Aiming at the problems of large redundancy of time domain signals,low diagnostic accuracy and poor generalization ability of rolling bearing fault diagnosis under strong noise environment,a new fault diagnosis method based on multi-scale evolution of multi-joint attention mechanism and multi-residual convolution block was proposed.Transition blocks of wide and narrow kernel convolutions and multiple joint attention mechanisms were used to supplement features of deep convolutions,to reduce feature loss and ensure the quality of feature maps.Through the allocation of channel and spatial attention weights,different weight parameters were provided for the convolution layer for adaptive feature refinement.The proposed method was tested and analyzed in the bearing data sets of Case Western Reserve University and Southeast University respectively.The results show that the accuracy of conventional classification is more than 99.75%,and it can also reach more than 98.5% in the case of strong noise interference.The average classification accuracy of classification in variable working conditions is more than 90%.The proposed diagnosis method has good fault diagnosis effect,generalization ability and anti noise performance.