Machine Learning-Based Predictive Maintenance: A Cost-Oriented Model for Implementation

Abstract Predictive Maintenance (PdM) is a condition-based maintenance strategy (CBM) that carries out maintenance action when needed, avoiding unnecessary preventive actions or failures. Machine learning (ML), in the form of advanced monitoring and diagnosis technologies, has become increasingly attractive. Implementing ML-based PdM is a difficult and expensive process, especially for those companies which often lack the necessary skills and financial and labour resources.Thus, a cost-oriented analysis is required to define when ML-based PdM is the most suitable maintenance strategy. The implementation of this strategy involves investment costs in IT technologies, in addition to costs incurred from traditional maintenance activities depending of the performance of the ML model classifier; however, no previous research consider both costs in the economic evaluation of PdM.This paper aims to provide a mathematical model where investment costs are included and the ML performance is evaluated in terms of the probability to correctly intercept faults. A error matrix is used to quantify costs due to maintenance actions. Moreover, the mathematical model provides a cost-based quantitative method, based on the Receiver Operating Characteristics (ROC) curve. This optimizes the decision threshold of the ML model classifier, which allows the maintenance costs to be minimized in comparison to traditional decision threshold optimization methods. Based on the mathematical model, a useful Decision Support System (DSS) that guides PdM implementation is introduced. Finally, the DSS is applied to a real case study to illustrate its applicability.