A Novel Approach For Moving Window Size Selection Utilizing Recursive Pca

The moving window principal component analysis (MWPCA) technique has been widely applied for dynamic statistical process monitoring of multivariate processes. To deal with abnormal situations with small fault symptoms, in our previous work, a novel incipient fault diagnosis method based on the Kullback-Leibler divergence (KLD) has been developed. Although certain algorithms have been given to select the size of moving window qualitatively, there have been no results concerning the quantitative analysis methods. Therefore, in this paper, a novel approach for moving window size selection is proposed. The optimal window size is determined as one that minimize the sum of squared prediction errors (SSPE) of eigenvalues of the covariance matrix. Additionally, utilizing the mapping between forgetting parameter and window size, the recursive PCA (RPCA) with forgetting parameter is utilized to reduce computation consumption. The feasibility and advantages of the proposed approach is demonstrated utilizing a multivariate system.