A fault diagnosis method for intermediate bearings based on improved variable mode decomposition and wide-kernel deep convolutional neural networks

Aiming at the issues of the intermediate shaft bearings， where background noise makes fault feature extraction difficult and diagnostic recognition rates low， a novel fault diagnosis approach is proposed. This method combines the Hippopotamus Optimization Algorithm （HO） improved Variational Mode Decomposition （VMD） with the Deep Convolutional Neural Networks with Wide First-layer Kernel （WDCNN）. Initially， this method constructs a multi-dimensional joint index， formed by integrating envelope spectrum energy， kurtosis， and permutation entropy， serves as the fitness function. The HO algorithm is then applied to optimize and determine the key parameters of VMD. After decomposing the original signal， effective modal components are selected and reconstructed following the correlation coefficient-sample entropy criterion. Finally， the reconstructed signal is fed into the WDCNN model for fault diagnosis. Experimental results from the Case Western Reserve University bearing dataset show that the reconstructed signal achieves a diagnosis accuracy of 98.30%. It also outperforms support vector machines， backpropagation neural networks， and one-dimensional convolutional neural networks in terms of accuracy. For data from a self-built intermediate bearing fault simulation test bench， the average accuracy of multiple diagnoses is 96.4%， with the average accuracy of various classification methods optimized by 18.9% and stability by 52%. In conclusion， the proposed method can eliminate non-linear interference like noise and significantly enhance the classification accuracy of fault diagnosis models.