Abstract: A real time on-line fault detection algorithm is set up and realized,which mainly employs the theory of time series analysis.A key concept for fault detection is a whiteness test on the computing residual vector over the evolution time window.A standard autocorrelation whiteness test is used,where the residuals are assumed to be white,and a 99% confidence interval threshold is adopted in the autocorrelation domain.The off-nominal test then is reduced to a thresholding technique performed on the lag terms of the residuals autocorrelation.Therefrom a fault detection scheme is designed,which includes two steps,the first is to declare an event for any model whenever the first ten lags of its residuals autocorrelation exceeding a 99% confidence interval,the second is to declare a fault condition if at least M individual ARMA models declared an event in N measured parameters ( M≤N ).Verification with a number of fire-test data of real engines shows that the algorithm is very effective and applicable for practical engineering.