A Systematic Methodology for Gearbox Health Assessment and Fault Classification

• Main Authors: Jay Lee, David Siegel, Hassan Al-Atat
• Format: Article
• Language: English
• Published: The Prognostics and Health Management Society 2011-01-01
• Series: International Journal of Prognostics and Health Management
• Subjects: Data-driven methods for fault detection; diagnosis; prognosis
• Online Access: http://www.phmsociety.org/sites/all/modules/pubdlcnt/pubdlcnt.php?file=http://www.phmsociety.org/sites/phmsociety.org/files/phm_submission/2009/ijPHM_11_002.pdf&nid=188

A systematic methodology for gearbox health assessment and fault classification is developed and evaluated for 560 data sets of gearbox vibration data provided by the Prognostics and Health Management Society for the 2009 data challenge competition. A comprehensive set of signal processing and feature extraction methods are used to extract over 200 features, including features extracted from the raw time signal, time synchronous signal, wavelet decomposition signal, frequency domain spectrum, envelope spectrum, among others. A regime segmentation approach using the tachometer signal, a spectrum similarity metric, and gear mesh frequency peak information are used to segment the data by gear type, input shaft speed, and braking torque load. A health assessment method that finds the minimum feature vector sum in each regime is used to classify and find the 80 baseline healthy data sets. A fault diagnosis method based on a distance calculation from normal along with specific features correlated to different fault signatures is used to diagnosis specific faults. The fault diagnosis method is evaluated for the diagnosis of a gear tooth breakage, input shaft imbalance, bent shaft, bearing inner race defect, and bad key, and the method could be further extended for other faults as long as a set of features can be correlated with a known fault signature. Future work looks to further refine the distance calculation algorithm for fault diagnosis, as well as further evaluate other signal processing method such as the empirical mode decomposition to see if an improved set of features can be used to improve the fault diagnosis accuracy.