Abstract: This paper proposes an improved Wave-U-Net-based blind source separation algorithm that integrates neural networks to address the issues of algorithmic incompatibility and low computational efficiency in traditional blind source separation—problems arising from long-term non-stationary mixed signals and potential statistical dependence among source signals. First, the original Wave-U-Net is restructured into a feature extraction module, and an encoder–decoder architecture is constructed. Second, a two-stage skip connection mechanism is introduced to form a nested Wave-U-Net model. Third, high-quality training signals are generated using a Bounded Component Analysis method based on Multivariate Nonlinear Chirp Mode Decomposition, followed by targeted training to obtain separated source signals. Simulation results demonstrate that the proposed algorithm accurately separates statistically dependent, long-term non-stationary signals and achieves superior accuracy and performance compared with conventional blind source separation methods. Finally, the algorithm is applied to external flight signals of helicopters, successfully separating the main rotor and tail rotor components—providing a critical data foundation for subsequent fault localization and diagnosis of helicopter rotor systems.