Customizable Asymmetric Loss Functions for Machine Learning-based Predictive Maintenance

In many predictive maintenance scenarios, the costs for not accurately detecting or anticipating faults can be considerably higher than the cumulative costs for inspections or premature maintenance. However, conventional symmetric loss functions widely used in machine learning cannot reflect such different costs. In this paper, we propose a method to construct asymmetric loss functions for regression tasks that are capable to better reflect this cost imbalance during the training of machine learning models and that allow for modeling the loss function to a) precisely match the cost relation for both kinds of errors where they can be estimated, b) control the impact of outliers, and c) manage the risk of over- or underestimation of the target variable, even when exact costs for (at least) one side are not known. We demonstrate on a realistic data set that the customized asymmetric loss functions can significantly reduce the impact of overestimations of the remaining useful life and can help to take more informed decisions on maintenance planning, leading to more cost-efficient production processes.