Deep Learning-Aided Subspace-Based DOA Recovery for Sparse Arrays

Nir ShlezingerarXiv 2023 · DOI · arXiv

TL;DR

This paper introduces Sparse-SubspaceNet, a deep learning approach that enables accurate direction-of-arrival (DoA) estimation from sparse antenna arrays, even when the array is miscalibrated or sources are coherent. By learning a surrogate covariance matrix, it bridges the gap between data-driven deep learning and traditional subspace methods, offering improved performance and interpretability.

What Problem Does This Paper Address?

Estimating the direction of arrival (DoA) of signals is a fundamental task in array signal processing. Sparse arrays, with fewer antennas than traditional uniform linear arrays (ULAs), can theoretically resolve more sources than their element count by reconstructing a "virtual" ULA. However, existing subspace-based DoA estimators (like MUSIC or ESPRIT) rely on strong assumptions: non-coherent sources, perfect array calibration, and sufficient data. Sparse arrays exacerbate these issues because the virtual array elements are inaccessible, making calibration difficult and increasing sensitivity to noise and limited data. This paper tackles the challenge of performing robust DoA estimation from sparse arrays in the presence of coherent sources, array miscalibrations, and limited snapshots, while retaining the interpretability of subspace methods.

Methods

The core innovation is Sparse-SubspaceNet, a hybrid model-based and data-driven approach. It leverages a deep neural network (DNN) to learn how to compute a "surrogate covariance matrix." This surrogate matrix is designed to mimic the covariance of a calibrated, full virtual ULA, making it suitable for standard subspace DoA estimators.

The process involves three stages:

  1. Linear Transformation (G): A matrix G is used to "lift" the sparse array observations to the dimensions of the virtual ULA. This matrix preserves known element values and uses zero-padding for unknown ones.
  2. Covariance Focusing & Autoencoder: The empirical autocorrelation of the transformed data is computed and then fed into a denoising convolutional autoencoder. This autoencoder, trained with parameters ψ, learns to produce a surrogate covariance matrix.
  3. Surrogate Covariance Formation: The output of the autoencoder is reshaped and combined with a small regularization term (ε) to form a Hermitian positive-definite surrogate covariance matrix.

For training, the authors adapt the ESPRIT subspace method, making it differentiable to allow for backpropagation. They introduce a Sort Mean-Squared Error (SMSE) loss function, which is permutation-invariant and computationally feasible for a large number of sources, unlike previous approaches. This allows the DNN to be trained to produce a surrogate covariance that yields accurate DoA estimates when processed by ESPRIT.

Datasets

The paper describes training and evaluation using simulated data. The authors generated datasets of observed sequences (X) and their corresponding DoAs (θ). They do not mention using publicly available datasets or introducing new benchmarks.

Key Insights and Ideas

  • Bridging Deep Learning and Subspace Methods: The paper successfully demonstrates how deep learning can enhance traditional signal processing techniques by learning to overcome their inherent limitations, particularly in challenging scenarios like sparse and miscalibrated arrays.
  • Learned Covariance for Interpretability: Instead of a purely end-to-end deep learning approach for DoA estimation, Sparse-SubspaceNet learns a covariance matrix. This preserves the interpretability of subsequent subspace methods, allowing users to visualize the spectrum (as shown in Fig. 3) and understand how DoAs are recovered.
  • SMSE Loss for Scalability: The introduction of the SMSE loss function is crucial for enabling training with a large number of sources, a common scenario where traditional permutation-invariant losses become computationally intractable.
  • Robustness to Miscalibration and Coherence: The results highlight that the learned surrogate covariance effectively compensates for array miscalibrations and the presence of coherent sources, which are significant challenges for conventional methods.

Key Results

  • Sparse-SubspaceNet shows only a minor performance degradation compared to using a full ULA, even when detecting more sources than available antennas.
  • It significantly outperforms model-based matrix completion methods for multiple sources, handling both coherent and non-coherent signals effectively.
  • The method demonstrates robustness to array miscalibrations, even surpassing model-based subspace methods applied to a full (miscalibrated) ULA in some scenarios.
  • Qualitative results (Fig. 3) show that Sparse-SubspaceNet produces a spectrum from which Root-MUSIC can reliably identify DoAs, similar to applying Root-MUSIC directly to a calibrated full ULA.

Limitations and Future Work

  • Fixed Dimensionality Lifting Matrix (G): The paper mentions that the matrix G is fixed. Future work could involve learning G from data alongside the DNN, potentially further improving performance.
  • Element Placement Learning: The authors suggest that the ability to convert a sparse array estimator into a trainable model opens possibilities for learning the optimal placement of array elements for specific DoA estimation tasks.
  • Hyperparameter Tuning: The architecture involves hyperparameters like Tmax and ε, which would require careful tuning for optimal performance.

The paper concludes by emphasizing the successful integration of data-driven learning with model-based subspace methods to achieve robust DoA estimation from sparse arrays, preserving interpretability.

Summary updated

BibTeX
@misc{shlezinger2023deep,
  author = {Nir Shlezinger},
  title = {{Deep Learning-Aided Subspace-Based DOA Recovery for Sparse Arrays}},
  year = {2023},
  eprint = {2309.05109},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2309.05109},
}

Loading related content...

Figure