Scholarlink

<p>This study proposes a Policy-Governed Agentic AI for Closed-Loop Manufacturing Control (PGAI-CLMC) framework to support zero-defect manufacturing by innovatively integrating heterogeneous industrial data in a risk-aware manner. <b>PGAI-CLMC fuses information from sensors</b>, Manufacturing Execution Systems (MES), and Enterprise Resource Planning (ERP) systems by temporally aligning anomaly events, semantically mapping operational context, and constructing a traceable digital thread across shop-floor and enterprise layers. At the edge, deep neural networks perform real-time anomaly severity classification and short-horizon performance prediction, while adaptive statistical safeguards—based on Exponentially Weighted Moving Average (EWMA), Cumulative Sum (CUSUM), and capability-aware dynamic limits—continuously recalibrate thresholds under non-stationary conditions. <b>At the cloud level</b>,<b> the proposed policy-governed agentic orchestration layer operationalizes explicit execution constraints (e.g.</b>,<b> SOP-</b>,<b> risk-</b>,<b> and capability-aware gating) and coordinates multi-source evidence aggregation</b>,<b> digital-twin validation</b>,<b> and human-in-the-loop (HITL) review to produce auditable</b>,<b> execution-ready control actions with evidence and rule traceability.</b> Evaluation on 25,275 real-world manufacturing records demonstrates a 22% improvement in anomaly classification accuracy, a 96% reduction in false alarms, a 16% increase in monitoring robustness, and a 19.5% increase in overall operational efficiency. <b>Beyond predictive performance</b>,<b> PGAI-CLMC significantly improves decision reliability and accountability</b> by explicitly linking detected anomalies to MES/ERP execution flows through governed policies and audit trails. The proposed framework is designed to be transferable to other discrete and process manufacturing settings where trustworthy real-time decision-making requires sensor–MES–ERP fusion.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An integrated framework featuring policy-governed agentic AI for closed-loop manufacturing control with multi-source sensor–MES–ERP

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Shih-Chang Lin,
  • Chung-Yang Chen

摘要

This study proposes a Policy-Governed Agentic AI for Closed-Loop Manufacturing Control (PGAI-CLMC) framework to support zero-defect manufacturing by innovatively integrating heterogeneous industrial data in a risk-aware manner. PGAI-CLMC fuses information from sensors, Manufacturing Execution Systems (MES), and Enterprise Resource Planning (ERP) systems by temporally aligning anomaly events, semantically mapping operational context, and constructing a traceable digital thread across shop-floor and enterprise layers. At the edge, deep neural networks perform real-time anomaly severity classification and short-horizon performance prediction, while adaptive statistical safeguards—based on Exponentially Weighted Moving Average (EWMA), Cumulative Sum (CUSUM), and capability-aware dynamic limits—continuously recalibrate thresholds under non-stationary conditions. At the cloud level, the proposed policy-governed agentic orchestration layer operationalizes explicit execution constraints (e.g., SOP-, risk-, and capability-aware gating) and coordinates multi-source evidence aggregation, digital-twin validation, and human-in-the-loop (HITL) review to produce auditable, execution-ready control actions with evidence and rule traceability. Evaluation on 25,275 real-world manufacturing records demonstrates a 22% improvement in anomaly classification accuracy, a 96% reduction in false alarms, a 16% increase in monitoring robustness, and a 19.5% increase in overall operational efficiency. Beyond predictive performance, PGAI-CLMC significantly improves decision reliability and accountability by explicitly linking detected anomalies to MES/ERP execution flows through governed policies and audit trails. The proposed framework is designed to be transferable to other discrete and process manufacturing settings where trustworthy real-time decision-making requires sensor–MES–ERP fusion.