Advance Prediction of Adverse Digressions in Continuous-Time Systems Using ANN Kernels: A Generic Approach Instantiated in Steel Manufacturing

摘要

A domain-independent generic methodology is developed for online prediction of rapid adverse digressions in continuous-time systems, preceding the actual incipience of such digressions. The complete methodology itself consists of three stages. The first two stages are domain specific and involve statistical analysis and standard prediction tools like Artificial Neural Networks. The third stage that transforms the ANN outputs into a reliable measure of the instantaneous digression of the system is generic across domains and is the contribution of this work. The core novelty enabling this transformation is a paradigm shift in assessment of the ANN output, from that of "classification" to that of "continuity". The development of this methodology is performed on a specific industrial process-continuous casting in steel manufacturing-and described in this paper. This can be applied with minor customization to continuous-time systems in domains such as the biological, industrial processes, vehicles, and economic and financial systems.