基于多尺度 PCA 的工业过程故障预测

The fault prognosis problem for continuous processes with hidden performance degradation and input faults is studied. It is assumed the degradation process develops slowly and the input fault occurs suddenly and varies rapidly. Based on the multi-scale principal component analysis model, a fault prognosis method is proposed for the performance degradation process. We first apply discrete wavelet decomposition to a segment of historical data under normal operation condition, and then perform the principal component analysis on the wavelet coefficients for each scale. After multilayer wavelet decomposition, the degradation can be detected by the low frequency coefficients model. Then the degraded extent is estimated by a reconstruction-based method and predicted by an exponential smoothing approach. At last, the remaining useful life is predicted. A case study on continual stir tank reactor (CSTR) shows the efficiency of the proposed approach.