CPT-SPT Correlation Analysis Based on BP Artificial Neural Network Associated with Partial Least Square Regression

Abstract

Most of the correlations of CPT-SPT have been widely investigated based on a statistical method without considering the multicollinearity among the influenced factors. Therefore, one model combining back propagation neural network (BP ANN) with partial least square regression (PLS), which could consider the multicollinearity influence, is proposed for building CPT-SPT correlation. The sensitivity analysis based on the ANN model and the PLS model is conducted, and the four most sensitive factors are obtained, i.e., cone resistance (q_c), soil behavior type (SBT), friction resistance (fs) and soil behavior index (Ic). Further, these four most sensitive factors plus fine content (Fc%) and effective stress $(\sigma_{o}^{{\prime }} )$ are adopted as the input factors of the combined model (BP ANN associated PLS). And 362 group data are collected from New Doha Port for building the combined model. The result shows the combined model has a correlation index R² of 0.83311 and further demonstrates the combined model is effective. Finally, additional 50 group data are applied to verify the combined model. The result indicates that it can improve the correlation coefficient between the predicted (q_c/pa)/N value and the measured one from 0.6429 to 0.7523 and reduce the relative error from 28% down to 21% compared with the solely PLS model. Given the advantage of the combined model, it can provide a more effective and reliable method for CPT-SPT correlation analysis in practice.