What DGA Doesn’t Tell You

Every asset owner, every reliability engineer, and by now most claims professionals can read a Dissolved Gas Analysis report. Acetylene means arcing. Ethylene means heat. Hydrogen means partial discharge, or a hundred other things. Run the numbers through a Duval Triangle, compare them against IEEE C57.104-2019 or IEC 60599, and the transformer tells you what it has been doing.

That is the value of DGA, and it is real. As a screening tool for a fleet in service, nothing else comes close for the money. It is cheap, it is repeatable, and it catches developing faults months or years before anything else will.

But screening a healthy fleet and investigating a destroyed unit are two entirely different exercises, and DGA is optimised for the first one.

When a transformer fails — when there is a claim, a coverage question, a subrogation target, an outage report due to a regulator, or a plant that cannot come back online — the questions change. Nobody is asking whether there is a fault anymore. The fault is obvious; it is on the ground in front of you. Four questions replace it:

DGA cannot, on its own, answer any of those four.

We have opened units whose gas data described a condition that was entirely real, but was not the condition that destroyed them. We have opened units whose last sample was unremarkable. Neither outcome makes DGA a bad tool. It makes it a tool with a defined scope, being asked to carry a case.

A dissolved gas analysis is a description of a symptom. A forensic investigation is an account of a cause. Confusing the two is the single most common error in transformer failure files.

DGA Tells You Something Is Wrong. It Doesn’t Tell You Where.

Gas dissolves in oil, and oil circulates. By the time a sample reaches the lab, the fault gases generated at a specific point inside the tank have been diluted, convected and mixed through the oil volume — a process that takes hours to days, and that leaves you with something approximating a tank average rather than a position.

A gas profile characterises a fault by energy and temperature. Which winding, which phase, which end, inside or outside the coil — none of that is in the number. Two failures with entirely different origins, and entirely different parties behind them, can arrive at the laboratory looking much the same.

In a screening context this hardly matters — you flag the unit and go and look. In a failure investigation it matters enormously, because location is very often the whole question. Something that was built into the unit and something introduced to it years later can present as the same reading, and they are not remotely the same outcome for the parties involved.

It Classifies Faults. It Does Not Find Them.

Ask a Duval Triangle whether your transformer has a problem, and it has no answer for you. That is not the question it was built for.

Duval’s first triangle has no zone for “healthy”. Every point that can be plotted on it lands in a fault category — partial discharge, thermal, discharge, or a combination of them. Plot the ratios from a sound unit and a fault type comes back, because a fault type is the only output the geometry has. Later triangles added zones for stray gassing, which helps at the margins, but does not change the shape of the problem.

Some of the older ratio schemes were more candid about it. Doernenburg will not return a diagnosis at all until the gas concentrations clear defined thresholds; below them it declines to answer, and “no diagnosis” is a common outcome. That gate is a tacit admission — the interpretation only means anything once a fault is already known to be there.

Which leaves the question that actually matters, is there a fault here at all, sitting outside the interpretation methods entirely. It gets settled by concentration levels, by rate of change, and by judgement about that specific unit. Separate from the chart, harder than the chart, and not made by the chart.

There is a quieter version of the same problem. When a result lands where a scheme has no answer for it, it usually gets written down as normal. But “the transformer is fine” and “this method has nothing to say about this result” are two very different findings. In most files they look identical.

A Gas Profile Is A Sum, Not A Sequence

This is the limitation that does the most quiet damage to investigations.

Dissolved gases accumulate. The oil holds a running total of everything that has happened to it since the last time it was processed — it does not timestamp anything. There is no chronology in the numbers.

That has a specific and serious consequence after a catastrophic event: the high-energy failure generates so much gas that it buries whatever was there before it. The low-energy signature of the initiating condition — the slow thermal fault, the developing partial discharge, the mechanical weakness that had been there for two years — is drowned in acetylene from the final flashover. Post-failure DGA on a badly failed unit very often reports, in effect, “this transformer failed catastrophically.” Which everyone already knew.

The distortion runs in the other direction too. If the pressure relief operated, if oil was ejected, if the tank ruptured, a significant fraction of the gas record left the building before anyone drew a sample. Post-failure numbers can overwhelm the history or lose it — and the result looks equally authoritative either way.

There is a second-order version of this worth naming. DGA is usually described as an early warning, and for slowly developing conditions it earns that description. But it remains a record of what has already occurred — the gas exists because something has already happened to produce it. Where a condition develops over years, the warning arrives with room to act. Where it develops over days, it arrives with the damage.

Cause and consequence look identical in a gas sample. Separating them is not a laboratory task.

The obvious objection at this point is to sample more often. It is a fair objection, and it is usually where the conversation stops — because the next question does not have a general answer.

How often? Annually is common fleet practice and is no help at all against something that develops between visits. Monthly costs more than most fleets will fund, and still would not have caught a condition that ran its course in a fortnight. The right interval is not a number. It depends on the unit, its duty, what it is protecting, what has already changed about it, and what the consequence of losing it would be. Anyone who answers that question with a single figure has not understood it.

The Failure May Have Occurred Somewhere The Oil Never Went

Main-tank DGA sees the main tank. There are important parts of a transformer it does not see at all, or sees only indirectly and late.

On-load tap changers on many designs live in their own compartment on their own oil system — a serious OLTC event can be well advanced while main-tank DGA reads clean. Condenser-type bushings are generally their own sealed system, with their own failure physics and their own diagnostics. Mechanical winding displacement from a through-fault may generate very little gas while leaving the unit one event away from destruction. Core and frame grounding problems, cooling system failures, protection that did not operate, and moisture distribution in the solid insulation all sit largely outside the gas picture.

The reverse is also true, and is a trap in its own right: a degraded barrier between an OLTC compartment and the main tank puts tap changer gas into the main tank sample, where it reads as a fault in the active part. Gas that appears in the oil is not automatically evidence of anything that happened in the oil.

“Normal” Is A Statistical Statement, Not A Clean Bill Of Health

The current generation of DGA guidance is explicitly population-based. Screening limits describe where a unit sits relative to a large fleet — they are percentiles, not physics. A result inside the normal band means the transformer is unremarkable compared to its peers. It does not mean nothing is developing.

Several ordinary situations degrade that further. Oil that has been processed or degassed has had its history erased, and the trend that would have carried the diagnostic information starts again from zero. Some fluids and some in-service conditions produce stray gassing at moderate temperatures — hydrogen, methane, ethane — that mimics a low-energy fault while representing no damage at all. Ester fluids do not behave like mineral oil and cannot be read against mineral-oil criteria without generating both false alarms and false comfort; they have their own guidance for a reason. And rate of change is at least as meaningful as any absolute number — which requires a clean baseline history that, in the units that fail, frequently does not exist.

Context compounds it. The same numbers on two different units can mean opposite things. A new, heavily loaded unit in a wind farm has legitimate reasons for a hydrogen figure that would be alarming on an older, lightly loaded transmission transformer. Screening criteria do not know which one they are looking at. Nor do they know that the oil was changed two years ago — which resets the record, and can leave an unremarkable profile sitting on top of damage that never went anywhere.

“The DGA was normal at the last service” is one of the most common sentences in a transformer claim file. It is much weaker evidence than it sounds, in both directions.

Transformers Rarely Fail For One Reason

Failures propagate. One condition creates the circumstances for the next, and by the time a unit comes apart there may be three or four things wrong with it that did not arise independently.

Nearly every interpretation scheme assumes a single fault. When more than one is present, the gas signature is a blend — and blends do not resolve cleanly into a category. The result still looks like an answer. It is an average of several.

Which of the several was the origin, and which are consequences of it, is the question the file actually turns on. It is not a question a gas profile can be made to answer, no matter how carefully it is read.

The Sample Is Evidence, And It Can Be Compromised

In an investigation, a DGA result is not just data — it is an exhibit, and it inherits every weakness of how it was obtained.

Sampling technique, container type, headspace, transport, temperature and time to analysis all move the numbers. Extraction method varies between laboratories. Reproducibility between labs is wider than most people assume when they compare two reports side by side — industry guidance has acknowledged comparative testing in which essentially identical samples returned key gas concentrations differing by a factor of three or more.

And after a failure, the physical circumstances almost guarantee irregularity: the unit is de-energised, possibly drained, possibly moved, sometimes already partially disassembled by well-meaning responders before anyone with a forensic mandate arrives.

A gas number with no defensible chain of custody behind it is a number that opposing counsel will be delighted to see in your report.

A Fault Type Is Not A Liable Party

This is the point that matters most to insurers, and it is the one DGA structurally cannot reach.

Suppose the interpretation is airtight and everyone agrees the unit suffered a high-energy discharge. That finding is compatible with a defect built in at the factory. It is compatible with something done, or not done, on site. It is compatible with an event that arrived from the network, with a protection system that failed to do its job, and with a condition that had been developing quietly for years. It is also compatible with several of those at once — which, in the units we open, is more often what actually happened.

Each of those points at a different party. The gas profile separates none of them, and no amount of re-reading the Duval Triangle will make it do so. Neither will working down a list of usual suspects — because the cases that end up in dispute are precisely the ones where the obvious candidates have already been eliminated, and the answer sits somewhere less convenient.

Attribution comes from elsewhere, and never from one place. It is assembled — out of what survived the failure, out of the record the unit carries with it, and out of what the damage itself says about the order in which things happened — into a single account that has to hold all of the evidence, including the parts that do not fit.

That is a discipline. It is not a test you order.

It is also not deterministic. Two competent engineers can work the same failure and reach different conclusions — not because one of them is careless, but because they weighted the evidence differently. That is precisely why the reasoning has to be visible in the report and not just the finding. A conclusion nobody can follow is a conclusion nobody can defend.

The Lightning Strike That Wasn’t

A file reached us with the cause apparently already settled. The unit had been through a lightning event, the gas data was consistent with that account, and the explanation had hardened before anyone opened anything.

It did not survive the teardown. What the physical evidence described was a defect that had been in the unit since it was installed, compounded by damage to the supply cables that pre-dated the storm entirely. The apportionment moved — between the manufacturer and parties who had worked on the site — and the matter resolved before litigation costs were incurred.

The DGA had not been wrong. It had been answering a narrower question than anyone was asking of it. That distinction is worth a great deal of money and is almost never drawn in the first week.

Choose The Expertise Before You Need It

The practical risk in transformer failures is not that DGA is unreliable. It is that DGA is available — fast, familiar, and already in the file — while the evidence that would actually establish cause is perishable and disappears in the first week.

Sites get cleared. Units get scrapped or shipped. Components get separated from the unit they came from. Oil gets drained without sampling protocol. The window in which a failure can still be reconstructed is short, and it closes before most files are formally opened.

What can usefully be decided in advance is not a procedure. It is who gets called, and how early. Scope follows the questions the case will eventually have to answer — and those questions are rarely obvious in the first week. A coverage dispute, a subrogation target and a regulator’s outage report pull an investigation in three different directions, and evidence preserved for one is routinely destroyed by the work done for another. Deciding which of those the site visit is actually serving, before anyone touches the unit, is most of the job.

The rest is competence and record. High-voltage plant, tap changers, bushings, protection systems, battery storage and mission-critical facility topologies are genuinely different problems, and the differences are not superficial. And whatever is concluded has to survive in the forum where it lands — which means a report that is objective, transparent about its own method and limits, and explicit about where fact ends and expert opinion begins, in the form contemplated by ASTM E620.

DGA is where a transformer investigation starts. Treating it as where the investigation ends is what turns a recoverable loss into an unexplained one.