excavation face throughout the entire shield construction cycle was examined, revealing the spatiotemporal evolution patterns of the micro-scale force chains of soil particles from a micro-mechanical perspective. It is demonstrated by the research outlined above that the developed continuous-discontinuous coupled model for shield tunneling not only significantly improves computational efficiency but also accurately captures the large-deformation behavior of soil during excavation. It possesses analytical capability for the micro-mechanical response in critical zones. The proposed shield continuous-discontinuous coupling model significantly enhances our comprehension of soil deformation and stress transfer mechanisms during shield tunnelling. This model offers vital theoretical underpinnings and technical assistance, thereby facilitating the optimisation of shield construction parameters and guaranteeing construction safety.

The unclear impact of temperature on rockfill dam settlement and the lack of a solid basis for selecting temperature parameters in prediction models are problematic. These issues significantly limit the accuracy and applicability of deformation monitoring models for rockfill dams. For this reason, a method for decomposing the settlement components of rockfill dams, along with an intelligent prediction approach, is proposed. The Bayesian optimization (BO) algorithm is employed to optimize the hyperparameters of the Bayesian dynamic linear model (BDLM), enabling a comprehensive exploration of the correlation between rockfill dam settlement and temperature factors. Based on this, a BO-BDLM-based decomposition model is constructed to quantify the contribution of the temperature factor to settlement behavior. Spatiotemporal analysis is conducted to uncover the evolution patterns of various influencing components, revealing the underlying mechanism by which temperature affects settlement. Furthermore, both a full-feature model and a simplified prediction model are developed to predict settlement, and their prediction accuracies are compared. The contribution of the temperature factor is quantitatively assessed using the SHapley Additive exPlanations (SHAP) method. Example analyses demonstrate that our BO-BDLM significantly improves performance and accurately isolates the temperature factor consistent with rockfill dam deformation characteristics. The temperature component contributes approximately 2%-4% of total settlement but accounts for 38.39% of model importance. This pivotal factor substantially enhances prediction accuracy. By quantitatively assessing temperature influence and establishing its selection basis, our study offers valuable insights for the safety evaluation of rockfill dams and related engineering projects.

Leakage from buried water pipelines in loess areas can induce soil erosion, ground deformation, and collapse, threatening water conveyance systems and surrounding environments. To clarify leakage-induced failure mechanisms and improve continuous monitoring and anomaly identification, this study investigated a typical buried water pipeline in a loess area. A pipe–soil numerical model was established using CFD–DEM coupling, with loess discrete-element parameters calibrated by laboratory direct shear tests. Distributed fiber-optic monitoring units were embedded in the model to examine the effects of leakage location, burial depth, leakage size, and internal pipe pressure on soil erosion, cavity development, surface deformation, and fiber-optic strain response. A GAN-CLSA model, integrating GAN-based data augmentation with a CNN-LSTM-Attention classifier, was further developed to identify three anomalies: pipeline leakage, foundation pit excavation, and rockfall impact. The results show progressive erosion and deformation evolution during pipeline leakage. Fiber-optic strain responses effectively characterize the location and development of abnormal zones. After GAN-based sample augmentation, the GAN-CLSA model converged stably. The training and validation accuracies stabilized at approximately 95%, while the class-wise test accuracies ranged from 88.6% to 97.5%.

Accurate prediction of seepage pressure head is essential for assessing the integrity of concrete‐faced rockfill dams (CFRDs), particularly for CFRDs serving as reservoir basins of pumped‐storage power stations, where daily water‐level fluctuations of tens of meters make seepage losses a direct threat to generation efficiency. Existing data‐driven methods struggle to jointly model the complex coupling among environmental drivers and the long‐range temporal dependencies that characterize seepage monitoring sequences. To address this issue, we propose Deep‐Autoformer, a multistep prediction model that embeds progressive series decomposition as an internal network primitive and replaces standard self‐attention with a Deep‐AutoCorrelation mechanism. In this mechanism, the single‐layer query–key projections are deepened into a three‐layer fully connected network (DeepFFN), strengthening the extraction of nonlinear temporal patterns. Evaluated on three monitoring points of a pumped‐storage CFRD over a 240‐day horizon, Deep‐Autoformer achieves R 2 values of 0.9325, 0.9426, and 0.9353, outperforming Informer, Transformer, Reformer, LSTM, and GRU by margins of 0.04–0.16 in R 2 . Ablation experiments confirm that DeepFFN, the autocorrelation mechanism, and progressive decomposition each contribute independently to prediction accuracy. The model provides a practical tool for real‐time seepage safety monitoring and preventive maintenance of pumped‐storage CFRDs.

Distributed optical fiber strain sensing (DOFSS) for loess landslide monitoring depends fundamentally on cable–soil deformation compatibility. This study investigated the water-content-dependent interfacial coupling mechanism through pullout tests and continuum–discontinuum coupled numerical simulations. The results show that the interface exhibits progressive failure and strain-softening behaviors. As water content increases, both peak pullout force and displacement decrease, indicating reduced coupling capacity. The simulations successfully reproduced the pullout response, capturing soil displacement and bond breakage around the cable. Finally, a deformation coupling coefficient,K, and a strain-energy-based coupling coefficient,S, were introduced to quantify compatibility. This framework improves the evaluation of cable–soil coupling and provides a basis for designing and interpreting fiber-optic monitoring systems in rainfall-sensitive loess slopes.

Accurate calculation of dam seepage flow based on engineering inspection results is essential for assessing operational status and ensuring structural stability. This study combines the equivalent permeability coefficient with artificial intelligence to establish a model for predicting the seepage volume of concrete face rockfill dams (CFRD) under the condition of panel cracks. Firstly, the finite element method was used to obtain the calculation results of dam seepage flow under different working conditions. Second, a sensitivity analysis was conducted to quantitatively determine how panel-crack parameters affect the dam's seepage discharge under face slab cracks conditions. Then, a model for predicting dam seepage flow under panel crack conditions based on Genghis Khan shark optimizer and Long short-term memory network was developed to predict the seepage flow under special operating conditions of CFRDs. The study establishes a basis for key directions in face slab crack detection and introduces a new approach for modeling dam seepage under face slab crack conditions.

Acquiring effective vibration responses from arch dams is a critical prerequisite for analyzing their dynamic characteristics and performing damage diagnosis. However, arch dam vibration signals typically exhibit low signal-to-noise ratio (SNR) and non-stationary characteristics. Existing denoising methods struggle to suppress noise while fully preserving the modal features within the signal. To address the aforementioned issues, this paper proposes a multi-channel spatial-temporal U-Net (MC-ST-UNet) framework driven by structural response consistency for arch dam vibration signal denoising. First, a multi-point joint U-Net architecture is designed, integrating spatial-temporal attention mechanisms with Transformer modules to construct spatial correlation modeling and temporal context aggregation modules. This constrains the structural response consistency of multi-point vibration signals from arch dams. Second, an improved complex ideal ratio mask is designed to preserve the phase information of arch dam vibration signals, addressing the issue of modal feature loss commonly caused by generic denoising algorithms. Finally, a multidimensional physical constraint loss function based on arch dam mechanics is developed. This establishes a dual-domain (frequency-time) collaborative optimization framework to ensure the physical authenticity of denoised signals while maximally preserving modal characteristics. Numerical validation is conducted using a 6-degree-of-freedom (6-DOF) dynamic reference system, complemented by engineering application verification using field data from an actual arch dam. Six mainstream denoising methods serve as baseline models for comparative analysis. Results demonstrate that the proposed MC-ST-UNet model consistently outperforms baseline models in denoising performance. Numerical validation reveals an approximately 40% improvement in SNR compared to the optimal baseline model. This method provides high-quality data support for subsequent modal identification and damage diagnosis in arch dam structural health monitoring.

This study addresses the limitations of traditional feedback analysis methods for dam construction materials, which suffer from low accuracy, long computation times, and an inability to capture micromechanical properties. We propose a novel macro-micro parameter joint intelligent feedback analysis model for rockfill materials, driven by dam deformation monitoring data, that efficiently and accurately determines the macro and micro parameters of these materials. By employing an intelligent inverse analysis model, researchers can derive the macro and micro material parameters of the discrete-continuum coupling model, aiding in the optimization of design standards and guiding dam construction and operation. To enhance this process, we construct an adaptive surrogate model using a Runge-Kutta optimizer (RUN) and an extreme gradient boosting (XGBoost) algorithm. This model captures the complex nonlinear relationship between macro and micro parameters and dam settlement, reducing the need for time-consuming numerical simulations. By leveraging deformation monitoring data from panel rockfill dams, the RUN-XGBoost algorithm effectively addresses the inverse analysis problem. The results demonstrate that this intelligent inverse analysis model can rapidly and accurately determine rockfill dam parameters, improving the precision of macro-micro parameter calculations and enabling a comprehensive investigation of the mechanical evolution of rockfill materials, with implications for structural safety analysis.

A large number of pumped storage power stations have been planned and constructed worldwide in recent years. Considering the rapid rate and large amplitude of reservoir water level fluctuations, an improved deformation monitoring model adapted to the operational characteristics of pumped storage dams is proposed. Firstly, the influence factors in the classical deformation prediction model are improved by introducing the water-level change rate term. To reduce the adverse effect of multicollinearity among influence factors on prediction accuracy, kernel principal component analysis (KPCA) is adopted to optimize the factor combination for the monitoring model. Secondly, to overcome the limitation of deterministic prediction in existing dam safety monitoring models, deep learning and interval prediction are integrated. This study proposes an improved deformation influence factor model considering water-level change rate and establishes a deterministic–interval joint prediction framework for pumped storage dam deformation. A case study shows that the MAE of the proposed model is reduced by an average of 24.98% relative to the compared model. The proposed fusion model delivers high-precision deterministic predictions of dam deformation and generates corresponding prediction intervals to quantify predictive uncertainty. It can provide more comprehensive support for the safety monitoring and evaluation of pumped storage dams.

The measured dynamic response of concrete arch dams under seismic excitation is a typical time series that contains rich information about structural conditions. Safety monitoring based on dynamic responses of arch dam structures is highly important for the timely detection of structural damage and ensuring dam safety. In this study, a PSO-LSTM-based model for safety monitoring and damage identification of arch dam structures was proposed. The method was centered on the long short-term memory (LSTM) neural network, and key hyperparameters were adaptively tuned by the particle swarm optimization (PSO) algorithm to improve monitoring accuracy for nonlinear and nonstationary structural dynamic responses. Structural damage was identified through residual analysis combined with the 3σ anomaly detection criterion. Numerical simulations and shaking table model test cases of an arch dam were introduced for validation. The proposed method was compared with the standalone LSTM model and the SSA-LSTM model in terms of the root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R2), and damage identification accuracy. The results showed that the proposed PSO-LSTM method achieved greater accuracy in monitoring the safety of arch dam dynamic responses and effectively identified structural damage, thereby verifying its effectiveness.

Modal identification is a fundamental topic within the domain of structural health monitoring (SHM). Nonetheless, the diverse types of concrete dams and the variability in operating conditions present considerable challenges in accurately identifying their modal parameters when subjected to ambient excitation. This study proposes a three-stage physics-informed neural networks (TSPINNs) framework aimed at achieving highprecision automatic identification of the modal parameters of concrete dam structures under ambient excitation. The framework is composed of three principal stages: Initially, a singular value decomposition (SVD) network is developed, that uses SVD and matrix differentiation principles, to automatically ascertain the modal order. A physics-constrained blind source separation neural network is subsequently designed, that incorporates the statistical characteristics of blind source separation (BSS) and a frequency-domain attention mechanism, to effectively segregate ambient excitation signals. Finally, a vibration equation solving network is constructed, grounded in dynamic knowledge, to accurately identify the modal parameters of the isolated signals. The results, validated through numerical experiments on a 6-degree-of-freedom (6-DOF) mass-spring-damper system and empirical data from two actual concrete dams, demonstrate that TSPINNs can effectively distinguish and identify all physical modes. While the accuracy of frequency and mode shape identification is comparable to that of traditional methods, the identification accuracy of the damping ratio in the 6-DOF system is notably enhanced by 82.25 %. These findings underscore the synergistic advantages of integrating physics-informed machine learning with data-driven approaches in the dynamic parameter identification of complex engineering structures, offering a novel methodology for single and continuous dynamic monitoring of concrete dam structures.

To apply distributed optical fiber strain sensing (DOFSS) to the monitoring of rainfall-induced landslide deformation of landslide-prone soil slopes, this study proposes a deployment scheme and installation technique for strain-sensing optical fiber cables in soil slopes based on existing engineering experience and validates them through laboratory-scale physical model tests. The results show that direct burial of optical fiber cables in boreholes is a suitable installation method. A 2.0 mm tight-buffered optical fiber cable anchored with 5.0 mm heat-shrink-tube anchors was adopted as the fixed sensing configuration in the laboratory tests. In the laboratory model tests, the locations of strain peaks measured by the optical fiber cables were spatially associated with the visibly deformed regions, indicating that the strain anomalies can be used to approximately locate internal strain-concentration zones. The slope toe was identified as the location where localized failure was first observed, and the second rapid increase in strain at the slope toe may be regarded as a potential precursor of accelerated localized deformation under the tested condition. Borehole spacing has a significant influence on monitoring accuracy and is recommended to be controlled within 20–48% of the horizontal length of the potentially unstable slope zone. The variations in volumetric water content, earth pressure, and wetting front migration at different locations of the slope are highly correlated with rainfall infiltration and the evolution of slope surface erosion.

System reliability assessment of buried pipelines subjected to landslide hazards is computationally challenging due to high-dimensional spatial variability and complex nonlinear large-deformation responses. This paper proposes AK-SYS-SIS, integrating Sequential Importance Sampling (SIS) into the Active Learning Kriging framework for system reliability analysis. The method constructs an approximate optimal sampling density to adaptively guide candidates toward critical boundaries of multiple failure modes, while a composite learning strategy selectively updates only the most critical surrogate component. Two stochastic large-deformation finite element frameworks, RCEL-OLHS and GCRCEL-OLHS, are developed within the Coupled Eulerian-Lagrangian (CEL) technique to simulate earthquake-induced and rainfall-induced landslide–pipeline interactions, respectively. These frameworks incorporate Karhunen–Loève random field discretization, strain-softening constitutive models via Abaqus VUSDFLD/UMAT subroutines, modified Mohr-Coulomb plasticity, and soil rotating anisotropy, with GPU-accelerated random field generation for the rainfall scenario. Benchmark validations on series, parallel, and hybrid systems demonstrate that AK-SYS-SIS improves efficiency by 4–6 orders of magnitude over Monte Carlo Simulation with relative errors below 3%. In engineering applications, the method identifies system failure probabilities with fewer than 100 finite element evaluations, reducing computational cost by 94%–96.5% compared to direct simulation.

The deformation characteristics of rockfill materials and unexpected deformation exceeding projections during the service process of rockfill dams with face slabs are unclear. A discrete-continuous numerical simulation model for a rockfill dam with a face slab is developed to conduct mesoscopic analysis in locally representative areas within the dam body. The evolution mechanism of the macro- and mesoscopic mechanical characteristics of rockfill materials in different parts of a dam body throughout service is revealed. By comparison with continuous numerical simulation methods, the reliability of the discrete-continuous coupling simulation method and its advantages in the study of micromechanical properties were further verified. The micromechanical characteristics of the rockfill materials were analyzed. This study shows that the discrete-continuum coupling method aligns with the stress-strain calculation results of the continuous method, which is in line with engineering practice. With the application of a load, the particle position changes the distribution of the discrete domain structure, thereby affecting the macroscopic stress-strain state of the panel rockfill dam. This method holds significant theoretical importance for understanding the mechanical properties of rockfill materials and managing the deformation state of rockfill dams with face slabs.

The real-time and accurate prediction of dam deformation behavior is crucial for ensuring the safe operation of dams. Most existing models for predicting dam deformation behavior are built under the assumption of no missing data. Nonetheless, such an optimal scenario is seldom realized in practice, given the influence of subjective and objective considerations like human observational inaccuracies and malfunctions in monitoring instruments. This study introduces a framework designed for feature imputation and label prediction of dam deformation behavior at the same time in the presence of missing monitoring data. The novel approach entails formulating the issue utilizing a graph framework, where observations and features are denoted as separate nodes within a bipartite graph linked by edges that signify observed feature values. This formulation allows for treating feature imputation as a task of predicting missing edges, and label prediction as a task of predicting node values. These tasks are independently solved using Graph Neural Networks (GNN) without interference between them. The test results on two deformation datasets of an arch dam show that, the approach proposed in this study yields 50% lower mean absolute error for imputation tasks and 15% lower for label prediction tasks, compared with other advanced methods.

The deformation of arch dams represents a highly complex nonlinear process governed simultaneously by internal structural conditions and diverse external environmental factors, making accurate prediction and continuous monitoring indispensable for ensuring structural safety assessment. In the existing study, single prediction models or two-stage ensemble models exhibit significantly insufficient generalization capability under the coupled effects of complex environmental factors. Moreover, directly inputting the residual series into the model for correction often overlooks the dependency between residuals and environmental factors, leading to a correction process that lacks physical interpretability. Furthermore, dam deformation monitoring data are generally affected by instrumental noise and random environmental disturbances, which further reduces prediction accuracy. To address and overcome these limitations, this study introduces a multiorder machine learning fusion framework enhanced with environmentally influenced residual clustering correction. The proposed framework integrates several complementary algorithms, including the Sparrow Search Algorithm (SSA)-optimized Long Short-Term Memory (LSTM) network, a Whale Optimization Algorithm (WOA)-driven Variational Mode Decomposition (VMD) noise suppression module, and a K-means clustering-based grouping strategy for second-order residuals using key environmental variables such as water level and temperature. In the first stage, an SSA-optimized LSTM is utilized to generate first-order deformation predictions. In the second stage, WOA-optimized VMD decomposes the first-order prediction sequence into multiple modal components, and the component most correlated with the first-order prediction sequence is identified using Pearson correlation analysis. This selected component, when combined with environmental variables such as water level, temperature, and time, is subsequently input into an ELM for second-order prediction. In the third stage, K-means clustering is applied to classify second-order residuals under different environmental conditions, thereby revealing heterogeneity in residual patterns. Rolling predictions for each residual cluster are then performed using SSA-LSTM. By integrating residual predictions with second-order outputs, third-order deformation fitting values are obtained, enabling dynamic, adaptive, and highly precise monitoring. Application to a real-world concrete arch dam demonstrates that the proposed framework achieves superior accuracy, robustness, and adaptability across multiple modeling stages, offering a comprehensive and practical solution for deformation safety monitoring under complex and variable environmental conditions.