DGT-former: a dual-channel time-frequency fusion model for remaining useful life prediction of rolling bearings

Significant breakthroughs have been achieved in the prediction of Remaining Useful Life (RUL) of rolling bearings based on deep learning. However, existing deep learning methods still have limitations. First, most of them adopt single-modal input or ignore the correlation between the different input data, which restricts the ability of extracting key degradation features; second, there are still challenges in modelling long-term temporal dependencies. To address these issues, this paper introduces a deep learning prediction framework called DGT-former, which incorporates a Dual-channel Gated Convolutional Network (DGConv), a Temporal Squeeze-and-Excitation module (TSE) and stacked Informer encoders. The network includes time- and frequency-domain channels, with the TSE attention module embedded in both to enhance representation of key degradation stages. Time-domain signals are converted to frequency-domain signals via Fast Fourier Transform (FFT), and both signals are input into the DGConv. The dual-channel features are concatenated and fed into the stacked Informer encoders. Moreover, applying its ProbSparse self-attention mechanism can effectively capture the long-term dependencies in sequences, and finally the RUL prediction results are obtained by mapping the fully connected layer. Experimental analysis on two rolling bearing datasets and comparison with other methods show that the proposed method outperforms them in accuracy and generalisation.