Modeling of Soil Swelling via Regression and Neural Network Approches

摘要

Damage due to soil swelling is very noticeable in wide spectrum of structuressuch as roads building, canal linings, landfill liners, etc. In order to control or overcome such damage, swelling soil are commonly stablized either mechanically or chemically. To evaluate severity of swelling and to design for the best and most economical stabilization strategy, an accurate assessment of the swell potential is required. This report uses reasonable-sized database representing 413 soils retrieved from 45 different projects covering 28 counties in Kansas to develop prediction models. Neural network-based models and various statistical models were developed and compared for their prediction. Additionally, the reliability of model predictions were examined using an additional 101 data sets. In the second phase, predictions obtained using the developed neural network models along with the experimental database were used to produce a reliability (probability of success) factor matrix. This matrix is used to assign a specific confidence level to predictions obtained from the developed neural network models in order to classify the soil under consideration as of swelling or non-swelling type. Results obtained from this study showed that neural network-based swelling potential prediction models provide significant improvements in prediction accuracy over statistical counterparts.