Performance enhancement of a multiple model adaptive estimator

This paper describes various performance improvement techniques for a multiple model adaptive estimator (MMAE) used to detect and identify control surface and sensor failures on an unmanned research flight vehicle. The MMAE uses a bank of Kalman filters that predict the aircraft response to a given input, with each model based on a different failure hypothesis, and then forms the residual difference between the predicted and actual sensor measurements for each filter. The MMAE uses these residuals to determine the probabilities of the failures that are modeled by each of the Kalman filters. Initially the MMAE identified most failures within one second and all within four seconds of onset, but with various performance improvement techniques, the identification time was reduced to less than two seconds. The techniques that will be described are removal of “β dominance” effects, bounding the hypothesis conditional probabilities, retuning the Kalman filters, increasing the scalar penalty for measurement residuals, decreased probability smoothing, and increased residual propagation. The noted performance improvement was mostly due to removing the “β dominance” effects, lower bounding the hypothesis conditional probabilities, increasing the scalar penalty for measurement residuals, and retuning of the Kalman filters