Online detection for bearing incipient fault based on deep transfer learning.

• We propose a new deep transfer learning model to enhance common feature representation of bearings. • We propose a new health indicator with better description of bearing degradation trend. • We propose a new transfer learning-based online detection method of bearing incipient fault. In order to achieve effective online detection of bearing incipient fault, it's necessary to adaptively extract representative features to incipient fault. However, the traditional feature extraction methods are less adaptive to online detection problem. In this paper, a new online detection method of incipient fault based on deep transfer learning is proposed. In the offline stage, a three-channel data set is first built by merging time/frequency/time-frequency domain information. Second, a new transfer learning model is constructed on auxiliary bearings data from a pre-trained VGG-16 model built on image data. Through a fine-tuning process, common deep features are extracted and the detection model is trained by using support vector data description. In the online stage, deep transfer features of target bearing are directly extracted, and final detection results are obtained. The experimental results on the bearing dataset of IEEE PHM Challenge 2012 show the comparative performance of the proposed method. [ABSTRACT FROM AUTHOR]