As a critical structure in hydraulic engineering, the seepage characteristics of earth-rock cofferdams are vital to engineering safety. However, traditional deterministic models and statistical approaches face considerable challenges when dealing with high-dimensional, nonlinear, and dynamically changing seepage data. To address this issue, a momentum-accelerated stochastic conjugate gradient algorithm (SCGMA) based on nonconvex stochastic optimization problems is presented. By integrating the advantages of stochastic gradient descent (SGD) and conjugate gradient methods, the algorithm incorporates momentum acceleration technology and a modified secant equation, ensuring that the search direction has two key features: sufficient descent and trust-region. Additionally, it leverages both gradient and function value information, which collectively improve its adaptability and global convergence performance. Numerical results demonstrate that SCGMA excels in machine learning (ML) tasks, significantly accelerating convergence speed and improving solution accuracy. This provides an efficient and precise optimization tool for seepage analysis in earth-rock cofferdams, contributing to enhanced precision and efficiency in hydraulic engineering safety monitoring.