Computational & Technology Resources
an online resource for computational,
engineering & technology publications
Civil-Comp Proceedings
ISSN 1759-3433
CCP: 74
PROCEEDINGS OF THE SIXTH INTERNATIONAL CONFERENCE ON THE APPLICATION OF ARTIFICIAL INTELLIGENCE TO CIVIL AND STRUCTURAL ENGINEERING
Edited by: B.H.V. Topping and B. Kumar
Paper 33

Prediction of Collapse Potential via Artificial Neural Networks

G. Habibagahi and M. Taherian

Department of Civil Engineering, Shiraz University, Iran

Full Bibliographic Reference for this paper
G. Habibagahi, M. Taherian, "Prediction of Collapse Potential via Artificial Neural Networks", in B.H.V. Topping, B. Kumar, (Editors), "Proceedings of the Sixth International Conference on the Application of Artificial Intelligence to Civil and Structural Engineering", Civil-Comp Press, Stirlingshire, UK, Paper 33, 2001. doi:10.4203/ccp.74.33
Keywords: geotechnical, collapse, neural network, unsaturated soils, volume change.

Summary
Collapse defined as the additional deformation of soils when wetted is believed to be responsible for damage to the buildings resting on compacted fills as well as failure in embankments and earth dams. Despite the considerable amount of work done in this area, the functional relationship between various soil parameters and the amount of collapse deformation is not well established and the exact interrelationship is still a matter of speculation. Neural network as a computational tool has proved to be capable of establishing a relationship between a series of input data and the corresponding outputs no matter how complex this relationship may be. Hence, this method is employed in this paper to investigate the collapse potential of unsaturated soils.

In this paper, three different types of neural networks, namely, conventional back- propagation neural network (BPNN), recurrent neural network (RNN), and generalized neural network (GRNN) are employed as computational tools to predict the amount of collapse and to investigate the influence of various parameters on the collapse potential. In order to arrive at a robust neural network, a comprehensive database is required apriori. Therefore, 192 series of single oedometer test were carried out on three soils with different initial conditions and inundated at different applied pressures to serve as the required database. The soils tested were classified as clay with low plasticity having different gradation curves. Initial dry densities of the samples varied from to and initial water contents were varied from 4.9% to 16.9%. Collapse was measured at inundation pressures of 100, 200, 400 and 800 kPa. Furthermore, other similar data available in the literature were included to arrive at a more comprehensive database. To this end, a total of 138 sets of data from test results performed on eight different soils and reported in the literature were added to the database to arrive at a total of 330 sets of data.

A single hidden layer was adopted for BPNN and RNN. Currently, there is no rule to determine the optimum number of hidden neurons. However, there are two approaches to arrive at the optimum number of hidden neurons. The first approach starts with a network with a large number of hidden neurons and then "pruning" the network by reducing the number of hidden neurons to arrive at the final network architecture. The second approach, on the contrary, starts with a network with a minimal number of neurons in the hidden layer and increases the network size in steps by adding a single hidden neuron each time and examining the network performance. This process is continued until there is no further improvement in the network performance. In this study, the latter approach was adopted to determine the number of hidden neurons. Based on this approach, a network with six neurons had the best performance for back-propagation type network, BPNN6, and a network with four neurons had the best performance for recurrent type network, RNN4. GRNN has a fixed number of neurons in the hidden layer, equal in number to the number of training datasets present in the database. BPNN6 had the best performance in predicting the testing datasets while GRNN had the best performance for the training datasets. Generalization capability is of utmost importance in any modelling technique. BPNN6 showed the best prediction capability for testing datasets (best generalization capability) and having a reasonable performance for the training datasets as well. Therefore, BPNN6 was selected as the superior network for assessing the collapse potential of unsaturated soils.

Next, efficacy of the selected neural network was verified by comparing the predicted results with some of the existing empirical relationships. These empirical relationships had been obtained by different investigators via statistical analysis of the collapse data. The comparison indicates superiority of the proposed approach in terms of the accuracy of prediction. Moreover, by analysing the network connection weights, relative importance of different parameters on collapse potential was assessed. Based on this analysis, for a given soil type, the initial dry unit weight, , is the most important factor influencing collapse potential followed by pressure at wetting and initial water content. It was also concluded that relative importance of soil properties affecting collapse potential, in a descending order, are Atterberg limits (liquid and plastic limits) followed by coefficient of uniformity and clay content. Coefficient of curvature, , is the least important factor. The proposed network is therefore a suitable tool to assess different placement conditions (initial dry density and placement water content) as well as stress level on collapse potential for a given soil. On the other hand, the network may be used to assess effectiveness of a soil improvement procedure on collapse potential by performing an appropriate parametric study.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description