In order to solve the problem of low prediction accuracy caused by insufficient data mining in the prediction task of remaining useful life (RUL) of rolling bearings in industrial production, a multi-channel fusion method for predicting the remaining life of rolling bearings was proposed. In this method, the original vibration signal is denoised and feature enhanced by complementary ensemble empirical mode decomposition (CEEMD) is taken as input. A three-channel network model was constructed, and three different neural networks were introduced: Temporal convolutional networks (TCN), convolutional long short-term memory network (ConvLSTM), and bidirectional gated recurrent unit neural network (Bi-GRU), which differentially extracts features from multiple dimensions such as time series, space, and receptive field. The multi-head attention mechanism (MA) is added on the basis of the structure to readjust the output weight of the network and accelerate the convergence speed of the model. Finally, a feature fusion output module was designed to predict the remaining life of rolling bearings. Experimental verification was carried out on two datasets and compared with the advanced models in other literatures. The results show that the proposed model can capture the bearing life degradation curve more accurately, and is better than the comparison model in a variety of evaluation indicators.