Abstract: Space target maneuver data serves as a foundational element for space mission planning, orbital prediction, and space situational awareness. However, within the data acquired by current space target awareness systems, maneuver events are significantly sparser compared to non-maneuvering trajectories, leading to severe class imbalance in deep learning training datasets. The imbalance impedes model convergence and limits the generalization capability of deep learning networks in maneuver detection tasks. To address the scarcity of maneuver data and the resulting data imbalance, a generative learning framework tailored for space TLE data are given in this paper. By analyzed the physical dynamics and statistical characteristics of space target maneuvers, GMM architecture are constructed. Training is conducted exclusively on labeled maneuver segments from real TLE datasets for GMM, then the space target maneuver data of TLE format are generated by trained models. Experimental results demonstrated that the generated data closely adhere to known orbital mechanics and process better training for deep learning model in Support for space missions.