Artificial neural network-based flight control using distributed sensors on fixed-wing unmanned aerial vehicles

Conventional control systems for autonomous aircraft use a small number of precise sensors in combination with classical control laws to maintain flight. The sensing systems encode center of mass motion and generally are set-up for flight regimes where rigid body assumptions and linear flight dynamics models are valid. Gain scheduling is used to overcome some of the limitations from these assumptions, taking advantage of well-tuned controllers over a range of design points. In contrast, flying animals achieve efficient and robust flight control by taking advantage of highly non-linear structural dynamics and aerodynamics. It has been suggested that the distributed arrays of flow and force sensors found in flying animals could be behind their remarkable flight control. Using a wind tunnel aircraft model instrumented with distributed arrays of load and flow sensors, we developed Artificial Neural Network flight control algorithms that use signals from the sensing array as well as the signals available in conventional sensing suites to control angle-of-attack. These controllers were trained to match the response from a conventional controller, achieving a level of performance similar to the conventional controller over a wide range of angle-of-attack and wind speed values. Wind tunnel testing showed that by using an ANN-based controller in combination with signals from a distributed array of pressure and strain sensors on a wing, it was possible to control angle-of-attack. The End-to-End learning approach used here was able to control angle-of-attack by directly learning the mapping between control inputs and system outputs without explicitly estimating or being given the angle-of-attack.