On-line Fault Detection and Classification for a Compressor Process in the Oxygen Plant

In this paper, a data-driven model is proposed for on-line monitoring a process with highdimensional variables, outliers, and time-varying characteristics. In this research, principal component analysis (PCA) is used to eliminate collinearity between process variables. After that, fuzzy rules are generated by using the compressed data and an outlier rejection clustering algorithm, named distancedbased fuzzy c-means (DFCM), from which a feasible solution can be obtained to reflect the actual data gatherings. When new event emerge, the data are collected for next model update. An adaptive PCA algorithm is utilized to accommodate the new event data without recalculating the trained data. The known event rules can be transferred to the new PCA subspace by rotating and shifting coordinates of the subspace. Therefore, only new event data need to be clustered on the new subspace. The proposed approach has been applied to monitor a compressor process of the steel plant. Results show the challenges of process monitoring can be effectively dealt with.