Principal component analysis variable selection for logistic models in plant hazard event prediction applications

In some process control applications, providing plant operators with even ten minutes of lead time to address potential hazard events could be the difference between catastrophic failure and graceful recovery. In this work, logistic models are developed to predict propylene compressor surge. Principal components analysis is used to select explanatory variables which exhibit high variance in the thirty minutes prior to a compressor surge event. Three types of models are developed: a main model based on principal components, a significant variables model and a model using interactions. Instances of each of the three model types were created using training data from three hazard events to create a total of nine predictive models. Each of the nine models was tested against data from a fourth hazard event. One of the nine models was capable of predicting surge events with a reasonable false alarm rate. The remainder of the models either did not identify any pre-hazard conditions or over-diagnosed by generating far too many false alarms. Future work will apply normalization and other techniques including the use of Process Plant Computing Limited´s parallel coordinate plotting software.