Fault detection strategy based on difference of score reconstruction associated with principal component analysis

The statistical process control based on principal component analysis (PCA) usually assumes that the underlying data generation process is independent and identically distributed (I.I.D.). When PCA is applied to detect faults in a process with multimodal structure or nonlinear monitored variables, its fault detection performance will descend. Aiming at the above limitations of PCA, a fault detection strategy based on difference of score reconstruction associated with PCA (Diff-PCA) is proposed in this paper. First, an input space is decomposed into two subspaces: principal component subspace (PCS) and residual subspace (RS) using PCA. Next, the reconstructed score vectors of each score vector are computed respectively through k nearest neighbors (kNN) rule in PCS and RS, and then a difference vector of score reconstruction can be also obtained. At last, the statistic values of the difference vectors are monitored to detect faults. Diff-PCA is capable of not only reducing the influence of multimodal and nonlinear characteristics, but also eliminating the autocorrelation of the statistic and improving the fault detection rate (FDR). The efficiency of the proposed strategy is implemented in two simulated cases and in the Tennessee Eastman (TE) processes. The experimental results indicate that the proposed method outperforms the conventional PCA, Kernel PCA( KPCA), Dynamic PCA (DPCA) and the fault detection method based on k nearest neighbors (FD-kNN). © 2019, Editorial Department of Control Theory & Applications South China University of Technology. All right reserved.