BRENNEN

Industries / Energy, water and waste / Electric transmission and distribution / Transformers and switchgear

Monitoring in service

Monitoring is the work of watching transformers once they are in service, mostly by measuring the gases that faults leave dissolved in the insulating oil, so that a fault is found before it becomes a failure.

5

Milestones achieved, from 1921 to 2015.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1921

    A relay that catches gas

    A fault inside an oil-filled transformer, such as sparking, breaks down the oil into bubbles of gas. Max Buchholz of Kassel, Germany, patented a device that collects the rising bubbles under a hood so that they lift a flap and close a contact, which sounds an alarm or trips the switch that cuts the transformer off before the damage spreads. He filed in Germany on 26 September 1921, and his US patent was granted on 13 September 1927. Makers such as the German firm EMB still build Buchholz relays for oil-filled transformers, now some with added sensors.

    image-ppubs.uspto.govweidmann-electrical.com

    Credit

    Max Buchholz

  2. 1977

    Gas sensors on the tank

    A fault inside an oil-filled transformer breaks down oil and paper, and the gases it makes dissolve in the oil. Testing for them, called dissolved gas analysis (DGA), began in the late 1960s on oil samples sent to a laboratory, usually once a year, so a fault could grow for up to 12 months unseen, by GE's account. Guy Bélanger and Michel Duval of Hydro-Québec described a monitor that stays on the transformer: hydrogen passes from the oil through a membrane into a fuel-cell-type cell, which gives a continuous reading. They tested it on a 20 MVA transformer (MVA, megavolt-amperes, is the unit of a transformer's power rating), and by 1989 Duval described two commercial hydrogen detectors, one of them read remotely from the substation.

    doi.orgdoi.orggrouper.ieee.orggevernova.com

    Credit

    Guy Bélanger and Michel Duval, Hydro-Québec research institute (IREQ)

  3. 1978

    Shared rules for reading gases

    In 1978 the IEEE issued its first guide to the gases found in oil-immersed transformers and what they mean for whether a unit can stay in service (C57.104), and the IEC issued Publication 599 on interpreting the analysis. Both turn gas levels and ratios between gases into a likely type of fault. Engineers still call their successors "the gas guides."

    doi.orgdoi.orggrouper.ieee.org

    Credit

    IEEE Transformers Committee; IEC Technical Committee 10

  4. 2001

    Inspected faults made public

    The IEC revised its guide as IEC 60599 in 1999, with new limits for the gas ratios and clearer definitions of normal and alarm levels. The revision rested on databases of DGA results from faulty equipment that had been opened and inspected in service, and in 2001 Michel Duval and A. dePabla published those databases so that maintenance staff could check their own readings against them.

    doi.orgdoi.org

    Credit

    Michel Duval (Hydro-Québec) and A. dePabla; IEC Technical Committee 10, Working Group 13

  5. 2015

    A world count of failures

    CIGRE, the international council of grid engineers, gathered 964 major failures from 1996 to 2010 at 58 utilities in 21 countries, covering 167,459 transformer-years (one transformer in service for one year). A major failure meant more than seven days out of service. Substation transformers failed at about 0.53% a year, and generator step-up transformers, which raise a power plant's output to grid voltage, at about 0.95%. Windings caused the most major failures, bushing failures had the worst effects outside the tank, including fire and explosion, and the data showed no clear rise in failure rate with age.

    Lessonage alone is a weak guide to which transformer will fail next.

    e-cigre.org

    Credit

    CIGRE Working Group A2.37

Watch for

  1. Language-model agents diagnose faults

    A team at Xi'an Jiaotong University built a system in which three AI agents, running on a large language model and guided by a knowledge graph of past fault cases, diagnose transformer faults. By the authors' own test it was right on 86% of 100 fault cases, which they put at 33% better than a BERT language model.

    Achieved whenanother group tests it on a utility's cases whose faults were confirmed by inspection, and publishes the score.

    zkxb.xjtu.edu.cn

    Claimed by

    Lin Jinshan and co-authors, Journal of Xi'an Jiaotong University, October 2025

  2. Machine learning misses early faults

    On 1,097 DGA records labeled by experts, supplied by the Austrian company VUM, nine machine-learning models were compared. They caught severe faults reliably, but 83.3% of moderate cases were read as normal, and the best model, XGBoost, scored a macro F1 of 0.757 (a 0-to-1 score that weighs each class equally).

    Achieved whena different team finds the same gap on another utility's data, or shows a method that closes it.

    tugraz.at

    Claimed by

    Tochukwu Udeh, Alexander Pirker and Markus Prossegger (Carinthia University of Applied Sciences and VUM), February 2026

Open

  1. 01

    An open DGA test set

    A public collection of DGA results from transformers that were later opened and inspected, each with the fault that was found, and a scored comparison on it of the IEC 60599 ratios, the Duval triangle (a chart that reads the fault type from three gases) and machine-learning models, reported separately for early and severe faults.

    The brief
    Credit · unclaimed
  2. 02

    A public transformer fire record

    A public table of fires and explosions at power transformers worldwide since 2015, one row per event: date, place, voltage and rating where reported, the part that failed (bushing, tap changer, winding or other), the consequences, and whether monitoring had warned beforehand, each with links to the reports it came from.

    The brief
    Credit · unclaimed

Open 01 · the brief

An open DGA test set

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public collection of DGA results from transformers that were later opened and inspected, each with the fault that was found, and a scored comparison on it of the IEC 60599 ratios, the Duval triangle (a chart that reads the fault type from three gases) and machine-learning models, reported separately for early and severe faults.

Why it is still open

A 2026 benchmark of nine models used a private set of 1,097 records, of which only 17 were labeled as faults. The public IEC TC 10 databases of inspected faults date from 2001. A search for an open, inspection-confirmed set used as a shared benchmark found none.

tugraz.atdoi.org

Built with what exists

The IEC TC 10 databases in Duval and dePabla (2001); the case studies in IEEE C57.104-2019 and in published conference papers; a language model that pulls gas tables out of PDFs and cites the page (see the Note "Findings that cite notes"); Python with scikit-learn and XGBoost.

  1. First step

    Transcribe the IEC TC 10 cases from the 2001 paper into a table and score the IEC 60599 ratios on them.

  2. Done looks like

    At least 300 inspected cases, each with seven gas values, the fault found and a source page, with a second checker agreeing on at least 95% of values, and every method scored by fault class with the code public.

Open 02 · the brief

A public transformer fire record

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public table of fires and explosions at power transformers worldwide since 2015, one row per event: date, place, voltage and rating where reported, the part that failed (bushing, tap changer, winding or other), the consequences, and whether monitoring had warned beforehand, each with links to the reports it came from.

Why it is still open

CIGRE's survey publishes only rates and shares, and its review found that national surveys usually lack where in the transformer a failure began and why. A search for a public, event-by-event record of transformer fires found none outside databases of nuclear-plant fires.

e-cigre.org

Built with what exists

News archives searchable through the open GDELT project; published incident reports from regulators and system operators, such as the US Nuclear Regulatory Commission's daily event notifications; a language model that drafts each row and cites the sentence it came from (see the Note "Findings that cite notes"); two human checkers.

  1. First step

    Collect one year of news reports of transformer fires in one country and code the first 30 events.

  2. Done looks like

    At least 200 events from 2015 to 2025 in at least 20 countries, each with a source link and the failed part recorded or marked "not reported," with a second checker agreeing on at least 90% of part codes, published under an open license.