This paper describes a strategy that allows the identification of gross errors for pyrolisis reactor measurements. The problem formulation is not specially suited for a particular case but has a wide range of application. A reactor model is formulated in terms of heat and mass balances and input-output mappings based on available measurements. The adjustment is done using historical data and a rigorous reactor simulation program. Neural networks trained with a Robust Back Propagation algorithm, relating variables in the convective zone, are essential to identify gross errors in crossover temperatures. The evaluation of the proposed scheme for gross error detection/identification shows a good performance.