Extreme Learning Machine-Based Operational State Recognition: A Feasibility Study with Mechanical Vibration Data

Abstract

The benefits of digital twins and accurate near real-time on-site condition monitoring of heavy machinery or load-bearing structures are undeniable. Both demand computationally light and accurate models based on continuously measured data. Extreme Learning Machine (ELM) algorithm provides the means for building accurate and fast predicting classification models. Therefore, the feasibility of the ELM algorithm for building models for near real-time operational state recognition of a rotating machine was studied. Three different models, called one, two, and six cycles, built using the ELM algorithm were compared with corresponding models trained using Support Vector Machine (SVM) and linear regression (LR) algorithms based on their accuracy and prediction times. The comparisons show that the SVM algorithm produces the best accuracy, but with the cost of high prediction times. The LR models have the lowest prediction time. In contrast, the ELM model for the two cycles presents better performance than the corresponding LR and SVM models when the combination of accuracy and the prediction time is considered. The great benefit of the ELM method comes from its mathematical properties: new data can be added to the ELM model without the need to retrain the whole model, and the model is competent to take strong nonlinearities into account. Thus, the possibilities of the ELM algorithm to act as a novelty detector in operational state recognition shall be investigated.