Icosahedral CNN for DoA Estimation of EM Wave Sources Arriving Spherical Antenna Array Incorporating the Impact of Mutual Coupling

Spherical antenna array (SAA) is usually employed anytime hemispherical scan coverage is needed, because it generates the same level of directivity in all scan angles. It is applicable in satellite communications, vehicular technology, and aerospace. Hence, the direction-of-arrival (DoA) estimation of electromagnetic (EM) waves impinging an SAA with undefined mutual coupling requires academic and industry attention. This paper presents a method for DoA estimation of EM waves via icosahedral convolution neural network (CNN) over steered response power with phase transform power map calculated from EM waves arriving SAA. Icosahedral CNN exhibits equal variation to 60 rotational icosahedral symmetries, representing an appreciable approximates continuous space approximates of spherical rotations, and implementable by the general 2D convolutional layers, exhibiting smaller cost of computation than spherical CNNs. Furthermore, as against the fully connected layers following icosahedral convolutions, soft-argmax function, which is a differential form of the popular argmax function, is employed. This make it possible to compute DoA in form of regression, which represents the convolutional layers output in form of probability distribution. Numerical experimental results shows that the proposed method with fitting equal variation of the problem perform better than the previous methods, in terms of robustness, lower cost of computation, and a root mean square errors of localization lesser than 11 degree, even in harsh environments.