Enhancing recursive-PCA-based process monitoring using t-SNE visualization

Abstract. Certain industrial chemical processes are known to change over time but are still regarded as normal; this can lead to false alarms when monitored using a non-adaptive process monitoring technique. Adaptive process monitoring techniques can improve product quality and minimize downtime by reducing false alarms. Since most industrial chemical processes are time-variant, standard principal component analysis (PCA) is not the best method for monitoring a slowly changing process. It is crucial to predetermine the data's characteristics to choose the best process monitoring technique. It has been discovered that t-distributed stochastic neighbour embedding (t-SNE) is a valuable data visualisation method. This paper uses t-SNE to visualise simulated data from a continuously stirred tank reactor (CSTR) process. It has been found that the visualized data using t-SNE is worth monitoring with recursive PCA due to its non-linear data characteristics. Consequently, sample-wise recursive PCA is utilized to identify slow changes caused by variations in ambient temperature during the day and night. Compared to standard PCA, the recursive PCA model performs noticeably better in reducing false alarms due to its effective adaptation to slow changes.