Aiming at the problems that the existing rolling bearing remaining useful life (RUL) prediction methods have a single feature extraction capability and cannot fully utilize the spatio-temporal information embedded in the data, a rolling bearing remaining useful life prediction method based on CEEMDAN combined with TCN-BiLSTM-MultiHead-SelAttention is proposed. Simultaneous extraction of spatial feature information hidden in time series-based features using a dual network of Temporal Convolutional Neural Network - Bidirectional Long Short Time Memory Network (i.e. TCN- BiLSTM). The degradation feature information is enhanced through multi-head self-attention weighting, followed by the prediction of the remaining useful life (RUL) of bearings through fully connected layers. The proposed network fully utilizes bearing features extracted from the frequency domain and one-dimensional time series. When combined with signal processing techniques, these features enable a multi-dimensional characterization of the bearing degradation trend. Based on the present methods, the RUL prediction intervals can be efficiently quantified without relying on physical or statistical a priori information about the bearings. The superiority of the proposed method is verified by an example study of rolling bearings. The results show that the method proposed has a higher prediction accuracy and is more inclined to over-prediction, which is conducive to carrying out predictive maintenance.