An early fault online detection model of rolling bearing based on deep attention convolutional autoencoder and multi-decision fusion under variable operation conditions

Abstract

A method based on model pre-training, fine-tuned transfer learning, and multi-decision fusion is proposed to achieve high-precision online early fault detection of rolling bearing under complex and variable operation conditions. Firstly, a novel attention mechanism is designed by combining the improved multi-head attention mechanism with rotary position embedding, and the Deep Attention Convolutional Autoencoder (DACAE) is constructed to extract bearing feature. Secondly, a self-supervised pre-training and fine-tuning strategy is used to features transfer, and combining data reconstruction error screening and enhancement algorithm to complete model optimization. Finally, various online detection results of algorithms are integrated, and multi decision voting mechanism is used to complete the detection task. Different bearing datasets are carried out, and the results show that the proposed method can effectively identify the early fault of rolling bearings, and reduce the false alarm rate under different working conditions, which has high robustness and reliability in the industry.