While there is consensus among stakeholders (FHWA, State DOT’s, Academia, Industry) that better data on bridge performance is required, there is limited infrastructure in place to make the best possible use of the data. This is a considerable threat to the success and existence of the shift towards data-driven approaches. Data without a plan for extracting value quickly becomes a burden. One technique commonly used for deterioration modeling is to use non physics-based models to predict performance based on existing data. Each empirical model form (e.g., Markov chain, Artificial Neural Networks, etc.) comes with a particular set of strengths, weaknesses, biases and assumptions. Unfortunately, there is a lack of consensus on the best applications of model forms, or even a framework to select and use models. Beyond that, the complexity of influences on bridge performance make it very difficult to identify causal relationships between observable behavior and driving factors.

The primary goal of this research is to establish a robust, flexible framework for integrating quantitative data collected from operating structures to provide reliable performance assessments and forecast remaining service life (i.e., descriptive relationships) for structures. This research will address the problem of model selection for bridge performance data through a multiple model framework that accounts for various model approaches, as opposed to excluding them. Over time, the proposed framework may be a viable approach for identifying causal relationships of bridge attributes and inputs to bridge performance.

The proposed research will achieve the stated goal through investigating varying parameter sets within a given model form, as well as multiple model forms, and looking at the ability of the model(s) to predict bridge performance. This will help to address the gap between what is observable (i.e., condition, nondestructive evaluation, member actions, global movements, etc) and what is desired (i.e., capacity, remaining service life, etc). Each model or set of models will be updated based on the observations using a Bayesian model updating approach. For multiple model forms, a set of mathematical rules must be implemented to move between model forms. The process will implicitly weigh model forms and parameter sets that are better predictors of the observed responses. From this population of multiple models, unobservable characteristics (like estimates of remaining service life) can be predicted in a probabilistic sense.