基于无监督特征选择的半导体质量检测方法

Semiconductor Quality Inspection Method Based on Unsupervised Feature Selection

摘要: 半导体质量检测数据存在高维冗余、类别不平衡及标签获取成本高的特性，导致传统有监督检测方法工业适用性受限，而传统无监督特征选择方法往往忽略全局结构或缺乏冗余量化机制，难以适配高维强非线性耦合的数据需求。基于此，本文提出一种面向半导体高维制造数据的无监督特征选择方法FSSC-DCOR (Feature Selection by Spectral Clustering and Distance CORrelation coefficient)。该方法结合谱聚类、距离相关系数与贪心策略三种技术，以特征为聚类对象，通过谱聚类挖掘特征内在关联结构并筛选高信息量候选特征，利用距离相关系数矩阵量化非线性冗余，最终通过贪心策略保留低冗余、高区分度的核心特征子集。该方法无需依赖标注标签即可完成高维数据有效降维，适配半导体场景标签稀缺的现实需求。实验结果表明，在SECOM半导体数据集上，所提方法的性能度量指标均优于传统特征选择方法。

Abstract: Semiconductor quality inspection data exhibits characteristics of high-dimensional redundancy, class imbalance, and high cost of label acquisition, resulting in limited industrial applicability of traditional supervised detection methods. In contrast, conventional unsupervised feature selection methods either ignore the global structure or lack redundancy quantification, making it difficult to meet the requirements of high-dimensional and strongly nonlinearly coupled data. To address this issue, an unsupervised feature selection method named FSSC-DCOR (Feature Selection by Spectral Clustering and Distance CORrelation coefficient) is proposed for high-dimensional semiconductor manufacturing data. This method combines three techniques: spectral clustering, distance correlation coefficient, and greedy strategy. Taking features as clustering objects, it mines the intrinsic correlation structure of features through spectral clustering to select high-information candidate features, quantifies nonlinear redundancy using a distance correlation coefficient matrix, and finally retains a core feature subset with low redundancy and high discriminability via the greedy strategy. Without relying on labeled data, the method can achieve effective dimensionality reduction of high-dimensional data, adapting to the practical demand of label scarcity in semiconductor scenarios. Experimental results demonstrate that on the SECOM semiconductor dataset, the performance metrics of the proposed method are all superior to those of traditional feature selection methods.