Intelligent fault diagnosis for high-speed bearing towards unbalanced samples via convolutional weight adaptive network

Abstract

High-speed bearings are often required to undertake long-term operation under unsatisfactory scenarios such as heavy load condition, and the raw vibration signals from the high-speed bearings are usually acquired with strong instability. In addition, the fault samples are unbalanced which far less than the healthy samples. Conventional intelligent fault diagnosis methods are subject to skew large samples, leading to the degradation of diagnosis performance. For this purpose, a convolutional weight adaptive network is proposed in this paper. Firstly, a multi-scale feature extraction network is constructed for extracting multi-scale fault features and excavating useful hidden information. Afterwards, the feature weight self-adaptive module is developed to dynamically fuse multi-scale fault features to heighten the contribution of the high-related features and to diminish the effect of the non-related features. Finally, the modified Focal loss is designed to re-balance the cost of various types of small fault samples and large healthy samples during the training process, making the model pay more attention to the samples which are few and easily confused. The experimental analysis by using vibration data of high-speed bearing demonstrates the feasibility and effectiveness of the proposed intelligent fault diagnosis method under unbalanced samples.