Bearing Remaining Useful Life Prediction Based on CICAE and ResConv1D-LSTM

Bearings play a critical role in industrial machinery, and accurate predictions of the remaining useful life (RUL) of bearings are essential for ensuring the safe operation of rotating equipment. However, traditional autoencoders (AEs) and CNN-based approaches tend to assign equal weights to all feature channels for degradation feature extraction, which inevitably leads to the critical degradation characteristics being obscured by noise or characteristics of other channels. Moreover, the additive operation of traditional residual connection inherently limits the interblock coordination, leading to redundant feature learning and constrained hierarchical representation capacity. To address these limitations, a novel two-stage deep learning framework is proposed. In the first stage, a channel-information-constrained AE (CICAE) develops an information entropy-based constraint to focus on the degradation-relevant channels while suppressing irrelevant ones. In the second stage, a ResConv1D-LSTM network integrates multiscale convolutional blocks for progressive residual learning with several long short-term memory (LSTM) layers for temporal modeling, so as to comprehensively characterize the degradation trend, and then, multiple fully connected layers are designed for RUL prediction. Experimental results demonstrate that the proposed bearing RUL prediction method outperforms the existing approaches and is applicable in real-world scenarios.