面向 GNSS 拒止环境的极少参考点无线电指纹定位方法

To tackle the significant decline in positioning accuracy caused by sparse reference points in Channel State Information (CSI) fingerprinting, this paper presents a fingerprint augmentation model called L2C GAN (Location-to-CSI Generative Adversarial Network). The proposed method reconstructs fine-grained radio maps by generating virtual fingerprints based on the learned non-linear mapping between spatial locations and Radio Frequency (RF) features. The methodology consists of three key stages. First, a Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is employed to preprocess the raw CSI amplitude data. This approach adaptively identifies and retains dominant signal clusters while eliminating outliers arising from environmental noise and hardware instability, thereby ensuring the purity and spatial consistency of the training data. Second, to accommodate convolutional operations, high-dimensional CSI matrices are segmented along subcarrier indices and reshaped into single-channel 2D feature maps. Third, the L2C GAN architecture is constructed using a conditional generation mechanism. To address the low dimensionality of spatial coordinates, a Fourier-feature-based Positional Encoding mechanism is designed to map 2D coordinates into high-dimensional embedding vectors. Both the generator and discriminator integrate Self-Attention modules to capture long-range dependencies among subcarriers, enabling the model to learn global context. Experimental validation is conducted using the OpenCSI dataset in an extremely sparse setting with only 2% (88) real reference points. The results demonstrate that L2C GAN effectively enhances radio map reconstruction. The average positioning error decreases from 2. 01 m to 1. 67 m, representing a 16. 92% improvement over the baseline without augmentation. Furthermore, experiments with varying parameter settings show that augmenting with 400 virtual reference points yields the optimal positioning accuracy of 1. 67 m. Comparative results also indicate that the proposed method significantly outperforms state-of-the-art augmentation techniques, such as SSIM Aug (Structural Similarity-based Augmentation), AF DCGAN (Amplitude Feature Deep Convolutional GAN), and LESS (Adaptive Fingerprint-based Localization with Less Site Survey). Notably, the model exhibits robust performance even under limited hardware configurations, achieving a positioning error of 1. 92 m with only two receiving antennas, which surpasses the unaugmented four-antenna baseline.