Evaluating unmanned aerial vehicle (UAV) survivability is crucial when UAVs are required to perform missions in hostile areas. There are complex spatiotemporal interactions among entities in hostile areas; therefore, evaluation of the survivability of a UAV flying along a specific route needs to effectively fuse spatiotemporal information. It is difficult to clarify how information is fused and how threats accumulate along the route. We present a novel solution for building a learnable evaluation model that can extract the required knowledge directly from the data. In this approach, hostile scenarios are decomposed into various threat entities, threat relations (TRs) and UAVs, where a TR is the relation between a threat entity and a UAV. We propose a data-driven evaluation model named the sequential threat inference network (STIN), which can learn TRs and perform spatiotemporal fusion to evaluate survivability. We validate the model in multiple scenarios that contain threat entities of different types, quantities and attributes. The results show that the STIN is superior to the baseline models in various situations. Specifically, the STIN can automatically generalize learned knowledge to scenarios with different numbers of threat entities without retraining. In the generalization experiment, the error increases little when the STIN is directly used in the new scenarios where the number of entities is larger than in the training scenarios.

当需要无人机在敌对地区执行任务时，评估无人机的生存能力至关重要。在敌对地区的实体之间存在复杂的时空相互作用。因此，对沿特定航线飞行的无人机的生存能力进行评估需要有效融合时空信息。很难弄清楚如何融合信息以及沿途的威胁如何累积。我们提出了一种新颖的解决方案，用于构建可学习的评估模型，该模型可以直接从数据中提取所需的知识。在这种方法中，敌对情景被分解为各种威胁实体，威胁关系（TR）和UAV，其中TR是威胁实体与UAV之间的关系。我们提出了一种数据驱动的评估模型，称为顺序威胁推断网络（STIN），可以学习TR，并进行时空融合以评估生存能力。我们在包含不同类型，数量和属性的威胁实体的多种情况下验证模型。结果表明，在各种情况下，STIN均优于基线模型。具体来说，STIN可以自动将学习到的知识概括为具有不同数量威胁实体的情况，而无需重新培训。在一般化实验中，当实体数量大于训练场景的新场景中直接使用STIN时，误差几乎不会增加。结果表明，在各种情况下，STIN均优于基线模型。具体而言，STIN可以自动将学习到的知识概括为具有不同数量威胁实体的情况，而无需重新培训。在一般化实验中，当实体数量大于训练场景的新场景中直接使用STIN时，误差几乎不会增加。结果表明，在各种情况下，STIN均优于基线模型。具体而言，STIN可以自动将学习到的知识概括为具有不同数量威胁实体的情况，而无需重新培训。在一般化实验中，当实体数量大于训练场景的新场景中直接使用STIN时，误差几乎不会增加。