Abstract: This work proposed an online detection method of abnormal data based on multilayer perceptron (MLP) and rolling-learning prediction mechanism. In this method, latest historical data with fixed length of data window is used to train an MLP model, and then a one-step-ahead prediction is obtained with the trained MLP model. Secondly, a confidence interval with probability p is calculated with the help of the one-step-ahead prediction and the model residual. New measurement is identified as normal one, if it falls inside the prediction interval; or an abnormal record when it is located out of the prediction interval. Instead of the real measurement, the prediction value is used to update the historical data if abnormal data occurs. Furthermore, through on-line test of real process data collected from a 300 MW coal-fired power generation unit, the effectiveness of the proposed method was verified.

Key words:  artificial neural network (ANN), multilayer perceptron (MLP), rolling-learning prediction, abnormal data, online detection