Electrolyzer diagnostics

An alarm tells the operator that an abnormal condition needs a timely response, whether that condition is a measured value, a calculated one or a rate of change. Diagnostics should do something different: assess whether the plant is entering a risky state, narrow down which mechanism is most likely to be driving it, and indicate how urgent the response really is.

In electrolyzers, that distinction matters because many consequential events begin as slow drifts in permeability, thermal response, voltage behaviour, impurity sensitivity, or balance-of-plant performance. That is the ground electrolyser predictive maintenance has to cover without the rotating-machinery signals other industries lean on.

Condition monitoring in most industries rests on rotating or lubricated machinery, and vibration and oil analysis have no subject on an electrolyser stack. What the stack offers instead falls into four families, and electrolyzer diagnostics is largely the work of reading them together.

Electrical signals carry the most information per measurement. Cell voltage under a known current density and temperature is the closest thing to a direct observation of stack condition, and its distribution across the population separates uniform ageing from something local. The total across the stack is the sum of those cells, so it moves by a fraction of a per cent when one of them drifts, which is why the distribution and the total answer different questions.

Gas-side signals describe the barrier. Hydrogen in the oxygen stream, compared at a comparable operating point over weeks, tracks separator or membrane condition, and its instantaneous value belongs to the safety system while its trend belongs to the maintenance programme. Hydraulic signals cover differential pressure, flow distribution and circulation. Chemical signals cover feedwater quality and, on alkaline systems, electrolyte conductivity and concentration. Thermography still applies, to the electrical connections and busbars rather than to the stack itself.

Electrolyzer signals are coupled by electrochemistry, transport, temperature, and pressure. If diagnostics ignore that structure, they can mistake noise for failure or miss the early stages of a real mechanism because the raw tags still look acceptable in isolation.

Physics context helps the model ask a better question: is the current plant behaviour consistent with what the system should be doing under this load, temperature, and process condition? That is a stronger basis for early warning than pattern matching alone, and it is what keeps the alert count survivable, which is the subject of reducing false positives in industrial anomaly detection.

A cell voltage reading on its own says very little. The same stack in the same condition reads differently at a different current density, a different temperature or a different pressure, and a renewable-coupled plant changes all three through the day. Comparing two raw readings taken a month apart mostly measures the difference between two operating points.

Normalisation is the step that turns a measurement into evidence: comparing like with like by correcting to a reference condition, which needs a reference to correct against. A polarisation curve captured under controlled conditions at commissioning, and repeated periodically, is what makes that correction defensible rather than assumed. Without one, the correction is made against a supplier curve for a nominally similar stack, and the uncertainty in that substitution can be the same size as the change being measured.

This is the quiet reason diagnostic programmes underperform. The instrumentation is often adequate and the reference was never captured, which is a decision taken at commissioning by people who were not thinking about a conversation three years later.

Better diagnostics change how teams plan interventions. Operators can derate, shift operating windows, inspect the right subsystem first, and avoid turning every anomaly into an emergency response.

That is why earlier diagnostics matter commercially as well as technically. They support production planning, reduce unnecessary shutdowns, and make maintenance planning more targeted. Where the symptom is energy rather than availability, the same reasoning applies to why a plant reads more kWh per kg than its datasheet.

Diagnostics identifies what is changing now and what is most likely driving it. Prognostics estimates how much operating margin remains at the current duty. Predictive maintenance is the decision that follows from both, weighing consequence, spares, outage cost and the remaining term of the offtake.

The three are frequently collapsed into one word, and the collapse hides where the weakness is. A maintenance plan built on a weak diagnostic layer produces dates nobody trusts, because the underlying question of what is actually wrong was never answered. Ranking mechanisms rather than asserting one is the honest form: several faults, a sensor error or an omitted variable can produce similar symptoms, and a ranked list with the supporting evidence is more use to an engineer than a single confident label.

Some mechanisms sit outside this entirely. A gasket that fails, a power electronics fault or a sudden loss of cooling develops faster than any maintenance decision can respond to, and those belong to protection and interlock systems. Precursors sometimes exist and are worth looking for, but claiming to predict the event itself is a reliable marker of a vendor overselling. The gaps this leaves between what the control system reports and what an operating decision needs are enumerated in what a control system does not show on an electrolyser plant.