Expert Knowledge–Guided Bayesian Belief Networks for Predicting Bridge Pile Capacity

Abstract

Bridge pile capacity is a vital criterion used to assure the durability and stability of a bridge pile foundation. In fact, reliably predicting the pile capacity plays a significant role in supporting data-driven decisions for the design, construction, and quality assurance of bridge piles. While previous studies have examined black-box machine learning (ML) models for bridge pile capacity prediction, little-to-no studies were directed to integrating expert knowledge and large bridge pile data to develop an easy-to-interpret white-box ML model for estimating bridge pile capacity. Therefore, this study proposed a novel white-box expert knowledge–guided Bayesian belief network (BBN) to accurately estimate bridge pile capacity. The proposed BBN was developed based on (1) a comprehensive bridge pile data set of 2,735 data points collected from a large bridge project, (2) expert knowledge obtained from eight bridge and geotechnical experts using the systematic three-round Delphi method, (3) a variety of data preprocessing methods, and (4) parametric Bayesian learning applied to different graphical models. The performance of four different BBN models was assessed and compared based on an unseen testing set to evaluate the generalizability of the proposed BBN model. Model evaluation results indicated that the optimal BBN is a tree-augmented Bayesian network that can estimate the discretized capacity of bridge piles with an accuracy of 90.51%. The proposed BBN model was further validated by testing its generalizability performance on another data from a different location. This study contributed to the body of knowledge by providing a novel, intrinsically interpretable, and robust data-driven expert knowledge–guided model for accurately estimating the bearing capacity of bridge piles. Ultimately, this paper aims to attract more research and practical attention toward developing knowledge-based white-box models for advancing the predictive analytics of bridge pile-related data and decisions.