In flash smelting, concentrate melts in two to three seconds and quality depends on precise slag control. JXCC built its furnace-side AIoT platform on MOI — sensors and images in one engine, vision models running as in-database UDFs — cutting furnace-bottom copper loss by 80%.
Founded in 1979 and headquartered in Nanchang, Jiangxi Copper Group is a Fortune Global 500 enterprise and one of China's largest copper producers, spanning mining, smelting, manufacturing and international trade. Its Guixi smelter is China's largest modern copper smelter and the country's first flash smelting plant, world-leading in scale and among the lowest-cost globally.
At the Guixi smelter the core process is flash smelting and the core equipment is the flash furnace. Deeply dried powdered concentrate, below 0.3% moisture, is mixed with air or oxygen at the burner and injected at 60–70 m/s from the top of the reaction shaft. Suspended in gas, the particles complete decomposition, oxidation and melting of the sulphides in two to three seconds. The molten mixture falls into the settler below, where matte and slag separate — and because the slag layer still carries a high copper content, reducing copper in slag is a problem every smelter works on.
Controlling the matte level is the effective lever over both copper-in-slag and slag-in-copper. Sensing the slag and matte layer heights accurately, and using those heights to schedule tapping of matte and slag from the right port at the right time, is what determines whether matte grade can be held while feeding the downstream converter properly.
The industry's usual method is a measuring rod, exploiting the relatively static surface at the settling zone and the different adhesion of matte and slag to the rod. Tapping is then scheduled by manual calculation from those readings. It works, but it is bounded by how often a person can measure and how well they can compute.
In data terms this is a textbook AIoT problem that has to solve OLTP, OLAP, time series and AI at once. Business interactions must respond in under a second; sensors and cameras supply level and temperature data while ETL brings in converter-side and historical tapping records; time-series ingest peaks at 250,000 rows per second; and analytical models over all of it must produce furnace condition assessment and tapping recommendations, with the camera data requiring AI vision on top.
MatrixOne's converged architecture let the project stand up a single, efficient data foundation rather than assembling one.
As a converged database it supports OLTP, OLAP and time series properly, covering the real-time, consistency and stability requirements of the workflow system, the high-frequency time-series writes, and the real-time analytics behind the indicators — in one place. Because it is built on a single storage engine and modelling approach, there are no ETL chains to operate between components.
Its cloud-native separation of storage and compute puts the storage layer on shared NFS/S3 protocols, so unstructured data is stored and managed alongside everything else. The camera data in this project is carried directly by the database storage layer.
User-defined functions let AI models implemented in code be packaged as UDFs inside the database, which removes the separate management and operations burden that a standalone AI algorithm layer would otherwise carry.
On this data architecture the project's implementation cycle came in around 60% shorter than planned, with the data architecture itself built in about a week. End to end — from acquisition through computation to tapping control strategy — response stays within five seconds.
On the production side, control over copper-slag output improved to the point that furnace-bottom copper fell by 80%, matte level precision is held, and tapping-floor operations are substantially automated.
For Jiangxi Copper this was a genuinely new approach: large-scale IoT acquisition with analytics and AI replacing manual operation, delivered without sinking into a swamp of IT complexity.
The same approach is now being rolled out across other stages of the group's smelting operations, on the same AIoT data foundation.