Leveraging context information for machine health prognosis with meta-feature-based fitness estimation

Predictive maintenance for machines typically involves model-based condition monitoring systems. However, many state-of-the-art methods are not designed to work under varying production plans or machine operating conditions. Contextual data, such as machine parameters, input material details, and environmental factors, can provide valuable insights into the effectiveness of existing models in changing environments. However, research in this area is limited. This article takes a broad view on adaptive ensemble learning for machine health prognosis from context information-based fitness estimation. Prognostic models are leveraged to increase accuracy under varying operating conditions by using spatial proximity computations between new contextual data features and data from their training phases. Several experiments aim to answer open research questions from previous work. In particular, this study extends the base algorithm from a single one to a set of five common regression algorithms and two classification algorithms and the demonstration domain from turbofans to a simplified synthetic and a milling scenario. In addition, the effect of the available time horizon for online prediction of the context information is evaluated. The results show that the adaptive ensemble method consistently improves degradation percentage estimation accuracy over both single models and conventional averaging ensembles. For classification tasks with highly imbalanced data, such as the milling scenario, the method offers marginal gains, indicating limited benefit from cross-validation weighting in such cases.