Multivariate statistical methods such as principal component analysis (PCA) and partial least squares (PLS) have been widely applied to the statistical process monitoring (SPM) of chemical processes and their effectiveness for fault detection is well recognized. These methods make use of normal process data to define a tight normal operation region for monitoring. In practice, however, historical process data are often corrupted with faulty data. In this paper, a new process monitoring method is proposed that is composed of three parts: (1) a preanalysis step that first roughly identifies various clusters in a historical data set and then precisely isolates normal and abnormal data clusters by the k-means clustering method; (2) a fault visualization step that visualizes high-dimensional data in 2-D space by performing global Fisher discriminant analysis (FDA), and (3) a new fault diagnosis method based on fault directions in pairwise FDA. A simulation example is used to demonstrate the performance of the proposed fault diagnosis method. An industrial film process is used to illustrate a realistic scenario for data preanalysis, fault visualization, and fault diagnosis. In both examples, the contribution plots method, based on fault directions in pairwise FDA, shows superior capability for fault diagnosis to the contribution plots method based on PCA. (C) 2005 American Institute of Chemical Engineers.