LAUSR

Abstract Current strategies for quality-related process monitoring mainly focus on a single quality indicator. For multiple related indicators, traditional algorithms extract the same quality-related features from variable spaces while neglecting the specific features of each indicator. Considering the correlation among these quality indicators, essential information can be captured in common features without being affected by the noise pattern of each indicator. By contrast, specific features are also needed for accuracy prediction. In this work, an end-to-end multiple quality-related model is proposed to monitor indicators jointly on the basis of a multi-task learning framework. Apart from the predictive loss of these quality indicators, this model finds the correlation among the extracted features according to the soft parameter-sharing strategy. After that, quality-related and quality-unrelated statistics are calculated to detect faults. Finally, the proposed method is evaluated by different cases in the Tennessee–Eastman and wind turbine blade icing processes.