A deep ensemble dense convolutional neural network for rolling bearing fault diagnosis

Abstract

The accurate identification of rolling bearing fault based on unbalanced data has always been a challenge in the field of fault diagnosis. In some practical scenarios, since the machine is in the normal state most of the time, data imbalance will inevitably be encountered. For this purpose, a deep ensemble dense convolutional neural network (DEDCNN) is developed in this paper. First, dense convolutional neural network (DCNN) is used as a basic classifier to learn representative features. A striking characteristic of DCNN is that the learned features of each layer can be reused by all subsequent layers. Second, an adaptive boosting algorithm is used to integrate multiple DCNN classifiers to construct a DEDCNN. Third, the parameter transfer training mechanism of an improved adaptive boosting algorithm is designed and applied to the DEDCNN, in which the learned parameter information of the ith DCNN classifier is transferred to train the (i + 1)th DCNN classifier, in order to speed up the training process and improve diagnostic ability. Extensive data imbalance experiments are conducted to demonstrate the effectiveness of the proposed method. The results demonstrate that the proposed method exceeds the capabilities of the existing approaches.