ARTIFICIAL NEURAL NETWORK APPROACH FOR ESTIMATION OF SHEAR - STRENGTH PARAMETERS OF SOIL

Abstract

The shear strength parameters of soil such as cohesion (c) and internal friction angle () govern important problems like bearing capacity of deep or shallow foundations, hill slope stability as well as slopes of dams and embankments, design of retaining structures etc. in most of the civil engineering practices. The shear strength of a soil mass is the internal resistance per unit area that the soil mass can offer to resist failure and sliding along any plane inside it. In geotechnical design practice, two important considerations that need careful examination are whether construction will cause deformation of the soil and/or instability due to shear failure. Therefore, an engineer has to ensure that the structure is safe against shear failure in the soil that supports it and does not undergo excessive settlement. Knowledge about the stress-strain behaviour, deformation and shear strength of the soil is essential. In order to determine these shear strength parameters of soil, laboratory tests such as direct shear test and the triaxial test are used. In this study, artificial neural network (ANN) with the feed-forward back-propagation algorithm is proposed for predicting the shear strength parameters of the soils. To develop this ANN based model, different soil parameters such as gravel % (GP), sand % (SP), silt % (STP), clay % (CP), dry density (DD) and plasticity index (P1) obtained through laboratory tests for soil samples from different parts of India, have been used as input parameters to predict the shear parameters of the soil. Fuzzy c-means (FC) clustering technique has been adopted for classifying the total dataset into training, testing and verification dataset. Different neural network architectures have been designed with single hidden layer and double hidden layers by varying the number of neurons. Levenberg-Marquardt algorithm was used in the training and testing of these different neural architectures. The training dataset was used to train various network architectures while the testing dataset was used to control the overtraining of the network and to evaluate the accuracy of the networks. The performance of the networks was evaluated by determining both the training and testing data accuracies in terms of correlation coefficient (R) and root mean squared error (RMSE. in case of data sampling using fuzzy clustering method, the neural network with architecture 6-9-2 having single hidden layer produced best correlation coefficient(R) and RMSE values for both training and testing. The R values for training and testing in this case are 0.726 iv and 0.860 respectively for prediction of both C and 0. While for double hidden layers case, the neural network 6-1-3-2 produced best results in terms of R and RMSE values. The R values in Nothis case for training and testing are 0.764 and 0.789 respectively for prediction of both C and . The results reveal that the performance of ANN model remains unaffected irrespective of the number of hidden layers (single or double) used in this study, though the neural architecture differs in both the cases. Also, it can be inferred from the results that indirect estimation of shear strength parameters of soils is feasible with reasonable accuracy using ANN technique.