An Acoustic Approach to Drone Identification Using Machine Learning is an open-access Naval Postgraduate School work by Rachel M. Kenagy.
The increasing use of drones in contested areas around the globe highlights the critical need for rapid and precise drone detection, localization, and identification. The miniature near-resonant MEMS acoustic vector sensor developed at NPS can detect drones and determine the direction of incoming sound but cannot identify the sound’s source. To address this limitation, this research aims to enhance the sensor with a real-time identification system utilizing machine learning (ML).
For this study, the author used a dataset comprising 350 audio files of drones and background sounds. These files were pre-processed into overlapping one-second segments, totaling 22,488 samples. The author extracted 27 features from each sample, including signal and frequency statistics and mel-frequency cepstral coefficients. The author employed correlation and neighborhood component analysis to improve ML performance and reduce the feature set from 27 to 13.
The author trained and tested 32 ML algorithms using standard MATLAB parameters to establish baseline performance across various models. Six algorithms achieved over 90% accuracy with acceptable prediction speeds. Further optimization did not enhance accuracy or speed. Based on their performance, accuracy, and potential for extension to multi-class classification, we recommend integrating a simple neural network with the rectified linear unit (ReLU) activation function and the ensemble bagged trees algorithm with the sensor.
Publication Date– June 2024
An Acoustic Approach to Drone Identification Using Machine Learning contains the following major sections:
- Introduction
- Background
- Methods
- Results
- Conclusion
Post Image Credit- Adobe Stock by EnelEva (made with AI)
