Unbalanced Few-Sample Fault Diagnosis Approach for Mechanical Equipment Based on Multi-Scale Class-Weighted Balanced Network

Mechanical equipment operating in extreme environments is prone to damage critical components, thus affecting its performance and remaining service life, which leads mechanical equipment diagnosis to be a focus. Data-driven fault diagnosis approaches have been recognized as the emerging research directions. However, they are limited by the fact that only few-sample unbalanced datasets can be applied in real industrial backgrounds, leading to serious overfitting and performance degradation. Therefore, an unbalanced few-sample fault diagnosis approach for mechanical equipment based on multi-scale class-weighted balanced network is proposed to improve the diagnosis performance by gradually resolving the quantitative and qualitative imbalance among classes. In the feature extraction stage, multi-scale feature fusion convolution module and channel class balancing mechanism are designed to enhance the ability of recognizing few classes. In the network optimization stage, the quantitative imbalance among classes and the accuracy imbalance within classes are firstly calculated to weight the cross-entropy loss, and then the triplet loss in metric learning is added, making it possible to ignore the quantitative and qualitative imbalance and guide the network to efficiently learn the features of all the classes. Experiments are conducted on the Case Western Reserve University bearing dataset to validate the effectiveness. The results show that even under the most extreme unbalanced few-sample condition, the average fault diagnosis accuracies of the proposed approach under all operating conditions are still above 94%, which are better than comparison approaches.