关键词: 健康指标, 早期故障检测, 深度可分离卷积, Savitzky-Golay滤波器, 自编码器

Abstract: A method of unsupervised bearing health indicator and early fault detection was proposed. A deep separable convolutional auto-encoder model was designed to effectively extract bearing state features， where the outputs of the encoder were used as bearing state features. Then， the Bray-Curtis distance was used to calculate the distance between the degenerated state features and the healthy state features as the bearing state health indicator（BC-HI）. An early fault detection method combined with Savitzky-Golay filter was proposed based on BC-HI. The abnormal threshold was obtained according to the trend of health indicators to judge the occurrence of early faults. In order to verify the effectiveness and generalization ability of the proposed method， the experiments were carried out on datasets of bearing accelerated life tests. The experimental results show that the health indicator proposed may reflect the degradation trend of bearings， and is sensitive to early faults， and has strong generalization ability. Compared with the methods such as isolated forest and support vector machine， the first fault detection time is earlier and the false alarm rate is lower， so it has certain application values. 

Key words:  , health indicator； early fault detection； depth separable convolution； Savitzky-Golay filter； auto-encoder