Improving Non-Line-of-Sight Identification in Cellular Positioning Systems Using a Deep Autoencoding and Generative...

Area of Science:

• Wireless communication
• Signal processing
• Machine learning for localization

Background:

• Positioning services are vital for integrating physical and digital information, with 5G technology enhancing their role in cellular networks.
• Non-line-of-sight (NLoS) propagation significantly degrades positioning accuracy by affecting angle and delay measurements.
• Accurate indoor positioning remains a challenge, particularly in complex environments like factories with limited base station coverage.

Purpose of the Study:

• To develop an advanced deep learning model for enhancing positioning accuracy in 5G networks.
• To address the challenges posed by non-line-of-sight (NLoS) propagation in cellular positioning systems.
• To validate the model's performance and generalization capabilities in diverse indoor 5G environments.

Main Methods:

• A deep autoencoding channel transform-generative adversarial network (DACT-GAN) model was proposed.
• The model utilized line-of-sight (LoS) samples for training to extract latent features.
• A discriminator was employed within the GAN framework to identify non-line-of-sight (NLoS) signals.

Main Results:

• The proposed DACT-GAN model demonstrated improved positioning accuracy in 5G indoor and factory scenarios.
• Compared to state-of-the-art methods, the model reduced device resource utilization.
• The model achieved a 2.15% higher area under the curve and a 12.6% reduction in computing time.

Conclusions:

• The DACT-GAN model effectively mitigates the impact of NLoS propagation on positioning accuracy.
• The approach shows significant improvements in resource efficiency and computational speed.
• This method offers a promising solution for future positioning terminals in commercial and industrial Internet of Things (IoT) applications.