Abstract: In response to the problem of low accuracy of fault diagnosis caused by the complexity of rolling bearing vibration signals of helicopter auto-tilter and the difficulty of extracting tiny features of bearing fault signals by traditional convolutional neural network, an improved method of LeNet-5 network is proposed. First, a new feature extraction layer is added to the LeNet-5 network and a parallel feature extraction framework is formed, aiming to enhance the extracting minute feature’s ability of network and alleviate the problem of low accuracy of helicopter fault diagnosis. Moreover, to improve the stability of the model and accelerate its convergence, a dropout layer and an adaptive parameter algorithm are adopted. Finally, experiments are carried out with the dataset of research group and the public dataset of Western Reserve University. The experiment results demonstrate that the improved LeNet-5 network has a higher test accuracy compared with the original LeNet-5 network model, with 99.6% in the research group dataset and 100% in Western Reserve University dataset, which verifies that the model has a higher accuracy rate in the fault diagnosis of rolling bearings.