Due to the inability to collect sufficient fault samples of a certain fault type of gears and bearings failures, the data exhibits a long tail distribution, making it impossible to effectively construct a neural network diagnosis model. When the federal learning method is introduced to solve the above long tail problem, the feature information of the tail fault type sample cannot be effectively extracted. In view of the problems, this paper proposes an improved federated learning method. Firstly, the diagnosis model is retrained by using federal features to improve the fault feature extraction ability of tail samples. Secondly, the CBAM (convolutional block attention module) attention mechanism is introduced to improve the ResNet (residual network) network model in federated learning, boosting its ability and efficiency of extracting key local feature information of channel and space. Thirdly, the traditional convolution is replaced by asymmetric convolution to enhance the ability and efficiency of extracting asymmetric feature information of samples. Finally, the interval calibration algorithm is used to optimize the classification margin of the network model to obtain higher diagnostic accuracy and efficiency. The experimental analysis based on the measured fault samples of gears and bearings shows that the proposed improved federated learning method can effectively improve the average and highest accuracy, by 8.78% and 3.40%, respectively.