Fault Diagnosis of Aero-Engine Inter-Shaft Bearings Using a Parallel SE-Depthwise Separable Convolutional Neural Network and Transformer Model

Inter-shaft bearings are a significant component within an aero-engine. Previous automated aero-engine bearing fault diagnosis methods mainly training and testing on datasets obtained from standard test bench experiments, such as the Case Western Reserve University (CWRU) rolling bearing dataset. The fault data in these datasets often significantly differ from those obtained during the actual operation of aero-engines, thereby limiting the accuracy and reliability of the diagnostic systems. This paper proposed a novel automatic fault diagnosis method for aero-engine bearings. The proposed method used Fast Fourier Transform (FFT) and Variational Mode Decomposition (VMD) to preprocess bearing fault signals. These signals were then input into a parallel model consisting of a Depthwise Separable Convolutional Neural Network with a Squeeze-and-Excitation Attention Mechanism (SE-DSCNN) and a Transformer. These models independently extracted spatial and temporal features, and finally employed a Cross-Attention Mechanism (CAM) to integrate these features and output the fault classification results. The proposed method in this paper was trained and tested on a real aero-engine inter-shaft bearing failure dataset from the Harbin Institute of Technology (HIT). Meanwhile, we also conducted experiments on the CWRU dataset to facilitate comparisons with the previous work. The results demonstrated that the method proposed achieved a test accuracy of 98.65% on the HIT dataset and a test accuracy of 99.82% on the CWRU dataset. The performance of the proposed method surpassed that of previous related work, proving the effectiveness and reliability of this method in the task of aero-engine bearing fault diagnosis.