Multilayer Adaptive Denoising Network for Incipient Fault Diagnosis of Gear System

The critical issue for early fault diagnosis of gearboxes is to extract fault features as earlier as possible. However, the signal-to-noise ratio (SNR) of the early fault features is low, and they are easily buried by noise. To overcome this problem, this paper proposes a novel multilayer and multiscale adaptive denoising convolutional neural network that utilizes a multilevel filtering structure to search for the optimal filter parameters at different scales, thereby optimizing the denoising performance. The proposed method combines the principles of Wiener filtering and the idea of the multi-scale optimization, which can break through the limitation of traditional filtering and denoising algorithms that can only solve for a single filter parameter. Consequently, the proposed denoising algorithm can achieve better performance. The spalling fault experiment results demonstrate that the proposed method can effectively find the optimal filtering parameters at different scales when dealing with complex noise signals, and the denoised signals have a higher SNR.