Self-Supervised Learning with Variational Autoencoders for Anomaly Detection in Autonomous Drone Fleets

However, Unmanned Aerial Vehicles (UAVs) are increasingly being employed for different applications such as surveillance, logistics, disaster response, and smart city management, and thereby, the requirement for robust detection of anomalies and path optimization techniques has arisen. Existing traditional UAV monitoring systems typically depend on anomaly detection based on a rule or an unsupervised clustering method, i.e., DBSCAN and One-Class SVM, balancing high false positives and slow response time, correction, and suboptimal trajectory. To tackle these issues, we develop an AI-powered UAV anomaly detection and adaptive navigation framework where Isolation forests and MDPs are used to predict UAV trajectory, and DQNs are used for adaptive navigation. Experimental results in a simulated smart city arena with a multi-UAV fleet show that our proposed framework achieves 94% accuracy, 92% precision, and 91% recall, which are much superior to the existing approaches, such as One Class SVM (87% accuracy) and DBSCAN (80% accuracy). Additionally, the system compensates for false positives by 30%, improves energy efficiency by 15%, and reduces UAV collision risks by 20%, with high compatibility with real-time UAV fleet management. In the future, the scalability and the adaptability of the UAV-based smart city operation will be furthered through federated learning-based decentralised UAV coordination, reinforcement learning for autonomous swarm intelligence, and hybrid optimisation techniques for improving energy efficiency.