Neural networks and nonlinear SPC

Statistical process control procedures are receiving significant attention in response to increasing demands for improved process performance and improved product reproducibility. In particular the multivariate statistical approaches of partial least squares and principal component analysis have been found to be useful to provide for effective predictive model development from highly dimensioned and ill conditioned monitored process data. This paper addresses an alternative approach using artificial neural networks to provide a process representation that is capable of characterising nonlinear behaviour. In particular the problem of process fault detection is considered using a feature detection network topology to extract from the process data important attributes that indicate in a meaningful way the presence of process malfunctions. The ability of the method to detect process faults is demonstrated by application to a comprehensive simulation of an industrial polymer reactor and two industrial fermentation processes.<
>