A0594
Title: Comparative evaluation of control charts for detecting mean and covariance shifts: $T^2$, $|S|$, and autoencoder
Authors: Sungim Lee - Dankook University (Korea, South) [presenting]
Abstract: Recent advances in information and communication technology have enabled real-time collection of process data, increasing interest in data-driven process monitoring. While univariate control charts commonly monitor both the mean and variance, most multivariate approaches focus solely on mean shifts. However, with the widespread use of multivariate and high-dimensional data, there is a growing need for monitoring methods that can also detect changes in covariance structures. The aim is to propose and compare three control charts: Hotellings $T^2$ chart for detecting mean shifts, the generalized variance($S$) chart for monitoring covariance changes, and an autoencoder (AE)-based chart that leverages neural network representations. To address the large data requirement of AE-based models, a simplified architecture is employed, tailored for moderate-dimensional settings. Extensive simulation studies are conducted under varying covariance structures to evaluate the anomaly detection performance of these methods. In addition, their practical applicability is assessed using real-world data. The results provide insights into the strengths and limitations of each approach under different data conditions.
