PDMBench: A Standardized Platform for Predictive Maintenance Research

ICLR 2026 Conference Submission13245 Authors

Keywords: Predictive maintenance, benchmark, evaluation

Abstract: Predictive maintenance (PdM) is critical for industrial reliability and cost-efficiency, yet fragmented datasets, inconsistent evaluation protocols, and incompatible preprocessing pipelines hinder progress. We introduce PDMBench, a standardized and extensible platform for exploring and evaluating machine learning models on multimodal time-series data across diverse industrial settings. PDMBench integrates 14 curated datasets spanning bearings, motors, gearboxes, and multi-component systems, capturing real-world complexities such as irregular sampling, heterogeneous sensor modalities, and varying fault modes. To enable fair and reproducible comparison, we design a unified preprocessing pipeline that normalizes signal quality, extracts consistent features, and standardizes input representations, bridging the gap between models requiring handcrafted features and those operating on raw sequences. The benchmark covers two core tasks, fault classification and remaining useful life prediction, and includes 22 models ranging from traditional classifiers to cutting-edge transformers. Models are evaluated across three dimensions: prediction, uncertainty, and efficiency. The PDMBench web interface supports interactive dataset exploration, model comparison, and diagnostic analysis. Experimental results reveal no universal best model, with performance varying by dataset, task, and component type, underscoring the importance of standardized benchmarking. PDMBench enables rigorous, scalable, and interpretable research for real-world predictive maintenance by aligning data, models, and metrics in a reproducible platform.

Primary Area: datasets and benchmarks

Submission Number: 13245