WTNET: DIRECTION-ASSISTED DRONE ANOMALY DETECTION

The safety of drones is critical for expanding applications in logistics, agriculture, and disaster response. This paper proposes a novel fusion strategy WTNet, combining TASTgramMFN and WaveNet networks. Building upon the TASTgramMFN architecture, our approach integrates temporal dynamic modelling, spectral analysis, and attention mechanisms to achieve acoustic-based drone anomaly detection. Unlike traditional binary classification approaches, the proposed framework adopts an 18-class fine-grained state modelling strategy-incorporating flight direction and fault types-to address acoustic variability under normal operations. Experiments conducted on the ICSV31 Challenge dataset demonstrate that the proposed method WTNet outperforms both the baseline WaveNet (60.72%) and the standalone TASTgramMFN (91.65%), achieving an AUC of 92.14%.