Bearing fault diagnosis under different operating conditions based on cross domain feature projection and domain adaptation

This paper focuses on the poor adaptability of fault diagnosis model under different operating conditions and a new transfer learning frame for diagnosis based on Joint Geometrical and Statistical Alignment (JGSA) is presented to solve this problem. Based on the extraction of sub-band energy in frequency, JGSA model is used to create two coupled projecting matrices and map training and test data into two subspaces. Data distribution shift between different domains is reduced statistically and geometrically in projecting spaces. Then Support Vector Machine (SVM) is established on the projecting feature space subsequently. The framework used in this paper is more adaptive for complex industrial process since it can be conducted on different domains without the prior whether they are similar or not. The bearing experiments results under different operating conditions show that the proposed framework based on JGSA works well when data distributions of different domain arc similar and it can promote the performance of general classifier when distribution divergence between different domains is large.