DOA-CNN: An automatic system error calibration model for enhancing accuracy in direction of arrival estimation of radio frequency signals

Abstract

This paper presents a research proposal on a deep convolutional neural network for the problem of direction of arrival estimation of radio frequency signals (called DOA-CNN). The DOA-CNN model is designed with multiplication layers to enhance strong features of the data through convolutional stacks enabling the DOA classification accuracy. The evaluation considers several factors affecting the accuracy of DOA estimation for uniform linear array (ULA), including antenna element position errors, and amplitude and phase errors caused by transmission path deviations in the receiver. The analysis and comparison of DOA-CNN with CBF, Capon, MUSIC, Root-MUSIC, and ESPRIT methods and other machine learning methods show that, considering the ideal configuration of the ULA array and the receiver, the Root-MUSIC and ESPRIT methods achieve the best accuracy since they can directly compute the DOA, while the other methods estimate the DOA via angular spectrum, leading to accuracy dependent on the spectral resolution. However, considering ULA errors and transmission path deviations in the receiver, the proposed DOA-CNN model outperforms in terms of accuracy compared to traditional methods and processes faster than some other machine learning models.