Techniques enabling early detection and diagnosis of faults are important in the processing industries. This paper emphasises a technique for early fault detection and diagnosis based on dynamic fault data and a two-step fault detection and diagnosis framework. The approach shows various advantages over alternative methods including prompt fault detection and localisation, applicability to large-scale systems without the need for excessive computing resources, and a modular architecture that allows plant sections to be treated individually. In the proposed method, the large-scale plant is broken up into sections and a Petri net based on real time data is used to locate the particular section of the plant in which the fault originates. This Petri net then activates secondary neural networks, which diagnose the exact location of the fault in that particular plant section. Applicability of the proposed technique is demonstrated through a pilot plant case study. (C) 2004 Elsevier Ltd. All rights reserved.