Traditional deep networks used for radar High-Resolution Range Profile (HRRP) target recognition usually ignore the inherent characteristics of the target

which results in the limited capability to learn effective features for classification task. To address this issue

a novel nonlinear feature learning method

called Robust Variational Auto-Encoder model (RVAE) is proposed. According to the stable physical properties of the average profile in each HRRP frame without migration through resolution cell

RVAE is developed based on variational auto-encoder

and such model is able to not only explore the latent representations of HRRP but also reserve structure characteristics of the HRRP frame. We use the measured HRRP data to show the effectiveness and efficiency of our algorithm.