Apprentissage automatique et acquisition des connaissances : deux approches complémentaires pour les systèmes à base de connaissances. Application au système "ACASYA" d'aide à la certification des systèmes de transport automatisés
Habib Hadj‐Mabrouk · HAL (Le Centre pour la Communication Scientifique Directe) · 1992
This PhD thesis presents a contribution to the improvement of the usual methods of safety analysis used within the framework of the certification of automated transport systems (ATS). The mission of the certification experts consists in assessing the safety character of a new ATSby evaluating the completeness of the accident scenarios envisaged in the manufacturer's safety study. The certification assistance methodology developed is based on the joint and complementary use of knowledge acquisition and machine learning. The method of knowledge acquisition chosen has proved effective in extracting and formalizing historical knowledge from safety analysis but insufficient to acquire in detail the expert certification approach which is highly intuitive and scalable. To fill this gap, our study focused on the use of machine learning techniques. The difficulty of defining and choosing a learning system adapted to the requirements of an industrial application led us to propose a characterization of the learning process. ACASYA is the software environment developed to support the certification assistance methodology. It is made up of two main modules: CLASCA and EVALSCA, respectively dedicated to the classification and evaluation of accident scenarios. CLASCA which we have entirely designed is a symbolic-digital, inductive, incremental, non-monotonous and interactive learning system. EVALSCA, developed around the CHARADE rules learning system, aims to suggest to certification experts possible failures not considered by the manufacturer and likely to jeopardize the safety of a new ATS. To date, ACASYA has proven the interest of the methodology to formalize, exploit and perpetuate the know-how of the certification expert, in order to tend towards the exhaustiveness of the security analysis. In contrast to diagnostic assistance systems, ACASYA can be seen as a tool to help prevent breakdowns, located downstream of conventional forecasting methods of safety analysis.