Open Multimode Process Monitoring: A Dual Memory-Based Continual Dictionary Learning Method

The traditional process monitoring method follows a static learning paradigm based on closed-set assumptions. However, with new modes constantly emerging, static models are limited to generalizing known modes. If new mode data is used to update the model, it can result in catastrophic forgetting. Existing solutions require strong similarity between modes, severely limiting industrial applications. In response to these issues, this article proposes a dynamic dictionary learning paradigm for open multimode process monitoring. First, within a probabilistic framework, the general form of the regularization constraint is derived as a standard quadratic form based on parameter weights, which is applicable to most monitoring frameworks. The optimization sensitivity weight is defined, and an analytical expression is derived to accurately memorize the important parameters of multiple modes. Then, for low similarity of modes, a mode information diversity perception-based sampling method is proposed. Based on this, a few-shot replay mechanism is introduced to memorize the direct features of historical modes, thereby reinforcing representational ability. The complementary dual memory of atoms and samples overcomes the mode similarity constraint. Finally, A benchmark evaluation protocol for continual monitoring of the open multimode process is proposed for the first time. A numerical simulation system and an industrial wheel production system verified the superiority.