Physics-Informed Deep Learning for Pilot Parameter Estimation and Pilot-Induced Oscillation Characterization

This thesis investigates the issue of loss-of-control in flight which is a driving contributor to fatal aviation accidents. The two main contributors tackled in this research are human pilot error and categories of pilot-induced-oscillations. The primary objective is to develop models that can capture the mathematical parameters that relate to human pilot control under the widely used McRuer Mathematical pilot model. The goal is that if the mathematical parameters that relate to human pilot control can be monitored during flight then the pilot’s input to the control system and pilot-induced-oscillations (PIO) can be monitored during flight to avoid loss-of-control. This research employs Physics Informed Neural Networks (PINN) for the purpose of pilot parameter estimation and pilot control estimation. With much of the research utilizing model-based approaches towards this problem that experience sensitivity in the estimation, this method seeks to investigate the gaps in the research with data-driven approaches towards this problem where the larger robustness of the PINN over the model-based approaches with new flight data is seen. This method is employed on simulation data along with flight simulator data operated by a volunteer pilot with great success where the pilot’s input to the control system is modeled accurately. This research also employs proven benchmark machine learning models for the purpose of pilot-induced-oscillation monitoring with the use of the Neal-Smith Criterion. These models help add context to any pilot parameter estimation as a direct metric related to PIO can be generated to determine if the parameter estimation is in the regime of PIO or not. The use of a machine learning model for PIO detection helps to bridge the gap in the research at being able to estimate the PIO metric for any given parameter estimation, as well as combine the concepts of pilot parameter estimation and PIO estimation. It is recommended to research various online learning methods that may show more robustness to different flight profiles as pre-training is not required. There also may be areas where an adaptive controller could be created to intervene with the pilot when the estimated pilot parameters are in the PIO regime. The adaptive controller could help to shift the parameters such that pilot flies with PIO safe parameters. Lastly, there also may be areas in the generative machine learning field to bridge the gap between the simulation and flight simulator environment to obtain even more accurate flight simulator testing results.