Rolling bearing is a key component in a large number of rotating machines，and its remaining useful life (RUL) prediction is related to whether an equipment can operate safely and stably.In order to solve the current problem of low accuracy of RUL prediction，a prediction method combining temporal convolutional autoencoder (TCAE) and Informer (TCAE-Informer) network in frequency domain is proposed for RUL of rolling bearings.The method designs a TCAE for the frequency-domain signals of different time samples of rolling bearings，and adaptively extracts the depth features that better reflect the full-life degradation law of rolling bearings.An Informer network model is built，and with the help of its learning advantage in long sequence information，the mapping relationship between the depth features and the RUL of rolling bearings is effectively fitted to realize the RUL prediction function of rolling bearings.XJTU-SY bearing dataset is used to compare and validate the evaluation indexes of three RUL prediction results.The results show that the proposed method is able to achieve more accurate RUL prediction effect under different working conditions compared with many existing methods，which proves that it has superior RUL prediction ability and generalization ability.The anti-interference test is carried out for different methods.In the test，the proposed method shows better RUL prediction effect under different noise conditions，which proves that it has superior anti-interference ability of RUL prediction.