转动惯量是航天器姿态控制系统中的关键参数，对转动惯量的精准、快速辨识是实现高精度航天器姿态控制的前提。然而，在实际航天任务中，转动惯量往往呈现强非线性高动态的时变特征，为转动惯量辨识带来了巨大困难。本文面向自主、高效的在轨辨识需求，针对传统机理模型方法建模精度受限，数据驱动方法对数据质量要求高、对分布外场景敏感等问题，提出一种数据与机理融合的转动惯量辨识与在轨更新方法。首先，针对因燃料消耗和附件展开引起的航天器质心偏移，建立了转动惯量时变特性的物理模型，并结合域随机化方法生成了高保真的训练数据。其次，本文提出一种融合Bi-GRU与注意力机制的时变转动惯量辨识模型，并基于观测扰动构建的噪声对抗训练机制，实现了复杂噪声条件下的稳定辨识。然后，基于伪标签监督和物理约束损失设计了星上模型的自主更新策略，实现了离线模型的轻量化在线参数更新。最后，将数据模型与机理模型融合的辨识策略嵌入姿态闭环控制回路，形成了由控制回路到数据池、由数据池到辨识回路、再由辨识回路到控制回路的星上自主工作模态。仿真结果表明，本文所提方法能将转动惯量辨识误差控制在1%以内，在有限数据存储与信息时延影响下的姿态控制精度优于1°。

The moment of inertia is a critical parameter in spacecraft attitude control systems, and its accurate and rapid identification is essential for achieving high-precision attitude control. However, in practical space missions, the moment of inertia often exhibits strongly nonlinear and highly dynamic time-varying characteristics, posing significant challenges for accurate estimation. To meet the demands for autonomous and efficient on-orbit identification, this paper proposes a hybrid data–physics-driven approach for moment of inertia estimation and in-orbit model updating. First, a physics-based model is established to capture the time-varying characteristics of the moment of inertia caused by fuel consumption and appendage deployment, and high-confidence training data are generated using a domain randomization technique. Second, a novel Bi-GRU and attention-enhanced neural network is proposed for accurate identification of time-varying inertia, and a noise-adversarial training scheme is designed based on observation disturbance modeling to ensure robustness under complex noise conditions. Then, an onboard lightweight model update strategy is developed, incorporating pseudo-label supervision and physics-constrained loss to enable efficient online parameter updates of the offline model. Finally, the proposed identification strategy, integrating both data-driven and physics-based models, is embedded into the attitude control feedback loop, forming a closed-loop autonomous onboard operation paradigm linking the control module, data buffer, and identification module. Simulation results show that the proposed method achieves inertia estimation errors within 1% and attitude control accuracy better than 1° under limited data storage and communication delay conditions, demonstrating strong potential for real-world onboard applications.