Machine Learning-Based Direction-of-Arrival Estimation Exploiting Distributed Sparse Arrays
Saidur R. Pavel, Md. Waqeeb T. S. Chowdhury, Yimin D. Zhang, Dan Shen, Genshe Chen
Abstract
Distributed sparse arrays, consisting of multiple subarrays, facilitate a higher number of degrees of freedom and enhanced direction-of-arrival (DOA) estimation performance beyond what is offered by single uniform linear arrays. When the array elements in each subarray are sparsely located, the covariance matrix is sparse with missing entries. Covariance matrix interpolation is commonly exploited to fill in the missing elements of the covariance matrix. Such techniques, however, would degrade or fail when array imperfection occurs, such as imperfect calibration or knowledge in sensor gain, phase, position, and inter-element mutual coupling. Such imperfections affect the matrix interpolation and impede accurate DOA estimation. To address these issues, we propose a neural network structure, which is trained to learn the relationship between the input sparse covariance matrix and the true signal directions. The neural network is trained based on minimizing a loss function between the predicted neural network output and the actual output so that the network is forced to extract the essential feature, rendering accurate DOA estimation results.