Predictive maintenance in predicting machine failures at unilever: A log-based approach using machine learning algorithm

Abstract

Log-based predictive maintenance creates opportunities for companies to apply proactive maintenance strategy using existing data without expensive hardware installation. The primary goal of this research is to develop a predictive maintenance model that predicts machine failure up to 7 days. The prognostics model is trained using real log datasets collected from 2 identical bottling machines including log messages, events and information taken from computerized maintenance management system (CMMS) at Unilever. Subsequently, maintenance decisions are decided at the beginning of the planning period with a rule-based model that consists of spare parts cost, lead time, labor costs, failure cost and production quantity using prognostics information from the predictive model. In total, 12 different models were trained using 2 balancing techniques, 3 machine learning algorithms and 2 datasets. The challenges posed by imbalanced and limited dataset were overcome using 3 fine-tuning techniques: Grid Search cross validation, ensemble learning methods and precision-recall curve optimization. The final prognostics model is able to predict machine failure with near-perfect accuracy and precision up to 3 days. Finally, hypothesis testing was conducted to validate the model's transferability on other identical machines.