LENART Marcin

PhD graduated
Team : LFI
Departure date : 10/02/2020
https://lip6.fr/Marcin.Lenart

Supervision : Marie-Jeanne LESOT, Andrzej BIELECKI

Co-supervision : PEDRISOR Teodora (Thales), REVAULT D'ALLONNES Adrien (Univ. Paris VIII)

Sensor Information Scoring for Decision-Aid Systems in Railway Domain

In this thesis, we investigate the problem of assessing information quality produced by sensors. Information quality is an abstract concept which is considered important in many fields as high-quality information is necessary for decision-making systems. Its assessment highly depends on the type of information, its context and considered domain. It is usually decomposed into different criteria, called dimensions, that allow to capture and combine different aspects of a piece of information. This thesis focuses on the case of information produced by sensors, i.e. devices measuring an aspect of reality and translating it into a digital value. Indeed, sensors, usually used in networks, do not always provide correct information and the scoring of this information is needed. This thesis proposes to exploit the sensor specificity to define a dedicated, and yet generic, scoring method. Existing approaches for information scoring in the case of sensors are usually based on ground truth or meta-information, which significantly limits their genericity: they are often difficult to obtain and make the approaches appropriate only for specific sensors, exploiting their unique characteristics. We propose an approach that deals with these difficulties by providing a model designed to be sensor-generic, not dependent on ground truth and dependent only on easy-to-access meta-information, exploiting only attributes shared among the majority of sensors. The proposed model is called ReCLiC from the four dimensions that it considers: Reliability, Competence, Likelihood and Credibility. Informally, the ReCLiC model takes as input a log file of sensor entries and aims at attaching each log entry with a numerical evaluation of quality of this entry: this quality is understood as the trust that can be put in the message content of the log entry, measured by considering the source, the content and the context of this message, which are the three main components defining a piece of information. We discuss in depth the requirements of the four proposed dimensions on which ReCLiC relies and propose motivated definitions for each of them. Furthermore, we propose an implementation of the generic ReCLiC definition to a real case, for specific sensor in the railway signalling domain: we discuss the form of the four dimensions for this case and perform a formal study of the information scoring behaviour, analysing each dimension separately. The proposed implementation of the ReCLiC model is experimentally validated using realistic simulated data created from a real dataset in the railway domain. The proposed experimental protocol allows to control various quality issues as well as their quantity, in four distinct scenarios of problematic log files. This experimental study that includes a study of the parameters shows that the proposed ReCLiC model has the desired behaviour and in particular the ability to assign low trust scores to the simulated noisy entries. Finally, the ReCLiC model is used to analyse a real dataset where quality problems are detected and discussed. A new visualisation method is proposed to show multiple trust scores from many sensors at the same time. This visualisation allows to observe trust propagation which shows how low-quality messages can impact other information. In addition, the notion of trust dynamics is introduced and analysed based on this example.

Defence : 10/02/2020 - 14h - Krakow (Poland)

Jury members :

SCHERER Rafał (Politechnika Częstochowska) [Rapporteur]
SMITS Grégory (IRISA, Lannion) [Rapporteur]
SZMUC Tomasz (AGH, Cracovie)
MARSALA Christophe (LIP6-SU, Paris)
BIELECKI Andrzej (AGH, Cracovie)
LESOT Marie-Jeanne Lesot (LIP6-SU, Paris)
REVAULT D'ALLONNES Adrien (Univ. Paris VIII, Paris)
PEDRISOR Teodora (Thales)

2018-2020 Publications

  • 2020
  • 2019
  • 2018
    • M. Lenart, A. Bielecki, M.‑J. Lesot, T. Petrisor, A. Revault D'Allonnes : “Dynamic Trust Scoring of Railway Sensor Information”, ICAISC 2018 - 17th International Conference on Artificial Intelligence and Soft Computing, vol. 10842, Lecture Notes in Computer Science, Zakopane, Poland, pp. 579-591, (Springer) (2018)