Enhancing multivariate quality monitoring: Integrating LSTM and MLS-SVR with Max-MCUSUM frameworks

This study enhances quality control in sugar production by monitoring solution color, grain size, and moisture content. These critical characteristics often exhibit autocorrelation, which can cause false alarms during statistical monitoring. To address this, we integrated Long Short-Term Memory (LSTM) and Multioutput Least-Squares Support Vector Regression (MLS-SVR) models with Max-MCUSUM control charts. The machine learning models were utilized to minimize Mean Squared Error (MSE) and extract residuals free of autocorrelation before applying Phase I and Phase II process monitoring. Results demonstrate that both LSTM and MLS-SVR successfully mitigated data autocorrelation. In the observed dataset, the LSTM-based Max-MCUSUM chart identified more Phase II out-of-control signals than the MLS-SVR-based chart. While Phase I established initial statistical control, Phase II detected three out-of-control signals primarily driven by variances in solution color and grain size. Despite these minor deviations, multivariate process capability indices (MCp and MCpk > 1) confirm the overall production reliably meets strict quality standards. Ultimately, leveraging LSTM or MLS-SVR residual-based Max-MCUSUM charts provides a highly effective framework for managing continuous production quality and pinpointing areas for process optimization.