Professional Documents
Culture Documents
Undrained Lateral Load Capacity of Piles in Clay Using Artificial Neural Network
Undrained Lateral Load Capacity of Piles in Clay Using Artificial Neural Network
www.elsevier.com/locate/compgeo
Technical communication
Received 23 January 2006; received in revised form 9 August 2006; accepted 24 August 2006
Available online 17 October 2006
Abstract
This paper describes the application of the artificial neural network model to predict the lateral load capacity of piles in clay. Three
criteria were selected to compare the ANN model with the available empirical models: the best fit line for predicted lateral load capacity
(Qp) and measured lateral load capacity (Qm), the mean and standard deviation of the ratio Qp/Qm and the cumulative probability for
Qp/Qm. Different sensitivity analysis to identify the most important input parameters is discussed. A neural interpretation diagram is
presented showing the effects of input parameters. A model equation is presented based on neural network parameters.
Ó 2006 Elsevier Ltd. All rights reserved.
Keywords: Artificial neural network; Lateral pile load capacity; Statistical criteria; Sensitivity analysis
0266-352X/$ - see front matter Ó 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compgeo.2006.08.006
S.K. Das, P.K. Basudhar / Computers and Geotechnics 33 (2006) 454–459 455
Nomenclature
correlation of coefficient (R). But, it is difficult to assess the 2. Results and discussion
prediction of the model in terms of under-prediction or
over-prediction based on the R value. While describing pre- In the present study, the data are taken from Rao and
diction of pile load capacity based on cone penetration test Suresh Kumar [22]. The data consist of diameter of pile
(CPT), Briaud and Tucker [16] have emphasized that other (D), depth of pile embedment (L), eccentricity of load (e),
statistical criteria should be used along with the correlation undrained shear strength of soil (Su) as the inputs and
coefficient. Abu-Farsakh and Titi [17] have used the mean the lateral load as the output. In the present study as 38
(l) and standard deviation (r) of ratio of predicted pile data points are available, 29 data points were considered
capacity (Qp) to the measured pile capacity (Qm) as impor- for training and 9 for testing as shown in Table 1 and the
tant parameters in evaluating different models. Therefore, piles considered here are short and rigid piles [22]. All the
it will be desirable to have certain other statistical measures variables (inputs and output) are normalized in the range
that are unbiased, and have different form in order to test [1, 1] before training. A multilayer feed back propagation
the effectiveness of the developed models in terms of their neural network is used with hyperbolic tangent sigmoid
predictability criteria. function as the transfer function. The network is trained
The ANN is considered as a black box system as it is (learning) with Levenberg–Marquardt (LM) algorithm as
unable to explain the underlying principles of prediction it is efficient in comparison to gradient descent back prop-
and the effect of inputs on the output. Therefore, interpre- agation algorithm [18,23,24]. The ANN has been imple-
tation of weights may be considered to the subject of mented using Matlab [25] and its neural network toolbox
future research [18]. Recently, a number of investigators [24].
have advocated the use of connection weights to interpret The biggest challenge in successful application of ANN
the input variable contributions in neural networks [19– is when to stop training. If training is insufficient then the
21]. network will not be fully trained, where as if training is
With the above in view in the present study a feed for- excessive then it will memorize the training patter or learn
ward neural network based predictive model for capacity noise. So it will not generalize for a new set of data. In
of laterally loaded piles under undrained condition has training of a network as the number of epoch (iteration)
been developed. The three criteria are selected to compare increases there is decrease in error during training, but
the ANN model with the available empirical model: the for testing set data initially there is decrease in error up
best fit line for predicted load capacity (Qp) and measured to a certain iteration, and then the error goes on increasing
capacity (Qm), the mean and standard deviation of the or remains constant. Hence, the training is stopped when
ratio Qp/Qm and the cumulative probability for Qp/Qm. the testing phase error increases even though error during
A neural interpretation diagram has been proposed relat- the training phase may go on decreasing. This is known
ing the parameters affecting the pile capacity. A model as early stopping criteria [26,27] and is implemented using
equation has been presented based on the connection neural network toolbox of Matlab [25] by comparing the
weight. MSE value for training and testing simultaneously.
456 S.K. Das, P.K. Basudhar / Computers and Geotechnics 33 (2006) 454–459
Table 1
Summary of training and testing data for lateral load capacity of pile (from [22])
Training data Testing data
2
D (mm) L (mm) e (mm) Su (kN/m ) Qm (N) D (mm) L (mm) e (mm) Su (kN/m2) Qm (N)
6.35 146.10 19.10 38.80 69.50 13.50 190.00 0.00 24.00 128.00
13.00 260.00 0.00 24.00 225.00 20.40 300.00 50.00 3.40 38.00
12.50 130.00 0.00 24.00 106.00 18.40 300.00 50.00 3.40 42.50
13.50 300.00 50.00 3.40 30.00 25.40 300.00 50.00 4.00 58.00
13.50 300.00 50.00 4.00 36.00 13.00 132.10 33.80 38.80 53.00
13.50 300.00 50.00 5.50 50.00 18.00 300.00 50.00 4.00 49.00
13.50 300.00 50.00 7.20 64.00 18.00 300.00 50.00 5.50 65.00
18.00 300.00 50.00 10.00 89.00 18.00 300.00 50.00 7.20 87.00
18.00 300.00 50.00 3.40 39.00 12.30 300.00 50.00 4.00 35.00
20.40 300.00 50.00 4.00 46.00
12.30 300.00 50.00 5.50 44.00
18.40 300.00 50.00 4.00 51.00
18.00 300.00 50.00 10.00 116.50
33.30 300.00 50.00 3.40 78.50
33.30 300.00 50.00 5.50 110.50
12.30 300.00 50.00 3.40 29.50
6.35 139.70 25.40 38.80 65.50
12.30 300.00 50.00 7.20 58.00
12.30 300.00 50.00 10.00 81.00
18.40 300.00 50.00 5.50 65.50
18.40 300.00 50.00 7.20 86.50
18.40 300.00 50.00 10.00 114.00
20.40 300.00 50.00 5.50 59.50
20.40 300.00 50.00 7.20 76.50
20.40 300.00 50.00 10.00 87.00
25.40 300.00 50.00 7.20 90.00
25.40 300.00 50.00 10.00 151.61
25.40 300.00 50.00 3.40 50.00
25.40 300.00 50.00 5.50 75.00
The number of neurons in the hidden layer is varied and ing good correlation, the value of R2 shows the same to be
the minimum root mean square error (RMSE) value of very poor (0.22). In Fig. 1 comparison of the predicted pile
9.34 is observed with two hidden layer neurons. So with load capacity using ANN, Broms’ and Hansen’s method
four (4) inputs (D, L, e and Su) and two (2) hidden layer and their deviation from the observed one has been made.
neuron and single (1) output, the ANN is described as a It is observed that the ANN predicted values are less
4–2–1 network architecture. The results of ANN models scattered and are close to observed values signified by its
are compared with that of Brom’s and Hansen method in
terms of the correlation coefficient (R) and coefficient of
efficiency (R2). The R2 is defined as 250
E1 E2
R2 ¼ ð1Þ ANN Training data
E1 200 ANN Testing data
Broms Method
Predicted load capacity (Qp )
closeness to the line of equality. The mean (l) and standard followed by the Broms method and ANN. This reveals that
deviation (r) of Qp/Qm are important indicators of the even though the ratio of Qp/Qm are not very much scattered
accuracy and precision of the prediction method [17]. the mean value (0.58) is far from the expected value (1.0).
Under ideal conditions an accurate and precise method However, when comparing P50 value it can be seen ANN
gives the mean value as 1.0 and the standard deviation to model is the best one (P50 = 1.01 for training and 1.001 test-
be 0. A l value greater than 1.0 indicates over prediction ing) followed by the Broms method (P50 = 1.124). The ANN
and under prediction otherwise. The best model is repre- model having P90 values (P90 = 1.17 for training and 1.156
sented by l value close to 1.0 and r close to 0. The other for testing) close to 1.0 shows the effectiveness of the ANN
method like cumulative probability of the Qp/Qm [17] has prediction compared to Broms method (P90 = 1.381) and
been also considered for the evaluation of different models. Hansen method (P90 = 0.835).
The ratio Qp/Qm is arranged as per their values and the The weights and biases of the final network are pre-
cumulative probability is calculated from the following sented in Table 2. The weights and biases can be utilized
equation: for sensitivity analysis and framing an ANN model in
i equation form. The same will be discussed in following
P¼ ð2Þ sections.
nþ1
where i is the order number given to the Qp/Qm ratio; n is 3. Sensitivity analysis
the number of data points.
If the computed value of 50% cumulative probability (P50) Sensitivity analysis is of utmost concern for selection of
is less than unity, under prediction is implied; values greater important input variables. Different approaches have been
than unity means over prediction. The ‘best’ model is corre- suggested to select the important input variables. The Pear-
sponding to the P50 value close to unity. The 90% cumulative son correlation coefficient is defined as one of the variable
probability (P90) reflects the variation in the ratio of Qp/Qm ranking criteria in selecting proper inputs for the ANN
for the total observations. The model with Qp/Qm close to 1.0 [19,28], Goh [29] and Shahin et al. [27] have used Garson’s
is the better model. Fig. 2 shows the variation of Qp/Qm with algorithm [30] in which the input-hidden and hidden-
cumulative probability (%) for different ANN model with output weights of trained ANN model are partitioned
both training and testing data and the results have been com- and the absolute values of the weights are taken to select
pared with that of empirical methods. Based on l value, the important input variables. So it does not provide infor-
ANN in both training (1.04) and testing (1.10) show good mation on the effect of input variables in terms of directly
prediction followed by the Broms method (1.148). The value or inversely related to the output, and details with the
of r (0.117) is found to be minimum for the Hansen method example have been presented in Goh [29]. Olden et al.
[21] presented a connection weights approach in which
1.8 the actual values of input-hidden and hidden-output
Hansen weights are taken. It adds the products across all the hid-
1.6 Broms den neurons. The results of such a study using different
ANN approaches are presented here as follows.
1.4 Table 3 shows the cross correlation of inputs with the Qm
Qp/Qm
0.8 Table 3
Cross correlation between inputs and the measured load capacity
0.6 D L e Su Qm
P90
P50 D 1.00 0.50 0.43 0.52 0.11
0.4 L 1.00 0.79 0.93 0.18
0 20 40 60 80 100
e 1.00 0.76 0.57
Cumilative Probability (%) Su 1.00 0.36
Qm 1.00
Fig. 2. Cumulative probability plots of Qp/Qm for different methods.
Table 2
Connection weights and biases for pile capacity (Qp)
Neuron Weights (wik) Biases
Input 1 Input 2 Input 3 Input 4 Output bhk b0
Hidden neuron 1 (k = 1) 1.4121 0.2503 0.6842 3.7774 2.2413 2.9916 0.8645
Hidden neuron 2 (k = 2) 0.7871 1.9661 2.5228 1.8843 2.0593 1.6005 –
458 S.K. Das, P.K. Basudhar / Computers and Geotechnics 33 (2006) 454–459
Table 4
Relative importance of different inputs as per Garson’s algorithm and connection weight approach
Parameters Garson’s algorithm (%) Connection weight approach
Relative Ranking of inputs as Sj values as per Connection Ranking of inputs as per
importance per relative importance weight approach relative importance
D 17.02 3 1.54 4
L 15.78 4 4.61 1
E 23.20 2 3.66 3
Su 44 1 4.59 2
are for short piles, it can be concluded that the cross corre- 1
lation does not show the physical relationship between
input and output. The sensitivity analysis for the model as
per Garson’s method [29,30] and Olden et al. [21] connec- 2 A
tion weight approach to find out important input parame-
ters are presented in Table 4. The Su is found to be the Qp
most important input parameter followed by e, D and L O
3
as per Garson’s method. It can also be seen that as per
the Olden et al. [21] method, L and Su are equally, the most
important input parameters followed by e and D. So it can B 1- D
4 2- L
be concluded that the interpretation of the weights to find 3- e
out the important input parameters based on Garson’s 4- Su
algorithm [29,30] and connection weight approach [21]
matches the physical meaning for lateral load carrying
Fig. 3. The NID showing axons representing connection weights and
capacity of piles. effects of inputs on Qp.
weights by grey lines and the thickness of the lines is pro- Qpn ¼ fsig b0 þ wk fsig bhk þ wik X i ð3Þ
k¼1 i¼1
portional to their magnitude. The relationship between
the input and output is determined in two steps. Positive where Qpn is the normalized (in the range 1 to 1 in this
effect of the input variables is depicted by positive input- case) Qp value; b0 is the bias at the out put layer; wk is
hidden and positive hidden-output weights, or negative the connection weight between kth neuron of hidden layer
input-hidden and negative hidden-output weights. The and the single output neuron; bhk is the bias at the kth neu-
positive input-hidden and negative hidden-output and neg- ron of hidden layer; h is the number of neurons in the hid-
ative input-hidden and positive hidden-output weight indi- den layer; wik is the connection weight between ith input
cate the negative effect of the input variables. So unlike variable and kth neuron of hidden layer; Xi is the normal-
absolute multiplication of weights, in this case multiplica- ized input variable i in the range [1, 1] and fsig is the sig-
tion of actual weights of input-hidden and hidden-output moid transfer function.
indicates the effect of that input variable on the output. Hence, the model equation for the output can be formu-
The input having positive effect on the output is repre- lated based on the trained weights of the ANN model. In
sented with grey circles and that having negative effect with this study, such a model equation for lateral load carrying
white circles. For the present example with the weights as capacity of pile was established using the values of the
obtained and shown in Table 2, a NID is presented as weights and biases shown in Table 2 as per the following
shown in Fig. 3. It can be seen from Fig. 3 that the inputs expressions.
D, L and Su have positive contribution to the Qp values
and e has negative effects on the Qp value. Thus it inferred A1 ¼ 2:9916 þ 1:4121D þ 0:2503L þ 0:6842e þ 3:7774S u
that D, L and Su are directly and e is indirectly propor- ð4Þ
tional to Qp value. So it can be seen that NID is an effective
A2 ¼ 1:6005 0:7871D þ 1:9661L 2:5228e 1:8843S u
method in indicating the physical relationship between
inputs with the output. ð5Þ
S.K. Das, P.K. Basudhar / Computers and Geotechnics 33 (2006) 454–459 459
eA1 eA1 [5] Broms BB. Lateral resistance of piles in cohesive soils. J Soil Mech
B1 ¼ 2:2413 ð6Þ Found Eng, ASCE 1964;90(SM2):27–63.
eA1 þ eA1
[6] Meyerhof GG. Bearing capacity and settlement of pile foundations.
eA2 eA2 J Geotech Engrg, ASCE 1976;102(3):196–228.
B2 ¼ 2:0593 A ð7Þ
e 2 þ eA2 [7] Goh ATC. Empirical design in geotechnics using neural networks.
Geotechnique 1995;45(4):709–14.
C 1 ¼ 0:8645 þ B1 þ B2 ð8Þ
[8] Goh ATC. Modeling soil correlations using neural networks.
C1 C 1
e e J Comput Civil Engrg, ASCE 1995;9(4):275–8.
Qpn ¼ ð9Þ [9] Goh ATC. Pile driving records reanalyzed using neural networks.
eC1 þ eC1
J Geotech Engrg, ASCE 1996;122(6):492–5.
The Qpn value as obtained from Eq. (9) is in the range [10] Chan WT, Chow YK, Liu LF. Neural network: an alternative to pile
[1, 1] and this needs to be denormalized as driving formulas. J Comput Geotech 1995;17:135–56.
[11] Teh CI, Wong KS, Goh ATC, Jaritngam S. Prediction of pile
Qp ¼ 0:5ðQpn þ 1ÞðQp max Qp min Þ þ Qp min ð10Þ
capacity using neural networks. J Comput Civil Eng, ASCE
where Qp max and Qp min are the maximum and minimum 1997;11(2):129–38.
[12] Rausche F, Moses F, Goble GG. Soil resistance predictions from pile
values of Qp respectively in the data set. dynamics. J Soil Mech Found Div, ASCE 1972;98:917–37.
[13] Lee IM, Lee JH. Prediction of pile bearing capacity using artificial
6. Conclusions neural networks. Comput Geotech 1996;18(3):189–200.
[14] Abu-Kiefa MA. General regression neural networks for driven piles
The following conclusions can be drawn from the above in cohesionless soils. J Geotech Geoenv Engrg, ASCE 1998;124(12):
1177–85.
studies: [15] Nawari NO, Liang R, Nusairat J. Artificial intelligence techniques
for the design and analysis of deep foundations. Electron J
(1) The three criteria that were selected to compare the Geotech Eng 1999. http://geotech.civeng.okstate.edu/ejge/ppr9909/
performances of developed ANN model with the index.html.
available empirical model (the best fit line for pre- [16] Briaud JL, Tucker LM. Measured and predicted axial response of 98
piles. J Geotech Eng 1988;114(9):984–1001.
dicted load capacity (Qp) and measured capacity [17] Abu-Farsakh MY. Assessment of direct cone penetration test
(Qm), the mean and standard deviation of the ratio methods for predicting the ultimate capacity of friction driven piles.
Qp/Qm and the cumulative probability for Qp/Qm) J Geotech Geoenv Eng 2004;130(9):935–44.
showed that, the developed ANN model is more effi- [18] Goh ATC, Kulhawy FH, Chua CG. Bayesian neural network
cient compared to empirical models of Hansen and analysis of undrained side resistance of drilled shafts. J Geotech
Geoenv Eng, ASCE 2005;131(1):84–93.
Broms. [19] Wilby RL, Abrahart RJ, Dawson CW. Detection of conceptual
(2) The sensitivity analysis based on cross correlation model rainfall-runoff processes inside an artificial neural network.
does not reflect the physical relationship between Hydrol Sci 2003;48(2):163–81.
the input and output. As per Garson’s algorithm [20] Olden JD, Jackson DA. Illuminating the ‘‘black box’’: understanding
undrained shear strength, and according to connec- variable contributions in artificial neural network. Eco Model
2002;154:135–50.
tion weight approach, length and the undrained shear [21] Olden JD, Joy MK, Death RG. An accurate comparison of methods
strength are almost equally important inputs. Hence, for quantifying variable importance in artificial neural networks using
it matches the physical meaning for lateral load carry- simulated data. Eco Model 2004;178(3):389–97.
ing capacity of short piles. [22] Rao KM, Suresh Kumar V. Measured and predicted response of
(3) Based on NID and connection eight approaches it is laterally loaded piles. In: Proceedings of the sixth international
conference and exhibition on piling and deep foundations, India;
observed that lateral load capacity of pile is increase 1996. p. 1.6.1–1.6.7.
with increase in diameter, length and un-drained [23] Juang CH, Elton DJ. Prediction of collapse potential of soil with
cohesive strength of soil and decreases with increase neural networks. Trans Res Record 1997;1582:22–8.
in eccentricity of load application. [24] Demuth H, Beale M. Neural network toolbox. The MathWorks, Inc.;
(3) A model equation is presented based on the trained 2000.
[25] MathWork Inc. Matlab User’s Manual. Version 6.5. Natick (MA):
weights of the ANN. MathWork Inc.; 2001.
[26] Basheer IA. Empirical modeling of the compaction curve of cohesive
References soil. Can Geotech J 2001;38(1):29–45.
[27] Shahin MA, Maier HR, Jaksa MB. Predicting settlement of shallow
[1] Poulos HG, Davis EH. Pile foundation analysis and design. New foundations using neural network. J Geotech Geoenv Eng, ASCE
York: Wiley; 1980. 2002;128(9):785–93.
[2] Matlock H, Reese LC. Generalized solutions for laterally loaded [28] Guyon I, Elisseeff A. An introduction to variable and feature
piles. Trans ASCE 1962;127:1220–48. selection. J Mach Learn Res 2003;3:1157–82.
[3] Portugal JC, Seco e Pinto PS. Analysis and design of pile under lateral [29] Goh ATC. Seismic liquefaction potential assessed by neural network.
loads. In: Proceedings of the 11th international geotechnical seminar J Geotech Eng, ASCE 1994;120(9):1467–80.
on deep foundation on bored and auger piles, Belgium; 1993. p. 309– [30] Garson GD. Interpreting neural-network connection weights. Artif
13. Intel Expert 1991;6(7):47–51.
[4] Hansen B. The ultimate resistance of rigid piles against transversal [31] Ozesmi SL, Ozesmi U. An artificial neural network approach to
force. Copenhagen: Danish Geotechnical Institute; 1961. Bulletin No. spatial modeling with inter specific interactions. Eco Model
12. p. 5–9. 1999;116:15–31.