Recent advances in electronics and wireless communications have enabled the concept of a wireless sensor network (WSN), a set of sensing devices that cooperate through wireless communications to extract information about an observed physical phenomenon. A WSN allows to deploy more sensors and cover a wider area than a conventional sensor array, and to exploit the spatial characteristics of the sensor signals to enhance the performance of its assigned signal processing task. However, wireless links consume a significant amount of energy and introduce additional noise, posing several challenges for algorithm design. Since sensor nodes are usually powered by batteries, energy efficiency is a crucial aspect. With respect to the additional noise introduced by the communication process, an algorithm needs robustness to ensure a good performance in practical situations.
This thesis develops quantization strategies for signal estimation in WSNs that address both challenges. For energy efficiency, the proposed strategies are focused on minimizing the energy spent in wireless communication while respecting a limit on the signal estimation performance, a problem known as bit depth allocation. Regarding robustness, their focus is on generalizing a distributed signal estimation algorithm to improve its resiliency to noisy links.
The first and second parts of the thesis develop several algorithms to solve bit depth allocation problems for linear minimum mean squared error (MMSE) signal estimation. They are shown to achieve significant energy savings, and their computational efficiency, closeness to the optimal solution and scalability are studied and compared. In particular, the proposed CCP-linMMSE method is shown to be superior in these aspects to generic and previously presented methods. Additionally, for the application of speech enhancement in a wireless acoustic sensor network (WASN), the energy savings and performance of greedy algorithms are evaluated through the use of both simulated and real-life recordings.
The third part focuses on distributed signal estimation in a WSN with noisy links. It first studies the energy-vs-performance trade-offs when nodes share their fused sensor signals with reduced signal bandwidth and bit depth. It then designs new fusion rules that consider the presence of additive noise in the signals exchanged between nodes. This leads to a generalization of the existing distributed adaptive node-specific signal estimation (DANSE) algorithm towards a version that is resilient to noisy links, referred to as the N-DANSE algorithm. The convergence of the N-DANSE algorithm is proven, and its superiority in the context of WSNs with noisy links is validated through numerical simulations.
In conclusion, this thesis has developed novel quantization strategies that improve the energy efficiency and robustness to noisy links of signal estimation in WSNs. This research enables a flexible adjustment of energy consumption and estimation performance, allowing to increase the lifetime of the network, and more reliable operation. It also lays the foundations for further refinements, such as energy models for the wireless links that capture more complex wave propagation effects and packet retransmission, and bit depth allocation integrated directly in the framework of distributed signal estimation.
10/6/2026 13:30 - 15:30
Aula van de Tweede Hoofdwet, 01.02, Kasteelpark Arenberg 41, 3001 Heverlee
