Multi-sensor besed framework for gear condition monitoring

Abstract

In recent years, there has been considerable interest in developing efficient machine diagnostics and prognostics tools for quantitative estimation of systems condition and remaining useful life. Often, it is beneficial to combine several measures into a single feature for machine condition monitoring purposes. Selection of appropriate features represents a key step to satisfy machine condition monitoring requirements. Gearboxes represent one of those complex systems where classification of fault stages and types (diagnostics) and remaining useful life prediction (prognostics) remain a challenging task. This thesis focuses on certain aspects of engineering tribology related to gearbox components diagnostics and prognostics based on multi-sensor measurements. A dynamic large-scale mechanical system test-bed has been designed, built, and commissioned. This apparatus is based on the accessory gearbox of the GE J85 turbojet engine, which operates in a number of aircrafts, such as the Canadair CT-114 Tutor. The test rig is equipped with a high speed data acquisition system along with a variety of sensing technologies such as vibration, sound, and acoustic emission transducers in addition to thermocouples, power cells and loading mechanisms. Various attributes are compared for detecting faulty gears and a non-parametric statistical method is used as a quantitative measure of transmission quality. The feature importance level is determined by the significant difference level; and the independent coefficient of the candidate feature is used to compare and rank different time and frequency features. An optimal feature set is then evaluated using the support vector machine classification method by considering a monotonically increasing classification rate. In addition, the selected feature subset has the potential to achieve a better recognition rate than those selected by other heuristic methods such as the mutual information method. This thesis also introduces two metrics which identify the appropriate prognostic feature: load stability ratio and degradation value. The two criteria can be used to compare candidate prognostic features to determine which are most useful for prognosis. An optimization-based method is then used to obtain the optimal feature. The optimized feature can be used with a degradation path modeling to estimate RUL (remaining useful life) for the specific gear system.