<p>Oriented to the challenge on weak feature extraction of dynamic properties, a new remaining useful life (RUL) prediction method based on copula entropy (CE) and convolutional neural network - bidirectional gated recurrent unit - attention mechanism (CNN-BiGRU-AM) hybrid architecture was proposed. Firstly, CE was introduced to measure the nonlinear correlation between extracted features and RUL in order to achieve efficient screening and enhancement of multi-sensor features. Subsequently, a multi-scale one-dimensional CNN was addressed to extract time-frequency deep features for enhancing the expressive ability of health degradation indicators. Finally, the bidirectional temporal dependencies of degradation were built by the BiGRU network, and attention mechanism was adopted to grasp key degradation features to perceive the degradation process. XJTU-SY and FEMTO-ST public bearing datasets were applied to validate the proposed CE-CNN-BiGRU-AM method. Compared with GRU, long short-term memory (LSTM), transformer and temporal convolutional network (TCN), the results indicated that the proposed model could achieve higher prediction accuracy and robustness. This investigation demonstrated a potential application method of RUL prediction under complex working conditions, which provided reliable technical support for the intelligent operation and maintenance of facilities in intelligent manufacturing.</p>