Answers are grounded in this paper and cite their sources.
Open-set recognition of ship-radiated noise is challenging when unknown target classes and propagation-channel mismatch occur together. We propose a framework that combines supervised contrastive training on paired channel views with Fused8, a logistic-regression gate that integrates classifier logits and four derived statistics for unknown rejection. We evaluated the framework on ShipsEar using a recording-disjoint five-fold protocol. With paired-view training, Fused8 achieved an overall accuracy (OA) of 66.00% on measured recordings and 59.93% on recordings filtered with a held-out simulated channel response. The corresponding no-rejection baselines achieved 56.83% and 51.54%. Compared with a logits-only gate fitted using the same samples, Fused8 improved OA by 4.03 percentage points on the measured recordings and 3.27 percentage points on the simulated target channel. These results show that adding the derived statistics improves rejection-based recognition in both conditions and that the fusion advantage persists under the simulated channel mismatch evaluated here.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: