Angle of Arrival and Centimeter Distance Estimation on a Smart UWB Sensor Node

Tobias Margiani, Silvano Cortesi, Milena Keller, Christian Vogt, Tommaso Polonelli, Michele MagnoIEEE Trans. Instrumentation & Measurement 2023 · DOI

TL;DR

This paper presents a compact, low-power UWB sensor node capable of estimating both distance and angle of arrival (AoA) with centimeter-level precision. By leveraging the Phase Difference of Arrival (PDoA) feature of the new Qorvo DW3220 UWB chip and employing machine learning techniques, the authors achieve significant improvements in AoA accuracy, overcoming limitations like front-back ambiguity and multi-path interference.

What Problem Does This Paper Address?

Precise and low-power indoor localization is a critical requirement for many emerging IoT applications, from robotics to smart manufacturing. While Ultra-Wideband (UWB) technology offers promising centimeter-level accuracy and low power consumption, existing systems often struggle with multi-path effects and non-isotropic antennas, necessitating numerous fixed anchors for reliable positioning. Current UWB systems primarily provide scalar distance measurements, lacking directional information. This paper addresses the challenge of achieving accurate, directional (vectorial) UWB localization using a compact, single-board solution, specifically focusing on improving Angle of Arrival (AoA) estimation.

Methods

The core of this work is the development of a novel sensor node built around the Qorvo DW3220 UWB transceiver. This chip integrates a dual-antenna interface and on-chip Phase Difference of Arrival (PDoA) estimation, a key feature for AoA determination. The authors designed a compact PCB with two antennas spaced at approximately half a wavelength (2.31 cm) to facilitate PDoA measurement.

To overcome the inherent limitations of raw PDoA measurements, such as front-back ambiguity and non-linear behavior at certain angles, the paper explores two machine learning (ML) approaches:

  • Regression Models: Gradient boosting (XGBoost) and Neural Networks (NNs) were trained to predict the AoA estimation error and the actual distance between UWB modules. These models use PDoA, TDoA, Channel Impulse Response (CIR), and signal quality indicators as input features.
  • Classification Models: A two-step approach was also investigated, where an initial classification step divides the 360° AoA range into 90° zones. Subsequent regression or direct calculation within these zones is then performed.

The authors also conducted extensive simulations using ANSYS HFSS and SIwave to model the antenna coupling and current density distribution on the PCB, aiding in the understanding of potential RF non-idealities.

Datasets

The primary dataset used for training and evaluation was collected using the developed sensor node in a controlled indoor environment. This dataset comprises:

  • 19,880 UWB exchanges (TWR measurements).
  • Measurements at 11 fixed distances ranging from 50 cm to 5.5 m.
  • For each distance, the double antenna module was rotated in 1° steps, with at least five TWR measurements performed at each step.
  • Collected data includes CIR, TDoA, PDoA, transmission quality, signal strength, and computed distance.

A secondary "evaluation" dataset of 6,651 samples was also generated under different environmental conditions to test the generalization capabilities of the ML models.

Key Insights and Ideas

  • DW3220's PDoA is a Game Changer: The integrated PDoA capability of the DW3220 chip significantly simplifies hardware design and reduces the need for complex synchronization between multiple receivers, which was a bottleneck in previous UWB AoA systems.
  • ML is Crucial for Accuracy: Raw PDoA measurements are insufficient for accurate AoA estimation beyond a limited angular range (±45°). Machine learning models are essential for compensating for non-linearities, front-back ambiguity, and environmental effects.
  • Environmental Learning: The ML models appear to learn to compensate for environmental effects (reflections, multi-path) rather than purely the antenna's physical behavior, suggesting that model re-training might be necessary for significantly different environments.
  • Two-Step Approach for Robustness: A classification-then-regression approach shows promise for improving generalization, especially in unknown environments, by first identifying the operative zone before applying specific correction models.
  • Compactness Achieved: The design successfully integrates dual antennas and the UWB transceiver onto a small 3 cm x 3.5 cm PCB, making it suitable for various IoT applications.

Key Results

  • AoA Accuracy: The system achieves an average angular accuracy of 0.59° within the ±45° range around the center, with a Mean Squared Error (MSE) of 0.59° and a Mean Absolute Error (MAE) of 5.21° at a 3m distance. For all evaluated distances, the maximum MSE was 2.40° and MAE was 8.46°.
  • Error Correction: ML models significantly improve AoA estimation, reducing errors by up to 35x in some cases. The best NN model achieved an MSE of 1.52° on test data after optimization.
  • Distance Accuracy: The corrected TWR distance estimation also improved, with the NN model achieving an MSE of 0.29 cm on the train dataset and 2.24 cm on the test dataset.
  • Stability: PDoA measurements showed good stability over different distances and time, with results spread as a Gaussian distribution. Averaging 10 AoA estimations reduced the standard deviation to below 1°.
  • Front-Back Ambiguity: The ML models effectively solved the front-back ambiguity, allowing for estimation across the full 360° range.

Limitations and Future Work

  • Generalization: The ML models showed reduced performance on the secondary evaluation dataset compared to the test dataset, indicating that generalization to entirely new environments without re-tuning is challenging.
  • Environmental Dependence: The models primarily learn to compensate for environmental effects, suggesting that for truly "plug-and-play" solutions, further research into models that capture antenna physics more directly might be needed.
  • 180° Zone Issues: Accuracy significantly drops in the zone around 180° AoA, with high MSE values observed even after ML correction.
  • Future Work: The authors suggest exploring the generalizability of ML models with larger datasets, combining multiple dual-antenna modules for 3D localization, and investigating alternative ML architectures or training strategies.

Summary updated

BibTeX
@article{margiani2023angle,
  author = {Tobias Margiani and Silvano Cortesi and Milena Keller and Christian Vogt and Tommaso Polonelli and Michele Magno},
  title = {{Angle of Arrival and Centimeter Distance Estimation on a Smart UWB Sensor Node}},
  journal = {IEEE Trans. Instrumentation \& Measurement},
  year = {2023},
  doi = {10.1109/tim.2023.3282289},
}
Figure