AI Harness in Smart Manufacturing (Manufacture)

1. Introduction

1.1 Background

Smart manufacturing is one of the industry directions with the highest barrier to entry for AI Harness adoption. The root cause is not insufficient model capability, but the physicality of the acceptance criteria: a single scheduling error costs inventory and delivery dates, a single missed inspection costs a recall, and a single mis-tuned process parameter costs a batch of scrap. These costs cannot be absorbed by "trying again," so smart manufacturing's requirements on Harness naturally lean toward observability (L5) and constraint-governance (L6).

From a standards perspective, smart manufacturing is one of the most heavily standardized manufacturing paradigms globally, and it has already formed a three-tier, citable specification skeleton:

• Maturity tier: 《Smart Manufacturing Capability Maturity Model》(GB/T 39116-2020) and 《Smart Manufacturing Capability Maturity Assessment Methods》(GB/T 39117-2020), which define the five maturity levels and assessment methods.
• Digital twin tier: the ISO 23247 series 《Automation systems and integration — Digital twin framework for manufacturing》, Part 1:2021 overview and general principles, Part 4:2021 information exchange, Part 5:2026 digital thread, Part 6:2026 digital twin composition.
• Integration and asset tier: IEC 62264 (equivalent to ANSI/ISA-95) enterprise—control system integration five-layer model; RAMI 4.0 (DIN SPEC 91345:2016-04) reference architecture model and the Asset Administration Shell AAS (IEC 63278-1:2023).

The significance of this skeleton is that it has already defined what a manufacturing system should look like. AI Harness does not need to reinvent the wheel; it only needs to align its own six-layer capabilities to this skeleton.

1.2 Definition and Scope

AI Harness in this direction refers to the engineered delivery layer that, in smart manufacturing scenarios, undertakes the four task types of process, scheduling, quality inspection, and digital twin, transforming model output into manufacturing instructions and judgment results that can be issued, traced, and regression-verified.

1.3 Positioning within the AI Harness System

图 1-1｜智能制造 AI Harness 六层定位与瓶颈

数据来源：基于本文分析绘制的示意图。

Leading layers: L2 Tools & Execution + L3 Orchestration & Control. Bottleneck layers: L5 Evaluation & Observability + L6 Governance & Security.

It should be particularly noted that, unlike the four directions of Media / Creative / AI Web Drama / AI Animation, smart manufacturing's quality bottleneck is not at L4. The state of the manufacturing process is already persisted by MES, SCADA, and the digital thread; the model does not need to "remember" the temperature of the previous batch of steel. The real difficulty lies in:

1. Inconsistent indicator systems (L5): Yield, OEE, and miss rate are measured differently across enterprises, and there is no regression-verifiable evaluation set, which makes it hard to prove the difference between "with a large model" and "without a large model."
2. Data sovereignty and explainability (L6): Production-line data is the enterprise's core asset and companies are unwilling to let it leave the network; meanwhile, process engineers are reluctant to adopt process-parameter adjustment suggestions from black-box models.

1.4 Industry Status and Published Quantitative Metrics

《Smart Manufacturing Capability Maturity Model》(GB/T 39116-2020) has broken the requirements down into 228 specific capability requirements. According to public reports from the China Electronics Standardization Institute, more than 150,000 enterprises have conducted online self-assessment and 1,500 enterprises have completed on-site assessment.

The table below lists the published quantitative metrics found; all are annotated with source and calibration level, and citation must not omit the calibration note:

Note: Among the data above, enterprise-case items are mostly vendor self-reported or local-selection report calibration, which constitutes selective disclosure; citation should use the phrasing "as disclosed by the specific disclosing party (enterprise or institution name)" rather than assertion.

2. Glossary

3. Case Studies

3.1 HBIS Group Tangsteel APS Intelligent Scheduling

3.1.1 Background

HBIS Tangsteel's cold-rolling lines comprise 1 pickling-rolling line, 1 continuous annealing line, and 3 galvanizing lines, with many automotive-steel varieties and complex processes. The foreign scheduling system introduced in 2015 could only aggregate data; planners had to manually sort through 3~4 hours of work each day and produce more than 10 reports, and scheduling plans relied heavily on personal experience, making it hard to cope with a high-mix, small-lot order structure.

3.1.2 Solution

On December 26, 2022, the independently developed APS system went live as one of the 142 specialized models of the "Hebei Taihang Iron & Steel Large Model." This large model was jointly developed by Hebei Province, organizing 13 AI companies including Huawei, Alibaba, and Baidu together with 39 steel enterprises, covering 142 scenarios across raw materials, ironmaking, steelmaking, sales, and more.

Breakdown from the Harness perspective:

• L1: Process constraints, line capabilities, and order structure are structured into the scheduling context, replacing the planner's experiential memory.
• L2: APS integrates with MES and inspection/test systems, and scheduling results are issued with work-order semantics.
• L3: Multi-line (pickling-rolling—continuous annealing—galvanizing) collaborative sequencing is a typical multi-stage DAG orchestration.
• L5: On-time delivery rate, inventory turnover, and line efficiency are used as regression metrics for continuous verification.

3.1.3 Results

• Scheduling generation time compressed from 3~4 hours to 0.5 hours.
• Cold-rolling line product inventory reduced 15%.
• Production efficiency up 20%.
• On-time order delivery rate 100%.

Source: Hebei Daily / People's Daily Online, 2025-04-16. Calibration reflects local party-media reporting of the enterprise's deployment.

3.2 Feite (Tianjin) Industrial Multimodal Vertical Intelligent Inspection

3.2.1 Background

Inspection of automotive parts and finished-vehicle paint surfaces has long faced three difficulties: many defect types (scratches, particles, orange peel, sags, etc.), non-uniform standards (manual visual inspection is affected by fatigue and subjectivity), and high cost of misses (defects passing to downstream processes or to the customer). Traditional machine-vision approaches require a large number of annotated samples per defect type and incur high changeover costs.

3.2.2 Solution

It adopts a fused architecture of "industrial vertical large model + multimodal perception" with cloud—edge—terminal collaboration:

• Terminal side: high-precision optical and structured-light sensors for acquisition.
• Edge side: a dedicated AI inference engine performs real-time recognition and classification.
• Cloud side: trains the vertical large model, using transfer learning and incremental training for continuous iteration.

This case was selected as a typical AI application case of the Ministry of Industry and Information Technology (MIIT) in 2025 (285 cases selected nationwide).

Breakdown from the Harness perspective:

• L1: The defect map and inspection standards constitute the judgment context, avoiding "the model defining what a defect is by itself."
• L2: A terminal—edge—cloud three-tier inference toolchain, where the edge side handles real-time judgment and the cloud side handles model iteration.
• L5: The miss rate and first-pass yield form a clear two-metric evaluation loop — note the tension between them: lowering the miss rate usually raises the over-reject rate, so both metrics must be observed simultaneously.

3.2.3 Results

• Inspection miss rate reduced to below 0.1%.
• First-pass yield raised to above 96%.
• Annual labor-cost savings of more than 50% per production line.
• Serving leading companies such as BYD, Volkswagen, China Resources Pharmaceutical, and over 300 others, with a 33% domestic market share (vendor self-reported calibration).

Source: Tianjin Port Free Trade Zone, 2025 (publicity of MIIT typical cases). Market share and savings ratios are vendor self-reported calibration.

3.3 Intelligent Industrial Data Analysis System for Paint Shops

3.3.1 Background

The paint shop is one of the stages in complete-vehicle manufacturing with the highest energy consumption and the highest defect-rework cost. Traditional practice relies on manual, experience-based parameter tuning; energy consumption and defect rate are coupled in a complex way, so point optimization often sacrifices one for the other.

3.3.2 Solution

• Vision layer: a 20-megapixel industrial camera + YOLOv5 for steel-code recognition and detection of 12 types of spray defects.
• Operations layer: multi-dimensional condition monitoring of vibration, temperature, and current + DNN-based predictive maintenance.
• Twin layer: process knowledge graph and digital twin, accumulating 300+ process rules and 200+ equipment models.

The system was selected for the 2025 "Data Elements ×" competition's collection of excellent project cases.

Breakdown from the Harness perspective:

• L4: The 300+ process rules and 200+ equipment models are essentially persisted domain state, depending on no model session memory.
• L5: Energy consumption, false-alarm rate, defect rate, and rework cost form a multi-dimensional evaluation surface, rather than a single accuracy metric.
• L6: Already deployed at NIO and BMW Brilliance, it involves cross-enterprise data boundaries and requires strict tenant and permission isolation.

3.3.3 Results

• Energy consumption reduced 15%~20%, per-vehicle energy down 3.5 kWh, annual energy-cost savings of RMB 1.5 million.
• Fault warnings issued 72 hours in advance, with a false-alarm rate below 3%, unplanned downtime reduced 60%, and annual maintenance-cost savings of RMB 450,000.
• Steel-code recognition and 12-type spray-defect detection accuracy of 98%~99%, spray-defect rate reduced 80%, and annual rework-cost savings of RMB 2 million.
• Already deployed at NIO and BMW Brilliance, production efficiency up 20%~30%, product first-pass yield up 5 percentage points, annual savings exceeding RMB 2.63 million, and annual carbon-emission reduction of about 800 tonnes.

Source: Jilin Provincial Administration of Government Services and Digitalization, 2026-05 (2025 "Data Elements ×" competition's collection of excellent project cases).

4. Practice Standards

4.1 AGENTS.md Specification

4.1.1. AGENTS.md (Smart Manufacturing · Manufacture Direction)

# AGENTS.md —— 智能制造（Manufacture）

## 角色与边界
- 你运行在制造执行环境之上，负责工艺建议、排产方案、质检判定与孪生查询四类任务。
- 你不得直接操控物理设备。所有作用于产线的指令必须经 MES/APS 以工单语义下发，并留痕。
- 你不得把核心工艺参数、配方、产线节拍数据输出到工厂网络之外。
- 判定标准由工艺工程师与质量工程师定义，你只负责执行与举证，不负责设定阈值。

## 环境假设
- 存在 MES、APS、SCADA、QMS、设备台账与工艺卡系统，通过私有 MCP 服务目录暴露只读接口。
- 存在机器视觉平台（相机 + 缺陷检测模型），可返回缺陷坐标、类型与置信度。
- 存在数字孪生与工艺规则库（本案例为 300+ 工艺规则、200+ 设备模型）。
- 模型推理为私有化部署或混合云边缘部署；生产数据不出厂。
- 存在 scripts/ 目录承载确定性计算：SPC 统计、OEE 计算、报表生成、导出封装。

## 上下文加载顺序（Context Budget）
1. 本文件（方向级）与组级 AGENTS.md
2. 本次任务的判定标准与阈值（不可裁剪）
3. 工艺卡、检验标准、BOM 与工艺路线（ISO 23247-3 可观测制造元素属性列表作为 Schema）
4. 相关设备台账与 AAS 子模型（IEC 63278-1:2023）
5. 历史批次数据与最近一次同类任务的检查点
6. 通用工艺知识（可裁剪）
规则：工艺卡与检验标准属于锚定上下文，任何情况下不得被裁剪。

## 工具契约
- 只读类（自由调用）：查询工单、查询设备实时值、查询质检记录、查询孪生状态、检索工艺文档。
- 分析类（调用后必须举证）：排产求解、缺陷分类、参数寻优、异常检测。
- 写操作类（工单语义 + 二次确认 + 留痕）：下发排产、修改工艺参数、判定合格/不合格、创建工单。
- 写操作必须映射为 IEC 62264 L3 的事务语义，禁止绕过 MES 直接写入设备层。
- 确定性计算一律走 scripts/：OEE、SPC 控制限、良率统计、能耗核算、报表导出。禁止让模型逐 token 生成统计结果。

## 任务执行流程（SOP）
1. 确认任务类型（工艺 / 排产 / 质检 / 孪生）与判定阈值，写入任务台账。
2. 加载工艺卡与检验标准，锁定版本号。
3. 取数：明确时间窗、产线、批次、设备范围；记录取数 SQL 或接口调用。
4. 分析：调用分析类工具，输出候选方案与置信度。
5. 校验：对照判定阈值逐项比对；多目标冲突时（如良率 vs 节拍）显式列出权衡。
6. 人工签核：提交候选 + 证据，由工艺或质量工程师确认。
7. 下发：以工单语义写入 MES/APS，记录操作人与时间。
8. 归档：写入数字主线，保存版本、轨迹、成本与耗时。

## 验证与证据要求
- 每条结论必须附：数据来源（系统 + 时间窗 + 口径）、计算脚本路径、数值结果、与阈值的差值。
- 禁止只给"建议提高温度 5℃"这类无证据结论；必须给出依据批次与预期影响区间。
- 引用外部效果数据必须标注口径层级（官方 / 权威媒体 / 协会 / 厂商案例 / 研报 / 第三方估算）。
- 漏检率与过杀率必须成对报告，不得只报漏检率。

## 失败与升级策略
- 数据缺失或口径不一致：停止分析，输出《数据缺口清单》，升级到数据工程师。
- 分析结果落在阈值边界（±5%）内：标注为"边界结论"，强制人工复核。
- 写操作二次确认未通过：立即中止，保持现场，生成待办工单。
- 同一问题连续 2 次求解不达标：升级到工艺工程师，并附上失败样本。
- 升级必须携带任务 ID、阶段、检查点、证据路径、已尝试处理。

## 安全与合规红线
- 生产数据不出厂；核心工艺参数与配方严禁外传。
- 不得打通 IT/OT 边界；与设备层的交互必须经 L3 及以上。
- 操作全量留痕，留痕记录纳入审计。
- 涉及安全联锁（SIS）的参数，智能体不得给出建议，一律转人工。
- 单次任务算力与调用预算设上限，超出即中断并上报。

## 禁止事项
- 禁止绕过 MES/APS 直接下发设备指令。
- 禁止把工艺参数、配方、客户订单信息写入外部服务或日志。
- 禁止在没有判定阈值的情况下给出合格/不合格结论。
- 禁止用模型记忆替代数字主线中的状态。
- 禁止编造标准编号与条款；不确定写 [待填写]。

## 输出格式
任务编号 / 任务类型 / 数据范围 / 判定阈值 / 分析结果（含置信度）/ 权衡说明 / 证据路径 / 待签核项 / 遗留问题

## 评估与自检
- 本次结论是否可复现（相同输入重跑是否得相同结果）？
- 漏检率与过杀率是否成对报告？
- 写操作是否都有工单号与操作人？
- 引用的每个数值是否标注了口径来源？
- 本次任务是否已纳入评估集（含失败样本）？

4.2 SKILL.md Specification

4.2.1. SKILL.md (Smart Manufacturing · Scheduling Plan Delivery)

---
name: manufacture-aps-schedule-delivery
description: 智能制造排产交付技能。当需要在订单交期、物料齐套、设备产能、人员排班等多重约束下生成产线排产方案，并完成约束校验、冲突权衡说明、工单语义下发与归档时使用。适用于钢铁冷轧、离散制造混线、电子制造等多品种小批量场景。
version: 1.0
created: 2026-09-12
---

# 智能制造 · 排产方案交付

## 适用场景
- 在多重约束下生成覆盖全工序、每日、每班、每台设备的排产计划。
- 已有排产系统但准确率不足，需要 AI 重排并给出可解释的约束冲突说明。
- 需要把排产结果以工单语义下发到 MES/APS，并留痕归档。

## 前置条件
- 已加载方向级 AGENTS.md 与组级 AGENTS.md。
- 可访问 MES/APS 只读接口，能获取订单、BOM、工艺路线、设备状态、人员班次。
- 已确定排产目标优先级（交期优先 / 产能优先 / 库存优先），并经计划主管确认。
- 已确定硬约束清单（物料齐套、设备模具、安全库存、交期承诺）与软约束权重。
- 已确定具名签核人（计划主管）。

## 输入
| 输入项 | 说明 | 必需 |
|---|---|---|
| 排产时间窗 | 起止日期与班次粒度 | 是 |
| 订单清单 | 订单号、数量、交期、优先级、客户 | 是 |
| 工艺路线 | 工序序列、标准工时、换型时间 | 是 |
| 资源清单 | 设备、模具、人员、可用产能 | 是 |
| 约束配置 | 硬约束清单与软约束权重 | 是 |
| 目标优先级 | 交期 / 产能 / 库存的排序 | 是 |

## 输出
- 排产方案（工单序列 + 设备 + 班次 + 起止时间）
- 约束冲突说明（哪些约束被满足、哪些被妥协、代价是什么）
- 关键指标（计划准确率、产能利用率、预计库存、交期达成率）
- 校验证据（脚本路径、指标计算过程）
- 签核记录与归档记录

## 执行步骤
1. 读取订单与资源数据，记录取数时间窗与接口调用。
2. 校验数据完整性：缺物料齐套信息或缺设备状态则停止，输出缺口清单。
3. 加载硬约束与软约束权重，构建求解输入。
4. 调用求解器生成候选方案 2~3 套（不同目标优先级）。
5. 用 scripts/ 中的计算脚本核算指标：产能利用率、交期达成率、预计库存、换型次数。
6. 生成约束冲突说明：明确指出被妥协的约束及其代价。
7. 提交计划主管签核。
8. 签核通过后，以工单语义写入 APS/MES，记录工单号与操作人。
9. 归档到数字主线，保存方案版本、求解输入、指标与轨迹。

## 质量标准（DoD）
- 全部硬约束被满足，或每一条被妥协的硬约束都有显式说明与签核。
- 指标由脚本计算而非模型生成，脚本路径可追溯。
- 方案可复现：相同输入重跑得相同结果。
- 工单下发有工单号、操作人、时间戳。
- 引用基准数据（如"排产准确率 50%~70% → 99%+"）时标注来源层级。

## 常见失败与处理
- 数据口径不一致（订单单位与 BOM 单位不符）：停止求解，回到数据治理环节。
- 求解超时：缩小时间窗或降低求解精度，并在输出中标注精度下降。
- 多目标冲突无解：不要强行给方案，输出冲突矩阵交人工决策。
- 换型时间被忽略导致方案不可执行：把换型时间列入硬约束后重求解。
- 计划准确率提升但库存上升：说明存在目标冲突，需人工确认优先级。

## 示例
任务：某冷轧产线（1 条酸轧线、1 条连退线、3 条镀锌线）三日排产。
输入：订单 87 张、交期分布 3~15 天、设备可用率 92%、班次 3 班。
执行：求解 2 套方案（交期优先 / 库存优先），脚本核算指标，输出冲突矩阵。
结果：交期优先方案准时交付率 100%、库存 +3%；库存优先方案准时交付率 94%、库存 -12%。
签核：计划主管选择交期优先方案，工单号 APS-2026xxxx，归档至数字主线。

4.3 Deployment Checklist

5. Summary

AI Harness in smart manufacturing has three characteristics that differ from the content-side directions:

First, it does not need to solve the consistency problem, but it does need to solve the trustworthiness problem. The state of the manufacturing process is already persisted by MES, SCADA, and the digital thread, so L4 is not the bottleneck. What actually blocks deployment is L5 — inconsistent indicator systems and a missing evaluation set, making it impossible to prove "how much benefit AI actually brought" — and L6 — data sovereignty and explainability, which make process engineers reluctant to adopt black-box suggestions.

Second, it has the most complete standard skeleton to align to. GB/T 39116-2020 provides the maturity grading, the ISO 23247 series provides the digital twin and digital thread framework, IEC 62264 provides IT/OT boundaries and transaction semantics, and IEC 63278-1:2023 provides the asset representation. Harness's design should be about "alignment" rather than "invention." In particular, the GB/T 39116-2020 level-3 "integration level" is a prerequisite for adopting multi-agent orchestration, a point that is often overlooked in engineering practice.

Third, its benefits are real but the calibration is complex. Tangsteel's APS compressed scheduling time from 3~4 hours to 0.5 hours, cut inventory 15%, raised efficiency 20%, and achieved 100% on-time delivery; Feite (Tianjin) reduced the miss rate to below 0.1% and raised first-pass yield to above 96%. These figures come from local party-media reports, MIIT typical cases, and vendor self-reports with different calibration levels; citation must annotate them and must not elevate them uniformly to "industry-wide average levels."

Information-Gap Statement

6. References

1. Introduction to Smart Manufacturing Capability Maturity Assessment (GB/T 39116-2020, GB/T 39117-2020) — China Electronics Standardization Institute. https://www.cc.cesi.cn/service/show-2478.aspx
2. ISO 23247-1:2021 Automation systems and integration — Digital twin framework for manufacturing — Part 1: Overview and general principles — ISO. https://www.iso.org/standard/75066.html
3. ISO 23247-5:2026 Digital twin framework for manufacturing — Part 5: Digital thread — ISO. https://www.iso.org/standard/87425.html
4. ISO 23247-6:2026 Digital twin framework for manufacturing — Part 6: Digital twin composition — ISO. https://www.iso.org/standard/87426.html
5. Automation systems and integration — Digital twin framework for manufacturing Part 3: Digital representation of manufacturing elements (national standard plan, identical adoption of ISO 23247-3:2021) — National Public Service Platform for Standards Information. https://std.samr.gov.cn/gb/search/gbDetailed?id=E18C68D8D2A34FFDE05397BE0A0AE2CE
6. ISA-95 architecture and layers — Siemens. https://www.siemens.com/zh-tw/technology/isa-95-framework-layers
7. White Paper: Terminal Automation (IEC 62264 part structure) — TIC4.0, 2025. https://tic40.org/wp-content/uploads/2025/12/White-Paper-Terminal-Automation-1.pdf
8. RAMI 4.0: The Reference Architecture Model for Industrie 4.0 — ERP Information. https://www.erp-information.com/rami-4-0
9. Structure of the Administration Shell — Plattform Industrie 4.0 / ZVEI, 2016. https://industrialdigitaltwin.org/wp-content/uploads/2021/09/01_structure_of_the_administration_shell_en_2016.pdf
10. How a steel company's "strongest brain" was forged (HBIS Tangsteel APS) — Hebei Daily / People's Daily Online, 2025. https://he.people.com.cn/n2/2025/0416/c192235-41197850.html
11. This city's 2 cases selected into MIIT typical AI application cases (Feite Tianjin) — Tianjin Port Free Trade Zone, 2025. https://www.tjftz.gov.cn/contents/6302/381807.html
12. 2025 "Data Elements ×" competition collection of excellent project cases · paint-shop intelligent industrial data analysis system — Jilin Provincial Administration of Government Services and Digitalization, 2026. https://zsj.jl.gov.cn/ztzl/2026sjysdsjlsfs/yxal/202605/t20260519_3631673.html
13. Hai'an launches AI-powered quality inspection pilot — China Daily, 2025. https://subsites.chinadaily.com.cn/nantong/2025-12/16/c_1149318.htm
14. Two cases from Qingdao Economic Development Zone on the MIIT typical-case list (Siruizhuoyuan, Qingdao Port) — Qingdao Economic and Technological Development Zone, 2026. http://qda.qingdao.gov.cn/dtyw/dtxx_114/202607/t20260727_10685848.shtml
15. AGENTS.md official site — Agentic AI Foundation (Linux Foundation). https://agents.md/
16. Agent Skills Specification — agentskills.io. https://agentskills.io/specification