Monitoring of metallurgical reactors by the use of topographic mapping of process data
Although principal component analysis has been applied widely for monitoring plant performance in a broad range of industrial processes, it is a linear technique that tends to break down when processes exhibit significant non-linear behaviour. In this paper a non-linear multivariate fault diagnostic system is proposed for metallurgical reactors, based on the use of hidden target mapping neural network to project the data to a three-dimensional subspace that can be visualized by a human operator. As is shown by way of a case study, the normal operating region can be defined by means of historic data confined by a convex hull. Subsequent process faults or novel data not projected to the normal operating region are automatically detected and visualized, while a sensitivity analysis of the data can aid the operator in locating the source of the disturbance.