Dynamic vision-enabled contactless machine fault diagnosis with neuromorphic domain generalization

In the recent years, breakthroughs in data-driven approaches have significantly enhanced the accuracy and efficiency of mechanical equipment fault diagnostics. Generally, vibration signals obtained through contact-based measurement remain the primary data source for mechanical fault detection methodologies. However, limitations are noted in many industrial scenarios such as small installation space, high temperature, etc. In comparison, contactless vibration sensors are expected to overcome the limitations and have broader application prospects. The bio-inspired event camera realizes contactless dynamic vision through asynchronous triggering mechanisms, demonstrating microsecond temporal resolution, wide illumination adaptability, and lower data redundancy than frame-based imaging. These characteristics facilitate contactless vibration monitoring in machinery. In practice, event-based measurements exhibit strong dependence on sensor pose variations, ambient illumination, and the operating conditions of mechanical equipment. Consequently, dynamic vision-based fault diagnosis typically exhibits poor generalization performance. To overcome these limitations, a novel dynamic vision-enabled fault diagnosis method for mechanical equipment is proposed in this paper. The event camera is used to capture the vibration signal in a contactless manner. A specially designed neuromorphic model is proposed for spatiotemporal signal processing and feature extraction. A domain generalization technique based on kernel methods is further proposed, which significantly enhances the model generalization performance. The experimental validation across two rotating machine datasets have validated the diagnostic robustness of the proposed method. Benefiting from the contactless nature, the proposed method shows promising industrial application prospects, while ensuring the fault diagnosis accuracy.