Real Time Predictions Of Adverse Digressions In Critical and Noisy Industrial Processes Using LSTMs

Abstract

Natural and manmade continuous-time dynamical systems are susceptible to adverse digressions, i.e. periods of rapid deterioration of performance. Advance prediction of such adverse digressions, if attained successfully, would be of great value to human civilization, and in particular to critical industrial processes, because such predictions will trigger either corrective or mitigating action. Machine Learning techniques, specifically the “learning-functionality-from-data” paradigm of supervised learning, is one of the most appropriate mechanisms for attaining the same. Prior work has shown that conventional feedforward Artificial Neural Networks (ANNs), and its methodologically close Extreme Learning Machines (ELMs), have been able to achieve a fair measure of success in this direction. However, both these techniques work on taking parameter inputs at a single time step of a running process. It is conjectured here that more accurate and deeper functional relationships between process-parameters and concomitant adverse digressions are encapsulated within time-sequences of these parameter-vectors. The authors here use Machine Learning techniques for sequence modelling, specifically Long Short Term Memory (LSTM) networks, to extract these relationships and demonstrate the above conjecture, taking as example a noisy, critical process within the steel manufacturing chain.