Gearbox Fault Diagnosis using Novel Dilated CNN and Piecewise Loss Function under Unbalanced Data

Abstract

Planetary gearbox, as key components of rotating machinery, constantly works under harsh environment, which would cause various fault modes. Accurate and intelligent gearbox fault diagnosis methods help to improve the maintenance economy and reduce reliance on research experience. Convolutional neural network (CNN) has been fully applied in rotating equipment fault diagnosis due to the advantages of weight sharing and supervised learning ability. However, most researches focus on the balanced samples, and it is expensive to collect enough fault samples in practical situation. In order to improve the diagnostic performance under unbalanced samples, a new method based on CNN and piecewise cross entropy (PCE) is proposed. Firstly, a novel multi-scale dilated CNN (NMDCNN) is constructed using zigzag dilated convolution to enrich the coverage of the fields of view and avoid the loss of information. Secondly, PCE is utilized to balance the misclassification cost between health samples and fault samples. The experiment of planetary gearbox data proves the superiority and validity of proposed method.