A self-adaptive class-imbalance TSK neural network with applications to semiconductor defects detection

This paper develops a hybrid approach integrating an adaptive artificial neural network (ANN) and a fuzzy logic system for tackling class-imbalance problems. In particular, a supervised learning ANN based on Adaptive Resonance Theory (ART) is combined with a Tagaki-Sugeno-Kang-based fuzzy inference mechanism to learn and detect defects of a real large highly imbalanced dataset collected from a semiconductor company. A benchmark study is also conducted to compare the classification performance of the proposed method with other published methods in the literature. The real dataset collected from the semiconductor company intrinsically demonstrates class overlap and data shift in a highly imbalanced data environment. The generalization ability of the proposed method in detecting semiconductor defects is evaluated and compared with other existing methods, and the results are analyzed using statistical methods. The outcomes from the empirical studies positively indicate high potentials of the proposed approach in classifying the highly imbalanced dataset posing overlap class and data shift.