A CNN-based Method for Multisensor and Multiscale Fault Diagnosis of Marine Electric Thruster Bearings under Noise Conditions

Bearings are indispensable supporting parts of the marine electric thruster (MET). Due to long-term work in harsh marine environments, bearings are more prone to failures. However, the diagnosis of MET bearings using vibration signals is a difficult task since the collected signal is nonlinear and nonstationary, and its fault features are easily swallowed by environmental noise. This paper proposes a novel multisensor and multiscale convolutional neural networks (MM-CNN) method to extract different scale features under low signal-to-noise ratios and diagnose the faults of MET bearings. Firstly, the principal component analysis is introduced to fuse multiple signals so as to reduce redundant information. Secondly, a multiscale convolutional module is designed for the feature extraction at different scales by four groups of convolution kernels with different dilate rates. The experimental results on the MET bearing dataset demonstrate that the proposed method outperforms the other state-of-the-art deep learning methods in terms of accuracy and antinoise ability.