For all kinds of rotating machinery, rotating machinery, gearboxes, rotating tools and other components are prone to bearing operation problems due to long-term vibration and wear. In addition, in the actual operation of the factory machine, there is often no manpower for data collection and classification. Therefore, it is necessary to design a system that can pre-process and classify unlabeled data. This research proposes a system that uses multi-scale entropy for feature extraction and complex scale vector data to describe bearing vibration signal analysis. In this study, the IMS bearing database was used for testing. The experimental results can accurately determine when the bearing is abnormal and remind the user that the bearing is damaged. And in this way, unsupervised learning can be realized, pre-processing and analysis can be performed by itself.