IDENTIFICATION OF SOUNDS BASED ON THE HILBERT-HUANG TRANSFORM FOR THE TASK OF DETECTING UAVs

Abstract

This paper investigates the identification of drone acoustic signatures via the Hilbert–Huang Transform combined with machine- and deep-learning techniques. After high-pass filtering at 80 Hz, each 3-s recording is split into 25 ms frames with 50 % overlap; every frame undergoes DCT pre-whitening, HHT processing, and extraction of 13 MFCCs. The resulting averaged feature vectors—drawn from 2 075 yes_drone and 266 unknown samples—are fed into a lightweight two-layer perceptron (64 + 32 ReLU units, sigmoid output). On a held-out test set the network reaches an overall accuracy of 0.95, with a drone recall of 0.99 and an F1-score of 0.78 for the unknown class. These results outperform a baseline MFCC + SVM system and approach the performance of deeper CNN architectures, while remaining computationally suitable for real-time embedded deployment. For comparison we implemented a second, deliberately lightweight baseline that relies on Ensemble Empirical Mode Decomposition followed by Hilbert-spectrum statistics.