In 2026 the standard most utilities lean on for transformer fault diagnosis is being rewritten. IEC 60599 stands at Edition 4.0 (2022, stability date 2026); Edition 5 is in IEC TC 10 under maintenance team MT 20, its working draft registered as a committee draft on 4 December 2025 and publication forecast for 2027. A University of Tehran comparison by Aliyari, Samimi and Sanaye-Pasand, added to IEEE Xplore on 24 March 2026, ran four interpretation techniques over 2,538 oil samples. The Duval triangle held the best accuracy and F1 score, the pentagon generalized worse, and the ratio method was more certain where it answered.
What each transformer fault diagnosis method needs as input
Three families dominate service practice, and each falls silent — not cautious, silent — when its input is absent.
| Method | Input it needs | Strongest use | Where it fails |
|---|---|---|---|
| Ratio schemes (Doernenburg, Rogers, IEC 60599) | 3–4 gas ratios, all above detection limits | IEC 60599:2022 fault types PD, D1, D2, T1, T2, T3 | Undefined outside the zone table; a pattern missing every zone returns nothing |
| Duval triangle / pentagon | CH4, C2H4, C2H2 for the triangle; five gases for the pentagon | Fast agreement on classic mineral-oil patterns | A lookup: causes sharing a zone are not separated |
| Machine-learning classifiers | Those features plus trend, load and a labelled training set | Ranking candidates from incomplete or noisy input | Accuracy belongs to the dataset, not the method |
Where does each method stop answering?
Ratio schemes: no output outside the table
The IEC ratio scheme defines only 11 of the 27 ratio-code combinations its three ratios can produce (IEC 60599:2022). Outside all of them the method has no opinion — and “no diagnosis” on a report reads as reassurance. It is not.
Doernenburg will not judge a sample until one gas exceeds a minimum concentration, and every ratio is undefined when its denominator drops below the detection limit: with acetylene under the floor, C2H2/C2H4 does not exist. IEC 60567:2023’s roughly ±15 % acceptance target is then wide enough to move a borderline ratio across a zone edge, so the measurement uncertainty budget belongs here. Our guide to DGA gas ratios covers these edges.
Duval triangle and pentagon: a lookup over a fixed gas set
The Duval triangle reads three gases and returns a zone. That fixed input is both its strength and its edge: distinct root causes can land in the same zone, and the plot names the zone, not the ranking inside it.
It is not the tool for every gas combination. Oils that stray-gas, tested under ASTM D7150, can plot into a discharge region and manufacture a fault that does not exist.
Machine-learning classifiers: a model is only its training set
Classifiers change the shape of the answer, not the input: they consume the same measurements plus load and history and return a ranked list. Reported accuracy swings widely because each study uses its own sample count and fault-class definition, so a published figure belongs to that dataset, never a specification.
A model trained only on confirmed faults has never seen a healthy oil that stray-gasses. Our note on machine learning in DGA covers what it adds.
The order that holds up in service
Sequence matters more than any single method. Confirm the sample first, with every gas above its detection limit. Compute the ratio scheme you report against, recording “out of scope” rather than “normal” when it returns nothing. Plot the Duval triangle as a cross-check: where the two disagree, that disagreement is the finding. Only then bring in a classifier.
What a method returns is evidence, not a verdict: it is bounded by the sample it came from. Opening a transformer stays an engineering judgement.
Frequently asked questions
Which transformer fault diagnosis method should I try first?
Start with the ratio scheme your specification names, usually IEC 60599, then plot the Duval triangle as a cross-check. If the ratios return nothing, record that as out of scope, not normal.
What if every gas ratio falls outside the zone tables?
Then that method has no opinion on this sample. Check each gas is above its detection limit, confirm sampling followed IEC 60567, and lean on hydrogen trend while you re-sample sooner.
PAS DGA for continuous hydrogen monitoring
Method choice matters less than input quality. A dissolved hydrogen channel adds a rate of change between samples and an early rise that precedes any classifiable ratio. The DGA-900 extends that to nine gases plus moisture, and the dissolved hydrogen sensor compliance guide shows how a continuous channel sits beside the laboratory schedule and the oil sampling procedure feeding it.
Tell us what your transformer fleet looks like and we will help place the measurement points.
Sources
- IEC 60599:2022, Ed. 4.0, IEC/TC 10. Edition 5: TC 10 MT 20, committee draft registered 4 December 2025, publication forecast 2027.
- F. Aliyari, M. H. Samimi, M. Sanaye-Pasand, “Comparative Accuracy of IEC Ratio Approaches and Duval Triangle/Pentagon in Transformer Fault Diagnosis from DGA Data,” ICEMD 2025, Tehran; IEEE Xplore 11441907. Research comparison, not a product claim.
- IEC 60567 (sampling and analysis); CIGRE TB 783; ASTM D7150 (stray gassing).