Distributed Signal Processing Algorithms for Acoustic Sensor Networks

In recent years, there has been a proliferation of wireless devices for individual use to the point of being ubiquitous. Recent trends have been incorporating many of these devices (or nodes) together, which acquire signals and work in unison over wireless channels, in order to accomplish a predefined task. This type of cooperative sensing and communication between devices form the basis of a so-called wireless sensor network (WSN). Due to the ever increasing processing power of these nodes, WSNs are being assigned more complicated and computationally demanding tasks. Recent research has started to exploit this increased processing power in order for the WSNs to perform tasks pertaining to audio signal acquisition and processing forming so-called wireless acoustic sensor networks (WASNs). Audio signal processing poses new and unique problems when compared to traditional sensing applications as the signals observed often have temporal characteristics that fluctuate rapidly requiring significantly faster transmission and computation speeds. In order for WASNs to cope with these increased demands, it is incontrovertibly apparent that efficient algorithms need to be developed. These algorithms must not only be computationally efficient but must also take into consideration several network wide design constraints in order to efficiently allocate the limited resources of the nodes. The main aim of this thesis is, therefore, to incorporate these constraints into the design of distributed signal processing algorithms for WASNs. The envisaged WASNs are primarily tasked with signal estimation that is exemplified by binaural noise reduction systems that utilize additional remote microphones. These systems focus on speech enhancement and noise reduction as their signal estimation task, where spatial cue preservation plays a role. The general problem statement of this thesis, the current state of the art and the mathematical framework for (distributed) signal estimation are given in the introductory chapters. WSNs will first be introduced in a centralized framework, where all of the information is aggregated and processed at a central location in order to perform signal estimation, e.g., for speech enhancement. This is then extended to a distributed framework, where the processing is distributed across the nodes and where each node performs a (possible node-specific) signal estimation task. As a general aim, the distributed algorithms developed in this thesis look to perform the same estimation as in the centralized framework while reducing the amount of data exchange between the nodes. The first part of this thesis focuses on a concrete application, investigating the improvements that can be achieved in terms of noise reduction performance and spatial cue preservation by adding a single remote sensor (microphone) to a binaural noise reduction system, e.g. binaural hearing aids or cochlear implants. The increase in the noise reduction performance by the addition of a remote sensor signal serves as the motivation for later chapters where many nodes consisting of multiple sensors are available. In the case where multiple remote microphones or nodes are available, a method for choosing which (sub)set of signals or nodes should be included in the estimation for maximum benefit is also discussed. A method is introduced that uses a so-called utility measure which limits the effect of node removal and can be used to aid in node addition. While first introduced in a centralized scenario, the utility is extended to that of a distributed scenario. However, due to the distributed nature of the signal estimation, the utility is shown to represent a bound as opposed to an exact quantity in the centralized scenario. The second part of this thesis introduces linear compression and fusion rules of the distributed adaptive node-specific signal estimation (DANSE) algorithm for use in heterogeneous and mixed-topology WSNs. These mixed-topology WSNs are divided into a set of smaller substructures that are based on either clique or cluster formation. It is shown, that the nodes are able to converge to the same optimal solution as if they were to receive all of the uncompressed signals from every other node. Supporting techniques are also introduced relating to different broadcast strategies and topology formation. In the third part of this thesis, each node implements a novel method to linearly compress its sensor signals in order to transmit to the other nodes in the WSN. This leads to the introduction of the topology-independent DANSE (TI-DANSE) algorithm. While the TI-DANSE algorithm is first introduced in a fully connected topology, the convergence properties are shown to be applicable in any topology, as long as the nodes have access to a network wide summed signal. An attractive attribute of the TI-DANSE algorithm is that since it relies on a network wide summed signal, it is less sensitive to link failures and also becomes applicable in WSNs with dynamic topologies. A method to compute this network wide summed signal is proposed that relies on a maximum of two transmissions per node. Finally, conclusions are given reiterating the contributions of the thesis as well as exploring possible future research directions.