A neural network with nuisance attribute projection: a novel method for bearing fault diagnosis under variable speed

Abstract

Rolling bearings are a widely used component in rotating machinery, and deterioration of their performance may lead to machine breakdown and accidents. However, because rolling bearings often work under time-varying working conditions it is a great challenge to accurately separate fault features from working condition features. This paper presents a new neural network framework for variable speed fault diagnosis. The proposed neural network can eliminate the influence of condition information during training. Firstly, samples of the same category under different working conditions are distinguished. Secondly, a projection matrix is designed at the last layer of the neural network to project the features extracted through forward propagation to a hyperplane of the same latitude, so that the fault features are more obvious. Finally, a network model that can distinguish fault features from working conditions is obtained through back propagation training. In addition, we propose a multi-loss function, which can reduce the within-class distance and enlarge the between-class distance. The method is validated using bearing datasets under time-varying rotation speeds. The results show that our method has higher accuracy with a simple structure, and is a great improvement on the traditional method of bearing fault diagnosis.