Deep learning soft sensor technology plays an important role in industrial process monitoring, but data distribution shifts caused by concept drift severely affect the predictive performance and generalization ability of models. Most existing methods adopt global static one-time modeling, which struggles to track dynamic changes in online operating conditions, or use online local modeling that is difficult to fit high-dimensional and strongly coupled complex industrial data. To address this issue, this paper proposes an online adaptive updating method for deep learning soft sensors based on Mutual Information Optimized Just-in-Time Fine-Tuning (MI-JITFT). First, a feature selection and weighting strategy based on mutual information optimization is designed to eliminate redundant noise and enhance the effectiveness of similar data retrieval by quantifying the importance of auxiliary variables. Then, just-in-time learning is combined with deep learning to dynamically fine-tune the deep learning soft sensor model by retrieving similar historical samples in real-time, thereby overcoming the degradation of predictive performance caused by concept drift. Experimental results on the penicillin fermentation process dataset demonstrate that the proposed method can effectively improve the working condition adaptive capability of various deep learning models.

Soft sensor, Concept drift, Online update, Just-in-time learning, Mutual information