ScholarWorks@경상국립대학교

초록

Bearing faults, critical failures in rotating machinery, can result in substantial downtime and maintenance costs, often manifesting as abnormal vibration patterns. However, raw vibration signals are typically high-dimensional and noisy, making direct fault classification challenging. This study proposes a high-precision fault diagnosis framework that integrates feature selection based on ANOVA(Analysis of Variance) F-statistics and Critical Sampling Set (CSS) methodology to identify informative features and samples. CSS is constructed from post-processed data using Kriging (KRG) and Latin Hypercube Sampling (LHS), and the selected samples are independently fed into Gated Recurrent Unit (GRU) and EfficientNet classifiers. Multi-domain features are extracted from vibration signals across the time domain (RMS, kurtosis, skewness), frequency domain (Fast Fourier Transform (FFT), power spectrum), and time-frequency domain (wavelet transform, empirical mode decomposition, Hilbert transform). Sample importance is quantitatively evaluated using multiple criteria, including Mahalanobis distance, Kullback-Leibler divergence, silhouette coefficient, Pattara score, and Fisher criterion. ANOVA F-statistics are applied to assess the discriminative power of each feature and guide feature selection. The proposed framework is validated on the MAFAULDA (Machine Faults Detection by Learning from Acoustics and Vibrations) benchmark dataset, which includes normal, cage, ball, and outer race fault classes. Experimental results demonstrate improved classification accuracy and computational efficiency compared to baseline approaches. The features selected by ANOVA exhibit strong alignment with fault-related frequency components and statistical properties, enhancing both interpretability and model performance. This work empirically demonstrates the importance of effective sample and feature selection in vibration-based fault diagnosis and offers a scalable foundation for expanding to high-dimensional sensor data applications in future research.

키워드