Professional Documents
Culture Documents
https://doi.org/10.1007/s00366-019-00847-5
ORIGINAL ARTICLE
Abstract
Due to a considerable level of uncertainty describing the pile–soil behavior, many pile capacity prediction methods have
focused on correlation with in situ tests. In recent years, artificial neural networks (ANNs) have been applied successfully
in many problems in geotechnical engineering, especially, axial pile capacity estimation for driven and drilled shaft piles.
Training neural networks is a crucial task that needs effective optimization algorithms. The most popular algorithm is a
back-propagation method (BP), which is based on a gradient descent that can trap in local minima. The paper proposes a
new artificial neural network (ANN) in which the learning is performed using a recent teaching–learning-based optimiza-
tion algorithm (TLBO), improving axial capacity predictions. The model is trained and validated on 479 data sets for a wide
range of uncemented soils and pile configurations, obtained from the literature. Results show that the considered TLBO-ANN
model outperforms other state-of-the-art models in the prediction accuracy and the generalization capability. For instance,
we obtained a coefficient of determination R2 = 0.941 and a variance accounted for VAF = 94.09% for TLBO-ANN while
R2 = 0.871 and VAF = 87.31% for the classical BP-ANN. In addition, error investigation with log-normal approaches dem-
onstrates that the probability that predictions fall within a ± 25% accuracy level for TLBO-ANN model is 0.93 and that for
BP-ANN model is 0.75. The proposed TLBO-ANN model predicts pile capacity with more accuracy, less scatter, and higher
reliability.
Keywords Ultimate capacity · ANN · TLBO-ANN · Pile load tests · SPT data · Failure zone
1 Introduction
* Mohammed Hachama Pile foundation is the best way to carry out the load issue
hachamam@gmail.com from the substructure. Thousands of piles are installed
Amel Benali around the world every year. The World Trade Organization
benali_amel4@yahoo.fr (OMC) reveals that the global production grew-up by 3% in
Aouda Bounif 2007 amounting to 4.7 trillion US dollars, while growth in
boun_aouda2@yahoo.fr 2015 was nearly 15%. Recent years witnessed an increas-
Ammar Nechnech ing interest in the axial pile capacity or pile settlement and
ammar.nechnech@gmail.com new computation procedures. Numerous research programs
Mourad Karray have been conducted either by the military corps (US Army
mourad.karray.benhassen@usherbrooke.ca Corps of Engineers) [1–3], or civil [4–8].
Pile behavior under axial loading is a complicated process
1
Department of Civil Engineering, University of Djilali involving inter-related movements of both the pile and the
Bounaama, Khemis Miliana, Algeria
supporting soil because the load is transferred from the pile
2
Laboratory of Pure and Applied Mathematics, University to the supporting medium. A large set of related parameters
of M’Sila, M’Sila, Algeria
are generally difficult to estimate. Thus, determining axial
3
Faculty of Civil Engineering, University of Science pile capacity is still an active research problem. Many meth-
and Technology, Beb Ezzouar, Algiers, Algeria
ods have been developed, which can be divided into many
4
Department of Civil Engineering, University of Sherbrooke, classes: analytical, semi-empirical, empirical, and numerical
Sherbrooke, Canada
13
Vol.:(0123456789)
Engineering with Computers
methods [9]. Most of them use simplified hypotheses which used by GAs. Parameter tuning is a crucial factor which
approximate the behavior of the system pile/soil. Conse- highly affects the performance of the algorithms.
quently, models imprecision and parameters uncertainties Recently, Rao et al. [15–19] proposed a teaching–learn-
may be an obstacle in the design of effective computation ing-based optimization (TLBO) algorithm which does not
methods. require any specific parameter. Based on this algorithm,
Empirical methods are widely used techniques that relate different techniques and applications have been proposed:
on correlations with in situ tests [10–12]. Such techniques Temur and Bekdas [28] for optimum design of reinforced
use simple statistical tools (e.g., regression analysis) which concrete retaining walls, Gandomi and Kashan [29] in shal-
cannot account for the complex interdependence of the low footing optimization, Oser and Temur [30] to optimize
involved factors such as soil stratification, soil–pile interac- the pile groups under vertical loads for a safe and economi-
tion, and distribution of soil resistance along the pile shaft. cal design, and Daloglu et al. [31] for optimum design of
Artificial neural networks (ANNs) are empirical mod- braced steel space frames including soil/structure.
eling techniques inspired by the functioning of the human This paper presents an TLBO-optimized artificial neural
brain. Unlike most available statistical methods, ANNs do networks for modeling axial pile capacity. The new hybrid
not need a predefined model of the relationship between the model is used to predict the ultimate capacity of both driven
inputs and the outputs. Rather, they learn the structure and and drilled shaft piles (the most common pile types used in
parameters of the model from the available data [13]. In the country strategic building, industry blocks, oil platforms,
geotechnics for instance, Momeni et al. [14] applied ANN in etc.) embedded in uncemented soils based on SPT data. As
predicting the bearing capacity of deep foundations. A feed- far as we know, similar hybrid model TLBO-ANN has been
forward back-propagation network has been implemented used in one single previous work in [32] for oil flow predic-
using 36 PDA tests performed on various concrete piles in tion, drilling, and production.
different project sites. Findings indicate the feasibility of To train our model, we collected a large database from
ANN in predicting ultimate, shaft and tip bearing resist- many sources [33]. We used corrected blows number in the
ances of piles. failure zone obtained using a statistical analysis of a numeri-
Recently, ANNs have been improved thanks to several cal model [33]. We achieved an analysis of the parameters
modern heuristic algorithms, which fall into two groups: (population size and iteration number) effect on the overall
evolutionary, e.g., genetic algorithms (GAs), and swarm result, and compared the our architecture with a classical
intelligence-based algorithms, e.g., particle swarm optimi- one using back-propagation algorithm, in addition to five
zation (PSO) and artificial bee colony (ABC) [16]. Such previous approaches. We derived and validated three new
algorithms have been successfully applied in civil engineer- simplified but accurate analytical models of the axial pile
ing [20, 21]. Momeni et al. [22] developed an PSO-ANN capacity expressed in terms of parameters.
model to predict unconfined compressive strength of rocks; The paper is organized as follows. Section 2 explains
the obtained results improved previous approaches. Arma- the ANN architecture and the TLBO learning algorithm
ghani et al. [23] predicted strength and elasticity modulus of to our model of axial pile capacity. In Sect. 3, we evaluate
granite through an ANN enhanced with the imperialist com- our model (TLBO-ANN) and compare it to an ANN trained
petitive algorithm (ICA). Momeni et al. [24] explored the with the to a back-propagation BP-ANN, in addition to some
GA-ANN model to predict the axial pile capacity obtaining empirical models. The paper ends with a concluding section.
results in close agreement with measured bearing capacities.
Armaghani et al. [25] established a hybrid model PSO-ANN
to estimate the ultimate bearing capacity of rock-socketed 2 Modeling axial pile capacity
piles. Their model offers a higher degree of accuracy com-
pared to conventional ANN. Moayedi et al. proposed an We develop here an ANN model that deals with two cases
adaptive neuro-fuzzy inferences system (ANFIS) optimized of pile installation, namely displacement and drilled shaft
with GA and PSO [26]. The results indicate higher reliability piles. This model combines a multi-layer perception (MLP)
of the optimized GA-ANFIS model in estimating friction network with the TLBO algorithm.
capacity ratio ( 𝛼 ) in driven shafts. Moayedi and Armaghani
[27] optimized an ANN with imperialism competitive algo- 2.1 Data
rithm (ICA) to estimate bearing capacity of driven pile in
cohesionless soil. The obtained results showed high reliabil- The data used to calibrate and validate the ANN model are
ity of the developed ICA-ANN model. All these probabilistic obtained from the literature (different sources), including
algorithms require common controlling parameters such as 335 and 144 static load tests on driven piles and drilled shaft
population size, number of generations, etc.; in addition to piles, respectively (Benali et al. [33]). We conducted experi-
specific parameters, such as mutation and crossover rates ments at different non-cohesive sites. The displacement tests
13
Engineering with Computers
include compression loading of steel and concrete piles, Table 1 Ranges of the database constituents
driven statically (jacked-in) or dynamically into the ground. Pile types
Usually, data division into training and testing sets is carried
Driven piles Drilled shaft
out arbitrarily, as there is no universal. Here, 70% and 30%
of whole datasets, 335 and 144 datasets, are used for the Min Max Min Max
training and testing phases, respectively, as recommended
Bext (m) 0.10 1.00 0.45 1.53
by Swingler [34].
Bint (m) 0.00 0.90 0.00 0.00
Bbase (m) 0.10 1.58 0.45 1.85
2.2 ANN architecture
Bhead (m) 0.10 1.00 0.45 1.85
D (m) 3.00 55.40 6.60 60.00
A neural network is a set of neurons interconnected with a
NShaft 3.00 64.36 5.00 78.77
each other. Connection weights can be thought of as the
Nbb 5.35 98.00 5.00 81.95
memory of the system and biases values. The disposition
Qt (MN) 0.12 11.50 0.85 17.50
of the weighted connection of these neurons is called archi-
Shaft material Steel Concrete Concrete Concrete
tecture. For instance, the multi-layer perceptron consists of
three adjacent layers: input, hidden and output [35]. How-
ever, Funahashi [36] proved that any problem can be approx-
imated using an ANN model with one hidden layer, which have failed if it experiences a settlement equal to 5% of its
is the choice we made in our model. Our MLP network has nominal dimension plus the elastic compression of the shaft
nine input neurons and one output neuron. The optimal num- ( QL∕Ep Ab ). In accordance with [41], the base failure zone
ber of neurons in the hidden layer has to be determined. is taken equal to height diameter above and four diameters
below pile tip. Nevertheless, the corrected bows number in
2.3 Inputs–outputs the failure zone Nb is taken as in Benali et al. [33]:
Nine factors affecting the pile capacity are considered as N b = 0.6N𝛽 + 0.4N𝛼 ,
inputs for our ANN: the embedment length (D), two cor-
rected blows numbers ( NShaft , Nb ), pile shaft material SM where N𝛽 is the corrected average blow counts within the
(1 for steel and 2 for concrete), pile classification PC (dis- lower zone and N𝛼 is the corrected average blow counts
placement or drilled shaft), and four diameters (to account within the upper zone.
the variety of pile shapes): Bext (external or shaft diameter),
Bbase (pile base diameter), Bint (internal diameter) and Bhead 2.4 Training
(pile head diameter). It is noted that, NShaft is defined as the
geometric mean of the values measured during the SPT test Training a network is the adaptation of free network param-
along the shaft pile. Table 1 reports the ranges of the data- eters using training inputs and the corresponding target val-
base constituents. ues. This is achieved by minimizing the network error (or
The output of the ANN is the ultimate pile capacity fitness function):
( Qt ) of driven piles, defined as the load corresponding to
1 ∑( )2
P
the plunging failure for the well-defined failure cases. For MSE = tj − o j , (1)
the cases where the failure load is not clearly defined, we P j
determine the failure load from the results of pile load tests
through a unique criterion. According to [37], Geotechnical where tj and oj are the target and the output values of the jth
Design-General rules, a small-diameter pile is considered pattern and P is the number of patterns.
to have failed if it experiences a settlement equal to 10% The most used algorithm for training is back-propagation
of its nominal dimension. On the other hand, the ultimate (BP). This algorithm is iterative, computationally simple,
capacity of a large-diameter pile is accessed, following the and quite efficient in performing a local search for an opti-
recommendations of [38, 39], when the pile settlement (S) mum solution, but with a slow convergence speed. As in
equals to all learning techniques, BP faces two main problems: the
̇ ̇ possible overfitting of training data and mathematical opti-
S = (QL)∕(Ep Ap ) + B∕30 (12), mization on large-scale engineering problems.
where Q is the test load, L is the pile length, Ap is the cross- Teaching–learning-based optimization (TLBO) is a popu-
sectional area of the pile, B is the pile diameter, and Ep is lation-based algorithm which simulates the teaching–learn-
the modulus of elasticity of the pile material. According to ing process of the classroom. A group of learners is consid-
Reese and O’Neill [40], drilled shaft pile is considered to ered as population and different subjects offered to learners
13
Engineering with Computers
are considered as different design variables of the optimiza- Table 2 Some architectures performances for an MLP trained with a
tion problem. Learner’s result is analogous to the fitness back-propagation algorithm (15 iterations)
value of the optimization problem. The TLBO algorithm Architecture RTrain MSE RTest
uses two basic modes of the learning: (1) through teacher
9 − 20 − 1 0.90 0.0089 0.92
(known as teacher phase) and (2) interacting with the other
9 − 15 − 1 0.91 0.0090 0.94
learners (learner phase). The TLBO is used to minimize (or
9 − 10 − 1 0.93 0.0098 0.94
to maximize) a given function f. When applied to ANN, f
9−7−1 0.91 0.0110 0.92
will be the MSE which depends on the network weights. It is
9−6−1 0.89 0.0120 0.83
worth noting that the TLBO requires only common control
9−5−1 0.86 0.0180 0.90
parameters (population size and number of generations).
2.4.1 Teacher phase
�
using trial-and-error approach. We tested 1, 2, 3, … 2I + 1
Xj,k,i = Xj,k,i + Dj,k,i , (3) hidden-layer nodes, where I is the number of input vari-
ables, because 2I + 1 is the upper limit needed to map any
where Xj,k,i
′
is the updated value of Xj,k,i which is accepted
continuous function [13]. Table 2 shows the performances of
only if the fitness function is improved, i.e., f (Xj,k,i
�
) < f (Xj,k,i ). some architectures. The best architecture corresponds to the
model with 10 hidden nodes (lowest prediction error 0.0098
2.4.2 Learner phase and largest correlation coefficient RTrain = 0.93 ; RTest = 0.94 )
(Table 1).
A learner interacts randomly with other learners for enhanc-
ing his or her knowledge. Two different learners k1 and k2 3.2 TLBO‑ANN
are randomly selected. Learner modification is expressed
as follows: We achieved an extensive study to determine the optimal
{ controlling parameters: the population size NPop and the
Xj,k1 ,i + ri (Xj,k1 ,i − Xj,k2 ,i ), if f (Xj,k1 ,i ) < f (Xj,k2 ,i ),
�
Xj,k = number of iterations Niter. A small population may fail to
,i Xj,k1 ,i + ri (Xj,k2 ,i − Xj,k1 ,i ), if f (Xj,k2 ,i ) < f (Xj,k1 ,i ). converge to a global solution, whereas a large one may cause
1
(4)
a delay in convergence and an increase in training time. The
Xj,k ,i is accepted if it gives better fitness function.
′
optimal value is typically established via a sensitivity test
1
13
Engineering with Computers
numbers: 1000, 1500, 2000, 2500, 3000, 5000, 8000, 10,000, As shown on Fig. 6, the TLBO-ANN has a better conver-
15,000, 20,000, 25,000 and 30,000. Figure 2 shows that the gence rate. For example, the TLBO and BP algorithms con-
MSE decreases when the number of iterations increases. verge to 0.007 and 0.009 in the 15000th and 15th iteration,
We choose Niter = 10000 which corresponds to and error respectively. In fact, the BP algorithm is trapped in a local
of 0.7% (MSE = 0.007). minimum as the energy stop decreasing over the 15th itera-
tion. The TLBO-ANN is slower than the BP algorithm, but
3.3 Empirical models gives better results.
We evaluated and compared the two algorithms using R2,
We applied several empirical models for the estimation of MSE, and P value [46]. Results are in Table 5, Figs. 7 and 8.
axial pile capacity: Meyehof [42], Aoki and Velloso [43],
Shioi and Fukui [34], Robert [44], and Bouafia [45] (see 3.5 Comparison between BP‑ANN, TLBO‑ANN,
Table 3). We estimated their constants a, b, c, and e by fit- and empirical methods
ting the empirical models to the training data. The obtained
values are given in Table 4. The TLBO-ANN model is compared with ANN model and
We evaluated the six empirical SPT-based methods on six SPT-based methods, using the following criteria: the
70 datasets, not included in the training datasets. Results are determination coefficient ( R2 ), the ratio of predicted to the
given in Figs. 3, 4, and 5. We carry out a regression analysis experimental bearing capacities (mean 𝜇 and standard devia-
to fit a line of predicted versus measured pile capacities. tion 𝜎 ), and the probability of predicting the pile capacity
The quality of the fitting is measured using a correlation within ± 25% accuracy (P). The later is obtained as the area
coefficient R. We obtained R equals to 0.920, 0.857, 0.839, beneath the log-normal distribution within a range equal to
0.850, 0.817 and 0.832, for Meyerhof, Shioi and Fukui, Aoki 0.75Qm ≤ Qp ≤ 1.25Qm . Results are summarized in Table 6.
and Velloso, Robert (driven piles), Robert (drilled shaft), Considering the R2 criteria, the developed TLBO-ANN
and Bouafia, respectively. All methods showed good per- model gives the best results. On another hand, the ideal
formances. Nevertheless, Meyerhof’s method shows the value of the ration Qp ∕Qm is one. The TLBO-ANN tends to
best results for driven pile axial capacity estimation, where overestimate the measured pile capacity by 9% on average
Robert and Bouafia methods produced good performance where this average is 17% for BP-ANN. On contrary, Rob-
for drilled shaft axial capacity. ert for bored piles, Bouafia, and Shioi and Fukui methods
underestimate the pile capacity by 12% , 16.5% , and 15% on
13
Engineering with Computers
Fig. 4 Measured versus
predicted pile capacity—Shioi
and Fukui’s method (left) and
Bouafia’s method (right)
average, respectively. Meyerhof appears to give good results We applied our model to predict the ultimate capacity of
with an underestimation average of 4% . Aoki and Velloso driven and drilled shaft piles embedded in uncemented soils
method underestimates the axial pile capacity by 29% . The based on SPT data. A comparison with other state-of-the-
TLBO-ANN has the highest log-normal distribution prob- art techniques, including an ANNs model with standard
ability values of 0.9261. back-propagation algorithm, proved that our TLBO-ANN
model achieved the best predictions. Indeed, the TLBO-
ANN ranked first w.r.t. the selected performances tests
4 Conclusion with a determination coefficient R2 = 0.94 , a ratio of the
predicted to the estimated pile capacity 𝜇 = 1.09 , and a vari-
Numerous approaches have been proposed in the literature to ance accounted for VAF = 94 ; while R2 = 0.87 , 𝜇 = 1.17
predict the ultimate capacities of piles, ranging from simple and VAF = 87 for the classical BP-ANN. In addition, the
empirical formulations to more sophisticated hybrid mod- probability that predictions fall within ± 25% accuracy level
els. In this paper, we proposed a new ANNs model, where is 0.93 for TLBO-ANN and 0.75 for BP-ANN. In the future,
the learning stage is achieved using the TLBO algorithm. we would perform a comparative study of optimization
13
Engineering with Computers
Fig. 6 Performance curve—
TLBO-ANN/BP-ANN
Fig. 7 Measured versus
predicted pile capacity of the
BP-ANN: training (left) and
testing (right)
Fig. 8 Measured versus
predicted pile capacity of the
TLBO-ANN: training (left) and
testing (right)
N.HL number of hidden layers, N.NHL number of neurons in hidden layer, 𝜇 momentum value, NP popula-
tion size, Maxit maximum number of iteration, R2 coefficient of determination, P P value
13
Engineering with Computers
Qp predicted pile capacity, Qm measured pile capacity, VAF variance accounted for, D.P. driven piles, D.S.P.
drilled shaft piles
⎡ 0.99950584 ⎤
⎢ ⎥
⎢ 2.02640275 ⎥
⎢ ⎥
⎢ 4.99996893 ⎥
⎢ ⎥
⎢ 5 ⎥
⎢− 4.9964599⎥
b1 = ⎢ ⎥, b2 = [− 4.9992929].
⎢− 4.5583188⎥
⎢ ⎥
⎢− 3.0332818⎥
⎢ ⎥
⎢− 4.9984299⎥
⎢ ⎥
⎢− 4.9999978⎥
⎢ ⎥
⎣ 4.65363322 ⎦
13
Engineering with Computers
13
Engineering with Computers
40. Reese LC, O’Neill MW (1988) Drilled Shaft: Construction Proce- 45. Bouafia A (2004) Évaluation de la résistance en pointe des pieux
dures and Design Methods. FHWA-HI-88-042, Federal Highway dans le sable à partir de l’essai SPT. In: The international geotech-
Administration, Washington, D.C nical conference, Beyrouth, pp 19–22
41. Eslami A, Fellenius BH (1997) Pile capacity by direct CPT and 46. MacKay DJC (1992) Bayesian interpolation. Neural Comput
CPTu methods applied to 102 cases histories. Can Geotech J 4(3):415–447
34:886–904
42. Meyerhof GG (1976) Bearing capacity and settlement of pile Publisher’s Note Springer Nature remains neutral with regard to
foundations. ASCE J Geotech Eng 102(3):1–19 jurisdictional claims in published maps and institutional affiliations.
43. Aoki N, Velloso D (1975) An approximate method to estimate the
bearing capacity of piles. In: Proceedings of the 5th Pan-Amer-
ican conference on soil mechanics and foundation engineering,
Buenos-Aires
44. Robert YA (1997) Few comments on pile design. Can Geotech J
34:560–567
13