UFCNet: U-shaped Fully Connected Network for Improving Direction of Arrival Estimation Accuracy in Electronic Intelligence Systems

In practical applications, it is crucial to eliminate errors, which are significantly introduced in the estimation of the direction of arrival (DOA) by the presence of array defects such as element position errors or phase/gain inconsistencies. This paper proposes a U-shaped deep neural network called UFCNet, which employs fully connected layers exclusively, to automatically calibrate the phase and amplitude errors present in the covariance matrix. The output of UFCNet is a nearly ideal covariance matrix, which is then utilized by the MUSIC or rootMUSIC algorithms to accurately estimate the DOA of signals. Simulation results demonstrate that UFCNet effectively mitigates phase and amplitude errors, thereby providing a nearly perfect covariance matrix for the rootMUSIC algorithm to estimate the DOAs of incoming signals with high precision. In comparison to other traditional methods, the combination of UFCNet and rootMUSIC achieves the best performance in terms of DOA estimation accuracy.