In the field of wireless communications and satellite navigation, radio frequency interference —whether intentional through spoofing or jamming, or accidental due to environmental conditions— represents one of the most critical challenges for signal integrity and availability. Accurate estimation of the angle of arrival (AoA) of a signal is a fundamental technique for detecting and locating anomalous sources. However, real environments rarely offer ideal line-of-sight (LoS) conditions. Multipath propagation in non-line-of-sight (NLoS) scenarios introduces distortions that significantly degrade the performance of classical methods and purely data-driven approaches.
To overcome this limitation, a promising approach emerges: hybrid learning that combines deep networks with known physical constraints. Instead of treating the neural network as a black box, a plane wave model is incorporated that imposes coherence between phase differences across antennas and the predicted angles. This allows the model to learn environment-invariant representations under LoS conditions, while adapting to multipath variability in NLoS. Additionally, a classifier is introduced in the latent space that distinguishes between LoS and NLoS samples, applying physics-based loss only to the former to avoid overburdening learning in complex scenarios. The use of domain incremental learning (DIL) enables generalization across NLoS environments with different scatterer distributions, achieving error reductions of up to 6° in configurations with few examples.
This type of hybrid architecture not only improves the robustness of AoA estimation but also opens the door to critical applications such as interference localization in mobile communication systems, defense, or critical infrastructures. For these solutions to be viable in production, custom software development is required that integrates artificial intelligence models with real-time signal processing. At Q2BSTUDIO, we accompany companies in creating custom applications that leverage artificial intelligence techniques to solve complex localization and detection problems. Our team also offers AWS and Azure cloud services to deploy these models at scale, as well as business intelligence services that allow visualizing and analyzing interference data using tools like Power BI.
Furthermore, the incorporation of autonomous AI agents capable of reconfiguring antennas or changing frequencies in response to interference represents a natural evolution of this technology. In environments where cybersecurity is a priority, such as protecting navigation signals or military communications, robust AoA estimation becomes a cornerstone to counter spoofing attacks. At Q2BSTUDIO, we offer AI for businesses specialized in perception and localization solutions, combining expertise in radio frequency hardware with deep learning algorithms. The key lies not only in understanding the physical problem but also in implementing systems that operate reliably under real conditions, where multipath and interference are the norm, not the exception.
