A Supervised Learning Framework for Joint Estimation of Angles-of-Arrival and Number of Sources

Machine learning is a promising technique for angle-of-arrival (AOA) estimation of waves impinging sensor array. However, the majority methods proposed so far only consider known, fixed number waves, i.e., sources (NOS). This paper proposes machine-learning-based estimator designed case when NOS variable and hence unknown priori. The comprises framework single-label classifiers. Each classifier predicts if are present within certain randomly selected segments array’s field view (FOV), resulting from discretising FOV with (FOV) resolution. classifiers’ predictions combined into probabilistic angle spectrum, whereupon AOAs estimated jointly by applying probability threshold whose optimal level learned data. estimator’s performance assessed using new metric: joint AOA success rate. Numerical simulations show that low SNR (−10 dB), resolution (2°) yields higher rate than high (1°), whereas opposite applies mid (0 dB) (10 SNRs. In nearly all simulations, except one at resolution, outperforms MUSIC algorithm maximum allowed error approximately equal to (or larger than)