Abstract: With the rapid development of deep-sea resource exploitation and marine engineering, underwater structures face increasingly complex vibrational and acoustic environments. Consequently, investigating the vibroacoustic coupling responses of these structures under complex stochastic excitations becomes critically important. Traditional deterministic analysis methods are inadequate for accurately characterizing the influences of multiple uncertainties inherent in real marine environments, such as random loads, material parameters, and geometric deviations, on acoustic performances. To address this, a hybrid numerical method integrating physics-informed kernel function neural networks (PIKFNNs) with the direct probability integral method was proposed for the efficient uncertainty quantification of structural random vibroacoustic responses in marine environments. Specifically, the near-field sound pressure, initially computed via the finite element method, was used as training data to drive the physics-informed kernel function dependent on system stochastic parameters, such as the sound speed, the density, and the excitation force, to learn the acoustic field mapping. Subsequently, the predicted sound pressure treated as a random variable was then employed to solve the corresponding probability density integral equation, thereby to achieve high-precision prediction of the statistical characteristics of the sound pressure level. The effectiveness of the proposed method was validated through numerical examples. Furthermore, the influence of stochastic parameters on acoustic performance was investigated in scenarios involving unbounded sea areas and deep-sea settings. This method provides a novel technical tool for the acoustic reliability assessment of complex marine equipment and for research concerning underwater multi-structure interaction problems.