An Investigation Into Fault Diagnosis of Planetary Gearboxes Using A Bispectrum Convolutional Neural Network

Abstract

To improve the efficiency and accuracy of fault diagnostics of planetary gearboxes, an intelligent diagnosis approach is proposed based on deep convolutional neural networks (CNNs) and vibration bispectrum (BSP). Rather than using raw vibration signals, BSP is appreciated as the input for the CNN models (denoted as BSP-CNN) because the BSP allows nonlinear feature enhancement and noise reduction. In addition, transfer learning (TL) is accompanied to address the challenges of CNN difficulties. The proposed BSP-CNN is verified firstly to diagnose a number of common faults including gear states: normal, tooth wear, tooth root crack, tooth breakage and missing tooth, achieving an accuracy of 97.36% in identifying different faults. Then, its TL capability is evaluated based on the sun gear faults datasets. The classification accuracy of the planet gear faults is over 95.1%. After the transfer learning, the classification accuracy of the sun gear fault is still higher than 97.9%, and the computational time consumed by proposed method is also less compared to other diagnosis methods. This article has twofold contributions: first, the development of a BSP-based CNN model for fault diagnosis; andsecond, the extensive evaluation of CNN-TL methods for monitoring and diagnosing planetary gearboxes.