This paper describes a contribution to improving the usual safety analysis methods used in the certification of railway transport systems. The methodology is based on the complementary and simultaneous use of knowledge acquisition and machine learning. The purpose is contributed to the generation of new accident scenarios that could help experts to conclude on the safe character of a new rail transport system. The method of analysis and evaluation is centered on the summarized failures (SFs) which are involved in accident scenarios capitalized. A summarized failure (SF) is a generic failure produced by the combination of a set of basic failures which has the same effect on the performance of the system. Each scenario brings into play one or more SFs. The purpose is to automatically generate a recognition function for each SF associated with a scenario class. The SF recognition function is a production rule which establishes a link between a set of facts (parameters which describe a scenario or descriptors) and the SF fact. A base of evaluation rules can be generated for each class of scenarios. The SF deduction stage requires a preliminary phase during which the rules which have been generated are transferred to an expert system in order to construct a scenario evaluation knowledge base. The evaluation knowledge base is exploited by forward chaining by an inference engine and generates the summarized failures (SFs) which must enter into the description of the scenario which is to be evaluated.