A Data-Driven nonparametric control chart for multivariate serially correlated data monitoring in advanced industrial scenarios

In modern industrial manufacturing systems, multi-sensor networks are extensively deployed to monitor equipment operational status and environmental conditions in real time, generating massive multivariate time series data with complex serial correlations. Such data, common in applications such as production line quality control and fault diagnosis, exhibit significant autocorrelation and non-stationary characteristics, making traditional multivariate process monitoring methods inadequate. Most conventional control charts fail to account for the inherent temporal dependencies among variables and struggle with multivariate and unknown underlying distributions, limiting their practical utility in dynamic industrial environments. To resolve this challenge, developing a robust nonparametric multivariate statistical process control (MSPC) scheme, leveraging historical data to its full extent, has become imperative. In the present study, we introduce a machine learning-based nonparametric framework designed specifically to monitor serial correlation data stream without distribution assumption, especially in some complex manufacturing processes. Our approach aims to extract the maximum amount of information from the historical data after the serial correlation has been removed. It then conducts effective monitoring through the multivariate exponentially weighted moving average (MEWMA) procedure combined with a random forest (RF) model. Numerical studies and real-world case show that the proposed method exhibits favorable performance in most scenarios.