K-SVD-based WVD enhancement algorithm for planetary gearbox fault diagnosis under a CNN framework

This paper presents a new method of planetary gearbox fault diagnosis by dealing with and analyzing vibration signals. This study contributes to the realization of automatic diagnosis using a convolution neural network (CNN) to process time-frequency distributions (TFDs) transformed from vibration time series. In order to solve the problem of non-stationary working states and strong noise interference in industrial applications, a K-singular value decomposition (K-SVD) is used to enhance the resolution of TFDs obtained by Wigner–Ville distribution (WVD), a typical time-frequency transform algorithm. The simulation results indicate that K-SVD can not only reduce the effects of cross-terms on WVDs but can also eliminate noise, which makes the fault characteristics outstanding in the time-frequency domain. The enhanced WVDs improve the accuracy of fault diagnosis in a classification framework based on the CNN that can extract features adaptively and obtain a high degree of discrimination between different fault conditions. Finally, the effectiveness of the proposed method is verified by a prototype experiment with roller bearings and a scale test rig of a planetary gearbox from a ship unloader. Moreover, a priority confusion matrix is proposed as a visualization tool with which to evaluate the performance of a fault diagnosis model. The results open the possibility of extrapolating the method to the fault diagnosis of other mechanical parts.