@

초록

In light of ongoing advancements in smart manufacturing, there is a growing need for intelligent fault diagnosis methods that maintain reliability under noisy, high-variability operating conditions. Conventional feature selection strategies often struggle when data contain outliers or suboptimal feature subsets, limiting their diagnostic utility. This study introduces a density-based feature space optimization (DBFSO) framework that integrates feature selection with localized density estimation to enhance feature space separability and classifier efficiency. Using k-nearest neighbor density estimation, the method identifies and removes low-density feature vectors associated with noise or outlier behavior, thereby sharpening the feature space and improving class discriminability. Experiments using roll-to-roll (R2R) manufacturing data under mechanical disturbances demonstrate that DBFSO improves classification accuracy by up to 36-40% when suboptimal feature subsets are used and reduces training time by 60-71% due to reduced feature space volume. Even with already-optimized feature sets, DBFSO provides consistent performance gains and increased robustness against operational variability. Additional validation using a bearing fault dataset confirms that the framework generalizes across domains, yielding improved accuracy and significantly more compact, noise-resistant feature representations. These findings highlight DBFSO as an effective preprocessing strategy for intelligent fault diagnosis in intelligent manufacturing systems.

키워드