Fault detection based on auto-regressive extreme learning machine for nonlinear dynamic processes

Through utilizing the extreme learning machines (ELM) in modeling the nonlinear dynamic relationship of the time-series data, a novel fault detection approach based on auto-regressive ELM (ARELM) is proposed for nonlinear dynamic processes. The ARELM model attempts to predict each data vector as a nonlinear mapping of its previous measurements, the nonlinear dynamic relations in the time-series samples can thus be interpreted. The residual between the actual data vector and its prediction could be monitored instead for detecting the abnormalities in the defined nonlinear dynamic mechanism. Inherited from the advantages of the ELM, the calculation burden of offline training and online monitoring procedures is much reduced in contrast to kernel based approaches. The efficiency and superiority of the ARELM-based method over other state-of-the-art fault detection techniques for dynamic processes have been demonstrated. (C) 2021 Elsevier B.V. All rights reserved.