DeepAoANet: Learning Angle of Arrival from Software Defined Radios with Deep Neural Networks

Zhuangzhuang Dai, Yuhang He, Tran Vu, 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 (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 Angle-of-Arrival (AoA) solution. We propose a Deep Learning approach to 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 the number of impinging signals and realized mean absolute AoA errors of less than 2°.

Introduction

Determining the origin of a radio signal—Direction Finding—is a cornerstone of modern wireless sensing, with applications ranging from indoor localization to jammer detection. However, two major hurdles persist in real-world deployments. First, the "Multipath Mess": in indoor environments, signals bounce off walls and furniture, creating "ghost" signals that confuse classical algorithms like MUSIC (Multiple Signal Classification). Second, the "Sim-to-Real" Gap: most existing deep learning approaches for AoA are trained on synthetic, "perfect" data. When these models meet real-world hardware—with its inherent noise, antenna coupling, and gain inconsistencies—their performance often collapses. Furthermore, classical methods usually require the user to know the number of incoming signals a priori, which is rarely possible in dynamic, real-world scenarios.

Methodology

The proposed DeepAoANet framework processes the sample covariance matrix derived from the In-phase and Quadrature (IQ) samples of a four-antenna array. The covariance matrix is a strategic choice because it encapsulates the phase differences between antennas—the primary indicator of direction—while filtering out the specific modulation or content of the signal.

Impact of antenna array size and spacing on the relative power distribution across different angles.

The architecture features a multi-head design to address the ambiguity of signal counts. A classification head determines if there is one or two impinging signals, while two separate regression heads predict the specific angles ($\theta_1, \theta_2$). The researchers evaluated two variants: DeepAoANet-FC (a standard fully-connected multilayer perceptron) and DeepAoANet-CNN (which treats the covariance matrix as a multi-channel image). Training involved a two-stage fine-tuning process, initially prioritizing the classification of the number of signals before refining angular precision.

Data Collection and Augmentation

A standout contribution of this work is the release of a massive real-world dataset. The hardware setup utilized the KerberosSDR, a low-cost unit with four coherent RTL-SDR dongles, and LoRa beacons as transmitters. Data was collected in diverse settings, including open areas, offices, corridors, and indoor-to-outdoor scenarios. To achieve high angular resolution, the team employed "phase augmentation"—mathematically shifting real-world samples to create virtual angles—and added synthetic Additive White Gaussian Noise (AWGN) to simulate different distances and signal strengths. The final dataset contains over 2.6 million snapshots, providing a robust benchmark for the community.

Key Results and Insights

DeepAoANet achieved a mean absolute error (MAE) of less than 2°. Notably, while classical MUSIC only works well within a narrow Field of View (approx. ±45°), DeepAoANet maintained high accuracy across a much wider range (±74°).

Polar representation of relative power for varying phase shifts, illustrating the spatial response of the antenna array.

The CNN variant proved significantly more robust at very low Signal-to-Noise Ratios (below -5dB), suggesting that convolutional layers are better at extracting structural phase patterns from noisy radio data than standard dense layers. Furthermore, at 0dB SNR, DeepAoANet maintained a 2.5° RMSE, whereas traditional Support Vector Regression (SVR) errors spiked to over 10°. The model is also highly efficient, with inference taking only ~8ms on a Raspberry Pi 4, demonstrating that sophisticated deep-learning-based direction finding can run on inexpensive, single-board computers in real-time.

Conclusion and Future Work

DeepAoANet demonstrates that a hybrid deep learning approach can effectively bridge the gap between theoretical signal processing and real-world SDR hardware. By treating the number of sources as a classification problem and their locations as a regression problem, the system achieves superior robustness in multipath environments.

Despite its success, the current model is capped at identifying a maximum of two simultaneous sources. Future work will focus on scaling the model to handle more sources and exploring frequency-agnostic architectures to enable operation across different radio bands 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 Angle-of-Arrival (AoA) solution. We propose a Deep Learning approach to 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 Light-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 $2^{\circ}$.
BibTeX
@inproceedings{dai2022deepaoanet,
  author = {Zhuangzhuang Dai and Yuhang He and Tran Vu 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},
}
Figure