Multi-head de-noising autoencoder-based multi-task model for fault diagnosis of rolling element bearings under various speed conditions

Fault diagnosis of rolling element bearings (REBs), one type of essential mechanical element, has been actively researched; recent research has focused on the use of deep-learning-based approaches. However, conventional deep-learning-based fault-diagnosis approaches are vulnerable to various operating speeds, which greatly affect the vibration characteristics of the system studied. To solve this problem, previous deep-learning-based studies have usually been carried out by increasing the complexity of the model or diversifying the task of the model. Still, limitations remain because the reason of increasing complexity is unclear and the roles of multiple tasks are not well-defined. Therefore, this study proposes a multi-head de-noising autoencoder-based multitask (MDAM) model for robust diagnosis of REBs under various speed conditions. The proposed model employs a multi-head de-noising autoencoder and multi-task learning strategy to robustly extract features under various speed conditions, while effectively disentangling the speed- and fault-related information. In this research, we evaluate the proposed method using the signals measured from bearing experiments under various speed conditions. The results of the evaluation study show that the proposed method outperformed conventional methods, especially when the training and test datasets have large discrepancies in their operating conditions.