An efficient method of parameter prediction is needed for sensor validation of Space Shuttle main-engine (SSME) parameters during real-time safety monitoring and post-test analysis. Feedforward neural networks (FFNN) have been used to model the highly nonlinear and dynamic SSME parameters during startup. Due to several problems associated with the use of feedforward networks, radial-basis-function neural networks (RBFNN) were investigated in modeling SSME parameters. In this paper, RBFNNs are used to predict the high-pressure oxidizer turbine discharge temperature, a redlined parameter, during the startup transient. Data from SSME ground test firings were used to train and validate the RBFNNs. The performance of the RBFNN model is compared with that of a FFNN model, trained with the Quickprop learning algorithm. In comparison with the FFNN model, the RBFNN-based model was found to be more robust against variations in architecture and network parameters, and was faster to train. In addition, the performance of the RBFNN model during nominal operation and during simulated input sensor failures was found to be robust in the presence of small deviations in the input.
