TCRNN: A Cross-domain Knowledge Transfer Acoustic Bearing Fault Diagnosis Method for Data Unbalance Issue

It is a non-contact method to diagnose axle box bearing health with the acoustic signal obtained by an acoustic sensor. However, in practice, the acoustic data collected can degrade the diagnostic performance due to data unbalance. At the same time, it is difficult to be applied the deep learning-based fault diagnosis model in reality, due to the lack of bearing acoustic dataset for training. To address these challenges, a Transfer Learning Convolutional Recurrent Neural Network (TCRNN) method based on a transfer learning framework is proposed in this paper. More specifically, the diagnostic knowledge from faulty bearing vibration data is transferred to a classifier network based on the bearing acoustic dataset, thus enhancing the diagnostic capability of the acoustic data-based fault diagnosis model under the data unbalance. The Convolutional Recurrent Neural Network (CRNN) is first trained in the source domain using a public bearing vibration dataset. Then, the trained network parameters are frozen and are used to transfer the fault diagnosis knowledge from the bearing vibration data. Finally, the frozen network is fine-tuned with an unbalanced acoustic dataset in the target domain. The experiments carried out on imbalanced datasets demonstrate the outstanding performance of the acoustic fault diagnosis model proposed in this study.