Aiming at the problems that the Adaptive Correlation Algorithm (ACA) uses single theoretical threshold and the Mahalanobis distance correction method is unreasonable, we proprose an improved adaptive correlation algorithm for real-time rocket engine fault detection on the basis of existing research. The improvements are: 1) The empirical threshold obtained from the historical data of the engine is used together with the theoretical threshold as judgment criterion; and 2) Mahalanobis distance is corrected by removing 1 to 3 parameters that contribute the largest deviation to data mean. Under the given false detecting rate of 5%, simulation data demonstrate that the improved algorithm can give timely and accurate fault detection result, and can effectively solve false alarm problem when outliers exist.