An intelligent non-optimality self-recovery method based on reinforcement learning with small data in big data era

Batch processes have attracted extensive attention as a crucial manufacturing way in modern industries. Although they are well equipped with control devices, batch processes may operate at a non-optimal status because of process disturbances, equipment aging, feedstock variations, etc. As a result, the quality indices or economic benefits may be undesirable using the pre-defined normal operation conditions. And this phenomenon is called non-optimality here. Therefore, it is indispensable to timely remedy the process to its optimal status without accurate models or amounts of data. To solve this problem, this study proposes an intelligent non-optimality self-recovery method based on reinforcement learning. First, the causal variables that lead to the non-optimality are identified by developing a status-degraded Fisher discriminant analysis with consideration of sparsity. Second, on the basis of self-learning mechanism, an intelligent self-recovery method is proposed using the reinforcement learning to automatically adjust the set-points of the causal controlled variables. The self-recovery action is taken iteratively through the Actor-Critic structure with two neural networks. In this way, effective actions are taken to remedy the process to its expected status which only require small data. Finally, the efficacy of the proposed method is illustrated by both numerical case and a typical batch-type manufacturing process, i.e., the injection molding process.