Classifier-Grounded Small Language Models for Predictive Maintenance

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Machine Health Copilot is a predictive-maintenance system designed to make machine-learning diagnostics more useful and trustworthy for equipment operators. The system separates prediction from explanation: a numerical model determines the machine's condition, while the language layer is grounded in that structured assessment rather than reasoning directly from raw sensor readings. This design is intended to reduce omitted conditions, altered values, and unsupported conclusions while keeping the system's outputs traceable. The prototype combines a multilabel CatBoost classifier with a fine-tuned Qwen3-1.7B small language model and is evaluated on the AI4I 2020 predictive-maintenance benchmark with additional constructed stress-test cases. The study compares raw-reading and classifier-grounded language-model configurations and evaluates condition completeness, numerical fidelity, unsupported-question handling, and behavior on overlapping and boundary cases.