Predictive maintenance for electrolysers, and why the usual playbook does not transfer

The predictive maintenance discipline was built on rotating machinery. Vibration spectra reveal bearing defects, imbalance and misalignment. Thermography finds electrical and mechanical hot spots. Oil analysis reveals wear metals and contamination. All three depend on a mechanism that rotates, rubs or is lubricated.

An electrolyser stack is a series assembly of electrochemical cells with no rotating parts, no lubricant and very little to vibrate. Applying a conventional predictive maintenance platform to it produces a system watching the pumps and blind to the asset that represents most of the capital and nearly all of the replacement cost.

This is why searching for electrolyser predictive maintenance tends to return either generic condition-monitoring material or vendors selling measurement hardware. The generic material is about a different physical problem.

Electrical. Cell voltage at a given current density and temperature is the primary condition indicator for a stack, and the distribution across cells carries far more information than the mean. Recording that distribution is what Yunify CVM exists to do. Uniform drift across the population is consistent with expected ageing. A small number of cells departing from their neighbours under identical conditions points at something local. Because a stack contains hundreds of cells in series, a single cell drifting substantially barely moves the aggregate reading, which is why stack voltage alone is a weak instrument for this purpose.

Gas side. Hydrogen content in the oxygen stream is both a safety measurement and a condition measurement. Its trend at a fixed load, over weeks, describes separator or membrane condition, and on a PEM stack it is often an early sign of membrane thinning, though what arrives first depends on pressure, temperature, load, analyser placement and the failure mode. The instantaneous value belongs to the safety system; the trend belongs to the maintenance programme, and conflating the two is a common mistake in both directions.

Hydraulic and thermal. Differential pressure across the stack, flow distribution, and temperature spread between inlet and outlet and between cells describe whether the assembly is being fed evenly. Maldistribution shows up as a thermal pattern before it shows up as a performance loss.

Chemical. Electrolyte conductivity and concentration in alkaline systems, and feedwater quality in all of them. Contamination is one of the few mechanisms that can move quickly, and it is usually visible upstream before it is visible in cell behaviour.

Some mechanisms build slowly and are genuinely predictable, which is the territory physics-informed prognostics covers. Gradual separator or membrane degradation, catalyst ageing and slow contamination effects develop over weeks to months and leave a trend behind them. These are where advance warning is realistic and where the maintenance decision has time to be planned.

Others develop over hours or days: an electrolyte excursion, a rapid contamination event, a developing seal leak. These are detectable but the useful response is operational rather than a maintenance plan.

A third group is effectively instantaneous. A gasket that fails, a power electronics fault, a sudden loss of cooling. These belong to protection and interlock systems, and claiming to predict them is one of the reliable markers of a vendor overselling.

Being explicit about which group a given failure mode belongs to is the most useful thing to settle at the start, because it sets what advance warning is available to be bought.

Circulation pumps, compressors, blowers, chillers, dryers and valves are conventional rotating and actuated equipment, and the standard techniques work on them unchanged. Vibration monitoring on the larger machines, current signature analysis on motors, thermography on electrical connections, and performance trending on heat exchangers and dryers.

This half is worth taking as seriously as the stack. Published reliability work on electrolyser plants consistently finds that a substantial share of downtime originates outside the stack, in equipment that is individually well understood. The stack is where the replacement cost sits; the balance of plant is frequently where the availability goes.

The two halves also interact. A circulation pump degrading changes flow distribution, which changes temperature spread, which changes cell behaviour. Treating them as separate monitoring systems means that chain gets diagnosed as a stack problem, which is the difference between diagnostics and alarming.

Start from failure modes that have actually cost this plant something, not from a list of available techniques. For each one, establish how it currently gets discovered, how much notice exists today, and what the consequence costs. Those three numbers are what any later claim of improvement is measured against.

Establish a baseline while the plant is healthy. Condition monitoring compares against a reference, and a reference captured after something started drifting is worth very little. A polarisation curve at defined conditions, a cell voltage distribution and a thermal profile at commissioning are all cheap at the time and impossible to recover later.

Decide what an alert is for. An alert that names an asset, a probable cause and an action gets used. A threshold crossing on a chart gets acknowledged and forgotten, and once operators learn that most alerts are noise, the programme is finished regardless of how good the underlying detection is. That failure mode is covered in more depth on the false positives page.

Yunify is built around this shape: cell-level measurement where the stack signals live, multi-physics models to attribute a change to a mechanism rather than to a suspicion, and an output written as an instruction. The general point stands independently of the tool, which is that the electrochemical half of an electrolyser maintenance programme needs different instruments and different reasoning from the mechanical half.