Rolling bearings fault diagnosis under various loads conditions based on GWJO-VMD-MAFCNN

Abstract

In response to the unclear fault characterization of rolling bearing vibration signals due to their nonlinear and nonstationary characteristics, a rolling bearing fault diagnosis method based on grey wolf jackal optimization algorithm variant mode decomposition multi-activation function convolutional neural networks (GWJO-VMD-MAFCNNs) is proposed. The multi-activation function convolutional networks (MAFCNNs) utilize both Tanh and Softmax activation functions to achieve better diagnostic results. Meanwhile, we leverage the GWJO to optimize the parameters σ and K of VMD, employing the refined VMD for fault diagnosis of the original bearing signals. Subsequently, the permutation entropy of the multiple intrinsic mode function components decomposed by VMD is used as a sample, which is then input into the MAFCNN for training. The experimental results show that the GWJO-VMD-MAFCNN achieved a 100% accuracy rate in all five trials, demonstrating the best diagnostic performance. Additionally, the proposed MAFCNN had the smallest validation loss, with the GWJO-VMD-MAFCNN reaching a validation loss of 2.13e⁻⁶, indicating that the proposed method has a high degree of accuracy. The comparative test data demonstrates that the performance of the GWJO-VMD-MAFCNN is superior to the existing optimization techniques.