Probabilistic digital twin for decision-making under uncertainty: Sensing, diagnosis, prognosis and optimization

Abstract

The probabilistic digital twin approach is explored in this dissertation as a tool for reliable decision making under uncertainty for rotorcraft maneuvering. Four objectives are pursued: (1) optimization of information gain for the purpose of placing sensors on a component of interest; (2) helicopter flight parameter optimization for mission planning in the presence of uncertain damage to critical components; (3) rotorcraft controls decision-making under uncertainty for time-varying control; and (4) course correction for rotorcraft with damage-affected dynamics. The primary contribution of this research is the investigation of the probabilistic digital twin methodology and its use for stochastic optimization in applications for improving reliability in operation. Digital twins are virtual models of physical systems, calibrated over time with sensor measurements to reflect changes in the real-world asset. First, the problem of optimal sensor placement is addressed for the construction of more accurate predictive models. The methodology employs a two-step optimization approach, first to optimize the input conditions of where to obtain high-fidelity simulations and then to optimize the sensor placement on the physical asset to maximize information gain for multi-fidelity models. The methodology is demonstrated for placing pressure sensors on an experimental panel in Mach 4 flow. With the sensor placement optimized, the next objective focuses on mission planning for a helicopter in maintenance-free operation under damage. A Bayesian calibration methodology is used to update the knowledge of the damage and propagate the uncertainty to the model prediction. With an up-to-date damage estimate, flight parameters (velocities) of an upcoming mission are optimized; improved reliability is observed over a non-digital twin approach. The next objective focuses on stress-aware, time-dependent maneuver control in the presence of damage. The optimal maneuver minimizes the stress experienced by the critical component. The last objective extends the probabilistic digital twin to address cases where the damage affected the dynamics of the rotorcraft. A Bayesian neural network is deployed to quantify the model prediction uncertainty, and this uncertainty is propagated along with the uncertainty in the damage diagnosis to obtain a probabilistic prognosis of the flight dynamics. The probabilistic dynamics prediction is incorporated in a stochastic optimization that corrects the course to the original trajectory after veering off course due to damage.