Covariance Matrix-Based DOA Estimation Using CNN: Performance Across Narrow and Wide Grids

Deep learning has emerged as a powerful tool for enhancing Direction-of-Arrival (DOA) estimation, offering resilience to noise, model mismatches, and computational inefficiencies inherent in traditional methods. This paper proposes a convolutional neural network (CNN)-based approach that leverages covariance matrices of received signals to extract spatial features critical for precise DOA estimation. The proposed model demonstrates robust performance across narrow and wide angular grids under varying Signal-to-Noise Ratio (SNR) conditions. Notably, it achieves a Root Mean Square Error (RMSE) below 3.5 degrees for the narrow grid and within 24 degrees for the wide grid, even at SNR levels as low as -20 dB. The study highlights the model’s ability to generalize effectively while acknowledging performance gaps for extreme noise and broader angular separations. A comparative analysis with traditional methods underscores the computational efficiency and adaptability of the CNN framework. Future work will focus on hybrid models, real-world dataset expansion, and architectural optimizations to enhance scalability in next-generation sensing and communication applications.