Implementation of statistical process control framework with machine learning on waveform profiles with no gold standard reference

Abstract

Condensation water temperature profiles are collected from a curing process for high-pressure hose products. The shape of those profiles resembles sine waves with diminishing amplitudes. A gold standard wave profile does not exist. Instead some wave profiles with various frequency and amplitudes are deemed normal for the water release operation. To the best of our knowledge, the current practice and research on SPC do not provide a solution for monitoring wave profiles of this kind. We leveraged existing methods, tools, algorithms that can be found in open source or commercial software for quick response to this type of problem. The proposed SPC implementation framework first converts waveform profiles from the time domain to the frequency domain. Then a set of phase I IX control charts is constructed based on a Partition Around Medoids (PAM) clustering method. A Support Vector Machine (SVM) classifier is then used to label a new profile to its associated group for phase II monitoring so that the IX chart associated with a homogeneous group can provide better process monitoring. Overall 146 water temperature profiles were collected in phase I process, while 39 profiles were captured in phase II process. Out of those 39 profiles, 6 of which were recognized as abnormal waveform profiles by quality engineers and our judgements. The proposed framework with machine learning and SPC implementation in the frequency domain works well during phase I control charting with low false alarm rates. The proposed framework also outperforms the other profile analysis methods in phase II control charting in term of high detection rate of abnormal profiles.