Hybrid Models for Situational Awareness of an Aerial Vehicle from Multimodal Sensing

Abstract

An integrated sensing system is presented for flight state awareness of fixed-wing aircraft. The hardware of the system consists of a lightweight and nonobtrusive sensor network that houses a multitude of piezoelectric and strain transducers in a distributed, large areal coverage setting. The sensor network is embedded onto the surface of a model morphing unmanned aerial vehicle wing with variable trailing edge camber. The data collected by the sensor network form the input of a set of physics and machine learning models that work in tandem to infer several state variables of flight in near real-time. Capabilities of the system are characterized via controlled wind tunnel experiments and presented through a custom interface that provides a side-by-side comparison of the ground truth, as captured by commercial sensors and video cameras, and predictions of the trained models. It is demonstrated that the system identifies safety-critical flight conditions with very high accuracy, paving the way for a novel aircraft instrumentation method that is lighter weight and more capable and reliable in certain aspects.