基于大语言模型的超宽带雷达多任务学习方法OA
UWB-LLM:ultra-wideband radar multi-tasking learning based on large language model
超宽带(UWB,ultra-wideband)雷达通过无线感知实现人数检测和生命体征监测,现有方法依赖统计特征或轻量级神经网络,在多种任务间的迁移与泛化能力有限.鉴于大语言模型(LLM,large language model)的优异跨模态学习能力,提出了基于大语言模型的超宽带雷达多任务微调框架,结合混合专家(MoE,mixture of experts)与低秩适应(LoRA,low-rank adaptation)方法,将雷达信号的时序特征映射至大语言模型嵌入空间,并通过参数高效微调(PEFT,parameter-efficient fine-tuning).在自采集数据集和公开数据集上,分别对人数检测任务及呼吸、心电图(ECG,electrocardiogram)与连续血压 3 种生命体征信号估计任务进行训练和评估.实验结果表明,所提框架在人数检测准确率和 3 种生命体征信号估计任务的相关系数较现有算法平均提升了 37.62%、7.47%、18.16%和14.70%.
Ultra-wideband(UWB)radar enables people counting and vital signs monitoring with wireless sensing.How-ever,existing methods rely on statistical features or lightweight neural networks,which have limited ability to migrate and generalize across multiple tasks.In view of the powerful cross-modal learning capability demonstrated by large lan-guage model(LLM),a UWB radar multi-task learning framework based on LLM was proposed.This approach success-fully mapped the temporal features of radar signals into the embedding space of LLM,and by integrating the mixture of experts(MoE)mechanism and low-rank adaptation(LoRA)strategy,UWB-LLM performed parameter-efficient fine-tuning(PEFT)for multi-task learning.Experiments were conducted on both self-collected and publicly available datasets for people counting task as well as respiration,electrocardiogram(ECG),and continuous blood pressure estimation tasks,respectively.Compared with the state-of-the-art algorithms,UWB-LLM achieves average accuracy improvements of 37.62%for people counting and 7.47%,18.16%,and 14.70%for the three vital sign estimation tasks.
饶翀;姜夕康;郭嘉航;李蕾;张琳
北京邮电大学人工智能学院,北京 100876北京邮电大学人工智能学院,北京 100876北京邮电大学人工智能学院,北京 100876北京邮电大学人工智能学院,北京 100876北京邮电大学人工智能学院,北京 100876||北京市大数据中心,北京 101160
信息技术与安全科学
大语言模型超宽带雷达人数检测生命体征监测多任务学习混合专家
LLMUWB radarpeople countingvital signs monitoringmulti-task learningMoE
《物联网学报》 2026 (2)
53-64,12
国家自然科学基金资助项目(No.61971056) The National Natural Science Foundation of China(No.61971056)
评论