Human-Centered Maintenance Decision-Making in Smart Manufacturing Enhanced by English Technical Documents and Human-in-the-Loop Collaboration

INTRODUCTION: Predictive maintenance is essential in human-centered smart manufacturing, yet existing methods often convert sensor data into fault diagnosis or RUL prediction without effectively transforming English technical documents into actionable maintenance constraints. OBJECTIVES: This study aims to develop an interpretable human-in-the-loop maintenance decision-making framework that integrates equipment states, English technical clauses, and human feedback to support constraint-aware maintenance actions. METHODS: An English technical document-enhanced framework is proposed, consisting of the English Document-Induced Constraint Learning (ELIC) module and the Human Feedback Loss Optimization (HFLO) module. ELIC models maintenance clauses from manuals, SOPs, and safety documents as state–document–action feasibility constraints. HFLO converts preference and risk feedback into optimization signals. A constraint-aware inference strategy combines task prediction scores, document feasibility scores, and feedback consistency scores to rank candidate actions. RESULTS: Experiments on MetroPT-3 and XJTU-SY under the controlled protocol show that the proposed method achieves the lowest SP-CVR on both datasets and the highest F1 on MetroPT-3. On XJTU-SY, TCN–Transformer achieves a lower RMSE, while the human-in-the-loop baseline achieves a higher SP-FAS. Ablation results show that ELIC, HFLO, and constraint-aware inference affect different task and protocol-consistency metrics. CONCLUSION: The proposed framework provides a traceable approach for jointly incorporating English technical clauses and feedback into maintenance action ranking under the constructed controlled protocol.