Convolutional Neural Network-Based DOA Estimation Using Non-uniform Linear Array for Multipath Channels

In this paper, a novel convolutional neural network (CNN) was designed for DOA estimation, which could deploy in radio-electronics systems for improving the accuracy and operation efficiency. The proposed model was evaluated with different hyper-parameter configurations for optimization, and then a suitable model was compared with other existing models to demonstrate its preeminence. Regarding dataset generation, our work considered the influence of both Gaussian noise and multipath channels to DOA estimation accuracy. According to the analysis, in frame of this study, the model with 5 conv-blocks, 48 filters, and a filter size of 1 × 7 achieved the best performance in terms of accuracy (75.27 % at + 5 dB SNR) and prediction time (10.1 ms) that notably outperformed two other state-of-the-art CNN model-based DOA estimation techniques. © 2020, ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering.