< Back to previous page
Distributed Signal Processing Algorithms for Sensor and Actuator Networks
Book - Dissertation
Technological progress in current analogue and digital devices together with the miniaturization of hardware have pushed sensing and actuating devices to new heights. Devices can perform more computationally intensive digital signal processing (DSP) tasks for a similar or even lower power budget and more devices can be placed on the same surface area, enriching DSP algorithms with more inputs/outputs and hence more degrees of freedom. When these devices have to perform a specific DSP task, such as the estimation or generation of a signal of interest, cooperation between devices with different locations can significantly enhance the overall performance, especially in adverse situations. The devices (or nodes) then need to be connected using either wired or wireless connections and form a sensor and actuator network (SAN). Compared to conventional fixed sensor or actuator arrays with a clearly localized position, SANs have access to more spatially distributed sensors and actuators, with more spatial degrees of freedom. This thesis will focus specifically on estimating and generating signals using the multi-channel Wiener filter (MWF) within both wireless and wired SANs. When wireless connections are favored, e.g., for an ad-hoc placement of the devices, the term wireless sensor and actuator network (WSAN) is often used. Specifically in acoustic applications where the sensors and actuators consist of microphone and loudspeaker arrays respectively, the network is referred to as wireless acoustic sensor and actuator network (WASAN). However, when wired connections are already in place as is the case for cell-free massive multiple input multiple output (CFmMIMO) systems where a wired SAN can be formed using the fronthaul connections between access points (APs) with up to 100 antennas in currently deployed mobile networks, wired connections should be used. Both applications require high processing rates for sampling, processing and transmission, compared to traditional low-rate SANs. It is therefore critical to develop efficient algorithms under which these nodes can cooperate and enhance their estimation/generation tasks, under tight bandwidth, complexity and energy constraints. Although aggregating and processing all uncompressed signals and parameters in one place (e.g., in a fusion center) seems straightforward, this demands a large communication bandwidth, consumes a lot of power and forms a single point of failure. This thesis therefore proposes novel distributed DSP algorithms, where nodes (microphone/loudspeaker arrays or APs) only exchange compressed signals and/or parameters and share the heavy computational burden of the centralized processing using in-network processing. The performance of centralized processing will be seen as an absolute benchmark for these distributed algorithm, and is often reached when the distributed algorithms have time to converge iteratively on stationary data. The first part of the thesis presents distributed algorithms designed for application in a WAS(A)N. In this part node-specific MWF-based speech enhancement in a WASN and sound zoning in a WASAN are considered as DSP techniques. The algorithms are node-specific, since the speech signals to be estimated are different for each node (but still originating from the same sources), the local sound zones can be chosen independently at each node or local loudspeaker arrays have node-specific power constraints. The DSP tasks define centralized solutions using all available sensors/actuators and they are compared to stand-alone solutions, i.e., as if the nodes are not aware of the signals received/transmitted by other nodes. The design of all the distributed algorithms in this part follows a top-down strategy, under which it will be shown that the nodes iteratively (in a time-recursive fashion) converge to the centralized solution, by exchanging certain signals or parameters using the available wireless links. The second part of this thesis focuses on developing distributed algorithms designed for application in CFmMIMO systems. The considered DSP tasks are MWF-based channel estimation, uplink receive combining, downlink transmit precoding and power allocation, taking into account channel estimation errors and different channel models. The presented distributed algorithms either use a network center as central processing unit to transmit the compressed data to and make final decoding/precoding decisions, or consider user-centric processing, whereby the decoding decisions are taken in the network by a user-specific AP. Similarly, convergence of the distributed algorithms to optimal solutions is shown, while strongly reducing the communication overhead over the available fronthaul links. Finally, the last chapter provides the conclusions, summarizes the contributions of this thesis and further discusses possible future research directions.