时序数据图像化:战术意图识别及可移植框架  

Timing data visualization:tactical intent recognition and portable framework

作　　者：宋亚飞[1] 李乐民 权文[1] 倪鹏 王科 SONG Yafei;LI Lemin;QUAN Wen;NI Peng;WANG Ke(Air and Missile Defense College,Air Force Engineering University,Xi’an 710051,China;Science and Technology on Complex Aviation Systems Simulation Laboratory,Beijing 100076,China)

机构地区：[1]空军工程大学防空反导学院,陕西西安710051 [2]复杂航空系统仿真重点实验室,北京100076

出　　处：《通信学报》2024年第8期149-165,共17页Journal on Communications

基　　金：国家自然科学基金资助项目(No.61806219,No.61703426,No.61876189);陕西省自然科学基金资助项目(No.2021JM-226);陕西省高校科协青年人才托举计划基金资助项目(No.20190108,No.20220106);陕西省创新能力支撑计划基金资助项目(No.2020KJXX-065)。

摘　　要：通过将时序编码为图像,提出了一种结合曲线滤波技术和EfficientNetV2图像识别网络的鲁棒且可移植的战术意图识别框架。曲线滤波技术可以有效地减少大量时域特征、模型参数和训练时间的冗余,基于此,提出了一种改进的格拉姆角场方法将时序编码为图像,提高了卷积神经网络的特征提取能力。EfficientNetV2网络能够有效地处理意图图像,并成为预训练模型,使得在不同系统之间进行迁移学习成为可能。实验结果表明,所提框架相对于机器学习及深度学习等方法提高了0.99%以上的准确率,具有更好的性能、可扩展性、鲁棒性和可迁移性。By transforming time series into images,a robust and transferable tactical intent recognition framework was proposed,which integrated curve filtering technology and the EfficientNetV2 image recognition network.Curve filtering technology effectively reduced redundancy in numerous time-domain features,model parameters,and training time,an enhanced Gramian angular field(GAF)method was proposed to encode time series into images,enhancing the feature extraction capabilities of convolutional neural networks.The EfficientNetV2 network was adept at processing intent images and could serve as a pre-trained model,facilitating transfer learning across different systems.Experimental results demonstrate that the proposed framework achieves over 0.99%higher accuracy compared to machine learning and deep learning methods,exhibiting superior performance,scalability,robustness,and transferability.

关 键 词：时序编码 意图识别 图像分类 曲线滤波 格拉姆角场 EfficientNetV2 

分 类 号：V219[航空宇航科学与技术—航空宇航推进理论与工程] TP173[自动化与计算机技术—控制理论与控制工程] V917[自动化与计算机技术—控制科学与工程]