Application and Interpretable Research of Capsule Network in Situational Understanding

In the context of multi-agent collaborative adversarial scenarios, the accurate and rapid assessment of situations is a crucial prerequisite for unmanned clusters to achieve autonomous decision-making. Leveraging deep learning techniques, multi-agent systems can achieve precise understanding of complex situations. However, the inherently non-interpretable black-box structure of deep learning makes it challenging to apply in domains with stringent security requirements. In this paper, we propose a threat situation classification network based on Capsule Networks to categorize different scenario situations, and conduct a comprehensive analysis of the interpretability of this network. The network introduces a novel convolutional 'Flatten Layer' to ensure that feature capsules are distributed within planes that maintain the same relative spatial relationships as the input image. This establishes the characteristic plane matrix heatmaps and the characteristic volume matrix heatmaps, which, along with the coupling coefficient matrix heatmaps, collectively demonstrate the network's sparse interpretability during the classification process. Experimental results show that the proposed network can effectively accomplish situation classification tasks while maintaining interpretability, providing insights for research in situation understanding in domains with high-security requirements.