Sturdy RF Direction Finding in WSNs Under Multipath and Non-Line-of-Sight Conditions Using Machine Learning
TL;DR
A sturdy, learning-based direction-finding framework that jointly detects and classifies the propagation condition (LOS, weak-NLOS, strong-NLOS) from low-level array statistics, and regresses the angle of arrival using a hybrid convolutional–recurrent network trained to be invariant to multipath-induced covariance distortion is presented.
Abstract
Accurate radio-frequency (RF) direction finding (DF) is a cornerstone capability for localization, tracking, and situational awareness in wireless sensor networks (WSNs). Classical subspace and beamforming-based angle-of-arrival (AoA) estimators such as MUSIC and ESPRIT degrade sharply once the propagation channel departs from the idealized single-path, line-of-sight (LOS) assumption on which they are built. In dense, cluttered, or urban WSN deployments, multipath fading and non-line-of-sight (NLOS) propagation routinely corrupt the array manifold, biasing angle estimates by tens of degrees and destabilizing downstream localization. This paper presents a sturdy, learning-based direction-finding framework that jointly (i) detects and classifies the propagation condition (LOS, weak-NLOS, strong-NLOS) from low-level array statistics, and (ii) regresses the angle of arrival using a hybrid convolutional–recurrent network trained to be invariant to multipath-induced covariance distortion. The proposed pipeline ingests the sample covariance matrix and eigen-spectrum of a uniform circular sensor array (UCA), which offers unambiguous 360° azimuthal coverage well suited to WSN nodes with no prior knowledge of source half-plane, applies a lightweight convolutional feature extractor to capture spatial correlation structure, and uses a bidirectional long short-term memory (Bi-LSTM) stage together with an attention layer to resolve angular ambiguities introduced by correlated multipath returns. An auxiliary NLOS-classification head regularizes training and enables condition-aware error correction at inference time. Extensive simulations across a Saleh–Valenzuela multipath channel model, varying signal-to-noise ratio (SNR), number of multipath taps, K-factor, and array geometry, show that the proposed model reduces root-mean-square angular error by 46–68% relative to MUSIC, ESPRIT, and a plain feed-forward DoA regressor, while sustaining sub-3° RMSE down to 0 dB SNR and under severe NLOS conditions. We further demonstrate resilience to sensor-count reduction and model compactness suitable for resource-constrained WSN nodes.