The purity alarm is a threshold, not a warning.

A plant alarm layer is rationalised around conditions an operator can act on, and for gas crossover the condition that can be acted on is usually defined close to the safety limit. SCADA is watching for that condition rather than for slow physical drift, so the response window an operator gets is narrow by design rather than by oversight.

That serves notification well and planning poorly. If the first meaningful signal comes close to the trip point, the plant has already lost most of its operating flexibility. None of this displaces the protection layers: existing alarms, trips, interlocks and OEM operating limits remain authoritative, and anything modelled on top of them is advisory.

The two product gases are separated by a diaphragm or a polymer membrane, and that barrier is not perfectly impermeable. A small quantity of hydrogen diffuses continuously into the oxygen side, which is ordinary behaviour rather than a fault.

What matters operationally is how the fraction moves with load. Hydrogen production scales with current. Permeation is driven largely by the concentration and pressure difference across the barrier, and while it does rise somewhat at higher current through supersaturation at the catalyst layer and the temperature that comes with load, it does not rise in proportion. Divide a weakly rising permeation rate by a linearly falling production rate and the fraction climbs as the plant turns down.

That places the tightest gas quality margin in the low-load region, which is precisely where a solar-coupled plant spends its mornings and evenings. It is also where the plant would most like to keep running, because those hours are the difference between capturing a resource profile and truncating it.

Yunify models membrane permeability and gas transport behaviour with first-principles relationships, then corrects those predictions with operating data. The system is not waiting for a binary alarm. It is tracking whether the plant is moving toward a physically unsafe direction.

In this case, the useful signal is not just a measured concentration value. It is the shape of the trajectory: whether H2-in-O2 is drifting in a way that matches membrane degradation and whether that drift is accelerating under load, temperature, or impurity conditions.

As baseline permeation increases with age, the load at which the hydrogen fraction reaches its safety margin moves upward. The plant does not fail. It quietly loses low-load hours, and that shows up in the production record before it shows up in any condition report.

Turndown loss is therefore worth tracking as a condition signal in its own right. A minimum stable load that has crept up over months, compared at a comparable temperature and pressure, describes barrier condition without waiting for the concentration itself to approach anything alarming.

The measured concentration needs the same treatment. Hydrogen in oxygen read at whatever operating point the plant happened to occupy is mostly a record of the operating point. Compared at a fixed load, temperature and differential pressure over weeks, it becomes a description of the barrier. Analyser health belongs in that picture too, since a drifting analyser and a drifting membrane produce the same trend on a chart.

The business value is straightforward. Where the drift is visible early enough, plant teams can rotate equipment, adjust loading and plan intervention windows instead of absorbing an unplanned stop.

That shifts the conversation from reactive alarm handling to asset strategy. In practice, the difference is emergency shutdown versus controlled action with lower production loss and less operator stress.

The lead time is a design objective rather than a guarantee. Deployments are scoped for a two to six week warning window, and what a given plant sees depends on its stack chemistry, load profile and the instrumentation already in place.

Warning time is only worth what can be done inside it. Days before a trip, the options are narrow: derate, shift the operating window upward, and prepare for an unplanned stop. Weeks before, the options widen to scheduling a stack rotation into a period the resource is poor anyway, ordering spares against a lead time rather than an emergency, and holding the plant at a load that protects margin while the replacement is arranged.

The instrumentation that supports it is modest. A hydrogen-in-oxygen analyser with a maintained calibration record, differential pressure and temperature logged at the same resolution, load history retained rather than averaged away, and cell voltage where the stack provides it. The common failure is not missing sensors but a historian configured to compress, since a trend reconstructed from five-minute averages cannot recover the operating points the normalisation needs.

The warning window is a design objective rather than a promise. Deployments are scoped for a two to six week target, and what a given plant sees depends on its stack chemistry, its load profile, its instrumentation and how much history it kept before anyone started looking.

A rising normalised crossover trend supports a membrane or separator hypothesis. It does not prove one. Differential pressure, temperature, load history, seal condition, a pinhole rather than general thinning, and the analyser itself can all move the same measurement, and several of them can move it at once late in life.

The order in which signals move carries information worth using. Crossover rising while cell voltage holds or falls slightly points towards the barrier, since a thinner membrane lowers ohmic resistance even as it leaks more. Cell voltage rising while crossover holds points elsewhere, usually towards catalyst or transport. Both rising together suggests more than one mechanism.

None of this is a repair. A bonded membrane electrode assembly is not serviceable at the membrane layer, so the practical question becomes how much operating margin remains and when replacement should be planned, which is a remaining useful life question with a commercial answer rather than a maintenance one.