DeepAoANet: Learning Angle of Arrival From Software Defined Radios With Deep Neural Networks

Zhuangzhuang Dai, Yuhang He, Vu Tran, Niki Trigoni, Andrew Markham

Abstract

Direction finding and positioning systems based on RF signals are significantly impacted by multipath propagation, particularly in indoor environments. Existing algorithms like MUSIC perform poorly in resolving Angle of Arrival (AoA) in the presence of multipath or weak signal regimes. We propose DeepAoANet, a deep learning approach for deriving AoA from a single snapshot of multichannel data from low-cost Software-Defined Radios (SDRs). By employing a hybrid classification-regression architecture, our method estimates up to two AoAs accurately. We demonstrate the approach using a four-element Universal Linear Array (ULA) in various Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) environments. Our proposed method achieves mean absolute AoA errors of less than 2° and demonstrates high reliability in determining the number of impinging signals.

Introduction

Determining the Angle of Arrival (AoA) of a radio signal is a cornerstone of indoor localization, drone tracking, and emergency response systems. Traditional methods, such as the Multiple Signal Classification (MUSIC) algorithm, rely on the mathematical orthogonality between signal and noise subspaces. However, these classical approaches often falter in real-world indoor settings where multipath propagation creates "ghost" signals and low Signal-to-Noise Ratios (SNR) obscure the primary signal. Furthermore, traditional algorithms typically require prior knowledge of the number of incoming signals.

DeepAoANet addresses these challenges by leveraging the raw capabilities of digitally sampled RF frontends. Unlike previous deep learning models trained on idealized synthetic data, DeepAoANet is designed to handle the hardware imperfections and noise inherent in affordable, off-the-shelf SDR modules.

Methodology

The DeepAoANet framework adopts a "physical layer" approach to signal processing. Instead of processing raw In-phase and Quadrature (IQ) samples directly, the system calculates the Sample Covariance Matrix (SCM). This matrix captures essential phase correlations between antenna elements while remaining agnostic to the specific modulation of the transmitted data.

Theoretical impact of antenna count and spacing on the angular power distribution, illustrating the resolution limits of a 4-element array.

The architecture features a hybrid output head that performs joint tasking. One part of the network performs classification to determine the number of impinging signals (e.g., zero, one, or two sources), while separate regression heads estimate the specific angles ($\theta_1, \theta_2$) for those sources. This two-stage approach is optimized via a weighted loss function that prioritizes classification accuracy before fine-tuning angular regression.

Data Collection and Augmentation

The researchers utilized a KerberosSDR rig with a four-element Universal Linear Array (ULA). They collected approximately 50,000 snapshots of LoRa signals across diverse environments, including offices and corridors. This dataset was augmented to 2.6 million snapshots by synthetically shifting phases and introducing varying levels of Gaussian noise to ensure robustness against real-world variability.

Key Insights and Performance

A significant advantage of DeepAoANet is its ability to treat the SCM as a multi-channel image (in the CNN variant), allowing the network to learn spatial relationships and antenna coupling effects that are difficult to model analytically.

One of the primary challenges in linear array design is the "edge problem," where angular resolution degrades as the signal source moves toward the axis of the antennas (near $\pm90^\circ$). DeepAoANet demonstrates superior stability across a wide Field of View ($\pm70^\circ$) compared to traditional Support Vector Regression (SVR) methods.

Polar plot showing the broadening of the beam pattern as the steering angle increases, highlighting the "edge problem" in linear arrays.

Experimental Results

  • High Accuracy: The system achieved a Mean Absolute Error (MAE) of less than 2°.
  • Robustness: High performance was maintained at SNRs as low as -10dB, a regime where both MUSIC and SVR-based methods typically fail.
  • Real-Time Performance: The inference engine is computationally efficient, running at over 60 samples per second on a Raspberry Pi 4, making it viable for mobile edge devices.
  • Generalization: In blind laboratory tests using a Vicon motion capture system for ground truth, the model successfully tracked multiple moving transmitters it had never encountered during training.

Conclusion

DeepAoANet provides a robust, real-time solution for AoA estimation using low-cost hardware. While currently limited to two simultaneous sources and sensitive to specific hardware configurations, it represents a significant step toward practical, deep-learning-based RF sensing. Future work will focus on semi-supervised learning to scale the number of detectable sources and developing more generalized models that can adapt to varying frequencies and antenna geometries without extensive retraining.

Summary updated

Abstract

Direction finding and positioning systems based on RF signals are significantly impacted by multipath propagation, particularly in indoor environments. Existing algorithms (e.g MUSIC) perform poorly in resolving Angle of Arrival (AoA) in the presence of multipath or when operating in a weak signal regime. We note that digitally sampled RF frontends allow for the easy analysis of signals, and their delayed components. Low-cost Software-Defined Radio (SDR) modules enable Channel State Information (CSI) extraction across a wide spectrum, motivating the design of an enhanced AoA solution. We propose a Deep Learning approach for deriving AoA from a single snapshot of the SDR multichannel data. We compare and contrast deep-learning based angle classification and regression models, to estimate up to two AoAs accurately. We have implemented the inference engines on different platforms to extract AoAs in real-time, demonstrating the computational tractability of our approach. To demonstrate the utility of our approach we have collected IQ (In-phase and Quadrature components) samples from a four-element Universal Linear Array (ULA) in various Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) environments, and published the dataset. Our proposed method demonstrates excellent reliability in determining number of impinging signals and realized mean absolute AoA errors less than 2°.
BibTeX
@inproceedings{dai2022deepaoanet,
  author = {Zhuangzhuang Dai and Yuhang He and Vu Tran and Niki Trigoni and Andrew Markham},
  title = {{DeepAoANet: Learning Angle of Arrival From Software Defined Radios With Deep Neural Networks}},
  booktitle = {IEEE Access},
  year = {2022},
  url = {https://openreview.net/forum?id=wbN8mDkRP-P},
}

Loading related content...

Figure