Robust Convolutional Neural Network Resistant to Overfitting for Classifying Bearing Fault Modes

초록

Bearings are precise components used in mechanical systems to prevent friction and support loads arising from rotational movements along an axis. Bearings primarily fail due to defects such as inner ring faults, outer ring faults, and rolling element faults. Unlike traditional vibration signal analysis, deep learning techniques combining convolutional neural networks with short-time Fourier transforms have been known to classify even minor defects. However, previous studies have not been conducted under conditions that strongly cause overfitting. This paper mainly discusses about how to prevent overfitting in convolutional neural networks for bearing fault mode classification. We constructed a dataset with simultaneously measured axial and radial acceleration signals during bearing operation. Simple signal analysis showed that two different signals are physically very similar, confirming that differences in training and validation performance due to changing measurement directions indicate overfitting. In this paper, we prevented overfitting and improved performance using a convolutional neural network based on auto-encoders.

키워드