Fault feature extraction for roller bearings based on DTCWPT and SVD

Aiming at difficulty in extracting fault feature from raw non-stationary and complex vibration signal with noise interference as well, a feature extraction method for roller bearings based on double-tree complex wavelet package transform (DTCWPT) and singular value decomposition (SVD) is proposed. DTCWPT is used to extract the component which expresses the fault feature most obviously among all the components of the decomposed signal. A one dimension signal can be transformed into a matrix through continuous truncation. By performing SVD on the matrix, singular values are obtained which can present the inherent characters of the matrix. To evaluate the classifying performance of proposed feature, Fisher measure is introduced and computed. Four roller bearings operating conditions such as inner race spalling, outer race spalling, roller element spalling and normal are simulated in experiment rig to test the performance of the proposed feature. The result suggests that the mean of singular values performs better in distinguishing the above four conditions of roller bearings than traditional characters such as root mean square (RMS), kurtosis and sample entropy.