Bearing Fault Diagnosis Based on Multi-task Learning

In recent years, machine learning techniques have been successfully applied to analyzing vibration signal in bearing fault diagnosis problems. However, the bottleneck to improve the diagnosis performance is the lack of valid domain knowledge about bearing fault in the classification or regression model, especially when the collected fault data is insufficient. Moreover, inadequate fault information will result in unstable diagnosis result. To solve this problem, a new bearing fault diagnosis method based on multi-task learning is proposed in this paper. The intuitive point is the bearing fault with similar crack size and loads can provide useful domain information for each other. By considering the diagnosis model on one bearing as a task, this method uses marginal spectrum by Hilbert-Huang Transform as features, and constructs multi-task learning model on multiple related diagnosis tasks with different fault types to improve their diagnosis performance simultaneously. Compared to the single diagnosis task, the proposed method shares the domain information between various fault types and then builds a more efficient diagnosis model. Experimental results on CWRU bearing data set show that, compared with some traditional machine learning-based diagnosis methods, the proposed method can effectively improve the diagnosis accuracy and robustness under different working conditions and it's especially suitable for dealing with lots of tasks with insufficient data. Even with insufficient features, the proposed method gets the accuracy from 77.50% to 94.17%, which indicates the multi-task diagnosis can get help from related bearing fault information.