Quality Control of Batch Processes Using Natural Gradient Based Model-Free Optimization

Model-based optimization (MBO) has been widely applied for quality control of batch processes; however, it is not easy to obtain a globally effective and accurate quality model with affordable effort. Instead of building a quality model, model-free optimization (MFO) uses process data directly, which is more efficient and economic for quality control of batch processes. Considering the complex nonlinearity and dynamics in batch processes, a quality control scheme using natural gradient based optimization is proposed in this paper. Optimization algorithm is developed from the aspect of manifold in non-Euclidean space. An approximation method is derived for the calculation of the natural gradient, and a multivariate iterative sensitivity matrix based on Riemannian geodesic distance is proposed to obtain a novel adaptive stepping strategy. The proposed quality control scheme has been verified in the injection molding process. A set of comparison tests are presented to demonstrate the feasibility and effectiveness of the proposed method.