Transfer learning-based multi-fidelity modeling method for multimode process monitoring

Data-driven modeling techniques have been widely applied in industrial systems for process monitoring. However, these models heavily rely on data accuracy and completeness. Challenges emerge when the mode characteristics of the system alter due to equipment deterioration (such as heat exchanger fouling, component wear, catalyst deactivation) or after maintenance activities (like cleaning, repair, replacement, etc.). Data collected from the old mode (before the mode change) no longer accurately reflects the characteristics of the new mode (after the mode change). This presents a significant challenge for multimode process modeling, as the new mode model cannot directly utilize old mode data when there is insufficient training data for the new mode. To address this issue, we propose a novel transfer learning-based multi-fidelity modeling (TL-MFM) method. The key innovation of this method lies in its fusion of limited high-fidelity data from the new mode with sufficient low-fidelity data from the old mode to construct a robust monitoring model. By leveraging a model transfer framework that optimizes the transfer of relevant information across fidelity levels, the TL-MFM method enhances the adaptability of the monitoring model. The effectiveness of the TL-MFM method is validated through a case study on a real-world condenser in a steam turbine generator set.