摘要
•A multimodal PHM framework fuses sensor, image, and text data via a cache-augmented MLLM.•Cache-Augmented Generation improves grounded, context-aware repair guidance.•Gate-to-gate analysis links maintenance actions to per-unit carbon intensity.•The sheet-metal case study reduces repair time by 64.2% and per-unit carbon intensity by 8.29%.•The framework supports equipment efficiency gains with quantified energy-carbon impacts.
Unplanned equipment downtime undermines production efficiency and energy performance, particularly in sheet-metal manufacturing, by increasing non-productive energy waste and associated carbon emissions. Traditional maintenance strategies often lack condition-aware, timely interventions, and existing Prognostics and Health Management (PHM) systems may not effectively integrate multimodal information for actionable decision-making. This study proposes an integrated industrial informatics framework that bridges operational technology (OT) and information technology (IT) by combining Transformer-based time-series anomaly detection with Multimodal Large Language Models (MLLMs). Addressing the limitations of isolated data silos, the system orchestrates heterogeneous inputs including sensor traces, visual imagery, and textual descriptions to enable comprehensive diagnostics and generate customized repair guidance. A Cache-Augmented Generation (CAG) mechanism is embedded to improve traceability, factual grounding, and response efficiency of maintenance recommendations. Unlike Retrieval-Augmented Generation (RAG), CAG pre-loads verified domain knowledge into the model’s key-value cache, enabling efficient inference without real-time retrieval. Experimental results demonstrate a 95.73% accuracy rate in anomaly classification, a BLEU score of 42, and a BERTScore of 0.84 for recommendation quality. The system also achieves a 64.2% reduction in repair time, improves overall equipment effectiveness (OEE) by 13.22%, and decreases per-unit carbon intensity by 8.29%, while absolute per-shift emissions increase by 11.3% due to higher throughput — a trade-off between carbon efficiency and total output volume, and per-shift output rises 21.3% by reducing non-productive standby time and increasing effective operating hours. Overall, the framework provides traceable, data-driven maintenance support while quantifying energy–carbon impacts at the machine level, offering a practical decision-support pathway to improve both equipment efficiency and environmental performance.
