Automated algorithm selection for Predictive Maintenance: Advances and challenges

Abstract

Applications of Predictive Maintenance (PdM) in manufacturing systems with changing operating conditions still face significant challenges. In particular, the selection and application-specific configuration of prognostic algorithms often require expert knowledge and substantial computational resources, limiting scalability and broad adoption. Automated Machine Learning (AutoML) and Meta-Learning offer promising strategies to address these barriers; however, existing approaches frequently remain misaligned with the practical requirements of PdM in real-world industrial environments. This paper presents a systematic literature review of Meta-Learning techniques in the context of PdM. We first analyze the typical model development pipeline and emphasize the need for increased automation. Furthermore, general challenges associated with implementing PdM in industrial settings are discussed. After formalizing the problem as a Combined Algorithm Selection and Hyperparameter Optimization (CASH) task, a detailed literature analysis is conducted. The core contribution of this work is a structured assessment of Meta-Learning methods applied to time series forecasting and anomaly detection–two fundamental tasks in PdM. The review demonstrates the potential of Meta-Learning to improve algorithm and hyperparameter selection in PdM scenarios, while simultaneously identifying critical research gaps: (i) the underutilization of unsupervised approaches in low-label environments, (ii) the absence of adaptive methods capable of addressing dynamic industrial conditions, and (iii) the lack of robust integration strategies for deployment in operational settings. These findings provide a roadmap for future research at the intersection of Meta-Learning and industrial PdM.