Deep Learning Based DOA Estimation With Trainable-Step-Size LMS Algorithm

In this paper, we investigate the trainable-step-size least mean square (TSS-LMS) algorithm, which combines the LMS algorithm model with deep learning for the direction of arrival (DOA) estimation using adaptive filtering. Although the existing variable-step-size LMS approaches have been proposed to improve the DOA estimation accuracy, they are generally unsuitable for scenarios with the limited number of snapshots. Therefore, we propose a deep learning-based algorithm driven by the LMS model in this work, which incorporates the trainable step size. More specifically, the iterative mathematical model based on the LMS algorithm is unfolded into a deep network, where each iteration corresponds to a layer of this network. Based on this structure, the iteration step size in each layer is set as a trainable variable, allowing for adaptation during the training of the network. Through deep learning with the received signal dataset, the TSS-LMS network improves the adaptability and DOA estimation accuracy of the LMS algorithm for the limited snapshots scenario. Simulation results show that our proposed method is effective and outperforms existing algorithms in terms of DOA estimation performance and computational complexity trade-off.