Real-time Analytics
China Iron and Steel Research needed a practical way to apply large-model and visual AI capabilities across metallurgical research and manufacturing, especially for metallography analysis, surface defect detection, and long-process material tracking that older methods struggled to deliver cost-effectively.
Reported impact
China Iron and Steel Research established a metallurgy large-model platform that delivered measurable breakthroughs on core industrial scenarios while demonstrating a workable domestic AI stack for central-enterprise manufacturing innovation.
Baidu AI Cloud and China Iron and Steel Research built a metallurgy-focused large-model platform around an IPS three-layer architecture and a private deployment product stack. The design combines Qianfan, Baige, Yijian, Zhenzhi, Kunlun, and related private-cloud capabilities so model building and application delivery can move together from platform to business scenario. Instead of treating the large model as a generic assistant, the joint team centered the implementation on a 'process-aware' metallurgy model that could validate real industry workflows. In practice, this meant using platform capabilities to stand up and verify models for metallography analysis, hot-rolled surface defect detection, and large-scene cross-domain material tracking. The stack also optimized Kunlun and the major Baidu platforms for fast delivery after contract signing. By tying model training, visual analysis, report generation, and process tracking to concrete metallurgy scenarios, the project moved from abstract AI experimentation into an operational industrial analytics system.
Metallography images, steel-surface defect images, and material-tracking scene data feed into the private deployment stack, Qianfan and Baige support model training and validation, and the finished workflows generate reports, defect analysis, and cross-domain tracking outputs for metallurgy operations.
China Iron and Steel Research wanted to build a metallurgy-industry large-model platform and empower the industry with a 'process-aware large model'. (original: 中国钢研与百度等企业深化合作，建设基于人工智能技术的冶金行业大模型平台，面向行业提供“流程感知大模型”，赋能冶金行业。)
For material tracking, the system reached 100 ms response time, MOTA of at least 95%, and MOTP of at least 90%. (original: 模型响应速度100ms，MOTA(跟踪准确度)≥95%，MOTP(跟踪精度)≥90%。)
In metallography analysis, the model reached MIOU of at least 90% and PA of at least 96%, with one ferrite-plus-austenite model recording MIOU 91.44% and accuracy 97.30%. (original: 大模型性能指标MIOU(平均交并比)≥90%，PA(像素准确率)≥96%。完成铁素体+少量奥氏体数据的模型训练，模型性能指标： MIOU：91.44%；准确率:97.30%。)
In surface defect detection, the model achieved precision of at least 95%, recall of at least 94%, and a processing speed of 2000 images in 100 seconds. (original: 模型精确率≥95%，召回率≥94%；处理速度达到100秒2000张图片。)
The implementation realized the full landing of a private deployment product system under an IPS three-layer architecture and fused Qianfan, Baige, Yijian, Zhenzhi, and related platform capabilities into applications. (original: 通过IPS三层架构的整体规划，实现了万源私有化产品体系的全面落地。从业务场景出发，依据千帆、百舸、一见、甄知等平台的核心能力，实现从产品到应用的深度融合。)
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Baidu AI Cloud Sample May 6, 2025 Weiyi Manufacturing Weiyi Manufacturing needed to reduce reliance on exhausting manual visual inspection in 3C factories because workers had to inspect thousands of parts from many angles every day, which caused fatigue, unstable quality,… Miss rate fell below 0.1%. BCC · Baidu Algorithm Engine · Intelligent Industrial Quality Inspection Solution 
NVIDIA Sample Aug 25, 2026 Siemens Siemens needed to help industrial teams overcome fragmented engineering tools, labor shortages, and slow design and maintenance workflows so they could design, simulate, and operate complex products and factories faster… 10%–15% reductions in capital expenditure NVIDIA CUDA · NVIDIA DGX B200 · NVIDIA DGX H200 
In-house Sample Aug 13, 2026 Lockheed Martin Static flight-hour tables and usage metrics ignored real-time consumption, causing chronic under-resourcing, limited stock effectiveness, and a reactive sustainment supply chain that could not keep pace with modern… predict component failure up to 72 hours before it occurs LogC2 dashboard · deep-learning models · digital data fabric 