A novel two-dimensional progressive domain adaptation framework for cross-domain remaining useful life prediction

To achieve the Remaining Useful Life (RUL) prediction of mechanical equipment working in cross-domain scenarios, a Two-dimensional Progressive Domain Adaptation framework (TDPDA) is proposed. This framework develops a progressive spatio-temporal feature alignment strategy combining Maximum Mean Discrepancy (MMD) and adversarial learning to enhance the resistance to domain transfer. A feature extraction module based on two-dimensional feature-mixing MLP and multiscale sequence decomposition is tailored to adapt to this strategy, aiming to better extract effective degradation information from origin data. Finally, a stacked dilated convolution blocks-based RUL prediction module is constructed to effectively predict the RUL of equipment in cross-domain scenarios. The effectiveness of TDPDA is validated on the aircraft engine dataset and the IEEE PHM Challenge 2012 dataset. Experimental results demonstrate that the proposed method outperforms existing approaches in handling cross-domain RUL prediction tasks. •A novel Two-Dimensional Progressive Domain Adaptation Framework (TDPDA) is proposed, comprising a feature extraction module, an RUL prediction module, and a domain discriminator, to achieve effective cross-domain RUL prediction.•Developed a two-dimensional progressive spatio-temporal feature alignment strategy and a feature extraction module, and introduced an adversarial learning mechanism to enhance the TDPDA’s ability to learn domain-invariant features and minimize domain shift.•A dilated convolution block is constructed, and based on this, an RUL prediction module is built to achieve effective cross-domain prediction.•The results demonstrate its superiority in RUL prediction tasks under cross-working conditions.