Multigrained Hybrid Neural Network for Rotating Machinery Fault Diagnosis Using Joint Local and Global Information

Deep learning (DL) models, such as multilayer perceptrons (MLPs) and convolutional neural networks (CNNs), have strong feature representation and nonlinear mapping capabilities, and their effectiveness has been demonstrated in fault diagnosis. However, fault features usually occur at different scales and are always disturbed by noise, making it difficult for DL-based models to learn local and global information in mechanical vibration signals. To address this issue, a multigrained hybrid neural network named MgHNN is proposed to extract robust features that seamlessly integrate CNN into vision MLP. First, the short-time Fourier transform (STFT) is performed on original vibration signals to obtain time–frequency images, and each image is then divided into multiple nonoverlapping patches. Second, a novel multigrained feature representation (MFR) block is proposed by constructing hierarchical residual-like connections within one single wave block, which is more suitable for learning hierarchical local and global feature representations among different image patches. Third, we propose a depthwise wave (DWwave) block by integrating depthwise convolution and feature concatenation operations, which can make MFR block better focus on the local information and effectively overcome the vanishing gradient problem. Finally, experimental results on two fault diagnosis datasets demonstrate that the MgHNN has improved diagnostic accuracy and reduced model complexity compared to state-of-the-art models. The results of noise interference experiments indicate that the MgHNN exhibits superior robustness against noise.