UC San Diego

Condition based monitoring (CBM) is a subset of structural health monitoring (SHM) that is focused on monitoring vibration signals generated by rotating machines in situ and processing the data by various techniques designed to extract damage-sensitive features as a means of performing damage (e.g., bearing and gear failure) presence, location, or extent. To date, a wide variety of CBM techniques have been documented in the literature and are well accepted in the CBM community. The literature provides current technical means for extracting damage features and in some cases the trending of features in run-to-failure experiments under constant mechanical parameters such as load and rotational speed, but it lacks any statistical analysis on the effects that varying parameters of the mechanical systems for binary damage states have on detectability. Specifications on data acquisition and choice of algorithm parameters used in extracting the damage-sensitive features remain somewhat vague. This thesis attempts to provide a better global understanding of how variability in the damage detection problem affects the features as a precursor for future work in pattern recognition and optimal detection of damage to rotating machines. This thesis compares the various features under varying conditions and computational parameters in a statistically rigorous way using receiver operating characteristic curves, which compare the probability of detection vs. the probability of false alarms as a means to improve detectability for future use in embedded systems. This thesis also introduces a new damage feature which demonstrates superior detection performance when compared to traditional damage feature for use in detecting worn tooth gear box damage