A new two‐stage feature‐enhanced hierarchical network framework for multimode quality‐related process monitoring

Abstract

Multimode quality‐related process monitoring plays a key role in ensuring the economic benefits of enterprises. The lack of data mode labels, poor model interpretability, and weak robustness of quality‐related features have brought challenges to process monitoring. This paper proposes a two‐stage feature‐enhanced hierarchical network framework (TFHN). First, random masking and random noise are applied to generate augmented sample pairs. The self‐supervised clustering network (SSCN) is constructed to extract mode features and perform clustering. Then, a prior‐knowledge‐guided quality‐related feature enhancement network (PKG‐QFEN) is constructed for each cluster. Combining mutual information and mechanism knowledge to build a quality‐related feature enhancement module (QFEM). One‐dimensional decoupled convolutional kernels extract variable‐level features, which are then fed into the QFEM for feature interaction and quality‐related feature enhancement. Finally, Kullback–Leibler divergence is employed to constrain the feature distributions and construct the monitoring statistic. Our analysis shows that, in the lack of mode labels, self‐supervised signals constructed via data augmentation can extract representative features of each mode under feature‐consistency constraint. The fusion of prior knowledge and data can improve the interpretability of model and the generalization of quality‐related features. TFHN shows good potential in multimode quality‐related process monitoring.