Abstract

Detection of faults in the early stages for rotating machinery is important for optimizing maintenance chores and avoiding severe damages to other parts. An approach based on Dictionary learning for sparse representation aiming at gearbox fault detection is proposed. A gearbox vibration signal database with 900 records considering the normal case and nine different faults is analyzed. A dictionary is learned using a training set of signals from the normal case. This dictionary is used for obtaining the representation of signals in the test set considering either normal or faulty condition vibration signals. The dictionary based representation is analyzed for extracting features useful for detection of faults. The analysis is performed considering different load conditions. Additionally the Analysis of Variance (ANOVA) is performed for ranking the extracted features. Results are promising as there are significant statistical differences between the normal case and each of the recorded faults. Comparison between faults also shows that faults tends to group into several clusters in the feature space where classification of faults could be feasible.