Ethylene Compressor Monitoring

Principal component analysis (PCA), which is widely used in process monitoring, performs best when the system variables are linearly correlated. However, in practice, the variables are often nonlinearly related, and may be subject to periodic forcing or disturbances, which compromise the performance of conventional PCA. In model-based PCA (MBPCA), multivariate statistics are used to analyze the portion of the observed variance that cannot be predicted using a model of the process, and thus significantly enhances the attainable diagnostic resolution. In this paper, MBPCA is used for fault detection monitoring of an ethylene compressor, which operates under a significant periodic disturbance caused by the ambient temperature. An analytical expression is derived to predict the limits of identifiable faults given bounds on the parametric model uncertainty.