Robust preliminary-summation-based principal component analysis for non-Gaussian processes with outliers

Preliminary-summation-based PCA (PS-PCA), was recently proposed to handle the non-Gaussian features of industrial processes. However, when PS-PCA is applied to the monitoring of data with outliers, the “summation infection” phenomenon occurs, which makes PS-PCA ineffective. To eliminate the influence of outliers, this paper proposes a novel robust PS-PCA (RPS-PCA) which distinguishes outliers from the faulty data using the consecutive detection results and then removes them. Because RPS-PCA only eliminates the influence of outliers in the normal process data while retaining the outliers in the faulty data, it can use the outliers to improve its fault-detecting ability.