Information Technology in Geo-Engineering

Shield tunneling construction is a predominant method in metro infrastructure development, where the wear of tunnel boring machine (TBM) cutters is a critical issue that slows down its construction. The factors affecting cutter wear are multifaceted and complex. Traditional empirical formulas tend to oversimplify, being limited to specific engineering contexts and exhibiting low generalizability. Conventional cutter replacement is immensely time-consuming and labor-intensive. This study explored the application of machine learning techniques to predict cutter wear, utilizing six different models: Back Propagation Neural Network (BPNN), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), and Convolutional Neural Network (CNN). A comparative analysis is established to evaluate the suitability of these algorithms for this task. Data sourced from the 15.2m diameter Chunfeng Road Tunnel project in Shenzhen, China, was enriched by proportionally distributing cumulative wear across each tunnel ring, followed by outlier removal using boxplot methodology and normalization of the data. Hyper-parameter optimization, including grid search techniques, is applied to model optimization. The outcomes revealed the 1D-CNN model’s superior ability to predict cutter wear trends, yielding a test set R2 of 0.872 and MAPE of 0.077, outperforming other machine learning models. Additionally, in cumulative cutter opening sections, the 1D-CNN model only has three sections with error percentages exceeding the 30% margin, satisfying the accuracy requirements. Extensive experimental investigations revealed that the 1D-CNN is preferentially selected for this cutter wear prediction task, providing empirical insights beneficial for cutter replacement decisions in construction.

Navi Mumbai International Airport is being constructed by flattening a 92-m high Ulwe hill. Drilling & blasting has been adopted for the rock excavation to flatten the hill. The blasting operation at this site has various specific challenges, including the presence of residential houses, high-rise buildings, and electric lines nearby. It is imperative to control the ground vibration within safety limits when the blasting is done near the structures. This is done by devising the controlled blasting pattern, mostly devised using a predictive model. Genetic algorithm approach predicts the output parameters under various combinations of the input parameters. The predictive model is initially trained in this approach, and the fitness value is defined based on the expected vibration level. The various combinations of input parameters are chosen from a population generated using the genetic algorithm. Mutation and cross-over are further performed to maximize fitness. The developed predictive model for the construction site using this approach was accurate up to the acceptable level. Hence, the predictive model was further used to devise the controlled blasting pattern for the critical blasting operations at the site.

Leakage of water supply pipelines is a common cause of urban road collapse. A database containing 96 pipeline leakage cases in China has been established to study the correlation between pipeline leakage and road collapse. A series of statistical analysis on the distribution pattern and influence factors of road collapse was carried out, and water pipe diameter, pipe buried depth and the surrounding soil category were identified as three key factors affecting the road collapse probability. Then a rapid assessment model for road collapse was established based on the above cases through logistic regression analysis. The model was then applied to zoning the collapse susceptibility of roads in Yangpu District, Shanghai. The results show that the zoning matched well with the spatial distribution of local road collapse cases, indicating the effectiveness of the proposed model.

Theoretically expressed as the ratio of demand to capacity, the Factor of Safety (FoS) is defined to ensure the safety of structures and has been widely used in the deterministic design of underground excavations in rocks. However, FoS alone is inadequate to ensure the stability of tunnel openings. A measure of reliability or probability of failure is needed to account for the uncertainties in ground properties. This research tries to provide a connection between the probability of failure (Pf) and the Factor of safety (FoS), which is a very important guide for determining an appropriate target FoS to be used in tunnel design. A sensitivity analysis is conducted initially to determine which parameters have the most important effects on the determination of safety and reliability. The study considered five correlated randomized parameters, two limit state functions based on convergence and stress, and two generally used reliability methods, the First-order Reliability Method (FORM) and Monte Carlo Simulation (MCS). The result clearly indicated a negative relation between the Pf and the FoS. Moreover, distinct Pf vs. FoS curves were obtained for the two limit state functions, and the differences are attributed to the shapes of the equations and the variables involved. The Pfs for a displacement-based limit state is generally higher than the Pfs for a load-based limit state for the same FoS. This research offers valuable review and insights into a more pragmatic reliability-based design on UTIs by providing FoS for a given Pf.

National Highways is a major infrastructure owner in the UK. It owns and manages the strategic road network in England, comprising 4,500 miles of motorways and major A roads. It has a budget of £27.5bn for investment over a five-year period and is responsible for the construction, maintenance and operation of assets on the network, including the road pavement, structures, drainage and earthworks. This paper presents a risk-based approach to the prioritization of inspection of the earthworks assets using a data-driven approach. A risk-based method of prioritization enables National Highways to make informed decisions in an efficient and consistent manner. The method uses asset and operational data from different sources, as well as GIS algorithms and Python scripts to combine and process the data sets. The output is a consumable table that is used by geotechnical specialists to inform the frequency of inspections of the earthworks asset. The method has been kept deliberately transparent, simple and modular to ensure that it can be easily understood by those responsible for decision making and can adapt to changes in requirements.

In this paper, a novel particle tracking method is proposed to investigate the kinematics of Leighton Buzzard sand (LBS) particles exhibiting slight- and medium-level fragmentation by combining PointConv and PointNetLK networks. Firstly, a series of image processing algorithms were employed on the raw CT slices to facilitate reproduction of particle morphology. Subsequently, all particles were represented as point sets, down sampled and grouped, yielding substantial datasets for neural network training and testing. Additionally, Gaussian noise was generated and introduced into the particle point sets to enhance the network robustness. Then, the PointConv network was implemented to efficiently match the particles under different strains. This success was attributed to the good preservation of morphological features of these particles and the excellent capacity of PointConv to capture morphological features. Next, the PointNetLK network was trained and incorporated with the particle correspondence obtained by PointConv, aiming to determine the optimal transformation matrix for the corresponding particles. Finally, the predictive results are evaluated based on the visualization of the transformed point cloud, and the particle kinematics is analyzed.

The design of rock tunnels using drilling and blasting methods involves a multitude of interdisciplinary domains and exhibits a complex hierarchy of concepts. The related textual content spans diverse categories, presenting a vast amount of information with a rich and complex hierarchical structure, thereby escalating the difficulty in organizing knowledge. This study adopted an improved Latent Dirichlet Allocation (LDA) model to integrate and classify knowledge in the realm of tunnel support design. Considering the characteristics of relevant texts in the field of support design, the improved LDA model considering the semantic weights of words was proposed. Topic mining results show that under the semantically weighted LDA model, the model has higher consistency and lower confusion when the number of topics is between 10 and 14, so this study sets the number of text topics as 12. In addition, the consistency and confusion performance of the LDA model considering semantic weighting is better than that of the LDA model without semantic weighting. This study contributes to a more profound understanding of the professional terminology and concepts within the tunnel support design domain and aids in downstream tasks such as constructing domain ontologies.

The lighting environment significantly affects the drivers’ safety and comfort and vehicles’ carbon emissions in the tunnel. Due to the lack of direct quantitative descriptions of continuous driving behaviors in tunnels, previous studies failed to explain how the integrated lighting environment affects the driving strategies and traffic flow characteristics in the tunnel. Motivated by this gap, a VR-based experiment framework was proposed. This study selected the correlated color temperature (CCT) and road surface luminance, the two optical parameters affecting tunnel driving behaviors most, to distinguish the lighting environments. Nine virtual tunnels were created in Unity with three representative values for each parameter. Based on the Action Point Model (APM), the continuous driving behaviors in tunnels were discretized into sets of action points. Forty-four participants were asked to perform the virtual driving task, and the frequency, amplitude, and location of their action points were analyzed. Results show that high luminance and high CCT can lead to smoother driving. As drivers get closer to the lead vehicle, the impact of lighting environments on drivers weakens. Detailed experimental methodology, behavioral explanations underlying these findings, and practical implications are also discussed in the paper.

With the current growing size of cities, it is difficult to realize real-time inspections of fixed municipal facilities by manual inspections. There are various ways to realize automated inspections of municipal facilities, but these methods usually focus only on the state of the municipal facilities themselves, ignoring the negative impacts of peripheral obstacles on safety and availability. Therefore, this paper proposes a monocular camera-based algorithm for inspecting obstacles around municipal facilities. This algorithm estimates the pose of known inspection objects in the image, thereby inverting the spatial information. In addition, to reduce the cost of dataset acquisition, this paper also proposes an algorithm for generating a dataset for position estimation using Blender. The method in this paper can fully utilize the cameras on the existing inspection equipment to achieve the inspection of obstacles around municipal facilities without incurring additional hardware costs.

Integral abutment bridges (IABs) derive multiple benefits from the absence of deck expansion joints, including increased redundancy, resilience, and lower construction and maintenance costs. Nevertheless, thermally induced contraction and expansion of IABs’ superstructure results in the cyclic soil structure interaction that is largely responsible for developing detrimental bridge approach settlement. This study delves into the complexities of the soil-structure interaction by physically testing a downscaled model of the existing integral bridge, subjecting it to 100 cycles of thermal loading, and using digital imaging correlation (DIC), a technique originating in fluid mechanics to capture the evolution of the underlying deformation mechanisms that develop in the backfill soil behind the IABs’ abutments. The results of this study directly contribute to advancing the knowledge of thermally-induced soil structure interaction in IABs by emphasizing the relevance of the underlying granular soil failure and collapse for devising engineering solutions for improved performance of IABs.

In the realm of underground engineering development, industries such as rail transit, utility tunnels, railways, and highways emerge as a pivotal strategic focus for numerous nations. The information integrity within the design, construction, and operational phases of underground engineering projects remains a challenge, attributable to the sheer volume of information and the intricate array of information sources, thereby compromising the accuracy of information dissemination. In the paper, the current standard frameworks of rail transit, utility tunnel, railway, and highway engineering are then reviewed, and the different standard frameworks are compared with the National BIM Standard System (NBIMS) to analyze the current frameworks from the system integrity and standard level. The methodologies that could be employed in summarizing the underground engineering information modeling standard framework are scrutinized from the perspectives of technical and application standards. The rock tunnel engineering is exemplified to demonstrate the concept of formulating the rock tunnel engineering information modeling standard system, and the rock tunnel information is classified into five aspects: geographical geology, surrounding environment, design, construction, and monitoring.

In geo-engineering, it is crucial to understand the large and complex datasets for project success. Traditional textual and 2D visualizations methods are no longer adequate for complex geotechnical data. Our proposed system uses Large Language Models (LLMs) for extracting information and Augmented Reality (AR) for advanced visualization, making geotechnical data analysis and comprehension more intuitive. Utilizing LLM, the system can process the geo-engineering reports to identify and organize important information. The extracted information is then represented through AR in an interactive 3D format, achieving a more intuitive user experience. This paper explains our approach from methodology to practical evaluation, demonstrating the adaptability of LLM to various report structures and the AR integration for accessible visualization. Preliminary evaluations have been conducted on the performance of the system, which has positive outcomes on diverse reports. It indicates the capability of our proposed system to handle and present geotechnical data effectively. By combining LLMs and AR, we aim to provide a tool that can enhance decision-making and offer a better understanding of geo-engineering information, which is fundamental for the safety and efficiency of engineering projects.

The compressional tectonic setting between the Southern Tianshan and Western Kunlun Mountains results in dynamic geological activity in the Tarim Basin. Drilling operations face significant challenges due to the complexity of lithology and stress variations in the Ordovician. This study analyzes well diameter, cave morphology, whole rock mineral and physicochemical properties, and rock strength to investigate the mechanism of Ordovician wellbore instability. The results indicate that fractured carbonate rocks predominate in the Ordovician. These rocks have low linear expansion and high rolling recovery rates. Wellbore instability was caused by extrusion tectonics, fractured formation, and stress concentration. Fractured formations with weak cementation and loose structure are more susceptible to shear failure. Numerous high-angle fractures and weak planes can also contribute to drilling fluid leakage. Therefore, when encountering Ordovician carbonate formations, it is important to use drilling fluids that enhance sealing and consolidating properties to improve formation integrity and rock strength. We recommend increasing the density of the drilling fluid appropriately. Efforts should be made to simplify the drill string, optimize drilling parameters, and adopt a less drilling and more reaming approach when drilling through deep Ordovician carbonate rock to ensure downhole safety.

Rock fractures exhibit roughness that influences the mechanical and hydraulic properties of the rock mass, which are important to various underground projects. This paper aims to develop a method to estimate the Joint Roughness Coefficient (JRC) of rock fractures using terrestrial laser scanners (TLS) and deep neural network for quick and safe investigation. While TLS offers fast acquisition of 3D point cloud, errors are inevitably contained in the scan data. By analyzing scan data of a granite outcrop, the effect of these errors on JRC measurement is investigated. In addition, a deep neural network is trained to estimate the error-corrected JRC of the scan data. The performance of the developed estimator is evaluated, and its implication is guided as a result.

Recent studies on slope failures in the Midwestern states have shown a substantial difference between the measured strength at the design stage and the back-calculated strength of the failed slopes. The reduced strength may be related to the region’s geological history, and it is called the wet-drained-fully softened shear strength. Laboratory techniques applied to evaluate this strength are time-consuming and expensive. On the other hand, field techniques have not yet been appropriately developed to evaluate this strength for soils above the groundwater table. This study proposes an innovative CPT-based technique to evaluate the fully softened shear strength in the field by utilizing a special wetting cone to inject water into the soils at the desired depth and achieve soil disturbance resulting from the penetration of the cone. Field CPT test results from test locations showed a significant decrease in tip resistance and sleeve friction by wetting the soils, indicating that the innovative use of the CPT method could measure the fully softened strength.

This study aims to determine the stress and ground load in the initial tunnel lining excavated in weathered claystone. This is achieved by analyzing data from field monitoring using vibrating wire pressure cells and strain gauges. The study focuses on the construction of the Samarinda Tunnel, a 400-m-long shallow road tunnel with a cross-section of 12.8 m by 10.5 m. The tunnel is designed to accommodate two-lane one-way traffic from Alimuddin Street to Kakap Street in Samarinda City, East Kalimantan Province, Indonesia. The monitoring results indicate that the initial lining of the tunnel is safe, and its behavior appears stable. The load exerted on the lining varies from 3.6–7.3 kPa, much lower than Terzaghi’s rock load theory suggests. By examining the constructed P-M diagram, it has been determined that the steel rib can handle the maximum stress it experiences $$(\sigma_{\max } = 74.7\,{\text{MPa}})$$ ( σ max = 74.7 MPa ) , , and its factor of safety (FoS = 3.1) still exceeds the required value of FoS = 1.5 for temporary support.

Debris flows are one of the most dangerous types of mass movements due to their extremely rapid evolution, long travel distances, and significant impact forces. Structural measures, together with mapping and land use planning, are usually deployed to mitigate the effects of such events on relevant areas. These structures can also be used to collect relevant information regarding the occurrence of potentially critical events to activate appropriate countermeasures to ensure people’s safety. The monitoring system presented in this paper is designed to be installed on flexible barriers, a widespread intervention to deal with debris flows. The system is composed of a wide range of different sensors, following a multi-parameter approach, to validate the collected data and disseminate accurate information concerning the ongoing event. Exploiting IoT-based communication technologies and automatic procedures for data acquisition and elaboration allows for the identification of the impacts on the barrier with a real-time approach, activating appropriate actions to mitigate the related risk. Two different events are presented, involving a system installed near the A27 Highway in Northern Italy, highlighting the system’s reliability and effectiveness in detecting potentially critical events.

InSAR (Interferometric Synthetic Aperture Radar) is a microwave remote sensing technique that uses the phase shift of radar signals acquired at different timeframes to measure or monitor ground deformation. InSAR has many implications, such as monitoring ground deformation caused by natural- or geo-hazards, e.g., earthquakes, volcanoes, landslides, anthropogenic activities, groundwater pumping, underground mining, and hydrocarbon extraction. InSAR can also be utilized to study infrastructure displacements and environmental changes, such as monitoring changes in surface water level, mapping floods, soil moisture contents (at a shallow depth), and deforestation. The first significant application of SAR is the deployment of real-aperture radar interferometry to study the topography of the Moon in the early 1970s. However, InSAR was not widely used due to the limitations of computation capacity and the sparse available SAR data until the early 1990s. A major milestone for InSAR applications came in the 1990s when researchers used SAR data to measure ground deformation induced by the Landers Earthquake in California, and one of the publications landed on the cover of Nature magazine. This landmark achievement brought widespread recognition to the potential of InSAR for mapping ground deformation. Over the past two decades, the computation power and availability of SAR data have improved considerably with the launch of more satellites carrying SAR sensors. This paper presents a brief introduction to the history and fundamentals of InSAR, as well as case studies of its applications in the geo-engineering fields, including landslide displacement monitoring and underground excavation-induced ground subsidence mapping.

Wellbore stability is critical in oil and gas development, particularly due to shale hydration-induced instabilities. To address this challenge, this study introduces a novel laboratory methodology for real-time, spatially resolved monitoring of hydration-induced strains in shale cores using distributed fiber optic sensors. This technique offers a detailed mapping of hydration strain distribution over time, enhancing our understanding of hydration response characteristics. By employing distributed fiber optic sensing technology, the study overcomes the limitations of conventional monitoring methods and provides precise and efficient capture of the hydration process. The research goes beyond static observations by experimentally monitoring the impact of potassium chloride solutions with varying concentrations on the hydration-induced strain, shedding new light on evaluating drilling fluid performance. The findings reveal that the inhibitory properties of soaking solutions significantly influence shale hydration behavior, with higher inhibitory capacities slowing down the rate of hydration-induced expansion, as evidenced by reduced strain rates or lower strain amplitudes. Moreover, the non-uniform distribution of hydration strain across different monitoring points highlights the heterogeneous nature of shale surface hydration. The presence of rock fractures, particularly in Core 2, affects strain measurement and can lead to the observation of negative strains in the hydration strain waterfall plot. This information is crucial for interpreting the dynamic shifts in shale properties amid hydration and evaluating the inhibitive effectiveness of drilling fluids.

For over 30 years, digital advancements have transformed the world, and geoengineering is no exception. From readily available aerial data to machine learning-powered soil analysis, BIM design, and real-time construction monitoring, these technologies empower our field. However, innovations often follow a hype cycle, with inflated expectations giving way to disillusionment before, hopefully, reaching true productivity. This cycle presents strategic opportunities and challenges for academia and practitioners alike, depending on timing. While the digital transformation of geoengineering tools will undoubtedly continue, the key lies not in the technology itself but in efficiently harnessing its potential to solve the fundamental problems that make geoengineering so captivating.

Due to frequent climate-induced collapses in the railway track substructure, engineers must provide engineering solutions using more robust techniques beyond traditional fieldwork and GIS mapping. Machine learning (ML) techniques have recently been increasingly applied in geotechnical engineering. In this paper, we compare the performance of three algorithms: Particle Swarm Optimization (PSO), Bayesian Optimization (BOA), and Grid Search Optimization (GSO) for the autonomous evaluation of karst deformation along the railway track. Our assessment revealed the performance of the models as follows: PSO showed R2 = 0.959, RMSE = 1.265, and MAE = 0.882; BOA established R2 = 0.965, RMSE = 0.958, and MAE = 0.801; and GSO exhibited R2 = 0.950, RMSE = 1.389, and MAE = 0.976. These results indicate that BOA, having the highest R2 and minimal error values of RMSE and MAE, demonstrated the most optimal capacity for predicting karst deformation under the influence of large climate indices along the railway track.

Determining the tunnel portal boundary mileage is one of the key parameters in tunnel design. Accurately predicting the tunnel portal boundary mileage is significant for enhancing tunnel portal design efficiency. Methods for constructing two features based on discrete terrain point cloud data and tunnel position information are proposed in this study. Furthermore, two features are uniformly extracted along the tunnel design alignment, forming temporal evolution features. Moreover, three AI algorithms (LSTM, GRU, TCN) are selected to establish the prediction model of tunnel portal boundary mileage based on the optimal hyperparameters obtained through a five-fold cross-validation and grid search algorithm. The research results reveal that the predictive accuracy of the LSTM, GRU, and TCN network models on the training sets are 88%, 87%, and 55%, respectively. The tunnel portal boundary mileage prediction models established by the LSTM and GRU algorithms demonstrate better predictive performance. When applied to a new dataset, it is found that the predictive accuracy of the LSTM and GRU models is 86%. Therefore, the LSTM and GRU models are suitable for predicting the tunnel portal boundary mileage. The research offers valuable guidance for designers in selecting tunnel portal boundary mileage and provides a novel approach for predicting relevant parameters in tunnel design using AI technologies.

Recognizing water leakage is a crucial task in the daily operation of shield tunnels. Conventional machine learning methods applied in tunnel maintenance rely heavily on labeled images, a process that demands significant time and manual labor. Self-supervised learning (SSL) offers a cost-effective solution by enabling training with a limited amount of labeled data. In this study, we propose a novel SSL model, namely, SSRecNet, to recognize water leakage using a small number of labeled images. Initially, unlabeled images are employed to pretrain the feature extraction process, which is achieved through the restoration of noisy images. Subsequently, a small number of labeled images are utilized to fine-tune the feature extraction and train the classifier. Finally, ablation experiments are conducted to assess the impact of pretraining on enhancing the accuracy of the proposed SSL model. The outcomes demonstrate that SSRecNet attains a commendable accuracy rate of 100% on the training set and 81.75% on the test set. Ablation experiments confirmed that pretraining the feature extraction by SSRecNet can significantly enhance the model’s performance.

A comprehensive understanding of TBM cutter wear is crucial for formulating effective boring plans and determining optimal intervals for cutter replacement. Cutter wear is a complex and non-linear phenomenon influenced by various factors. This study proposes a data-driven model for predicting TBM cutter wear values, incorporating TBM operational parameters, mechanical characteristics, and geological factors. Utilizing data from the Shenzhen Metro Chunfeng Road project, a dataset comprising ten parameters and 6445 data was curated, with the Transformer model employed as the predictive framework. The findings reveal that the Transformer model achieves a mean squared error (MSE) of 0.004, mean absolute error (MAE) of 0.0313, and correlation coefficient (R2) of 0.8544. Comparative analysis with the Long Short-Term Memory (LSTM) baseline model indicates the superior performance of the Transformer model across evaluation metrics. This suggests that the Transformer model exhibits enhanced capabilities in a data-driven context. The ability to predict cutter wear for the subsequent time interval holds significance for guiding construction activities.

The exploitation of underground mineral deposits is dominantly accomplished using drilling and blasting. The large-scale blasting with higher explosive consumption is practiced to achieve the production targets. However, such large-scale blasting induces ground vibration, a hazard to structural safety and environmental nuisance. The magnitude of ground vibration can be reduced by devising a controlled blasting pattern. Such a pattern is devised by developing the ground vibration predictive models. Machine learning based on the Random Forest predictive model was developed in this study to predict the ground vibration induced by underground blasting. The number of trees in this predictive model influences prediction accuracy and computational time. Accordingly, it was optimally selected to maximize the R2 and minimize the RMSE values. Once the predictive model was developed, the predictions were made for the testing data set. The predictions were also made using various empirical relationships. The comparison of the predictive models shows that RF is the best predictive model with R2 and RMSE values of 0.94 and 0.438 mm/s, respectively. This is because the RF-based predictive models consider interdependency among the parameters, whereas the empirical relationships directly use the input parameters to predict ground vibration. Because of the better accuracy of the RF predictive model, it is being used for day-to-day ground vibration prediction at the experimental site.

With emerging data-intensive technologies, industry automation has become promising in different fields, including the construction sector. Reinforcement learning has been applied to optimize conventional tunneling processes to minimize instabilities and excavation time. This study aims to take advantage of offline reinforcement learning through the soft actor-critic method, in which policies are evaluated and improved with offline datasets of the transitions occurring within the environment. The proposed method shows capabilities for encouraging exploration while generating actions, minimizing instabilities during the excavation, and allowing the transfer of this knowledge to different tunneling environments.

The Konkan Railway, excavation work completed in 1998 and serving India since then, faces climate-related challenges such as heavy monsoon rains, rockfalls, and slope failures that threaten railway operations in regard of train accidents and traffic interruptions. While the railway has implemented various geotechnical strengthening measures, concerns arise as climate change leads to more extreme weather events. Flattening slopes helps improve the stability of the railway track by reducing the risk of landslides, soil erosion, and boulder failures. A flatter slope contributes to safer and smoother train operations and is less prone to geological hazards. Railway infrastructure management involves numerous challenges, particularly in ensuring the stability and safety of slopes adjacent to the tracks. This research focuses on applying machine learning (ML) techniques in slope reconstruction works using controlled blasting to address these challenges. Blast-induced slope failure or rockfall during blasting operations are critical concerns that may interrupt normal traffic operations. They need accurate prediction for effective risk mitigation and rescheduling of traffic operations. Initially, at three locations of Konkan Railways, ML models have been developed using 490 datasets with the most inputs out of thirteen parameters using multicollinearity and Logistic Regression techniques based on minimum Akaike Information Criterion (AIC) values. This paper narrates the use of the best-trained Random Forest model for the rapid assessment of such failure during day-to-day blasts at two slope reconstruction sites of the Konkan Railway.

Due to the poor adaptability of TBM to geological formations, there are problems in actual construction, such as unclear mechanisms and laws of rock-machine interaction, insufficient exploration and perception lag of geological conditions, and strong subjectivity in equipment excavation parameter control. This article is based on the TBM excavation dataset that is publicly available for the Yinsong Project. The main monitoring parameters during the TBM excavation process are divided into three categories: control parameters, state parameters, and performance parameters. The outliers are identified and removed using the Turkey box plot, and then the data is dimensionally reduced using the APCA method. On this basis, a rock machine interaction model (i.e. the relationship between the three types of parameters) for the TBM excavation process was established using the random forest algorithm that integrates Bayesian optimization methods. Finally, with TBM advance speed and specific energy as optimization objectives, the main control parameters of TBM were intelligently optimized using the multi-objective genetic algorithm NSGA II. After optimization, the specific energy was reduced by 5.55%, and the advance rate was increased by 65.79%. The conclusion shows that the above optimization method can achieve efficient and intelligent TBM excavation and has certain engineering application prospects.

Accurate prediction of wall deflection induced by braced excavations is of great importance to prevent potential damage to retaining structures and the surrounding environment. Although previous studies have utilized spatiotemporal deep learning algorithms for this purpose, they overlooked the underlying deflection mechanism when developing these algorithms. To address this limitation, this paper proposes a novel physics-guided architecture of neural networks (PGA-NN) for predicting wall deflection where the physical mechanism is hardcoded in the neural network architecture. In the PGA-NN, a physical intermediate layer and a monotonicity-preserving long short-term memory (LSTM) cell are integrated through understanding the underlying physical mechanism of wall deflection. The performance of the proposed PGA-NN model was verified using data from an excavation project. Results indicate a significant superiority of the proposed PGA-NN model over the baseline model. The PGA-NN model demonstrates a strong capability for accurately forecasting wall deflections in advance by considering the physical mechanism.

Discontinuity recognition is important for the description of rock discontinuities. The accuracy and efficiency of traditional recognition methods based on 3D point clouds are often affected by manual parameter settings during the process. This paper proposes a deep learning intelligent discontinuity recognition method based on a fast, maximum consistent color mapping of the point cloud. The main idea is to quickly assign different colors to different discontinuity planes according to the normal vector while keeping the color within the same discontinuity plane as uniform as possible. Then, the recognition of discontinuities in 3D point clouds is converted to the recognition of discontinuities with different colors in 2D images. Sharp points of discontinuities are also detected to facilitate the recognition. The Mask R-CNN neural network is used to train and recognize discontinuities in 2D images. Finally, the recognized 2D discontinuities are mapped back to the 3D point cloud. The results show the proposed method can achieve intelligent recognition of rock discontinuity with high efficiency and full automation without manual intervention during the process, and the recognition effect is more consistent with manual judgment than traditional methods.

Physics-informed neural network (PINN), featuring its merits in surrogate modeling and inverse analysis, provides a new paradigm for geo-engineering. PINN has been successfully applied in the simulation of continuum fields, such as solid mechanics, fluid mechanics, and thermodynamics. However, its application in fracture modeling also needs more attention since fracture simulation is an important and difficult problem in both scientific and engineering domains. This study adopts the partial differential equation of the phase-field fracture model as the physical constraint of PINN. We develop a staggered scheme for phase-field fracture modeling in 1D scenarios. Two networks are adopted for approximating fracture phase field and displacements, respectively. The results corroborate its correctness with analytical solutions. Losses of two neural networks are compared, and it is found that this multi-network method can effectively circumvent the problem of pathological gradients. It is also more promising for more complex problems such as multiphysics fracturing cases.

The key to improving the geological adaptability of TBM is to understand the interaction between the rockmass and the TBM operational data. Since the cutterhead loads, i.e., torque and thrust, are crucial for the safe operation of TBM, a deep learning model based on the two-stage attention mechanism has been established to explore the synergy effects of cutterhead loads with the changes of control parameters and rockmass during TBM excavation. The results show that (1) the proposed model predicts thrust and torque with R2 of 0.86, 0.75, and MRE of 0.25 AND 0.31, respectively. (2) Thrust and torque are highly dependent on the level of TBM cutterhead rotation speed and penetration rate. (3) The rockmass condition contributes about 10 percent to the accuracy improvement of the cutterhead loads prediction model.

The study investigates the interpretability and explainability of machine learning models in geotechnical engineering, focusing on case studies on rock mass quality and rock type predictions from drilling data. By applying permutation feature importance, Shapley values, direct model interpretation, and partial dependency plots, features such as overburden thickness, tunnel width, feeder pressure, and penetration rate were identified as significant influencers on model predictions. These insights ensure the models’ predictions are consistent with geotechnical physical expectations, thus improving their practical applicability. The research underscores the necessity for transparent and interpretable machine learning solutions in geotechnical engineering, enhancing the reliability and application of such models in critical construction operations.

Uneven settlement and associated issues like joint water leakage in shield tunnels operating in soft clay are prevalent, significantly impacting the safety of tunnel operations and the structural service life. The accuracy of idealized theoretical models in predicting the uplift-settlement of shield tunnels is limited. This paper focuses on the calculation method for the above-ground uplift of the overlying shield tunnel caused by grouting. It conducts parameter analysis for expansion, strata, and tunnel parameters regarding the longitudinal deformation calculation method. The study establishes a database of final uplift values for shield tunnels, utilizes the Keras deep learning framework to develop an LSTM model, and applies it to predict the uplift in a grouting case study in Ningbo Metro 2. The predictive analysis results demonstrate that the predicted curve closely matches the measured values, indicating a certain level of effectiveness. The peak value of the calculated curve is more consistent with the measured than predicted. Outside the grouting range, the predicted results are slightly higher than the measured values, while the calculated results are slightly lower. Considering both the predicted and calculated results better reflects the actual situation.

This study aims to identify hidden patterns within seismic data collected during tunnel excavation using a mini-TBM in the laboratory. Seismic data in the form of acoustic emissions were recorded and analyzed, with features extracted from both time and frequency domains. Four distinct unsupervised machine learning techniques - namely, K-means, Density-based spatial clustering of applications with noise (DBSCAN), Gaussian mixture model (GMM), and Hierarchical clustering - were employed for analysis, and their outcomes were compared. The findings illustrate that while each algorithm presents advantages and drawbacks in uncovering patterns, Hierarchical clustering generally yields superior results compared to the others. Despite being computationally expensive, Hierarchical clustering offers notable advantages, such as robustness to outliers and noise, making it a reliable choice for clustering tasks, particularly in detecting hidden patterns within seismic data during TBM excavation. The methodology adopted in this research holds the potential for assessing the performance of TBMs in real-world tunnel excavation projects.

It is beneficial to know the soil type on the go during excavation to ensure optimal Earth pressure balance tunnel boring machine (EPBM) operation. This encouraged many researchers to perform data-driven analyses to develop prediction models. These models utilize the complex mapping hidden in the operational data collected by TBM sensors that relate them to the excavated ground condition. However, why data-driven prediction models such as machine learning (ML) can capture the ground condition using the operational parameters is a question that is often unanswered due to their highly complex or black-box nature. On the other hand, more quality ground-truth data is often needed along the tunnel alignment to ensure optimal training for the supervised ML algorithms. In such cases, a semi-supervised algorithm can provide a better generalization of the relationship between operational parameters and excavated tunneling ground. To this end, this research attempted to investigate the behavior of the operational parameters while they are encountering different ground conditions based on the predictions made by a successfully developed semi-supervised machine learning model to identify soil type during EPBM operation.

The subsurface profile is usually mapped based on discrete borehole logs, geophysical investigation, and cross-borehole techniques. The geological profile is organized after making several assumptions and does not necessarily capture the ground's natural variability. Increasing the borehole density impacts the project's economics. The actual ground profile often varies from the predicted profile, leading to hazards during tunneling. In the case of a tunnel boring machine (TBM) excavated tunnel, probing could be deployed to understand the geometrical aspects that are about to be encountered at the tunnel face with the advancement. However, the probing data might not always represent the geometrical within the tunnel envelope. Researchers have proposed various machine learning (ML) based models in the last two decades to predict the ground geology in front of the TBM face. This paper attempts to review such models for different types of TBM, including earth pressure balance (EPB), slurry shield, hard rock, open-face, and dual mode. The first of this article summarizes the significant contributions published related to ML in TBM tunneling while critically evaluating the applicability of each mode. The second part of this proposal proposes a new supervised ML model that can be used in tunneling in different rock masses and applies to all the TBM types. TBM's continuously monitored parameters like face pressure, torque, revolutions per minute (rpm), advance rate, and total advance are considered inputs to the model. Such parameters, along with the extreme gradient boost (XGBoost) method, are utilized to understand the geotechnical characteristics of the material ahead of the tunnel face. The dataset will be used with input parameters and target variables as engineering properties of geological units, such as uniaxial compressive strength and porosity at and ahead of the tunnel face. The proposed model is expected to enhance the overall accuracy and interpretability of the predictions.

The disc cutter is the most frequently damaged and consumed component in the TBM boring process. Especially under large sections of soft and hard composite strata, the frequent changes in the cutter load cause accelerated wear of the cutter, affecting the construction efficiency and cost. Thus, rapid and precise perception of cutter wear contributes to optimizing cutter replacement and operational strategies. This study proposes an intelligent prediction framework for cutter wear, considering the spatial movement properties of TBM in composite strata. Firstly, the cutter movement distance was calculated by a geometric relationship, and then a standardized dataset was established by the tunneling data preprocessing. Subsequently, a cutter wear prediction model, DCSPN, is proposed based on the cutting properties. Taking a project in China as a case study, the proposed model showed a prediction relative error of more than 80% and an R2 of more than 0.85 in the testing set. The results show that the proposed framework improves the perception accuracy of disc cutter wear in composite strata.