Imbalanced fault diagnosis of planetary gearboxes based on noise enhancement and threshold adaptive Siamese decoupled network.

In the case of sufficient and balanced training data, the intelligent diagnosis models can accurately determine the state of the planetary gearbox and play a significant role in ensuring its healthy operation. However, the planetary gearbox operates normally for much longer than the moment of failure in practical engineering, which makes the sample size of fault state extremely small and the training data imbalanced. As a result, the model fail to detect the extremely small samples and thus serious fault missed diagnosis. In order to improve the performance of imbalanced diagnosis of planetary gearboxes with containing extremely small samples, this paper proposed an imbalanced fault diagnosis method for planetary gearboxes based on noise enhancement and threshold adaptive Siamese decoupled network. Firstly, the extremely-samples are enhanced into small samples by adding noise appropriately, and a set of metrics are proposed to evaluate the quality of the enhanced samples. Then, the Siamese network is constructed, and the special input requirements of the Siamese network are used to expand the training data again, which solves the problems of poor generalization and missed diagnosis caused by small samples and imbalance. Finally, a threshold adaptive and multi-scale decoupled convolution is proposed to improve the Siamese network and further improve the diagnostic performance. It is verified by imbalanced planetary gearbox data. On the imbalanced training data with small samples, the diagnostic accuracy of the proposed method was up to 98.33 %. In the extreme cases of high imbalance (fault / total < 5%) and small sample size of fault (only 3 samples per class), the diagnostic accuracy still reached 71.11%. It shows that the proposed method has great advantages and potential in imbalanced fault diagnosis with small samples.