Predictive Maintenance (PdM) in industrial IoT seeks to prevent the costly and unsafe consequences of undetected equipment degradation. Time-series anomaly detection (TSAD) is a practical approach to this problem, and Automated Machine Learning (AutoML) can reduce the effort required to configure pipelines for such applications. Prior AutoML systems often: i) impose new programming interfaces or complex external configuration files; ii) limit optimization to algorithm-level hyperparameters; and iii) do not encode PdM-specific requirements. To address these limitations, we present a modular, open-source, and extensible end-to-end AutoML platform that jointly optimizes pipeline stages, preserves common interfaces, and offers declarative control over incident-driven policies. The platform implements 24 TSAD methods across four implementation favors and supports multimodal input data. Evaluation results show that the platform efficiently identifies competitive configurations across heterogeneous PdM scenarios and, when combined with multimodal signals, yields improvements in prediction performance of up to 19%.
Papadopoulos et al. (Thu,) studied this question.