Multi-Layer DLV for Quality-Relevant Monitoring and Root Cause Diagnosis

Quality-relevant root cause diagnosis is essential for the quality improvement and maintenance of dynamic processes. However, the traditional dynamic latent variable (DLV) modeling methods are mainly unsupervised ones that extract dynamic relations from one dataset (process data only). In this paper, in order to extract latent dynamics between two datasets (process data and quality data), a multi-layer DLV based quality anomaly online monitoring and root cause diagnosis method is proposed. A solution of dynamic inner CCA for modeling two group datasets is provided, then quality-relevant dynamic variations, process residuals, and quality residuals are isolated. The dynamic variations are subsequently decomposed to dynamic and static ones to form a clear decomposition. Based on these decompositions, a multi-layer DLV-based quality-relevant fault monitoring method is proposed. Then, a contribution plot in the MLDLV framework is defined to diagnose the possible qualityrelevant faulty candidates that are used in the subsequent transfer entropy-based root cause diagnosis. Finally, the experimental results on the Tennessee Eastman benchmark demonstrate the effectiveness of the proposed method.