Design and Simulation of Fault Tolerant Flight Control Schemes Implemented on a Parallel and Distributed Computational Cluster

摘要

In recent years, there has been an increase in the use of Unmanned Aerial Systems (UAS) in the civilian sector for various purposes. As these platforms are constrained in terms of payload and capacity, they are typically equipped with a minimal sensor suite and the use of redundant sensors is uncommon. This research effort describes the design and simulation of a Neural Network (NN) based fault tolerant flight control approach for sensor and actuator failures, implemented on a parallel and distributed computational architecture. The inter process communication is implemented using BSD sockets and Message Passing Interface (MPI). For the purpose of the sensor failure detection, identification and accommodation (SFDIA) task, it is assumed that the pitch, roll and yaw rate gyros onboard the aircraft are without physical redundancy. The SFDIA task is accomplished through the use of a set of four neural networks, named Main Neural Network (MNN) and a set of three De-Centralized Neural Networks (DNNs), providing analytical redundancy for the pitch, roll and yaw gyros. The purpose of the MNN is to detect any failure on the three sensors, while the purpose of the DNNs is to identify the failed sensor and subsequently to facilitate failure accommodation by providing estimates of the sensor measurements. The actuator failure detection, identification and accommodation (AFDIA) scheme also features the MNN, for detection of actuator failures, along with three Neural Network Controllers (NNCs) for providing the compensating control surface deflections to neutralize any failure induced pitching, rolling and yawing moments. All NNs continue to train online, on top of an offline trained baseline network structure, using the Extended Back-Propagation Algorithm (EBPA), with data from a pilot-in-the loop flight simulation. Experiments indicate that the distributed architecture is capable of learning the behavior of the sensors (roll, pitch and yaw gyros) and is able to detect and identify failures on them. Additionally, it has also been shown that the distributed architecture is able to provide compensating control surface deflections to recover from failures on the actuators of the aircraft.