This paper presents a diagnostic module developed by IFP and tested off-line on a FCC (Fluid Catalytic Cracking) pilot plant. The method uses four successive complementary techniques. They enable to go step by step from the observations to a sentence in natural language describing the faults. First, a quantitative causal model is elaborated from a quantitative behavioural model. Causality is obtained from the structure of each equation. Then, global and local alarms are generated using residuals (differences between measures and outputs of the model) and fuzzy logic reasoning. Then, a hitting set algorithm is applied to determine sets of components or equipment which are suspected to have an abnormal behaviour. Finally, expert human operator knowledge about those components is used to identify the fault(s) and produce messages for the operators. This software is currently tested off-line on the FCC pilot plant at IFP. The performance of the diagnostic module is illustrated on four practical scenarios of abnormal behaviour. This work is conducted as part of the CHEM EC funding project.