DEMON style neural networks front-end features for passive sonar classification

다국어 초록 (Multilingual Abstract)

This study proposes a novel neural network front-end feature based on conventional sonar signal processing. It simplifies the extraction of Detection Envelope Modulation On Noise (DEMON)gram, a method used in passive sonar signal processing, by implementing it with two consecutive Short-Time Fourier Transform (STFT) operations. This converts the 1-dimensional sonar signal into a 2-imensional feature that can effectively capture the frequency modulation characteristics of cavitation generated by propellers. This DEMONgram-based frontend feature, when combined with conventional Mel spectrogram-based features in audio classification, can demonstrate higher performance. Experimental results on the ShipsEar dataset show that the proposed method achieves an accuracy of 81.0 %, a 5.8 % point improvement over the conventional Mel spectrogram-based features, thus demonstrating its effectiveness in passive sonar signal classification tasks.