Discovering key factors and causalities impacting bridge pile resistance using Ensemble Bayesian networks: A bridge infrastructure asset management system

Abstract

Bridges are one of the critical infrastructure systems and play a critical role in supporting the economic development of nations. During the planning, design, and construction phases of bridges, it is important to ensure that the different bridge components satisfy the desired performance for avoiding catastrophic failure. One example is the foundation system with deep driven piles. A crucial consideration when planning, designing, and constructing bridge foundation system is to accurately estimate the pile resistance (i.e., capacity). However, this process constitutes a considerable challenge for bridge managers and engineers due to the uncertainties, subjectivities, and biases involved in providing accurate estimates. In fact, the current design process of bridge pile foundations is still performed in a relatively simple and deterministic manner. Thus, there is still a lack of studies that were conducted to effectively represent the causalities between different variables impacting the bridge pile resistance using objective, data-driven methods rather than subjective, expert-based methods. Therefore, this paper addresses this knowledge gap by proposing a data-driven, knowledge-discovery Bayesian network model as an expert system to (1) identify the key factors impacting the performance of bridge piles through data-driven knowledge discovery; (2) capture the causal (direct and indirect) relationships between the different variables that influence bridge pile resistance, and (3) predict bridge pile resistance in support of decision making in bridge design and construction. Multiple Bayesian network models were developed using two distinct learning paradigms based on a pile dataset with 18 variables and 2,375 data samples from a large bridge project. One paradigm is to run a regular process for learning graphical models using Bayesian structure learning, while the other one is to couple Bayesian networks with bagging ensemble learning. Model validation results showed that the best model learned by bagging ensemble learning paradigm achieves the highest prediction accuracy of 89.78% for predicting bridge pile resistance. Additionally, the developed Bayesian network models identify 8 critical factors affecting bridge pile resistance. These factors are categorized into four groups, including the external applied force on pile head, pile mechanical property (i.e., compressive stresses in bridge piles), hammer property and operation characteristics (e.g., hammer blow count), and pile cap elevation. The observations from this study would assist bridge managers and practitioners in decreasing the uncertainties that impact the estimation of pile capacity, and thus help in balancing cost-effectiveness with good performance. The paper adds to the body of knowledge by providing a great deal of insights related to the factors and causalities influencing the performance of deep bridge foundations to ultimately help in the plan, design, and construction phases of bridges.