Abstract: If there exist gross errors in the soft sensor modeling data， the accuracy of the model is largely affected. Therefore， for the data set to be used in the modeling process， a method of gross error detection of 3MAD-MMMD based on cluster analysis is proposed to process the data before modeling. The data of different variables in different time from the seamless pipe perforation process is collected. Then these data are arranged into a matrix. The 3MAD algorithm is used first to eliminate the large error of single-variables to get the new data matrix. Based on the Euclidean distance formula， the distance is then obtained from the data in matrix to another which is closest to it of the same variable. Finally， d_○med， the median value of all variables’ closest distance， is treated as testing standards to detect gross error of new data matrix. It can be seen from the experimental and simulation results that the gross errors in the collected data sets are effectively eliminated in the 3MAD-MMMD detection method.

Key words: soft sensor modeling；gross error, cluster analysis, 3 median absolute deviation, modified median minimum distance