Collaborative Inversion of 3D Parameters for Precessing Cones via Cross‐View Dual‐Stream BiGRU and Geometric Decoupling

ABSTRACT

Traditional radar micro‐Doppler (m‐D) feature extraction suffers from single‐station occlusion and noise, while end‐to‐end (E2E) deep learning lacks physical interpretability and generalisation. This paper proposes a collaborative 3D parameter inversion method for processing cones using a cross‐view dual‐stream BiGRU (CV‐DS‐BiGRU) network and spatial geometric decoupling. First, the network employs a cross‐attention mechanism to repair fragmented trajectories under low signal‐to‐noise ratios. Second, a differential ratio elimination method algebraically decouples the nonlinear system, yielding closed‐form solutions for 3D structural and kinematic parameters without high‐dimensional searches. Simulations demonstrate that the proposed cascaded architecture maintains robust accuracy under noise. Furthermore, it overcomes the discretisation effects of pure E2E models, achieving superior zero‐shot physical generalisation in out‐of‐distribution (OOD) scenarios.