170 / 2023-11-21 11:27:43

A novel Transformer model based on dynamic convolution and ProbSparse self-attention for RUL prediction of bearings

Remaining useful life prediction, Transformer, Dynamic convolution, ProbSparse self-attention mechanism, Rolling bearing

Special Sessions > SS01 - Federated and Transfer Learning for Prognostics and Health Management of Complex Systems

The health of rolling bearings is related to the normal operation of rotating machinery. Accurately predicting the remaining useful life (RUL) of bearings is the key to avoiding the failure of bearings and system. In this paper, a new dynamic convolution Transformer model with ProbSparse self-attention mechanisms is proposed to extract advanced degradation characteristics from complicated vibration signal for accurately RUL of bearing, which is called dynamic ProbSparse self-attention Transformer (DPT) model. First, the cumulative amplitudes of frequency domain are computed as the network inputs. Then, a dynamic convolutional layer is constructed in Transformer architecture with ProbSparse self-attention mechanism to enhance the long-distance feature capturing capability. Finally, the high-level representation is fed back to a linear regression network for estimating bearing’s RUL. The proposed DPT model is validated through using two datasets. Experimental results present that the proposed DPT network surpasses the other models in RUL predicting, which possesses higher precision and computational efficiency.