Traditional intent recognition algorithms of power lower-limb prosthesis often use machine learning algorithm classifiers, which require manual extraction in feature selection. In this paper, using deep learning algorithms in motion intent recognition research, the algorithm is improved on the basis of the traditional convolutional neural network, and is more suitable for the motion intent recognition based on the short-term behavior sample data, while suppressing the application of deep learning algorithms in motion intent recognition. The sliding window preprocessing is performed on the intent recognition data set. The purpose is to augment the data of the time series samples. Amplifying the target data set can make the training set more abundant and comprehensive, improve the accuracy of recognition. The improved convolutional neural network is used to perform feature learning and classification on the augmented data set. The experimental results show that the recognition rate of the improved algorithm under the 13 types of motion patterns reaches 93%.