An Adaptive Reweighting Method Based on the Effective Area for Fault Severity Classification under Class Imbalance

In real-world scenarios, fault severity data follows an imbalance distribution, meaning that normal and low-level faults constitute a large portion of the training data. In contrast, high-level fault instances are relatively scarce. This imbalance leads to the decision hyperplane of classifiers to skew in favor of higher fault levels. Typically, reweighting methods ensure that hyperplanes fairly divide the feature space by assigning weights based on the number of instances. However, we found that the number of instances fails to represent the effective area within a fault level in the feature space. To tackle this issue, we first analyze the distribution and their correlations of different instances within each class. Second, we calculate the effective area by treating instances of the same class as identically distributed random variables. Third, we introduce a novel weighting approach, with weights inversely proportional to the effective area. This method results in more flexible weights, allowing for adaptive adjustments based on feature optimization during the training process. Experimental results demonstrate that the proposed method outperforms the comparative methods on a real-world fault severity dataset and enhances fault diagnosis performance under class imbalance.