Few-shot fault diagnosis of rolling bearing under variable working conditions based on ensemble meta-learning

Abstract

Accurate fault diagnosis of rolling bearing under variable working conditions can ensure that the rotating machinery run in a safety, reliable and efficient way. In this paper, we propose an ensemble meta-learning (EML) model for few-shot fault diagnosis of rolling bearing. The few-shot high-dimensional vibration signal of rolling bearing are transformed into grayscale image with different fault labels. In addition, episodic training framework is applied to make up different target datasets for meta-learning of few-shot fault samples Considering the discrepancy of vibration signal under different state parameters, ensemble learning framework is utilized to combine different meta-learning models, so as to achieve general accurate diagnostic result. Two publicly available rolling bearing datasets are used to testify the effectiveness of the proposed EML method. The results show that the proposed model can quickly update and iterate with few-shot high-dimensional samples, and demonstrate higher diagnostic accuracy and adaptability under variable working conditions when compared with other individual models.