In practice, an industrial polyethylene process produces various products, even often developing new production grades for market demand. Therefore, the process has not only multiple operating conditions, but also time-varying characteristics. In addition, the process measurements inevitably are redundant and noisy. It is a challenging problem for on-line classifying the operating conditions in the industrial process. In this paper, principal component analysis (PCA) is applied to the reference data set to reduce the dimensions of variables and eliminate the collinearities among process measurements. Since outliers are inevitable in a real plant data set, they significantly stretch the cluster centers and covariances and reach an unreliable solution. In this paper, the distance-based fuzzy c-means (DFCM) algorithm is proposed. A boundary distance for each cluster is derived for identifying outliers, which should be discarded from the reference data set. Before the on-line classification, the statistic Q and T-2 of new data have to be evaluated. If any one of the statistics is out of its control limits, it indicates the new data do not belong to the PCA subspace and they should be collected for the next model update. In this paper, the blockwise recursive formulas for updating the means and covariance matrix are derived. By utilizing the updated means and covariance, a new PCA subspace that accounts for all events is derived recursively. In addition, through rotating and shifting the coordinates of the PCA subspace, the cluster parameters on the new subspace can be directly transferred from the previous one. The proposed method was successfully applied to monitor a polyethylene process with multiple production grades and time-varying characteristics.