Since the existing data-driven remaining useful life (RUL) prediction methods of rolling bearings still need a lot of prior knowledge to extract features, construct health indicators, and set fault thresholds, a direct RUL prediction method based on time convolution network (TCN) with multi-head attention mechanism is proposed. In this method, the short-time Fourier transform (STFT) of the original vibration signal is used as the input of the stack noise reduction automatic encoder (SDAE) to get the depth feature representation, and then input it to the attention TCN for RUL prediction. Finally, an example is verified in the rolling bearing data set of PRONOSTIA. The results show that the prediction error-index MAE and MAPE of this method are 53.92% and 46.13% lower than those of the other four methods, respectively, and the score index is 52.98% higher than that of these methods.