Failure diagnostics and general decrease of accident rate at power plants is a major task of energy generation industry, and solution of it provides reliable energy supply country wide and technological progress in mechanical engineering. Along with some other crucial means, the task could be solved by means of teaching the maintenance staff based on accidents that have already occurred. That is no secret that everywhere in the world due to indecision or misinterpretation a huge number of accidents have happened merely because the personnel were not aware of similar cases at other power plants. Nevertheless with the development of computational technologies and mathematical algorithms the role of personnel in some cases has been reduced to observation and action in critical situation, while the rest is performed by machines: various types of diagnostics and prediction of failure systems based on artificial neural networks are widely applied and developed. However, in order to train these systems, it is absolutely required to know the reasons that could lead to and consequences that could follow some deterioration in turbo generator sets performance. The aim of the present paper is to give statistical analysis of turbo generator sets failure reasons based on open source data presented by Russian and foreign researchers and analysts in the field. The statistical data could be used to perform classification and ranking of failure reasons in terms of frequency of occurrence, possibility to identify or detect, etc. and the paper also gives brief listing of possible ways of detection or identification of failure modes and possible consequences for the main units of a turbo generator.