A multi-objective feature selection and self-paced ensemble framework for semiconductor defect detection

english

2025

volume 20, issue 4

class imbalance   class overlap   defect detection   machine learning   multi-objective feature selection   quality inspection   self-paced ensemble   semiconductor manufacturing

Univerza v Mariboru, Fakulteta za strojništvo
(Obvezni izvod spletne publikacije)

In semiconductor manufacturing, defect detection is commonly performed using high-dimensional process data. These data often exhibit class imbalance and class overlap, which create challenges for achieving reliable classification performance. To address these issues, this study proposes a multi-objective feature selection and self-paced ensemble (MOFS-SPE) framework. The framework employs a multi-objective evolutionary algorithm based on decomposition (MOEA/D) for feature selection. In this process, the area under the precision–recall curve (AUPRC) and the R-value are used as objective functions to identify feature subsets that are highly relevant to quality outcomes. In addition, the framework integrates the self-paced ensemble (SPE) with tree-based classifiers to handle imbalanced and overlapping data. Experiments conducted on a real semiconductor manufacturing dataset (SECOM dataset) demonstrate the effectiveness of the proposed approach. Compared with using the full feature set, the selected features increase the area under the receiver operating characteristic curve (AUROC) from 0.685 to 0.770 and the AUPRC from 0.932 to 0.972. When applying the SPE framework, the specificity of the decision tree model improves from 0.048 to 0.667, thereby enhancing the reliability of identifying defective products. Overall, the proposed framework provides a useful reference for intelligent quality inspection in semiconductor production environments.

06.07.2026