Output details

Improved fault diagnosis in multivariate systems using regression-based reconstruction

A new regression-based technique for fault identification and isolation is described. If a fault condition is detected in the score variables the fault signature is estimated using a neural network. This is then removed from the

corresponding score variables, allowing separation of normal/abnormal variation from the process variables. This can successfully isolate the fault signature, overcoming inherent limitations of conventional variable reconstruction and

supports the possibility of handling faults with a nonlinear signature. The new method compares well with a projection-based alternative using experimental data from a reactive distillation unit supplied by collaborators at the University of Dortmund, Germany.