Coasting of the Fueling Trolley System: Dynamic Modeling and Uncertainty Assessment

Abstract

Structural health monitoring in industrial systems is challenging due to complex component interactions, limited accessibility, and sparse sensing. In such settings, condition monitoring methods must remain reliable under limited information and significant model-form uncertainty. Bayesian state-space modeling addresses this by representing system interactions stochastically and explicitly quantifying measurement uncertainty. In particular, Kalman filters provide real--time state updates from sensor data while accounting for both process and measurement noise, yielding a practical framework for uncertainty quantification and robust inference.

This thesis develops a physics--based dynamic model of a Fueling Machine Transport trolley to predict the behavior for condition monitoring purposes. Using the real life coasting data, the unknown model parameters are estimated using the Kalman filter method. Using Kalman filtering with data--driven noise identification, we compare two force estimation formulations: a simplified 1-DOF model that identifies a net resistive force and a 2-DOF model that separates two coupled force components. Numerical studies show that the 1-DOF formulation is more accurate and robust under the single sensor setting, whereas the 2-DOF model exhibits multiple convergences and strong prior sensitivity due to structural unobservability.

The coasting data is collected in the form of Gray code signal via the encoder installed on the trolley. Experimental work decodes Gray-code encoder signals and aligns them with event logs to produce position/coasting trajectories that agree with plant records. The single degree of freedom approach is used to find the resistive force in the system. These findings support deploying the simplified estimator with routine consistency checks and route level trend tracking. Practical recommendations include monitoring Q_F as a health indicator, periodic re-estimation of Q and R, and optional secondary sensing to improve identifiability. Overall, the thesis demonstrates that carefully chosen models, paired with principled uncertainty quantification, can deliver robust, interpretable estimates for SHM under sparse sensing.