基于平衡采样的主动半监督学习人类活动识别研究  

Research on active semi-supervised learning for human activity recognition based on balanced sampling

作　　者：郇战[1] 刘艳 李志新 董晨辉 周帮文 秦王盛 HUAN Zhan;LIU Yan;LI Zhixin;DONG Chenhui;ZHOU Bangwen;QIN Wangsheng(School of Microelectronics and Control Engineering,Changzhou University,Changzhou 213000,China;School of Computer Science and Artificial Intelligence,Aliyun School of Big Data,Changzhou University,Changzhou 213000,China)

机构地区：[1]常州大学微电子与控制工程学院,江苏常州213000 [2]常州大学计算机与人工智能学院阿里云大数据学院,江苏常州213000

出　　处：《传感器与微系统》2024年第7期63-66,71,共5页Transducer and Microsystem Technologies

基　　金：国家自然科学基金面上资助项目(61976028);国家自然科学基金资助项目(61772248)。

摘　　要：基于可穿戴传感器的人类活动识别研究逐渐受到人们的广泛关注。本文提出了一种基于平衡采样的主动半监督模型,在挑选样本进行标注时,将样本的不确定性和多样性一并考虑在内,挑选出类别平衡的不确定性样本。确保训练后的模型对每个类都有很好的识别性能,从而提升整体分类结果。同时,为了全部利用标记和未标记样本的信息,将主动学习和半监督学习相结合,利用损失项信息不断更新网络参数,提升模型在低注释下的识别性能。该模型在2个公开数据集上得到了验证,在确保获得较优分类准确率的同时,可以大大减少样本的人工标注工作。Human activity recognition research based on wearable sensors has gradually attracts widespread attention.An active semi-supervised model based on balanced sampling is proposed.When selecting samples for labeling,the uncertainty and diversity of the samples are taken into account,and the uncertain samples with balanced categories are selected.Ensure that the trained model has good recognition performance for each class,thereby improve the overall classification results.At the same time,in order to fully utilize the information of labeled and unlabeled samples,active learning and semi-supervised learning are combined,the network parameters are continuously updated by using the loss item information to improve the recognition performance of the model under low annotation.The model has been verified on two public datasets,which can greatly reduce the manual labeling work of samples while ensuring better classification accuracy.

关 键 词：主动学习 半监督学习 查询策略 人类活动识别 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]