Pub Date : 2026-01-12DOI: 10.1016/j.tust.2026.107447
Zekun Li , Georgios Maragkos , Miaocheng Weng , Fang Liu , Bart Merci
Smoke propagation in passages with combined horizontal and inclined sections is examined through theoretical analysis and CFD simulations. The study explores the effects of heat release rate (HRR), passage geometry, and fire source location on buoyancy-driven flow, ventilation behavior, and upstream smoke flow, including heat outflow. Results show that geometric parameters such as ceiling height and elevation height play a crucial role in governing the smoke movement patterns. A critical induced airflow velocity, called ‘heat confinement velocity’, is identified, beyond which upstream heat outflow is effectively suppressed. Beyond the parameters mentioned above, it also accounts for the influence of fire position, especially near the lower entrance and near inclined sections. Additionally, a simplified theoretical model is established to estimate the induced airflow velocity, as well as a scaling relationship between this induced velocity and the heat confinement velocity, for HRR values above and below the critical value as classically defined for tunnels. Finally, a significant impact is demonstrated if the location of the fire is close to an inclined section of the passage, with much weaker upstream smoke flow and much stronger flow into the inclined section of the passage. These findings are useful for performance-based smoke control design in inclined and semi-inclined underground spaces.
{"title":"Numerical study of smoke movement and heat confinement under the influence of the stack effect in passages with horizontal and inclined sections","authors":"Zekun Li , Georgios Maragkos , Miaocheng Weng , Fang Liu , Bart Merci","doi":"10.1016/j.tust.2026.107447","DOIUrl":"10.1016/j.tust.2026.107447","url":null,"abstract":"<div><div>Smoke propagation in passages with combined horizontal and inclined sections is examined through theoretical analysis and CFD simulations. The study explores the effects of heat release rate (HRR), passage geometry, and fire source location on buoyancy-driven flow, ventilation behavior, and upstream smoke flow, including heat outflow. Results show that geometric parameters such as ceiling height and elevation height play a crucial role in governing the smoke movement patterns. A critical induced airflow velocity, called ‘heat confinement velocity’, is identified, beyond which upstream heat outflow is effectively suppressed. Beyond the parameters mentioned above, it also accounts for the influence of fire position, especially near the lower entrance and near inclined sections. Additionally, a simplified theoretical model is established to estimate the induced airflow velocity, as well as a scaling relationship between this induced velocity and the heat confinement velocity, for HRR values above and below the critical value as classically defined for tunnels. Finally, a significant impact is demonstrated if the location of the fire is close to an inclined section of the passage, with much weaker upstream smoke flow and much stronger flow into the inclined section of the passage. These findings are useful for performance-based smoke control design in inclined and semi-inclined underground spaces.</div></div>","PeriodicalId":49414,"journal":{"name":"Tunnelling and Underground Space Technology","volume":"171 ","pages":"Article 107447"},"PeriodicalIF":7.4,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145957106","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1016/j.tust.2026.107459
Zhihan Zhou , Yanhong Xi , Jun Mao , Guilan Yu , Wanki Chow
This study systematically investigates the fluctuating characteristics and temperature field distribution of carriage fires through a series of scaled-down (1:8) modeling experiments. The fire development process was observed to exhibit two distinct stages: a fuel-controlled stage characterized by monotonically increasing maximum temperatures with heat release rate, and a ventilation-controlled phase showing a decreasing trend in maximum temperature. Parametric studies revealed that the maximum temperature during the fuel-controlled phase is primarily governed by heat release rate and ventilation factor, while demonstrating a weak correlation with gradient. A power law relationship was established between dimensionless maximum temperature, dimensionless heat release rate, and wall heat loss. Furthermore, a temperature decay prediction model incorporating slope-opening coupling effects was developed, featuring two key coefficients: Coefficient , dependent on gradient (θ), exhibited a 12.8 % variation rate as θ increased from 0 to 0.0764, reflecting the enhancement of buoyancy-driven flow. Coefficient , determined by opening geometry (proportional to the 0.15th power of the opening height-to-width ratio), was attributed to strong turbulent mixing effects. By integrating the maximum temperature and decay models, a comprehensive temperature field prediction model was developed for inclined railway carriage fires with lateral openings (W/H: 0.45–4.55) and gradients (θ: 0–0.0764). The model demonstrated high accuracy, with prediction errors consistently below 10 %. This work provides an improved predictive framework for temperature fields in inclined railway carriage fires and offers a valuable theoretical foundation for train fire safety design.
{"title":"Coupled ventilation-slope effects on flame dynamics and temperature distribution in high-speed train compartment fires","authors":"Zhihan Zhou , Yanhong Xi , Jun Mao , Guilan Yu , Wanki Chow","doi":"10.1016/j.tust.2026.107459","DOIUrl":"10.1016/j.tust.2026.107459","url":null,"abstract":"<div><div>This study systematically investigates the fluctuating characteristics and temperature field distribution of carriage fires through a series of scaled-down (1:8) modeling experiments. The fire development process was observed to exhibit two distinct stages: a fuel-controlled stage characterized by monotonically increasing maximum temperatures with heat release rate, and a ventilation-controlled phase showing a decreasing trend in maximum temperature. Parametric studies revealed that the maximum temperature during the fuel-controlled phase is primarily governed by heat release rate and ventilation factor, while demonstrating a weak correlation with gradient. A power law relationship was established between dimensionless maximum temperature, dimensionless heat release rate, and wall heat loss. Furthermore, a temperature decay prediction model incorporating slope-opening coupling effects was developed, featuring two key coefficients: Coefficient <span><math><mrow><mi>α</mi></mrow></math></span>, dependent on gradient (θ), exhibited a 12.8 % variation rate as θ increased from 0 to 0.0764, reflecting the enhancement of buoyancy-driven flow. Coefficient <span><math><mrow><mi>β</mi></mrow></math></span>, determined by opening geometry (proportional to the 0.15th power of the opening height-to-width ratio), was attributed to strong turbulent mixing effects. By integrating the maximum temperature and decay models, a comprehensive temperature field prediction model was developed for inclined railway carriage fires with lateral openings (W/H: 0.45–4.55) and gradients (θ: 0–0.0764). The model demonstrated high accuracy, with prediction errors consistently below 10 %. This work provides an improved predictive framework for temperature fields in inclined railway carriage fires and offers a valuable theoretical foundation for train fire safety design.</div></div>","PeriodicalId":49414,"journal":{"name":"Tunnelling and Underground Space Technology","volume":"171 ","pages":"Article 107459"},"PeriodicalIF":7.4,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145957104","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1016/j.tust.2026.107458
Weisong Liu , Jun Zhang , Rui Ba , Weiguo Song
The guides can help pedestrians find exits, thereby improving evacuation efficiency and reducing casualties in emergency. This study investigates how guides affect crowd evacuations via controlled experiments and modeling. To investigate the influence of guide on crowd evacuation, a Social Force Model with Guidance (SFMG) has been established by embedding the guidance term into the social force framework. The interaction mechanism between a guide and pedestrians was studied by performing crowd movement experiments under different guidance modes (dynamic/static), movement speeds and crowd densities. The guidance attraction force formula involving the above variables has been proposed. It is revealed that the guidance attraction field is influenced by the guidance mode and speed. Subsequently, the simulations in a subway platform under varying visibility conditions were conducted and the influence of the initial layout of guides on crowd evacuation efficiency was studied. The results revealed that arranging guides in the areas far from the exit can facilitate the crowd evacuation. The distance between the guide’s initial position and exit was denoted as D. in the simulation analyses. A variable DR (distance ratio) calculated by the ratio of D to the length of the platform subzone was adopted to quantity guide’s initial position. Dynamic guide: The higher DR results in the shorter evacuation time under low visibility. But the optimal initial position of guide shifts slightly closer to the exit with increasing visibility. Static guide: The U-shaped relationship between evacuation time and DR is observed. The optimal position falls within the 40 %∼60 % DR. These findings are helpful to design indoor emergency guidance plan, and the optimal positioning rules are transferable to common building layouts.
{"title":"Mode-Dependent optimal positioning of evacuation Guides: An Experimental–Modeling study on static and dynamic guidance effect","authors":"Weisong Liu , Jun Zhang , Rui Ba , Weiguo Song","doi":"10.1016/j.tust.2026.107458","DOIUrl":"10.1016/j.tust.2026.107458","url":null,"abstract":"<div><div>The guides can help pedestrians find exits, thereby improving evacuation efficiency and reducing casualties in emergency. This study investigates how guides affect crowd evacuations via controlled experiments and modeling. To investigate the influence of guide on crowd evacuation, a Social Force Model with Guidance (SFMG) has been established by embedding the guidance term into the social force framework. The interaction mechanism between a guide and pedestrians was studied by performing crowd movement experiments under different guidance modes (dynamic/static), movement speeds and crowd densities. The guidance attraction force formula involving the above variables has been proposed. It is revealed that the guidance attraction field is influenced by the guidance mode and speed. Subsequently, the simulations in a subway platform under varying visibility conditions were conducted and the influence of the initial layout of guides on crowd evacuation efficiency was studied. The results revealed that arranging guides in the areas far from the exit can facilitate the crowd evacuation. The distance between the guide’s initial position and exit was denoted as <em>D.</em> in the simulation analyses. A variable <em>DR</em> (distance ratio) calculated by the ratio of <em>D</em> to the length of the platform subzone was adopted to quantity guide’s initial position. Dynamic guide: The higher <em>DR</em> results in the shorter evacuation time under low visibility. But the optimal initial position of guide shifts slightly closer to the exit with increasing visibility. Static guide: The U-shaped relationship between evacuation time and <em>DR</em> is observed. The optimal position falls within the 40 %∼60 % <em>DR</em>. These findings are helpful to design indoor emergency guidance plan, and the optimal positioning rules are transferable to common building layouts.</div></div>","PeriodicalId":49414,"journal":{"name":"Tunnelling and Underground Space Technology","volume":"171 ","pages":"Article 107458"},"PeriodicalIF":7.4,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145957105","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-10DOI: 10.1016/j.tust.2026.107449
Chengwen Wang, Xiaoli Liu, Weiqiang Xie, Yanlin Su, Yingtong Ju
The construction of twin shield tunnels has become increasingly prevalent in densely populated urban areas. Accurately predicting the surface settlement induced by twin-shield tunnelling is of great significance for risk mitigation and refined settlement control. This study proposes a novel intelligent approach that integrates numerical modelling, empirical formula, and automated machine learning (AutoML) to predict surface settlement troughs induced by twin-shield tunnelling. Using a well-validated numerical model that considered 11 input parameters (including geological, geometric, and operational factors), 2000 settlement trough datasets were generated through numerical modelling. Subsequently, an improved superposition method was applied to extract six characteristic control parameters of the settlement troughs, thereby constructing a high-quality dataset. A multi-output AutoML model was then developed to predict the control parameters of the twin-tunnel-induced settlement troughs. Compared with six conventional machine learning models and two classical ensemble strategies, the AutoML model exhibited superior predictive accuracy and generalization capability, achieving average R2 values of 0.9977 and 0.9835 for the training and test sets, respectively. The Shapley Additive Explanations (SHAP) method was employed to analyze the interpretability of the AutoML model. The results highlight the significant influence of construction parameters (e.g., tunnelling contraction ratio) on the maximum settlement, as well as the regulatory effects of geometric parameters (tunnel diameter, burial depth, and twin-tunnel spacing) on the shape of the settlement trough, thereby providing valuable guidance for design optimization and precise construction control. Finally, the proposed AutoML model was validated using five real-world engineering cases, where the predicted settlement troughs closely matched the measured data, thereby confirming the robustness, reliability, and practical applicability of the model and demonstrating its promising potential for engineering practice.
{"title":"Intelligent prediction of surface settlement troughs induced by twin shields tunnelling: Insights from a numerical modelling-empirical formulation-interpretable automated machine learning fusion method","authors":"Chengwen Wang, Xiaoli Liu, Weiqiang Xie, Yanlin Su, Yingtong Ju","doi":"10.1016/j.tust.2026.107449","DOIUrl":"10.1016/j.tust.2026.107449","url":null,"abstract":"<div><div>The construction of twin shield tunnels has become increasingly prevalent in densely populated urban areas. Accurately predicting the surface settlement induced by twin-shield tunnelling is of great significance for risk mitigation and refined settlement control. This study proposes a novel intelligent approach that integrates numerical modelling, empirical formula, and automated machine learning (AutoML) to predict surface settlement troughs induced by twin-shield tunnelling. Using a well-validated numerical model that considered 11 input parameters (including geological, geometric, and operational factors), 2000 settlement trough datasets were generated through numerical modelling. Subsequently, an improved superposition method was applied to extract six characteristic control parameters of the settlement troughs, thereby constructing a high-quality dataset. A multi-output AutoML model was then developed to predict the control parameters of the twin-tunnel-induced settlement troughs. Compared with six conventional machine learning models and two classical ensemble strategies, the AutoML model exhibited superior predictive accuracy and generalization capability, achieving average <em>R</em><sup>2</sup> values of 0.9977 and 0.9835 for the training and test sets, respectively. The Shapley Additive Explanations (SHAP) method was employed to analyze the interpretability of the AutoML model. The results highlight the significant influence of construction parameters (e.g., tunnelling contraction ratio) on the maximum settlement, as well as the regulatory effects of geometric parameters (tunnel diameter, burial depth, and twin-tunnel spacing) on the shape of the settlement trough, thereby providing valuable guidance for design optimization and precise construction control. Finally, the proposed AutoML model was validated using five real-world engineering cases, where the predicted settlement troughs closely matched the measured data, thereby confirming the robustness, reliability, and practical applicability of the model and demonstrating its promising potential for engineering practice.</div></div>","PeriodicalId":49414,"journal":{"name":"Tunnelling and Underground Space Technology","volume":"171 ","pages":"Article 107449"},"PeriodicalIF":7.4,"publicationDate":"2026-01-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928515","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-10DOI: 10.1016/j.tust.2026.107446
Jianming Zhang , Kebin Shi , Peixuan Lin , Lei Li , Qiyong Mao , Haibo Jiang , Xinjun Yan , Jingwei Gong
During the construction planning phase, accurately predicting the construction duration of long-distance tunnels built using Tunnel Boring Machines (TBM) is critical for optimizing construction organization and controlling costs. However, the uncertainty of geological conditions and the variability of tunneling efficiency pose challenges in making precise predictions during the planning phase. To address this issue, this study proposes a Monte Carlo model based on Latin Hypercube Sampling (LHS), incorporating the uncertainties in surrounding rock distribution and the evolution of tunneling efficiency. The prediction process is divided into two core stages. The first stage involves integrating borehole data and surrounding rock information obtained from preliminary geological surveys. Using a Markov chain corrected by Bayes’ formula, the uncertainty in geological spatial characteristics is continuously deduced. In the second stage, we first propose a tunneling efficiency decay factor (e) and couple it with the uncertainty in the surrounding rock distribution to establish simulation rules for the construction duration of long-distance TBM tunnels. Subsequently, the Monte Carlo method under LHS sampling is applied for the duration simulation. Finally, two targeted model transfer strategies are proposed to enhance the model’s applicability across different projects. The effectiveness of the proposed method was validated using the Xinjiang KS super‑long tunnel as a case study. The results demonstrated: (1) After considering the spatial distribution uncertainty of geological conditions and parameter e, the proposed model accurately forecasted the construction duration of long‑distance TBM tunneling, and the average prediction error was less than 4 days. Moreover, the model outperformed existing approaches in accuracy and robustness, and exhibited excellent stability and lower computational resource requirements. (2) Global sensitivity analysis indicated that uncertainty in surrounding rock distribution was the primary driver of duration fluctuations, and the proposed model effectively reduced the impact of this uncertainty on construction duration. Dynamic sensitivity further showed that as the excavation distance increased (beyond 6700 m), the sensitivity index of e reached 0.25–0.40, which significantly impacted construction duration. Furthermore, introducing e reduced the prediction error range by 76.47 %–95.83 %. (3) The proposed model exhibited good transferability, and the effectiveness of both model transfer strategies was demonstrated on the new project. This approach provides a valuable reference for predicting construction durations of long-distance TBM tunneling projects in complex geological conditions.
{"title":"Transferable prediction of TBM long-distance tunneling construction duration considering uncertainties in surrounding rock distribution and the evolution of tunneling efficiency","authors":"Jianming Zhang , Kebin Shi , Peixuan Lin , Lei Li , Qiyong Mao , Haibo Jiang , Xinjun Yan , Jingwei Gong","doi":"10.1016/j.tust.2026.107446","DOIUrl":"10.1016/j.tust.2026.107446","url":null,"abstract":"<div><div>During the construction planning phase, accurately predicting the construction duration of long-distance tunnels built using Tunnel Boring Machines (TBM) is critical for optimizing construction organization and controlling costs. However, the uncertainty of geological conditions and the variability of tunneling efficiency pose challenges in making precise predictions during the planning phase. To address this issue, this study proposes a Monte Carlo model based on Latin Hypercube Sampling (LHS), incorporating the uncertainties in surrounding rock distribution and the evolution of tunneling efficiency. The prediction process is divided into two core stages. The first stage involves integrating borehole data and surrounding rock information obtained from preliminary geological surveys. Using a Markov chain corrected by Bayes’ formula, the uncertainty in geological spatial characteristics is continuously deduced. In the second stage, we first propose a tunneling efficiency decay factor (<em>e</em>) and couple it with the uncertainty in the surrounding rock distribution to establish simulation rules for the construction duration of long-distance TBM tunnels. Subsequently, the Monte Carlo method under LHS sampling is applied for the duration simulation. Finally, two targeted model transfer strategies are proposed to enhance the model’s applicability across different projects. The effectiveness of the proposed method was validated using the Xinjiang KS super‑long tunnel as a case study. The results demonstrated: (1) After considering the spatial distribution uncertainty of geological conditions and parameter <em>e</em>, the proposed model accurately forecasted the construction duration of long‑distance TBM tunneling, and the average prediction error was less than 4 days. Moreover, the model outperformed existing approaches in accuracy and robustness, and exhibited excellent stability and lower computational resource requirements. (2) Global sensitivity analysis indicated that uncertainty in surrounding rock distribution was the primary driver of duration fluctuations, and the proposed model effectively reduced the impact of this uncertainty on construction duration. Dynamic sensitivity further showed that as the excavation distance increased (beyond 6700 m), the sensitivity index of <em>e</em> reached 0.25–0.40, which significantly impacted construction duration. Furthermore, introducing <em>e</em> reduced the prediction error range by 76.47 %–95.83 %. (3) The proposed model exhibited good transferability, and the effectiveness of both model transfer strategies was demonstrated on the new project. This approach provides a valuable reference for predicting construction durations of long-distance TBM tunneling projects in complex geological conditions.</div></div>","PeriodicalId":49414,"journal":{"name":"Tunnelling and Underground Space Technology","volume":"171 ","pages":"Article 107446"},"PeriodicalIF":7.4,"publicationDate":"2026-01-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928516","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-10DOI: 10.1016/j.tust.2026.107448
Mohammad Matin Rouhani , Jamal Rostami
Prediction of the cutting forces acting on disc cutters is essential for accurate modeling and performance assessment of hard rock tunnel boring machines (TBM). This study investigates two methodologies to enhance the quality of force prediction: the modification of the Colorado School of Mines (CSM) model and the application of advanced machine learning algorithms. The modified CSM model presents rock-type-specific formulas for sedimentary, metamorphic, and igneous rocks, utilizing dimensionless parameters including the Lame brittleness index, internal friction angle, and wave velocity ratio. Three state-of-the-art machine learning architectures, including SAINT (Self-Attention and Intersample Attention Transformer), TabNet, and TabM, are tested with hyperparameter optimization carried out using the Geometric Mean Optimization and Reptile Search Optimization algorithms. The modified CSM model shows statistically significant improvement compared to the original CSM model for all rock groups (p < 0.001), with the most notable enhancement for igneous rocks. Among the machine learning models, GMO-SAINT achieved the highest accuracy for normal force prediction (R2 = 0.98 for training and 0.96 for testing), while RSO-TabNet performs best for rolling force prediction (R2 = 0.94 for testing). SHAP analysis shows that tip width and cutting depth are the two primary factors that affect normal and rolling forces, respectively, while UCS consistently emerges as a secondary factor with all models. Overall, this combined methodology offers a more reliable cutting force estimation for improving TBM performance prediction.
{"title":"Predicting disc cutter forces for hard rock TBM cutterhead modeling: a comparative analysis of modified CSM semi-theoretical model and hybrid deep learning approach","authors":"Mohammad Matin Rouhani , Jamal Rostami","doi":"10.1016/j.tust.2026.107448","DOIUrl":"10.1016/j.tust.2026.107448","url":null,"abstract":"<div><div>Prediction of the cutting forces acting on disc cutters is essential for accurate modeling and performance assessment of hard rock tunnel boring machines (TBM). This study investigates two methodologies to enhance the quality of force prediction: the modification of the Colorado School of Mines (CSM) model and the application of advanced machine learning algorithms. The modified CSM model presents rock-type-specific formulas for sedimentary, metamorphic, and igneous rocks, utilizing dimensionless parameters including the Lame brittleness index, internal friction angle, and wave velocity ratio. Three state-of-the-art machine learning architectures, including SAINT (Self-Attention and Intersample Attention Transformer), TabNet, and TabM, are tested with hyperparameter optimization carried out using the Geometric Mean Optimization and Reptile Search Optimization algorithms. The modified CSM model shows statistically significant improvement compared to the original CSM model for all rock groups (p < 0.001), with the most notable enhancement for igneous rocks. Among the machine learning models, GMO-SAINT achieved the highest accuracy for normal force prediction (R<sup>2</sup> = 0.98 for training and 0.96 for testing), while RSO-TabNet performs best for rolling force prediction (R<sup>2</sup> = 0.94 for testing). SHAP analysis shows that tip width and cutting depth are the two primary factors that affect normal and rolling forces, respectively, while UCS consistently emerges as a secondary factor with all models. Overall, this combined methodology offers a more reliable cutting force estimation for improving TBM performance prediction.</div></div>","PeriodicalId":49414,"journal":{"name":"Tunnelling and Underground Space Technology","volume":"171 ","pages":"Article 107448"},"PeriodicalIF":7.4,"publicationDate":"2026-01-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928416","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-09DOI: 10.1016/j.tust.2025.107441
Haoran Wang , Chengchao Guo , DingFeng Cao , Jin Tang , Fuming Wang
In this study, a sensing-inversion method was proposed to investigate the mechanical response mechanisms of shield tunnels under heavy rainfall conditions, integrating displacement monitoring, distributed fiber optic sensing, and a strain–displacement-internal force recursive inversion method. Physical model tests were conducted to simulate interactions between heavy rainfall, soil strata, and tunnel structures. Laser displacement sensors and distributed optical fibers were used to monitor dynamic structural deformations and strains. An inversion model based on elastic foundation curved beam theory was developed to quantitatively analyze tunnel deformation evolution, load development mechanisms, and internal force distribution characteristics. The results indicate that the proposed inversion method improved accuracy by over 80% compared to conventional models and effectively captured radial displacements and internal force distributions. Under rainfall loading, the tunnel lining exhibited elliptical deformation and settlement, accompanied by compressive stresses at the crown and invert. The region of compressive stress expanded with increasing overburden thickness, whereas tensile stress developed at the haunches. The compressive stress at the crown exceeded that at the invert. When the tunnel was deeply buried, longer rainfall infiltration paths delayed structural responses to water penetration. Furthermore, deep overburden facilitated the dispersion localized stress concentrations in the lining caused by rainfall.
{"title":"New sensing-inversion integrated method for mechanical behavior analysis of shield tunnels during heavy rainfall","authors":"Haoran Wang , Chengchao Guo , DingFeng Cao , Jin Tang , Fuming Wang","doi":"10.1016/j.tust.2025.107441","DOIUrl":"10.1016/j.tust.2025.107441","url":null,"abstract":"<div><div>In this study, a sensing-inversion method was proposed to investigate the mechanical response mechanisms of shield tunnels under heavy rainfall conditions, integrating displacement monitoring, distributed fiber optic sensing, and a strain–displacement-internal force recursive inversion method. Physical model tests were conducted to simulate interactions between heavy rainfall, soil strata, and tunnel structures. Laser displacement sensors and distributed optical fibers were used to monitor dynamic structural deformations and strains. An inversion model based on elastic foundation curved beam theory was developed to quantitatively analyze tunnel deformation evolution, load development mechanisms, and internal force distribution characteristics. The results indicate that the proposed inversion method improved accuracy by over 80% compared to conventional models and effectively captured radial displacements and internal force distributions. Under rainfall loading, the tunnel lining exhibited elliptical deformation and settlement, accompanied by compressive stresses at the crown and invert. The region of compressive stress expanded with increasing overburden thickness, whereas tensile stress developed at the haunches. The compressive stress at the crown exceeded that at the invert. When the tunnel was deeply buried, longer rainfall infiltration paths delayed structural responses to water penetration. Furthermore, deep overburden facilitated the dispersion localized stress concentrations in the lining caused by rainfall.</div></div>","PeriodicalId":49414,"journal":{"name":"Tunnelling and Underground Space Technology","volume":"171 ","pages":"Article 107441"},"PeriodicalIF":7.4,"publicationDate":"2026-01-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928513","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-09DOI: 10.1016/j.tust.2025.107440
Junjie Liu , Qing Ai , Lulu Zhang , Junyi Zhu , Hui Wang , Xingchun Huang , Yong Yuan
Monitoring data from underwater tunnels are critical for operations and maintenance. However, they are often corrupted by noise from water level fluctuations, and the degradation process within them is difficult to extract, which limits the utility of these data. To address this issue, this study proposes a data-physics integration model for predicting tunnel convergence considering water level fluctuations and lining structure degradation. In the data-driven part, an improved Seasonal and Trend decomposition using Loess (STL) is developed to separate seasonal and trend components while accounting for gradual stiffness degradation of the tunnel lining, thereby producing more realistic time-variant seasonal component. In the physics-based part, a probabilistic degradation model is constructed on the modified rigid ring model, with parameters dynamically updated via a dynamic Bayesian network. By embedding the physics-based degradation model into the STL framework, the proposed approach enhances the prediction accuracy of trend component and strengthens physical interpretability. Comparative analysis using convergence monitoring data from a real underwater tunnel shows that, the proposed integration model achieves higher prediction accuracy and better captures the underlying degradation mechanism.
{"title":"Data-physics integration model for predicting tunnel convergence subject to water level fluctuations and lining structure degradation","authors":"Junjie Liu , Qing Ai , Lulu Zhang , Junyi Zhu , Hui Wang , Xingchun Huang , Yong Yuan","doi":"10.1016/j.tust.2025.107440","DOIUrl":"10.1016/j.tust.2025.107440","url":null,"abstract":"<div><div>Monitoring data from underwater tunnels are critical for operations and maintenance. However, they are often corrupted by noise from water level fluctuations, and the degradation process within them is difficult to extract, which limits the utility of these data. To address this issue, this study proposes a data-physics integration model for predicting tunnel convergence considering water level fluctuations and lining structure degradation. In the data-driven part, an improved Seasonal and Trend decomposition using Loess (STL) is developed to separate seasonal and trend components while accounting for gradual stiffness degradation of the tunnel lining, thereby producing more realistic time-variant seasonal component. In the physics-based part, a probabilistic degradation model is constructed on the modified rigid ring model, with parameters dynamically updated via a dynamic Bayesian network. By embedding the physics-based degradation model into the STL framework, the proposed approach enhances the prediction accuracy of trend component and strengthens physical interpretability. Comparative analysis using convergence monitoring data from a real underwater tunnel shows that, the proposed integration model achieves higher prediction accuracy and better captures the underlying degradation mechanism.</div></div>","PeriodicalId":49414,"journal":{"name":"Tunnelling and Underground Space Technology","volume":"171 ","pages":"Article 107440"},"PeriodicalIF":7.4,"publicationDate":"2026-01-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928514","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-08DOI: 10.1016/j.tust.2026.107445
Jingqi Cui , Shunchuan Wu , Haiyong Cheng , Xiaowei Hou , Jiaxin Wang , Weihua Liu , Chaoqun Chu
Accurate prediction of tunnel water inflow is critical for ensuring construction safety and risk control in tunnel engineering. However, traditional regression methods face significant challenges, including limited sample sizes, imbalanced data, complex feature interactions, and difficulty in engineering deployment. To address these issues, this study proposes an intelligent prediction framework that integrates data augmentation, model optimization, interpretability, and online deployment, and additionally possesses strong adaptability to dynamic field conditions. First, the SMOGN undersampling method is employed to balance and augment the training dataset, effectively expanding sparse samples and suppressing the influence of outliers, thereby enhancing the model’s generalization ability. Subsequently, LightGBM is improved through Optuna-based hyperparameter optimization and Analytic Hierarchy Process (AHP)-based feature weight adjustment, forming the AHP-OP-LightGBM hybrid model. This approach reduces prediction error by 15.89 % while aligning feature weights more closely with physical constraints. Compared with conventional optimization strategies, the model demonstrates superior capability in representing hydrogeological characteristics due to the dual mechanism of automated hyperparameter tuning and feature weight correction. Correlation analysis and SHAP-based interpretability further clarify the nonlinear synergistic mechanisms governing the coupled geomechanical-hydrological processes controlling tunnel water inflow. To support engineering application, a cloud-deployed online prediction system is developed using web technologies, integrating SHAP for transparent decision support. Additionally, an incremental learning module is incorporated to accommodate dynamic data variations. Validation using a small set of local incremental samples yields a maximum prediction error of only 1.9169 m3/h, demonstrating strong compatibility and accuracy across different engineering scenarios. Comparative experiments show that, on average, the proposed model reduces prediction error by 39.65 % and improves fitting accuracy by 18.43 % compared with traditional regression methods. Overall, this study provides a high-precision, interpretable, and generalizable intelligent solution for predicting tunnel water inflow under complex geological conditions.
隧道涌水的准确预测是保证隧道工程施工安全和风险控制的关键。然而，传统的回归方法面临着样本量有限、数据不平衡、特征交互复杂、工程部署困难等重大挑战。针对这些问题，本研究提出了一种集数据扩充、模型优化、可解释性和在线部署于一体的智能预测框架，并对动态现场条件具有较强的适应性。首先，采用SMOGN欠采样方法对训练数据集进行平衡和扩充，有效扩展稀疏样本，抑制离群值的影响，从而增强模型的泛化能力。随后，通过基于optuna的超参数优化和基于AHP （Analytic Hierarchy Process）的特征权值调整对LightGBM进行改进，形成AHP- op -LightGBM混合模型。该方法将预测误差降低了15.89%，同时将特征权重与物理约束更紧密地对齐。与传统优化策略相比，该模型具有自动超参数整定和特征权值校正的双重机制，具有较好的表征水文地质特征的能力。相关分析和基于shap的可解释性进一步阐明了控制隧道涌水的耦合地质力学-水文过程的非线性协同机制。为了支持工程应用，利用web技术开发了一个云部署的在线预测系统，集成了SHAP以提供透明的决策支持。此外，还包含了一个增量学习模块，以适应动态数据变化。使用一小组局部增量样本进行验证，最大预测误差仅为1.9169 m3/h，显示了不同工程场景的强兼容性和准确性。对比实验表明，与传统回归方法相比，该模型的预测误差平均降低了39.65%，拟合精度平均提高了18.43%。总体而言，本研究为复杂地质条件下的隧道涌水预测提供了高精度、可解释性和通用性的智能解决方案。
{"title":"An interpretable and adaptive tunnel water inflow prediction method using data augmentation and AHP-Enhanced OP-LightGBM","authors":"Jingqi Cui , Shunchuan Wu , Haiyong Cheng , Xiaowei Hou , Jiaxin Wang , Weihua Liu , Chaoqun Chu","doi":"10.1016/j.tust.2026.107445","DOIUrl":"10.1016/j.tust.2026.107445","url":null,"abstract":"<div><div>Accurate prediction of tunnel water inflow is critical for ensuring construction safety and risk control in tunnel engineering. However, traditional regression methods face significant challenges, including limited sample sizes, imbalanced data, complex feature interactions, and difficulty in engineering deployment. To address these issues, this study proposes an intelligent prediction framework that integrates data augmentation, model optimization, interpretability, and online deployment, and additionally possesses strong adaptability to dynamic field conditions. First, the SMOGN undersampling method is employed to balance and augment the training dataset, effectively expanding sparse samples and suppressing the influence of outliers, thereby enhancing the model’s generalization ability. Subsequently, LightGBM is improved through Optuna-based hyperparameter optimization and Analytic Hierarchy Process (AHP)-based feature weight adjustment, forming the AHP-OP-LightGBM hybrid model. This approach reduces prediction error by 15.89 % while aligning feature weights more closely with physical constraints. Compared with conventional optimization strategies, the model demonstrates superior capability in representing hydrogeological characteristics due to the dual mechanism of automated hyperparameter tuning and feature weight correction. Correlation analysis and SHAP-based interpretability further clarify the nonlinear synergistic mechanisms governing the coupled geomechanical-hydrological processes controlling tunnel water inflow. To support engineering application, a cloud-deployed online prediction system is developed using web technologies, integrating SHAP for transparent decision support. Additionally, an incremental learning module is incorporated to accommodate dynamic data variations. Validation using a small set of local incremental samples yields a maximum prediction error of only 1.9169 m<sup>3</sup>/h, demonstrating strong compatibility and accuracy across different engineering scenarios. Comparative experiments show that, on average, the proposed model reduces prediction error by 39.65 % and improves fitting accuracy by 18.43 % compared with traditional regression methods. Overall, this study provides a high-precision, interpretable, and generalizable intelligent solution for predicting tunnel water inflow under complex geological conditions.</div></div>","PeriodicalId":49414,"journal":{"name":"Tunnelling and Underground Space Technology","volume":"171 ","pages":"Article 107445"},"PeriodicalIF":7.4,"publicationDate":"2026-01-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928517","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Subsurface cavities in soft soil poses significant geotechnical challenges to the rapid expansion of urban underground spaces. This study presents a comprehensive framework for quantifying cavity-induced deformation amplification during deep excavation and develops an intelligent inversion system for cavity characterization using monitoring data. Through 1,800 finite element simulations, parametric analyses reveal that cavities located within 0.5 times the excavation depth (He) horizontally from the diaphragm wall and at depths of 1.5 He to 2.25 He constitute the most critical influence zone, amplifying horizontal wall displacement by up to 1.67 times and ground settlement by up to 2.2 times. K-means clustering analysis identifies five distinct settlement deformation patterns (Modes I–V) strongly correlated with cavity size and location. A convolutional neural network (CNN) based inversion model is developed to predict cavity dimensions and positions from deformation monitoring data, achieving over 85 % accuracy (R2 > 0.85) on test datasets. The model demonstrates robust performance under soil parameter uncertainties modeled with random fields, maintaining acceptable prediction accuracy when spatial variability is considered. This integrated framework provides a practical tool for real-time cavity detection and risk mitigation in deep excavation projects within cavity-bearing strata, offering valuable guidance for construction safety management in complex urban geological conditions.
{"title":"Intelligent identification and deformation analysis of subsurface cavities in deep excavations using CNN-based inverse modeling","authors":"Wei Zhang , Ya-Dong Xue , Jin-Zhang Zhang , Gang Zheng , Zeng-Zhi Qian , Yu-Xin Zhai","doi":"10.1016/j.tust.2025.107419","DOIUrl":"10.1016/j.tust.2025.107419","url":null,"abstract":"<div><div>Subsurface cavities in soft soil poses significant geotechnical challenges to the rapid expansion of urban underground spaces. This study presents a comprehensive framework for quantifying cavity-induced deformation amplification during deep excavation and develops an intelligent inversion system for cavity characterization using monitoring data. Through 1,800 finite element simulations, parametric analyses reveal that cavities located within 0.5 times the excavation depth (<em>H</em><sub>e</sub>) horizontally from the diaphragm wall and at depths of 1.5 <em>H</em><sub>e</sub> to 2.25 <em>H</em><sub>e</sub> constitute the most critical influence zone, amplifying horizontal wall displacement by up to 1.67 times and ground settlement by up to 2.2 times. K-means clustering analysis identifies five distinct settlement deformation patterns (Modes I–V) strongly correlated with cavity size and location. A convolutional neural network (CNN) based inversion model is developed to predict cavity dimensions and positions from deformation monitoring data, achieving over 85 % accuracy (R<sup>2</sup> > 0.85) on test datasets. The model demonstrates robust performance under soil parameter uncertainties modeled with random fields, maintaining acceptable prediction accuracy when spatial variability is considered. This integrated framework provides a practical tool for real-time cavity detection and risk mitigation in deep excavation projects within cavity-bearing strata, offering valuable guidance for construction safety management in complex urban geological conditions.</div></div>","PeriodicalId":49414,"journal":{"name":"Tunnelling and Underground Space Technology","volume":"171 ","pages":"Article 107419"},"PeriodicalIF":7.4,"publicationDate":"2026-01-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928518","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号