Neural models are built to predict the temperature of the steel sheet at the exit of an induction furnace in a galvanizing line, with respect to the power applied and to the operating conditions. The second section briefly describes the induction furnace in the galvanizing line and presents the information available for modeling. The third section recalls the structure of the one hidden layer feedforward neural network basically used for system identification and describes the Gauss-Newton training rule based on a quite general error criterion. The following section derives different training algorithms from three robust forms of the general criterion. In the last part, the three robust learning rules are employed on the available data. Results are compared with those of the standard Levenberg-Marquardt update rule. The best model obtained is then pruned to improve the generalization ability of the network.

A particular point of a study for quality monitoring and improvement, and for condition diagnosis on a hot dipped galvanising line (Soliac, Florange, France) is presented. This point deals with the supervision of sensors and processes used to produce galvannealed sheet. The galvanising tower of the plant is briefly presented in section 1. Section 2 introduces the redundancy to supervise sensors or processes. In section 3, two examples of possible redundancy on the galvannealing section are described. Detection tests are finally presented, with special attention to the robustness to outliers.