【作者】 王磊；

【导师】 郭贤生；

【作者基本信息】 电子科技大学 ， 信号与信息处理， 2021， 硕士

【摘要】 随着物联网技术的蓬勃发展,基于室内位置服务的市场需求与日俱增。室内定位中的指纹式定位方法因其在复杂的室内环境中抗干扰能力强得到广泛应用。传统的指纹式室内定位算法假设离线建库阶段的信号强度分布与在线定位阶段的信号强度分布是一致的。然而,信号强度分布经常因环境变化、设备异构等因素发生变化,此外无线接入点的缺失或新增会导致特征维度异构。迁移学习通过将源域的知识迁移到目标域中可以减轻环境变化和设备异构引起的信号强度分布变化。然而,传统的迁移学习定位算法未充分提取可迁移特征,存在知识迁移不充分问题。针对该问题,本文研究基于深度迁移学习的室内定位技术,主要研究工作如下:针对现有迁移学习定位算法存在的知识迁移不充分问题,本文提出了一种基于全局和局部结构一致性约束的迁移学习定位算法。在潜在子空间中,通过最小化源域和目标域的边缘和条件概率分布差异,以及最大化两个域的方差,约束全局结构的一致性,使得域间均值差异减小且映射后样本尽可能分离;通过最小化类内方差和最大化类间方差,以及保持数据局部的邻域结构,约束局部结构的一致性,使得映射后带相同标签样本尽可能接近,不同标签样本尽可能远离,且保留迁移前后数据的邻域关系。实验结果验证了该算法的有效性。本文提出了一种基于同构深度迁移网络的室内定位算法,利用深度神经网络可学习反映域间不变因素的深度可迁移特征的能力完成知识迁移,克服了传统的迁移学习定位算法通过浅层映射学习浅层表示特征而不能充分迁移知识的缺点。该深度迁移网络通过匹配希尔伯特空间中域间的均值嵌入和相关子域间的均值嵌入,以及匹配域间的协方差来减轻环境变化和设备异构引起的信号强度分布变化带来的定位性能下降。实验结果验证了该深度迁移网络的有效性。针对传统异构迁移学习定位算法通过浅层线性映射学习浅层公共特征而不能充分迁移知识的问题,本文提出了一种基于异构深度迁移森林的室内定位算法。通过采集少量的目标域带标签数据,以两个深度神经网络作为特征映射层,分别学习源域和目标域的一个公共特征,然后利用概率决策森林可以保留数据结构的特性,采用概率决策森林作为预测层用于联合迁移和定位。实验结果证明了该算法在域间分布不同且特征维度不同情景下的定位性能。更多还原

【Abstract】 With the vigorous development of Internet of things technology,the market demand based on indoor localization services is growing with each passing day.The fingerprint localization method in indoor localization is widely used because of its strong ability for resistance to disturbance in complex indoor environments.The traditional fingerprint localization algorithm assumes that the signal strength distribution in the offline database construction phase is consistent with that in the online localization phase.However,the distribution of the signal strength often changes due to environmental changes,heterogeneous devices and other factors.In addition,the lack or addition of wireless access points will lead to heterogeneous feature dimensions.By transferring the knowledge from the source domain to the target domain,transfer learning can reduce the change of distribution of the signal strength caused by environmental changes and heterogeneous devices.However,the traditional transfer learning localization algorithm does not fully extract the transferable features,which leads to insufficient knowledge transfer.To solve this problem,this thesis studies the indoor localization technology based on deep transfer learning.The main research work is as follows:Aiming at the problem of insufficient knowledge transfer in existing transfer learning localization algorithms,this thesis proposes a transfer learning localization algorithm based on global and local structural consistency constraints.In the potential subspace,by minimizing the discrepancy of marginal and conditional probability distribution between the source and target domain,and maximizing the variance of the two domains,the consistency of the global structure is constrained,so that the mean difference between the domains is reduced and the samples are separated as much as possible after mapping;by minimizing the intra class variance and maximizing the inter class variance,and maintaining the local neighborhood structure of the data,the consistency of the local structure is constrained,so that the mapped samples with the same label are as close as possible,and the samples with different labels are as far away as possible,and the neighborhood relationship of the data before and after transferring is preserved.Experimental results verify the effectiveness of the algorithm.In this thesis,an indoor localization algorithm based on homogeneous deep transfer network is proposed,which can learn the deep transferable features reflecting the invariable factors between domains to complete the knowledge transfer by the deep neural network,and overcomes the shortcomings of traditional transfer learning localization algorithm,which only learn the shallow representation features through the shallow mapping and can not transfer the knowledge fully.The deep transfer network can reduce the localization performance degradation caused by the change of distribution of the signal strength caused by the change of environment and heterogeneous devices by matching the mean embedding between domains and the mean embedding between related subdomains in Hilbert space,and the covariance between domains.The experimental results verify the effectiveness of the deep transfer network.Aiming at the problem that traditional heterogeneous transfer learning localization algorithm can not transfer knowledge sufficiently by learning shallow common features through shallow linear mapping,this thesis proposes an indoor localization algorithm based on heterogeneous deep transfer forest.By collecting a small amount of labeled data in the target domain,two deep neural networks are used as feature mapping layers to learn a common feature of the source domain and the target domain respectively.The probabilistic decision forest can retain the characteristics of data structure,and it is used as the prediction layer for joint transferring and localization.The experimental results show that the localization performance of the algorithm is better in the situation of different distribution between domains and different feature dimensions.更多还原