UMix-Mamba: A Dual-Pathway Model for Accurate Prediction of Key Parameters in Erythromycin Production

In bio-intelligent manufacturing, accurate forecasting of critical metabolic parameters in erythromycin production remains underexplored. Traditional methods struggle with long-term, multivariate time-series prediction, and deep learning techniques are seldom applied in this domain. To overcome these challenges, we propose a novel model, UMix-Mamba, which enhances the U-Net framework by replacing its convolutional modules with the more parameter-efficient Mamba module. To address non-stationarity in the industrial process, we introduce an innovative differential decomposition approach and apply a channel convolution module to minimize the impact of redundant information. Our model significantly improves prediction accuracy, providing a more efficient and reliable tool for optimizing erythromycin production processes.