Dynamic vision-based machinery intelligent fault diagnosis with robustness on camera positions

The event cameras have been successfully developed and applied in different areas in the past years. With the significant advantages of high temporal resolution, low latency, etc ., many challenging tasks have been addressed, such as drone vision, fast object detection and so forth. The event cameras are promising for machine condition monitoring and fault diagnosis, since the vibration effects can be well captured by the bio-inspired vision technology. Using the well-established big data-driven methodologies, the intelligent vision-based fault diagnosis model can be readily built. However, the vision signal suffers from remarkable disturbances in positions, views, distances, etc . Variations in signal collection can easily compromise the model performance, which frequently occur in practice. In this paper, an dynamic vision-based machinery intelligent fault diagnosis method is proposed with special attention on camera position robustness. By creating the self-generated event sample variants, a self-supervised event learning method is proposed for domain generalization regarding camera positions. For processing the easily collected unlabeled parallel data at different camera positions, a cross-supervision learning method is proposed for domain alignment. Extensive experiments on the dynamic vision-based rotating machine fault diagnosis test rig are carried out for validations. The results show the proposed method can effectively extract domain-invariant features from the dynamic vision data, and identify the machine fault modes at different camera positions. It is validated that the proposed method offers a novel perspective and promising tool for robust machine fault diagnosis using vision data from event cameras.