DOA Estimation using Multi-head Self-Attention with Relative Positional Encoding and Hybrid Multi-Objective Grey Wolf Optimization

Abstract

Direction of Arrival (DOA) estimation is the process of estimating the incident angle of waves from a radiating or reflecting source using an antenna array. DOA estimation is highly important in several areas, such as navigation, remote sensing, radar, sonar and wireless communication. Nevertheless, noise, interference and multipath effects complicate accurate DOA estimation. Deep learning (DL) techniques, which model the intricate relationships in signal data, can overcome these challenges. This work presents a new DOA estimation method that uses a hybrid LSTM-CNN feature extraction pipeline, multi-head attention with GWO-mRMR-based feature selection, and an ANN. The method includes improvements such as positional encoding in multi-head attention and an adaptive learning rate in the ANN. It is trained and evaluated on synthetic signal data generated by a Uniform Linear Array (ULA) model. In the evaluation, it achieves MAE (2.15), MSE (15.68), RMSE (3.96), and R² score (0.98). This accuracy can be attributed to the robust feature extraction pipeline, which captures spatial and temporal features effectively, and to the model’s focus on the most relevant and informative features. Additionally, the method achieves lower MSE and RMSE than MUSIC, SVR and ESPRIT and a clear advantage at low SNR, while the classical subspace methods remain competitive at high SNR.