Manufacturing Analytics: KPIs, Data & Implementation Guide制造分析深度指南：生产 KPI、数据与实施

Quick Answer: What Is Manufacturing Analytics?快速回答：什么是制造分析？

1. Manufacturing Analytics Purpose, Scope and Decisions1. 制造分析的目的、范围与决策

Start with a decision, not a dashboard. A plant manager may need to protect daily output; a line leader may need to reduce microstops; quality may need to contain a defect family; maintenance may need to prioritize inspections; finance may need to explain conversion-cost variance. Each decision has a different owner, horizon, latency, grain and acceptable uncertainty. Write those elements before selecting technology or KPIs.

应从决策而不是仪表板开始。工厂经理可能需要保障日产出，产线负责人需要减少微停机，质量团队需要控制某类缺陷，维护团队需要确定检查优先级，财务需要解释转换成本差异。每项决策都有不同责任人、时间范围、延迟、粒度与可接受不确定性，应在选择技术或 KPI 前写清楚。

Bound the production system by site, area, line, cell, machine, product family, routing, shift and calendar. Define the observation period, planned production time, excluded states, rework treatment and data cutoff. Separate discrete, batch and continuous processes where event semantics differ. A useful question is specific: “Which verified losses prevented Line 2 from meeting yesterday’s approved plan, and which owner can investigate each loss?”

按工厂、区域、产线、单元、设备、产品族、工艺路线、班次与日历界定生产系统。定义观察期、计划生产时间、排除状态、返工处理与数据截止点；当事件语义不同时，应区分离散、批次与连续流程。有用问题应具体，例如：“哪些已验证损失导致 2 号线昨天未达到批准计划，各损失由谁调查？”

2. Manufacturing Analytics vs MES, IIoT and Dashboards2. 制造分析与 MES、IIoT 及仪表板的边界

MES records and coordinates production execution: orders, operations, material consumption, labor, states and completions. SCADA, PLCs and historians capture process signals and machine states. ERP holds plans, orders, inventory, standard costs and financial actuals. Quality and maintenance applications hold inspections, defects, dispositions, assets and interventions. Manufacturing analytics connects governed evidence from these systems; it does not replace their transactional or control authority.

MES 记录并协调生产执行，包括订单、工序、物料消耗、人工、状态与完工；SCADA、PLC 和历史库采集过程信号与设备状态；ERP 保存计划、订单、库存、标准成本与财务实际；质量和维护应用保存检验、缺陷、处置、资产与干预。制造分析连接这些系统中的受治理证据，但不取代其交易或控制权限。

An operations dashboard is a presentation surface: it monitors approved measures and exceptions. Analytics includes the semantic model, reconciliation, diagnosis, experiment and validation behind that surface. IIoT provides connectivity and observations; analytics converts observations into decision evidence. Keep these boundaries explicit so a visual alert is not mistaken for a validated root cause or an approved control action.

运营仪表板是展示界面，用于监控批准指标与异常；分析还包括界面背后的语义模型、对账、诊断、实验与验证。IIoT 提供连接和观测，分析把观测转为决策证据。应明确边界，避免把视觉告警误当成已验证根因或批准控制行动。

3. Manufacturing Data Sources and the Production Event Model3. 制造数据来源与生产事件模型

Typical inputs include ERP orders and costs; MES operations, counts and genealogy; PLC, SCADA or historian tags; quality inspections and nonconformances; computerized maintenance records; warehouse moves; engineering specifications; labor and shift calendars; and energy meters. Preserve source timestamps, ingestion timestamps, time zone, unit, quality flag and original value. Sensor frequency does not imply analytical value: aggregate only after retaining the event needed for traceability.

典型输入包括 ERP 工单与成本，MES 工序、计数与谱系，PLC、SCADA 或历史库标签，质量检验与不合格，计算机化维护记录，仓库移动，工程规范，人工与班次日历，以及能源表。应保留来源时间、摄取时间、时区、单位、质量标志与原值。传感器频率不等于分析价值；只有在保留可追溯事件后才进行聚合。

Model events such as order released, operation started, state changed, unit completed, material issued, inspection recorded, defect opened, downtime classified and maintenance performed. Give every event a stable identifier, subject, event type, effective time and provenance. Link events to conformed dimensions for site, asset, product, routing, operation, order, batch, shift, reason and calendar. This supports multiple facts without forcing signals, units, defects and costs into one duplicating wide table.

应建模工单释放、工序开始、状态变化、单位完工、物料发料、检验记录、缺陷开启、停机分类与维护执行等事件。每个事件都要有稳定标识、对象、类型、生效时间与来源，并连接工厂、资产、产品、路线、工序、工单、批次、班次、原因与日历等一致维度。这样能支持多个事实表，避免把信号、单位、缺陷与成本塞进会重复数据的宽表。

4. Master Data, Production Context and Time Semantics4. 主数据、生产语境与时间语义

Machine names, product codes, reason codes and routings drift across plants and systems. Create governed mappings while preserving the original identifiers and effective dates. Resolve whether an “asset” means a physical machine, a logical work center, a line or a replaceable component. Product versions, recipes, tooling and engineering changes must be effective-dated because expected rate, yield and process limits can change over time.

设备名称、产品代码、原因代码与工艺路线会在工厂和系统之间漂移。应建立受治理映射，同时保留原标识和生效日期。必须明确“资产”指物理设备、逻辑工作中心、产线还是可更换部件；产品版本、配方、工装与工程变更也须按生效日期管理，因为预期速度、良率与过程限制会随时间变化。

Time is a manufacturing dimension, not a formatting choice. Distinguish event time, system time and ingestion time; define shifts that cross midnight, planned breaks, maintenance windows, warm-up, changeover and daylight-saving behavior. Late-arriving events should restate affected periods under a documented policy. For cycle duration, specify whether blocked, starved, off-shift and paused time is included. Publish the calendar and state precedence used by every KPI.

时间是制造维度，而不是格式选择。应区分事件时间、系统时间与摄取时间，定义跨午夜班次、计划休息、维护窗口、预热、换型及夏令时行为。迟到事件应按书面政策重述受影响期间。计算周期时，要说明是否包含阻塞、缺料、非班次和暂停时间，并公布所有 KPI 使用的日历与状态优先级。

5. Build a Balanced Manufacturing KPI System5. 构建平衡的制造 KPI 体系

No single metric describes a factory. Use a hierarchy: outcome measures for service, output, quality, cost and safety; driver measures for flow, equipment, labor, material and process; and guardrails for data coverage, controls and unintended effects. Select only metrics tied to a named decision. A throughput gain that raises defects or creates excess downstream WIP is not an improvement.

没有一个指标能描述整座工厂。应建立层级：服务、产出、质量、成本和安全的结果指标；流动、设备、人工、物料与过程的驱动指标；以及数据覆盖、控制和意外影响的护栏指标。只选择与明确决策关联的指标。若提高吞吐量同时增加缺陷或下游在制品，就不能称为改进。

6. An Eight-Step Manufacturing Analytics Implementation6. 制造分析八步实施路线

1. Frame the decision. Name the user, action, horizon, value and guardrails.界定决策。明确用户、行动、时间范围、价值与护栏。
2. Bound the process. Select one value stream, line or loss family.限定流程。选择一个价值流、产线或损失类别。
3. Inventory sources. Profile ownership, latency, history and access.盘点来源。分析责任、延迟、历史与访问。
4. Define grain and semantics. Approve events, states, dimensions and formulas.定义粒度与语义。批准事件、状态、维度与公式。
5. Reconcile a baseline. Match counts, time, quantities and cost to control totals.对账基线。把计数、时间、数量和成本与控制总额匹配。
6. Release drill-through. Show KPI, trend, distribution, exceptions and evidence.发布下钻。展示 KPI、趋势、分布、异常与证据。
7. Run the action loop. Assign owner, due date, disposition and result.运行行动闭环。分配责任人、截止日、处置与结果。
8. Validate and scale. Compare outcomes, monitor drift and reuse governed components.验证并扩展。比较结果、监测漂移并复用受治理组件。

Deliver in thin vertical slices. One reconciled line with evidence and a real review cadence is more valuable than an enterprise model nobody trusts. Scale only after operators, engineers, quality, maintenance and finance agree on the baseline and can reproduce representative cases.

应按薄切片交付。一个完成对账、具备证据并进入真实评审节奏的产线，比无人信任的企业级模型更有价值。只有当操作、工程、质量、维护与财务同意基线，并能复现代表性案例后才扩展。

7. Throughput, Cycle Time, Takt, WIP and Bottlenecks7. 吞吐量、周期时间、节拍、在制品与瓶颈

Throughput is accepted output per chosen time unit, not every completion signal. Define good quantity, equivalent units for product mix and the process boundary. Cycle time may mean machine cycle, operation elapsed time or order lead time; label it precisely and report distributions. Takt is available production time divided by required demand for the same horizon and scope. It is a demand rhythm, not automatically the demonstrated machine rate.

吞吐量是所选时间单位内的合格产出，而不是每个完工信号。应定义良品数量、产品组合的当量单位与流程边界。周期时间可能指设备周期、工序经过时间或订单提前期，必须准确标注并报告分布。节拍是同一时间与范围内的可用生产时间除以需求，它是需求节奏，并不自动等于设备实际速度。

Analyze WIP by state, age, location, product and next constraint. Little’s Law—WIP = throughput × flow time—can be a consistency check only when the system is stable and units and boundaries align. Identify constraints from sustained queue, utilization, blocking and starvation evidence, not from one busy snapshot. Test whether improvement at the apparent bottleneck increases accepted system output rather than merely moving inventory downstream.

应按状态、库龄、位置、产品与下一约束分析在制品。小法则“在制品＝吞吐量×流动时间”只有在系统稳定、单位和边界一致时才能作为一致性检查。应通过持续队列、利用率、阻塞和缺料证据识别约束，而不是依赖一张繁忙快照。还要验证表面瓶颈的改进是否增加系统合格产出，而非只把库存向下游移动。

8. OEE, Downtime and Changeover Analysis8. OEE、停机与换型分析

Overall equipment effectiveness is commonly expressed as Availability × Performance × Quality. Availability compares run time with planned production time; Performance compares actual output rate with an approved ideal rate; Quality compares good output with total output. The detailed OEE calculation page should own formula variants, worked calculations and edge cases. Here, OEE is one diagnostic lens inside the wider production system.

设备综合效率通常表示为“可用率 × 性能率 × 质量率”。可用率比较运行时间与计划生产时间，性能率比较实际产出速度与批准理想速度，质量率比较良品与总产出。详细的OEE 计算页面负责公式变体、计算示例与边界情况；本页把 OEE 作为更广生产体系中的一个诊断视角。

Create a mutually exclusive state model and a governed loss tree: planned stop, breakdown, setup, changeover, adjustment, minor stop, reduced speed, startup reject and production reject. Preserve unknown and unclassified time rather than silently redistributing it. For changeovers, compare comparable product transitions and separate last-good-to-first-good time from planned sanitation, approval and ramp-up. Pareto charts should retain duration, frequency and exposure because a rare long failure and frequent microstops require different action.

应建立互斥状态模型和受治理损失树：计划停机、故障、设置、换型、调整、短暂停机、降速、启动废品与生产废品。必须保留未知和未分类时间，不能静默分摊。换型应比较可比产品切换，并区分“最后一件良品到第一件良品”的时间与计划清洁、批准、爬坡。帕累托图要同时保留持续时间、频率与影响，因为低频长故障与高频微停机需要不同措施。

9. Yield, Scrap, Rework and Quality Analytics9. 良率、报废、返工与质量分析

First-pass yield measures units that pass a defined step without rework divided by units entering that step. Rolled throughput yield combines step-level probabilities and exposes hidden factory rework. Scrap rate needs a denominator—started, produced, inspected, material weight or cost—and a clear disposition date. Defect rate must define inspection opportunity because products with more checks otherwise appear worse.

一次通过率衡量无需返工即通过指定步骤的单位数占进入该步骤单位数的比例。滚动通过率组合各步骤概率，能暴露隐藏返工。报废率必须说明分母是投产、产出、检验、材料重量还是成本，并定义处置日期。缺陷率要定义检验机会，否则检查更多的产品看起来反而更差。

Connect nonconformance to product revision, material lot, supplier lot, machine, tool, cavity, recipe, operator qualification, environmental condition and process window only where identifiers and timing support the join. Use control charts for process stability, not as decorative threshold lines. A correlation between temperature and defect does not prove temperature caused the defect; confirm data alignment, confounding, mechanism and a controlled change before claiming root cause.

只有当标识和时间支持连接时，才把不合格与产品版本、物料批次、供应商批次、设备、刀具、模腔、配方、操作资质、环境条件和过程窗口关联。控制图用于判断过程稳定性，而不是装饰性阈值线。温度与缺陷相关并不证明温度导致缺陷；声称根因前必须确认数据对齐、混杂因素、作用机制与受控变更。

10. Schedule Attainment, Capacity and Constraint Analytics10. 计划达成、产能与约束分析

Schedule attainment compares actual accepted completions with the approved frozen schedule using matching product, quantity, operation and period. Preserve schedule versions and the freeze time; comparing actuals with a plan repeatedly changed after execution hides instability. Report early, on-time, late, short and overproduction separately. Link misses to verified loss categories such as material shortage, quality hold, equipment, labor, engineering or priority change.

计划达成使用匹配的产品、数量、工序与期间，把实际合格完工与批准冻结计划比较。必须保留计划版本与冻结时间；若把实际与执行后反复修改的计划比较，会掩盖不稳定。应分别报告提前、准时、延迟、短缺与超产，并把偏差连接到已验证的物料短缺、质量冻结、设备、人工、工程或优先级变化。

Capacity is not one number. Distinguish theoretical, calendar, demonstrated, effective and committed capacity, with product mix, crew, tooling, maintenance and yield assumptions. Utilization can be high while service deteriorates if variability and queues consume the buffer. Scenario analysis should disclose constraints and uncertainty; it supports planner review but does not release orders or approve a production schedule.

产能不是一个数字。应区分理论、日历、实际证明、有效与已承诺产能，并说明产品组合、班组、工装、维护和良率假设。如果波动和队列耗尽缓冲，即使利用率很高，服务也可能恶化。情景分析要披露约束和不确定性；它支持计划人员评审，但不会释放工单或批准生产计划。

11. Maintenance, Reliability and Predictive Evidence11. 维护、可靠性与预测证据

Connect asset hierarchy, operating hours, failure events, work requests, work orders, parts, condition signals and production consequences. MTBF is operating time divided by failures for a defined repairable population; MTTR is repair or restoration time divided by repair events under an explicit clock. State whether waiting for labor, parts or approval is included. Segment by failure mode and operating context rather than averaging unlike assets.

应连接资产层级、运行小时、故障事件、工作请求、维修工单、备件、状态信号与生产后果。MTBF 是特定可维修总体的运行时间除以故障数；MTTR 是在明确计时规则下的维修或恢复时间除以维修事件。要说明是否包含等待人工、备件或批准，并按故障模式和运行语境细分，而不是平均不同资产。

Predictive maintenance requires a labelled outcome, adequate failure history, lead-time window, cost-sensitive threshold and validation against a baseline policy. Random train/test splits can leak future information; use time-aware and asset-aware evaluation. Monitor false alarms, missed failures, warning lead time and maintenance capacity. Predictions prioritize investigation; authorized maintenance processes decide inspection, shutdown, parts and work execution.

预测性维护需要标记结果、足够故障历史、提前预警窗口、考虑成本的阈值，以及相对于基线策略的验证。随机训练测试切分可能泄漏未来信息，应采用时间与资产感知评估。还要监测误报、漏报、预警提前量和维护产能。预测用于确定调查优先级；检查、停机、备件与维修执行仍由授权维护流程决定。

12. Manufacturing Cost, Material and Energy Analytics12. 制造成本、物料与能源分析

Reconcile standard and actual material, labor and overhead at the appropriate order, operation and period grain. Build variance bridges for price, usage, yield, mix, rate, efficiency and volume without double counting. A production loss may affect several financial accounts, so keep operational exposure separate from posted financial impact until finance validates allocation and period treatment.

应在适当工单、工序与期间粒度对账标准和实际物料、人工与制造费用。建立价格、用量、良率、组合、费率、效率与数量差异桥接，避免重复计算。同一生产损失可能影响多个财务科目，因此在财务验证分配和期间处理前，应把运营影响与已入账财务影响分开。

For material, compare issued, consumed, returned, scrapped and backflushed quantities with unit conversion and lot traceability. For energy, allocate meter data only at a defensible level and report kWh per accepted equivalent unit with load, product and weather context where relevant. Avoid distributing a plant total to every unit with false precision. Improvements should be evaluated across cost, output, quality and sustainability guardrails.

物料分析应在单位转换和批次追溯基础上比较发料、消耗、退料、报废与倒冲数量。能源仅在可辩护层级分配，并在相关时结合负载、产品与天气语境，报告每合格当量单位的千瓦时。不能以虚假精度把工厂总量分配给每件产品。改进应同时按照成本、产出、质量与可持续性护栏评估。

13. Anomaly Detection and Root-Cause Analysis13. 异常检测与根因分析

An anomaly is an observation that differs from an expected distribution or rule; it is not automatically a fault. Establish the comparison group, seasonality, operating regime, product and maintenance state before setting limits. Store model version, features, training window, threshold and confidence. Route alerts by expected consequence and give reviewers the preceding state, comparable history and raw evidence.

异常是偏离预期分布或规则的观测，并不自动等于故障。设置限制前应确定比较组、季节性、运行模式、产品与维护状态。保存模型版本、特征、训练窗口、阈值与置信度；按预期后果路由告警，并为复核者提供之前状态、可比历史和原始证据。

A disciplined root-cause workflow moves from symptom to stratification, timeline, candidate mechanisms, tests and verified countermeasure. Use Pareto, five-why or fishbone techniques to structure inquiry, not to manufacture certainty. Validate joins, clocks and denominators first. Then seek evidence that the proposed cause precedes the effect, explains the mechanism, survives relevant controls and changes the outcome when addressed.

严谨根因流程从症状进入分层、时间线、候选机制、测试和经验证对策。帕累托、五问或鱼骨图用于组织调查，而不是制造确定性。应先验证连接、时钟与分母，再寻找证据证明候选原因先于结果、解释机制、在相关控制下仍成立，并在处理后改变结果。

14. Data Governance, Validation and Analytical Controls14. 数据治理、验证与分析控制

Every published measure needs a business owner, technical owner, formula, grain, source fields, state rules, calendar, refresh, tolerance and change history. Display freshness, coverage and unknown classifications near the result. Reconcile production counts with MES or ERP control totals, elapsed time with the calendar, downtime states with total planned time, and cost with the ledger or approved operational record.

每个发布指标都需要业务责任人、技术责任人、公式、粒度、来源字段、状态规则、日历、刷新、容差与变更历史。在结果附近显示新鲜度、覆盖率与未知分类。把生产计数与 MES 或 ERP 控制总额对账，把经过时间与日历对账，把停机状态与计划总时间对账，把成本与总账或批准运营记录对账。

Apply role-based access, least privilege and purpose limitation to production, employee and supplier data. Preserve lineage from result to query, transformation and source. Test duplicate events, missing intervals, impossible state overlaps, negative duration, counter resets, unit mismatches and late data. When definitions change, version and restate deliberately. Auditability is part of analytical quality, especially when evidence affects safety, quality release, labor or financial reporting.

对生产、员工与供应商数据实施基于角色的访问、最小权限与目的限制。保留从结果到查询、转换和来源的血缘。测试重复事件、缺失区间、不可能的状态重叠、负时长、计数器重置、单位不匹配与迟到数据。定义变化时要有版本并有意重述。可审计性是分析质量的一部分，尤其当证据影响安全、质量放行、劳动或财务报告时。

15. Worked Example: Diagnose a Production Shortfall15. 示例：诊断生产短缺

Consider a hypothetical packaging line whose frozen daily plan is 10,000 accepted equivalent units. The first report shows 8,600 accepted units and attributes the gap to “downtime.” Reconciliation finds that 38 minutes of duplicated machine states inflated downtime, while two products used an outdated ideal rate. After corrections, availability is less important than the original chart suggested and reduced speed becomes the largest measured loss. All figures are illustrative, not benchmarks.

假设某包装线冻结日产计划为 10,000 个合格当量单位。初始报告显示合格产出 8,600，并把缺口归因于“停机”。对账发现 38 分钟重复设备状态夸大了停机，两个产品还使用过时理想速度。修正后，可用率不再像原图显示的那么重要，降速成为最大已测损失。所有数字仅为示例，不是基准。

The team stratifies reduced-speed periods by product, tooling, crew and material lot. A recurring pattern appears after changeover for one product-tool combination, but the sample is small. Engineers inspect the approved setup record and run a controlled observation. They find a tooling alignment step is inconsistently documented. After the standard is clarified and training completed, the next four comparable runs show shorter ramp-up with no quality deterioration.

团队按产品、工装、班组与物料批次分层降速期间。某产品—工装组合在换型后出现重复模式，但样本很小。工程师检查批准设置记录并进行受控观察，发现一项工装对准步骤记录不一致。标准澄清并完成培训后，随后四次可比运行显示爬坡缩短，且质量没有恶化。

The result is recorded as provisional evidence, not universal causality. The owner, intervention date, comparable-run criteria, output, yield and safety guardrails remain visible. If performance later drifts, the team can reproduce the data and determine whether product mix, tool wear or another condition changed. This is the difference between an attractive dashboard story and an accountable analytics loop.

结果被记录为暂定证据，而不是普遍因果。责任人、干预日期、可比运行标准、产出、良率与安全护栏保持可见。若表现以后漂移，团队可以复现数据并判断产品组合、刀具磨损或其他条件是否变化。这正是漂亮仪表板故事与可问责分析闭环的区别。

16. AI Manufacturing Analytics With Multi-Source Evidence16. 基于多源证据的 AI 制造分析

AI can help discover tables, map candidate fields, generate analytical queries, summarize loss patterns, compare cohorts and flag quality failures. It should not invent machine states, reason codes, ideal rates, root causes or financial impacts. Natural-language questions require a governed semantic layer so “output,” “downtime” and “yield” resolve to approved definitions. Retain source, cutoff, filters, joins, transformations and model versions with every material answer.

AI 可以帮助发现表、映射候选字段、生成分析查询、总结损失模式、比较群组并标记质量失败。它不能虚构设备状态、原因代码、理想速度、根因或财务影响。自然语言问题需要受治理语义层，使“产出”“停机”和“良率”解析到批准定义。每个重大答案都应保留来源、截止点、筛选、连接、转换与模型版本。

InfiniSynapse fits as an analytics and intelligence layer: connect approved ERP, MES, warehouse, historian, quality, maintenance and cost data; ask cross-source questions; plan and generate queries; inspect results; and preserve reviewable evidence. It is not an MES, PLC, SCADA, production scheduler, machine controller or automatic work-order system. Human and authorized operational controls remain the system of action.

InfiniSynapse 适合作为分析与智能层：连接批准的 ERP、MES、数仓、历史库、质量、维护与成本数据，提出跨源问题，规划和生成查询，检查结果并保留可复核证据。它不是 MES、PLC、SCADA、生产排程器、设备控制器或自动工单系统；人员与授权运营控制仍是行动系统。

This page is part of the broader supply chain analytics content framework. Connect upstream decision context through demand planning, material policy through inventory optimization, and cross-network events through supply chain visibility.

本页属于供应链分析主题内容体系。上游决策语境连接需求计划，物料政策连接库存优化，跨网络事件连接供应链可视化。

17. Common Manufacturing Analytics Failures and Checklist17. 制造分析常见失败与检查清单

• Define the decision, owner, scope, grain, time model and materiality.定义决策、责任人、范围、粒度、时间模型与重大性。
• Map ERP, MES, machine, quality, maintenance, material and cost evidence.映射 ERP、MES、设备、质量、维护、物料与成本证据。
• Approve master-data mappings, state precedence, calendars and unit conversions.批准主数据映射、状态优先级、日历与单位转换。
• Reconcile counts, durations, quantities and value to control totals.把计数、时长、数量与价值对账到控制总额。
• Publish formula, freshness, coverage, unknowns and drill-through with every KPI.随每个 KPI 公布公式、新鲜度、覆盖、未知项与下钻。
• Separate anomaly, hypothesis, validated cause and approved action.区分异常、假设、验证原因与批准行动。
• Compare interventions against a frozen baseline and quality, safety and cost guardrails.根据冻结基线以及质量、安全和成本护栏比较干预。

Frequently Asked Questions常见问题

Sources and Evidence Notes资料来源与证据说明

This guide uses official manufacturing and measurement-science references for data infrastructure, KPI categories, factory outcomes, production domains and reliability validation. The worked example and all numbers in it are explicitly hypothetical, not performance benchmarks.

本指南使用官方制造与测量科学资料支持数据基础设施、KPI 类别、工厂结果、生产领域与可靠性验证。示例及其中所有数字均明确为假设，不是绩效基准。

• NIST — Advanced Manufacturing Data Infrastructure and AnalyticsNIST — 先进制造数据基础设施与分析
• NIST — Smart manufacturing performance metric classificationNIST — 智能制造绩效指标分类
• NIST — Smart Manufacturing Systems Design and AnalysisNIST — 智能制造系统设计与分析
• Microsoft Learn — Intelligent factories and manufacturing outcomesMicrosoft Learn — 智能工厂与制造结果
• Oracle — Manufacturing analytics production domainsOracle — 制造分析生产领域
• NIST — Prognostics and Health Management validationNIST — 预测与健康管理验证