Abstract
Batch processes in manufacturing industries often adapt new products to meet the changing market demands. Dynamic modeling with limited data for new products may lead to inaccurate results. One solution is to extract useful knowledge from historical production data that can be applied to the new grade. However, past historical data from different product grades are often imbalanced. In this study, a meta-learning subspace identification (meta-SID) scheme is proposed to quickly learn a model of the new grade. It can extract more robust common knowledge from historical batches with imbalanced datasets; then the extracted common parameters can be directly transferred into a subspace identification for a new grade of product. The new SID model can be quickly trained even with a limited number of batch data. A numerical example and a case of a real industrial polyvinyl chloride process demonstrate the effectiveness of the proposed algorithm.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have