In the industrial applications of fault diagnosis, it is often difficult to obtain enough fault data, and the shortage of fault samples is always a thorny problem. The performance of deep learning method will be affected due to incomplete fault information under the condition of data shortage. To solve the above problem, an attention residual network is proposed with feature fusion at different depth levels under small sample condition. Firstly, this method takes deep residual network as the main work, inverse bottleneck structure is used to avoid network degradation and fault information loss. Secondly, through dense connections, the original features can be reused in different deep layers to enrich fault information. Finally, different attention mechanisms that focus on the different fault features are introduced to the residual network, as well as the local and global features are weighted and fused to improve the accuracy of fault diagnosis. The centrifugal pump test bench is used to carry out the small sample fault diagnosis experiment. Compared with CNN, ResNet, DRSN, the diagnosis accuracy of the model proposed in this paper is 0.9958, which is higher than other comparison methods, and verifies the feasibility of the method.