Imbalanced UWB LOS/NLOS Identification Based on Back Propagation Neural Network

Ultra-wide band (UWB) positioning technology is a promising solution for indoor wireless localization. However, the positioning accuracy of UWB-based positioning system is significantly affected by non-light-of-sight (NLOS) components of the channel impulse response (CIR). Therefore, identifying the NLOS components of the CIR is important to UWB positioning. Recently, machine learning techniques have been applied to NLOS recognition based on large-scale balanced datasets in which the number of NLOS signals is comparable to that of line-of-sight (LOS) signals. However, the acquisition of NLOS components are more difficult than that of LOS components in practice. So large-scale datasets are usually unavailable. In this paper, we propose an Imbalanced learning method based on Back Propagation Neural Network (IL-BPNN) to recognize UWB LOS/NLOS signals in the presence of imbalanced datasets in which the number of NLOS signals is not sufficient. In the proposed IL-BPNN method, three different classifiers using oversampling, focal loss function and classification weighting are used to process the imbalanced data independently. In addition, a voting mechanism is employed to select the optimal classification results from the outputs of the above three classifiers. Simulation results show that the proposed method has higher precision and recall compared with the existing feature-based and CIR-based SVM methods.