From Tap-to-Tap Records to Predictive Maintenance: The Data Already in Your Historian

Most EAF steelmakers already collect the data they need for predictive maintenance. They just don't know it yet.

Every Level 2 system logs hundreds of parameters per heat — power profiles, electrode positions, hydraulic pressures, cooling water temperatures, gas flows, and dozens of discrete I/O states. This data accumulates in historians and SQL databases, rarely queried beyond basic reporting and regulatory compliance.

The Maintenance Paradox

Maintenance in a melt shop operates on two timescales. Planned maintenance follows fixed intervals: electrode columns every N heats, delta closures every M heats, roof and sidewall panels on a calendar schedule. Unplanned maintenance — the kind that stops production — happens when something fails between intervals.

The paradox is that fixed-interval maintenance is simultaneously too frequent (replacing components with remaining useful life) and not frequent enough (missing early degradation that falls between inspections). The operators know this. They develop intuitions about which furnace "sounds wrong" or which hydraulic system "feels sluggish." But these intuitions don't scale across shifts, and they can't be documented or transferred.

What the Data Shows

When we analyze historian data at partner plants, three patterns consistently emerge:

1. Electrode consumption signatures. Abnormal electrode wear shows up as subtle changes in the regulation pattern — faster slipping, wider current oscillations, asymmetric phase behavior — hours before the electrode becomes visibly short or breaks. A classifier trained on historical breakage events can flag at-risk electrodes with 80%+ precision at 6–8 hours lead time.

2. Hydraulic degradation. Roof swing and electrode clamp hydraulics degrade gradually, but the degradation is masked by the control system compensating. By tracking the gap between commanded and actual positions over time, we can identify cylinders and valves approaching failure well before they cause a production stop.

3. Cooling circuit fouling. Panel cooling water flow rates and delta-T measurements, logged every few seconds, contain information about internal fouling rates. Simple trend models predict when a circuit will need cleaning with enough lead time to schedule it during a planned outage.

Implementation Without Disruption

The key insight is that none of this requires new sensors, new PLCs, or new network infrastructure. The data is already being collected. What's missing is the analytical layer that transforms raw time series into actionable maintenance signals.

At PRAX, we deploy this layer as a software overlay on existing data infrastructure. The system reads from the historian, runs models on a standard compute node, and pushes alerts to the maintenance planning system via standard APIs. The entire deployment typically touches zero automation code and requires no furnace downtime.

The Bigger Picture

Predictive maintenance is often positioned as an end in itself. We see it as the entry point to a broader shift: from reactive operations to anticipatory ones. Once a plant has the infrastructure to predict equipment failures, the same infrastructure can predict quality deviations, energy anomalies, and process drift. The maintenance use case justifies the investment; the operational intelligence that follows is where the real value compounds.