Multi-step remaining useful life prediction of rotating machinery using variational mode decomposition and attention-based sequence to sequence model

Abstract

Accurate prediction of the remaining useful life (RUL) of rotating machinery is critical for predictive maintenance and operational safety in industries. A novel multi-step RUL prediction framework for rotating machinery is proposed in this paper, integrating variational mode decomposition (VMD) for noise reduction and an attention-based sequence to sequence (Seq2Seq) model that employs multi-head self-attention for end-to-end prediction from multi-state parameters to RUL. The VMD algorithm adaptively decomposes raw vibration signals into intrinsic mode functions (IMFs), effectively isolating noise and enhancing feature extraction by selecting dominant modes based on energy and correlation thresholds. The denoised features are then fed into a Seq2Seq architecture, in which the encoder captures temporal dependencies in the sensor data, and the decoder, augmented with multi-head self-attention, dynamically weights salient features across time steps to improve long-term forecasting accuracy. The proposed model demonstrates superior performance on both the C-MAPSS aero-engine simulation dataset and the XJTU-SY experimental rolling bearing degradation dataset, achieving mean absolute error (MAE) reductions ranging from 17.5% to 39.7% across different prediction horizons and subsets compared to the best-performing baseline. These results highlight the enhanced accuracy and robustness of the proposed approach for multi-step RUL prediction of rotating machinery, providing a reliable solution for industrial applications.