Transfer Learning for Soft Sensors in Process Industries: A Review and Future Perspectives

Abstract Transfer learning (TL) has become an important method for data-driven modeling whereby labeled data in target domains are limited. It has emerged as a promising approach for soft sensor enhancement by effectively leveraging knowledge from related domains. This work presents a systematic review of recent advances in TL for soft sensor modeling in process industries. Cross-domain transferable information is categorized and analyzed from both data- and mechanistic-level perspectives. Industrial applications involving specific forms of transferable information under various target-domain data scenarios are surveyed along with a detailed discussion of strategies to mitigate negative transfer. Furthermore, benchmark data sets for evaluating TL performance are summarized. The review establishes fundamental principles for method selection and scenario-specific adaptation and provides practical guidelines for designing transfer strategies. Finally, key challenges and future research directions for TL-based soft sensor modeling are outlined, including model interpretability, dynamic TL, and data privacy, to guide further research in the process industries.