This paper deals with a multi-agent architecture that we have developed for signal understanding. The architecture we propose has been developed in the framework of the AITRAS ESPRIT project of the EEC which aims at designing a new shell for the implementation of real time signal interpretation systems. The strong points of the architecture are : - the connection between low level signal and pattern recognition modules on one hand and symbolic interpretation on the other, - a mixed structure combining a blackboard model with explicit communication between modules by a message passing mechanism. In this respect, this new architecture extends the specialist society model that we developed earlier (Gong, 1989), - the integration of hypothetical reasoning. In this paper, we will mainly concentrate on the first demonstrator of the project. It concerns eddy current inspection for the non destructive examination of steam generator tubes in nuclear power plant. We will emphasize the enhancement that Artificial Intelligence technology provides in order to deal with such applications.