Multi-window Drift Detection Method Based on Integrating Outlier Identification and Input/Output Space Information with Its Application

Abstract

In soft sensing application of actual industrial process, it is necessary to accurately identify samples that represent the working condition drift. Therefore, a multiwindow concept drift detection method is proposed in this paper, which integrates outlier identification and input/output space information. Firstly, a principal component analysis (PCA) model and a decision tree (DT) model are built by using historical data. Then, in the first window, the process data are standardized on the basis of 3σ criterion to remove outliers. In the second window, the T <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> and SPE of the new sample are calculated to determine whether it is a drift sample in the input space. In the third window, the prediction error of the new sample is calculated to determine whether it is a drift sample in the out space. Finally, we take the union of the concept drift samples detected in the input and output space to obtain the new updating sample set. At the same time, the PCA model and the decision tree model are updated after the number of the updating samples reach a pre-set threshold. The experimental results verify the effectiveness of the proposed algorithm in the actual industrial datasets.