Research on Methods for Balancing Component Sample Categories Under Rule Constraints

To tackle the issue of low classification accuracy caused by imbalanced sample categories and the local optima problem in classification models for electronic components in aerospace products, this paper introduces a rule-constrained majority class hybrid under-sampling technique for sample category balancing. First, hash encoding is used to uniformly map high-dimensional features to a fixed number of low-dimensional spaces. This process reduces sample collision probabilities. Second, under defined rule constraints, a hybrid under-sampling method combining random under-sampling and the NearMiss-2 algorithm balances the samples of large-scale electronic components in aerospace products. This ensures a uniform sample distribution and clear category boundaries. Finally, the performance of component classification models using different data encoding and under-sampling methods was compared with KNN, Random Forest, and SVM as base classifiers. The results show that the proposed method enhances classification accuracy across different models. Optimized with hash encoding and hybrid under-sampling, the models achieved a classification accuracy of 94.28%. This effectively addresses the problem of low accuracy due to imbalanced data.