An Unsupervised Rotating Machinery Fault Diagnosis Method Based on Multi-Scale Feature Residual Neural Network

Abstract

Abstract In complex settings, noise affects rotating parts like bearings and gears, weakening fault signals and complicating feature selection, resulting in redundancy. To address this, a multi-scale residual neural network is proposed for machinery fault diagnosis with domain adaptation. Using residual connections, it fuses vibration signal features from a multi-scale network for a global view. The method employs maximum mean discrepancy and entropy boundaries for adaptation, enhancing signal classification. Successful unsupervised cross-domain fault diagnosis is shown in experiments. Future work aims to refine the network architecture and enhance generalization through advanced data augmentation.