International Conference on Sensing, Measurement & Data Analytics in the era of Artificial Intelligence (ICSMD 2022)

In bearing fault diagnosis field, training data in different feature space (heterogeneous data), resulted from variable working conditions of rotating machinery, inevitably leads to performance degradation of a well-trained model. Aiming at this problem, the paper presents a new heterogeneous domain adaption (HDA) strategy based on cross-domain structure preserving projection (CDSPP).  Ready for fault diagnosis, a new feature extraction strategy combines noise resistant correlation (NRC)and intrinsic time-scale decomposition (ITD) is proposed to enhance the robustness of signal features. Then, heterogeneous fault vectors from target and source domains are fed into CDSPP model to align the feature distribution by projecting two domains into a common low-dimensional space. The final experiments shows that this method can effectively correct the distributional drift among different feature types and prove that this method is expected to be new technique for boosting the performance of heterogeneous transfer in fault diagnosis task.

bearing fault diagnosis, heterogeneous domain adaptation, noise resistant correlation, cross-domain structure preserving projection