Bayesian network based machine learning for design of pile foundations

Abstract

Realistic estimation of the uncertainty associated with the bearing capacity, which is often represented by the uncertainty of a model bias factor, is important to reliability-based design of pile foundations. Due to the existence of cross-site variability, the statistics of a model bias factor may vary from one site to another. Also, as the number of site-specific load test data is often very limited, it is difficult to obtain the site-specific statistics of the model bias factor. This paper aims to establish a Bayesian network based machine learning method to develop site-specific statistics of the model bias factor utilizing information from both the regional and site-specific load test data, through which the resistance factor for design of the pile foundation can be calibrated. The suggested method has been verified using a comprehensive load test database for design of driven piles in Shanghai, China. It is found that a few site-specific pile load test data can significantly reduce the uncertainty associated with the model bias factor and hence increase the cost-effectiveness of the pile design. The method suggested in this paper lays a sound foundation for site-specific reliability-based design of pile foundations, and provides useful insight into the planning of site-specific load tests for design of pile foundations.