The existing fault diagnosis methods using artificial neural networks do not deal with time vary-ing symptoms of faults. But in real plants, because the effect of the fault propagates continuously with time, the changing values of various sensors provide valuable information for on-line fault diagnosis. This paper describes a new method in which time varying values of the measurements are used as the symptoms of the faults to train the artificial neural network. A fault diagnosis system for a distillation column based on the proposed method can accurately diagnose both single and multiple faults.