The history of the development of technical systems is characterized with permanent increase of system complexity and requests for higher reliability. In many cases adequate modeling of complex plants is not affordable. As the consequence, data based models are widely used, but the computational time and the quality can still be too poor for reliable fault diagnosis. In this paper, an iterative approach to fault diagnosis in complex systems is proposed, with objective to obtain efficient and reliable fault diagnosis. The presented methods have been developed in the framework of an industrial project in the field of engine test benches, from which examples are shown.