Model Regularization for High-Mix Control

Past solutions to the multicontext run-to-run control problem are examined, especially in how they attempt to resolve ill-posedness of the model that leads to a lack of observability. The specific models examined in this paper are assumed to contain a combination of terms, each partitioned by a different context. It is shown that none of the past approaches adequately addresses the observability issue. This lack of observability has been known to cause sporadic excursions in practice due to the phenomena of estimation error drift. In this paper, a regularization scheme is presented that is shown to address this issue. A novel aspect of the proposed approach is the use of model augmentation versus model reduction or forcing reference values. The resulting regularized models are amenable to a variety of recursive estimation schemes. Last, simulations and production data are used to validate the behavior of the proposed approach.