Time-series failure prediction on small datasets using machine learning

Abstract

Condition-based maintenance is a decision-making strategy using condition monitoring information to optimize the availability of operational plants. In this context, machine learning techniques are useful and have been used in predicting the remaining useful life (RUL) of equipment to ensure the overall safety and reliability of the system through maintenance policies and, consequently, reducing costs arising from the failure. These databases are not large which is tricky for data-driven models. In this study, we consider five different databases containing the failure times from distinct real-world equipment. Here, four different regression algorithms were compared for RUL prediction, namely: Support Vector Regression (SVR), Decision Tree (DT), Multilayer Perceptron (MLP) and K-Nearest Neighbors (KNN). Furthermore, aiming to improve the data quality, the Empirical Mode Decomposition (EMD) was used, which is responsible for pre-processing the input data used on the predictive modeling. We optimize the models hyperparameters using grid-search cross-validation algorithm and the performance of each model is compared using the normalized root mean squared error (NRMSE). Considering the datasets analyzed, KNN model proves to be the most promising to perform the prognostic task in small datasets, adapting itself to the distinct characteristics of the different databases. In addition, we mention the better performance after optimizing the hyperparameters, which avoided overfitting problems and had a low computational cost for the problems analyzed here.