Double Branch Rebalanced Network for Class Imbalance Learning in Process Fault Diagnosis
Yao Ma, Hongbo Shi, Shuai Tan, Yang Tao, Bing Song · Industrial & Engineering Chemistry Research · 2023
In the process of industrial production, the equipment is operating at normal conditions for most of the time, failures are generally rare. Therefore, the real-world data set collected from industrial processes always have skewed distributions; the data in normal conditions are more than that in fault conditions, which result in a class imbalance (CI) problem. Skewed data distributions will affect the ability of feature learning of the monitoring model. To overcome the CI problem, in this paper, a method named double branch rebalanced network (DBRN) is proposed. DBRN is a two-stage method, which learns two decision boundaries by designed convolutional neural network branches. First, in the resampling branch, a novel resampling method is proposed to rebalance the classes from the data-level. Second, a cost adaptive reweighting strategy is developed in the reweighting branch to rebalance the cost of each class. Finally, a fusing learning strategy is designed to fuse these two branches. To describe the proposed method, the experiment-based Tennessee Eastman (TE) data set is constructed with different imbalance ratios. The experiment results show that the proposed DBRN method achieves better performance in CI fault diagnosis.