Dynamic vision-based machine vibration sensing and fault diagnosis with signal alignment and feature clustering

In recent years, intelligent fault diagnosis methods for machines have been widely developed. The contact accelerometer has been popularly used with high measurement accuracy. However, in many industrial applications, non-contact sensors are often preferred due to practical constraints. The event camera is a novel bio-visual sensing technology for asynchronously capturing pixel-wise changes in brightness. It has various advantages including high measurement rate, exceptional resolution, wide dynamic range, etc., which thus has promising prospects in non-contact monitoring and fault diagnostic tasks. Despite these advantages, the high complexity of the dynamic vision data from the event cameras poses significant processing challenges, and the extraction of machine vibration information is of great difficulties. To address these challenges, this paper proposes a novel dynamic vision-based machine vibration sensing and fault diagnosis method. First, the dynamic vision data is reconstructed into event frame sequences. Next, a deep neural network is proposed to extract the micro-vibration information with feature clustering for enhancing model robustness. A signal alignment method is further proposed where the contact sensing data are used as a reference for optimizing model performance. Finally, the intelligent fault diagnosis is implemented with the estimated vibration data. Experimental validations are conducted with real rotating machine data, which demonstrate the promising applicability of the proposed method in non-contact machine vibration sensing and fault diagnosis.