Effectiveness of predicting tunneling-induced ground settlements using machine learning methods with small datasets

Prediction of tunneling-induced ground settlements is an essential task, particularly for tunneling in urban settings. Ground settlements should be limited within a tolerable threshold to avoid damages to aboveground structures. Machine learning (ML) methods are becoming popular in many fields, including tunneling and underground excavations, as a powerful learning and predicting technique. However, the available datasets collected from a tunneling project are usually small from the perspective of applying ML methods. Can ML algorithms effectively predict tunneling-induced ground settlements when the available datasets are small? In this study, seven ML methods are utilized to predict tunneling-induced ground settlement using 14 contributing factors measured before or during tunnel excavation. These methods include multiple linear regression (MLR), decision tree (DT), random forest (RF), gradient boosting (GB), support vector regression (SVR), back-propagation neural network (BPNN), and permutation importance-based BPNN (PI-BPNN) models. All methods except BPNN and PI-BPNN are shallow-structure ML methods. The effectiveness of these seven ML approaches on small datasets is evaluated using model accuracy and stability. The model accuracy is measured by the coefficient of determination (R²) of training and testing datasets, and the stability of a learning algorithm indicates robust predictive performance. Also, the quantile error (QE) criterion is introduced to assess model predictive performance considering underpredictions and overpredictions. Our study reveals that the RF algorithm outperforms all the other models with the highest model prediction accuracy (0.9) and stability (3.02 × 10^?27). Deep-structure ML models do not perform well for small datasets with relatively low model accuracy (0.59) and stability (5.76). The PI-BPNN architecture is proposed and designed for small datasets, showing better performance than typical BPNN. Six important contributing factors of ground settlements are identified, including tunnel depth, the distance between tunnel face and surface monitoring points (DTM), weighted average soil compressibility modulus (ACM), grouting pressure, penetrating rate and thrust force.     

Comparison of machine learning methods for ground settlement prediction with different tunneling datasets

This study integrates different machine learning (ML) methods and 5-fold cross-validation (CV) method to estimate the ground maximal surface settlement (MSS) induced by tunneling. We further investigate the applicability of artificial intelligent (AI) based prediction through a comparative study of two tunnelling datasets with different sizes and features. Four different ML approaches, including support vector machine (SVM), random forest (RF), back-propagation neural network (BPNN), and deep neural network (DNN), are utilized. Two techniques, i.e. particle swarm optimization (PSO) and grid search (GS) methods, are adopted for hyperparameter optimization. To assess the reliability and efficiency of the predictions, three performance evaluation indicators, including the mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient (R), are calculated. Our results indicate that proposed models can accurately and efficiently predict the settlement, while the RF model outperforms the other three methods on both datasets. The difference in model performance on two datasets (Datasets A and B) reveals the importance of data quality and quantity. Sensitivity analysis indicates that Dataset A is more significantly affected by geological conditions, while geometric characteristics play a more dominant role on Dataset B.     

Investigation of feature contribution to shield tunneling-induced settlement using Shapley additive explanations method

Accurate prediction of shield tunneling-induced settlement is a complex problem that requires consideration of many influential parameters. Recent studies reveal that machine learning (ML) algorithms can predict the settlement caused by tunneling. However, well-performing ML models are usually less interpretable. Irrelevant input features decrease the performance and interpretability of an ML model. Nonetheless, feature selection, a critical step in the ML pipeline, is usually ignored in most studies that focused on predicting tunneling-induced settlement. This study applies four techniques, i.e. Pearson correlation method, sequential forward selection (SFS), sequential backward selection (SBS) and Boruta algorithm, to investigate the effect of feature selection on the model's performance when predicting the tunneling-induced maximum surface settlement (S_max). The data set used in this study was compiled from two metro tunnel projects excavated in Hangzhou, China using earth pressure balance (EPB) shields and consists of 14 input features and a single output (i.e. S_max). The ML model that is trained on features selected from the Boruta algorithm demonstrates the best performance in both the training and testing phases. The relevant features chosen from the Boruta algorithm further indicate that tunneling-induced settlement is affected by parameters related to tunnel geometry, geological conditions and shield operation. The recently proposed Shapley additive explanations (SHAP) method explores how the input features contribute to the output of a complex ML model. It is observed that the larger settlements are induced during shield tunneling in silty clay. Moreover, the SHAP analysis reveals that the low magnitudes of face pressure at the top of the shield increase the model's output.     

基于循环神经网络的盾构隧道#br# 引发地面最大沉降预测

伴随着计算机技术的快速发展,机器学习等新兴算法正在被越来越多地运用于预测隧道掘进引发的地面最大沉降。在隧道施工过程中,由盾构机和地面监测点位采集的数据具有很强的序列化特征,而传统的机器学习算法对序列数据的处理存在一定的局限性。循环神经网络(RNN)具有极强的对时序型数据的处理能力,在视频识别、语音翻译等领域有着广泛的应用。采用两种RNN模型(LSTM、GRU)和传统的BP神经网络模型,以地质参数、几何参数和盾构机参数作为输入,对隧道施工过程中引发的地面最大沉降进行预测分析。结果显示,RNN对隧道沉降的预测结果优于传统的BP神经网络模型,并且RNN在连续未知区段的预测结果比BPNN更加稳定。     

基于改进BP网络的盾构隧道开挖地表沉降预测

盾构隧道施工中引起的地表沉降是衡量开挖方式是否合适的关键指标。文中在介绍BP神经网络及盾构施工引起变形情况的基础上,对基于BP神经网络的盾构隧道开挖引起的地表沉降预测进行了研究,考虑了训练样本中奇异数据的剔除,采用变步长的方法,并选取适当的动量项系数,综合考虑各种影响因素,建立了盾构隧道开挖引起的地表沉降预测的BP网络模型,并对广州地铁二号线进行了具体的预测分析。分析结果表明:理论计算结果与工程实际情况一致,误差小于5%,所建立的预测模型是令人满意的。     

Prediction of flyrock distance induced by mine blasting using a novel Harris Hawks optimization-based multi-layer perceptron neural network

In mining or construction projects, for exploitation of hard rock with high strength properties, blasting is frequently applied to breaking or moving them using high explosive energy. However, use of explosives may lead to the flyrock phenomenon. Flyrock can damage structures or nearby equipment in the surrounding areas and inflict harm to humans, especially workers in the working sites. Thus, prediction of flyrock is of high importance. In this investigation, examination and estimation/forecast of flyrock distance induced by blasting through the application of five artificial intelligent algorithms were carried out. One hundred and fifty-two blasting events in three open-pit granite mines in Johor, Malaysia, were monitored to collect field data. The collected data include blasting parameters and rock mass properties. Site-specific weathering index (WI), geological strength index (GSI) and rock quality designation (RQD) are rock mass properties. Multi-layer perceptron (MLP), random forest (RF), support vector machine (SVM), and hybrid models including Harris Hawks optimization-based MLP (known as HHO-MLP) and whale optimization algorithm-based MLP (known as WOA-MLP) were developed. The performance of various models was assessed through various performance indices, including a10-index, coefficient of determination (R²), root mean squared error (RMSE), mean absolute percentage error (MAPE), variance accounted for (VAF), and root squared error (RSE). The a10-index values for MLP, RF, SVM, HHO-MLP and WOA-MLP are 0.953, 0.933, 0.937, 0.991 and 0.972, respectively. R² of HHO-MLP is 0.998, which achieved the best performance among all five machine learning (ML) models.     

盾构隧道致地层沉降的物理模型试验研究

通过不同隧道埋深、支护压力和掘进速度的盾构隧道施工地表沉降的大型物理模型试验,总结不同条件下的地表沉降规律,分析土压力的变化特性,归纳不同条件下的地表沉降曲线;探讨隧道埋深、支护压力和掘进速度对地表沉降值的影响,推导地表横断面沉降槽计算的经验公式.结果表明:随着隧道埋深增加,地表沉降值减小,地表横向沉降槽影响范围加宽;...     

A hybrid machine learning model to estimate self-compacting concrete compressive strength

This study examined the feasibility of using the grey wolf optimizer (GWO) and artificial neural network (ANN) to predict the compressive strength (CS) of self-compacting concrete (SCC). The ANN-GWO model was created using 115 samples from different sources, taking into account nine key SCC factors. The validation of the proposed model was evaluated via six indices, including correlation coefficient (R), mean squared error, mean absolute error (MAE), IA, Slope, and mean absolute percentage error. In addition, the importance of the parameters affecting the CS of SCC was investigated utilizing partial dependence plots. The results proved that the proposed ANN-GWO algorithm is a reliable predictor for SCC’s CS. Following that, an examination of the parameters impacting the CS of SCC was provided.     

Machine learning and data-driven prediction of pore pressure from geophysical logs: A case study for the Mangahewa gas field,New Zealand

Pore pressure is an essential parameter for establishing reservoir conditions, geological interpretation and drilling programs. Pore pressure prediction depends on information from various geophysical logs, seismic, and direct down-hole pressure measurements. However, a level of uncertainty accompanies the prediction of pore pressure because insufficient information is usually recorded in many wells. Applying machine learning (ML) algorithms can decrease the level of uncertainty of pore pressure prediction uncertainty in cases where available information is limited. In this research, several ML techniques are applied to predict pore pressure through the over-pressured Eocene reservoir section penetrated by four wells in the Mangahewa gas field, New Zealand. Their predictions substantially outperform, in terms of prediction performance, those generated using a multiple linear regression (MLR) model. The geophysical logs used as input variables are sonic, temperature and density logs, and some direct pore pressure measurements were available at the reservoir level to calibrate the predictions. A total of 25,935 data records involving six well-log input variables were evaluated across the four wells. All ML methods achieved credible levels of pore pressure prediction performance. The most accurate models for predicting pore pressure in individual wells on a supervised basis are decision tree (DT), adaboost (ADA), random forest (RF) and transparent open box (TOB). The DT achieved root mean square error (RMSE) ranging from 0.25 psi to 14.71 psi for the four wells. The trained models were less accurate when deployed on a semi-supervised basis to predict pore pressure in the other wellbores. For two wells (Mangahewa-03 and Mangahewa-06), semi-supervised prediction achieved acceptable prediction performance of RMSE of 130–140 psi; while for the other wells, semi-supervised prediction performance was reduced to RMSE > 300 psi. The results suggest that these models can be used to predict pore pressure in nearby locations, i.e. similar geology at corresponding depths within a field, but they become less reliable as the step-out distance increases and geological conditions change significantly. In comparison to other approaches to predict pore pressures, this study has identified that application of several ML algorithms involving a large number of data records can lead to more accurate prediction results.     

A hierarchical system to predict behavior of soil and cantilever sheet wall by data-driven models

The study proposes a framework combining machine learning (ML) models into a logical hierarchical system which evaluates the stability of the sheet wall before other predictions. The study uses the hardening soil (HS) model to develop a 200-sample finite element analysis (FEA) database, to develop the ML models. Consequently, a system containing three trained ML models is proposed to first predict the stability status (random forest classification, RFC) followed by 1) the cantilever top horizontal displacement of sheet wall (artificial neural network regression models, RANN1) and 2) vertical settlement of soil (RANN2). The uncertainty of this data-driven system is partially investigated by developing 1000 RFC models, based on the application of random sampling technique in the data splitting process. Investigation on the distribution of the evaluation metrics reveals negative skewed data toward the 1.0000 value. This implies a high performance of RFC on the database with medians of accuracy, precision, and recall, on test set are 1.0000, 1.0000, and 0.92857, respectively. The regression ANN models have coefficient of determinations on test set, as high as 0.9521 for RANN1, and 0.9988 for RANN2, respectively. The parametric study for these regressions is also provided to evaluate the relative insight influence of inputs to output.     

Optimization of machine learning models for predicting the compressive strength of fiber-reinforced self-compacting concrete

Fiber-reinforced self-compacting concrete (FRSCC) is a typical construction material, and its compressive strength (CS) is a critical mechanical property that must be adequately determined. In the machine learning (ML) approach to estimating the CS of FRSCC, the current research gaps include the limitations of samples in databases, the applicability constraints of models owing to limited mixture components, and the possibility of applying recently proposed models. This study developed different ML models for predicting the CS of FRSCC to address these limitations. Artificial neural network, random forest, and categorical gradient boosting (CatBoost) models were optimized to derive the best predictive model with the aid of a 10-fold cross-validation technique. A database of 381 samples was created, representing the most significant FRSCC dataset compared with previous studies, and it was used for model development. The findings indicated that CatBoost outperformed the other two models with excellent predictive abilities (root mean square error of 2.639 MPa, mean absolute error of 1.669 MPa, and coefficient of determination of 0.986 for the test dataset). Finally, a sensitivity analysis using a partial dependence plot was conducted to obtain a thorough understanding of the effect of each input variable on the predicted CS of FRSCC. The results showed that the cement content, testing age, and superplasticizer content are the most critical factors affecting the CS.     

Prediction of maximum surface settlement caused by earth pressure balance (EPB) shield tunneling with ANN methods

In order to determine the appropriate model for predicting the maximum surface settlement caused by EPB shield tunneling, three artificial neural network (ANN) methods, back-propagation (BP) neural network, the radial basis function (RBF) neural network, and the general regression neural network (GRNN), were employed and the results were compared. The nonlinear relationship between maximum ground surface settlements and geometry, geological conditions, and shield operation parameters were considered in the ANN models. A total number of 200 data sets obtained from the Changsha metro line 4 project were used to train and validate the ANN models. A modified index that defines the physical significance of the input parameters was proposed to quantify the geological parameters, which improves the prediction accuracy of ANN models. Based on the analysis, the GRNN model was found to outperform the BP and RBF neural networks in terms of accuracy and computational time. Analysis results also indicated that strong correlations were established between the predicted and measured settlements in GRNN model with MAE = 1.10, and RMSE = 1.35, respectively. Error analysis revealed that it is necessary to update datasets during EPB shield tunneling, though the database is huge.     

A generic framework for geotechnical subsurface modeling with machine learning

This study introduces a generic framework for geotechnical subsurface modeling, which accounts for spatial autocorrelation with local mapping machine learning (ML) methods. Instead of using XY coordinate fields directly as model input, a series of autocorrelated geotechnical distance fields (GDFs) is designed to enable the ML models to infer the spatial relationship between the sampled locations and unknown locations. The whole framework using GDF with ML methods is named GDF-ML. This framework is purely data-driven which avoids the tedious work in the scale of fluctuations (SOFs) estimating and data detrending in the conventional spatial interpolation methods. Six local mapping ML methods (extra trees (ETs), gradient boosting (GB), extreme gradient boosting (XGBoost), random forest (RF), general regression neural network (GRNN) and k-nearest neighbors (KNN)) are compared in the GDF-ML framework. The results show that the GDFs are better than the conventional XY coordinate fields based ML methods in both accuracy and spatial continuity. GDF-ML is flexible which can be applied to high-dimensional, multi-variable and incomplete datasets. Among these six methods, GDF with ET method (GDF-ET) clearly shows the best accuracy and best spatial continuity. The proposed GDF-ET method can provide a fast and accurate interpretation of the soil property profile. Sensitivity analysis shows that this method is applicable to very small training dataset size. The associated statistical uncertainty can also be quantified so that the reliability of the subsurface modeling results can be estimated objectively and explicitly. The uncertainty results clearly show that the prediction becomes more accurate when more sampled data are available.     

Prediction of thermal conductivity of rock through physico-mechanical properties

The transfer of energy between two adjacent parts of rock mainly depends on its thermal conductivity. Present study supports the use of artificial neural network (ANN) and adaptive neuro fuzzy inference system (ANFIS) in the study of thermal conductivity along with other intrinsic properties of rock due to its increasing importance in many areas of rock engineering, agronomy and geo environmental engineering field. In recent years, considerable effort has been made to develop techniques to determine these properties. Comparative analysis is made to analyze the capabilities among six different models of ANN and ANFIS. ANN models are based on feedforward backpropagation network with training functions resilient backpropagation (RP), one step secant (OSS) and Powell–Beale restarts (CGB) and radial basis with training functions generalized regression neural network (GRNN) and more efficient design radial basis network (NEWRB). A data set of 136 has been used for training different models and 15 were used for testing purposes. A statistical analysis is made to show the consistency among them. ANFIS is proved to be the best among all the networks tried in this case with average absolute percentage error of 0.03% and regression coefficient of 1, whereas best performance shown by the FFBP (RP) with average absolute error of 2.26%. Thermal conductivity is predicted using P-wave velocity, porosity, bulk density, uniaxial compressive strength of rock as input parameters.     

Soil deformation induced by Double-O-Tube shield tunneling with rolling based on stochastic medium theory

The prediction of soil deformation during tunneling is very difficult for Double-O-Tube (DOT) shield tunnel construction, especially for the shield rolling. According to the characteristics of DOT shield tunneling and rolling, a calculation model of soil deformation due to tunneling-induced ground loss was established. Based on the stochastic medium theory, the theoretical solutions of soil deformations considering the rolling of DOT shield machine were derived by polar coordinate transformation and multi-subdomain integral method. The predicted surface settlement from the proposed solution is better agreement with the observed data than those obtained by the two previous methods (namely the equivalent excavated-area method (EAM) and the simple superstition method (SM)). In addition, only ground surface settlement can be estimated under no rolling of DOT shield machine using the two previous methods, while this proposed solution owns great progress in solving the subsoil deformation and the influences of rolling. In order to further study the influence of DOT shield rolling angle on soil deformation under different engineering conditions, the parameter sensitivity analyses regarding the tunnel depth h, the ground loss parameter ɛ and the influence zone angle β₀ were extensionally discussed.     

基于长短时记忆网络的盾构机刀盘扭矩实时预测

盾构机是土质隧道开挖的优质工程机械,被广泛应用在地铁建设。刀盘扭矩是保证盾构正常推进的关键参数,能被精确实时预测对预防灾难事故、确保施工正常推进具有极高的指导意义。针对现有扭矩预测多为计算平均值的问题,提出一种基于长短时记忆(Long-Short Term Memory,LSTM)网络的扭矩实时预测模型。首先通过分析盾构机状态参数与扭矩的相关性,选择一组关键状态参数,降低输入维度;然后建立LSTM扭矩预测模型;最后利用该模型在归一化后的实际数据集上进行验证并与BP网络模型对比。试验结果表明,该模型在测试集上均方差为0.002 81,平均绝对误差为0.036 7,均优于BP网络;该模型具有良好的预测能力与泛化性能,能够很好地拟合关键状态参数与刀盘扭矩之间的非线性关系。     

隧道拱顶下沉时序遗传算法神经网络预测模型

传统神经网络算法不可避免会出现局部极值问题,可能导致训练失败。因此,作者在分析了隧道拱顶下沉规律及其主要影响因素的基础上,采用基于遗传算法的神经网络建立了隧道拱顶下沉时序的预测模型。模型在綦万高速公路麒麟寺隧道施工中成功应用,结果表明采用基于遗传算法的神经网络能够有效避免局部极值问题,收敛速度较快,能对隧道拱顶下沉时序进行较为准确的预测。     

Modeling of bentonite/sepiolite plastic concrete compressive strength using artificial neural network and support vector machine

Plastic concrete is an engineering material, which is commonly used for construction of cut-off walls to prevent water seepage under the dam. This paper aims to explore two machine learning algorithms including artificial neural network (ANN) and support vector machine (SVM) to predict the compressive strength of bentonite/sepiolite plastic concretes. For this purpose, two unique sets of 72 data for compressive strength of bentonite and sepiolite plastic concrete samples (totally 144 data) were prepared by conducting an experimental study. The results confirm the ability of ANN and SVM models in prediction processes. Also, Sensitivity analysis of the best obtained model indicated that cement and silty clay have the maximum and minimum influences on the compressive strength, respectively. In addition, investigation of the effect of measurement error of input variables showed that change in the sand content (amount) and curing time will have the maximum and minimum effects on the output mean absolute percent error (MAPE) of model, respectively. Finally, the influence of different variables on the plastic concrete compressive strength values was evaluated by conducting parametric studies.     

Wavelet Analysis for tunneling-induced ground settlement based on a stochastic model

Current practice in predicting tunneling-induced ground settlement has some limitations in describing the time-dependent settlement process due to the existence of measurement error. In this study, settlement data was considered as time series by establishing a stochastic model, while measurement error was regarded as a stationary and normally distributed stochastic process. Furthermore, Wavelet Analysis was introduced to filter the measurement error and extract the actual settlement value, which is similar to denoising in signal processing. In addition, methods such as the unit root test, normality test and ANOVA, were used to testify whether the characteristics of the filtered part of settlement data were consistent with those of measurement error. As a result, an optimal selection of wavelet basis and decomposition level could be made when using Discrete Wavelet Transform. Finally, extensive instrumentation data obtained from a real tunnel project supported our model hypothesis and proved the feasibility of this approach, and decomposing at level 4 with wavelet D16 was proved to achieve the best performance.     

Application of machine learning algorithms for the evaluation of seismic soil liquefaction potential

This study investigates the performance of four machine learning (ML) algorithms to evaluate the earthquake-induced liquefaction potential of soil based on the cone penetration test field case history records using the Bayesian belief network (BBN) learning software Netica. The BBN structures that were developed by ML algorithms-K2, hill climbing (HC), tree augmented naive (TAN) Bayes, and Tabu search were adopted to perform parameter learning in Netica, thereby fixing the BBN models. The performance measure indexes, namely, overall accuracy (OA), precision, recall, F-measure, and area under the receiver operating characteristic curve, were used to evaluate the training and testing BBN models’ performance and highlight the capability of the K2 and TAN Bayes models over the Tabu search and HC models. The sensitivity analysis results showed that the cone tip resistance and vertical effective stress are the most sensitive factors, whereas the mean grain size is the least sensitive factor in the prediction of seismic soil liquefaction potential. The results of this study can provide theoretical support for researchers in selecting appropriate ML algorithms and improving the predictive performance of seismic soil liquefaction potential models.