Professional Documents
Culture Documents
Yeetendra Kumar
Research Scholar, Civil Engg. Deptt, MNNIT, Allahabad –211004, E-mail: yeetendra@rediffmail.com
Vijay Kumar
Research Scholar, 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
Key words: Unified soil classification system, artificial neural network, percentage of particle sizes, liquid limit, plastic
limit, plasticity index
ABSTRACT: Percentage of particle sizes, liquid limit, plastic limit and plasticity index are some of the basic
information of soil needed to classify it in Unified Soil Classification System (USCS). Though aforementioned
properties are sufficient for engineering classification of soil but we may include other properties also which may cause
behavioral alteration in a specific soil. Present study is an approach of applying Artificial Neural Network (ANN)
modeling technique by embracing some physical properties indicator of in-situ state of soil and some index properties
of soil to predict specific class. Though it is site specific phenomenon but efforts were made to include both classes
that are coarse and find grained soils in the study. The results obtained from ANN models found satisfactory in
agreement with the Unified Soil Classification System and may be used to predict specific soil class. Further it is
suggested that incorporating other soft computing tools in connection with ANN may prove far better.
ID No./ pp. 5
Geotec Hanoi 2011 October ISBN 978-604-82-000-8
This system divides the soil into two main categories 3.2 Back-Propagation Neural networks
coarse grained and fine grained soil and their behavior is We may understand the BP system by the following
controlled by grain size distribution and plasticity diagram.
characteristics. There are in all 15 groups of soils in dj(n)
which 8 (GW;GP;GM;GC;SW;SP; SM and SC) belongs
to coarse grained soil and 6 (ML;CL;OL;MH; CH and
OH) and remaining one (Pt) is categorized as peat. Based Y 0= +1
on cumulative mass retained on 0.075 mm sieve (out of wj0(n)= bj(n)
coarse fraction) coarse grained soil is again divided into
gravel and sands. Further narrow downing of the gravel ej(n)
v j(n) φ( ) Yj(n) -1
and sand is done on the basis of boundary conditions of wji(n)
fines i, e, if fraction passing No. 200 sieve < 5% group yi(n)
symbol may be GW:GP:SW:SP, if >12% group symbol
may be GM:GC:SM:SC, if 12% ≤ fraction passing No.
200 sieve ≥ 5%use of dual symbol required. Other than
above mentioned rules use of plasticity chart devised by
Casagrande (1948), coefficient of uniformity C u,
coefficient of curvature Cc and plasticity index Ip is
recommended.
Contrary to coarse grained soils, fine grained soils are
divided solely by the plasticity chart itself and further Figure 1 Signal flow graph of neuron j.
subdivision are made on the basis of plasticity
characteristics. The A – line in the plasticity chart The error signal ej(n) of neuron j at iteration n may be
possessing the equation calculated by
is generally used for separating the
clay from silt and organic soils from inorganic soils.
(1)
3. ARTIFICILIAL NEURAL NETWORK
The error energy of neuron j may be ½ e j2 (n). Thus total
Artificial neural networks can be most adequately error energy for all neurons in the output layer
characterized as ‘computational models’ with particular
properties such as the ability to adapt or learn, to
ξ
generalize or to cluster or organize data and in which
(2)
operation is based on parallel processing. Task of
generalization is achieved by training and testing through
Where set C includes all the neurons in the output layer
division of input and target vectors into two datasets.
of the network. Let N denote the total number of patterns
contained in the training set. The average squared error
3.1 Categories of Neural Network
energy is obtained by summing 𝛏 (n) over all n and then
Two types of networks are commonly used for function normalizing with respect to the size N
approximation. First is a back-propagation network and
second is a radial basis neural network. The back-
propagation algorithm for the design of multilayer (3)
perceptron may be viewed as the application of recursive
technique known in statics as stochastic approximation The objective of learning process is to adjust the free
wherein radial basis neural networks is like curve fitting parameters (like synaptic weights and bias levels) to
(approximation) problem in a high dimensional space minimize 𝛏av.
(Haykin 2006). In ongoing literature the discussion is
limited to BP network as it is used for model 4. STRATEGY FOR DATA & NETWORK
development. SELECTION
ID No./ pp. 5
Geotec Hanoi 2011 October ISBN 978-604-82-000-8
yk2
f
N3
f
wP
(4)
Where ∆w is weight update; 𝜕E/𝜕w represents the error
gradient, and η is the learning rate which determines the
size of weight adjustment. Ip
There are many ways to improve the rate of
convergence; first is by adjusting coefficient of learning Figure 2 Neural network with 9 x n x 1 architecture
rate η to an optimal value; second is by adding a
momentum term in the aforementioned error gradient; The first layer presents inputs bring in to the network,
and third is by normalizing dataset using following second layer namely hidden layer shows neurons in
equation (Rafiq et al. 2001; Kayadelen 2007; Gunadym processing ( in this case varying between 2 to 18 ),
2009). connecting links in between first and second layer are
synaptic weights with adaptable nature and the last layer
presents output that is four dimensional nature in this
case and discussed in next article. Networks were trained
ID No./ pp. 5
Geotec Hanoi 2011 October ISBN 978-604-82-000-8
for varying number of iterations till the minimum value vectors. Though bore hole depth up to 10 meters was
of MSE reached. considered for analysis but due to data from a single site
were adopted into the models as shown in fig. 2, there
5. DEMARCATION OF SOIL CLASS found less heterogeneity in soil showing slight variations
into the soil parameters and finally giving better results
ANN is designed to work within certain intervals of input as some of them are summarized in Table 3 below.
and output depending on transformation function. To The architecture 6-9-4 yielded comparatively better
meet this criterion soil class in USCS system is allotted results, MSE error from this models was 0.00379,
certain numbers or coded taking percentage of fine showing better performance.
contents in each soil class and plasticity characteristics as
the basis for demarcation. From the study of the bore - Table 3 ANN based results of some soil samples.
log chart, standard pattern of the percentage of particle of Dx Dy Dz wL wP IP Soil
different sizes, Atterberg limit and plasticity value were class
marked and designated with some values as shown in 87 45 1.3 0 0 0 SM
Table 1. 86 47 2.5 0 0 0 SM
100 80 7 26 19 7 CL
94 74 3.8 0 0 0 ML
Table 1 Demarcation of soil class
99 94 15 36 19 17 CI
SM ML CL CI
91 68 3.3 0 0 0 ML
Dx 85 90 95 100 98 85 6.8 32 20 12 CL
Dy 50 70 85 90 89 85 8.1 30 21 9 CL
Dz 1.5 3.5 10 17 96 79 13 36 18 18 CI
wL 0 0 27 37 93 72 3.2 0 0 0 ML
wP 0 0 17 21
IP 0 0 10 15 7. CONCLUSIONS
As shown in Table 1 above four kind of soil that are silty The strategy adopted for soil class designation found
sand (SM), inorganic silt (ML), inorganic clay of satisfactory though some adjustments were made at early
medium plasticity (CI) and inorganic clay of low to stages like, dissimilar input vectors were removed at
medium plasticity (CL)were used as learning objects for beginning itself and some of abrupt results were ignored.
the networks, and the input vectors as shown in fig. 2 This is because ANN works with the numbers not with
were used for training. To separate the output signal from the ranges or partitioned data. Use of computational
each other, the outputs of the four soil class were defined methods that works well with predefined ranges of input
as shown in Table 2. and target parameters (like fuzzy logic and
neurozzyfuzzy technique) may yield better results
Table 2 Learning paradigm without any compromise at early stages.
Dx Dy Dz wL wP IP Soil Four dimensional
class output 8. REFETRENCES
85 50 1.5 0 0 0 SM 0.9 0.1 0.1 0.1
90 70 3.5 0 0 0 ML 0.1 0.9 0.1 0.1 Baziar, M.H. &Ghorbani.A. (2005). Evaluation of lateral
95 85 10 27 17 10 CL 0.1 0.1 0.9 0.1 spreading using artificial neural networks. Soil
100 90 17 37 21 15 CI 0.1 0.1 0.1 0.9
Dynamics and earthquake engineering. 25: 1-9.
6. RESULTS AND DISCUSSIONS Cal, Y. (1995). Soil classification by neural
network.Advances in Engineering Software. 22: 95-97.
The results were interpreted on the basis of four Chua, C.G. and Goh, A.T.C. (2003). A hybrid Bayesian
dimensional outputs instead of single value, for example back-propagation neural network approach to
for any output value 0.0256, 0.854, 0.139 & 0.265 the multivariate modeling. Int. Jl. Numer.
maximum value is in set two that is 0.854 and since this Anal.Meth.Geomech., John Wiley & sons, 27, 651-
output matches with four dimensional output pattern of 667.
inorganic silt (ML) that is 0.1, 0.9, 0.1 & 0.1 so the same
grade of soil will be preferred for corresponding input
ID No./ pp. 5
Geotec Hanoi 2011 October ISBN 978-604-82-000-8
Goh, A.T.C. &Kulhawy, F.H. (2003). Neural network artificial neural networks. Expert Systems with
approach to model the limit state surface for reliability Applications.2 38: 5603-5613.
analysis.Can. Geotech. J., 4: 1235-1244. Shahin, M.A., Jaksa, M. B. &Maier, H.R. (2005). Neural
Goktepe, F., Arman, H. & Pala, M. (2010). A new network based stochastic design charts forsettlement
prediction.Can. Geotech. J., 42: 110-120.
approach for classification of clayey soil: a case study
Shangguan, Z., Li, S. &Luan, M. (2009).Intelligent
for Adapazari region, Turkey. Scientific Research and
forecasting method for slope stability estimation by
Essays. 5(15): 2037-2043.
using probabilistic neural network. The Electronic
Haykin, S. (2006), Neural Networks, 2nd Ed., Prentice
Journal of Geotechnical Engineering.13:1-10.
Hall, New Delhi, India.
Kim, C.Y., Bae, G.J., Hong, S.W., Park, C.H., Moon, Wan, S. & Yen, J.Y. (2006). The study of SSI problems
H.K. & Shin, H.S. (2001). Neural network based in an industrial area with modified neural network
prediction of ground surface settlements due to approaches. Int. J. Numer. Anal. Meth. Geomech. ,
tunneling. Comp. and Geot., 28: 517:547. 32(9), 1087-1106.
Kayadelen, C. (2008), Estimation of effective stress
parameter of unsaturated soils by using artificial neural
networks. Int. J. Numer. Anal.Meth.Geomech.32(9),
1087-1106.
Rajshekhran, S. and Pai, G.A.V. (2010), Neural
networks, fuzzy logic and genetic algorithms. PHI
learning private limited, New Delhi, India.
Rafiq, M.Y., Bugmann, G. and Easterbrook, D.J. (2001),
Neural network design for engineering applications.
Comput.Struct., 79, 1541-1552.
Sezer, A. (2011), Prediction of shear development in
clean sands by use of particle shape information and
ID No./ pp. 5