Detection of incipient tooth defect in helical gears using multivariate statistics

Multivariate statistical techniques have been successfully used for monitoring process plants and their associated instrumentation. These techniques effectively detect disturbances related to individual measurement sources and consequently provide diagnostic information about the process input. This paper investigates and explores the use of multivariate statistical techniques in a two-stage industrial helical gearbox, to detect localised faults by using vibration signals. The vibration signals, obtained from a number of sensors, are synchronously averaged and then the multivariate statistics, based on principal components analysis, is employed to form a normal (reference) condition model. Fault conditions, which are deviations from a reference model, are detected by monitoring Q- and T²-statistics. Normal operating regions or confidence bounds, based on kernel density estimation (KDE) is introduced to capture the faulty conditions in the gearbox. It is found that Q- and T²-statistics based on PCA can detect incipient local faults at an early stage. The confidence regions, based on KDE can also reveal the growing faults in the gearbox.