Learnable Wavelet Packet Denoising Network for Aero-Engine Fault Feature Detection

It is a challenging problem to extract aero-engine fault signals which contaminated by non-Gaussian noises under complex operation conditions. A Learnable Wavelet Packet De-noising Network (LWPD-Net) is proposed in this paper to address it. The highlights of LWPD-Net are to performs multi-scale decomposition of the original signal to extract features across various frequency bands, and more importantly the Double-Sharp threshold function with learnable parameter is employed to suppresses the non-Gaussian noise. Moreover, supervised learning strategy is adopted to learn LWPD-Net parameters for adaptively enhancing the fault features through conducting training samples from fault signals corrupted by non-Gaussian noises. Simulation experiments show that LWPD-Net can recover fault feature frequencies in the envelope spectrum for varying noise levels, and achieves satisfying fault feature extraction performances. Additionally, experiments conducted on aero-engine gear hubs confirm that the proposed method can recover the spectral features of vibration signals interfered with non-Gaussian noise under different operation conditions.