Robust temporal–spatial synthesis projection for process monitoring under multi-factor disturbances

Abstract

Process monitoring methods in industrial applications frequently encounter performance degradation challenges stemming from multi-factor disturbances, including mode switching, outlier interference, and dynamic variations. Existing approaches rarely demonstrate sufficient robustness to overcome the comprehensive disturbances generated by these factors. This paper proposes a robust temporal–spatial synthesis projection (RTSSP) strategy to enhance algorithmic robustness by considering both spatial and temporal information. A hybrid neighborhood–kernel similarity (HNKS) is defined by integrating both global distance and local neighborhood information, enabling comprehensive capture of spatial-scale features in multimodal data while leveraging neighborhood topological differences to suppress outlier influence. Additionally, RTSSP explores both the real manifold data structure and temporal information, which captures dynamic changes and learns the synthesis projection from the spatial–temporal dimension to extract the core features with high discriminative properties. Finally, experimental validation through numerical simulations and a two-phase flow process case demonstrates the significant advantages of the proposed method.