AbstractIn this work we present a summary of the review on data-driven soft sensors published in [1] together with a proposal of how to deal with the identified issues and challenges. We discuss the most common approaches for the development of soft sensors followed by a critical analysis of the main issues in the current soft sensor development. Currently, these are the time which has to be spent on the model development including data pre-processing and model building together with the effort which needs to be spent on periodical performance assessment and re-training of the model. Based on the identified problems we propose a solution based on a model development architecture which can accommodate different data pre-processing techniques, predictive modelling methods as well as approaches for model adaptation. The architecture is based on a structure which unifies several concepts from machine learning such as ensemble methods, local learning, meta learning and concept drift handling. Using the above mechanisms it provides means for automated data pre-processing, model validation, selection and adaptation which can be used to significantly simplify the soft sensor building and maintenance process.