以空調箱冷凍機效率異常、超純水系統壓差異常為例,展示可插拔語意執行框架(SAM)從感測觸發到跨域傳播的完整決策鏈。
以下影片說明本研究核心概念——如何透過可插拔領域知識模組,讓設施管理系統快速部署、跨域複用,並保留完整的決策稽核紀錄。
每一個 Domain Pack 封裝特定領域的觸發條件、語意規則、執行動作與稽核邏輯。同一套通用執行環境(SAM Runtime)可載入不同 Pack,無需重新開發即可處理新場域。
TIAA 是最小可執行語意單元。每個感測異常事件被解析為完整的四元組,執行完畢後透過 Propagation 機制向鄰近知識域廣播,維持整體系統語意一致性。
半導體廠潔淨室的空調箱依賴冷凍機維持恆溫恆濕。當冷凍效率下降,晶圓良率直接受影響。以下展示 HVAC PdM Pack 如何完成從感測到派工,再到跨域傳播的完整鏈路。
SHACL 語意約束規則:定義「什麼情況下算異常」,系統自動比對感測值是否觸發。
// SHACL 觸發規則 sh:ChillerCOPConstraint sh:path fab:currentCOP ; sh:maxInclusive 3.2 ; sh:minCount 1 . // Neo4j Cypher — 異常偵測 MATCH (s:Sensor)-[:MONITORS]->(c:Chiller) WHERE s.COP < 3.2 AND duration.between(s.lastNormal, s.now).minutes > 15 RETURN c.id, s.COP, s.timestamp
Issue {
id: "ISS-2026-0612-001",
type: "ChillerEfficiencyDegradation",
severity: "HIGH",
affectedAsset: "AHU-B3-07",
probableCause: ["refrigerant_leak", "fouling"],
impactZone: "CleanroomB3",
riskScore: 0.82
}
WorkOrder {
id: "WO-2026-0612-038",
priority: "URGENT",
type: "ChillerInspection",
sla: "4h",
linkedIssue: "ISS-2026-0612-001",
autoGenerated: true
}
PROV-O 稽核紀錄:記錄「誰、何時、根據哪筆資料、做了什麼決定」,可完整重建執行路徑。
// PROV-O 稽核紀錄
prov:Activity { id: "exec-WO-038" }
prov:wasAssociatedWith :Engineer_Lin ;
prov:wasInformedBy :Issue-001 ;
prov:startedAtTime "2026-06-12T14:52:03Z" ;
prov:used :ChillerSensorReading-4821 .
// 跨域傳播事件(SAM Propagation Bus)
PropagationEvent {
source: "HVAC-PdM-Pack",
trigger: "WO-038-completed",
timestamp: "2026-06-12T16:34:00Z",
targets: ["EnergyMonitoring", "ProductionScheduling", "AssetLifecycle"]
}
超純水(UPW)是半導體製程清洗用水,水質與水壓直接影響良率。以下展示同一套 SAM 框架換上 UPW Monitoring Pack 後的執行過程——框架完全不動,只換了知識模組。
SHACL 語意約束規則:定義濾膜更換的壓差門檻,系統自動比對是否觸發換濾流程。
sh:UPWPressureConstraint sh:path fab:differentialPressure ; sh:minInclusive 0.30 ; sh:message "Filter fouling threshold exceeded" . ΔP_in: 2.41 bar(進水) ΔP_out: 2.07 bar(出水) ΔP: 0.34 bar → TRIGGER ✓
Issue {
type: "UPWFilterFouling",
severity: "CRITICAL",
affectedFilter: "UPW-LOOP-F03",
estimatedTimeToFailure: "6.5h",
downstreamImpact: ["Wet-Etch-A", "CMP-B"],
riskScore: 0.91
}
Step 1: CREATE WorkOrder(type: EmergencyFilterChange, sla: 2h)
Step 2: QUERY Inventory("UPW-F03-membrane") → stock: 4 units ✓
Step 3: NOTIFY ProductionScheduler(
downtime: 90min,
stations: ["Wet-Etch-A", "CMP-B"])
Step 4: ESCALATE FacilityManager(reason: riskScore > 0.9)
PROV-O 稽核紀錄:同樣格式,確保 UPW 案例的決策路徑可被完整重建與稽核。
prov:Activity { id: "exec-WO-UPW-041" }
prov:wasAssociatedWith :Engineer_Chen ;
prov:wasAssociatedWith :Manager_Production ;
prov:used :FilterPressureReading-9923 ;
prov:used :InventoryQuery-result ;
prov:endedAtTime "2026-06-12T16:18:44Z" .
PropagationEvent {
source: "UPW-Monitoring-Pack",
trigger: "WO-UPW-041-completed",
timestamp: "2026-06-12T16:18:44Z",
targets: ["WaterQualityRecord", "ProductionScheduling", "SupplyChain"]
}
以下數據來自發表於 Advanced Engineering Informatics(AEI)之正式論文,以 HVAC 預測性維護情境為驗證場域。