Bearing fault diagnosis based on optimized variational mode decomposition and 1D convolutional neural networks

Due to the fact that measured vibration signals from a bearing are complex and non-stationary in nature, and that impulse characteristics are always immersed in stochastic noise, it is usually difficult to diagnose fault symptoms manually. A novel hybrid fault diagnosis approach is developed for denoising signals and fault classification in this work, which combines successfully variational mode decomposition (VMD) and a one-dimensional convolutional neural network (1D CNN). VMD is utilized to remove stochastic noise in the raw signal and to enhance the corresponding characteristics. Since the modal number and penalty parameter are very important in VMD, a particle swarm mutation optimization as a novel optimization method and the weighted signal difference average as a new fitness function are proposed to optimize the parameters of VMD. The reconstructed signals of mode components decomposed by optimized VMD are used as the input of the 1D CNN to obtain fault diagnosis models. The performance of the proposed hybrid approach has been evaluated using sets of experimental data on rolling bearings. The experimental results demonstrate that the VMD can eliminate signal noise and strengthen status characteristics, and the proposed hybrid approach has a superior capability for fault diagnosis from vibration signals of bearings.