A Practical Root Cause Diagnosis Framework for Quality-Related Faults in Manufacturing Processes With Irregular Sampling Measurements

In actual manufacturing processes, because of the difference of sampling intervals of different process and quality variables, most of the collected data have characteristics of multiple sampling rates. How to fully use those irregular sampling data to realize nonlinear dynamic causality modeling for root cause location and propagation path identification of quality-related faults remains an open question. For this end, a practical root cause diagnosis framework is developed for manufacturing processes with irregular sampling measurements, in which a $k$ -nearest mutual information (MI)-based optimal variable division method is proposed to obtain quality-related and quality-unrelated variables. Then, a sampling interval attention network is designed for adaptively adjusting the dynamic temporal relationships among consecutive samples. Subsequently, a gated recurrent neural network (RNN) for improved design of conditional Granger causality (GC) analysis models is designed for diagnosing root causes and identifying propagation pathways of quality-related faults. Finally, the hot rolling process (HRP) is taken as an example for verification, and the accuracy rate of root cause diagnosis reaches 90.91% and 92.31%, which indicates its practicability and feasibility.