Aiming at the problems that the existing convolutional neural network cannot fully extract the correlation features be? tween rolling bearing time domain signals， the large number of samples required for model training and the insufficient generaliza? tion， A new method for diagnosing multi-condition faults of rolling bearings based on an enhanced convolutional neural network model is proposed. The length of the bearing single?revolution fault characteristic signal is calculated according to the rolling bear? ing speed and sampling frequency， then the complete information of the single?revolution time domain signal is encoded by Gramian Angular Difference Field coding technology to generate the corresponding feature image， enabling the neural network can visually learn the time domain signal correlation features. The 7×7 deep convolutional layer of the ConvNeXt model is recon? structed by using the asymmetric convolution in the ACNet network model： that is， two 3×3， one 1×3 and one 3×1 asymmetric small convolution kernel are used to reconstruct the 7×7 convolutional layer in the form of a multi-branch structure combination，which enhances the feature extraction efficiency of the ConvNeXt model. The data augmentation module and learning rate decay strategy of the ConvNeXt model are improved to raise the generalization of the ConvNeX model under small-sample training， to build an enhanced deep convolutional neural network model IConvNeXt. Different fault diameters of Case Western Reserve Univer? sity， composite rolling bearing faults of Southeast University and variable speed bearing fault data sets of Ottawa， Canada are used for experimental verification， the results show that the proposed IConvNeXt model achieves a fault diagnosis rate of 100% for dif? ferent fault diameters and composite faults of rolling bearings， and a fault diagnosis rate of 99.63% for variable speed bearings. The proposed method is experimentally compared with RP+ResNet， RP+ IConvNeXt， time-frequency graph+DCNN， MLCNN LSTM， MTF+ IConvNeXt and other methods， the results were condicted to validate that the fault diagnosis effect of the pro? posed model is better than that of other methods under less sample training and has strong generalization performance.