Simulation of a Machine Learning Based Controller for a Fixed-Wing UAV with Distributed Sensors

Abstract

Recent research suggests that the information obtained from arrays of sensors distributed on the wing of a fixed-wing small unmanned aerial vehicle (UAV) can provide information not available to conventional sensor suites. These arrays of sensors are capable of sensing the flow around the aircraft and it has been indicated that they could be a potential tool to improve flight control and overall flight performance. However, more work needs to be carried out to fully exploit the potential of these sensors for flight control. This work presents a 3 degreesof-freedom longitudinal flight dynamics and control simulation model of a small fixed-wing UAV. Experimental readings of an array of pressure and strain sensors distributed across the wing were integrated in the model. This study investigated the feasibility of using machine learning to control airspeed of the UAV using the readings from the sensing array, and looked into the sensor layout and its effect on the performance of the controller. It was found that an artificial neural network was able to learn to mimic a conventional airspeed controller using only distributed sensor signals, but showed better performance for controlling changes in airspeed for a constant altitude than holding airspeed during changes in altitude. The neural network could control airspeed using either pressure or strain sensor information, but having both improved robustness to increased levels of turbulence. Results showed that some strain sensors and many pressure sensors signals were not necessary to achieve good controller performance, but that the pressure sensors near the leading edge of the wing were required. Future work will focus on replacing other elements of the flight control system with machine learning elements and investigate the use of reinforcement learning in place of supervised learning.

Reinforcement Learning to Control Lift Coefficient Using Distributed Sensors on a Wind Tunnel Model

“…These tests were performed using a semi-span wing of a WOT 4 Foam-E Mk2+ (Ripmax, Enfield, UK) radio control aircraft, instrumented with an array of 30 pressure and 4 strain sensors as per Fig. 1, which was the same wing model used in [20][21][22]. The wing was mounted on a ply and balsa wood half-fuselage built following the original outline design of the WOT 4 fuselage, with a larger tailplane for increased pitch control authority.…”

“…Within machine learning, supervised learning is a subcategory defined by its use of labelled datasets to train algorithms to classify or predict outcomes accurately. The results in [22] saw that ANNs were capable of both dealing with the size and high-complexity of the distributed sensing system and of adapting to different situations not included in their training set, especially when using the full sensor suite. One of the limitations of this approach was that the controllers aimed to match the performance of the training set instead of learning to optimise or minimise the error between the desired and the actual parameter.…”

“…The exploitation of novel sensors on UAVs requires either the design of new controllers which can directly use the raw information coming from the sensor system (e.g. [21], [22]), or the design of new processing algorithms that translate the sensor signals into parameters that will feed conventional (e.g. [23,24]) or alternative controllers (e.g.…”

“…Measurements captured detachment of the flow and non-linear behaviours such as hysteresis and rate dependent effects. This sensing array was then used for angle of attack control of a 1 degree-of-freedom (DOF) platform in wind tunnel experiments [21], and for airspeed control of a 3 DOF simulation model [22] by means of ANN-based controllers. The prior use of distributed pressure and strain sensing arrays for aerodynamic loads estimation and application on end-to-end control sheds light on the potential use of this sensing array to control aerodynamic coefficients during flight.…”

“…The work in [22] saw some of the benefits and limitations of using a supervised learning approach. Within machine learning, supervised learning is a subcategory defined by its use of labelled datasets to train algorithms to classify or predict outcomes accurately.…”