Non-Gaussian noise reconstruction and physics-guided filtering for aero-engine main bearing fault diagnosis under missing fault sample conditions

In recent years, deep learning has been extensively applied to fault diagnosis with remarkable achievements. However, due to factors such as safety constraints and high costs, it is extremely difficult to obtain aero-engine fault samples. Therefore, this article proposes a diagnostic framework for aero-engine main bearings based on non-Gaussian noise reconstruction and physics-guided filtering, which enables fault diagnosis without measured real fault data. First, a joint time–frequency-calibrated non-Gaussian noise synthesis method is proposed. This algorithm extracts higher-order cumulants and power spectral density features from healthy samples and reconstructs non-Gaussian background noise using Gram–Charlier series expansion and spectral tuning techniques. Subsequently, a physics-guided time–domain filter is constructed. By integrating structural resonance effects, periodic impact pulses, and complex modulation characteristics, this filter achieves deep modulation between the noise field and fault excitation pulses via convolution operators, thereby generating simulated fault data with minor differences from measured data. Finally, a generalized fault diagnosis network is constructed based on the dilated convolution network and contrastive learning to capture long-range temporal dependencies and construct a highly discriminative feature manifold, thereby achieving robust knowledge transfer from the simulated domain to the real domain. Experiments are conducted on the aero-engine bearing dataset, a self-built bearing dataset, and the Shandong University of Science and Technology dataset. Through extensive recognition experiments conducted on the three datasets, the proposed method demonstrates excellent diagnosis performance on tasks without real fault data.