Abstract: The dynamic on-line monitoring system (DMS) based on on-line training of ANN has been implemented successfully in a refinery for real-time,on-line and dynamic estimation of quality indexes of fractionator side draw,such as end point of raw-gasoline,pour point of diesel-oil,Reid vapor pressure,flash point,freezing point.It improves the control performance and optimizes the operation of the fractionator.To improve the predicted accuracy of DMS and fit the new production cases,a novel ANN training method, on-line training method,has been proposed in this article.This method overcomes the disadvantages of off-line training of ANN effectively.It can save lots of time and effort in contrast to the conventional off-line training of ANN.It improves the performance of DMS.This method includes three main steps:judge when ANN needs to be trained again;get new training patterns on-line;train ANN on-line.Steady state determination of the process is a key procedure in on-line training.A method based on the mathematical theory of evidence is used in steady state determination.This improves the training results of ANN and the reliability of DMS.