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.

The accuracy and efficiency of the equivalent source method (ESM) for underwater radiated-field prediction depend strongly on the position and number of the equivalent sources. The conventional uniform distribution (UD) is fixed before the field is observed, and the coherence-based optimum configuration (COC) ranks candidates by local pairwise correlations, blind to the singular directions the field excites. We propose the leverage-based effective-subspace configuration (LESC): it screens the singular directions the measured pressure excites, ranks candidates by leverage scores within that subspace, and prunes redundant sources by a coherence threshold, solving the strengths by cross-validated Tikhonov regularization. In free-space simulations of a baffled circular piston and a dual-hotspot planar source, LESC uses fewer, position-adaptive sources while cutting the 5000 Hz error by 73–$83\%$ compared with UD and by 73–$74\%$ against COC, and maintains an error of $2.33\%$ at 6300 Hz, where UD and COC degrade sharply, at one-third to one-half of COC’s configuration cost. An ablation identifies the effective singular-direction screening as the decisive ingredient. The ordering is reproduced at matched source counts and under colored and spatially correlated noise, reversing only where the source count is small relative to the excited-direction count. The study is confined to free-space radiation from planar sources, within which these advantages are established.