Noise Mask Network-based Feature Learning of Vibration Signals for Machinery Fault Diagnosis under Multiple Non-ideal Conditions

Although deep learning techniques have been successfully applied in machinery fault diagnosis, the key problems of feature learning under multiple non-ideal conditions have not been well solved, such as noise interference, limited labeled data, and class imbalance. In this study, noise mask network (NMNet) is developed for feature learning and machinery fault diagnosis under multiple non-ideal conditions. Firstly, a temporal masking block (TMB) is proposed for imbalanced data enhancement. Secondly, a temporal self-supervised learning framework (TSLF) is developed for noise filtering and solving the problem of limited annotation data simultaneously. In addition, an asymmetric multi-scale autoencoder (AMSA) is constructed for deep feature learning and signal reconstruction. The experiment results on rotor testbed demonstrate the effectiveness of NMNet for machinery fault recognition with several non-ideal conditions.