1751
Prediction of hard rock TBM penetration rate based on Data Mining techniques
Modèles de prévision du taux de pénétration de tunnelier dans les roches dures
Martins F.F., Miranda T.F.S.
Department of Civil Engineering, University of Minho, Campus de Azurém, 4800-058 Guimarães, Portugal
ABSTRACT: The aim of this work is to use Data Mining tools to develop models for the prediction of hard rock tunnel boring
machine (TBM) penetration rate (ROP). A database published by Yagiz (2008) was used to develop these models. The parameters of
the database were the uniaxial compressive strength (UCS), an index used to quantify the brittleness and toughness and denominated
peak slope index (PSI), the distance between the planes of weakness (DPW), the angle between tunnel axis and the planes of
weakness (α) and the output parameter rate of penetration (ROP). The R program environment was used as a modeling tool to apply
the artificial neural networks (ANN) and the support vector machine (SVM) algorithms and the corresponding models. These models
were compared with two equations presented by Yagiz (2008) and Yagiz and Karahan (2011). It was concluded that the ANN model
has the best performance. Moreover, these new models allowed computing the importance of the different input parameters for
predicting machine performance. It was concluded that PSI is the most important parameter and UCS is the less important parameter.
RÉSUMÉ : L'objectif de cette étude s’agit d'utiliser des outils de Data Mining en vue de développer des modèles de prévision de la
taux de pénétration d’un tunnelier dans les roches dures (ROP). Une base de données publiée par Yagiz (2008) a été utilisée pour
développer ces modèles. Les paramètres de la base de données comprend la résistance en compression uniaxiale (UCS), un index que
permettre mesurer la fragilité et la ténacité appelé d’index de pic maximal (PSI), la distance entre les plans de faiblesse (DPW), l'angle
entre l'axe du tunnel et le des plans de faiblesse (α) et le paramètre de sortie dénommé de taux de pénétration (ROP). L'environnement
du programme R a été utilisé comme un outil de modélisation pour appliquer les algorithmes des réseaux de neurones artificiels et des
machines à vecteurs de support et leurs modèles correspondants. Ces modèles ont été comparés à deux équations présentées par Yagiz
(2008) et Yagiz et Karahan (2011). On a conclu que le modèle des réseaux de neurones artificiels a été la meilleure performance. En
outre, ces nouveaux modèles ont permis le calcul de l'importance des différents paramètres d'entrée pour prévoir la performance de la
machine. Il a été conclu que l'PSI est le paramètre le plus important et l’UCS est le paramètre moins important.
KEYWORDS: tunnel boring machine, penetration ratio, data mining, machine learning
1 INTRODUCTION
The first question that arises when someone wants to excavate a
tunnel with a tunnel boring machine is to evaluate its
performance. However, this is a very complex task that requires
not only the choice of a performance parameter but also of
predictive models that require not only this parameter but also
other input parameters. Yagiz (2008) pointed out as relevant
performance parameters the penetration ratio (ROP), the ratio of
excavated distance to the operating time during continuous
excavation phase, and advance rate (AR), the ratio of both
mined and supported actual distance to the total time.
Nevertheless, according to the author, most of the forecasting
models are related with the prediction of the ROP. There are
many kinds of forecasting models. These include theoretical,
empirical, artificial neural network, fuzzy logic, genetic
algorithms and particle swarm optimization (Yagiz and Karahan
2011).
Yagiz (2008) using a statistical approach obtained a
predictive equation of ROP as a function of measured
engineering rock properties. Recently, Yagiz and Karahan
(2011) presented a new equation to estimate ROP using the
particle swarm optimization. Both studies included as
independent variables the uniaxial compressive strength (UCS),
an index used to quantify the brittleness and toughness and
denominated peak slope index (PSI), the distance between the
planes of weakness (DPW) and the angle between tunnel axis
and the planes of weakness (α). Their database consisted of 153
collected data sets related to Queens Water Tunnel # 3, stage 2,
New York City, USA.
The aim of this study is to develop models based on Data
mining techniques such as artificial neural networks (ANN) and
support vector machines (SVM) using the same database
presented by Yagiz (2008) and to compare the performance of
these models with the performance of the ones presented by
Yagiz (2008) and Yagiz and Karahan (2011).
2 REVIEW OF DATA MINING IN TUNNELLING
Feng et al. (2004) presented a novel machine learning method,
termed support vector machine (SVM), to obtain a global
optimization model in conditions of large project dimensions,
such as tunnels, small sample sizes and nonlinearity. A new idea
is put forward to combine the SVM with a genetic algorithm.
The results indicate that the established SVMs can appropriately
describe the evolutionary law of deformation of geo-materials at
depth and provide predictions for the future time steps with
acceptable accuracy and confidence. Liu et al. (2004)
introduced the SVM regression algorithm for the design of
tunnel shotcrete-bolting support parameters. Suwansawat and
Einstein (2006) attempt to evaluate the potential as well as the
limitations of ANN for predicting surface settlements caused by
EPB shield tunneling and to develop optimal neural network
models for this objective. Gajewski and Jonak (2006) presented
the results of a research work using ANN to classify the signals
of machining forces typical for particular worn cutting tools.
Javadi (2006) explored the capabilities of neural networks to
predict the air losses in compressed air tunneling. Yoo and Kim
(2007) demonstrated that an integrated Geographical
