Feature Selection for Data Classification in the Semiconductor Industry by a Hybrid of Simplified Swarm Optimization

In the semiconductor manufacturing industry, achieving high yields constitutes one of the pivotal factors for sustaining market competitiveness. When confronting the substantial volume of high-dimensional, non-linear, and imbalanced data generated during semiconductor manufacturing processes, it becomes imperative to transcend traditional approaches and incorporate machine learning methodologies. By employing non-linear classification models, one can achieve more real-time anomaly detection, subsequently facilitating a deeper analysis of the fundamental causes behind anomalies. Given the considerable dimensionality of production line data in semiconductor manufacturing, there arises a necessity for dimensionality reduction to mitigate noise and reduce computational costs within the data. Feature selection stands out as one of the primary methodologies for achieving data dimensionality reduction. Utilizing wrapper-based heuristics algorithms, although characterized by high time complexity, often yields favorable performance in specific cases. If further combined into hybrid methodologies, they can concurrently satisfy data quality and computational cost considerations. Accordingly, this study proposes a two-stage feature selection model. Initially, redundant features are eliminated using mutual information to reduce the feature space. Subsequently, a Simplified Swarm Optimization algorithm is employed to design a unique fitness function aimed at selecting the optimal feature subset from candidate features. Finally, support vector machines are utilized as the classification model for validation purposes. For practical cases, it is evident that the feature selection method proposed in this study achieves superior classification accuracy with fewer features in the context of wafer anomaly classification problems. Furthermore, its performance on public datasets further substantiates the effectiveness and generalization capability of the proposed approach.