When the adversarial training with learnable attack strategy network model(LAS-AT) is introduced into the rolling bearing fault diagnosis,the model has some problems such as insufficient feature extraction ability,low computational efficiency and low diagnostic accuracy.In view of these problems,an improved LAS-AT(CDEW-LAS-AT) was proposed.In order to improve the feature extraction ability of the LAS-AT,the efficient channel attention（ECA） was embedded into the feature extraction function module of the network structure，and the rolling bearing fault experiment was made to verify and analyze.The results show that the ECA attention mechanism has effectively improved the model′s feature extraction ability,convergence,stability and accuracy,with accuracy increases by 1459% and loss value decreases.In order to improve the computational efficiency of the model,the ordinary convolution in WideResNet was replaced by the deep separable convolution,and the weight parameters of the model were optimized.The results show that the deep separable convolution effectively reduces the number of parameters,computation and computational delay of the network,reduces the scale of the network structure,and improves the real-time and accuracy of network diagnosis.Finally,in order to improve the diagnostic accuracy again,the cuckoo search algorithm (CS) was used to optimize the parameters in the strategy and the disturbance parameters of the generated adversarial example.The convergence,stability and accuracy of the model are further improved by the cuckoo search optimization algorithm,and its accuracy reaches 94.52%，with an overall improvement of 37.88%.