SENSOR FAULT TOLERANT CONTROL SYSTEM USING ONLINE MULTI-LAYER FEED-FORWARD NEURAL NETWORK

Sensor Fault Detection Identification and Accommodation (SFDIA) is an important part of safety critical systems used in aircraft. SFDIA can be achieved either by hardware redundancy or analytical redundancy technique. The advantages like reduced complexity, cost and weight of analytical redundancy over hardware redundancy encourages the designers to follow the former technique. Analytical redundancy techniques could use either model based or non-model based approaches. Model based techniques include observer based residual generation, parity based and parameter based approaches [1]. Fuzzy decision-making and artificial neural networks are used for building analytical redundancy in non-model based approaches. Due to the learning and adaptation capability of Neural Network (NN) [2-4], applicability to nonlinear and multivariable systems, parallel distributed processing and hardware implementation, Artificial NNs are very appealing for the purpose of providing fault tolerance capabilities in a flight control system following sensor failures.\ud In this paper, the SFDIA is achieved by using a Main Neural Network (MNN) and n Decentralized Neural Networks (DNNs) for a system with n sensors. Here MNN is used to detect the fault and DNN is used for identifying the fault. The reconfiguration of faulty sensor can be achieved by feeding back the DNN estimate for the faulty sensor instead of sensor measurement to the flight control system. The SFDIA scheme is realized using MATLAB/SIMULINK® for closed loop decoupled linearised models (Refer Appendix A for the longitudinal and lateral motion models) of Aerosonde UAV [5-6] having pitch and roll angle autopilots with rate feedback [7]. The SFDIA algorithm is evaluated for pitch and roll rate sensor faults of constant bias type