A Probabilistic Graphical Framework Fusing Data for Model Updating and Decision Support.

Abstract

There is significant uncertainty in assessing the structural health and capabilities of a marine structure during both service life and after sustaining damage. Design-stage marine structural engineering models offer limited information on the as-built structure's health during service life. Despite copious amounts of data provided by structural monitoring techniques, synthesizing these different data types to update the design-stage models remains challenging. A novel decision support graph was created by extending a parametrically encoded Bayesian network data fusion framework to influence diagrams for data to decision. The data to decision framework combines observational and sensor through-life data to update the design-stage models. Once updated, these models provide predictions of future structural health and safety, decision support for inspection timing and extent, and decision support to emergency response teams for survival and mission objective satisfaction strategies.

To demonstrate the effectiveness of the Bayesian network parametrically encoded data fusion, a lognormal probabilistic fatigue initiation model was developed for a series of large stiffened metallic grillages; grillages consist of identical fatigue-critical details typical of vessel and platform structures. Monte-Carlo simulations were used to compare the Bayesian network's prognosis with the synthetic data. Evidence for inference includes data acquired from visual inspection, operating conditions, and an innovative stand-alone mechanical strain sensor, the Strain Amplification Sensor, developed as a part of this work. Results demonstrated that the Bayesian network produces better estimates for fatigue crack initiation through addition of various pieces of evidence. Successful prognosis led to the adaptation of the network to provide inspection guidance, and to aid in decision-making given a damaged marine structure.