In view of the low performance evaluation efficiency of neural architecture search (NAS) method in searching optimal structure,and the low diagnostic accuracy of hydraulic pump fault diagnosis due to poor model generalization ability,an improved Data-free NAS method was proposed.CAME optimizer and cosine annealing with warm restart algorithm were introduced to replace SGD optimizer and LambdaLR optimization algorithm,respectively,to improve the performance estimation and verification functions of Data-free NAS such as diagnostic accuracy and computational efficiency.Through the verified analysis of the hydraulic pump fault measured fault experiment,it can be seen that the improved method has obvious effectiveness and superiority over the original method；the CAME optimizer has obvious advantages in optimizing hyperparameters such as learning rate and momentum weight of the network,with accuracy and efficiency improved by 7.24% and 37.5%,respectively,with an accuracy up to 100%; the learning rate parameters can be improved by cosine annealing with warm restart algorithm,and the efficiency is improved by 81.25%.