Abstract

Fault detection in rotating machinery is important for optimizing maintenance chores and avoiding severe damages to other parts. Signal processing based fault detection is usually performed by considering classical techniques for alternative representation of significant signals in time domain, frequency domain or time-frequency domain. An approach based on dictionary learning for sparse representations of vibration signals aiming at gearbox fault detection and classification is proposed. A gearbox signal dataset with 900 records considering the normal case and nine fault classes is analyzed. A dictionary is learned by using a training set of signals from the normal case. This dictionary is used for obtaining the sparse representation of signals in the test set and the norm metric is used to measure the residual from the sparse representation. The extracted features are useful for machine learning based fault detection. The analysis is performed considering different load conditions. ANOVA statistical analysis shows that there are significant differences between features in the normal case and each of the faulty classes, and best ranked features form well separated clusters. An experiment of fault classification is developed using a support vector machine for multi-class classification of faults. The accuracy obtained is 95.1% in the cross-validation testing.