【摘要】 人工智能赋能卫星网络构建的“天感天算”模式，对广域遥感监测、应急快速响应及智慧城市运营等具有重要的战略意义。然而，受星地感知距离的客观约束，传统卫星直接感知面临精细化目标识别能力不足等瓶颈。针对此挑战，提出星地协作高能效智能推理方法，通过地面分布式设备实现目标近距离感知与特征提取，再将轻量化特征数据上传到卫星进行聚合推理。利用人工智能模型的内在鲁棒性，主动容忍“感知-通信”链路中的信息失真，从而在保障推理精度的前提下提高系统能效。首先，构建了“感知-通信”联合失真模型，量化推理性能与感知误差、传输误差的映射关系；其次，建立置信区域与传输误码率、协作设备数量的数学关系；最后，在此基础上设计优化问题，并在满足鲁棒性推理精度约束下，最小化地面设备总功率。仿真结果表明，相比于传统卫星直接感知方案，所提方法在保证同等推理精度下可显著提升系统能效。更多还原

【Abstract】 The “sky-sensing sky-computing” paradigm enabled by artificial intelligence in satellite networks holds significant strategic importance for wide-area remote sensing, emergency rapid response, and smart city operations. However, traditional satellite-based direct sensing faces bottlenecks in fine-grained target recognition capability due to the inherent constraint of satellite-ground sensing distance. To address this challenge, we propose an energy-efficient collaborative inference method for satellite-terrestrial networks, where ground-based distributed devices perform near-field target sensing and feature extraction, and then upload lightweight feature data to satellites for aggregated inference. By leveraging the inherent robustness of AI models, the system actively tolerates information distortion in the “sensingcommunication” link, thereby improving energy efficiency while maintaining inference accuracy. First, we establish a joint “sensing-communication” distortion model that quantifies the mapping relationship between inference performance and sensing error as well as transmission error. Second, we derive the mathematical relationship between confidence region and transmission bit error rate, as well as the number of collaborative devices. Finally, based on this foundation, we formulate an optimization problem to minimize the total power consumption of ground devices subject to robust inference accuracy constraints. Simulation results demonstrate that compared with conventional satellite direct sensing schemes, the proposed method significantly improves system energy efficiency while guaranteeing equivalent inference accuracy.更多还原