Home // Research // FSTC // Computer Sci... // News and Events // PhD defence: CONFIDENCE-BASED DECISION-MAKING SUPPORT FOR MULTI-SENSOR SYSTEMS


twitter linkedin facebook google+ email this page
Add to calendar
Event date: Thursday, 31 October 2019 03:00 pm - 05:00 pm
Place: Campus Belval, Maison du Nombre

Event Date: Oct. 31, 2019, 3 p.m. - Oct. 31, 2019, 5 p.m.

Event Location: Campus Belval, Maison du Nombre


We live in a world where computer systems are omnipresent and are connected to more and more sensors. Ranging from small individual electronic assistants like smartphones to complex autonomous robots, from personal wearable health devices to professional eHealth frameworks, all these systems use the sensors’ data in order to make appropriate decisions according to the context they measure.


However, in addition to complete failures leading to the lack of data delivery, these sensors can also send bad data due to influences from the environment which can sometimes be hard to detect by the computer system when checking each sensor individually. The computer system should be able to use its set of sensors as a whole in order to mitigate the influence of malfunctioning sensors, to overcome the absence of data coming from broken sensors, and to handle possible conflicting information coming from several sensors.


In this thesis, we propose a computational model based on a two layer software architecture to overcome this challenge.


In a first layer, classification algorithms will check for malfunctioning sensors and attribute a confidence value to each sensor. In the second layer, a rule-based proactive engine will then build a representation of the context of the system and use it along some empirical knowledge about the weaknesses of the different sensors to further tweak this confidence value.


Furthermore, the system will then check for conflicting data between sensors. This can be done by having several sensors that measure the same parameters or by having multiple sensors that can be used together to calculate an estimation of a parameter given by another sensor. A confidence value will be calculated for this estimation as well, based on the confidence values of the related sensors.


The successive design refinement steps of our model are shown over the course of three experiments. The first two experiments, located in the eHealth domain, have been used to better identify the challenges of such multi-sensor systems, while the third experiment, which consists of a virtual robot simulation, acts as a proof of concept for the semi-generic model proposed in this thesis.