Fault detection of FWTPs in coal‐fired power plants using K‐WD‐KPCA in consideration of multiple operation conditions

Abstract

AbstractIn coal‐fired power plants, the feed water treatment processes (FWTPs) supply qualified water to utility boilers. The faults of an FWTP may endanger the whole power plant. The classical PCA‐/KPCA‐based fault detection algorithms are valid mainly under single operation condition. When they are put to a practical FWTP, they cannot deal with the problems as treatment route switching, multiple operation conditions, process fluctuations, and process nonlinearity. In this paper, k‐means, wavelet denoise (WD), and KPCA are integrated together to form a new algorithm as K‐WD‐KPCA. It is expected to deal with the process nonlinearity through KPCA, cope with the multiple operation conditions through k‐means, and relieve the process fluctuations through WD. In the experiments on the data sets collected from a practical FWTP, the multiple operation conditions are classified into three categories; consequently, three KPCA models are trained and correspondingly scheduled for online condition matching. WD is further used to denoise the real‐time T2 and SPE statistics. Results show that the WD part of K‐WD‐KPCA algorithm can indeed lower the false alarm rate without reducing its fault detection performance. Finally, the proposed K‐WD‐KPCA algorithm is coded into a software platform and deployed to a coal‐fired power plant containing 2 × 1000 MW generation units. The effectiveness of the K‐WD‐KPCA algorithm is convinced through field application results.