High-dimensional categorical process monitoring: A data mining approach

The advent of industrial big data has provided an unprecedented opportunity to achieve a data-driven monitoring of large-scale complex processes. When a process involves massive categorical variables each evaluated by attribute levels rather than real numbers, which is common in modern manufacturing and service applications, the existing process monitoring methods typically fail in modeling the joint distribution of these categorical variables due to the curse of high dimensionality. To fill this research gap, we propose a novel data mining–based framework—a nonparametric method—for High-Dimensional (HD) categorical process monitoring. Specifically, a series of multiscale frequent patterns are particularly defined and quickly extracted to characterize both the significant individual behaviors and the major collective behaviors of HD categorical variables. Then all these discovered multiscale patterns, serving as informative surrogates of the original HD categorical data, are monitored sequentially from low scale to high scale via a principled and powerful multiple hypotheses testing procedure embedded with an alpha spending function and a false discovery rate approach. The superiority of our proposed method is validated extensively by numerical simulations and real case studies. It is capable of maintaining a desired false alarm rate when the process is normal and becoming very sensitive to many different kinds of process shifts.