Multi-Task U-Net with Class-Conditional Mahalanobis Scoring for Unsupervised Drone Anomaly Detection

Authors:

Abstract
With the growing use of drones, the importance of early detection of abnormal conditions has increased. However, in real-world scenarios, it is difficult to collect sufficient abnormal data in advance, making unsupervised anomaly detection approaches essential. Conventional autoencoder (AE)-based methods typically use reconstruction error as the anomaly score, but their detection performance degrades when abnormal samples are reconstructed as normal. To address this limitation, this paper proposes a multi-task U-Net that simultaneously performs reconstruction and classification. By applying structured loss functions—Supervised Contrastive (SupCon), Triplet, and Center losses—the proposed model shapes the latent representation into class-conditional Gaussian distributions and computes anomaly scores based on the Mahalanobis distance. Furthermore, the integration of Squeeze-and-Excitation (SE) blocks and an adaptive noise layer enables robust representation learning even under low signal-to-noise ratio (SNR) conditions. Experimental results on the ICSV31-AI Challenge drone acoustic dataset demonstrate that the proposed model achieved a macro Area Under the Curve (AUC) of 0.9630, representing over a 59\% relative improvement compared to the baseline WaveNet AE (0.6059) and significantly outperforming the single-task U-Net (0.5620). These results confirm that the proposed method enables precise, class-aware anomaly detection under unsupervised settings and holds strong potential for applications beyond drones, including various types of industrial equipment.

This publication has 9 references indexed in Scilit: