协同特征感知的传感器网络网内优化计算研究

In the WSN-based monitoring applications for high-dimension targets (e.g., bridge, dam, state of the backbone grid), the sensor nodes need to collaboratively detect the global features of their common target. This usually requires the sink node to gather all the raw data for centralized computation, which is much different from feature combination based on local raw data in traditional monitoring applications of WSNs. In order to reduce the plenty raw data transmission in WSNs, we need study the in-network computing for the centralized computation task, while guaranteeing the global feature's quality. This is a cross discipline, and has not get much studied..Innovations of this projects are two folds. On the one hand, based on the idea of in-network processing, we propose new thinking about the optimal computation assignment of incremental computations for centralized task. On the other hand, we proposed new concept of feature compressive sensing constrained by quality of the feature extracted instead of quality of the signal reconstructed. We'll try to sense the feature during the sampling stage, through constructing the relationship between the sparse measurement matrix, quality of the main sparse elements and the feature extracted. The project will finally take example of the collaborative sensing of mechanical response feature for structures, and realize both the in-network processing method and feature compressive sensing method for performance comparison, so as to provide some guidance for feature in-network computing in WSNs.

在基于无线传感器网络的大空间尺度目标（如桥梁、水坝、骨干电网）状态监测中，各传感器节点需协作完成对共同的目标体的整体特征感知。这往往要求汇聚节点进行基于所有源数据的集中式计算，而非传统传感器网络那样可以基于局部数据融合的特征拼合。为降低海量的网络源数据传输，我们需要在保证协同特征感知精度下，研究集中式计算任务的网内计算方法。这是一个交叉学科，目前国内外研究较为薄弱。.项目的创新在于：一方面基于网内处理的思想，提出集中式任务在迭代（包括并行迭代）计算模式下进行网内最佳计算分配的研究新思路；另一方面，提出以特征感知精度（而非源数据重建精度）为约束的“特征压缩感知”新理念，探索将特征感知直接建在采集阶段，建立稀疏测量矩阵、重构的主稀疏分量精度、以及相应特征感知精度三者的内在关联。项目最终将以结构力学动态响应特征的协同感知为例，揭示网内处理思想与特征压缩感知思想的性能比较，给出有价值的指导。

在基于无线传感器网络的大空间尺度目标（如桥梁、水坝、骨干电网）状态监测中，各传感器节点需协作完成对共同的目标体的整体特征感知。这往往要求汇聚节点进行基于所有源数据的集中式计算，而非传统传感器网络那样可以基于局部数据融合的特征拼合。为降低海量的网络源数据传输，我们需要在保证协同特征感知精度下，研究集中式计算任务的网内计算方法。.项目一方面基于网内处理的思想，提出集中式任务在迭代（包括并行迭代）计算模式下进行网内最佳计算分配的研究新思路，另一方面，提出以特征感知精度（而非源数据重建精度）为约束的“特征压缩感知”新理念，探索了将特征感知直接建在采集阶段，建立稀疏测量矩阵、重构的主稀疏分量精度、以及相应特征感知精度三者的内在关联。.项目取得的重要成果包括：1）实现了将典型特征计算——Hankel矩阵奇异值提取进行了递归计算拆分，并证明了该计算拆分不影响特征计算精度。2）从拆分的宽度、深度、厚度三个维度归纳了计算拆分模式、及对应的路由形态。3）证明了迭代计算下的最优路径。4）建立了特征压缩感知的计算方法，并分析比较了网内特征处理与特征压缩感知两种不同技术路径的优缺点，为特征网内计算在不同应用场景下的技术路径选择给出重要的指导意见。