Abstract: Reduced-order technology has been applied to the analysis on complex flow fields. Based on a fully connected autoencoder network, a multi-channel autoencoder and a hierarchical autoencoder were designed to achieve reduced-order modeling and flow feature extraction of unsteady flow fields in a vaneless diffuser of the centrifugal compressor. Then the results were compared with the flow field modes obtained via conventional proper orthogonal decomposition (POD). Results show that the multi-channel decomposition autoencoder and hierarchical autoencoder techniques address the issue of conventional autoencoder methods being unable to perform reduced-order flow field visualization, providing an intuitive representation of the evolution process of unstable flow fields in the vaneless diffuser and offering a new tool for analyzing unsteady flow fields.