A multi-scale gated spatial attention network for rotor fault diagnosis under various conditions and equipment

To address the challenges of data distribution discrepancies between the source and target domains and the insufficiency of on-site fault-labelled samples in rotor fault diagnosis, a novel method called multi-scale gated spatial attention network (MsGSAN) is proposed. Within this method, a gated spatial attention module is proposed, which helps the model adaptively adjust the input features and identify key spatial location information within the features, enhancing the feature representation ability. Feature extraction is performed using an improved multi-scale convolution module and attention module, obtaining rich fault feature scales and achieving the integration of both global and local features. Through the domain adaptation method, the distribution discrepancy of fault feature extraction between the source and target domains is minimised, achieving feature alignment in cross-domain diagnosis. The proposed method is tested using two different rotor equipment. The results demonstrate its higher accuracy in both cross-condition and cross-equipment rotor fault transfer tasks, offering a valuable reference for cross-domain rotor fault diagnosis.