A review of direction of arrival estimation based learning models for uniform linear arrays: Methods, challenges, and trends

The uniform linear array (ULA) remains the reference geometry for direction-of-arrival (DoA) estimation in wireless communications, radar, sonar, and electronic surveillance. Subspace estimators such as MUSIC and ESPRIT resolve closely spaced sources when the array is well calibrated and the assumed signal model holds, but their accuracy falls away at low signal-to-noise ratio (SNR), under coherent multipath, and when few snapshots are available. Learning-based estimators do not invert an assumed model. They instead fit a mapping from measured array data to angle, and this review organizes the resulting literature along the two axes that govern what such a mapping can represent. The first axis is the representation used to encode the received data, namely the sample covariance matrix, the spatial pseudo-spectrum, or raw in-phase and quadrature samples. The second is the strategy used to produce angle estimates, whether by classification, regression, multi-label prediction, or sequential prediction. Shallow, deep, hybrid, and physics-embedded architectures are compared within this framework, with particular attention to how much prior signal-processing structure each one retains. Dataset construction is treated as a design problem in its own right, covering simulation parameter selection, measurement campaigns, and the transfer of models trained on synthetic data to physical hardware. Accuracy, complexity, and generalization are examined jointly, since the surveyed studies rarely address all three at once. Two obstacles recur throughout. The absence of shared benchmarks makes reported figures difficult to compare, and most models provide little insight into how an angle estimate is formed. Physics-informed constraints, compact architectures for embedded platforms, and attention-based temporal models are discussed as responses to these obstacles in 5G/6G and automotive sensing.