Anomalous drone detection and localization with a single acoustic vector sensor

Proposed in this work is a method to detect and locate an anomalous drone by passively receiving and processing the drone’s noise with a single acoustic vector sensor on the ground. Specific spatio-temporal-frequency features of sound pressure and particle velocity are extracted from the multichannel noise data, formulating the designed input feature vector to train a machine learning model for anomaly detection. The detected anomalous drone is then located by array signal processing with a single sensor. In the field tests, a drone hovers in eight distinct directions with respect to the sensor. To evaluate the proposed algorithm, the artificial anomaly sounds are broadcasted by a loudspeaker carried by the drone. More practically, the noise from a drone with two fractured blades is also captured and processed, which is indistinguishable from the normal drone noise to human hearing. Experimental results show that the proposed algorithm achieves good performance at both anomaly detection (with an F1 score over 90%), and drone localization (with an average angular estimation error below 9deg).