一种利用Transformer的无人集群对抗态势要素识别方法

针对无人集群对抗问题中原始态势信息繁杂,难以准确识别集群阵型和集群运动趋势等态势要素的特点,为提高无人集群态势要素识别能力,设计了一种利用Transformer的无人集群对抗态势要素识别方法。基于Transformer模型的思想构建了可应用于无人集群对抗态势要素识别问题的Transformer-Decoder注意力层模型,实现良好的集群态势要素识别能力,设计层间注意力结构以改进提升了Transformer-Decoder的特征表达能力,进一步提高识别准确率。首先将集群态势序列信息输入到LSTM循环神经网络,编码成时序特征信息；然后使用Transformer-Decoder注意力模块和层间注意力模块提取集群的综合高阶态势信息,最后多维度分类网络和softmax层实现对多类态势要素的分类。实验结果表明:利用Transformer和层间注意力的无人集群对抗态势要素识别方法在态势要素分类问题上表现出良好的性能,能够同时对多类态势要素进行准确分类；相对于基线方法,利用Transformer和层间注意力的集群态势识别方法在集群阵型和运动趋势识别问题上具有更高的准确率。尤其在体现集群内部相对趋势的态势要素的分类问题上,该方法明显表现出更好的性能。

Recognition method for unmanned swarm adversarial situation elements using Transformer

Since the original situation information in the unmanned swarm confrontation problem is complicated, it is difficult to accurately identify the situation elements such as the swarm formation and swarm movement trend. In order to improve the identification ability for unmanned swarm situation elements, an identification method for unmanned swarm adversarial situation elements using Transformer was designed. On the basis of the Transformer model, a Transformer-Decoder attention layer model that can be applied to unmanned swarm confrontation situation element identification problem was constructed, so as to achieve a good ability to identify swarm situation elements. The inter-layer attention structure was designed to improve the feature expression ability of Transformer-Decoder to further improve the recognition accuracy. First, the situation sequence information of the unmanned swarm was input into LSTM and encoded into time sequence feature information. Then the Transformer-Decoder attention module and the inter-layer attention module were used to extract the comprehensive high-order situation information of the swarm. Finally multi-dimensional classification network and softmax layer were adopted to realize the classification of multiple situation elements. The experimental results showed that the unmanned swarm adversarial situation elements recognition method using Transformer and inter-layer attention exhibited good performance on situation element classification problem, and could accurately classify multiple situation elements synchronously. Compared with the baseline method, the swarm situation recognition method using Transformer and inter-layer attention had higher accuracy in the recognition of swarm formation and movement trend. Especially in the classification of situation elements that reflect the relative trends within the swarm, the proposed method clearly showed better performance.