复杂场景下基于稀疏表示的声振传感器网络协同分类识别

Aiming at the difficult issues on the long-term target recognition using acoustic and vibration sensor networks in complex scenes, considering the factors like temporal and spatial correlation, collaborative signal processing and energy consumption in sensor networks comprehensively, a novel approach combining classification with data compression is explored based on sparse representation and compressed sensing. The proposed approach firstly compresses the sensing feature data by random projection on the sensor nodes, and then recognizes the targets on compressed samples after transmitting to the central/backbone node by sparse representation, which can reduce the energy transmission of original data. The innovations are as follows: (1) A multi-task and multi-feature joint sparse representation classification model(MTMF-JSRC) with temporal and spatial correlation for multi-sensor classification(including linear and nonlinear model) is proposed. And for dimension reduction of features, the adaptive data compression method with dynamic feedback based on random projection is also proposed according to expert analysis, recognition rate and network conditions. Therefore, the target classification and data compression are associated effectively. We can investigate energy consumption and transmission cost simultaneously while maintaining the sufficient recognition rate of the target. (2) The sliding empirical mode decomposition method based on fast b-spline curve fitting is proposed, which reduces the computation cost while speeding up the decomposition, restraining end effect and mode mixing effectively in the decomposition process, then the robust features could be extracted effectively. (3) The over-complete dictionary library is built based on the above features, the fast and robust classification method based on augmented Lagrangian method and alternating direction method of multipliers is proposed. Finally, the proposed model and approaches make the fast and accurate target recognition available in complex scenes.

针对复杂场景下声振传感器网络长时间的目标事件识别难点，本课题综合考虑传感器网络的时空相关、协同信号处理、和能量消耗等因素，引入压缩感知与稀疏表示理论同时解决多传感器网络复杂场景监测中的分类识别和数据压缩问题。在传感器节点利用随机投影压缩声振特征数据，传到中心/骨干节点后再利用稀疏表示对压缩数据进行分类识别。具体创新内容为：(1)建立时空相关的多任务多特征线性与非线性联合稀疏表示分类模型，根据专家分析决策、识别精度及网络状况，提出基于随机投影的动态反馈自适应数据压缩方法，将目标识别和数据压缩有效关联，探索在达到一定目标识别率的同时降低声振传感节点的能耗与传输代价；(2)提出基于滑窗与快速B样条曲线拟合的经验模态分解方法，在加快分解速度的同时减少计算量，有效抑制模态混叠和端点效应现象，在此基础上提取噪声鲁棒性好的特征或特征组合；(3)提出基于增广拉格朗日法与交替方向乘子法的快速稳健求解算法。

针对复杂场景下声振传感器网络长时间的目标事件识别难点，本课题引入压缩感知与稀疏表示理论，同时解决多传感器网络复杂场景监测中的分类识别和数据压缩问题，平衡识别精度与网络能耗。主要研究内容包括时空相关的传感器网络联合稀疏表示分类模型，声振信号的特征提取，自适应随机投影数据压缩，多特征字典学习，快速稳健的稀疏表示分类求解。取得的研究成果如下：①建立了声振传感器网络多任务多特征稀疏表示分类理论模型，将目标识别和数据压缩有效关联，实现不同精度要求下声振信号的自适应采样。②提出了基于核稀疏表示的目标分类识别方法，通过构造单核与多核函数，将信号特征映射到高维核特征空间以实现线性可分，提升稀疏表示方法在复杂场景中的分类精度与适应性。③分析协同表示与稀疏表示的机理，将多任务和协同表示相结合，提出了基于最小残差加权模型的多任务协同表示分类识别方法。④结合随机投影和协同表示，提出了一种基于随机投影和协同表示的多任务目标识别模型，克服现有方法计算量大、耗时长的缺点。根据实际需求动态调整压缩比，平衡分类精度和计算代价。⑤结合字典学习，提出了基于Fisher判别字典学习的声振传感器网络协同分类识别算法，提高目标的分类精度。⑥针对Fisher判别字典时间复杂度高的问题，构造综合字典与分析字典分别进行信号的稀疏表示与判别重构，提出了基于多任务混合字典学习的声振传感器网络目标分类方法，进一步提升目标分类精度，降低算法复杂度。⑦提出了基于Clifford模糊支持向量机的多分类方法，构建了多维信号的多分类模型，可有效处理分类识别时存在的随机性和模糊性。发表学术论文19篇，其中SCI检索5篇，EI检索12篇，已授权专利2项，受理专利4项，培养硕士研究生7名。本项目的研究成果可应用于战场环境下的目标识别、交通流量监测与车辆识别、高速列车轨道伤损探测、农业病虫害监测、大型工业设备状态自动识别、以及大规模工程结构健康监测等场景。