Professional Documents
Culture Documents
Prediction of Compression Index of Cohesive Soil Using Neural Network Approach
Prediction of Compression Index of Cohesive Soil Using Neural Network Approach
Prediction of Compression Index of Cohesive Soil Using Neural Network Approach
1
Research Scholar, Civil Engg. Deptt, MNNIT, Allahabad –211004, E-mail: yeetendra@rediffmail.com
2
Assistant Professor, Civil Engg. Deptt, MNNIT, Allahabad –211004, E-mail:venkatesh@mnnit.ac.in
3
Research Scholar, Civil Engg. Deptt, MNNIT, Allahabad –211004, E-mail: vijay03c34@gmail.com
Abstract: Compressibility is an indicator of amount of compression due to squeezing out of pore water from a
soil mass subjected to lasting static loads. Calculation of secondary consolidation settlements envisages compression
index of soil which is the slope of void ratio versus logarithm of effective stress. Consolidation test, to determine
compression index may use either disturbed or undisturbed soil specimen but uncertainty and imprecision in both
the cases is always associated due to variety of reasons. Present study is the outcome of applying a fairy new method
of modeling consolidation behavior of clay bearing soil using artificial neural networks. Particle sizes and shapes,
water retaining capacity and stress history are some of the parameters influencing consolidation. Soil properties
indicative of aforementioned factors from different locations were included as input vectors. Further permutation
and combination were applied to these inputs to observe their effect on compression index. The findings based on
statistical evaluation of ANN results and experimental results showed the successful application of ANN modeling
technique to predict compression index of the soil.
∂E
∆ w=η (10)
∂W
Figure 2 Neural network with 12 x n x 1 architecture
∂E The first layer shows inputs bring in to the network,
Where ∆ wis weight update; represents the
∂W second layer (namely hidden layer) presents varying
error gradient, and η is the learning rate which numbers of neurons, lines connecting first and second
determines the size of weight adjustment. layer are synaptic weights having adaptable nature
and the last layer presents output which is
There are many ways to improve the rate of compression index in this case. Sixty percent of the
convergence; first is by adjusting coefficient of total data were used for training phase and remaining
learning rate η to an optimal value; second is by 40 percent data was used for testing and validation.
adding a momentum term in the aforementioned error Networks were trained for epochs, varying from 500
gradient; and third is by normalizing dataset using to 5000 till the minimum value of MSE reached.
following equation [Rafiq 2001, Kayadelen 2008,
Gunaydım 2009]. Results and Discussion
As stated earlier through twelve input vectors
U actual −U min networks were developed and trained, separately for
ed = (11) Cc values obtained through eq. 2 & 5.
U max −U min
At the end of analysis, the network outputs were post Table 1a Regression & MSE Values from eq.2 based
processed to convert the data back into de-normalized models
units. The ANN toolbox in MATLAB 7.10 (R2010a) S. Structur Overall MSE %Absolute
computer added software was used to perform the No e Regression Error
necessary computation in which LR and momentum 1 12-4-1 0.998 0.002701 6.245
were kept constant whereas connection weights were 2 12-6-1 0.999 0.000722 7.537
kept adjustable automatically for all the models. 3 12-10-1 0.999 0.000348 6.848
4 12-12-1 1.000 0.000260 8.088
Development of Network Architecture 5 12-18-1 1.000 0.000188 6.080
Models with single hidden layer of varying numbers
of neurons (2 to 18) were used in the Table 1b Regression & MSE Values from eq.5 based
analysis.Fig.2describes the way network were treated models
from given set of input and target parameters. S. Structur Overall MSE %Absolute
No e Regression Error
d 1 12-4-1 0.998 0.000359 6.245
2 12-6-1 1.000 0.000094 2.089
N 3 12-8-1 1.000 0.000029 2.763
4 12-12-1 0.999 0.000276 11.674
e0 5 12-18-1 1.000 0.000128 9.493
N1
Dx Table 1a & b presents MSE values for some of the
optimal models obtained from networks as depicted
Dy N2 in fig. 2 for Cc values through eq.2 & 5 respectively.
Further these networks were simulated from twenty
Dz externally applied inputs and percentage absolute
N3 yk cc error (summarized in conjunction with regression and
wf MSE values in Table 1a & b) was calculated which
form the basis for opting best network. A few of
ρf absurd results (in this case only one out of twenty)
were omitted for minimizing percentage absolute
wL error.
wp Nn
c
φ
of number of input vectors directly or indirectly
0.25 significant to the compression characteristics of the
f(x) = 0.99 x + 0 soil. It may be seen from Table 1a that network
R² = 0.99 architecture attributed 12-18-1 proved best network
Pedicted Compression Index
0.23
based on percentage absolute error. Coefficient of
0.21 correlation obtained for compression index value
through eq. 2 and ANN model was R2 = 0.9904 and
0.19 linear correlation was y = 0.99x + 0.0011 which
shows closeness of ANN results to the standard ones.
0.17
Network architecture 12-06-1 (eq. 5 based results)
0.15 gave an absolute error of 2.089 % which is less than
eq. 2 based best architecture (12-18-1) error that is
6.080 %. It clearly indicates dependency of
0.13
compression index on multi variable equation rather
0.13 0.15 0.17 0.19 0.21 0.23 0.25
single variable equation. Coefficient of correlation
Actual Compression Index and linear correlation were R2 = 0.9965 and y =
Figure 3a. Cc through Terzaghi & Peck and ANN 1.0069x - 0.0006 respectively for this model.
References
0.2 1. Azzouz, A.S., Krizek, R.J. and Corotis, R.B.
f(x) = 1.01 x − 0 (1976). “Regression analysis of soil
R² = 1 compressibility.” Soils and Foundations,
Pedicted Compression Index
0.18
Jpn. Soc. Soil Mech. Found. Eng., 16(2), 19-
29.
0.16
2. Bowles J.E. (1989). “Physical and
0.14 geotechnical properties of soils.” McGraw-
Hill Companies, New York.
0.12 3. Bowles J.E., 1996. “Foundation analysis and
design.” 5thEd. McGraw-Hill Companies,
0.1 New York.
4. Bourmas, G. and Tsakiri, M. (2008).
0.08 “Comparing a univariate time series
0.08 0.1 0.12 0.14 0.16 0.18 0.2 approach with neural networks to predict
deformation of soil mass.” Measuring the
Actual Compression Index changes, A joint symposium of FIG and
IAG, May 12-15, 2008, Lisbon, Portugal.
5. Chua, C.G. and Goh, A.T.C. (2003). “A
hybrid Bayesian back-propagation neural
Figure 3b. Cc through Azzouz et al and ANN network approach to multi variate
modeling.” Int. Jl. Numer. Anal. Meth.
From the Table 1a & b above, one optimal network Geomech., John Wiley & sons, 27, 651-667.
was selected (highlighted once) and results (de- 6. Doris, J.J., Rizzo, D.M. and Dewoolkar,
normalized) obtained through ANN models are M.M. (2008). “Forecasting vertical ground
summarized in fig. 3a & b which essentially surface movement from shrinking/swelling
compares actual compression index value through soils with artificial neural network.” Int. Jl.
single variable and multi variable equation to the Numer. Anal. Meth. Geomech., John Wiley
predicted compression index values through ANN & sons, 32, 1229-1245.
models. 7. Gunaydım,O. (2009). “Estimation of soil
compaction parameters by using statistical
Conclusion analyses and artificial neural networks.”
Though compression index values determined Environmental Geology 57, 203-215.
through single and multi-variable equations showed 8. Haykin, S. (2006). “Neural Networks.” 2nd
noticeable difference but ANN gave very close Ed., Prentice Hall, New Delhi, India.
results in both the cases. It is due to prudent selection
9. Juang, C.C., Jiang, T. and Christopher, R.A.
(2001). “Three-dimensional site
characterization: neural network approach.”
Geotechnique 51(9), 799-809.
10. Kaya, A. (2009). “Residual & fully softened
strength evaluation of soils using artificial
neural network.” Journal of Geotechnical
and Geological Engineering, Vol.27, 281-
288.
11. 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.
12. Kurup, P.U. and Dudani N.K. (2002).
“Neural network for profiling stress history
of clays from PCPT data.” Jl. of Geotech.
and Geoenv. Engineering, ASCE, 128(7),
569-579.
13. Penumadu, D., Zhao, R. and Frost, D.
(2000). “Virtual geotechnical laboratory
experiments using a simulator.” Int. Jl.
Numer. Anal. Meth. Geomech., John Wiley
& sons, 24, 439-451.
14. Rafiq, M.Y., Bugmann, G. and Easterbrook,
D.J. (2001). “Neural network design for
engineering applications.” Comput. Struct.,
79, 1541-1552.
15. Rajshekhran, S. and Pai, G.A.V. (2010).
“Neural networks, fuzzy logic and genetic
algorithms.” PHI learning private limited,
New Delhi, India.
16. Schmertmann, J.H. (1953). “Estimating the
true consolidation behavior of clay from
laboratory test results.” Proc. ASCE 79,
Separate 311, pp 26.
17. Sezer, A. (2011). “Prediction of shear
development in clean sands by use of
particle shape information and artificial
neural networks.” Expert Systems with
Applications, Elsevier, 38, 5603-5613.
18. Terzaghi, K. and Peck, R.B. (1967). “Soil
mechanics in engineering practice.” 2nd
Ed.,Wiley, New York, p 73.