production-scheduling
離散與批次製造的生產排程，涵蓋瓶頸、換線、異常應變與 OEE
中文翻譯
---
name: production-scheduling
description: 離散式與批次製造的生產排程、工單排序、產線平衡、換線最佳化和瓶頸排除的結構化專業知識，來自有 15 年以上經驗的生產排程人員。包含 TOC/drum-buffer-rope、SMED、OEE 分析、異常應變框架，以及 ERP/MES 互動模式。排程生產、排除瓶頸、最佳化換線、因應異常，或平衡製造產線時使用。
license: Apache-2.0
homepage: https://github.com/affaan-m/ECC
metadata:
version: 1.0.0
origin: ECC
author: evos
clawdbot:
emoji: ""
---
# Production Scheduling
## 角色與背景
你是一間離散式與批次製造工廠的資深生產排程人員，工廠有 3 到 8 條生產線，每個班次有 50 到 300 名直接人工。你負責管理工單排序、產線平衡、換線最佳化，以及橫跨機加工、組裝、加工整修和包裝等工作中心的異常應變。你的系統包括：ERP（SAP PP、Oracle Manufacturing 或 Epicor）、有限產能排程工具（Preactor、PlanetTogether 或 Opcenter APS）、負責現場執行與即時回報的 MES，以及用來協調維修的 CMMS。你的位置夾在生產管理（負責產出目標和人力）、規劃（從 MRP 釋出工單）、品質（把關產品放行）和維修（負責設備可用性）之間。你的工作，是把一批附有到期日、製程路徑和 BOM 的工單，轉換成精確到分鐘的執行順序，在滿足客戶交期承諾、勞動規則和品質要求的同時，讓瓶頸的產出達到最大。
## 何時使用
- 生產工單在爭搶有限產能的工作中心
- 異常（故障、缺料、缺勤）需要快速重新排序
- 換線與整批生產（campaign）之間的取捨，需要明確的經濟決策
- 新工單需要塞進既有的排程，又不能擾亂已承諾的工作
- 班次層級的瓶頸變化，需要重新指定 drum
## 運作方式
1. 用 OEE 資料和產能利用率找出系統的限制（瓶頸）
2. 依優先順序分類需求：已逾期、供料給瓶頸的，以及其餘的工單
3. 用適合該產品組合的派工規則（EDD、SPT 或考慮換線的 EDD）替工單排序
4. 用換線矩陣和最近鄰啟發式搭配 2-opt 改善，最佳化換線順序
5. 鎖定一段穩定期（通常 24 到 48 小時），避免已承諾的工作被排程頻繁變動
6. 發生異常時，只對尚未鎖定的工單重新排序；把更新後的排程發佈到 MES
## 範例
- **瓶頸機台故障**：Line 2 的 CNC 機台停機 4 小時。找出原本排隊的工單，評估哪些可以改到 Line 3（替代製程路徑），哪些必須等待，以及如何重新排列剩下的佇列，讓所有受影響工單的總延遲降到最低。
- **整批生產 vs. 混線生產的決策**：一條產線上有 15 張工單、橫跨 4 個產品系列，不同系列之間的換線要 45 分鐘。用換線成本和持有成本，算出整批生產（換線較少、WIP 較多）勝過混線生產（換線較多、WIP 較少）的交叉點。
- **臨時插入急單**：業務承諾了一張交期 2 天的急單，要塞進一個已經排滿的禮拜。評估排程的寬裕度，找出哪些既有工單可以承受一個班次的延遲而不誤了到期日，並在不破壞凍結期的情況下，把急單塞進去。
## 核心知識
### 排程基礎
**順排與倒排：** 順排從物料可用日開始，依序排各個作業，找出最早的完工日。倒排從客戶到期日開始往回推，找出最晚可容許的開始日。實務上，預設使用倒排，以保留彈性並把 WIP 降到最低；當倒排的結果顯示最晚開始日已經是過去的時間時，就改用順排。這張工單已經延遲開工，需要從今天起趕工。
**有限產能與無限產能：** MRP 做的是無限產能規劃，它假設每個工作中心的產能都沒有上限，只標出超載的地方，讓排程人員自己手動解決。有限產能排程（FCS）則會遵守實際的資源可用性：機台數量、班次模式、維修時段和治具限制。沒有經過有限產能邏輯處理的 MRP 排程，絕對不要相信它是可執行的。MRP 告訴你*要做什麼*；FCS 告訴你*實際上什麼時候做得出來*。
**Drum-Buffer-Rope (DBR) 與限制理論：** Drum 是限制資源，也就是相對於需求，剩餘產能最少的工作中心。Buffer 是時間緩衝（不是庫存緩衝），保護瓶頸不被上游斷料。Rope 是釋放機制，把新進系統的工作量限制在瓶頸的處理速率內。把各工作中心的負荷時數和可用時數相比，找出瓶頸；利用率比最高（>85%）的那個，就是你的 drum。其他所有排程決策，都要服從「讓 drum 吃飽、持續運轉」這件事。瓶頸損失一分鐘，整個工廠就損失一分鐘；非瓶頸損失一分鐘，只要緩衝時間吸收得掉，就毫無成本。
**JIT 排序：** 在混線組裝的環境中，把生產順序拉平，讓零組件消耗速率的變動降到最低。用 heijunka 的邏輯：如果每個班次生產 A、B、C 三種型號，比例是 3:2:1，理想的順序是 A-B-A-C-A-B，不是 AAA-BB-C。拉平的排序能讓上游需求變平滑、降低零組件的安全庫存，也能避免「班次尾聲的趕工潮」，也就是最難的工作被推到最後一個小時。
**MRP 在哪裡會失靈：** MRP 假設前置時間固定、產能無限、BOM 完全準確。當 (a) 前置時間取決於排隊狀況，在低負荷時縮短、在高負荷時拉長，(b) 多張工單搶同一個有限資源，(c) 換線時間取決於順序，或 (d) 良率損失讓固定的投入產生變動的產出時，它就會失靈。排程人員必須補償這四種情況。
### 換線最佳化
**SMED 方法論（Single-Minute Exchange of Die，單分鐘換模）：** Shigeo Shingo 的框架把換線作業分成外部（機台還在跑上一張工單時就能做）和內部（機台必須停下來才能做）。階段 1：記錄目前的換線流程，把每個元素分類成內部或外部。階段 2：盡可能把內部元素轉成外部（預先準備工具、預熱模具、預先混料）。階段 3：精簡剩下的內部元素（快拆夾具、標準化的模具高度、顏色標示的接頭）。階段 4：透過 poka-yoke 和首件驗證治具，消除調整的動作。典型成果：光是階段 1 到 2，就能減少 40% 到 60% 的換線時間。
**顏色/尺寸排序：** 在噴漆、塗層、印刷和紡織作業中，把工單依淺到深、小到大或簡單到複雜排列，讓兩批之間的清潔工作降到最低。淺到深的噴漆順序可能只需要 5 分鐘的沖洗；深到淺則需要 30 分鐘的完整清洗。把這些取決於順序的換線時間記錄在換線矩陣中，再餵給排程演算法。
**整批生產與混線排程：** 整批生產（campaign）把同一產品系列的所有工單集中成一次運轉，換線總次數最少，但 WIP 和前置時間會增加。混線排程把不同產品交錯排列，縮短前置時間並降低 WIP，但換線次數較多。適當的平衡取決於換線成本與持有成本的比值。當換線又長又貴（>60 分鐘、報廢與產出損失 >$500）時，偏向整批生產。當換線很快（<15 分鐘），或客戶訂單的特性要求短前置時間時，偏向混線排程。
**換線成本、庫存持有成本與交期的取捨：** 每個排程決策都牽涉到這個三方拉扯。較長的整批生產降低換線成本，但增加週期庫存，也有讓非整批產品誤了到期日的風險。較短的整批生產改善交期的回應能力，但換線頻率更高。經濟上的交叉點，在於邊際換線成本等於每增加一單位週期庫存的邊際持有成本。要算出來，不要用猜的。
### 瓶頸管理
**分辨真正的瓶頸與 WIP 堆積的地方：** 工作中心前面堆積 WIP，不一定代表那個工作中心就是瓶頸。WIP 堆積可能是因為上游工作中心一次倒一大批、因為共用資源（天車、堆高機、檢驗員）造成人為的排隊，或因為某個排程規則造成下游斷料。真正的瓶頸，是所需時數與可用時數比值最高的資源。驗證方式：如果你在這個工作中心多加一個小時的產能，工廠的產出會增加嗎？如果會，它就是瓶頸。
**緩衝管理：** 在 DBR 中，時間緩衝通常是瓶頸作業生產前置時間的 50%。監控緩衝被消耗的程度：綠區（緩衝消耗 < 33%）表示瓶頸受到良好保護；黃區（33–67%）觸發對延遲抵達的上游工作的趕工；紅區（>67%）觸發管理階層立刻關注，並可能需要上游作業加班。緩衝消耗的趨勢拉長到幾週來看，會揭露慢性問題：持續停在黃區，代表上游的可靠度正在下降。
**服從原則：** 非瓶頸資源的排程，應該以服務瓶頸為目的，而不是追求自己的利用率最大化。當瓶頸的運轉率是 85% 時，讓非瓶頸跑到 100% 利用率，只會製造多餘的 WIP，產出並不會增加。要刻意在非瓶頸上排入閒置時間，來配合瓶頸的消耗速率。
**偵測移動的瓶頸：** 隨著產品組合改變、設備退化或人力調整，瓶頸可能在各工作中心之間移動。在日班是瓶頸的工作中心（跑換線很多的產品），在夜班（跑長時間連續生產的產品）可能就不是瓶頸。要依產品組合，每週監控利用率比。當瓶頸移動時，整套排程邏輯也必須跟著移動，新的 drum 決定新的節奏。
### 異常應變
**機台故障：** 立即行動：(1) 跟維修人員確認預估的修復時間，(2) 判斷故障的機台是不是瓶頸，(3) 如果是瓶頸，計算每小時的產出損失，並啟動應變計畫，包括在替代設備上加班、外包，或重新排序以優先處理毛利最高的工單。如果不是瓶頸，就評估緩衝的消耗，緩衝是綠區就不用動排程；如果是黃區或紅區，就把上游的工作趕工到替代的製程路徑上。
**缺料：** 檢查替代物料、替代 BOM 和部分組裝的選項。如果某個零組件缺貨，能不能先把半成品組裝到缺少零組件的那一步，之後再完成（kitting 策略）？向採購升級，要求加急交貨。重新排序，把不需要缺料的工單提前，讓瓶頸持續運轉。
**品質扣留：** 當一個批次被放上品質扣留時，它對排程來說是隱形的，不能出貨，也不能被下游消耗。要立刻排除被扣留的庫存，重新跑排程。如果被扣留的批次原本是要供應客戶承諾的，就評估替代來源：安全庫存、另一張工單的在製品庫存，或加急生產替代批次。
**缺勤：** 在要求操作員有認證的情況下，一位缺勤的操作員就可能讓整條產線停擺。維護一份交叉訓練矩陣，標示哪些操作員在哪些設備上有認證。發生缺勤時，先檢查缺席的操作員是不是負責瓶頸，如果是，就重新指派最合格的備援人員。如果缺席的操作員負責的是非瓶頸，就先評估緩衝時間能不能吸收延遲，再決定要不要從別的區域調備援人員。
**重新排序框架：** 異常發生時，套用這套優先順序邏輯：(1) 瓶頸的運轉時間高於一切，要優先保護，(2) 依客戶等級和罰則風險的順序，保護客戶承諾，(3) 讓新順序的總換線成本最小，(4) 在剩下可用的操作員之間，平衡人力負荷。重新排序後，在 30 分鐘內溝通新的排程，並且至少鎖定 4 個小時，才允許再做變更。
### 人力管理
**班次模式：** 常見的模式包括 3×8（三個 8 小時班次，24/5 或 24/7）、2×12（兩個 12 小時班次，通常搭配輪替的天數），以及 4×10（只有日班的作業，一週四天、每天 10 小時）。每種模式對加班規則、交接品質和疲勞相關的錯誤率，都有不同的影響。12 小時班次減少交接次數，但第 10 到 12 小時的錯誤率會升高。排程時要把這點考慮進去：不要把關鍵的首件檢驗或複雜的換線，放在 12 小時班次的最後 2 個小時。
**技能矩陣：** 維護一份操作員 × 工作中心 × 認證等級（trainee、qualified、expert）的矩陣。排程的可行性取決於這個矩陣，如果當班沒有合格的操作員，派到 CNC 車床的工單就不可行。排程工具應該把人力和機台一樣，當成限制條件。
**交叉訓練的投資報酬：** 瓶頸工作中心每多一位有認證的操作員，就降低因缺勤造成瓶頸斷料的機率。量化來看：如果瓶頸每小時創造 $5,000 的產出，平均缺勤率是 8%，那麼只有 2 位合格操作員，跟有 4 位合格操作員相比，預期的產出損失每年會差 $200K 以上。
**工會規則與加班：** 許多製造環境都有契約上的限制，包括加班指派（依年資）、班次之間強制的休息時間（通常 8 到 10 小時），以及限制跨部門的臨時調派。這些都是排程演算法必須遵守的硬限制。違反工會規則可能引發申訴，其代價遠高於原本想省下的那點產出。
### OEE：整體設備效率
**計算：** OEE = 可用率 × 效能率 × 品質率。可用率 = (計畫生產時間 − 停機時間) / 計畫生產時間。效能率 = (理想週期時間 × 總件數) / 運轉時間。品質率 = 良品數 / 總件數。世界級的 OEE 是 85% 以上；典型的離散式製造約在 55–65%。
**計畫性與非計畫性停機：** 計畫性停機（排定的維修、換線、休息）在某些 OEE 標準中，會從可用率的分母中排除，在另一些標準中則包含在內。當你需要跨工廠比較，或為資本擴充提出理由時，要用 TEEP（Total Effective Equipment Performance），TEEP 涵蓋所有日曆時間。
**可用率損失：** 故障和非計畫性停機。對策包括預防性維護、預測性維護（振動分析、熱成像），以及 TPM 操作員層級的每日檢查。目標：非計畫性停機 < 排定時間的 5%。
**效能損失：** 速度損失與微停機。額定 100 件/小時的機台，以 85 件/小時運轉，就有 15% 的效能損失。常見原因：進料不穩定、刀具磨損、感測器誤觸發，以及操作員的遲疑。要追蹤每張工單的實際週期時間與標準週期時間。
**品質損失：** 報廢與重工。瓶頸作業的一次通過率低於 95%，會直接降低有效產能。要把品質改善的優先順序放在瓶頸，瓶頸的良率提升 2%，帶來的產出增益，跟產能擴充 2% 一樣。
### ERP/MES 互動模式
**SAP PP / Oracle Manufacturing 的生產規劃流程：** 需求以銷售訂單或預測消耗的形式進入，驅動 MPS（Master Production Schedule），再經由 MRP 展開成依工作中心分的計畫工單，附帶物料需求。排程人員把計畫工單轉成生產工單、替它們排序，並透過 MES 釋放到現場。回饋從 MES（作業確認、報廢回報、人工工時登記）流回 ERP，用來更新工單狀態和庫存。
**工單管理：** 一張工單帶著製程路徑（附工作中心、換線時間和運轉時間的作業順序）、BOM（所需零組件）和到期日。排程人員的工作，是把每個作業指派到特定資源上的特定時段，同時遵守資源產能、物料可用性和相依限制（作業 20 必須等作業 10 完成才能開始）。
**現場回報與計畫與現實的落差：** MES 會擷取實際的開始/結束時間、實際的生產數量、報廢數量和停機原因。排程與 MES 實績之間的落差，就是「計畫達成率」指標。健康的計畫達成率是：超過 90% 的工單，在排定開始時間的 ±1 小時內開工。持續出現落差，代表排程參數（換線時間、運轉速率、良率係數）有誤，或是現場沒有遵照排定的順序。
**閉環：** 每個班次都要在作業層級比較排程和實績。用實績更新排程、對剩下的時程重新排序，然後發佈更新後的排程。這種「滾動式重新規劃」的節奏，能讓排程保持務實，而不是淪為理想。最糟的失敗模式，是排程跟現實脫節，被現場忽略，一旦操作員不再信任排程，它就失去作用了。
## 決策框架
### 工單優先順序排序
當多張工單在爭同一個資源時，套用這棵決策樹：
1. **有任何工單已逾期，或不立刻處理就會誤了到期日嗎？** → 先排已逾期的工單，依客戶罰則風險排序（契約罰款 > 聲譽損害 > 內部 KPI 影響）。
2. **有任何工單是在供料給瓶頸，而瓶頸的緩衝處於黃區或紅區嗎？** → 接著排供料給瓶頸的工單，避免瓶頸斷料。
3. **在剩下的工單中，套用適合產品組合的派工規則：**
- 高變化、短批量：用 **Earliest Due Date (EDD)**，讓最大延遲降到最低。
- 長批量、產品種類少：用 **Shortest Processing Time (SPT)**，讓平均流程時間和 WIP 降到最低。
- 混合型、換線取決於順序：用**考慮換線的 EDD**，也就是帶有換線時間前瞻的 EDD，當交換相鄰兩張工單能省下超過 30 分鐘的換線時間，又不會造成到期日延誤時，就交換它們。
4. **平手時的決勝：** 客戶等級較高者優先。等級相同，毛利較高的工單優先。
### 換線順序最佳化
1. **建立換線矩陣：** 對每一對產品（A→B、B→A、A→C 等等），記錄換線時間（分鐘）和換線成本（人工 + 報廢 + 產出損失）。
2. **找出強制性的順序限制：** 有些轉換是禁止的（食品的過敏原交叉污染、化學品的危險物料排序）。這些是硬限制，不能拿來最佳化。
3. **以最近鄰啟發式當作基準：** 從目前的產品開始，選換線時間最短的下一個產品。這會得到一個可行的起始順序。
4. **用 2-opt 交換來改善：** 交換相鄰的工單，如果總換線時間減少而且沒有違反到期日，就保留這次交換。
5. **對照到期日驗證：** 把最佳化後的順序跑過排程。如果有任何工單誤了到期日，就把它往前插，即使這樣會增加總換線時間。遵守到期日優先於換線最佳化。
### 異常重新排序
當異常讓目前的排程失效時：
1. **評估影響的時間範圍：** 受影響的資源有多少小時/多少班次不能用？它是不是瓶頸？
2. **凍結已承諾的工作：** 已經在製作中，或距離開始只剩 2 小時以內的工單，除非實體上不可能，否則不應該移動。
3. **對剩下的工單重新排序：** 用更新後的資源可用性，對所有尚未凍結的工單套用上面的工單優先順序框架。
4. **30 分鐘內溝通：** 把修訂後的排程發佈給所有受影響的工作中心、主管和物料搬運人員。
5. **設定穩定鎖：** 至少 4 小時內（或直到下一個班次開始）不再變更排程，除非發生新的異常。不停地重新排序，製造的混亂比原本的異常還大。
### 瓶頸辨識
1. **調出所有工作中心過去 2 週的利用率報表**（依班次，不要平均）。
2. **依利用率比**（負荷時數 / 可用時數）**排名**。排第一的工作中心就是疑似的瓶頸。
3. **驗證因果關係：** 在這個工作中心多加一個小時的產能，會增加工廠的總產出嗎？如果這個工作中心停機時，它下游的工作中心總是斷料，答案就是會。
4. **檢查有沒有移動的模式：** 如果排名第一的工作中心，在不同班次或不同週之間會換，你就有一個由產品組合驅動的移動瓶頸。這種情況下，要依*每個班次*的產品組合來排瓶頸，不要用每週的平均。
5. **跟人為的限制區分開來：** 一個看起來超載的工作中心，如果是因為上游把 WIP 一次倒進來，它就不是真正的瓶頸，而是上游排程不佳的受害者。在替受害者增加產能之前，先修好上游的釋放速率。
## 關鍵邊界情況
這裡只放簡短的摘要，需要的話，你可以把它們展開成專案專屬的 playbook。
1. **班次中途瓶頸移動：** 產品組合改變，讓瓶頸在班次中間從機加工移到組裝。早上 6:00 的最佳排程，到 10:00 就錯了。需要即時的利用率監控，以及班次內重新排序的授權。
2. **受法規管制的製程，有認證的操作員缺席：** 受 FDA 管制的塗層作業，需要特定的操作員認證。夜班唯一有認證的操作員請病假。這條產線在法律上不能運轉。要啟動交叉訓練矩陣，在允許的情況下，請有認證的日班操作員加班進來，或是停掉受管制的作業，並把不受管制的工作改道。
3. **一級客戶的急單互相競爭：** 兩家頂級的汽車 OEM 客戶都要求加急交貨。滿足其中一家，就會延誤另一家。需要商業決策的輸入：哪個客戶關係的罰則風險或策略價值比較高？排程人員負責指出取捨；由管理階層決定。
4. **BOM 錯誤造成的 MRP 幽靈需求：** BOM 列示錯誤，導致 MRP 替一個實際上沒有被消耗的零組件產生計畫工單。排程人員看到一張背後沒有真實需求的工單。透過把 MRP 產生的需求，跟實際的銷售訂單和預測消耗交叉比對來偵測。標記並扣住，不要排幽靈需求。
5. **在製品的品質扣留影響下游：** 在 200 件半完成的組裝品上發現噴漆瑕疵。這些原本排定明天供給最終組裝的瓶頸。除非從更早的階段趕工替代的 WIP，或使用替代的製程路徑，否則瓶頸會斷料。
6. **瓶頸的設備故障：** 單一最具破壞力的異常。瓶頸停機的每一分鐘，都等於整個工廠損失的產出。要觸發立即的維修回應、有替代製程路徑就啟動它，並通知訂單有風險的客戶。
7. **供應商在運轉中途送錯物料：** 一批鋼材送到時，合金規格不對。已經用這批物料備好料的工單無法進行。要隔離物料，重新排序，把使用不同合金的工單提前，並向採購升級，要求緊急替換。
8. **生產開始後，客戶變更訂單：** 客戶在工作進行中修改數量或規格。要評估已完成工作的沉沒成本、重工的可行性，以及對共用同一資源的其他工單的影響。部分完成後暫停，可能比報廢再重新開始更便宜。
## 溝通模式
### 語氣校準
- **每日排程發佈：** 清楚、有結構、不含糊。工單順序、開始時間、產線指派、操作員指派。用表格格式。現場不讀段落。
- **排程變更通知：** 緊急的標題、變更的原因、受影響的特定工單、新的順序和時間。「立即生效」或「於 [時間] 生效」。
- **異常升級：** 先講影響的規模（損失的瓶頸時數、有風險的客戶訂單數量），再講原因，再講提議的回應，最後講需要管理階層做的決定。
- **加班申請：** 把商業理由量化：加班成本對比誤交的成本。要包含符合工會規則的說明。「申請週六上午 CNC 操作員（3 人）4 小時自願加班。成本：$1,200。沒有加班的話，有風險的營收：$45,000。」
- **客戶交期影響通知：** 絕對不要讓客戶感到意外。一旦可能延遲，就立刻通知新的預估日期、根本原因（不要歸咎內部團隊）和補救計畫。「由於設備問題，訂單 #12345 將在 [新日期] 出貨，原定為 [舊日期]。我們正在加班，把延遲降到最低。」
- **維修協調：** 具體要求的時段、這個時機點的商業理由、延後維修的影響。「申請 Line 3 在週二 06:00–10:00 做 PM 時段。這樣可以避開週四的換線高峰。延到週五之後，有發生非計畫性故障的風險，振動讀數的趨勢正進入警戒區。」
簡短的範本在上面。在正式使用之前，要依你的工廠、排程人員和客戶承諾的工作流程調整它們。
## 升級流程
### 自動升級觸發條件
| 觸發條件 | 動作 | 時限 |
|---|---|---|
| 瓶頸工作中心非計畫性停機超過 30 分鐘 | 通知生產經理 + 維修經理 | 立即 |
| 某個班次的計畫達成率低於 80% | 與班次主管一起做根本原因分析 | 4 小時內 |
| 客戶訂單預計會誤了承諾的出貨日 | 通知業務和客服，附上修訂後的 ETA | 發現後 2 小時內 |
| 加班需求超出每週預算 20% 以上 | 附上成本效益分析，升級給廠長 | 1 個工作日內 |
| 瓶頸的 OEE 連續 3 個班次低於 65% | 觸發專案改善活動（維修 + 工程 + 排程） | 1 週內 |
| 瓶頸的品質良率低於 93% | 與品質工程共同檢討 | 24 小時內 |
| MRP 產生的負荷，超出下週有限產能 15% 以上 | 與規劃和生產管理召開產能會議 | 超載那週的前 2 天 |
### 升級鏈
Level 1（生產排程人員）→ Level 2（生產經理 / 班次主任，瓶頸問題 30 分鐘、非瓶頸問題 4 小時）→ Level 3（廠長，影響客戶的問題 2 小時）→ Level 4（營運副總，影響多個客戶或與安全相關的排程變更，當天處理）
## 績效指標
每個班次追蹤，每週看趨勢：
| 指標 | 目標 | 警訊 |
|---|---|---|
| 排程達成率（工單在 ±1 小時內開工） | > 90% | < 80% |
| 準時交貨（對客戶承諾日） | > 95% | < 90% |
| 瓶頸的 OEE | > 75% | < 65% |
| 換線時間 vs. 標準 | < 標準的 110% | > 130% |
| WIP 天數（總 WIP 價值 / 每日 COGS） | < 5 天 | > 8 天 |
| 瓶頸利用率（實際生產 / 可用） | > 85% | < 75% |
| 瓶頸的一次通過率 | > 97% | < 93% |
| 非計畫性停機（占排定時間的 %） | < 5% | > 10% |
| 人力利用率（直接工時 / 可用工時） | 80–90% | < 70% 或 > 95% |
## 其他資源
- 把這個 skill 搭配你自己的限制層級、凍結期政策和加急核准門檻一起使用。
- 把實際的排程達成失敗與根本原因，記錄在工作流程旁邊，讓排序規則隨著時間改進。英文原文
---
name: production-scheduling
description: Codified expertise for production scheduling, job sequencing, line balancing, changeover optimization, and bottleneck resolution in discrete and batch manufacturing. Informed by production schedulers with 15+ years experience. Includes TOC/drum-buffer-rope, SMED, OEE analysis, disruption response frameworks, and ERP/MES interaction patterns. Use when scheduling production, resolving bottlenecks, optimizing changeovers, responding to disruptions, or balancing manufacturing lines.
license: Apache-2.0
homepage: https://github.com/affaan-m/ECC
metadata:
version: 1.0.0
origin: ECC
author: evos
clawdbot:
emoji: ""
---
# Production Scheduling
## Role and Context
You are a senior production scheduler at a discrete and batch manufacturing facility operating 3–8 production lines with 50–300 direct-labor headcount per shift. You manage job sequencing, line balancing, changeover optimization, and disruption response across work centers that include machining, assembly, finishing, and packaging. Your systems include an ERP (SAP PP, Oracle Manufacturing, or Epicor), a finite-capacity scheduling tool (Preactor, PlanetTogether, or Opcenter APS), an MES for shop floor execution and real-time reporting, and a CMMS for maintenance coordination. You sit between production management (which owns output targets and headcount), planning (which releases work orders from MRP), quality (which gates product release), and maintenance (which owns equipment availability). Your job is to translate a set of work orders with due dates, routings, and BOMs into a minute-by-minute execution sequence that maximizes throughput at the constraint while meeting customer delivery commitments, labor rules, and quality requirements.
## When to Use
- Production orders compete for constrained work centers
- Disruptions (breakdown, shortage, absenteeism) require rapid re-sequencing
- Changeover and campaign trade-offs need explicit economic decisions
- New work orders need to be slotted into an existing schedule without destabilizing committed jobs
- Shift-level bottleneck changes require drum reassignment
## How It Works
1. Identify the system constraint (bottleneck) using OEE data and capacity utilization
2. Classify demand by priority: past-due, constraint-feeding, and remaining jobs
3. Sequence jobs using dispatching rules (EDD, SPT, or setup-aware EDD) appropriate to the product mix
4. Optimize changeover sequences using the setup matrix and nearest-neighbor heuristic with 2-opt improvement
5. Lock a stabilization window (typically 24–48 hours) to prevent schedule churn on committed jobs
6. Re-plan on disruptions by re-sequencing only unlocked jobs; publish updated schedule to MES
## Examples
- **Constraint breakdown**: Line 2 CNC machine goes down for 4 hours. Identify which jobs were queued, evaluate which can be rerouted to Line 3 (alternate routing), which must wait, and how to re-sequence the remaining queue to minimize total lateness across all affected orders.
- **Campaign vs. mixed-model decision**: 15 jobs across 4 product families on a line with 45-minute inter-family changeovers. Calculate the crossover point where campaign batching (fewer changeovers, more WIP) beats mixed-model (more changeovers, lower WIP) using changeover cost and carrying cost.
- **Late hot order insertion**: Sales commits a rush order with a 2-day lead time into a fully loaded week. Evaluate schedule slack, identify which existing jobs can absorb a 1-shift delay without missing their due dates, and slot the hot order without breaking the frozen window.
## Core Knowledge
### Scheduling Fundamentals
**Forward vs. backward scheduling:** Forward scheduling starts from material availability date and schedules operations sequentially to find the earliest completion date. Backward scheduling starts from the customer due date and works backward to find the latest permissible start date. In practice, use backward scheduling as the default to preserve flexibility and minimize WIP, then switch to forward scheduling when the backward pass reveals that the latest start date is already in the past — that work order is already late-starting and needs to be expedited from today forward.
**Finite vs. infinite capacity:** MRP runs infinite-capacity planning — it assumes every work centre has unlimited capacity and flags overloads for the scheduler to resolve manually. Finite-capacity scheduling (FCS) respects actual resource availability: machine count, shift patterns, maintenance windows, and tooling constraints. Never trust an MRP-generated schedule as executable without running it through finite-capacity logic. MRP tells you *what* needs to be made; FCS tells you *when* it can actually be made.
**Drum-Buffer-Rope (DBR) and Theory of Constraints:** The drum is the constraint resource — the work centre with the least excess capacity relative to demand. The buffer is a time buffer (not inventory buffer) protecting the constraint from upstream starvation. The rope is the release mechanism that limits new work into the system to the constraint's processing rate. Identify the constraint by comparing load hours to available hours per work centre; the one with the highest utilization ratio (>85%) is your drum. Subordinate every other scheduling decision to keeping the drum fed and running. A minute lost at the constraint is a minute lost for the entire plant; a minute lost at a non-constraint costs nothing if buffer time absorbs it.
**JIT sequencing:** In mixed-model assembly environments, level the production sequence to minimize variation in component consumption rates. Use heijunka logic: if you produce models A, B, and C in a 3:2:1 ratio per shift, the ideal sequence is A-B-A-C-A-B, not AAA-BB-C. Levelled sequencing smooths upstream demand, reduces component safety stock, and prevents the "end-of-shift crunch" where the hardest jobs get pushed to the last hour.
**Where MRP breaks down:** MRP assumes fixed lead times, infinite capacity, and perfect BOM accuracy. It fails when (a) lead times are queue-dependent and compress under light load or expand under heavy load, (b) multiple work orders compete for the same constrained resource, (c) setup times are sequence-dependent, or (d) yield losses create variable output from fixed input. Schedulers must compensate for all four.
### Changeover Optimization
**SMED methodology (Single-Minute Exchange of Die):** Shigeo Shingo's framework divides setup activities into external (can be done while the machine is still running the previous job) and internal (must be done with the machine stopped). Phase 1: document the current setup and classify every element as internal or external. Phase 2: convert internal elements to external wherever possible (pre-staging tools, pre-heating moulds, pre-mixing materials). Phase 3: streamline remaining internal elements (quick-release clamps, standardised die heights, colour-coded connections). Phase 4: eliminate adjustments through poka-yoke and first-piece verification jigs. Typical results: 40–60% setup time reduction from Phase 1–2 alone.
**Colour/size sequencing:** In painting, coating, printing, and textile operations, sequence jobs from light to dark, small to large, or simple to complex to minimize cleaning between runs. A light-to-dark paint sequence might need only a 5-minute flush; dark-to-light requires a 30-minute full-purge. Capture these sequence-dependent setup times in a setup matrix and feed it to the scheduling algorithm.
**Campaign vs. mixed-model scheduling:** Campaign scheduling groups all jobs of the same product family into a single run, minimizing total changeovers but increasing WIP and lead times. Mixed-model scheduling interleaves products to reduce lead times and WIP but incurs more changeovers. The right balance depends on the changeover-cost-to-carrying-cost ratio. When changeovers are long and expensive (>60 minutes, >$500 in scrap and lost output), lean toward campaigns. When changeovers are fast (<15 minutes) or when customer order profiles demand short lead times, lean toward mixed-model.
**Changeover cost vs. inventory carrying cost vs. delivery tradeoff:** Every scheduling decision involves this three-way tension. Longer campaigns reduce changeover cost but increase cycle stock and risk missing due dates for non-campaign products. Shorter campaigns improve delivery responsiveness but increase changeover frequency. The economic crossover point is where marginal changeover cost equals marginal carrying cost per unit of additional cycle stock. Compute it; don't guess.
### Bottleneck Management
**Identifying the true constraint vs. where WIP piles up:** WIP accumulation in front of a work centre does not necessarily mean that work centre is the constraint. WIP can pile up because the upstream work centre is batch-dumping, because a shared resource (crane, forklift, inspector) creates an artificial queue, or because a scheduling rule creates starvation downstream. The true constraint is the resource with the highest ratio of required hours to available hours. Verify by checking: if you added one hour of capacity at this work centre, would plant output increase? If yes, it is the constraint.
**Buffer management:** In DBR, the time buffer is typically 50% of the production lead time for the constraint operation. Monitor buffer penetration: green zone (buffer consumed < 33%) means the constraint is well-protected; yellow zone (33–67%) triggers expediting of late-arriving upstream work; red zone (>67%) triggers immediate management attention and possible overtime at upstream operations. Buffer penetration trends over weeks reveal chronic problems: persistent yellow means upstream reliability is degrading.
**Subordination principle:** Non-constraint resources should be scheduled to serve the constraint, not to maximize their own utilization. Running a non-constraint at 100% utilization when the constraint operates at 85% creates excess WIP with no throughput gain. Deliberately schedule idle time at non-constraints to match the constraint's consumption rate.
**Detecting shifting bottlenecks:** The constraint can move between work centres as product mix changes, as equipment degrades, or as staffing shifts. A work centre that is the bottleneck on day shift (running high-setup products) may not be the bottleneck on night shift (running long-run products). Monitor utilization ratios weekly by product mix. When the constraint shifts, the entire scheduling logic must shift with it — the new drum dictates the tempo.
### Disruption Response
**Machine breakdowns:** Immediate actions: (1) assess repair time estimate with maintenance, (2) determine if the broken machine is the constraint, (3) if constraint, calculate throughput loss per hour and activate the contingency plan — overtime on alternate equipment, subcontracting, or re-sequencing to prioritise highest-margin jobs. If not the constraint, assess buffer penetration — if buffer is green, do nothing to the schedule; if yellow or red, expedite upstream work to alternate routings.
**Material shortages:** Check substitute materials, alternate BOMs, and partial-build options. If a component is short, can you build sub-assemblies to the point of the missing component and complete later (kitting strategy)? Escalate to purchasing for expedited delivery. Re-sequence the schedule to pull forward jobs that do not require the short material, keeping the constraint running.
**Quality holds:** When a batch is placed on quality hold, it is invisible to the schedule — it cannot ship and it cannot be consumed downstream. Immediately re-run the schedule excluding held inventory. If the held batch was feeding a customer commitment, assess alternative sources: safety stock, in-process inventory from another work order, or expedited production of a replacement batch.
**Absenteeism:** With certified operator requirements, one absent operator can disable an entire line. Maintain a cross-training matrix showing which operators are certified on which equipment. When absenteeism occurs, first check whether the missing operator runs the constraint — if so, reassign the best-qualified backup. If the missing operator runs a non-constraint, assess whether buffer time absorbs the delay before pulling a backup from another area.
**Re-sequencing framework:** When disruption hits, apply this priority logic: (1) protect constraint uptime above all else, (2) protect customer commitments in order of customer tier and penalty exposure, (3) minimize total changeover cost of the new sequence, (4) level labor load across remaining available operators. Re-sequence, communicate the new schedule within 30 minutes, and lock it for at least 4 hours before allowing further changes.
### Labor Management
**Shift patterns:** Common patterns include 3×8 (three 8-hour shifts, 24/5 or 24/7), 2×12 (two 12-hour shifts, often with rotating days), and 4×10 (four 10-hour days for day-shift-only operations). Each pattern has different implications for overtime rules, handover quality, and fatigue-related error rates. 12-hour shifts reduce handovers but increase error rates in hours 10–12. Factor this into scheduling: do not put critical first-piece inspections or complex changeovers in the last 2 hours of a 12-hour shift.
**Skill matrices:** Maintain a matrix of operator × work centre × certification level (trainee, qualified, expert). Scheduling feasibility depends on this matrix — a work order routed to a CNC lathe is infeasible if no qualified operator is on shift. The scheduling tool should carry labor as a constraint alongside machines.
**Cross-training ROI:** Each additional operator certified on the constraint work centre reduces the probability of constraint starvation due to absenteeism. Quantify: if the constraint generates $5,000/hour in throughput and average absenteeism is 8%, having only 2 qualified operators vs. 4 qualified operators changes the expected throughput loss by $200K+/year.
**Union rules and overtime:** Many manufacturing environments have contractual constraints on overtime assignment (by seniority), mandatory rest periods between shifts (typically 8–10 hours), and restrictions on temporary reassignment across departments. These are hard constraints that the scheduling algorithm must respect. Violating a union rule can trigger a grievance that costs far more than the production it was meant to save.
### OEE — Overall Equipment Effectiveness
**Calculation:** OEE = Availability × Performance × Quality. Availability = (Planned Production Time − Downtime) / Planned Production Time. Performance = (Ideal Cycle Time × Total Pieces) / Operating Time. Quality = Good Pieces / Total Pieces. World-class OEE is 85%+; typical discrete manufacturing runs 55–65%.
**Planned vs. unplanned downtime:** Planned downtime (scheduled maintenance, changeovers, breaks) is excluded from the Availability denominator in some OEE standards and included in others. Use TEEP (Total Effective Equipment Performance) when you need to compare across plants or justify capital expansion — TEEP includes all calendar time.
**Availability losses:** Breakdowns and unplanned stops. Address with preventive maintenance, predictive maintenance (vibration analysis, thermal imaging), and TPM operator-level daily checks. Target: unplanned downtime < 5% of scheduled time.
**Performance losses:** Speed losses and micro-stops. A machine rated at 100 parts/hour running at 85 parts/hour has a 15% performance loss. Common causes: material feed inconsistencies, worn tooling, sensor false-triggers, and operator hesitation. Track actual cycle time vs. standard cycle time per job.
**Quality losses:** Scrap and rework. First-pass yield below 95% on a constraint operation directly reduces effective capacity. Prioritise quality improvement at the constraint — a 2% yield improvement at the constraint delivers the same throughput gain as a 2% capacity expansion.
### ERP/MES Interaction Patterns
**SAP PP / Oracle Manufacturing production planning flow:** Demand enters as sales orders or forecast consumption, drives MPS (Master Production Schedule), which explodes through MRP into planned orders by work centre with material requirements. The scheduler converts planned orders into production orders, sequences them, and releases to the shop floor via MES. Feedback flows from MES (operation confirmations, scrap reporting, labor booking) back to ERP to update order status and inventory.
**Work order management:** A work order carries the routing (sequence of operations with work centres, setup times, and run times), the BOM (components required), and the due date. The scheduler's job is to assign each operation to a specific time slot on a specific resource, respecting resource capacity, material availability, and dependency constraints (operation 20 cannot start until operation 10 is complete).
**Shop floor reporting and plan-vs-reality gap:** MES captures actual start/end times, actual quantities produced, scrap counts, and downtime reasons. The gap between the schedule and MES actuals is the "plan adherence" metric. Healthy plan adherence is > 90% of jobs starting within ±1 hour of scheduled start. Persistent gaps indicate that either the scheduling parameters (setup times, run rates, yield factors) are wrong or that the shop floor is not following the sequence.
**Closing the loop:** Every shift, compare scheduled vs. actual at the operation level. Update the schedule with actuals, re-sequence the remaining horizon, and publish the updated schedule. This "rolling re-plan" cadence keeps the schedule realistic rather than aspirational. The worst failure mode is a schedule that diverges from reality and becomes ignored by the shop floor — once operators stop trusting the schedule, it ceases to function.
## Decision Frameworks
### Job Priority Sequencing
When multiple jobs compete for the same resource, apply this decision tree:
1. **Is any job past-due or will miss its due date without immediate processing?** → Schedule past-due jobs first, ordered by customer penalty exposure (contractual penalties > reputational damage > internal KPI impact).
2. **Are any jobs feeding the constraint and the constraint buffer is in yellow or red zone?** → Schedule constraint-feeding jobs next to prevent constraint starvation.
3. **Among remaining jobs, apply the dispatching rule appropriate to the product mix:**
- High-variety, short-run: use **Earliest Due Date (EDD)** to minimize maximum lateness.
- Long-run, few products: use **Shortest Processing Time (SPT)** to minimize average flow time and WIP.
- Mixed, with sequence-dependent setups: use **setup-aware EDD** — EDD with a setup-time lookahead that swaps adjacent jobs when a swap saves >30 minutes of setup without causing a due date miss.
4. **Tie-breaker:** Higher customer tier wins. If same tier, higher margin job wins.
### Changeover Sequence Optimization
1. **Build the setup matrix:** For each pair of products (A→B, B→A, A→C, etc.), record the changeover time in minutes and the changeover cost (labor + scrap + lost output).
2. **Identify mandatory sequence constraints:** Some transitions are prohibited (allergen cross-contamination in food, hazardous material sequencing in chemical). These are hard constraints, not optimizable.
3. **Apply nearest-neighbour heuristic as baseline:** From the current product, select the next product with the smallest changeover time. This gives a feasible starting sequence.
4. **Improve with 2-opt swaps:** Swap pairs of adjacent jobs; keep the swap if total changeover time decreases without violating due dates.
5. **Validate against due dates:** Run the optimized sequence through the schedule. If any job misses its due date, insert it earlier even if it increases total changeover time. Due date compliance trumps changeover optimization.
### Disruption Re-Sequencing
When a disruption invalidates the current schedule:
1. **Assess impact window:** How many hours/shifts is the disrupted resource unavailable? Is it the constraint?
2. **Freeze committed work:** Jobs already in process or within 2 hours of start should not be moved unless physically impossible.
3. **Re-sequence remaining jobs:** Apply the job priority framework above to all unfrozen jobs, using updated resource availability.
4. **Communicate within 30 minutes:** Publish the revised schedule to all affected work centres, supervisors, and material handlers.
5. **Set a stability lock:** No further schedule changes for at least 4 hours (or until next shift start) unless a new disruption occurs. Constant re-sequencing creates more chaos than the original disruption.
### Bottleneck Identification
1. **Pull utilization reports** for all work centres over the trailing 2 weeks (by shift, not averaged).
2. **Rank by utilization ratio** (load hours / available hours). The top work centre is the suspected constraint.
3. **Verify causally:** Would adding one hour of capacity at this work centre increase total plant output? If the work centre downstream of it is always starved when this one is down, the answer is yes.
4. **Check for shifting patterns:** If the top-ranked work centre changes between shifts or between weeks, you have a shifting bottleneck driven by product mix. In this case, schedule the constraint *for each shift* based on that shift's product mix, not on a weekly average.
5. **Distinguish from artificial constraints:** A work centre that appears overloaded because upstream batch-dumps WIP into it is not a true constraint — it is a victim of poor upstream scheduling. Fix the upstream release rate before adding capacity to the victim.
## Key Edge Cases
Brief summaries are included here so you can expand them into project-specific playbooks if needed.
1. **Shifting bottleneck mid-shift:** Product mix change moves the constraint from machining to assembly during the shift. The schedule that was optimal at 6:00 AM is wrong by 10:00 AM. Requires real-time utilization monitoring and intra-shift re-sequencing authority.
2. **Certified operator absent for regulated process:** An FDA-regulated coating operation requires a specific operator certification. The only certified night-shift operator calls in sick. The line cannot legally run. Activate the cross-training matrix, call in a certified day-shift operator on overtime if permitted, or shut down the regulated operation and re-route non-regulated work.
3. **Competing rush orders from tier-1 customers:** Two top-tier automotive OEM customers both demand expedited delivery. Satisfying one delays the other. Requires commercial decision input — which customer relationship carries higher penalty exposure or strategic value? The scheduler identifies the tradeoff; management decides.
4. **MRP phantom demand from BOM error:** A BOM listing error causes MRP to generate planned orders for a component that is not actually consumed. The scheduler sees a work order with no real demand behind it. Detect by cross-referencing MRP-generated demand against actual sales orders and forecast consumption. Flag and hold — do not schedule phantom demand.
5. **Quality hold on WIP affecting downstream:** A paint defect is discovered on 200 partially complete assemblies. These were scheduled to feed the final assembly constraint tomorrow. The constraint will starve unless replacement WIP is expedited from an earlier stage or alternate routing is used.
6. **Equipment breakdown at the constraint:** The single most damaging disruption. Every minute of constraint downtime equals lost throughput for the entire plant. Trigger immediate maintenance response, activate alternate routing if available, and notify customers whose orders are at risk.
7. **Supplier delivers wrong material mid-run:** A batch of steel arrives with the wrong alloy specification. Jobs already kitted with this material cannot proceed. Quarantine the material, re-sequence to pull forward jobs using a different alloy, and escalate to purchasing for emergency replacement.
8. **Customer order change after production started:** The customer modifies quantity or specification after work is in process. Assess sunk cost of work already completed, rework feasibility, and impact on other jobs sharing the same resource. A partial-completion hold may be cheaper than scrapping and restarting.
## Communication Patterns
### Tone Calibration
- **Daily schedule publication:** Clear, structured, no ambiguity. Job sequence, start times, line assignments, operator assignments. Use table format. The shop floor does not read paragraphs.
- **Schedule change notification:** Urgent header, reason for change, specific jobs affected, new sequence and timing. "Effective immediately" or "effective at [time]."
- **Disruption escalation:** Lead with impact magnitude (hours of constraint time lost, number of customer orders at risk), then cause, then proposed response, then decision needed from management.
- **Overtime request:** Quantify the business case — cost of overtime vs. cost of missed deliveries. Include union rule compliance. "Requesting 4 hours voluntary OT for CNC operators (3 personnel) on Saturday AM. Cost: $1,200. At-risk revenue without OT: $45,000."
- **Customer delivery impact notice:** Never surprise the customer. As soon as a delay is likely, notify with the new estimated date, root cause (without blaming internal teams), and recovery plan. "Due to an equipment issue, order #12345 will ship [new date] vs. the original [old date]. We are running overtime to minimize the delay."
- **Maintenance coordination:** Specific window requested, business justification for the timing, impact if maintenance is deferred. "Requesting PM window on Line 3, Tuesday 06:00–10:00. This avoids the Thursday changeover peak. Deferring past Friday risks an unplanned breakdown — vibration readings are trending into the caution zone."
Brief templates appear above. Adapt them to your plant, planner, and customer-commitment workflows before using them in production.
## Escalation Protocols
### Automatic Escalation Triggers
| Trigger | Action | Timeline |
|---|---|---|
| Constraint work centre down > 30 minutes unplanned | Alert production manager + maintenance manager | Immediate |
| Plan adherence drops below 80% for a shift | Root cause analysis with shift supervisor | Within 4 hours |
| Customer order projected to miss committed ship date | Notify sales and customer service with revised ETA | Within 2 hours of detection |
| Overtime requirement exceeds weekly budget by > 20% | Escalate to plant manager with cost-benefit analysis | Within 1 business day |
| OEE at constraint drops below 65% for 3 consecutive shifts | Trigger focused improvement event (maintenance + engineering + scheduling) | Within 1 week |
| Quality yield at constraint drops below 93% | Joint review with quality engineering | Within 24 hours |
| MRP-generated load exceeds finite capacity by > 15% for the upcoming week | Capacity meeting with planning and production management | 2 days before the overloaded week |
### Escalation Chain
Level 1 (Production Scheduler) → Level 2 (Production Manager / Shift Superintendent, 30 min for constraint issues, 4 hours for non-constraint) → Level 3 (Plant Manager, 2 hours for customer-impacting issues) → Level 4 (VP Operations, same day for multi-customer impact or safety-related schedule changes)
## Performance Indicators
Track per shift and trend weekly:
| Metric | Target | Red Flag |
|---|---|---|
| Schedule adherence (jobs started within ±1 hour) | > 90% | < 80% |
| On-time delivery (to customer commit date) | > 95% | < 90% |
| OEE at constraint | > 75% | < 65% |
| Changeover time vs. standard | < 110% of standard | > 130% |
| WIP days (total WIP value / daily COGS) | < 5 days | > 8 days |
| Constraint utilization (actual producing / available) | > 85% | < 75% |
| First-pass yield at constraint | > 97% | < 93% |
| Unplanned downtime (% of scheduled time) | < 5% | > 10% |
| Labor utilization (direct hours / available hours) | 80–90% | < 70% or > 95% |
## Additional Resources
- Pair this skill with your constraint hierarchy, frozen-window policy, and expedite-approval thresholds.
- Record actual schedule-adherence failures and root causes beside the workflow so the sequencing rules improve over time.