Evaluation of the Improved Extreme Learning Machine for Machine Failure Multiclass Classification

Abstract

The recent advancements in sensor, big data, and artificial intelligence (AI) have introduced digital transformation in the manufacturing industry. Machine maintenance has been one of the central subjects in digital transformation in the manufacturing industry. Predictive maintenance is the latest maintenance strategy that relies on data and artificial intelligence techniques to predict machine failure and remaining life assessment. However, the imbalanced nature of machine data can result in inaccurate machine failure predictions. This research will use techniques and algorithms centered on Extreme Learning Machine (ELM) and their development to find a suitable algorithm to overcome imbalanced machine datasets. The dataset used in this research is Microsoft Azure for Predictive Maintenance, which has significantly imbalanced failure classes. Four improved ELM methods are evaluated in this paper, i.e., extreme machine learning with under-sampling/over-sampling, weighted-ELM, and weighted-ELM with radial basis function (RBF) kernel and particle swarm optimization (PSO). Our simulation results show that the combination of ELM with under-sampling gained the highest performance result, in which the average F1-score reached 0.9541 for binary classification and 0.9555 for multiclass classification.