A novel quality-related dynamic slow feature analysis method for process monitoring

In recent decades, multivariate statistical process monitoring (MSPM) has established itself as a dominant technique in process monitoring and fault detection. However, traditional MSPM methods exhibit suboptimal performance in quality-related dynamic process monitoring, as they lack the capability to elucidate whether faults occurring in process monitoring correlate with quality variables or not. To solve the problem, a novel quality-related dynamic slow feature analysis (QRDSFA) method is proposed, where a new objective function is established based on dynamic partial least squares (DPLS) and SFA. Afterward, a combined method incorporating the Lagrange multipliers and a self-search strategy is introduced to solve the multi-objective optimization problems. Furthermore, a vector autoregression (VAR) model is introduced to detect anomalies in the dynamic process at the current moment. Finally, statistical indicators are developed using latent variables and residuals to track dynamic changes pertinent to quality variables. Comprehensive case studies involving numerical simulations, the Tennessee Eastman process, and electric servo systems demonstrate that the proposed methodology exhibits superior monitoring performance compared to other dynamic approaches.