Abstract: Existing random load identification methods generally suffer from ill-posedness near structural natural frequencies; that is, when the excitation spectrum overlaps with the natural frequencies, significant identification errors tend to occur. To address this issue, this study aims to improve the accuracy of random excitation spectrum identification by proposing a data-driven deep learning approach.First, a deep neural network model is constructed to establish a nonlinear mapping between response data and excitation spectra, forming a complete framework for random load identification. Subsequently, random excitation experiments are conducted on a fiber-reinforced polymer (FRP) beam model to obtain response data under multiple operating conditions, which are used for model training and validation. Finally, the proposed method is systematically evaluated through comparative analysis under different excitation scenarios. The results demonstrate that the proposed method achieves high-precision identification of random excitation spectra, with overall errors below 5% and correlation coefficients reaching 0.98. Moreover, the identification accuracy shows no significant degradation near the structural natural frequencies. It is concluded that the proposed deep learning approach effectively alleviates the ill-posedness inherent in traditional methods, significantly enhancing the stability and accuracy of random load identification, and providing a promising new pathway for inverse analysis of random loads in complex structures.