移动通信网 程然
高炉不会因为有人要核对 AI 的答案而暂停运转。炉内温度、煤气流和原料状态必须时刻处于受控之中——这正是工业 AI 与办公室聊天机器人的本质区别。南京钢铁集团有限公司(南钢)与华为及生态伙伴合作,将自主研发的"元野"钢铁大模型部署到炼铁、轧钢、质检和环保管控等核心生产环节,给出了工业智能化的一份务实答卷:炉温预测可提前两小时、准确率超过 90%,优质铁水产出提升逾 10%,年降本超过 1 亿元人民币。
概率AI与机理模型的双轮驱动
这一部署的关键,在于华为并非简单地把一个大模型"搬进"钢厂。元野系统将两类能力结合在一起:一类是擅长识别模式、预测走势的概率式 AI;另一类是依据温度、压力、化学成分等物理关系进行精确计算的确定性工业机理模型。前者从海量运行数据中捕捉规律,后者为预测结果提供独立的过程物理校验,两者互为支撑。
这种架构设计回应了工业 AI 落地的核心痛点。钢厂预测必须考虑设备行为、材料属性以及调整工艺可能引发的连锁反应,一个"看起来合理"的答案远远不够。确定性模型对给定输入产出明确结果,其精度取决于输入质量和物理模型的有效性——这既保证了可解释性,也界定了系统能力的边界。在这一框架下,工业智能体(Agent)本质上是调用模型、运行数据和预设工作流完成特定任务的软件应用,监控设备、预警异常、推荐排产调整,但并不脱离人的监督直接操控机器。
【关键数据】
- 炉温预测提前量: 2小时
- 炉温预测准确率: 超过90%
- 优质铁水产出提升: 超过10%
- 新产品研发周期: 缩短50%
- 年成本节约: 超过1亿元人民币(约1500万美元)

从高炉到环保的全流程渗透
高炉内部无法直接目视检查,操作人员长期依赖传感器读数和经验积累来判断炉况。元野系统融合炉内传感数据与冶金分析,模拟炼铁过程,让操作人员能够提前预判炉温变化并识别潜在问题。机器学习模型捕捉传感数据中正在形成的异常模式,物理模型则独立验证当前工况下高炉应有的运行状态——双重验证给了操作人员足够的时间窗口,在炉况恶化演变为生产事故之前介入处置。
在产线下游,任务形态随之变化。轧制钢板需要在快速变化的工况下精确控制,元野的轧制力预测模型响应时间达到毫秒级。质检环节则将机器人、成像技术与 AI 融合,南钢的自动化金相检验系统可完成试样研磨抛光、图像采集和钢材微观组织评级全流程。环保管控方面,面向煤气与烟气硫治理的 S-Link 智能体整合了 602 个工艺参数,结合图像与知识能力,实现统计预测、机理模型和数学优化的协同——提前一小时预测硫排放风险,数秒内计算出煤气掺混方案,并辅助排放配额分配。
工业AI的价值锚点
从更大范围看,南钢披露其新产品研发周期已经减半,年成本节约超过 1 亿元人民币。"数智化转型天然是以人为中心的,其目的不是替代人,而是释放人的潜能。"南钢数智融合发展部总经理、AI 研究院院长李金言的这一表述,为这场转型定下了基调。华为钢铁有色 BU 工业解决方案高级专家程永福则强调,AI 应贯穿生产、运营和管理全链条,让改进在企业内部持续累积。
南钢案例的行业启示在于:工业 AI 无法被简化为大语言模型的单点应用。有些功能需要概率式模式识别,有些则依赖确定性计算、数学优化和成熟的工业控制系统,真正的价值来自将它们有机整合。华为提供算力与 AI 技术底座,南钢贡献生产数据、机理模型和运营经验——这种"技术厂商+行业Know-how"的深度耦合模式,或将成为传统重工业智能化改造的主流路径,其示范效应值得持续追踪。
出处:Huawei and NISCO put AI to work in the steelworks
英文原文Markdown Content: ### _A Chinese steelmaker is combining probabilistic AI with deterministic industrial models, using Huawei technology to improve furnace operations, product quality and energy efficiency._
A blast furnace cannot pause while someone checks an AI answer. Its temperature, gas flow and raw materials have to remain under control. Intelligence earns its place here by helping people make better steel, with less energy, waste and disruption.
That is the practical test at Nanjing Iron & Steel Co., Ltd. (NISCO). Working with Huawei and ecosystem partners, the Chinese steelmaker has deployed its [YuanYe Steel Large Model](https://www.huawei.com/en/news/2026/10/nisco-yuanye-steelmodel) across operations ranging from ironmaking and rolling to quality inspection and environmental control.
Together, these technologies supply computing hardware, cloud infrastructure and AI models. NISCO's YuanYe system combines them with the steelmaker's own production data, metallurgical expertise and operational models.
The distinction is important. Huawei is not simply installing a chatbot in a steelworks. The deployment balances probabilistic AI, which can recognise patterns and predict outcomes, with deterministic industrial models that calculate how machinery and metallurgical processes behave according to defined physical relationships.
A steelworks presents a different problem from an office chatbot. Predictions must account for equipment behaviour, material properties and the consequences of changing a production process. A plausible answer is insufficient.
The probabilistic models identify patterns in operational data, interpret complex observations and estimate what may happen next. The deterministic models calculate relationships between variables such as temperature, pressure, chemical composition and material behaviour.
A deterministic model produces a defined result for a given set of inputs and assumptions. Its accuracy still depends on the quality of those inputs and the validity of the physical model. It does not make the entire production system infallible.
The architecture also illustrates why industrial AI cannot be reduced to large language models. Some functions require probabilistic pattern recognition; others depend on deterministic computation, mathematical optimisation and established industrial control systems. A useful system must bring these capabilities together.
An industrial agent in this context is a software application that uses models, operational data and programmed workflows to perform a particular task. It might monitor equipment, identify a developing problem or recommend a change to production scheduling. These agents are not necessarily autonomous systems with authority to operate machinery without human supervision.
Inside a blast furnace, operators must judge conditions they cannot inspect directly. Sensor readings provide evidence, but understanding what those readings mean has traditionally depended heavily on accumulated experience.
YuanYe combines internal sensor data with metallurgical analysis to simulate the ironmaking process. NISCO says its furnace temperature prediction and health diagnosis models allow operators to anticipate temperatures and identify problems [two hours ahead, with accuracy exceeding 90%](https://www.huawei.com/en/news/2026/10/nisco-yuanye-steelmodel).
This is where the combination of probabilistic and deterministic intelligence becomes particularly useful. Machine-learning models can identify developing patterns in sensor readings, while physics-based models provide an independent representation of how the furnace should behave under particular operating conditions.
That gives the people running the furnace time to investigate and intervene before deteriorating conditions become a production problem. NISCO reports an increase of more than 10% in prime-grade molten iron yield, alongside lower energy consumption per unit of output.
Further along the production line, the task changes. Rolling steel plate requires precise control under rapidly changing conditions. YuanYe's rolling-force prediction model responds within milliseconds, according to the partners.
Quality inspection brings robots, imaging and AI together. NISCO's automated metallographic inspection system handles sample grinding and polishing, image acquisition and grading of the steel's microscopic structure.
Environmental control adds another set of tightly connected variables. NISCO's S-Link Intelligent Agent for Gas and Flue Gas Sulfur Control uses the YuanYe architecture to combine 602 process parameters with image and knowledge capabilities.
The system brings together statistical forecasting, industrial mechanism models and mathematical optimisation. It predicts sulfur risks an hour ahead, calculates gas-blending options within seconds and helps allocate emission quotas.
These are applications in which probabilistic prediction helps operators anticipate changing conditions, while deterministic calculations and optimisation methods support operational decisions. The value comes from combining them rather than treating every industrial problem as a task for a large language model.
Across its broader programme, NISCO reports that new-product research and development cycles have been halved and [annual costs reduced by more than CNY100 million](https://www.huawei.com/en/news/2026/10/nisco-yuanye-steelmodel), approximately $15 million.
“Digital and intelligent transformation is inherently human-centric. Its purpose is not to replace people, but to unleash human potential,” said [Li Jinyan](https://www.huawei.com/en/news/2026/10/nisco-yuanye-steelmodel), general manager of NISCO's Digital and Intelligent Integrated Development Department and director of its AI Research Institute.
Cheng Yongfu, senior industrial solution expert at Huawei's Steel & Non-ferrous BU, places the emphasis on extending AI across production, operations and management, where improvements can accumulate throughout the business.
For now, the consequential work is happening on the factory floor. Huawei's computing and AI technology provides part of the foundation; NISCO's production data, deterministic process models and operating expertise give it an industrial purpose.