基于压缩域传感网数据流的状态识别算法研究

The status identification technology based on sensor network is applied to structural health monitoring, medical monitoring, geological disaster warning, military counter-terrorism and other fields widely. In theory, it belongs to data stream classification issues. In line with the particularities of sensor network data stream for the presence of time correlation, spatial correlation and vulnerable to interference noise, this project proposed an data stream summary acquired mechanism based on the anomaly detection algorithm in compressed domain, and an adaptive ensemble classification algorithm to realize to recognize the status of monitoring target based on the sensor network data stream. We plan to investigate the project with the following four aspects. Firstly, taking the advantage of time-spatial correlation in sensor network data stream, compressed sampling the data stream with separable compressive sensing, then implanting spectral anomaly detection in compressed domain to acquire the data summary with overcoming the effects of measurements noises. Secondly,according to the less traning samples , establishing the base classification based on information geometry theory.Thirdly, establishing an adaptive data stream ensemble classification algorithm with the information entropy, to classify the mulit-classification and detect the concept drift. And update the individual classification based on active learning algorithm. Lastly, putting up the experimental platform of through wall human detection with Ultra Wide-band (UWB) radar sensor network to analyze and evaluate the performance of the proposed algorithm. The research will further expand the application breadth and depth of sensor network and it has important theoretical and economic value.

基于传感网的实时状态识别技术广泛应用于结构健康监测、医疗监测、地质灾害预警、军事反恐等领域，理论上其属于数据流分类问题。本项目针对传感网数据流存在的时间相关性、空间相关性，及易遭受噪声干扰等特殊性，提出了基于压缩域异常检测算法的数据流概要获取机制及自适应集成分类算法，实现基于传感网数据流的状态识别。主要从四个方面展开研究:第一，运用分离式压缩感知实现数据流的压缩采样，在此基础上结合谱理论进行异常检测获取数据流概要；第二，在较少已标记样本前提下，研究基于信息几何学理论的基础分类器构造；第三，构建基于信息熵的自适应数据流集成分类算法，进行概念漂移检测与多类别同时识别，并利用自主学习理论实现个体分类器更新；第四，搭建基于UWB雷达传感网的多人体隔墙目标识别实验平台进行算法性能分析与评估。本研究成果将进一步拓展传感网技术的应用广度和深度，具有重要的理论和经济价值。

基于传感网的实时状态识别技术广泛应用于结构健康监测、医疗监测、地质灾害预警、军事反恐等领域，理论上其属于数据流分类问题。本项目针对传感网数据流存在的时间相关性、空间相关性，及易遭受噪声干扰等特殊性，开展了五个方面研究：第一，运用分离式压缩感知实现数据流的压缩取样，并在此基础上开展了数据流的异常检测算法；第二，针对数据流中存在的多类别分类问题，提出了基于模糊模式识别与遗传算法的多类识别算法；第三，根据数据流中广泛存在的非平衡样本问题，开展了基于深度学习的非平衡样本条件下多状态识别研究，并探索了单传感器和传感网数据流对非平衡样本条件的分类效果；第四，针对数据流中的概念漂移现象，分别研究了并行概念漂移检测算法和基于模糊积分集成算法的更新机制；最后搭建了基于UWB雷达传感网的人体状态识别实验平台，对提出的算法进行了性能分析与测试。本研究成果将传感网的大数据分析提供重要的理论基础。