Loading...
Artificial intelligence based adaptive interference mitigation for aviation communication systems
Citations
Altmetric:
Date
2025-12
Abstract
Radio frequency interference (RFI) increasingly affects aviation communication due to expanding wireless networks, UAV operations, and onboard electronics. Conventional mitigation relies on static filters or rule-based responses, which fail to adapt to dynamic spectral conditions. Aviation systems require uninterrupted RF communication during critical phases, demanding intelligent mitigation. Current approaches do not address the combined need for real-time detection, high classification accuracy, and adaptive response under airborne constraints. Literature offers machine learning models for RFI classification, but most struggle in dynamic environments. Classical models lack temporal context, and reinforcement learning has seen limited use in static settings. No integrated method combines detection and mitigation tailored to aviation. Prior work excludes multi-objective optimization, spectral restoration, or cumulative reward tracking. This thesis addresses these gaps through two AI-based frameworks designed for accurate classification and adaptive response in noisy spectral conditions.
The first method combines a Vision Transformer (ViT) for spectrogram classification with a Deep Q-Network (DQN) for adaptive mitigation. It achieved 92.1% classification accuracy, 12.6 dB SNR gain, and 0.38-second latency. The agent converged in 23 episodes, with stable performance across interference types and low RMSE. These results make it suitable for low-latency environments with moderate spectral complexity. It handled synthetic and real-world patterns with consistent output under short-duration signal disruptions. The second method integrates a Convolutional Spatio-Temporal Encoder (CSTE) with a Multi-Objective Reinforcement Learning (MORL) agent. It achieved 94.3% classification accuracy, 15.1 dB SNR gain, and 0.42- second latency, with convergence in 18 episodes. It generalized well under overlapping interference, maintained stable rewards, and improved spectral reconstruction. Scalarized rewards and attentionguided encoding supported reliable decision-making across variable RF conditions. ViT-RL excelled in decision speed and simplicity, suitable for systems with tight resource limits. The CSTE-MORL Shield outperformed in accuracy, SNR gain, and policy robustness. Ablation studies confirmed key contributions from attention modules and reward shaping. Confusion matrices, F1 scores, and heatmaps supported the Shield’s strength in complex interference. Each model addresses distinct operational needs based on spectral complexity.
These methods close critical gaps in aviation RFI mitigation through learning-based strategies. ViT-RL supports fast-response scenarios, while CSTE-MORL is more effective in high-noise environments. Together, they contribute to safer, more reliable aviation communication and form a foundation for intelligent RF management in future airborne systems.
Supervisor
Description
Publisher
University of Limerick
Citation
Files
Loading...
Malik_2025_Artificial.pdf
Adobe PDF, 9.27 MB
ULRR Identifiers
Funding code
Funding Information
Sustainable Development Goals
External Link
License
Attribution-NonCommercial-ShareAlike 4.0 International