ETD

• CNN
• Deep Learning
• Drone
• Keras
• LSTM
• Mel Spectrogram
• MFCC
• Spectrogram
• Tensorflow
• UAVs

In recent years, drone usage has risen significantly, increasing the potential for them to pose a threat. Due to their small size, they can be difficult to detect for many reasons. This thesis, conducted at Lambda Laboratories at MBDA Italia in La Spezia, focuses on analyzing and designing acoustic detection algorithms for unmanned aerial vehicles (UAVs).
The process involved deep learning to construct and train neural networks for drone acoustic detection. A dataset of 2496 drone and non-drone sounds lasting one second each was collected, primarily using a microphone. To build a larger dataset for the network, custom drones were designed and built.
A spectral analysis was performed to identify the main frequency bands of drone sounds and reduce noise components. Features such as MFCCs, spectrograms, and Mel spectrograms were extracted from the data and used to train CNN and LSTM neural networks.
The best neural network was then selected and used to predict the presence of drones in real-time by recording one second audio frames.
The obtained results were positive, especially from the use of LSTM neural networks with MFCCs as features, paving the way for future developments.