This thesis is located in an area of research IFSTTAR (French Institute of Science and Technology of Transport, Development and Networks) whose objective is to promote eco-driving by changing driving behavior to reduce fuel consumption. This is why defining the principles of efficient driving in the context of conduct and assists the driver so that it adopts these principles of conduct. More specifically, the objective of this thesis is to produce knowledge about driving behavior to explain the over or under fuel consumption. Modeling driving behavior will be based on data collected of natural situation driving (that is, gathered on the personal vehicle of the subject at its various commuting and this for several months).
This type of environmental data is extremely rich to know the real purpose in terms of driving, but the analysis is complicated by their heterogeneous nature. It is therefore interesting to study based on new approaches to knowledge discovery for the introduction of knowledge domain experts (Knowledge of laboratory cognitive science research to address the point of view of the conduct, the motives of the driver and engine experts for consumption) in this analysis process.
In the current research, expert knowledge in driving activity are often taken upstream to define observable but not throughout the analysis process. This is often due to the fact that statisticians argue about global indicators (number of braking,) and ergonomists on local situations (target behavior in a given infrastructure). In addition, this research often faces the problems of too many performance indicators that do not allow converging by data mining algorithms. Indeed, let the indicators are not relevant and they induce noise in the analysis, or they are too specific and are not representative.
The proposed approach is to define a methodology for analysis of combining several methods of data mining by introducing the knowledge of experts in the field in the following four steps:
Characterization of driving situations: for this step, we will build on research on cognitive modeling (Bellet Tattegrain) and Bayesian networks (Markov chains Dapzol 2001) by generalizing across situations. This will allow for homogeneous groups of situations (even same objective context and road driving).
Acquisition of knowledge: This step will acquire the knowledge of experts to identify and define relevant parameters that explain certain behaviors in specific situations.
Data preparation: to define a set of indicators potentially useful for discriminating between situations of a group they must then transfer factors experts in observable indicators on the data collected. For this, we will rely on expert knowledge in SPI and the results of research on driving activity (Ericsson, 2005 Mathern 2012).
Identification of relevant parameters: The central step of data mining must solve the problems of selection among all those previous indicators that actually explain the performance criterion. For this, knowledge extraction methods from data (data mining) will be used and the results will be validated by experts in driving.
Extraction rules knowledge production: the final step will summarize the various analyzes by a set of rules explaining the performance criterion (fuel consumption).
The advantage of this approach is to use expert knowledge and context of the activity to a precise and concrete representation of an event or situation of activity. This representation of the activity as a rule can be used to explain, predict and create a help system for driving that reduces fuel consumption.