Experiments were carried out with a pilot scale, wail cooled, fixed-bed reactor for benzen oxidation. The inlet concentration of benzene was varied in three ways : periodically, with PRBS, and by means of a step change. A model for the reactor is developed with use of the Karhunen-Loeve (K-L) expansion and neural networks. The K-L. expansion procedure used here acts as a preprocessor to achieve good data compression while preserving as much information about the measurements as possible. A recurrent multilayer feedforward network is then used to relate the coefficients of the K-L expansion to the operating conditions. The model developed is used for on-line prediction of the axial temperatures in a fixed-bed reactor and results show that the predictions are in good agreement with measurements.