Leveraging Deep Learning for Practical DoA Estimation: Experiments with Real Data Collected via USRP.

Area of Science:

• Signal Processing
• Machine Learning
• Wireless Communications

Background:

• Direction-of-Arrival (DoA) estimation is crucial for wireless systems.
• Traditional DoA methods face challenges with complex signal environments.
• Deep learning offers potential for improved DoA estimation accuracy.

Purpose of the Study:

• To experimentally validate deep learning models for DoA estimation.
• To compare performance using synthesized versus real-world radio data.
• To assess the impact of environmental factors on DoA accuracy.

Main Methods:

• Developed and trained deep neural networks (DNN) and convolutional neural networks (CNN) for DoA estimation.
• Utilized synthesized data incorporating mutual coupling and multipath effects.
• Employed realistic data collected via Universal Software Radio Peripheral (USRP).

Main Results:

• Deep learning models trained with USRP-collected data achieved superior DoA estimation accuracy.
• Synthesized data training resulted in less accurate estimations compared to real-world data.
• Indoor environments with strong non-line-of-sight (NLoS) signals degraded estimation performance.

Conclusions:

• Experimental validation confirms the efficacy of deep learning for DoA estimation.
• Real-world radio data is essential for robust and accurate deep learning-based DoA estimation.
• Environmental conditions, particularly NLoS propagation, significantly impact DoA estimation performance.