19th Asia-Pacific Vibration Conference (APVC2021)

In order to make full use of the effective information contained in unlabeled samples and improve the accuracy of fault diagnosis, a semi-supervised fault diagnosis method (CNN-Tri) based on improved convolutional neural network (CNN) and tri training method (Tri-training) is proposed. The method takes the time domain map of the fault vibration signal as the input, utilizes CNN to extract the features of the time domain map, obtains the one-dimensional features of the vibration signal, and trains the improved Tri-training to get three classifiers. Finally, the reliable unlabeled data and pseudo tags are selected by using the trained classifier to join the training set of CNN, and the final CNN model and three classifiers are obtained by repeated training. The experimental results show that the proposed method has good diagnostic performance in the case of labeled small samples.
 

Convolutional neural network (CNN),Tri-Training,Machine learning classifier,Loss function