A new intelligent bearing fault diagnosis model based on triplet network and SVM

Abstract Separating sensitive characteristic signals from original vibration data is an important challenge for rolling bearing fault diagnosis. Because it is difficult to obtain large number of damaged bearings, Rolling bearing fault datasets are often small sample datasets. For the classification of small sample rolling bearing fault datasets, we propose a coupling vibration data classification method based on triplet embedding. The method is divided into two steps: feature extraction and fault identification. First, build a triple embedding based on the CNN model to reduce the original vibration signal, and then train the SVM model for classification. Compared with traditional features and autoencoder, triplet network can learn the differences between samples. Make classification training easier and more accurate. We have evaluated the performance of this method through two bearing experiment examples. The experimental results show that this method is superior to stacked autoencoder, stacked denoising autoencoder and CNN.