Remaining useful life prediction of rolling bearings based on TET and DSRNet-AttBiLSTM

Here, to correctly extract degradation characteristics of bearings and make effective predictions for remaining useful life ( RUL ) of rolling bearings, a rolling bearing RUL prediction method based on transient extracting transform (TET) and DSRNet-AttBiLSTM was proposed. Firstly, after segmenting and recombining the original vibration signal, TET was performed for the recomposed signal to obtain its time-frequency map. Bilinear interpolation was used to reduce the dimensionality of time-frequency map, and channel splicing was performed for the reduced time-frequency map to obtain bearing time-frequency visualized features. Secondly, to correctly and effectively extract degradation features of rolling bearing, SConv and DConv basic modules containing depthwise separable convolution and spatial channel attention were constructed. Based on them, DSRNet was established to extract bearing degradation features in both spatial and channel dimensions. Once more, to make the bidirectional long-short term memory (BiLSTM) network pay more attention to input features with more important information during learning, an attention layer was constructed at the feature input end, and combined with BiLSTM to form an AttBiLSTM prediction module for HI calculation. Finally, the linear regression fitting was used to predict RUL of rolling bearings. The experimental results on PHM2012 dataset and XJTU-SY dataset showed that the proposed method can effectively predict RUL of rolling bearings. © 2024 Chinese Vibration Engineering Society. All rights reserved.