Practical CNN-Based Beam Selection for Wideband Single-Receiver Switched Beam Antenna

Abstract

In this paper, a beam selection method is presented for a switched beam antenna (SBA) specifically designed and built for long-range wireless communications. The proposed SBA operates with eight beams for automatic direction finding (DF) and 360-degree beam activation at 900-1400 MHz frequency range. Remarkably, this method employs a simple switch board and one sampler without synchronization requirement, avoiding the need for complex circuitry typically required for separate beam sampling. Instead, a convolutional neural network (CNN) processes the sampled signals to accurately determine the signal direction. The proposed CNN is trained by using a two-stage method with both artificial and real data to achieve fine-tuned accuracy of 99.79% with artificially generated data and 95% with real data for signal-to-noise ratios (SNR) above −10 dB. Additionally, the CNN is optimized to run on low-end hardware. A significant 16% accuracy enhancement compared to the direction estimation method based on conventional power measurement (PM) under identical conditions is demonstrated through real-time beam selection using a commercially available Raspberry Pi board. This proposed low-cost, wideband, high-accuracy, and highly efficient SBA is an excellent candidate for practical applications, showcasing the effectiveness and accuracy of data-driven methods in real world, especially in low SNR scenarios ( $\lt -10$ dB).

Keywords

• Smart antenna
• switched beam antenna (SBA)
• direction of arrival (DOA) estimation
• deep learning (DL)