LFT-Net: A Physics-Guided Transformer for Temporal Fusion of Low-Dimensional Radar Target Features in Confusable Target Recognition

To address the challenge of distinguishing confusable targets with significantly overlapping features in radar low-altitude surveillance, this paper introduces a Low-dimensional Feature Temporal Fusion Network (LFT-Net) based on the Transformer architecture. Current methods often face constraints stemming from the ambiguity of static data or the gradient vanishing phenomenon inherent in Recurrent Neural Networks (RNNs), making it difficult to see subtle distinctions in temporal evolution. LFT-Net addresses these limitations through three main ways: First, orthogonal feature decoupling employs a dual-branch projection strategy to explicitly separate static physical attributes from dynamic micro-motion patterns, fundamentally reducing interference from static features. Second, the physical scale-aligned Transformer encoder incorporates a dedicated alignment mechanism to suppress scale drift across heterogeneous feature spaces. While preserving physical semantic consistency, this encoder utilizes self-attention to precisely capture long-range periodic dependencies in micro-motion signals. Third, multi-scale temporal pooling operates in parallel to extract both continuous steady-state characteristics and transient impulse features from target signals. A series of experiments containing nine categories of typical targets reveal that LFT-Net attains a recognition accuracy of 84.7%. This represents a notable improvement of 13.9%, 4.9%, and 4.6% compared to the Long Short-Term Memory (LSTM) network, the state-of-the-art Transformer method MPT-SFANet, and the recent dual-branch method DPFFN, respectively. Moreover, LFT-Net offers a favorable complexity-performance trade-off, supporting its use in resource-limited radar systems.