Improved direction-of-arrival estimation method based on LSTM neural networks with robustness to array imperfections

Array imperfections severely degrade the performance of most physics-driven direction-of-arrival (DOA) methods. Deep learning-based methods do not rely on any assumptions, can learn the latent data features of a given dataset, and are expected to adapt better to array imperfections compared with existing physics-driven methods. Hence, an improved DOA estimation method based on long short-term memory (LSTM) neural networks for situations with array imperfections is proposed in this paper. Various analyses given by this paper demonstrate that the phase features are the key to DOA estimation. Considering the sequential characteristics of the moving target and the correlation of multi-frame data features, the LSTM neural networks are used to learn and enhance the phase features of sampled data. The DOA estimation accuracy and generalization capability are improved by mitigating the phase distortion using LSTM. Numerical simulations and statistical results show that the proposed method is satisfactory in terms of both the generalization capability and imperfection adaptability compared with state-of-the-art physics-driven and data-driven methods.