The increasing prevalence of drones has prompted a critical need for effective counter-drone systems. While ground-based methods have limitations, this research explores the potential of drone-based acoustic detection. By leveraging aerial platforms, we aim to overcome challenges associated with ground-based systems and develop a more robust and adaptable counter-drone solution. This paper presents a novel approach combining advanced signal processing, data augmentation, and deep learning to accurately detect drones through acoustic emissions. Our experimental results demonstrate superior precision compared to existing methods that employ similar techniques. To support our research, a comprehensive dataset of real-world UAV acoustic signatures has been collected.
Our work consisted of several key components: Data Collection using a simple Bluetooth earbud attached to a drone to capture audio recordings with and without an intruder drone; Data Preprocessing including data segmentation, normalization, and noise augmentation; Feature Extraction using Short-Time Fourier Transform (STFT) to generate spectrograms; Training multiple deep learning architectures, including fully connected networks, convolutional neural networks (CNNs), and transfer learning with pre-trained models; Hyperparameter Optimization using Optuna to fine-tune our models and preprocessing parameters; Creation of a Real-time system capable of detecting intruder drones from an aerial platform.
This comprehensive approach allows us to develop a flexible system capable of detecting the presence of intruder drones in real time.