In view of the complex working conditions， large load and low signal-to-noise ratio （SNR） of vibration signal of rolling bearing， a weak fault diagnosis model based on attention recurrence （AR） is proposed to construct digital capsule and fuse with capsule network （Caps）. In this model， the bidirectional long-short time memory neural network （Bi-LSTM） is introduced to extract the temporal features of the time-frequency diagram， and establish nonlinear association between capsules. Secondly， we use the attention recurrence mechanism to construct digital capsules to improve the influence of energy intensity changes in different time and frequency bands of time-frequency diagram. Then the attention recurrence and digital capsules constructed by dynamic routing mechanism are fused by 3D convolution adaptively to realize the diversity of feature extraction. Finally， the softmax classifier is used to map the fusion features to the output layer， to achieve the fault diagnosis of rolling bearing in high noise environment. The results show that this method has higher diagnostic accuracy than other diagnostic models on weak fault signals with small samples and low signal-to-noise ratio. Moreover， the model can effectively reduce the over fitting problem and strong generalization ability.