Novel Meta Mode-Adaptive Multihead Attention for Multimode Industrial Process Soft Sensing.

Abstract

Soft sensing technologies play a crucial role in industrial production. However, industrial process data typically exhibit complex multimode characteristics, which are often ignored by conventional soft sensing approaches that treat the data as originating from a single mode. To address this limitation, a novel meta mode-adaptive multihead attention (M-MAMHA) method is proposed for multimode soft sensing tasks. First, a MAMHA mechanism is employed to capture dynamic features from each mode while modeling intermode dependencies. The inclusion of adaptive weighting allows the model to assess the importance of each mode automatically, improving flexibility and feature fusion effectiveness. Next, a meta for multimode soft sensor (Meta4MSS) framework based on the Reptile meta-learning algorithm is designed, incorporating a parameter-specific adaptive learning rate. This supports the model in quickly adapting to distributional shifts and diverse modes across tasks, significantly enhancing generalization performance. The proposed method combines the strengths of attention and meta-learning, offering a flexible and robust solution for handling complex multimode industrial data. Experimental validation on two real-world industrial process datasets demonstrates that M-MAMHA achieves superior prediction accuracy and improved robustness compared to state-of-the-art soft sensing models. These results confirm the empirical effectiveness and broad usability of our approach in multimode industrial environments.