Research on UWB NLOS Identification and Mitigation Methods Based on Gramian Angular Field and Hybrid Fusion Neural Network

Ultra Wideband (UWB) signals offer strong penetration capabilities and high transmission rates, playing an important role in precise indoor positioning. However, the complex and dynamic nature of indoor environments often leads to Non-Line-of-Sight (NLOS) propagation and severe signal attenuation, resulting in significant ranging and positioning errors. Currently, Convolutional Neural Network (CNN) based approaches have made some progress in NLOS identification tasks. Nevertheless, traditional CNNs suffer from performance bottlenecks owing to their lack of the temporal feature modeling capabilities and limited representational diversity. To address this issue, this paper introduces the concept of feature fusion and proposes a multi-feature fusion CNN model that incorporates Gramian Angular Field (GAF) encoding. By applying GAF encoding, it converts a one-dimensional channel impulse response (CIR) signal into a two-dimensional image, capturing the temporal correlations within the signals. During model training, different CNN architectures are trained using these GAF images, and the best performing network is selected to construct a dual stream CNN that extracts both the time domain features of the original CIR signals and the frequency domain features of the GAF images. Meanwhile, physical layer parameters such as the maximum growth amplitude of the received CIR signals (maxGrowthCIR), the amplitudes of the first and second paths near the first path (FP_Ampl1, FP_Ampl2), and UWB signal noise are incorporated to build a multi-dimensional auxiliary feature set. These auxiliary features enhance the ability to distinguish NLOS scenarios that are difficult to identify using CIR features alone, further reducing NLOS errors. Additionally, a hybrid fusion method is used to capture correlations among channel features, mitigating the risk of data drowning caused by differences in input feature dimensions. Simulation results demonstrate that the proposed method achieves outstanding classification performance in both binary and multi-class NLOS tasks, with accuracy exceeding 95%. Compared with traditional algorithms, the proposed method reduces the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of distance estimation by 47.14% and 36.08%, respectively, achieving a coefficient of determination (R²) of 0.9493. Furthermore, based on the recognition results, the proposed method can reduce the root mean square error fraction and the mean absolute error of the distance error by 43.78% and 41.67% respectively compared with the existing methods. In the actual indoor positioning experiments, this method improves positioning accuracy by more than 83%, fully verifying its engineering application potential in complex indoor localization scenarios.