Cost-Sensitive Fault Identification in Predictive Maintenance Applications: a Case Study

The study focuses on health monitoring applications that enables companies to offer smart maintenance services on Fault Identification (FI) and Maintenance Indication (MI) tasks. This work aims to demonstrate that business information can be advantageously embedded in a modified machine learning workflow and that the cost related to the mis-classification error can be reduced. Three ensemble learning techniques are evaluated in order to minimise the mis-classification costs. A case study is analysed in order to show the performance of the proposed strategy in an industrial application of condition-monitoring. A combination of classification algorithms has been tested to check the cost reduction. In particular, Decision Trees, Nearest Neighbours, Linear SVM, Random Forest, AdaBoost, Naive Bayes and XGBoost have been considered. For cost-minimisation, the Voting and Stacking ensembles are employed, together with adaptations of the Minimum Bayes Risk (MBR) classifier. The proposed Maintenance-Based (MB) approach provided Cost Of Risk (COR) savings for the majority of the algorithms tested and the cost minimisation technique has been proven to be effective. Ensemble models provided the considered COR metrics under 110 ppm of the monitored system price.