Transformer-based outlier-tolerant adaptive beamforming algorithm

Abstract

Adaptive beamforming technology plays a crucial role in modern communication systems, but existing deep learning methods lack robustness against data acquisition outliers. We propose an outlier-tolerant adaptive beamforming method based on the Transformer architecture for a 16-element linear antenna array with uniform inter-element spacing. First, an array signal model accounting for mutual coupling errors and data acquisition outliers is established, and an in-depth analysis is conducted of the performance degradation mechanism caused by data outliers arising from multiple factors, including computational precision, quantization errors, and system implementation. Second, a Transformer-based outlier signal classifier is designed that intelligently recognizes different types of outlier signals by fusing signal angle information and antenna weight features. On a large-scale dataset containing 56,373,835 samples, the classifier achieves an accuracy of 99.9990%. Finally, a two-stage Transformer adaptive beamforming method incorporating an outlier-aware mechanism is proposed. By combining the outlier-aware preprocessing module with the Transformer-based beamforming method, the system’s robustness to datasets with outliers is significantly enhanced. Experimental results demonstrate that under ideal conditions, the Transformer-based method achieves improvements in main lobe accuracy of approximately 15% and 16% compared to the traditional Null Steering Beamforming (NSB) and the Gated Recurrent Unit (GRU)-based adaptive beamforming methods, respectively, with approximately 10% improvement in the Signal to Interference plus Noise Ratio (SINR) performance. More importantly, under outlier levels of 10%-30%, the proposed outlier-tolerant method achieves significant performance improvement of approximately 75% in both main lobe accuracy and null accuracy compared to the conventional Transformer method, with main lobe accuracy stably maintained within 0.5° and null accuracy controlled within 0.1°, effectively addressing the data outlier robustness problem faced by deep learning beamforming algorithms in practical applications.