Advance Predictions of critical digressions in a noisy industrial process- performance of Extreme Learning Machines versus Artificial Neural Networks

Abstract

Manmade continuous-time systems like vehicles, grids and industrial processes are susceptible to adverse digressions in performance which can result in losses to severe breakdowns. Traditionally, emergence of faults in systems was detected by algorithms based on specific sensory signals responding to the incipience of the fault. However, the fault itself is engendered by the fact that certain specific system state parameters acquire a specific combinatorial relationship preceding the fault, thus if the acquisition of this relationship can be detected immediately on formation, neutralization of the fault can be effected early – something particularly relevant to critical systems. The advent of the industrial IoT has made real time systemic data (state parameters) across a chain of processes available at computing platforms, and potentially enabled data-based algorithms – specifically Artificial Neural Networks – to predict such adverse digressions before actual initiation. However, most manmade systems are subject to drift, due to which ANNs trained on data acquired over a certain operational period lose accuracy going forward. Adaptive Critic systems are designed to enable ANNs to neutralize this drift effect, but these need frequent retraining hitting the constraint on computational time. Extreme Learning Machines have emerged as alternatives to ANNs with training times less by orders of magnitude, but their accuracy has to be tested against real noisy industrial data. This work investigates the accuracy of ELM performance versus that of ANNs for such data, and synthesizes ensembles of ELMs to provide accuracy at similar levels as ANNs. This facilitates the incorporation of ELM ensembles into Adaptive Critic frameworks for accurate pre-initiation prediction of faults and related control functions.