Introduction: Reliable identification of vessels from underwater acoustic signatures is relevant to navigation safety, maritime-domain awareness, and coastal protection, particularly in ports, chokepoints, and environmentally sensitive littoral areas where non-cooperative or anomalous traffic may pose operational risks. This study addresses recording-level recognition of underwater engines from broadband hydroacoustic data, aiming to infer engine presence, motor identity, operating condition, and composite motor–condition identity (MCID) labels under nuisance noise and transient events.
Methods: We used the Wolfset dataset, composed of 168 single-channel waveform audio file (WAV) recordings at 44.1 kilohertz (kHz), including absent, single-engine, and two-engine cases. The main experiments retained absent and single-engine recordings with WAV-grouped train, validation, and test splits. Each recording was segmented into overlapping 2-second windows and represented by Spec2, a compact two-channel input combining Low-Frequency Analysis and Recording (LOFAR)-like spectral cues with Detection of Envelope Modulation on Noise (DEMON)-like envelope-modulation information. We evaluated convolutional neural network (CNN), convolutional recurrent neural network (CRNN), and residual network (ResNet) baselines and proposed Spec2 Band-SoftMoE, a DEMON-gated dense soft mixture-of-experts (SoftMoE) model with contiguous frequency-band specialization and multitask heads for presence, motor, type, condition, noise, and transient labels.
Results: The proposed model achieved the best WAV-level composite score with four specialized bands, reaching 86.8%, with motor macro-F1 score of 94.0%, condition macro-F1 score of 96.0%, MCID macro-F1 score of 89.0%, and MCID accuracy of 90.0%. LOFAR-only and DEMON-only ablations confirmed that spectral and modulation cues are complementary. On two-engine out-of-domain (OOD) mixtures excluded from training, an auxiliary motor-bits head reached 61.9% macro-F1 score.
Conclusions: Moderate frequency-band specialization improved underwater engine recognition and reduced motor-condition confusions. The results support acoustic artificial intelligence (AI) as a tool for vessel identification, navigation support, and coastal surveillance, while motivating future mixture-aware models and adaptation to Distributed Acoustic Sensing (DAS) data.
Conflict of Interest: The authors declare no conflicts of interest.