Soft Sensor Modeling Based on Dual-Stream Multi-scale GRU with Feature Fusion Mechanism

The multi-level coupling of industrial process technologies results in process flows with multi-scale characteristics and dynamics. Moreover, long-term dependencies between processes make it challenging to fully extract temporal features. To address this, a multi-scale GRU-based soft sensing model within a dual-stream framework is proposed. On one hand, a dual-stream information extraction structure is designed to capture both local coupling relationships and global temporal dependencies, forming complementary information flows that enhance the prediction performance. On the other hand, efficient channel attention and temporal attention mechanisms are introduced in the local multi-scale and global temporal perception feature extraction components, respectively.These mechanisms dynamically explore the latent correlations between input and target features, capture key features, and assess the importance of different historical time points for predicting the target time point, thereby selecting critical time point information. Finally, predictive performance of the proposed algorithm is validated through its application to wastewater treatment processe and \({\text {CO}}_2\) absorption column.