Bearing Fault Diagnosis for Electrical Machines based on Multi-sensor Multi-level Fusion and Convolutional Neural Network

Abstract

Accuracy and robustness are two important aspects for bearing fault diagnosis under various conditions. As an effective approach, the fusion of multi-sensor data could deeply extract the fault information. However, current fusion networks mostly focus on fusing the information at the single level. The relevance and complementarity for the data collected from multi sensors cross different levels are ignored. To overcome the limited fault information in single level, a novel convolutional neural network model with multi-sensor multilevel fusion (MSLF-CNN) is proposed. MSLF-CNN is comprised of four feature extraction branches and a decision fusion branch. It first automatically extracts the fault features based on the data collected by a single sensor. Next, the extracted features from multi sensors at the data level, feature level, and decision level are fused to improve the data interactivity. Further, the coupled features are obtained. Last, the verification experiments are provided to show the excellence of MSLF-CNN, namely, the diagnosis accuracy has been extremely improved by the fusion of vibration-torque signals.