A ViTMCA network for cross-domain fault diagnosis of aeroengine gas path

Aero-engine gas path fault diagnosis faces cross-domain challenges due to differences in complex working conditions throughout the life cycle. Conventional diagnosis methods are difficult to cope with the inconsistency of data distribution between source and target domains, which leads to a significant decrease in the diagnostic effectiveness of the model on data from different domains. To improve the diagnostic performance for aeroengine gas path component in unlabelled cross-domain scenarios, a visual transformer multi-scale convolutional attention (ViTMCA) network structure is proposed for feature extraction and a loss function combined domain adversarial loss and feature alignment loss is developed for domain adaptation in this paper. The experimental results verify the adaptability and robustness of ViTMCA for cross domain fault diagnosis, achieves the highest accuracy of 95.84% in target domain diagnosis. The combination of multi scale convolution and self-attention mechanism proving the importance of local and global feature extraction and multi scale feature fusion for domain adaption. The integration of domain adversarial loss and Multi Kernel Maximum Mean Difference (MK-MMD) multi-layer feature alignment loss in the domain adaptive loss function realizes stronger domain invariant representation learning across source and target domain.