Predicting EPBM advance rate performance using support vector regression modeling

https://doi.org/10.1016/j.tust.2020.103520Get rights and content

Highlights

  • Support vector regression successfully modeled 85–92% of advance rate.

  • Significant difference in advance rate model between soil units.

  • Key influential parameters include cutterhead thrust, cutterhead torque, foam injection.

Abstract

Earth pressure balance shield tunnel boring machines (EPBM) are widely used in tunneling practice yet the mechanics that define ground-EPBM interaction, specifically the advance rate, are not well understood. In the study presented here, machine learning techniques including feature selection, support vector regression (SVR) and partial dependence plots (PDP), were successfully applied to EPBM data to develop and explain EPBM advance rate modeling through five widely varying soil types. The geotechnical conditions were implicitly incorporated into the analysis by developing soil formation-specific SVR models. The SVR models were highly successful in capturing AR behavior, exhibiting R2 values of 0.88–0.95 when independently evaluated with test data. Automatic feature selection revealed the same EPBM parameters of notable influence on AR across all ESUs, including net thrust, cutterhead torque, foam flow rate and screw conveyor torque. The SVR models, however, revealed considerably different relationships between these key parameters and AR, indicating that the soil plays a significant role in AR behavior. PDP analysis captured the sensitivity of AR to each key parameter as a function of parameter magnitude. The PDP results show that AR is positively correlated (increasing AR with increasing parameter value) and/or negatively correlated (decreasing AR with increasing parameter value) to varying degrees as a function of parameter value, all of which is strongly soil dependent.

Introduction

Earth pressure balance shield tunnel boring machines (EPBM) are widely used for soft-ground tunneling projects. The principle of EPB is to exert monitored pressure on the excavation face to counterbalance the earth and water pressures. This allows EPBMs to safely operate within the congested urban environment without causing ground settlement or damaging existing infrastructure. Considering the risks of tunneling in infrastructure-laden urban environment, understanding the mechanics of EPBMs, the inner workings of its operating components and the EPBM-soil interactions are of significant interest. EPBM operation involves several tasks including soil conditioning, controlling ground deformations by maintaining appropriate chamber pressure, navigating the project alignment, applying an effective combination of controllable parameters and processing the excavated materials through a depressurizing screw conveyor. The complex operational procedure makes it difficult to understand some of the mechanics of EPBMs; therefore, performance prediction of these machines remains a challenging task.

Performance of TBMs in hard rock tunneling has been investigated with various machine learning methods including: artificial neural networks (Yagiz et al., 2009, Salimi et al., 2016, Gholamnejad and Narges, 2010), support vector regression (Mahdevari et al., 2014), neuro-fuzzy methods (Grima et al., 2000), hybrid models (Armaghani et al., 2017) and particle swarm optimization (Yagiz and Karahan, 2011). However, to the authors’ knowledge, only limited studies have been published on the assessment of EPBMs performance using machine learning techniques (Mooney et al. 2018). Unlike open mode hard rock TBM tunneling where the advance rate is strictly related to the cutterhead-rock interaction, EPBM advance rate can be influenced by numerous factors unique to EPBMs, including soil conditioning, material flow through the cutterhead openings, excavation chamber material flow, magnitude and distribution of face pressure, and screw conveyor behavior. The machine learning models developed for hard rock TBMs do not account for these aspects. To this end, machine learning models unique to EPBMs are required.

Modern EPBMs record extensive information during excavation, including time histories of thrust forces (main jacks, articulation jacks, cutterhead support jacks), cutterhead and screw conveyor torque, conditioning injection rates and pressures, alignment data, and so forth. Detailed excavation data makes data-driven modeling a promising approach for EPBM performance evaluation and forecasting (Mooney et al., 2018). Data driven modelling and statistical learning can be useful in understanding complex problems where the mechanics of the system is elusive.

In this paper, we pursue EPBM performance characterization using data-driven models to identify the governing factors that affect EPBM advance rate (AR). AR, often referred to as penetration rate or rate of penetration in rock tunneling parlance, is not directly controlled by the operator. Rather, AR is a result of complex machine-ground and machine-muck interaction, the mechanics of which are yet undescribed (Mooney et al., 2018). To pursue data driven modeling of AR, we conduct a comprehensive feature evaluation using a feature selection technique (RRreliefF) to identify the best features describing EPBM performance. Using the selected features, EPBM AR is modeled with support vector regression (SVR), a well-established and accepted machine learning technique. SVR is considered a black-box technique wherein the explicit relations of inputs (independent variables, controllable EPBM parameters) and outputs (dependent variable, advance rate) are not available and cannot be directly explored. We overcome this by examining the influence of EPBM parameters (selected features) on EPBM performance using partial dependence plots (PDP). We employ PDPs to quantify the effect of each individual operation parameter on EPBM advance rate. The investigation of PDPs provides further insight into EPBM operation and the governing parameters that affect performance.

We used the recorded dataset from the Northgate Link tunneling project in Seattle, Washington to implement and validate our proposed data analysis framework. A brief description of this tunnel project and the dataset is provided first. Then, feature engineering and the study of EPBM parameters to identify the governing factors that affect the AR is presented. Finally, regression analysis using SVR and PDPs is presented in detail to quantify the effect of governing parameters on EPBM AR. In the current work, we focus on establishing the relation between EPBM data and the resulting AR, excluding the explicit consideration of geological/geotechnical information. We presume that the geotechnical parameters are implicit in the EPBM performance. Explicit inclusion of geotechnical information is limited by the sparse number of boring logs (order of 100) along the alignment. EPBM operation data is partitioned using the available but limited geological data in a way that each segment is consists primarily of a single engineering soil unit (ESU). In this way, we learn the influence of geological/geotechnical conditions by model examination and comparison across the ESU-dependent performance models. The study provides a novel contribution in that EPBM AR performance models for different soils have not been developed prior, and the framework implicitly shows how soil type influences the AR models.

Section snippets

Sound Transit Northgate Link tunneling project

We employed excavation data from the Sound Transit Northgate Link tunneling project (N125), in the city of Seattle, to conduct and validate our data-driven study of EPBM performance. Constructed by the joint venture Jay Dee Constructors, Coluccio, and Michels, the Northgate link extension project provides light rail service to connect the University District to Northgate. The underground part of the project is 5.6 km long and runs in twin tunnels from the University of Washington through

Feature selection and evaluation

A total of 107 EPBM parameters are considered in the data-driven modeling of AR. However, inclusion of all 107 parameters is not only computationally prohibitive but also can adversely affect the prediction performance due to the presence of irrelevant and redundant information (Caruana and Freitag, 1994, Maldonado et al., 2014). Feature (parameter) evaluation and optimal feature-subset selection is therefore employed to build efficient and accurate data-driven AR models.

Methodology

The selected features based on RReliefF results are used to train support vector regression (SVR) models of AR. Support vector machines (SVM) were originally developed as a classification algorithm that finds a linear hyperplane to separate the data samples of different classes. For many classification problems, however, there are infinite number of linear classifying hyperplanes. SVM selected the hyperplane with maximum separation (margin, distance) from two classes. This basic form of SVM can

Partial dependence plots

Complex models such as SVR act as black boxes and offer little insight into interpretability. Correspondingly, the relationship between the governing parameters and the predicted AR cannot be explored using such models. Friedman (2001) proposed a model agnostic approach, called the Partial Dependence Plot (PDP) to illustrate the change in average prediction as features (parameters) vary over their marginal distributions. In this regard, the PDP is a plot of model AR predictions for all possible

Conclusions

In the study presented here, feature selection, support vector regression and partial dependence plots were successfully applied to EPBM data to develop and explain EPBM advance rate modeling through various soil types. Feature evaluation using RReliefF produced a ranked list of EPBM parameters (features) that most influence AR in each ESU. Included in the most influential parameters for each ESU were cutterhead torque, net thrust (as a measure of cutterhead thrust), screw conveyor 1 torque,

CRediT authorship contribution statement

Soroush Mokhtari: Methodology, Software, Formal analysis, Investigation, Visualization, Writing - original draft. Michael A. Mooney: Conceptualization, Methodology, Resources, Supervision, Funding acquisition, Project administration, Writing - original draft, Writing - review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

We would like to acknowledge Sound Transit and Jay Dee Contractors, Inc. for providing access to Northlink project data. We particularly thank Dr. Ehsan Alavi and Dr. Lisa Mori for their thoughtful feedback on machine performance and their insights into the project.

References (22)

  • Godinez, R., Yu, H., Mooney, M.A., 2015. Earth pressure balance machine cutterhead torque modeling: learning from...
  • Cited by (25)

    • Time-series prediction of shield movement performance during tunneling based on hybrid model

      2022, Tunnelling and Underground Space Technology
      Citation Excerpt :

      Different types of strata posed different demands on the construction technology. The advancing speed of the EBP shield machine would be slower in the rock strata, and MT would be increased (Alp and Apaydin, 2019; Cheng et al., 2020; Kargar et al., 2020; Farrokh, 2020; Mokhtari and Mooney, 2020). Uncertainties exist in geological data sets (Juang et al., 2019; Lin et al., 2021d; Phoon and Kulhawy, 1999).

    • Tunnel boring machine vibration-based deep learning for the ground identification of working faces

      2021, Journal of Rock Mechanics and Geotechnical Engineering
      Citation Excerpt :

      Armaghani et al. (2017) found that the hybrid model performed better than the ANN in predicting the penetration rate of a hard rock TBM. By comparing machine learning and deep learning algorithm-based methods, Mokhtari and Mooney (2020) adopted feature selection, support vector regression, and partial dependence plots to develop an AI-based earth pressure balance (EPB) advance rate prediction model. Moreover, numerous convolutional neural network (CNN) models with robust image reorganization have been developed for tunnel defects or leakage detection in recent years (Huang et al., 2018; Chen et al., 2020, 2021; Yu et al., 2021; Zhao et al., 2021).

    View all citing articles on Scopus
    View full text