A local and global multi-head relation self-attention network for fault diagnosis of rotating machinery under noisy environments

Abstract

Fault diagnosis under noisy environments (FDUNE) for rotating machinery is a highly challenging task. In recent years, deep learning models have become research hotspots in the field of FDUNE. However, the existing FDUNE approaches suffer from a limitation that insufficient consideration of both local and global features in the feature extraction process leads to unsatisfactory diagnostic performance. In this paper, a local and global multi-head relation self-attention network (LGMHRSANet) is proposed to improve the diagnostic accuracy of rotating machinery under noisy environments, which integrates convolution and self-attention into the transformer form, enabling it to capture local features and global long-range temporal features from vibration signals. Two experimental cases on rolling bearings and gearboxes are implemented to verify the effectiveness of LGMHRSANet under noisy environments. Experimental results demonstrate that LGMHRSANet has superior diagnostic performance compared to other deep learning models, regardless of whether it is in a non-noise environment, or a strong noise environment. In addition, the adaptive performance analysis in the variable noise domain indicates that LGMHRSANet has good robustness in noisy environments.