Research on the acoustic-vibration response of underwater structures remains limited due to the prohibitive cost of experimental testing. In this study, the singular boundary method and finite element method are introduced to establish the numerical computational model for the acoustic-vibration response of underwater structures. However, the analysis of the probability characteristics of the acoustic-vibration response of underwater structures considering a random ocean environment is challenging. To overcome this, this paper extends the direct probability integral method (DPIM) [1], as a novel stochastic acoustic-vibration analysis method to address this issue. Furthermore, given the lack of sufficient statistical data for random factors and the high computational expense of the singular boundary method-finite element model, a back-propagation neural network improved by particle swarm optimization (BPNN-PSO) model [2] for predicting the stochastic acoustic-vibration response is established. Leveraging this model, a novel machine Learning-based DPIM is proposed. In the numerical example, the calculated results are compared with those of Monte Carlo simulation, which demonstrates the high accuracy of the BPNN-PSO-based DPIM. The results show the effects caused by different random factors of the ocean environment on the acoustic-vibration response of underwater structures.