Comparative Analysis of Different Classes of On-line State Estimators for Aerodynamics Angles and True Airspeed Sensors for Applications to the Sensor Failure Problem

Throughout aviation history, there have been numerous incidents due to sensor failure that have caused a range of issues from loss of control of the aircraft to crashes resulting in loss of human life. Although there are many hardware-based solutions to this problem, the threat of control hardware failure still exists. This work investigates the efficacy of implementing neural networks (NN) and Kalman filters (KF) to solve the accommodation portion of the sensor failure detection, identification, and accommodation (SFDIA) problem through on-line real-time estimation of specific aircraft dynamic parameters. The implementation of on-line estimation architectures into the aircraft flight control system provides multiple advantages such as cost effectiveness and drastic decrease in weight. The multilayer perceptron (MLP) NN, extended minimal resource allocation (neural) network (EMRAN), extended KF (EKF), and unscented KF (UKF) have been evaluated in this effort for the purpose of providing analytical redundancy (AR) for estimating the parameter of the ‘failed’ sensor in lieu of physical redundancy. Each NN-based and KF-based estimator was compared using preset criteria including estimation accuracy, time to perform, and complexity of the model. The overall results have shown that the NN-based sensor failure accommodation (SFA) schemes outperform the KF-based SFA schemes with no undetected faults nor false alarms and significantly smaller estimation errors. More specifically, the EMRAN-based neural estimator has the best performance of all four schemes followed by the MLP NN, UKF, and EKF, respectively. This research shows the great potential of analytical redundancy-based approaches as opposed to physical or hardware redundancy to improved aviation safety for preventing future crashes due to sensor failures