Modelling pile capacity and load-settlement behaviour of piles embedded in sand & mixed soils using artificial intelligence

Abstract

This thesis presents the development of numerical models which are intended to be used to predict the bearing capacity and the load-settlement behaviour of pile foundations embedded in sand and mixed soils. Two artificial intelligence techniques, the gene expression programming (GEP) and the artificial neural networks (ANNs), are used to develop the models. The GEP is a developed version of genetic programming (GP). Initially, the GEP is utilized to model the bearing capacity of the bored piles, concrete driven piles and steel driven piles. The use of the GEP is extended to model the load-settlement behaviour of the piles but achieved limited success. Alternatively, the ANNs have been employed to model the load-settlement behaviour of the piles.The GEP and the ANNs are numerical modelling techniques that depend on input data to determine the structure of the model and its unknown parameters. The GEP tries to mimic the natural evolution of organisms and the ANNs tries to imitate the functions of human brain and nerve system. The two techniques have been applied in the field of geotechnical engineering and found successful in solving many problems.The data used for developing the GEP and ANN models are collected from the literature and comprise a total of 50 bored pile load tests and 58 driven pile load tests (28 concrete pile load tests and 30 steel pile load tests) as well as CPT data. The bored piles have different sizes and round shapes, with diameters ranging from 320 to 1800 mm and lengths from 6 to 27 m. The driven piles also have different sizes and shapes (i.e. circular, square and hexagonal), with diameters ranging from 250 to 660 mm and lengths from 8 to 36 m. All the information of case records in the data source is reviewed to ensure the reliability of used data.The variables that are believed to have significant effect on the bearing capacity of pile foundations are considered. They include pile diameter, embedded length, weighted average cone point resistance within tip influence zone and weighted average cone point resistance and weighted average sleeve friction along shaft.The sleeve friction values are not available in the bored piles data, so the weighted average sleeve friction along shaft is excluded from bored piles models. The models output is the pile capacity (interpreted failure load).Additional input variables are included for modelling the load-settlement behaviour of piles. They include settlement, settlement increment and current state of loadsettlement. The output is the next state of load-settlement.The data are randomly divided into two statistically consistent sets, training set for model calibration and an independent validation set for model performance verification.The predictive ability of the developed GEP model is examined via comparing the performance of the model in training and validation sets. Two performance measures are used: the mean and the coefficient of correlation. The performance of the model was also verified through conducting sensitivity analysis which aimed to determine the response of the model to the variations in the values of each input variables providing the other input variables are constant. The accuracy of the GEP model was evaluated further by comparing its performance with number of currently adopted traditional CPT-based methods. For this purpose, several ranking criteria are used and whichever method scores best is given rank 1. The GEP models, for bored and driven piles, have shown good performance in training and validation sets with high coefficient of correlation between measured and predicted values and low mean values. The results of sensitivity analysis have revealed an incremental relationship between each of the input variables and the output, pile capacity. This agrees with what is available in the geotechnical knowledge and experimental data. The results of comparison with CPT-based methods have shown that the GEP models perform well.The GEP technique is also utilized to simulate the load-settlement behaviour of the piles. Several attempts have been carried out using different input settings. The results of the favoured attempt have shown that the GEP have achieved limited success in predicting the load-settlement behaviour of the piles. Alternatively, the ANN is considered and the sequential neural network is used for modelling the load-settlement behaviour of the piles.This type of network can account for the load-settlement interdependency and has the option to feedback internally the predicted output of the current state of loadsettlement to be used as input for the next state of load-settlement.Three ANN models are developed: a model for bored piles and two models for driven piles (a model for steel and a model for concrete piles). The predictive ability of the models is verified by comparing their predictions in training and validation sets with experimental data. Statistical measures including the coefficient of correlation and the mean are used to assess the performance of the ANN models in training and validation sets. The results have revealed that the predicted load-settlement curves by ANN models are in agreement with experimental data for both of training and validation sets. The results also indicate that the ANN models have achieved high coefficient of correlation and low mean values. This indicates that the ANN models can predict the load-settlement of the piles accurately.To examine the performance of the developed ANN models further, the prediction of the models in the validation set are compared with number of load-transfer methods. The comparison is carried out first visually by comparing the load-settlement curve obtained by the ANN models and the load transfer methods with experimental curves. Secondly, is numerically by calculating the coefficient of correlation and the mean absolute percentage error between the experimental data and the compared methods for each case record. The visual comparison has shown that the ANN models are in better agreement with the experimental data than the load transfer methods. The numerical comparison also has shown that the ANN models scored the highest coefficient of correlation and lowest mean absolute percentage error for all compared case records.The developed ANN models are coded into a simple and easily executable computer program.The output of this study is very useful for designers and also for researchers who wish to apply this methodology on other problems in Geotechnical Engineering. Moreover, the result of this study can be considered applicable worldwide because its input data is collected from different regions.