Abstract: Known as a compression algorithm, autoencoder is widely used for dimensional reduction and image denoising. It can be used in flow field identification and data processing as a reduced-order method. Moreover, massive labeled flow field data makes it promising to apply machine learning in fluid dynamics. Taking the flow around a cylinder as an example, the autoencoder model of the velocity field behind the cylinder is established to reduce the order and extract features from the original data. This model encodes the 1394 velocity components into 32-dimensional data, and trains them through self-supervised learning. For well-trained autoencoder models, the basis for evaluating them is to discuss whether the original flow field could be reconstructed. The application of the low dimensional encoded data is explored by correlating it with the flow field sensitive outputs, and a neural network for the regression of surface pressure of cylinder based on encoded data is established. It is verified that the result of the autoencoder has inherited the main information in the original velocity field. The root mean square error of the decoded velocity field compared with the original field is less than 0.02, and the root mean square error of the pressure coefficient regression network can be less than 0.1. The above results indicate that the autoencoder can be used in the future as a feature extraction and order reduction method of flow field.