Self-supervised fusion of deep soft assignments for multi-view diagnosis of machine faults

Abstract

Fault patterns are often unavailable for machine fault diagnosis without prior knowledge. This makes it challenging to diagnose the existence of machine faults and their types. To address this issue, a novel scheme of deep soft assignments fusion network (DSAFN) is proposed for the self-supervised multi-view diagnosis of machine faults. To enhance the robustness of the model and prevent overfitting, random noise is added to the collected signals. In each view, vibration features are extracted by a denoising autoencoder. Using the extracted deep features, a soft assignment fusion strategy is proposed to fully utilize both the public and complementary information of multiple views. Critical diagnosis missions, including novel fault detection and fault clustering, are accomplished through binary clustering and multi-class clustering of DSAFN, respectively. Two diagnostic experiments are conducted to validate the proposed method. The results indicate that the proposed method performs better than state-of-the-art peer methods in terms of diagnostic accuracy and noise robustness.