DOA Estimation by Feature Extraction Based on Parallel Deep Neural Networks and MRMR Feature Selection Algorithm

An active research topic in the array signal processing field is the direction of arrival (DOA) estimation, where establishing an efficient estimation of radio waves received from an antenna array is of great interest. Recent studies mostly focus on deep neural network (DNN) algorithms to improve the performance of DOA estimation in mmWave communication systems. These studies did not give importance to the diversity of features extracted from the data and the relationship between them, which affects the efficiency of DOA estimation. To improve the performance, we propose a new model for DOA estimation based on a parallel deep neural network (PDNN). It efficiently extracts spatial and temporal features, then fuses them and selects the distinct ones while removing the redundant ones. In parallel, the proposed model extracts spatial and temporal features using a convolution neural network (CNN) and long short-term memory (LSTM). Combining spatial and temporal features increases efficiency due to feature diversity. Then, a minimum redundancy and maximum relevance (MRMR) algorithm is proposed to select the distinct features. The performance of the proposed model is evaluated by conducting a wide range of experiments on uncorrelated narrowband data with different scenarios and based on the root mean square error (RMSE) coefficient. The results by computer simulations on a uniform linear array show that the proposed structure significantly increases the accuracy of DOA estimation and is more robust to changes in input parameters than traditional algorithms, such as estimation of signal parameters via invariant rotation (ESPRIT), classification of multiple signals (MUSIC), and DNN method. These results demonstrate the superiority of the proposed PDNN model over other techniques, especially in terms of accuracy.