STATISTICAL MONITORING FOR ENGINEERING SYSTEMS WITH COMPLEX DATA

Abstract

Modern industrial processes usually involve multiple related variables and demand effective statistical process control (SPC) schemes for abnormal detection. The challenges involved in current SPC schemes arise from several aspects: i) The variables’ distributions are usually unknown and very complex; ii) The variables’ dimension can be very large and even higher than data size; iii) As sensor technologies advance, for every variable, we can measure its signals during the continuous process operation, which are profile or functional signals; iv) The abnormal pattern is usually various and unpredictable, or even sparse. To address these challenges, (i) this thesis proposes to use nonparametric methods to design distribution-free monitoring schemes which have robust detection power for general abnormal patterns; (ii) This thesis proposes to use functional data analysis and machine learning methods for profile feature extraction, and then to construct monitoring schemes to detect sparse anomaly based on the features.