Speaker: Danyang Zhang

Title: Online Monitoring of Serially Correlated Multivariate Data in Multi-Process Systems With Machine Learning Control Charts: Application to Disease Detection in a U.S. Animal Production System

Abstract:  Advancements in machine learning (ML) have led to the incorporation of ML-based methods into Statistical Process Control (SPC) charts to enhance monitoring performance. Traditional ML-based SPC charts do not consider dependency between observations over time, such as that in time series data or longitudinal data. Recent approaches employ sequential data decorrelation procedures to address serial correlation prior to building ML-based SPC charts, but mainly focus on a single process, which typically requires extensive in-control (IC) data to accurately estimate IC parameters and develop a suitable model. Motivated by the presence of multiple processes sharing common structures within an animal production system, we propose a new SPC framework that leverages knowledge across multiple processes to enable more effective monitoring. Our proposed method borrows information from other processes within the same system to achieve better process control, while taking care of the correlation in the multivariate responses and the serial correlation over time. Using the distance-based support vector machine (DSVM) control chart as an example of ML-based control charts, our empirical and simulation studies demonstrate that the proposed framework improves monitoring performance in terms of average run length (ARL) compared to traditional MSPC methods and DSVM control charts with individual decorrelation estimates or none. The proposed approach improves the timely detection of process deviations while maintaining robust false alarm control, ultimately improving the reliability of multivariate process monitoring.