Field Notes: Finding a Vacuum Leak Before the Batch Leaves the Furnace

A batch of precision stainless parts comes out of vacuum-quench brittle. Somewhere between the austenitizing soak and the nitrogen quench, something went wrong — but the furnace’s own controls don’t say what. The heat-treat log shows a green checkmark. The lab hardness test, run hours later, shows the problem. By then the batch has already moved.

That gap — between "the cycle finished" and "we know why it failed" — is the one we keep running into on vacuum heat-treat lines, and it’s what we built a reference monitor to close: catch the root cause within seconds of it happening, correlate it against lab hardness automatically, and prove it against a real thermal cycle, not a slide. Two failure modes matter here, and they sit at opposite ends of the process: a vacuum leak during the soak, and an under-pressured nitrogen quench. Same symptom on the bench — brittleness — two completely different causes on the furnace.

The two ways a good batch goes bad

A vacuum drift during the austenitizing soak (typically somewhere in the 1000–1050°C range for martensitic stainless) lets oxygen and moisture reach the surface, depleting chromium at the grain boundaries — a soft, decarburized skin that chips at an edge. A slow nitrogen quench does the opposite kind of damage: linger too long in the 800–500°C band and chromium carbides precipitate along the grain boundaries instead, an intergranular fracture path that has nothing to do with surface condition at all. Both end up brittle. Neither shows up on a bulk hardness spec if you’re not watching the right window at the right moment.

So the monitor doesn’t wait for the batch to finish and then ask "was that in spec." It watches two live conditions continuously: chamber pressure during SOAK against a 0.05 mbar ceiling, and the transit time crossing 800→500°C during QUENCH against a 130-second ceiling — plus a predictive check that extrapolates the current cooling slope so a slow quench gets flagged before the band even finishes crossing. On a simulated leak, that caught it in 15.2 seconds. On a simulated under-pressured quench, the predictive check fired in 32.3 seconds — before the crossing was even complete.

flowchart LR
    PLC["Furnace PLC<br/>OPC UA server"] -- "DataChange<br/>subscription" --> RE["Rule engine"]

    RE -- "SOAK phase:<br/>vacuum > 0.05 mbar<br/>sustained 30s" --> A1["Alert: vacuum drift<br/>~15s"]
    RE -- "QUENCH phase:<br/>800→500°C > 130s,<br/>or trending slow" --> A2["Alert: slow quench<br/>~32s (predictive)"]
    RE -- "batch reaches<br/>COMPLETE" --> CERT["Thermal Certificate<br/>of Conformity"]

    LAB["Lab hardness result"] -. correlated into .-> CERT
    A1 -.-> CERT
    A2 -.-> CERT

Reading a PLC’s mind in real time, not after the fact

We build these against OPC UA rather than the vendor-proprietary protocol older PLC generations speak, and that choice matters more than it sounds like it should. A polling client asking "what’s the temperature now?" every second is placing a bet on how fast things change — and a fifteen-minute quench cycle can blow past an 800–500°C transient in under two minutes. OPC UA’s subscription model flips that: the PLC pushes a change the moment it happens, and the monitor’s rule engine reacts to it, not to a guessed interval. Catching a fault in seconds instead of at end-of-batch wasn’t a matter of polling faster. It came from not polling at all.

The bug that would have quietly lied to us

The first version subscribed to four separate tags — phase, temperature, vacuum, elapsed time — each written to the PLC in its own OPC UA write. That’s a torn read waiting to happen: a new phase value can arrive paired with the previous tick’s pressure reading, because nothing guarantees those four writes get delivered to a subscriber as one atomic unit. It surfaced exactly as you’d expect — a nominal batch, mid-soak, briefly reporting a vacuum breach that never actually happened, because the client caught a "SOAK" phase update next to a stale "still ramping down from atmospheric" pressure value from a half-second earlier. The fix was to stop treating four correlated tags as four independent facts: one JSON snapshot, written once per scan, subscribed to as a single unit. The false alarm didn’t come back.

The second one was subtler and arguably worse, because it wouldn’t have thrown an error — it would have silently overwritten a real result. The web dashboard reconnects across batches so it never needs restarting between cycles; a finished batch’s PLC process lingers a few ticks after reaching COMPLETE before the next one starts. Reconnect during that window and you read the same batch ID again, build a fresh monitor against an already-finished snapshot with no SOAK or QUENCH history in it, and it reports a clean PASS — replacing the FAIL certificate that was already sitting there, correctly, moments earlier. Neither bug would have shown up in a demo that only ran the happy path once. Both did, the moment the system ran continuously, which is the only way a furnace actually runs.

The throughline

None of this was exotic instrumentation. A hardness lab feeding back into a thermal certificate. A subscription instead of a poll loop. A single atomic tag instead of four independent ones. What made each piece worth getting right is that a heat-treat furnace doesn’t announce which of two very different failure modes just happened — it just goes quiet and hands you a part that’s wrong for a reason you have to go find, usually after the batch has already left the building.

That’s the seam we spend our time in at Simplico: precision manufacturing shop floors where a Mitutoyo caliper, a PLC, or a furnace controller is already generating the truth about a part, and the only thing missing is a system that listens to it in real time and puts it in front of SPC and ERP before a bad batch moves. If your shop is running IATF 16949, ISO 13485, or AS9100 audits on data that’s still getting hand-typed off an instrument or a controller screen, that’s exactly what our simpliFactory architecture review is built to find. Book one — it’s a 90-minute conversation, not a sales pitch.