Meta-learning-aided generalized anomaly detection for unmanned aerial vehicles from simulation to unseen reality

Abstract

Anomaly detection is critical to the safety and reliability of unmanned aerial vehicles (UAVs). However, traditional deep learning methods rely on independent and identically distributed (i.i.d.) assumptions and are susceptible to data distribution variations, while domain adaptation approaches are constrained by privacy and data collection costs, making it difficult to obtain target data in advance or may be inaccessible. This paper introduces a novel meta-learning-aided generalized anomaly detection (Meta-GAD) framework, which harnesses knowledge acquired from multiple simulated domains to enable robust UAV anomaly detection in real-world scenarios. First, a local–global feature joint modeling method based on one-dimensional convolutional neural network and convolutional block attention module (1D CNN-CBAM) is constructed, which leverages 1D CNN for extraction of local features and adaptively fuses local–global information via CBAM’s channel- and spatial-attention mechanisms, enhancing the model’s ability to fit complex UAV flight data. Second, a model-agnostic meta-learning (MAML) approach with a dual-gradient optimization strategy is designed, leveraging the 1D CNN-CBAM model as the base learner to learn domain-invariant representation via two gradient updates in inner-outer loops. Then, an adaptive detection strategy integrating anomaly feature enhancement and extreme distribution modeling is introduced to improve the performance of anomaly detection. Finally, the efficacy of the proposed framework is validated through model training on multiple simulated flight datasets and model testing on an unseen real flight dataset.