Fault modeling and prognosis based on combined relative analysis and autoregressive modeling

The conventional fault detection in general focuses on reactively detecting the significant changes and failure of the plant sate as indicated by confidence limit violation, which, however, is not indicative of a developing fault. In the present work, a fault modeling and prognostic strategy is developed for slowly time-varying autocorrelated fault processes. Several important issues are addressed, including how to evaluate the changes of process variations from normal to fault, how to quantify their influences on monitoring performance and how soon they will violate a confidence limit in the future. First, a combined relative analysis algorithm is proposed via reconstruction technique in the context of principal component analysis (PCA) based monitoring system to decompose the underlying fault effects. Then, based on the estimated fault directions, the associated fault magnitudes are calculated to estimate the fault effects along these directions from which a new monitoring index is defined to quantify the fault effects. An autoregressive (AR) model is then developed based on this new index for online fault prognosis. Its feasibility and performance are illustrated with both numerical and experimental data.