Few-Shot Gas Path Fault Diagnosis with Omni-Dimensional Attention based on Meta-Transfer Learning

Gas path datasets of aero-engine usually contain large amount of healthy condition samples but few fault samples especially rare fault samples, which are not suitable for training Deep Neural Networks(DNN). To improve diagnostic accuracy, Omni-dimensional Attention Meta-transfer Learning (OAMTL) is proposed. The OAMTL has a three-step process: pre-training, meta-transfer learning, and meta-adjustment. This approach allows rapid adaptation to new faults by using knowledge obtained from previously known faults. Firstly, we used Gramian Angular Difference Field (GADF) to convert each sensor signal into grayscale fault images and then aggregated vertically to form a single composite image. Secondly, on the basis of ResNet18, we embedded Omni-dimensional Attention model to effectively explore multi-dimensional information and identify the corresponding fault characteristics. Furthermore, we develop a novel algorithm for gas path faults diagnosis. This approach not only obtains more accurate fault features, but also mitigates the risk of overfitting, which improves diagnosis performance and accuracy. Finally, we used experimental data from the Aero Engine Academy of China to validate the proposed method. The experimental results validate that the average accuracy of the new fault datasets reaches $\mathbf{9 7 . 5 1 \%}$ and $\mathbf{9 8 . 3 9 \%}$ for 1 -shot and 5 -shot conditions.