Professional Documents
Culture Documents
Estimation of Liquid and Plastic Limit Using Artificial Neural Network Models
Estimation of Liquid and Plastic Limit Using Artificial Neural Network Models
Estimation of Liquid and Plastic Limit using Artificial Neural Network Models
Yeetendra Kumar
Research Scholar, Civil Engg. Deptt, MNNIT, Allahabad –211004, E-mail: yeetendra@rediffmail.com
Vijay Kumar
M.Tech Student, Civil Engg. Deptt, MNNIT, Allahabad –211004, E-mail: vijay03c34@gmail.com
K. Venkatesh
Asstt. Professor, Civil Engg. Deptt, MNNIT, Allahabad –211004, E-mail: venkatesh@mnnit.ac.in
ABSTRACT: The physical properties of clays are considerably influenced by the amount of water present in them. The
boundary water content at which the soil undergoes a change from one state to another is referred as consistency limits.
In practice the property of consistency is associated only with fine grained soils especially clays. The present study is
concentrated on estimation of liquid limit and plastic limit taking particle size distribution, SPT-N Value, natural
moisture content, specific gravity and dry density as independent variable using artificial neural network modeling
technique. The objective of this study is to investigate the practicability of using artificial neural network to model the
complex relationship between basic soil parameters with consistency limits. ANN models with varying input parameter
were used to investigate the best possible architecture to predict liquid limit and plastic limit. The analysis showed that
the results obtained from ANN models are in satisfactory agreement with the experimental results and ANN models may
be used to predict limits of the soil.
architecture, rules of learning and self organization (kurup through 75 and 2 micron IS sieve were selected one of the
et al 2002). input parameter. Likewise SPT-N value, natural moisture
content (w), specific gravity (G s) and dry density (ρd)may
Such neural networks have been widely used in be the factors which can influence consistency of soil, that
geotechnical engineering for modeling the soil and rock is the reason artificial neural network modeling used in
properties. In the past few years ANN modeling technique this study took aforementioned six independent variable as
got its way in swell pressure and soil suction behavior input vectors while as liquid limit (LL) and plastic limit
(Yusuf Erzin 2007), settlement prediction (Shahin et al (PL) as target vector. Five different models, in which the
2005) error prediction in triaxial apparatus (Dihoru et al. number of input variable changed from two to six as
2005), damage classification of bridges based on texture shown in table 1, were employed in the analysis. The
analysis (Kabir et al. 2008), stress strain behavior of sandy output of the models was liquid limit and plastic limit.
soils (Banimahd et al. 2005), and effective stress
parameter estimation (Kayadelen 2008) etc. Table 1. Network architecture with combination of input
SELECTION AND CONFIGURATIONS OF ANN S.No Inputs Structure Best
MODELS . structure
The feed-forward neural network that consists of 1 SPT-N , w 2-H1-2, 2-7- 9-2
multilayer perceptrons trained back-propagation 2-H1- H2-2,
algorithms were employed for this study. In the ANN 2 SPT-N , w, Gs 3-H1-2, 3-10- 8-2
analysis, dataset were generally normalized to obtain 3-H1- H2-2,
better convergence. Thus, the data set used in this study
3 SPT-N , w, Gs,ρd 4-H1-2, 4-12- 9-2
was scaled between 0 to 1 using the eq. (1) (Rafiq et al.
4-H1- H2-2,
2001; Kayadelen 2007; Gunadym 2009).
4 SPT-N , w, Gs, ρd , 5-H1-2, 5-14- 10-2
% passing 75 5-H1- H2-2,
micron
(1) 5 SPT-N , w, Gs, ρd , 6-H1-2, 6-14- 11-2
% passing 75 and 6-H1- H2-2,
2 micron
Range of normalization depends upon transformation Where H1 and H2 represents hidden layer 1 and 2
function. In the end of analysis, the network outputs were respectively (no. of neurons in H1 and H2 varies from 2 to
post processed to convert the data back into de-normalized 15)
units. In feed-forward neural networks, the data presented
to the input layer are propagated to the next layer by the TRAINING, TESTING AND VALIDATION PHASES
interconnections between the neurons. Simple processing Sixty percent of the total data were used for training phase
is performed on the data by the neurons of the receiving and remaining 40 percent data was used for testing and
layer prior to its being propagated on to the next layer. validation. training process was repeated for epochs 500,
This procedure is repeated until the transformed data reach 1000, 1500, 2000, 5000, 6000, 7000, 8000, 9000, 10000,
the output layer, completing the processing for that data 15000, 18000, 20000, 24000, and 25000 till the following
input. The ANN toolbox of MATLAB 7.4 computer added
conditions reached
software was used to perform the necessary computation
in which LR and momentum were kept constant whereas
(i) The minimum value of MSE reached
connection weights were kept adjustable automatically for
through
all the models. Models with single and double hidden
repeated training.
layer of varying numbers of neurons (2 to 15) were used in
(ii) The MSE of the training set is reached to
the analysis.
0.002
(iii)MSE started increasing after increasing epochs
PARAMETRIC ANALYSIS AND ARCHITECTURE (Noorzaei et al. 2007).
OF ANN MODELS
In practice the property of consistency is associated only RESULTS AND DISCUSSIONS
with fine grained soils especially clays. Some Next step after training process is the selection of best
classification systems draw the line between silt and clays networks architecture. Based on logical iteration, MSE and
based on particle size as determined from hydrometer test, regression plots, best architecture were defined for each
typically at 0.001 to 0.005 mm. Although such system can input combination as shown in table no 1. After selection
be useful, they can also be misleading because the biggest of best networks architecture, some datasets were used
difference between silt and clay is not their particle sizes, again to simulate the same network to check the
but their chemical structures (Coduto 2007). Keeping performance of best five networks for each input
above criterion in mind percentage of particle passing combination and to compare the predicted output with the
205 T11
actual output. Out of five, MSE plot of network 5-14-10-2 as shown in table no. 1, gave regression 0.686, 0.727,
is shown below: Figure 7. Comparison, based on graph 0.814, 0.857, & 0.832 for LL and regression 0.703, 0.740,
between actual and predicted value for the networks 1 to 5 0.836, 0.856 & 0.823 for PL (Figs 2a-6b)
Fig. 2a Comparison between actual and predicetd value of Fig. 3b Comparison between actual and predicetd value of
LL by model 2-7-9-2 PL by model 3-10-8-2
CONCLUSIONS
It may be concluded that the network containing one
hidden layer always underperformed in comparison to
network containing two hidden layer. Network 2-7- 9-2
gave the coefficient of regression 0.686 for LL and 0.703
for PL while as network 5-14- 10-2 gave the coefficient of
regression 0.857 for LL and 0.856 for PL, in between
these, network 3-10- 8-2 and 4-12- 9-2 gave the
Fig. 5b Comparison between actual and predicetd value of intermediate values of coefficient of regression that is the
PL by model 5-14-10-2 clear indication of betterment of network by increasing the
input from 2 to 5.
REFERENCES
Banimahd, M., Yasrobi, S.S. and Woodward, P.K. (2005).
Fig. 6a Comparison between actual and predicetd value of Artificial neural network for stress strain behavior of
LL by model 6-14-11-2 sandy soils: knowledge based verification. Computer
and Geotechniques 32: 377-386.
Coduto, D. P. (2007). Geotechnical Engineering. 1st Ed.,
Prentice Hall, New Delhi, India.
Dihoru, L., Wood, D. M., Sadek, T. and Lings, M. (2005).
A neural network for error prediction in a true triaxial
apparatus with flexible boundaries. Computer and
Geotechniques 32: 59-27.
Erzin, Y. (2007). Artificial neural network approach for
swell pressure versus soil suction behavior. Can.
Geotech. J. 44: 1215-1223.
Gulhati, S. K. and Dutta, M. (2008). Geotechnical Engg.
5th Ed., Tata McGraw-Hill, New Delhi, India.
Gunaydım, O. (2009).Estimation of soil compaction
parameters by using statistical analyses and artificial
Fig. 6b Comparison between actual and predicetd value of neural networks. Environ Geol 57:203-215.
PL by model 6-14-11-2 Kabir, S., Rivard, P. and Ballivy, G. (2008). Neural
205 T11