Akademik Veri Yönetim Sistemi
6th International Conference on Innovative Academic Studies, Konya, Türkiye, 12 - 13 Mart 2025, ss.376-383, (Tam Metin Bildiri)
Propeller sound classification plays a vital role in applications such as underwater vehicle navigation, condition monitoring, and marine surveillance. Deep learning techniques, particularly convolutional neural networks (CNNs), have emerged as powerful tools for analyzing complex acoustic signals. This study presents a deep learning-based approach for propeller sound classification using GoogLeNet, leveraging mel-spectrogram transformations to enhance feature extraction. The methodology involves preprocessing acoustic signals, converting them into spectrogram images, and training a GoogLeNet-based CNN model. Experimental results demonstrate the effectiveness of this approach, achieving a classification accuracy of 0.997 on the training dataset and 0.947 on the test dataset. Spectral analysis reveals that higher-frequency variations play a crucial role in distinguishing between different propeller blade configurations. The findings indicate that GoogLeNet’s multi-scale feature extraction and computational efficiency make it well-suited for propeller sound classification tasks. This study provides a foundation for real-time sound-based diagnostics in both military and civilian applications, with potential extensions through data augmentation and alternative deep learning architectures.