Adaptive State Filtering for Space Shuttle Main Engine Turbine Health Monitoring

Abstract

Real-time estimation of system states or parameters that are dife cult or expensive to measure directly is often needed for adaptive control or health monitoring purposes. A practical algorithm is proposed for adaptive state e ltering in nonlineardynamic systemswhen the state equations areunknown or too complex to model analytically. The state equationsare constructively approximated by using recurrent neural networks. The proposed algorithm is based on the predictor-update approach of the Kalman e lter, but a least-mean-square e lter implementation with an adaptive e lter gain is used. Furthermore, unlike the Kalman e lter and its nonlinear extensions, the proposed algorithm makes minimal assumptions regarding the underlying nonlinear system dynamics and their noise statistics. The e lter is used to estimate the high-pressure turbine discharge temperature of the space shuttle main engine, during setpoint changes and turbopump failures. The e lter is developed by using simulated engine data, and its performance is tested on both simulated and actual recorded space shuttle main engine transients. When the complexity of the problem studied is considered, the resulting e lter accuracy is shown to be quite acceptable. Further use of the adaptive e lter gain developed is to enable real-time detection of certain system failures, such as the turbopump failures of the space shuttle main engine. This concept is demonstrated also by using both simulated and experimental failure data.