An ensemble framework based on diversity enhancement and difference indicator for fault diagnosis of rotating components with unknown modes

Abstract

Rotating components often run continuously at high speed under heavy load, resulting in variable failure modes. Because a priori not-considered fault may occur during the actual operation, it is significant to develop methods that can identify both pre-known types of faults and unknown types of faults. In this study, an ensemble framework based on partial dense convolutional neural networks with multiple diversity enhancement strategies (MDE PD-CNN ensemble) is proposed. Firstly, PD-CNN is employed to improve the generalization ability of the base model. Variety PD-CNN are constructed under multiple diversity enhancement strategies. Furthermore, differences in the output of samples on different base models are measured to detect unknown faults. Both known and unknown faults can be accurately diagnosed based on the ensemble procedure with the difference indicator. Experiments on bearing and gear datasets are conducted to demonstrate the superiority of the proposed ensemble framework.