Machine Learning for failure prediction: A cost-oriented model selection

Abstract

Industry 4.0 technologies are changing industrial maintenance management. In particular, Machine Learning (ML) techniques have been increasingly used for failure prediction. In this context, the ML model performance influences maintenance costs, as it defines, for example, whether interventions will be carried out before or after the failure. Several studies describe the ML model process design, considering the model evaluation based on standard performance metrics, such as accuracy, recall, precision, and F1-Score. However, when the ML process is applied to multiple models with specific ML pipelines, choosing the ideal one based on these metrics does not always represent the lowest maintenance cost. Therefore, this work proposes an ML process for failure prediction based on binary classification models with a cost-oriented ML model selection phase. Finally, an application example of the proposed process was provided, exemplifying the potential of the proposed model for reducing maintenance costs.