Trainee Project
Energy Efficient Communication Scheduling for IoT-based waterbirds monitoring: Machine Learning Strategies
2022/02/01 - 2022/06/30
Wetlands cover about six percent of the earth's surface. Characterized by a unique biodiversity, they represent one of the richest and most diverse
ecosystems on our planet (Skinner & Zalewski, 1995). More than half of the wetlands have been destroyed over the last century. These environments are
still threatened today due to urbanization, agricultural intensification and pollution. Wetlands are home to a very wide variety of animal and plant species.
They play a major role in biodiversity, in particular, they are privileged places for tens of thousands of waterbirds of different species to winter or make a
temporary halt. Worldwide, there are about 10, 000 known bird species. Birds respond to possible changes in the environment and are good indicators of
biodiversity. Some of the waterbird species are threatened with extinction (Barrowclough, Cracraft, Klicka, & Zink, 2016). Fauna and flora monitoring in
general and birds in their natural habitat in particular, has several interests (Archaux, 2011):
▪ it provides a reliable indicator for regularly assessing the evolution of the number of endangered birds in an area ;
▪ birds can be seen as a tool for environmental and even public health monitoring ;
▪ it can allow avoiding significant economic losses.
A wireless multimedia sensor network (WMSN) (Akyildiz, Melodia, & Chowdhury, 2007) can be used to monitor migratory birds in their natural habitat
(wetlands). Acoustic and image sensors can be deployed in conjunction with other types of presence detection sensors to identify, recognize and count the
number of a threatened bird species based on their vocalizations (songs and cries) and/or photographs. Monitoring waterbirds in their natural habitat using a
wireless sensor network (WSN) must consider issues related to massive data collection (multimedia) as well as automatic analysis of the collected data
(Gaston & O'Neill, 2004). The collection of varied and massive data (in terms of volume and rate of acquisition) using a highly constrained infrastructure
such as a WSN is a real problem and requires considerable effort from the scientific community. Such applications have specific needs in terms of quality of
service, bandwidth and not least, storage and processing capacities. All these needs result in significant energy expenditure.
In order to accommodate these large bandwidth requirements, one solution consists in increasing the available bandwidth through high-performance
network services and protocols. These services may include optimized resource (communication) allocation mechanisms. Research on high data rate WSN
applications is still in the earlier stage. Quite a few systems have been deployed where additionally most of data gathering is performed manually. A
guaranteed data transfer rate has to be provided for these applications to allow for remote and automatic data acquisition.

The 2015 IEEE 802.15.4e amendment (Guglielmo, Brienza, & Anastasi, 2016) of the IEEE 802.15.4-2006 standard define a number of channel access
modes (MAC protocols) to suite Low Power Loss Networks (LLN) constrained nature. This includes TSCH, which received the attention of both academia
and industry researchers. In fact, the specification of the access schedule is left to higher protocol layers. Meanwhile and thanks to recent advances in GPU
technology and cloud computing, machine learning (ML) returned to the front stage as a response to complex and challenging problems from various
domains. While traditional heuristics produce "one-size fits all" solutions, ML is able to deal with complexity and allows adapting to the actual network state
(network dynamics) which avoids manual intervention (Kieu-Ha, et al., 2015).

In this thesis, we propose to review the ML-based scheduling algorithms (Hermeto, Gallais, & Theoleyre, 2017). Centralised algorithms are able to find
optimal solutions at the cost of reactivity while distributed ones adapt faster to network dynamic at the expense of precision. The aim is to compare the
different approaches to highlight the trade-off optimality/reactivity. This can lead to proposing a scheduling algorithm that leverage distributed and
centralized benefits. A proof of concept is to realise in order to demonstrate the feasibility of the proposed solution in the context of high data rate reporting
dictated by our waterbirds monitoring application.
Eco-Technic systems engineering