Deep Learning-Based Underwater Propeller Recognition Using Time-Frequency Representations of Acoustic Signals
CMES-COMPUTER MODELING IN ENGINEERING & SCIENCES, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.32604/cmes.2026.089269
- Dergi Adı: CMES-COMPUTER MODELING IN ENGINEERING & SCIENCES
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Compendex, INSPEC, zbMATH, Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest)
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Recep Tayyip Erdoğan Üniversitesi Adresli: Evet
Özet
Underwater propeller recognition supports maritime surveillance, defense, and environmental noise monitoring. Published classifiers report near-perfect accuracy on the public propeller blade-count dataset, yet all rely on segment-level cross-validation. Segments cut from one continuous recording then appear in both training and test folds, which leaks recording-specific information. This study presents a leakage-aware benchmark of deep learning for propeller blade-count recognition. Seven classifiers are re-evaluated under two protocols. The first is the established stratified segment-level protocol. The second is a blocked recording-level protocol with purging, which tests extrapolation across time. The models span four ImageNet transfer networks, a from-scratch multi-scale convolutional neural network (CNN), a raw-waveform one-dimensional CNN, and a feature-based support vector machine. Model selection uses inner validation only, and all comparisons carry paired statistical tests. Three findings emerge. First, no model reaches the published 99.80% reference level once model selection avoids the test fold. Second, the protocol changes stability far more than mean accuracy. Every spectrogram transfer model falls to 72%-79% on the first temporal block, while mean accuracy shifts by only two points. Third, the one-dimensional CNN with 0.46M parameters is the strongest and most robust model. It reaches 99.13 +/- 0.73% under the segment-level protocol and 98.27 +/- 2.89% under the blocked protocol. Gradient-weighted class activation mapping confirms that the networks read blade-pass harmonics below 1 kHz. Ablations show that log compression is essential, worth 19.8 percentage points, while augmentation gives no significant gain. The evaluation protocol, not the architecture, governs the reported accuracy on this dataset.