Adaptive Critic Design for Extreme Learning Machines applied to noisy and drifting industrial processes

Abstract

Natural and manmade continuous-time dynamic systems are susceptible to adverse digressions, i.e. periods of rapid deterioration of performance. Conceptually, digressions of a given type are engendered when specific parameters of the system get into definitive combinatorial relationships. Machine Learning (ML) techniques can in principle identify in real time the acquisition (formation) of such relationships and release warnings that consequently command actions that can return the system to normality. Here a manufacturing system, namely the continuous casting process of steel making, is the object of study and both Artificial Neural Networks (ANNs) and Extreme Learning Machines (ELMs) have been shown to perform the above function. A specific characteristic of industrial systems isprocess drift; in such processes the drift is induced into the inter-parameter data relationship learnt by the ML mechanism and the latter has to adapt to the evolving relationship else lose out on accuracy. Adaptive-Critic techniques can function as enablers for the ML mechanism to adapt to this drift. In this work two such Adaptive-Critic techniques are developed, the first for ANNs and then for ELMs, which are demonstrated to work successfully for the industrial process of interest. Importantly, this is the first development in public domain of an Adaptive-Critic technique using ELMs for adaptation to industrial drift. The techniques are generic and amenable for drifting processes in any industrial environment.