TY - JOUR
T1 - SAHG-WMRNet
T2 - A Structure-Aware Hierarchical Graph Network for Multifunction Radar Work Mode Recognition
AU - Zhou, Zixiang
AU - Liu, Chuyi
AU - Lang, Ping
AU - Dong, Jian
AU - Gao, Meijing
AU - Fu, Xiongjun
N1 - Publisher Copyright:
© 1965-2011 IEEE.
PY - 2026
Y1 - 2026
N2 - The pulse sequence-based multifunction radar (MFR) work mode recognition (WMR) is crucial for cognitive electronic warfare. However, the missing and spurious pulses, measurement errors, and limited labeled samples pose severe challenges to the WMR task. SAHG-WMRNet is proposed based on self-supervised pulse group segmentation (PGS) and semisupervised adaptive hierarchical graph nodes recognition. First, pulse sequence parameters are encoded with multiple feature maps in the high interclass separable latent space through self-supervised multiscale encoder, and the differences among feature maps are utilized to automatically segment the pulse sequence into multiple pulse groups with their individual modulation types via segmentation manner. Then, the initial pulse-level graph is constructed based on feature maps via K-nearest neighbor algorithm. Finally, a pulse and pulse group-level hierarchical graph is constructed through adaptive coarsening and refining based on the PGS results. By using hierarchical graph, the WMR task can be implemented in a semisupervised manner, utilizing a small amount of labeled data and explicit geometric relationships. The simulation results in one-shot and five-shot scenarios show that the proposed method can achieve better recognition accuracy with limited labeled samples under nonideal scenarios, compared to some existing methods.
AB - The pulse sequence-based multifunction radar (MFR) work mode recognition (WMR) is crucial for cognitive electronic warfare. However, the missing and spurious pulses, measurement errors, and limited labeled samples pose severe challenges to the WMR task. SAHG-WMRNet is proposed based on self-supervised pulse group segmentation (PGS) and semisupervised adaptive hierarchical graph nodes recognition. First, pulse sequence parameters are encoded with multiple feature maps in the high interclass separable latent space through self-supervised multiscale encoder, and the differences among feature maps are utilized to automatically segment the pulse sequence into multiple pulse groups with their individual modulation types via segmentation manner. Then, the initial pulse-level graph is constructed based on feature maps via K-nearest neighbor algorithm. Finally, a pulse and pulse group-level hierarchical graph is constructed through adaptive coarsening and refining based on the PGS results. By using hierarchical graph, the WMR task can be implemented in a semisupervised manner, utilizing a small amount of labeled data and explicit geometric relationships. The simulation results in one-shot and five-shot scenarios show that the proposed method can achieve better recognition accuracy with limited labeled samples under nonideal scenarios, compared to some existing methods.
KW - Hierarchical graph representation
KW - multifunction radar (MFR)
KW - self-supervised segmentation
KW - semisupervised learning
KW - work mode recognition (WMR)
UR - https://www.scopus.com/pages/publications/105039297677
U2 - 10.1109/TAES.2026.3693210
DO - 10.1109/TAES.2026.3693210
M3 - Article
AN - SCOPUS:105039297677
SN - 0018-9251
VL - 62
SP - 11879
EP - 11890
JO - IEEE Transactions on Aerospace and Electronic Systems
JF - IEEE Transactions on Aerospace and Electronic Systems
ER -