AI-driven predictive maintenance for a steel mill

In a steel mill, a single hour of unplanned downtime on the rolling line can cost hundreds of millions of dong, not to mention the safety risks of emergency repairs. A domestic steel producer wanted to move away from fixed-interval maintenance and run-to-failure repairs towards maintenance based on the actual condition of each asset. Zentrova is delivering the programme in two phases; phase one covers the 620 most critical assets in the rolling and raw materials areas.

The challenge

Equipment runs around the clock amid extreme heat, dust and heavy vibration. Fixed maintenance schedules meant many healthy components were replaced too early, while bearing and gearbox failures still struck without warning. Diagnostic know-how rested with a handful of senior engineers who were approaching retirement, and the company had no structured way to capture what they knew. Every month without change meant more lost production and more overtime spent on emergency repairs.

Our solution

• Industrial sensor network measuring vibration, temperature and motor current on motors, mills and conveyors, built to survive harsh conditions.
• Time-series data platform capturing billions of data points a day, with pre-processing at the edge to reduce network load.
• Machine learning models in Python that detect anomalies and estimate the remaining useful life of each asset, enriched with historical repair records.
• Expert knowledge captured digitally: diagnostic rules from senior engineers are encoded in the system and used to validate model output.
• Alerts built into the maintenance workflow, automatically raising work orders with the predicted cause and the parts to prepare.
• Wave-based validation: each equipment group runs in parallel with the old process for three months before it goes live.

Results

Rolling line 2, the first pilot area, has delivered measurable results after six months of operation:

• Unplanned downtime reduced by 41%.
• Failures flagged on average nine days in advance.
• Spare parts spending down 17% by avoiding premature replacement.
• Alert precision of 87%, with false alarms below 8%.

Phase two will extend coverage to the reheating furnaces and the compressed air system. Over the same period, responsibility for retraining and monitoring the models will move step by step to the client's own data team, supported by playbooks and joint review sessions with Zentrova engineers.

Our maintenance crews no longer chase breakdowns. They plan replacements for the scheduled stops instead. — Head of Mechanical and Electrical Engineering, client plant