分布式信号检测与识别的目的是在无线传感器网络中判断信号是否存在并确定其所属的类别。稀疏性在自然界中广泛存在，具体表现为在高维向量中，只有少数元素具有显著幅度，大多数元素近似为零。对稀疏性加以利用，有助于提升分布式检测与识别的性能。本文在无线传感器网络所面临的通信负载约束和机密性约束下，重点研究了基于稀疏性的分布式信号检测与识别方法。贡献如下：1）针对通信负载约束下的分布式信号检测问题，充分挖掘待检测信号的稀疏性，在并行结构网络中提出了基于高精度量化数据传输的审查-局部最大势检测方法，推导了该方法的检测性能与通信量的理论关系。为进一步降低通信负载，提出了基于单比特数据传输的似然比量化-局部最大势检测方法，在并行结构网络和树结构网络中分别推导了检测性能的定量衡量标准，并通过建立和求解优化问题得出了准最优量化门限。仿真实验表明，和已有方法相比，所提出的检测方法可在保证检测性能的前提下显著降低通信量。2）针对机密性约束下的分布式信号检测问题，利用待检测信号的稀疏性，提出了基于虚伪审查策略的局部最大势检测方法，对系统的检测性能和机密性水平进行了定量分析。为了在保证系统机密性的前提下实现最佳检测性能，建立了相应的优化问题，通过理论分析提供了准最优系统参数的设定方法。对于极度稀疏的待检测信号，所提出的方法可在相同系统机密性水平下达到与已有方法基本相同的检测性能，但是，前者放松了对信号稀疏度先验知识的需求，具有更强的可实现性。此外，为保证系统机密性并实现与审查-局部最大势检测方法相同的检测性能，基于虚伪审查策略的局部最大势检测方法所需的传感器个数仅为审查-局部最大势检测方法的1.21倍，实现成本较低。3）针对通信负载约束下的分布式信号识别问题，基于传感器与融合中心之间所发送数据的稀疏性，结合压缩感知技术和联邦平均算法的优点，先后在传输高精度量化数据和单比特数据的场景下提出了压缩感知-联邦学习算法和单比特压缩感知-联邦学习算法，实现了节点和融合中心之间的双向通信压缩。实验表明，相比于已有方法，所提出的方法可在相同的通信量下达到更高的识别准确度。

The aim of distributed detection and recognition in wireless sensor networks is to make decision regarding the existence of signals of interest and to identify the category to which an existing signal belongs. Sparsity widely exists in nature, which indicates that a high-dimensional vector only includes a few elements of significant values while others are mostly close to zero. A full exploitation of sparsity is conducive to the enhancement of the performances of distributed detection and recognition. In this dissertation, under the practical constraints of communication burden and system secrecy in wireless sensor networks, the problems of distributed detection and recognition are investigated by taking the sparsity into full account. The main contributions are listed as follows:1) For the problem of distributed detection under the constraints of communication burden, by fully exploiting the sparsity of the signals to be detected, Censoring Locally Most Powerful Test is proposed in parallel sensor networks with the transmission of high-precision data. For this method, the relation between the detection performance and the communication burden is theoretically derived. To further reduce the communication burden, Locally Most Powerful Test based on Quantization of Likelihood Ratios is proposed with 1-bit data transmitted in the network. The metric of the detection performance is derived in parallel sensor networks and tree-structured sensor networks, respectively, with which near optimal quantization thresholds are provided after formulating and solving an optimization problem. Simulation results verify that the proposed methods induce a lower communication burden to attain the same detection performance as the existing methods.2) For the problem of distributed detection under the constraints of system secrecy, utilizing the sparsity of the signals to be detected, Locally Most Powerful Test based on Falsified Censoring is proposed. Quantitive analysis of the detection performance and the secrecy level of the system is performed. To attain the optimal detection performance while ensuring system secrecy, an optimization problem is formulated and solved to determine the near optimal system parameters. For signals that are extremely sparse, under the same contraints of system secrecy, the existing method and the proposed one achieve almost the same detection performance. However, the latter does not require the sparsity level to be completely known, which is more practical. Besides, to ensure system secrecy and to attain the same detection performance as Censoring Locally Most Powerful Test with Q nodes, Locally Most Powerful Test based on Falsified Censoring only needs 1.21Q sensor nodes, which can be realized at a fairly low cost.3) For the problem of distributed recognition under the constraints of communication burden, based on the sparsity of the data transmitted between the sensor nodes and the fusion center, by combining the advantages of compressed sensing and Federated Averaging, two methods, i.e., Federated Learning based on Compressed Sensing and Federated Learning based on 1-Bit Compressed Sensing, are proposed when high-precision data and 1-bit data are transmitted in the network, respectively. The proposed methods enable bi-directional compression of the communication between the fusion center and the sensor nodes. Experimental results show that the proposed methods attain higher values of recognition accuracy under the same communication burden compared with the existing methods.