Journal of Chemical Engineering of Japan, Vol.47, No.12, 876-886, 2014
Multivariate Outlier Detection Approach Based on k-Nearest Neighbors and Its Application for Chemical Process Data
This paper proposes a novel method to accurately estimate the multivariate location and scatter for detecting outliers in the high-dimensional and complex contaminated data. Firstly, the abnormal degree of corresponding sample is characterized by a gamma index based on k-nearest neighbors. The smaller gamma index indicates the smaller distances from the sample to its neighbor samples and the higher probability for it to be a normal sample, while the higher probability to be an outlier. Secondly, based on the gamma index, a quasi-modified robust scaling is proposed to select the sub-sample data including the maximum normal data from the sample data. Continuing, the robust Mahalanobis distances are calculated based on the location and scatter of the sub-sample data and employed to distinguish between the normal data and outliers in the sample data. Finally, the proposed method is evaluated by using synthetic data, some standard benchmark data and a real industrial process data. The results show that the location and scatter of the sample data are calculated precisely and the outliers can be effectively detected and eliminated by the proposed method, which demonstrates its satisfactory ability to identify outliers and good prospect of application for chemical process data.
Keywords:Chemical Process;k-Nearest Neighbor;Mahalanobis Distance;Multivariate Data;Outlier Detection