Flight data anomaly detection is crucial for ensuring the flight safety of unmanned aerial vehicles (UAVs). However, the complex spatiotemporal correlation and the interference of random noise in flight data, coupled with the high cost associated with labeling anomaly data, present significant challenges to the accuracy of anomaly detection models. This study proposes an attention-based encoder-decoder hybrid network with adaptive thresholds (AEDHN-AT) unsupervised data-driven method for critical parameter anomaly detection of UAV flight data. First, a novel reconstruction-based model, AEDHN, is proposed. This model first utilizes bidirectional long short-term memory and one-dimensional convolutional neural network (BiLSTM-1D CNN) as the feature extractor to achieve multi-level feature extraction from UAV flight data. Building upon this, an attention mechanism is introduced to guide the feature extractor in selectively learning crucial information and features. Second, a dynamic threshold calculation method based on support vector regression (SVR) with residual smoothing is proposed to obtain dynamic thresholds. Finally, the effectiveness of the proposed method is verified through extensive experiments on real UAV flight data. The experimental results demonstrate that the proposed method exhibits superior anomaly detection performance compared to the baseline methods.

Unmanned aerial vehicle (UAV); flight data; anomaly detection; adaptive thresholds; feature extraction