STE-RS: A novel framework for robust anomaly detection in UAV flight data

Anomaly detection in flight data has become crucial for ensuring the safety and reliability of unmanned aerial vehicles (UAVs), particularly with the expanding UAVs application scenarios and increasing mission complexity. However, many existing anomaly detection methods frequently overlook inherent random noise and transient disturbances. Furthermore, these methods fail to explicitly capture the complex spatio-temporal relationships among multivariate flight data, resulting in more false negatives and false positives. In this paper, we propose a novel anomaly detection framework (STE-RS) that integrates spatial Granger causality (SGC), temporal distribution clustering (TDC), mask-enhanced (ME), and residual smoothing (RS) to effectively address these problems. Firstly, SGC models each flight variable using an independent neural network and captures the nonlinear causality among variables through sparsity constraints. TDC is designed to cluster flight data into latent distributions to handle heterogeneous temporal patterns. The mask-enhanced module leverages an attention mechanism to comprehensively capture spatio-temporal correlations. Finally, the Savitzky-Golay method is applied to mitigate the effects of random noise and disturbances. Anomaly detection is performed by comparing the smoothed residuals with a predefined anomaly threshold. Experimental results on the ALFA and Basic dataset demonstrate that STE-RS significantly outperforms several baseline methods across multiple evaluation metrics including Precision, Recall, and F1-score when facing different data distributions. Additionally, the framework supports root cause analysis, proving to be a scalable and reliable solution for UAV anomaly detection.