Auto-feature extraction for industrial machine learning based on system dynamics diagram

Abstract

In industrial data analytics, many useful features are implicitly reflected in system connection and interaction. Manual feature extraction is tedious and low-efficiency. An automated feature extraction approach based on system dynamics diagram is proposed in this paper. System mechanism and control policy are captured by system dynamics diagram (SDG). Four typical SDG structures identification algorithms are introduced, including anchor variable, single path, parallel branch, and closed-loop. Feature extraction operators can be configured according to graph structures. The approach is illustrated with a feedwater flow control system. Such a formalized graph structure based automated machine learning can significantly reduce feature extraction efforts.