A multi-instance Hoeffding Anytime Tree approach for interpretable failure detection in industrial systems

The advent of Industry 4.0 has transformed industrial systems through advances in IoT, machine learning, and big data analytics, enabling predictive maintenance to reduce downtime and improve reliability. A core challenge in predictive maintenance is that high-frequency sensor streams contain vast amounts of irrelevant or noisy data, while only certain temporal segments truly reflect equipment degradation. This paper introduces MIHATT, an application-oriented integration of Multi-Instance Learning (MIL) and the Hoeffding Anytime Tree (HATT) that frames failure detection as a time series classification task. MIHATT represents each operating sequence as an ordered bag of overlapping time windows, then iteratively isolates and reinforces the subsequences that capture the true failure concept. Leveraging HATT’s ability to dynamically grow and prune branches, the model adapts to evolving data distributions while remaining efficient for large-scale, high-frequency predictive maintenance data. Minimal preprocessing preserves a white-box structure, allowing direct interpretation of the decision tree to reveal which variables, thresholds, and time windows drive predictions. MIHATT is validated on two widely used predictive maintenance benchmarks–the NASA CMAPSS Jet Engine dataset and the Case Western Reserve University Bearing dataset–covering eight fault-detection scenarios. The results validate the MIHATT approach, achieving an average improvement in accuracy of 27% over existing MIL-based methods while supporting interpretation through explicit decision rules and temporal subsequence identification. All code and processed datasets are publicly available, ensuring reproducibility and supporting future research in interpretable predictive maintenance.