Neural network based sensor validation for reusable rocket engines

This paper describes a new method of sensor failure detection, isolation, and accommodation using a neural network approach. The proposed technique utilizes dimensionality reduction to develop the neural network for sensor failure detection and data recovery. In this work, the sensor validation scheme is applied in closed loop for detection and recovery of sensor faults in a simulation of the Space Shuttle main engine (SSME) control. The performance of the sensor validation scheme is evaluated based on its ability to detect sensor faults and provide estimates of the actual signals for the SSME engine controller after the sensor faults have been detected. In the simulation, a group of eleven sensors with known analytical redundancy on the fuel side of the SSME is selected for the sensor validation algorithm. Four sensor outputs in this group are used by the controller to regulate the key performance parameters such as thrust and mixture ratio of the engine. The generation of the training data and the training techniques are discussed. The simulation results show that the proposed sensor validation scheme can adequately compensate up to three sensor failures without major deterioration in performance. The effect of the delay time in detecting sensor failures and the transition from the failed sensor values to the estimated values for the controller are discussed. Abilities and constraints of the neural network-based sensor validation scheme are also discussed.