A Novel Transformer-Based Multivariate Spatio-Temporal Feature Fusion Method for UAV Actuator Anomaly Detection

The actuator plays a crucial role in controlling the flight attitude of unmanned aerial vehicles (UAVs), making timely anomaly detection essential for operational reliability and flight safety. However, existing methods often have difficulty jointly modeling the complex temporal dynamics and inter-variable spatial dependencies of multivariate actuator signals, while their computational complexity can limit real-time deployment. Moreover, most existing approaches rely on univariate or weakly coupled representations, making them less effective in detecting simultaneous faults across multiple actuators. To address these challenges, this paper proposes a Multivariate Spatio-Temporal Feature Fusion Transformer (STF_Tran) framework for anomaly detection in fixed-wing UAV actuators. Unlike conventional transformer-based multivariate anomaly detection methods that employ a shared attention mechanism to model heterogeneous dependencies, STF_Tran adopts a dual-branch architecture that separately encodes temporal dynamics and spatial correlations from multivariate actuator signals. A self-learning mechanism enables each branch to learn discriminative representations directly from normal operating data without requiring explicit fault labels, while a feature fusion module integrates the complementary spatio-temporal representations for anomaly reconstruction and scoring. Faults are identified by comparing the resulting anomaly scores with predefined thresholds, enabling the detection of diverse and simultaneous actuator anomalies. Experimental results demonstrate the effectiveness of STF_Tran, achieving F1 scores of 0.9944 and 0.9949 for deviation and stuck anomalies, respectively, and consistently outperforming several state-of-the-art methods.