Multiple kernel learning based feature selection for process monitoring

摘要

In the process of large-scale chemical engineering, more useful industrial process information can be obtained by increasing measuring variables, defined as system features. However, the increase in amount of features will lead to the high computation cost and reduce the efficiency of the process monitoring system. To solve this issue, those features that are redundant or bring an incorrect result should be removed before the process monitoring. Feature selection is the issue of selecting a subset of most informative features from the full set of features. A novel multiple kernel learning based method is proposed for the feature selection. Distinguished with the literature, the feature selection problem is transformed into an optimization problem in the multiple kernel learning, which avoids the drawback called the “monotonic” problem in conventional feature selection methods. Finally, the performance of the proposed method for feature selection in process monitoring is demonstrated through the experiment in Tennessee Eastman benchmark process.