Multi-source ensemble transfer learning-based unmanned aerial vehicle flight data anomaly detection with limited data: From simulation to reality

Abstract

Flight data anomaly detection is critical for ensuring the safety and reliability of unmanned aerial vehicles (UAVs). Traditional deep learning methods excel when sufficient data is available, but their performance significantly diminishes in data-scarce scenarios. Transfer learning is a promising solution; however, the performance of single-source transfer methods is often limited when there is a significant discrepancy between the source and target domains. This paper proposes a multi-source ensemble transfer learning-based anomaly detection (MSETL-AD) framework, aiming to transfer knowledge from multiple simulated domains to a real domain for anomaly detection in UAV flight data with limited data. First, a similarity calculation method based on dynamic time warping (DTW) is utilized to select simulated source domains that are similar to the target domain to mitigate the negative transfer problem. Second, a modeling strategy based on long short-term memory with attention mechanism (LSTM-AM) integrating transfer learning and fine-tuning techniques is proposed, which constructs a fundamental LSTM-AM prediction model for each source domain and then fine-tunes it using limited data in the target domain during the transfer process. Then, a similarity-based transfer weight assignment method is designed to guide multi-source domains for integration. Next, a similarity-guided dynamic threshold calculation method based on extreme value theory with residual smoothing is introduced to overcome random noise interference and realize adaptive anomaly detection. Finally, the effectiveness of the proposed method is validated through experiments using multiple simulated UAV flight datasets as the source domains and a real UAV flight dataset as the target domain.