797
Artificial intelligence for modeling load-settlement response of axially loaded (steel)
driven piles
Application de l’intelligence artificielle à la modélisation de la courbe effort-tassement des pieux
battus (en acier) soumis à un chargement axial
Shahin M.A.
Department of Civil Engineering, Curtin University, Perth WA, Australia
ABSTRACT: The design of pile foundations requires good estimation of the pile load-carrying capacity and settlement. Design for
bearing capacity and design for settlement have been traditionally carried out separately. However, soil resistance and settlement are
influenced by each other and the design of pile foundations should thus consider the bearing capacity and settlement in-separately.
This requires the full load-settlement behavior of piles to be well predicted. However, it is well known that the actual load-settlement
behavior of pile foundations can only be obtained by load tests carried out in-situ, which are expensive and time-consuming. In this
paper, artificial intelligence (AI) using the recurrent neural networks (RNN) is used to develop a prediction model that can resemble
the full load-settlement response of steel driven piles subjected to axial loading. The developed RNN model is calibrated and
validated using several in-situ full-scale pile load tests, as well as cone penetration test (CPT) data. The results indicate that the RNN
model has the ability to predict well the load-settlement response of axially loaded steel driven piles and can thus be used by
geotechnical engineers for routine design practice.
RÉSUMÉ: Le dimensionnement des fondations sur pieux nécessite une estimation précise de la capacité portante et du tassement d’un
pieu. Traditionnellement, la détermination de la capacité portante et du tassement d’un pieu est effectuée de manière séparée.
Cependant, la résistance du sol et le tassement du pieu sont interdépendants. Ainsi, le dimensionnement des fondations sur pieux
devrait considérer de manière simultanée la capacité portante et le tassement du pieu. Ceci nécessite une bonne prédiction de la courbe
effort-tassement du pieu. Cependant, il est bien connu que la courbe effort-tassement du pieu ne peut être obtenue que par des essais
de chargement du pieu in-situ, et qui sont coûteux et consommateurs en temps. Dans cet article, l’intelligence artificielle (IA) utilisant
les réseaux de neurones récurrents (RNN) est utilisée pour développer un modèle de prédiction qui simule la courbe effort-tassement
des pieux en acier soumis à un chargement axial à partir des essais in-situ. Le modèle RNN développé est calibré et validé en utilisant
plusieurs résultats d’essais de chargement de pieux in-situ, ainsi que des résultats d’essais pénétrométriques (CPT). Les résultats
obtenus indiquent que le modèle RNN a la capacité de prédire avec précision la courbe effort-tassement d’un pieu en acier chargé
axialement et il peut ainsi être utilisé dans la pratique par les géotechniciens.
KEYWORDS: artificial intelligence, recurrent neural networks, pile foundations, load-settlement, modeling.
1 INTRODUCTION
Bearing capacity and settlement are the two main criteria that
govern the design process of pile foundations so that safety and
serviceability requirements are achieved. Design for bearing
capacity is carried out by determining the allowable pile load,
which is obtained by dividing the ultimate pile load by an
assumed factor of safety. Design for settlement, on the other
hand, consists of obtaining the amount of settlement that occurs
when the allowable load is applied to the pile, causing the soil
to consolidate or compress. Design for bearing capacity and
design for settlement have been traditionally carried out
separately. However, Fellenius (1988) stated that: “
The
allowable load on the pile should be governed by a combined
appraoch considering soil resistance and settlement
inseparately acting together and each influencing the value of
the other
”. In addition, there is a strong argument regarding the
definition of the ultimate pile load and many methods have been
proposed in the litearture, some result in interpreted ultimate
loads that greatly depend on judgement and the shape of the
load-settlement curve (1980). Consequenlty, for design
purposes, the full load-settlement response of piles needs to be
well predicted and simulated; the designer can thus decide the
ultimate load and comply with the srevieability requirement.
Good prediction of the full load-settlement response of pile
foundations needs thorough understanding of the load transfer
along the pile length, which is complex, indeterminate and
difficult to quantify (Reese et al. 2006). The actual load-
settlement response of pile foundations can only be obtained by
carrying out load tests in-situ, which is expensive and time-
consuming. On the other hand, the load-settlement response of
pile foundations can be estimated using many methods available
in the literature. However, due to many complexities, available
methods, by necessity, simplify the problem by incorporating
several assumptions associated with the factors that affect the
pile behavior. Therefore, most existing methods failed to
achieve consistent success in relation to the predictions of pile
capacity and corresponding settlement. In this respect, the
artificial intelligence (AI) can be efficient as they can resemble
the in-situ full-scale pile load tests without the need for any
assumptions or simplifications. AI is a data mining statistical
technique that has proved its potential in many applications in
geotechncial engineering (see Shahin et al. 2009).
In this paper, the feasibility of using one of the most
commonly used AI techniques, i.e. recurrent neural networks
(RNN), is used for modeling the load-settlement response of
steel driven piles subjected to axial loading. To facilitate the use
of the developed RNN model for routine design by
practitioners, the model is translated into an executable program
that is made available for interested readers upon request.
2. OVERVIEW OF RECURRENT NEURAL NETWORKS
The type of neural networks used in this study are multilayer
perceptrons (MLPs) that are trained with the back-propagation
