Machine Learning-Aided Sparse Direction of Arrival Estimation

This article introduces a design and method to address the challenge of decreased accuracy in direction-of-arrival (DOA) estimation when utilizing planar sparse arrays with closely spaced sensors. It proposes a design leveraging Koch fractal geometry, which effectively reduces mutual coupling and enhances the detection of a greater number of sources compared with the number of sensors in the array. In addition, this article presents a technique that integrates machine learning (ML), specifically combining a deep stacked online sequential random vector functional link network with autoencoder (DSOSRVFLN-AE) and a robust multikernel random vector functional link network (RMKRVFLN), for DOA estimation. This technique involves feature extraction through unsupervised learning and DOA estimation via supervised learning. Results demonstrate superior performance, with a root mean square of 0.02 in 0-dB noisy environment, surpassing other prominent methods in terms of computational efficiency, learning speed, accuracy, and robustness. Similarly, figures depict that for a minimum resolution grid interval of 0.5°, the proposed DSOSRVFLN-AE-RMKRVFLN method estimates DOA with an MAE less than 0.02. The method effectively identifies multiple closely spaced targets (with a minimum angular separation of 0.5°) and is validated through digital architecture implemented on an Xilinx Virtex-5 field-programmable gate array (FPGA), indicating its simplicity and practicality.