Rockbursts represent a critical dynamic hazard in deep tunnel construction; however, the scarcity of labeled data poses significant challenges for accurate predictions. Hence, we reformulate the conventional four-class rockburst classification task into a binary-classification problem. A comprehensive rockburst dataset was compiled based on an extensive literature review. Six machine-learning algorithms-support vector machine (SVM), K-nearest neighbor (KNN), decision tree (DT), multilayer perceptron (MLP), random forest (RF), and extremely randomized trees (ETs)-were implemented and evaluated across multiple feature set configurations. The results are as follows: (1) feature set selection substantially affects predictive accuracy, with higher-dimensional feature combinations yielding superior performance; (2) ensemble methods(RF and ETs) outperform SVM and MLP by reducing variance and enhancing generalization on complex rockburst data; and (3) the binary-classification framework consistently outperforms the conventional four-class scheme, achieving accuracies above 0.80 by simplifying decision boundaries and reducing interclass ambiguity. These findings contribute to the development of a real-time online framework for rockburst risk prediction and offer valuable insights into proactive hazard mitigation in underground engineering.

Shield tail grease is crutial for the safety of shield tunneling in coastal urban areas, yet its performance degradation under seawater intrusion remains insufficiently studied. In this study, grease samples were subjected to immersion in freshwater and seawater at various concentrations. A series of laboratory tests were conducted, including rheological, adhesion, and cone penetration measurements. The effects of sea salt content and immersion time on grease properties were systematically studied, and microstructural changes under different conditions were compared using scanning electron microscopy (SEM). The results indicate that although seawater intrusion do not alter the rheological model of the grease (Herschel-Bulkley model) at 25 °C, it reduced the peak shear strength by 9.52%. With prolonged immersion time, the yield stress of the grease increases by 4.5% to 22.3%, while the viscosity coefficient and adhesion decreases by 4.5% to 14.4% and 85.2% to 87.2%, respectively. The cone penetration value exhibits a trend of initial increasing and then decreasing. At a sea salt content of 6.0%, the grease reaches its peak yield stress and adhesion, while cone penetration reaches minimum value. Microstructural analysis reveals that as immersion time increases, the grease porosity first decreases and then increases. This study provides a theoretical basis for understanding the performance degradation mechanisms of shield tail grease in marine environments, and offers practical guidance for ensuring tail seal safety.

With the rapid development of deep underground engineering (e.g., deep mining, geothermal exploitation, and high-level radioactive waste disposal), high-temperature granite in deep environments is often subjected to complex thermo-mechanical coupling effects, thus making its mechanical properties and failure mechanisms critical to engineering safety. Hence, to evaluate these aspects of high-temperature granite in uniaxial compression tests, we developed a new thermo-mechanical coupling model based on cohesive zone model in ABAQUS software. The numerical model not only solves the problem of heat transfer between cohesive elements, but also represents the grain composition of granite, and adhesion and occlusion between grains. We verified the effectiveness of the model by comparing its results with those of laboratory experiments. We used this numerical model to study the effects of mineral-grain size and boundary strength on the mechanical parameters, failure mode, and microfracture characteristics of high-temperature granite. With increasing mineral-grain size, both peak stress and peak strain first decreased, then increased, and finally decreased; the failure mode changed from tensile-shear mixed failure to shear-dominant one at temperatures of 150, 300, and 450 °C, and the proportion of intergranular cracks decreased. With an increase in the boundary strength, both peak stress and peak strain increased; the main failure cracks of granite became concentrated in the local range; and the proportion of intergranular cracks decreased constantly.

The deformation of joints in shield tunnels poses significant risks for tunnel safety, serviceability, and durability. Compared to circumferential joints, the deformation of longitudinal joints is more important for tunnel safety assessment. However, traditional tunnel inspection methods focus only on the overall convergence and dislocation between rings, neglecting the opening of longitudinal joints. Moreover, accurate detection of longitudinal joints is difficult, especially with visual data collected from field tunnels, because non-structural elements, such as cables and pipelines, often present similar visual features to longitudinal joints and thus interfere with their detection. The proposed method employs an advanced multi-modal deep learning model (an enhanced YOLOv8n, augmented with a Spatial-to-Depth (SPD) module and a Convolutional Block Attention Module (CBAM), and optimized with a grayscale-depth fusion strategy to robustly detect longitudinal joints in images, even under interference from non-structural elements and challenging lighting conditions. Once the joints are accurately identified and precisely localized in the image domain, the spatial correspondence between the images and point cloud data is established to enable boundary point extraction and coordinate transformation. This process isolates the relevant point cloud data associated with the detected joints and facilitates accurate calculation of the joint opening. Field experiments validated the effectiveness of the method in accurately locating longitudinal joints and quantifying their openings in an operational tunnel environment. The proposed method fills a critical gap in the current tunnel structural safety evaluations by offering an accurate and effective means of evaluating the deformation of longitudinal joints. This approach improves the accuracy and reliability of tunnel deformation assessments, and provides more critical and detailed data for tunnel structural safety evaluations. Future work will focus on refining the joint detection and boundary point extraction methods to address more complex tunnel environments and defects, as well as the need for comprehensive error analysis and uncertainty quantification.

Soil classification and model uncertainty significantly affect the accuracy of seismic liquefaction discrimination models. In this study, we developed a Bayesian Logistic Regression (BLR) model for soil classification and a Bayesian Adaptive Least Absolute Shrinkage and Selection Operator Logistic Regression (BALASSO-LR) model for liquefaction discrimination based on a cone penetration test with pore pressure measurement data. We aimed to evaluate the effects of two classification strategies, i.e., one-versus-one (OvO) and one-versus-rest (OvR), on the performance of a BLR model. Additionally, we evaluated the influence of liquefaction-related factors, including the soil behavior type index, I_c, and performed model uncertainty analysis to enhance the predictive reliability of the BALASSO-LR model. These models were then compared with the conventional simplified soil behavior classification method and the existing logistic regression (LR) models for liquefaction prediction. The results indicated that the BLR soil classification model using the OvR scheme achieved a prediction accuracy of 81.2%, representing improvements of 16.2% and 1.9% over the conventional Soil Behavior Type (SBT) method and the OvO scheme-based BLR model, respectively. The BALASSO-LR model for liquefaction discrimination achieved an accuracy of 84.1%. Omitting the soil classification index I_c decreased accuracy by 2.3%. When compared with the three simplified methods and existing LR models, BALASSO-LR exhibited an improvement of 5.7%-11.4% in the accuracy. In the uncertainty analysis, the No-U-Turn sampler (NUTS) algorithm with a prior distribution of N∼ (0,100) achieved the highest accuracy (84.1%), surpassing that of the Metropolis algorithm by 6%.

With the accelerating pace of urbanisation, urban flooding disasters are occurring with increasing frequency due to extreme weather events. Subway stations are particularly vulnerable to such flooding, where water ingress can lead to severe casualties and significant property losses. However, the spatiotemporal characteristics of water inundation in multi-level stations and their connecting tunnels remain inadequately understood. In this study, a scaled physical model comprising two multi-level subway stations and two tunnel segments was developed, and similarity ratios for the key parameters relevant to subway flooding experiments were derived. A series of physical experiments were conducted under varying water inflow rates to explore the entire process of station inundation and to analyse the propagation dynamics of floodwater through the stations and tunnels. The inundation process was categorised into three distinct stages: energy release diffusion, stable propagation, and cumulative rise. The multi-stage characteristics of the water level rise process were identified, including a previously unreported water level chasing effect occurring in the platform level of the connected station. A multi-stage quantitative model was developed to characterise the water level rise in multi-level stations and tunnels, and its accuracy was validated through physical experiments. These findings provide valuable references for flood disaster monitoring and early warning systems, personnel evacuation planning, and drainage system design in subway systems.