Bluetooth AoA estimation method based on LSTM with spatial-structural feature fusion

In direction estimation based on bluetooth low energy (BLE) angle of arrival, traditional signal processing methods and existing deep learning models often suffer from high computational complexity, poor robustness under low-signal-to-noise ratio (SNR) and multipath conditions, especially limited generalization to unseen angles. To address these issues, this paper proposes a lightweight deep learning method spatial-structural feature long short-term memory (SF-LSTM), which incorporates physics-inspired features to enhance spatial modeling while avoiding reliance on heterogeneous data sources or complex preprocessing. Moreover, a data augmentation method based on array manifold interpolation and physical-consistency perturbation is designed to generate pseudo-in phase and quadrature samples at intermediate angles, the proposed method is effectively alleviated angular sparsity in training data. Both simulations and physical experiments demonstrate that the SF-LSTM method achieves superior performance over existing methods in accuracy, robustness, and cross-distribution generalization. It exhibits particularly strong results in low-coverage angular regions, validating its potential for deployment in resource-constrained BLE indoor localization systems.