An Effective Outlier Detection Method for EAF Based on an Iterative Heterogeneous Ensemble

Jiong Zhang, Biao Wang, Hongjun Dai

Abstract: With the increasing applications of data-driven techniques in Electric Arc Furnace (EAF) steelmaking, outlier detection has drawn more attention than before. By removing outliers in EAF datasets, the performance of methods of process monitoring, process control, and system modeling can be improved. To overcome the drawback of single detection technique, several outlier ensembles have been proposed. However, structures of these outlier ensembles are fixed and only variance reduction is considered. From the perspective of bias-variance tradeoff, we can also construct outlier ensembles in terms of bias reduction to obtain stronger outlier ensembles. To this end, we propose an iterative outlier ensemble, in which heterogeneous base learners are used. By pruning the training set iteratively, we will obtain an effective ensemble since the influence of outliers has been alleviated. We apply the proposed outlier ensemble to EAF datasets to verify its performance. The experimental results have shown that our detection method outperforms other ensembles and single models in most cases.