MSGFN: Multiscale Gated Fusion Network for UWB NLOS/LOS Classifications

Abstract

In the research of ultrawideband (UWB) indoor positioning, non-line-of-sight (NLOS) signals constitute the core bottleneck leading to the degradation of positioning accuracy. Existing NLOS/line-of-sight (LOS) classification methods suffer from three key limitations: they fail to fully exploit the complex-value characteristics of UWB signals and rely mostly on single-channel amplitude inputs, thus losing phase and positive-negative information; their generalization ability is insufficient, resulting in a significant decline in classification performance in complex unknown environments; and they have serious parameter redundancy, making it difficult to adapt to low-power edge computing scenarios. To address the above problems, this article proposes a Multiscale Gated Fusion Network (MSGFN). Taking the real and imaginary parts of the channel impulse response (CIR) as dual-channel input, MSGFN captures the scattered fluctuation characteristics of NLOS signals and the stable trend of LOS signals, respectively, through parallel multiscale convolution branches. It introduces a gated attention mechanism to dynamically optimize the feature weights in temporal–spatial and channel dimensions and suppress noise interference. Besides, an A-Sigmoid-Log (ASL) activation function and absolute maximum pooling adapted to UWB complex-value signals are designed, and efficient parameter utilization is achieved by combining lightweight attention modules and strategic batch normalization. Experimental results show that the proposed model achieves a classification accuracy of 88.63% in unknown environments, outperforming traditional amplitude-based input methods. Meanwhile, with a lightweight design of only 97 761 parameters, it demonstrates superior parameter efficiency compared with similar models. This work provides an effective solution for UWB signal classification in low-power edge computing scenarios.