How to Select Algorithms for Predictive Maintenance: An Economic Decision Model and Real-World Instantiation

Predictive maintenance represents a promising application of Artificial Intelligence in the industrial context. The evaluation and selection of predictive maintenance algorithms primarily rely on statistical measures such as absolute and relative prediction errors. However, a purely statistical approach to algorithm selection may not necessarily lead to the optimal economic outcome, as the two types of prediction errors are negatively correlated, thus, cannot be jointly optimized, and are associated with different costs. As the current literature lacks corresponding guidance, we developed a decision model for industrial full-service providers, applying an economic perspective to selecting predictive maintenance algorithms. The decision model was instantiated and evaluated in a real-world setting with a European machinery company providing full-service solutions in the field of car wash systems. Building on sensor data from 4.9 million car wash cycles, the instantiation demonstrates the applicability and effectiveness of the decision model with fidelity to a real-world phenomenon. In sum, the decision model provides economic insights into the trade-off between the algorithms’ error types and enables users to focus on economic concerns in algorithm selection. Our work contributes to the prescriptive knowledge of algorithm selection and predictive maintenance in line with the consideration of different types of cost.