Adaptive Canonical Variate Analysis for Machine Fault Detection and Identification

Canonical variable analysis (CVA) has been extensively applied to monitor industrial processes. However, conventional CVA is unable to handle the characteristics of time-varying processes. To address the limitations of the standard CVA for systems with time-varying properties, adaptive canonical variate analysis (ACVA) together with the first order perturbation theory are employed in this paper to detect faults in rotating machines. A bias-corrected exponentially weighted moving average approach is used in combination with the ACVA-based contributions to identify variables that are closely related to faults. The performance of the proposed technique is illustrated with condition monitoring data acquired from an operational centrifugal compressor. The results indicate that the proposed method can effectively detect and diagnose faults in real industrial systems under time-varying operating conditions.