A novel transfer learning fault diagnosis method based on Manifold Embedded Distribution Alignment with a little labeled data

Accurate identification of rolling bearing faults is quite significant for the stable operation of mechanical systems. However, for practical diagnosis issues, it is difficult to obtain abundant labeled data due to the change of operating conditions and complex working environment, which puts forward higher requirements on the ability of the diagnosis methods. To tackle the mentioned problem, a novel transfer learning method based on a little labeled data is proposed, which uses bidirectional gated recurrent unit (BiGRU) and Manifold Embedded Distribution Alignment (MEDA). Firstly, frequency spectrum datasets are utilized to remove the redundant information of raw vibration signals. Secondly, the BiGRU network is constructed to generate auxiliary samples that are utilized as source domain. Finally, MEDA, as the most powerful non-deep transfer learning method, is applied to align the distribution of these auxiliary samples generated by BiGRU and the unlabeled samples from target domain. Experiment results indicate the excellent performance of the proposed method under a little labeled data.