A Cost-Sensitive Multi-scale Feature Multi-order Fusion Network for Bearing Fault Diagnosis Under Data Imbalance Conditions

Railroad transportation plays a vital role in the national economy, and fault diagnosis of its key components is of great significance in ensuring its safe operation. However, due to its long-term operation in a healthy state, it is tough to obtain fault signals, and this class imbalance data seriously degrades the diagnostic capability of intelligent diagnostic algorithms, leading to increased misclassification of minority classes. To address this problem, we propose a cost-sensitive multi-scale feature multi-order fusion network for bearing fault diagnosis under data imbalance conditions. Firstly, a lightweight multi-scale convolutional layer is employed to effectively extract the multi-scale features of the signal. Then the multi-order feature fusion is performed by a hornet to fully utilize the multi-scale information and enhance the ability to mine the limited diagnostic information of the minority class. And the excellent high-dimensional discriminative features are extracted through residual blocks. In addition, an improved cost-sensitive strategy is designed through the posterior probability distribution to rebalance the contribution of each class to the training process for improving the overall diagnostic performance of the model. Finally, the effectiveness and superiority of the proposed method are comprehensively verified by several sets of experiments with different imbalance rates.