Supervised learning models for health condition-based classification of remaining useful life in predictive maintenance: A preliminary study

This paper presents a study on the utility of supervised machine learning models in estimating the remaining useful life (RUL) of industrial equipment through health condition-based classification approach. A comparison of three main supervised learning models was made in term of accuracy as well as performance using the PHM08 Challenge Dataset from NASA Ames Intelligent Systems Division Diagnostics and Prognostics Group, as the training and test data. This dataset includes synthetic run-to-failure of turbojet engines. The data represents degradation trajectories on a small fleet that includes nine different engines with different initial health conditions. The real flight conditions recorded onboard a commercial jet were taken as inputs to the C-MAPSS model to support engine degradation simulation. The health condition-based classification approach for RUL estimation involves pre-processing of the input dataset and classification of RUL into three health indicative categories which are "red", "amber"and "green". The three machine learning models being compared are random forest, decision tree, and bagging classifier. The results show that random forest is the best among the three models. It has 91.4% accuracy in predicting the remaining useful life based on the proposed classification approach. Future works shall include adapting the model with real feed data related to predictive maintenance of equipment in other domains such as oil and gas industry.