Faults or special events which occur occasionally in continuous processes generate dynamic patterns in a large number of process variables. However, the patterns arising from the same fault can exhibit different time durations (depending on the operating conditions), magnitudes and directions. Any robust fault diagnosis method must be able to correctly classify these faults under these different conditions. This paper presents an off-line fault diagnosis method based on pattern recognition principles for multivariate dynamic data. The method consist of a filtering and scaling step, where the magnitude dependent information is removed, and a similarity assessment step via dynamic time warping (DTW). DTW is a flexible pattern matching method used in the area of speech recognition. The method presented in this paper is designed to classify faults independently of their magnitude, duration, direction and plant production level. As a further feature extraction step, principal component analysis is used to reduce the dimension of the multivariate problem and enhance the distance-based classification. Case studies from the Tennessee-Eastman plant are used to test the method and to illustrate its advantages and limitations.