Graph embedding deep broad learning system for data imbalance fault diagnosis of rotating machinery

The distribution of monitored data during the service life of machinery equipment is imbalanced, especially there is more monitoring data for health conditions than for failure conditions. Unfortunately, most existing intelligent fault diagnosis methods are built on the assumption of data balance and cannot effectively handle unbalanced data. Therefore, to solve the above problem, a graph embedding based deep broad learning system (GEDBLS) for data imbalance fault diagnosis of rotating machinery is proposed in this paper. Different from the traditional broad learning system (BLS), the designed GEDBLS not only utilizes the category and structure information of the data in the reconstruction process, but also allows learning the high-level abstract features of the vibration signal through a progressive encoding and decoding mechanism. In addition, GEDBLS considers category weights and intra-class tightness in the classification loss function for imbalanced data category classification. The effectiveness of the presented approach is verified via monitoring data from key equipment of rotating machinery. Experimental results indicate that compared with other methods, the proposed method has stronger ability of feature representation and imbalanced data processing.