| 摘要: |
| 针对航天测控数传资源预留任务中约束复杂、需兼顾各用户间公平性与决策可解释性的挑战，提出了一种融合TabNet与注意力机制的深度学习方法。该方法首先利用TabNet分别提取用户、资源和时序特征的深层表示，并同步输出特征重要性评分以增强模型可解释性；进而设计以资源为枢纽的用户-资源-时序三路注意力编码机制，实现用户特征、资源特征和时序特征的深度融合；最后将资源预留问题建模为马尔可夫决策过程，构建层次化注意力策略网络，依序确定资源、用户、开始时间与结束时间，从而生成完整的资源预留方案。仿真实验结果表明，相比传统机器学习算法，所提方法在资源利用率与加权用户满意度上分别提升23.86%与19.92%，消融实验进一步验证了注意力机制在均衡用户比例与捕获时序规律方面的关键作用，特征重要性排序与注意力权重的可视化也为决策过程提供了清晰依据。 |
| 关键词: 航天测控 数传资源 资源预留 深度学习 注意力机制 |
| DOI：10.20079/j.issn.1001-893x.260112005 |
|
| 基金项目: |
|
| A Space TT&C Data Transmission Resource Reservation Method Based on Deep Learning |
| DING Heng |
| (Southwest China Institute of Electronic Technology，Chengdu 610036，China) |
| Abstract: |
| For the challenges in space TT&C data transmission resource reservation，which involves complex constraints and the need to balance fairness among users and decision-making interpretability，a deep learning method integrating TabNet and attention mechanism is proposed.Firstly，the method uses the TabNet to extract the deep representation of user，resource and time series features respectively，and simultaneously outputs the feature importance score to enhance the model interpretability.Then，a three-way attention mechanism of user-resource-temporal is designed to realize the deep integration of user features，resource features and temporal features.Finally，the hierarchical attention strategy network is constructed by modeling the resource partitioning as a Markov decision process，and the complete resource reservation scheme is generated by determining the resource，user，start time and end time in sequence.The simulation experiments results demonstrate that，compared with traditional machine learning algorithms，the proposed method achieves 23.86% improvement in resource utilization and 19.92% enhancement in weighted user satisfaction.The ablation experiments further validate the pivotal role of the attention mechanism in balancing user proportions and capturing temporal patterns.The visualization of feature importance rankings and attention weights also provides a clear basis for decision-making processes. |
| Key words: space TT&C;data transmission resource resource reservation deep learning attention mechanism |
《电讯技术》编辑部 