Abstract
In this thesis, we address the problem of statistical inference in large-scale sensor networks observing
spatially varying fields. First, we revisit traditional single-sensor hypothesis testing. We then present a
multiple hypothesis framework to model spatial fields occurring in a multitude of practical signal process-
ing applications. Observing and monitoring phenomena that occur within a spatial field is essential to a
variety of applications. This includes tasks, such as, detecting occupied radio spectrum in shared spec-
trum environments, identifying regions of poor air quality in environmental monitoring, smart buildings
and different Internet of Things (IoT) applications. Many of these practical problems can be modeled
using a multiple hypothesis testing framework, with the goal of identifying homogeneous spatial regions
within which a defined null hypothesis (e.g. pollution remaining at tolerable level, radio spectrum being
unoccupied) is in place, and regions where alternative hypotheses are true. These regions can be formed
assessing observations made by multiple sensors placed at distinct locations. To be scalable for large-
scale sensor networks, we suggest to compute local test statistics, such as, p-values at each individual
sensor to avoid communication overhead due to a large number of sensors exchanging their raw mea-
surement data. Individual test statistics are fed to a Fusion Center (FC), which performs the inference. At
the FC, statistical inference is performed with a propose a method referred to as “Spatial Inference based
on Clustering of p-values (SPACE-COP)” that uses multiple hypothesis testing and Bayesian clustering
to detect occurring phenomena of interest within the spatial field. The method identifies homogeneous
regions in a field based on similarity in decision statistics and locations of the sensors. The number of
clusters, each of which is associated to a hypothesis, is determined by a newly derived Bayesian cluster
enumeration criterion that is based on the statistical model that has been derived in this project. An
EM-algorithm is developed to compute the probabilities that associate sensors with clusters. We present
two different decision criteria, for maximum performance (SPACE-COP) and control of false discoveries
(FDR SPACE-COP). The performance of the proposed methods is studied in a series of simulation ex-
amples and compared to competitors from the literature. Simulation results demonstrate the validity
of proposed SPACE-COP methods also for cases in which the assumption on underlying spatial shape of
alternative areas was clearly violated and true alternative areas followed arbitrary and even non-convex
shapes. In summary, the derived algorithms are applicable to large-scale sensor networks to perform
statistical inference and identify homogeneous regions in an observed phenomenon or field where the
null hypothesis does not hold.
spatially varying fields. First, we revisit traditional single-sensor hypothesis testing. We then present a
multiple hypothesis framework to model spatial fields occurring in a multitude of practical signal process-
ing applications. Observing and monitoring phenomena that occur within a spatial field is essential to a
variety of applications. This includes tasks, such as, detecting occupied radio spectrum in shared spec-
trum environments, identifying regions of poor air quality in environmental monitoring, smart buildings
and different Internet of Things (IoT) applications. Many of these practical problems can be modeled
using a multiple hypothesis testing framework, with the goal of identifying homogeneous spatial regions
within which a defined null hypothesis (e.g. pollution remaining at tolerable level, radio spectrum being
unoccupied) is in place, and regions where alternative hypotheses are true. These regions can be formed
assessing observations made by multiple sensors placed at distinct locations. To be scalable for large-
scale sensor networks, we suggest to compute local test statistics, such as, p-values at each individual
sensor to avoid communication overhead due to a large number of sensors exchanging their raw mea-
surement data. Individual test statistics are fed to a Fusion Center (FC), which performs the inference. At
the FC, statistical inference is performed with a propose a method referred to as “Spatial Inference based
on Clustering of p-values (SPACE-COP)” that uses multiple hypothesis testing and Bayesian clustering
to detect occurring phenomena of interest within the spatial field. The method identifies homogeneous
regions in a field based on similarity in decision statistics and locations of the sensors. The number of
clusters, each of which is associated to a hypothesis, is determined by a newly derived Bayesian cluster
enumeration criterion that is based on the statistical model that has been derived in this project. An
EM-algorithm is developed to compute the probabilities that associate sensors with clusters. We present
two different decision criteria, for maximum performance (SPACE-COP) and control of false discoveries
(FDR SPACE-COP). The performance of the proposed methods is studied in a series of simulation ex-
amples and compared to competitors from the literature. Simulation results demonstrate the validity
of proposed SPACE-COP methods also for cases in which the assumption on underlying spatial shape of
alternative areas was clearly violated and true alternative areas followed arbitrary and even non-convex
shapes. In summary, the derived algorithms are applicable to large-scale sensor networks to perform
statistical inference and identify homogeneous regions in an observed phenomenon or field where the
null hypothesis does not hold.
Original language | English |
---|---|
Qualification | Master's degree |
Awarding Institution |
|
Publication status | Published - 2019 |
MoE publication type | Not Eligible |