Deep Gaussian Hidden Markov Network for Robust HRRP Sequence Modeling and Target Recognition

High Resolution Range Profile (HRRP) sequence contains both target scattering structure and motion state information, presenting a promising application in radar target recognition. During the observation of non-cooperative targets, the uncontrollable angular sampling intervals between adjacent HRRPs lead to a mismatch between training and testing conditions, degrading the target recognition performance of existing methods. To address this problem, an end-to-end deep Gaussian Hidden Markov network (GaMaNet) for robust HRRP sequence recognition is proposed. Specifically, we construct a deep Gaussian Hidden Markov model (DGHMM) for robust temporal modeling. To address the challenge of inconsistent angular variations, the DGHMM incorporates the latent state as a probability distribution rather than a deterministic point, introducing controlled stochasticity to improve model robustness to temporal variations. In addition, an adaptive diagonally-structured state-space sequence (S4D) model is designed as the inference network of DGHMM to approximate the latent state posterior probability, where its dynamic mechanism modifies the model parameters to accommodate irregular angular sampling intervals. Furthermore, a prior-guided fusion recognition module is proposed to fuse the temporal features extracted by DGHMM and achieve recognition. By deriving the variational lower bound as the optimization objective, we jointly optimize feature extraction and recognition in an end-to-end framework, enabling direct learning of discriminative temporal features for target recognition. Experimental results based on the simulated and measured dataset demonstrate that our method outperforms state-of-the-art HRRP recognition approaches in terms of recognition performance and robustness under different scenarios.