Bearing Fault Diagnosis with Imbalance Dataset Based on Modified Bilateral-Branch Networks

The classical intelligent diagnosis methods based on deep learning generally require huge amount of labeled data with balanced samples among classes to train the model. However, for rolling bearing dataset in engineering scenarios, samples in normal condition occupy the vast majority of runtime data, while fault samples account for only a small fraction. The bilateral-branch networks (BBN) can effectively improve the recognition performance for the imbalance long-tailed dataset. This paper proposes a framework of modified BBN for bearing fault diagnosis with imbalance dataset, which aims to further enhance the diagnostic performance. Specifically, the modified BBN improves the diagnostic performance of imbalance tasks in three aspects. First, instance-class sampling strategy is adopted in the rebalancing branch to balance the attention of feature extraction among the imbalance classes. Second, parabolic increment curve is selected for the trend of trade-off parameter in the cumulative learning to alleviate the distribution distortion of the learned features. Third, τ-normalization is introduced for the trained classifier to further rebalance the classification among the imbalance classes via moving the decision boundaries. Experiments on a bearing dataset validate that the modified BBN proposed is superior to the existing methods, especially on extremely imbalance tasks.