A Hybrid Deep Learning and Particle Filter Framework for NLOS Mitigation in AI-Enhanced UWB Indoor Navigation

Abstract

Ultra-Wideband (UWB) is one of the best technologies for accurate indoor location, which is a key part of the Industrial Internet of Things (IIoT). But when there are Non-Line-of-Sight (NLOS) situations, its performance drops a lot. This research presents an innovative hybrid architecture that integrates a deep convolutional neural network (CNN) with a particle filter (PF) to effectively counteract non-line-of-sight (NLOS) effects. The CNN directly looks at the raw Channel Impulse Response (CIR) from UWB anchors to do both NLOS classification and range error regression at the same time. This data changes the measurement noise covariance of a PF in real time, which stops the filter from believing measurements that are wrong. In a difficult industrial testbed, our CNN model was put through a lot of tests and was able to identify NLOS with 98.5% accuracy. The hybrid framework cuts the average 3D positioning error down to 8.4 cm when used in the navigation system. This is a 76% reduction over a conventional Extended Kalman Filter (35.1 cm) and a 55% improvement over a regular PF (18.7 cm). Also, the 95th percentile error is cut down to 21.2 cm, which shows that it is quite reliable for use in industry.