Data-driven dynamic inferential sensors based on causality analysis

Considering the stringent requirements for product quality of complex industrial processes, the purpose of this study is to apply causality analysis to select causal features of quality-relevant variables; and then to improve the prediction performance and interpretability of inferential sensors. Based on the idea that low -dimensional causal features can approximate the underlying information of the process instead of the original high-dimensional measurements, feature causality analysis is proposed in this work. To describe dynamic information and extract efficient latent features, dynamic latent variable models are utilized to combine with feature causality analysis. After dynamic latent causal feature extraction, two kinds of inferential sensors are developed with extracted dynamic latent causal features. Several comparison studies have been implemented on the Tennessee Eastman benchmark process; the results show that the inferential sensors based on dynamic latent causal features obtain the best performance.