Paul HONEINE - Collaborative Learning in (Wireless) Sensor Networks -

Paul HONEINE

- Collaborative Learning in (Wireless) Sensor Networks -

Scientific Challenge

Collaborative information processing in sensor networks is becoming a very active field of research. Within conventional (centralized) sensor networks, sensors are only confined the role of collecting information measurements, while these data are collected through the network, and processed within a central unit, also known by the fusion center. Recent research reveals a new generation of sensors becoming more intelligent. These so-called smart sensors have the ability to collaborate with each other, by exchanging information in order to ensure optimal decision. As opposed to the centralized scheme, the proposed approach does not rely on a unique central processing unit whose damaging may lead to the failure of the entire system. The resulting system is thus very robust against a technical failure of the central unit or some sensors. Moreover, as information is collected and processed locally, within each sensor or within clusters of them, only relevant data are exchanged between smart sensors, hence, limiting the required channel communication bandwidth.

Wireless ad-hoc sensor networks have emerged as an interesting and important research area in the last few years. Wireless sensor networks cover a large spectrum of envisioned applications, varying from monitoring inhospitable habitats and disaster areas to operating indoors for intrusion detection and equipment monitoring. Many communities are interested with the potential applications of

these wireless sensor networks, and supporting future applications including military surveillance and tracking abilities. The building block of these networks is a wireless sensor mote, as with the Mica2 motes from Crossbow (see figure) used in our lab, with embedded sensing capabilities, memory and processing unit, as well as a wireless communication module and a power supply. Low cost and miniaturization of sensor nodes involve limited computational resources, power and communication capacities. Consequently, wireless ad hoc sensor networks require collaborative execution of a distributed task amongst a large set of motes.

Current Studies

Decentralized decision-making have been the focus of considerable research in the last decade. The large majority of the literature is based on the assumption that the probability distributions of the sensor observations lie within some known parametric family and seek to characterize the structure of optimal decision rules. Of course, this raises the question of how to determine an appropriate parametric family, and how to estimate parameters. Recent works in statistical machine learning have demonstrated the power and versatility of kernel-based methods for learning a problem on the basis of empirical data sets. This approach has been used successfully in a large class of problems, to name a few, spanning from the estimation of binary functions for detection and decision-making (with novelty detection techniques such as the one-class Support Vector Machines), to classification (Support Vector Classifier and kernel-based Fisher discriminant analysis), to functional estimation (Support Vector Regression and kernel-based least-squares methods), as well as unsupervised or non-labelled data learning (kernel-based Principal Component Analysis).

Most kernel machines are statistical learning algorithms that take advantage of the geometric and regularizing properties of reproducing kernel Hilbert spaces (RKHS). A key property behind such algorithms is that they can be expressed in terms of inner products only, involving pairs of input data. Replacing these inner products with a reproducing kernel provides an efficient way to implicitly map the data into a high, even infinite, dimensional RKHS and apply the original algorithm in this space. Because calculations are then carried out without making direct reference to the non-linear mapping of input data, this principle is commonly called the kernel trick. Kernel-based algorithms are computationally very efficient, and generally have their generalization performance guaranteed by the Statistical Learning Theory. Nevertheless, most kernel machines consider essentially centralized approaches to solve the learning problem, and therefore are not suitable for distributed learning in sensor networks.

We propose machine learning techniques for collaborative learning in wireless sensor networks, taking into account the limited resources in computation, memory usage, communication as well as power. We are interested in wireless sensor networks deployed in an environment to sense for modeling and monitoring given physical phenomena, such as gas, pollution, biochemical, temperature, humidity, acoustic field, motion, etc. It is worth noting that collaborative learning greatly enhances the accuracy of the estimation, and the detection performance.

We have been working on these problems for more than 10 years. Our main contributions are described in the following, chronologically from newer to older, with some selected publications:

•Recent Advances in Estimation and Evolution Tracking with Wireless Sensor Networks (2014-...)
•In-network (on-the-fly) Principal Component Analysis (2013-2016)
•Wireless Sensor Networks: Estimation and Evolution Tracking with Collaborative Learning (2008-2013)
•Illustrations are given at the end !!

Recent Advances in Estimation and Evolution Tracking with Wireless Sensor Networks (2014-...)

Our recent contributions are described in the following, with some selected publications. This is an ongoing work.

•Detection and estimation of the parameters of multiple gas sources in wireless sensor networks:

Gas Sources Parameters Estimation Using Machine Learning in WSNs.
Sandy Mahfouz, Farah Mourad-Chehade, Paul Honeine, Joumana Farah, and Hichem Snoussi
IEEE sensors journal
16 (14), 5795-5804, July, 2016
•Distributed estimation in intelligent camera networks, with an application on abnormal event detection:

• Optimal mobility for the sensors for the spatial regression:

In-network (on-the-fly) Principal Component Analysis (2013-2016)

(Work of PhD Student Nisrine Ghadban)

Principal component analysis is a very well-known statistical analysis technique. It determines principal axes that allow relevant data reduction. Thus, it is very useful in networks where a large number of measures is performed. However, such a technique needs high-complexity computation, which does not scales well in large networks. Our work proposes to overcome this issue, by describing several in-network strategies that estimate the principal axes with lower computational complexity. These strategies are either noncooperative or cooperative with information diffusion. The performance of the proposed strategies are illustrated on diverse applications, such as image processing and dimensionality reduction of time series in wireless sensor networks.

This is the first time that collaborative learning is considering an unsupervised learning problem.

Wireless Sensor Networks: Estimation and Evolution Tracking with Collaborative Learning (2008-2013)

Some Illustrations

Below is another example where we study the diffusion of a gas/pollution and we track its evolution in time. The space under investigation has an open door !

Collaborative Learning in (Wireless) Sensor Networks