EE Seminar: Deep Learning - Aided Subspace Methods for Direction-of-Arrival Estimation
הרישום לסמינר יבוצע בתחילת הסמינר באמצעות סריקת הברקוד למודל (יש להיכנס לפני כן למודל, לא באמצעות האפליקציה)
Registration to the seminar is done at the beginning of the seminar by scanning the barcode for the Moodle (Please enter ahead to the Moodle, NOT by application)
Electrical Engineering Systems Seminar
Speaker: Itamar Almog
M.Sc. student under the supervision of Prof. Anthony J. Weiss
Wednesday, 24th December 2025, at 15:00
Room 011, Kitot Building, Faculty of Engineering
Deep Learning - Aided Subspace Methods for Direction-of-Arrival Estimation
Abstract
Direction-of-Arrival (DoA) estimation is a fundamental task in array signal processing, it is known to degrade in practical regimes characterized by low signal-to-noise ratios, limited snapshots, source coherence, and array mismatches such as gain/phase errors, sensor position bias, and mutual coupling. We propose a data-driven front-end that restores a subspace-faithful covariance prior to applying classical estimators. The method ingests multi-lag autocorrelation features and uses a U-Net backbone with lightweight convolutional and Multilayer-Perceptron (MLP) heads to reconstruct a denoised, mismatch-compensated covariance matrix; its Eigen-Decomposition (EVD) is produced explicitly and supplied to downstream algorithms. Training combines a mean-squared reconstruction loss with a subspace-similarity objective that aligns the predicted and reference signal subspaces.
Comprehensive experiments cover varying numbers of sources, a broad span of Signal to Noise Ratio (SNR), both coherent and non-coherent conditions, and realistic levels of array imperfections. Using root-mean-square angular error as the primary metric, we benchmark beamforming, Minimum Variance Distortion less Response (MVDR), multiple signal classification (MUSIC), Root-MUSIC, Estimation of signal parameters via rotational invariant techniques (ESPRIT) and Unitary-ESPRIT with and without the proposed front-end. A single trained model generalizes across operating conditions and consistently improves downstream accuracy relative to applying the classical methods directly to raw sample statistics.
These results indicate that learning to reconstruct the covariance—rather than regressing angles end-to-end—provides a practical, plug-and-play enhancement to classical DoA pipelines. By restoring separation between signal and noise subspaces under coherence and model mismatch, the approach yields robust, accurate estimates while preserving compatibility with standard subspace-based techniques and producing an explicit EVD useful for subsequent processing.
-סמינר זה ייחשב כסמינר שמיעה לתלמידי תואר שני ושלישי-
This Seminar Is Considered A Hearing Seminar For Msc/Phd Students-
