Fault Diagnosis of Aero-Engine Bearings in Comprehensive Noise Environments Based on Multi-Scale Convolution Networks

For aero-engines operating in harsh environments, the presence of substantial noise poses challenges to effective fault diagnosis, a multi-scale attention residual network (MCRN) is proposed. Firstly, time-frequency images are obtained by short-time Fourier transform of the raw vibration signals. Secondly, multi-scale convolution is applied to feature extraction from the time-frequency images, and channel attention mechanism is employed to suppress irrelevant channels. Simultaneously, a spatial attention mechanism is employed to filter out useless spatial features. While considering the presence of noise interference, the concept of noise as an interval is proposed and injected into the signal. Lastly, the requirement of certain prior knowledge for interval noise determination is acknowledged, and an approach considering comprehensive noise injection is proposed. Through a series of comparative experiments, it is found that the MCRN proposed in this study can maintain a high accuracy rate and high robustness, which makes it suitable for applications in actual complex working conditions.