Performance evaluation of neural network based approaches for airspeed Sensor Failure Accommodation on a small UAV

Traditional approaches to sensor fault tolerance for flight control systems have been based on triple or quadruple physical redundancy. However, recent events have highlighted the criticality of "common mode" failures on the Air Data System (ADS). In fact, since the parameters of flight control laws are typically scheduled as a function of airspeed, incorrect readings from the ADS can lead to potentially catastrophic conditions. In this paper, we describe the evaluation of an analytical redundancy-based approach to the problem of Sensor Failure Accommodation following simulated failures on the ADS of a research UAV, using Artificial Neural Networks (ANNs). Specifically, two different neural networks are evaluated - the Extended Minimal Resource Allocating Network and a Multilayer Feedforward NN. These neural networks are trained and validated using experimental flight data from the WVU YF-22 research aircraft which was designed, manufactured, instrumented, and flight tested by researchers at the Flight Control Systems Laboratory at West Virginia University. The performance of the two approaches is evaluated in terms of the statistics of the tracking error in the estimation of the airspeed, as compared to actual measurements from the ADS, operating under nominal conditions.