【作者】 王双全；

【导师】 杨杰；

【作者基本信息】 上海交通大学 ， 模式识别与智能系统， 2008， 博士

【摘要】 传感器技术、无线通讯技术、嵌入式计算技术、分布式信息处理技术、微电子技术等领域的进步及相互结合,推动了无线传感器网络的快速发展。无线传感器网络将逻辑上的信息世界与客观上的物理世界连接起来,改变了人类与环境的交互方式,提供了利用逻辑信息来表述客观世界的一种有效的、便捷的方法。目前,无线传感器网络已经广泛地应用于环境智能、环境监控、工业制造、交通运输、军事工程等众多领域。作为无线传感器网络的一个重要应用领域,环境智能泛指能感知到用户的存在并为其提供智能化服务的电子环境和系统。环境智能的实现依赖于与用户行为密切相关的环境信息的采集和处理,依赖于对用户行为的分析、判断和推理,如:判断用户的位置、识别用户的行为、检测用户与环境的交互等,并在此基础上为用户提供智能化的服务。本文的研究主要针对环境智能中的两个研究重点:行为识别和目标定位。对于行为识别,介绍了基于环境变量的行为识别、人物交互式行为识别和穿戴式行为识别三种方法,并着重分析了穿戴式行为识别方法。穿戴式行为识别方法常采用监督学习方法,但该方法不具有异常检测能力和扩展学习能力。本文引入一类分类算法,利用组合高斯一类分类模型来判断行为是否是已知的。对于已知行为,采用加权支持向量机分类算法来识别其行为类别;对于未知行为,在分段算法的基础上,提取新行为的样本加入到识别系统中,扩展系统的识别能力。为了实现识别算法在传感器网络内的分布式实施,本文提出了一种基于移动代理的分布式分类方法,并将此方法运用到两种典型分类算法中。首先通过分解分类模型,将模型参数存放到对应的各个传感器节点上,实现分类模型的离散化;其次,在分类时,通过每个传感器节点计算自己的特征值和分类数据,实现分类操作的离散化;最后,利用移动代理串行访问各个节点,累积计算结果。与集中式分类算法相比较,基于移动代理的分布式分类算法可以减小带宽需求,平衡各个节点之间的计算、存储和能量消耗。对于目标定位,本文介绍人员定位和声音源定位。在人员定位中,常采用的方法往往需要用户穿戴传感器节点或在传感器节点上安装附加设备,本文提出了一种基于无线电波的非穿戴式定位方法,利用接收信号强度的变化来判断是否有人出现在一对无线电收发机之间,进而判断其位置。对于声音源定位,本文提出了一种源能量未知情况下的分布式声音源定位方法。文章结合增量梯度算法和基于能量比的声音源定位方法,通过分解和重组能量比定位方法的代价函数,得到适合分布式实施的迭代公式。采用合适的能量比个数以及迭代起始点位置,该方法可以获得与穷尽搜索方法近似的准确率,但只需要非常小的计算消耗。更多还原

【Abstract】 Recent advances in sensor technology, wireless communications, embedded computing, distributed information processing and micro electrical mechanical systems (MEMS) have enabled the development of wireless sensor network (WSN). WSN bridges the physical world and the logical world, provides a new interaction method between human and environment, and enables us to express the physical world by logical symbols in an efficient and convenient way. Nowadays, the WSN is widely used in ambient intelligence (AmI), environmental surveillance, manufacture, transportation and military engineering.As one important application area of WSN, AmI refers to the electronic environments that are sensitive and responsive to the presence of people. The realization of AmI systems relies on the collection of human information and environmental conditions, and, of course, on the analysis, judgement and inference of these information. With the context information such as positions, activities and the interaction between human and environment, the AmI system can provide intelligent services to the users.This thesis is focused on the activity recognition and object localization, which are two important research topics in AmI. As to activity recognition, three methods are introduced, which are environment variable based method, human-object interaction based method and wearable activity recognition method. For the wearable accelerometer based activity recognition, the supervised learning method is often used to recognize the human’s activities. However, this method cannot detect and recognize the unknown activities, and cannot extend the system’s recognition capability. In this thesis, the one-class classification algorithm is introduced and the combined Gauss one-class classification models are used to judge whether one activity is known. For the known activities, the weighted support vector machine (WSVM) is used to recognize their types. For the continuous unknown activities, based on the segmentation algorithm, training samples of new activities are selected and added into the existed recognition system to extend its recognition capability.In order to implement the recognition algorithms in the WSN in a distributed way, a mobile agent based distributed classification method is proposed and applied to two typical classification algorithms. The classification model is firstly decomposed and the model parameters are stored at related sensor nodes. During distributed classification, each sensor node calculates its own feature and classification result. A mobile agent is dispatched to visit all the sensor nodes serially and aggregate the results on them. Compared with the centralized classification algorithm, the proposed algorithm can reduce the bandwidth requirement, and balance the computation, storage and power consumption among sensor nodes.As to the object localization, the human localization and the acoustic source localization are introduced in this thesis. For human localization, the usual methods often needs wearing the sensors on or needs some additional devices. This thesis proposes a radio based human localization method without on-body sensor. The attenuation of received signal strength indicator (RSSI) is used to detect whether there is a person is standing between a pair of transceivers and determine his position accordingly.For the acoustic source localization, a distributed localization method with unknown source energy is proposed. With the combination of the incremental gradient algorithm and the energy ratios based acoustic source localization method, the cost function of the energy ratios based localization method is decomposed and reformulated to adapt to the distributed computation. Based on appropriate number of iteration and initial search locations, the proposed method can obtain approximate accuracy as that of the exhaustive search (ES) method with much less computation cost.更多还原