Explainable Adaptive Path Reconfiguration for Obstacle Avoidance in UAV Networks Using Transformer-Based Predictive Modeling and Bio-Inspired Optimization

초록

Critical challenges for uncrewed aerial vehicle (UAV) networks in flying ad-hoc networks (FANETs) operation include collision avoidance, energy efficiency (EE), and decision transparency in highly dynamic aerial conditions. This article proposes an explainable adaptive path reconfiguration (EAPR) framework by integrating Transformer-based trajectory prediction, secretary bird optimization algorithm (SBOA), and SHAP-driven interpretability, enabling predictive, self-aware UAV navigation. The Transformer model predicts future UAV and obstacle trajectories under multihead self-attention (MHSA) to proactively reconfigure before any potential collision risk. Then, the SBOA further optimizes the trajectories by minimizing a multiobjective cost function that jointly considers collision risk, energy consumption, and path deviation. It estimates the quantitative contribution of key features-distance, velocity, and orientation-to each decision made, guaranteeing human-interpretable autonomy. Extensive simulations for diverse UAV densities and obstacle environments prove that EAPR achieves 96.2% collision avoidance, 82.8% EE, and 0.128 path deviation with improvements of 15%-25% over particle swarm optimization (PSO), fire hawk optimization algorithm (FHOA), and Portia spider optimization algorithm (PSOA). Robustness tests further confirm resilience in case of communication loss and dense obstacle scenarios. This article establishes a unified paradigm of predictive, optimized, and explainable UAV swarm navigation for safe and transparent autonomous operations in next-generation aerial networks.

키워드