Sparse Signal Processing and Statistical Inference for Internet of Things

Abstract

Data originating from many devices within the Internet of Things (IoT) framework can be modeled as sparse signals. Efficient compression techniques of such data are essential to reduce the memory storage, bandwidth, and transmission power. In this thesis, I develop some theory and propose practical schemes for IoT applications to exploit the signal sparsity for efficient data acquisition and compression under the frameworks of compressed sensing (CS) and transform coding. In the context of CS, the restricted isometry constant of finite Gaussian measurement matrices is investigated, based on the exact distributions of the extreme eigenvalues of Wishart matrices. The analysis determines how aggressively the signal can be sub-sampled and recovered from a small number of linear measurements. The signal reconstruction is guaranteed, with a predefined probability, via various recovery algorithms. Moreover, the measurement matrix design for simultaneously acquiring multiple signals is considered. This problem is important for IoT networks, where a huge number of nodes are involved. In this scenario, the presented analytical methods provide limits on the compression of joint sparse sources by analyzing the weak restricted isometry constant of Gaussian measurement matrices. Regarding transform coding, two efficient source encoders for noisy sparse sources are proposed, based on channel coding theory. The analytical performance is derived in terms of the operational rate-distortion and energy-distortion. Furthermore, a case study for the compression of real signals from a wireless sensor network using the proposed encoders is considered. These techniques can reduce the power consumption and increase the lifetime of IoT networks. Finally, a frame synchronization mechanism has been designed to achieve reliable radio links for IoT devices, where optimal and suboptimal metrics for noncoherent frame synchronization are derived. The proposed tests outperform the commonly used correlation detector, leading to accurate data extraction and reduced power consumption.