Abstract: In this paper, based on the noise reduction capability and feature extraction capability of deep residual shrinkage networks with channel-wise thresholds (DRSNCW) module, combined with long short-term memory network (LSTM) and attention mechanism module, an end-to-end vibration signal-based bearing fault diagnosis model, named DRSNCW-LSTM, is proposed. In this model, LSTM module makes good use of the time-series characteristics of the signal to sufficiently extract the internal time-domain features of the vibration signal. Moreover, the introduction of attention mechanism enables the model to automatically extract important time-domain features for follow-up fault type recognition. The effectiveness of the proposed model is validated on the authoritative case Western Reserve University (CWRU) dataset, and experiments show that the proposed method can achieve more accurate bearing fault diagnosis than the state-of-the-art multi-scale convolutional neural network-LSTM (MCNN-LSTM) model without noise reduction processing. In particular, in the case of insufficient training data, the proposed method still achieves rather good bearing fault diagnosis with an average accuracy of 98.16%, which is an average improvement of 2.62% over the MCNN-LSTM.