Reliable Ship-Radiated-Noise Recognition via Protocol-Aware Evaluation and Bayesian Fusion

Reliable ship-radiated-noise recognition depends on both model design and evaluation protocol. On DeepShip, the common recording-level split still places training-seen ship identities in the test set; we define this overlap as a contamination rate of 68.1%. A ship-level protocol reduces contamination to 0 under our metadata-based identity grouping, and a controlled comparison with the same architecture and hyperparameters shows a 7.11-percentage-point (pp) accuracy difference. We also refine the constant-Q harmonic coefficient (CQHC) by formulating envelope–harmonic separation as non-negative constrained least squares and replacing the original one-step approximation with non-negative harmonic projection and regularized closed-form back-substitution. The three inputs, whose vertical axes represent linear frequency, Mel bands, and harmonic order, are encoded separately and fused at an intermediate level by soft mixture-of-experts gating. Bayesian hybrid attention provides posterior-based predictive uncertainty. Under the DeepShip recording-level protocol, accuracy is 88.64%, exceeding input stacking and decision averaging by 3.83 and 2.50 pp. The expected calibration error decreases from 0.0724 to 0.0413, and rejection at 90% coverage reduces the error by 27.0%. At the recording level, confidence-weighted voting reaches 94.17% over 120 test recordings. Without dataset-specific tuning, ShipsEar accuracy is 98.63% with a segment-random split and 95.05% with recording-disjoint evaluation.