To develop a fault detection device represents today one of the major challenges for manufacturers of robotic systems. The detection process requires the use of a number of sensors to monitor the operation of these systems. However, the cost and constraints in the implementation of these sensors often lead designers to optimize their numbers, leading to a lack of necessary measures for the detection of failures. One way to bridge this gap is to estimate non-measurable parameters from a mathematical model describing the dynamics of the real system. This paper presents an approach based on mixed data (measured data and estimated data) for the detection of failures in robotic systems. This detection is performed using a decision tree classifier type. The data used to learning from actions taken on the real system. This data is then enriched with data estimated from one observer based on an analytical model. This enrichment as attributes further aims to increase knowledge on the functioning of the classifier system and therefore improve the rate of good detection failures. An experiment on a system for actuating a robotic seat, showing the interest of our approach, will be presented at the end of the article.