Performance Advantages of Deep Neural Networks for Angle of Arrival Estimation

Oded Bialer, Noa Garnett, Tom Tirer

TL;DR

This paper demonstrates that Deep Neural Networks (DNNs) can solve the "Angle of Arrival" (AOA) estimation problem—identifying where signals are coming from—with the accuracy of a Maximum Likelihood estimator but at a fraction of the computational cost. Crucially, the proposed DNN outperforms classical signal processing methods like MUSIC and Alternating Projections when only a single "snapshot" of data is available and when the antenna array has hardware inaccuracies.

What Problem Does This Paper Address?

In fields like radar, sonar, and wireless communications, we often need to determine how many signal sources exist and their specific angles relative to an antenna array.

The "gold standard" for accuracy is the Maximum Likelihood (ML) estimator. However, ML is computationally "intractable": as the number of sources grows, the search space expands exponentially, making it impossible to run in real-time. To cope, engineers usually turn to "super-resolution" algorithms (like MUSIC or ESPRIT) or iterative approximations.

The catch? These alternatives often require multiple data observations (snapshots) to work well. In dynamic environments where sources or receivers are moving quickly, you might only get one snapshot. Furthermore, classical methods are brittle; they degrade quickly if the antenna array has slight physical imperfections or "mismatches."

Methods

The authors propose a Deep Neural Network (DNN) architecture to replace traditional iterative solvers.

  • Architecture: A fully connected (FC) network with 8 hidden layers. The input is a vector of $2N$ values representing the real and imaginary components of the signal received by $N$ antennas.
  • Multi-Task Learning: The network performs two tasks simultaneously:
  • Classification: It predicts the number of sources (from 1 to 4).
  • Regression: It estimates the specific angles for each possible number of sources.
  • Training: The model was trained using simulated data with varying Signal-to-Noise Ratios (SNR) and random signal coefficients. They used a specialized Root Mean Square Error (RMSE) loss function that calculates the distance between true angles and their closest predicted estimates.
  • Comparison: The authors compared the DNN against a suite of classical algorithms, including MUSIC (with spatial smoothing), Alternating Projections (AP), and Orthogonal Matching Pursuit (OMP).

Datasets

The study relies on simulated datasets generated according to a standard linear array model:

  • Array Configuration: 16 antenna elements (Uniform Linear Array) with $\lambda/2$ spacing.
  • Source Variety: Scenarios with 1 to 4 sources.
  • Signal Parameters: Sources had random phases and amplitudes.
  • Noise/Impairments: Data included additive white Gaussian noise and "mismatch" matrices to simulate hardware errors like phase offsets and channel crosstalk.

While no real-world field data was used, the simulations were designed to stress-test the algorithms under realistic hardware non-idealities.

Key Insights and Ideas

  • Breaking the Complexity Barrier: The DNN effectively "compresses" the expensive ML search into a series of matrix multiplications. While the ML estimator's complexity is $O(P^M)$ (where $P$ is the grid size and $M$ is the number of sources), the DNN's complexity is fixed and roughly seven orders of magnitude lower for four sources.
  • Single-Snapshot Superiority: Classical methods like MUSIC usually fail with a single snapshot unless they use "spatial smoothing," which shrinks the effective size of the antenna. The DNN learns to extract features from the raw signal that these methods miss, maintaining high resolution without sacrificing aperture.
  • Accidental Robustness: One of the most surprising findings was that the DNN was more robust to array inaccuracies (mismatch) than classical methods, even though it wasn't specifically trained on mismatched data. The authors attribute this to the "effective regularization" provided by training on a massive variety of source angles and noise levels.
  • Model Order Determination: Estimating how many sources exist is notoriously difficult with a single snapshot. The DNN's classification accuracy (90% at high SNR) significantly beat traditional information-theoretic criteria like MDL and AIC (72%).

Key Results

  • ML Performance: For two sources, the DNN's error curve perfectly matched the Maximum Likelihood estimator, reaching the theoretical Cramér-Rao Lower Bound.
  • Accuracy: In a 4-source scenario within a 50-degree field of view, the DNN achieved an RMSE of less than 0.5 degrees. This is significantly better than the 10-degree beamwidth of the array itself.
  • Speed: Because the DNN architecture is highly parallelizable, it is much faster than iterative methods like Alternating Projections when executed on a GPU.
  • Robustness: Under "high impairment" conditions (significant phase offsets and crosstalk), the DNN maintained a massive performance lead over the Alternating Projections algorithm.

Limitations and Future Work

  • Scaling: The current model is limited to a maximum of 4 sources. In dense environments (like electronic warfare), the number of sources could be much higher, requiring a more complex output structure.
  • Fixed Geometry: The network was trained for a specific 16-antenna array. If the antenna geometry changes, the network likely needs to be retrained or adapted via transfer learning.
  • Sim-to-Real Gap: While the authors simulated hardware impairments, the model has not yet been validated with data from physical hardware in the field.
  • Future Directions: The authors suggest that exploring different architectures, like "DenseNets," might provide similar performance with even fewer parameters.

Summary updated

Abstract

The problem of estimating the number of sources and their angles of arrival from a single antenna array observation has been an active area of research in the signal processing community for the last few decades. When the number of sources is large, the maximum likelihood estimator is intractable due to its very high complexity, and therefore alternative signal processing methods have been developed with some performance loss. In this paper, we apply a deep neural network (DNN) approach to the problem and analyze its advantages with respect to signal processing algorithms. We show that an appropriate designed network can attain the maximum likelihood performance with feasible complexity and outperform other feasible signal processing estimation methods over various signal to noise ratios and array response inaccuracies.
BibTeX
@inproceedings{bialer2019performance,
  author = {Oded Bialer and Noa Garnett and Tom Tirer},
  title = {{Performance Advantages of Deep Neural Networks for Angle of Arrival Estimation}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2019},
  url = {https://openreview.net/forum?id=JV9xE1Jcpk},
}

Loading related content...

Figure