A novel integration framework for degradation-state prediction via transformer model with autonomous optimizing mechanism

Accurate degradation-state prediction has been a prerequisite for formulating equipment maintenance strategies. Meanwhile, as the increasing timeliness requirement of the maintenance, long-sequence prediction is of great significance. However, accurate long-sequence prediction is still challenging for existing methods. To address the problem, this paper proposed a data-driven framework for state prediction of the degradation process. The framework consists of a multi-output encoder, a health indicator (HI) constructor, and a state predictor. Firstly, with the deployment of multiple activation functions, the encoder can extract multiple non-linear features simultaneously. Meanwhile, due to the limited prior knowledge under practical conditions, the encoder is designed to extract features from high-dimension space directly. Then, based on the auto-encoder mechanism, the extracted features are fused into a HI, which can indicate the degradation state of the object. Finally, a novel autonomous optimizing Transformer (AOT) combining the recurrent mechanism and the position embedding algorithm is proposed to predict the HI using the extracted feature sequences. The effectiveness of the proposed framework is verified through two whole-lifetime bearing datasets. Compared with some state-of-the-art degradation-state prediction approaches, the proposed method performs higher prediction accuracy.