Abstract: The supervision of many technical systems is often a challenging task due mainly to various nonlinearities. In this study, a multi-model approach for fault detection and diagnosis is proposed as an effective way since it allows to derive good process models valid over a wide range of operation and subsequently to detect changes of the current process behaviour. The diagnosis task is accomplished by decomposing the complex process into several sub-processes in order to generate a set of structured residuals. The validity of the approach is illustrated on the well known academic three tanks benchmark and different faults can be detected and isolated continuously, over several operating regimes.