Detecção de outliers em dados de alta dimensionalidade em processos químicos

Abstract

Chemical processes are characterized by high operational complexity, simultaneous monitoring of multiple process variables, as well as require high levels of reliability, stability, and rapid response to deviations. In this context, the early identification of outliers is essential to mitigate operational and financial impacts. This study proposes a method grounded in classical approaches for identifying out of control observations in sample groups, employing Principal Component Analysis (PCA) and Biplots to detect outliers in multivariate data from real petrochemical process. The proposed approach aims to address gaps in the existing literature by offering a solution applicable across different organizational hierarchies, including teams without advanced statistical expertise. The method relies on the Euclidean distance in the first plane of a GH-Biplot, which approximates the Mahalanobis distance, thereby providing a more suitable representation for correlated variables. Moreover, the methodology does not depend on distributional assumptions, making it robust and applicable to a wide range of scenarios. However, it required a clear reference of in control process behavior to allow develpoment. The visual nature of the approach further facilitates the detection of outliers and the identification of emerging patterns in high-dimensional data. The proposed method was evaluated using three real-world datasets. The first corresponds to the case study introduced in this work, while the other two are publicly available datasets, which are widely used in the literature, both containing explicitly labeled outliers. The results indicate that the proposed approach was effective in detecting outliers, achieving an accuracy greater than 70% across all three datasets evaluated.