Capturing sequence similarity using local spatiotemporal manifold-regularized dynamic network for semi-supervised industrial quality prediction

Due to the dynamics and label-scarcity of most industrial processes, semi-supervised dynamic quality prediction models have gradually become a research hotspot. For semi-supervised learning, the widely applied manifold regularization ignores the significant dynamics of process variables and the slow-varying properties of quality variables, leading to the failure of its manifold similarity assumption that similar inputs yield similar outputs. Moreover, its computational burden is heavy for applications. To deal with these issues, this study proposes a new local spatiotemporal manifold regularization (LSTMR) method. Specifically, LSTMR designs local spatial similarity and local temporal similarity with full consideration of the dynamics and slow-varying characteristics. The enhanced manifold regularization is obtained through similarity weighting to mine the latent information from unlabeled data. Meanwhile, the computational burden is significantly reduced by omitting unnecessary similarity calculation for sequences pairs. Finally, a dual-attention dynamic learning network (DADLnet) assisted by LSTMR is constructed for quality prediction. The DADLnet’ prediction objective is achieved by applying the prediction error term for labeled data and the LSTMR term for unlabeled data. The applications to an actual alumina digestion process exhibit the superiority of LSTMR-DADLnet.