A gross error detecting method based on fuzzy curve weighted MMMD clustering for soft sensor modelling data

Existence of gross errors in data samples of soft sensor modelling will result in a poor, inaccurate model. To overcome this problem, a new gross error detecting method based on fuzzy curve weighted modified median minimum distance (MMMD) clustering was proposed. In this method, fuzzy curve was used to determine the degree of importance of each auxiliary variable to the primary variable firstly. Then, in the similarity calculation of clustering, each auxiliary variable was weighted according to the calculation results. At last, clustering and gross error detection were done. This method was used to detect the gross errors in the data for building a 4-CBA concentration soft sensor model in PTA oxidation process. Results based on practical industrial data indicate the validity of the proposed method.