Development and Experimental Validation of Dynamic Bayesian Networks for System Reliability Prediction

Abstract

Vessels and marine structures are subjected to degradation during their service, jeopardizing structural safety and shortening their service life. Numerical models of such structural systems are developed and relied on to simulate and ensure system integrity. Such numerical models are the essential part of digital twins representing complex marine structures and providing enhanced forecasts of risk and lifecycle performance. Digital twins also require data fusion from observations or experiments to improve the numerical model agreement with the real-world structure. Due to the infeasiblity of full-scale testing of marine structures, scale experiments are developed but few of them reflect many of the properties of large and complex marine structures. Thus, an experiment must be designed to mimic the multiple degradation process and retain structural redundancy so that a single element failure will not remove all load carrying capacity. Dynamic Bayesian networks (DBN) expand the Ordinary Bayesian networks (BN) with slices representing the state of the system at different time intervals. DBN can model the degradation process of structure but its performance has not been validated by experiments. Therefore, the PhD research designs an experiment to mimic the properties of marine structure and develops a corresponding numerical model based on DBN whose performance is evaluated by the designed experiment.

To mimic the interdependence, redundancy and component-to-system level performance of marine structures in degradation, a hexagon tension specimen with four propagating fatigue cracks, one on each corner, is designed and tested. The applied loading cycles and corresponding crack lengths are recorded as the major time-varying data of degradation state. Two new methods of measuring crack length are developed based on computer vision and digital image correlation. A standard eccentrically-loaded single edge crack tension specimen is designed and tested to validate the performance of the developed computer vision-based method for measuring crack length. The results of the hexagon experiment demonstrate that the designed specimen successfully simulates the interaction among cracks and structural redundancy. To complement the test specimen, a DBN is constructed to predict the crack length with input observations. The network models the time-varying process of degradation with sequential slices. The task is divided into several steps including the first two steps as modeling single crack propagation with simulated observations, two cracks propagation considering dependence evaluated via simulated observations. The dependence among components are controlled by hyperparameters and are integrated into complex system behavior to reflect the structure from the component level to the system level. Then a DBN model is developed for four cracks propagation with dependence modeled by hyperparameters using Object-oriented Networks (OON) technology and evaluated by data gathered from the hexagon experiment. Finally, the dependence between crack length and stress is modeled in the fourth model based on the technology named Temporal Clone which is also evaluated via experimental data. The experimental data and developed numerical models provide support and guidance in the exploration of digital twin models.