BRENNEN

Industries / Industrial equipment and services / Industrial services / Pump and valve repair

Repair records

Repair records say who is allowed to repair a pump or valve, what was found and done each time, and how often each kind of equipment fails, so that plants can compare their results and judge whether their maintenance works.

6

Milestones achieved, from 1946 to 2026.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1946

    A code for repairs

    The National Board of Boiler and Pressure Vessel Inspectors, founded in 1919, first published the National Board Inspection Code (NBIC) in 1946. It sets rules for installing, inspecting, repairing and altering pressure equipment already in service, and how each repair is to be checked and documented; one of its four parts covers pressure relief devices, the valves that open to stop equipment bursting. The Board says most US and Canadian jurisdictions have adopted the code into law and that more than 5,200 repair organizations in over 60 countries use it.

    nationalboard.org

    Credit

    the National Board and its NBIC committee

  2. 1977

    A stamp for valve repairers

    In January 1977 the National Board's Executive Committee approved procedures for a certificate that lets a repair shop use the Board's "VR" stamp on the safety valves it repairs; a later Board presentation dates the program to 1976. To hold one, a shop needs a written quality system, and sample valves it has repaired are tested for opening pressure and flow at an accepted laboratory while a Board representative watches. Joseph Ball, head of the Board's pressure relief department, said in 2021 that the Board had tested almost 8,900 valves for the program, about 300 a year.

    LessonThe program answered repairs the Board had seen go wrong, such as wrong springs, seats lapped flat where the maker called for a bevel, and a nozzle plugged to stop a leak, which cut the valve's capacity; its pressure relief director also warned in 1983 that a valve repaired many times by different shops carries no assurance that its parts are still correct.

    nationalboard.orgnationalboard.orgnationalboard.orgnationalboard.org

    Credit

    the National Board

  3. 1984

    Energy companies pool failure data

    Eight energy companies in the OREDA project began collecting reliability and maintenance data on offshore equipment in 1981 and issued the OREDA Reliability Data Handbook in 1984, with further handbooks in 1992, 1997 and 2002. OREDA calls itself the cradle of ISO 14224, first published in 1999, the international standard for recording equipment reliability and maintenance data in one format so that plants, owners, makers and contractors can exchange and compare it. The 2016 edition, which covers petroleum, natural gas and petrochemical plants, calls its list of failure modes a "reliability thesaurus".

    oreda.comwebstore.ansi.orgwebstore.ansi.org

    Credit

    OREDA's member companies; ISO's committee for the petroleum industries

  4. 1989

    A council for valve repairers

    In the mid- to late 1980s the Valve Manufacturers Association of America (VMA) saw what its magazine later called a huge increase in spurious valve work, most notably bogus valves supplied to nuclear power plants, along with lawsuits over rebuilt valves sold as new. The service operations of its member companies formed a Valve Remanufacturers Council, which held its first official meeting in Houston on 24 May 1989 under its first executive director, Bill Sandler, to promote safety and set proper methods for repairing and rebuilding valves and actuators. In 1994 VMA's board proposed renaming it the Valve Repair Council, to set it apart from rebuilding and from surplus valve sales.

    valvemagazine.com

    Credit

    VMA and its members' service operations

  5. 1991

    Nuclear plants check maintenance

    The US Nuclear Regulatory Commission's maintenance rule, 10 CFR 50.65, published on 10 July 1991 and required at every plant by 10 July 1996, makes each nuclear power plant monitor the performance or condition of its important equipment against goals it sets, take corrective action when the goals are missed, and review the results at least every 24 months in light of industry-wide experience. It also requires plants to assess the added risk before doing maintenance.

    law.cornell.edu

    Credit

    US Nuclear Regulatory Commission

  6. 2026

    Pump failures trended nationwide

    In February 2026 Zhegang Ma and Sai Zhang of Idaho National Laboratory published, for the Nuclear Regulatory Commission, failure rates for motor-driven pumps at US commercial nuclear plants from 1998 to 2024. They drew on failure reports that plants file with the Institute of Nuclear Power Operations (INPO) in its IRIS database, formerly EPIX and ICES, where earlier studies had used the older NPRDS system. Over the latest 10 years, standby pumps were out of service less often and were less likely to fail during a 24-hour run.

    osti.gov

    Credit

    Idaho National Laboratory, INPO and the US Nuclear Regulatory Commission

Watch for

  1. Work orders from voice

    Fluke says the beta of the AI features in its eMaint maintenance software turns technicians' voice messages into structured work requests, answers questions from a plant's own work orders and manuals, and writes procedures and preventive maintenance tasks from equipment manuals, which it says can cut the time to create preventive maintenance tasks by 50% or more.

    Achieved whena customer or outside reviewer publishes the time saved, with the share of AI-written work requests and procedures that technicians had to correct.

    pressroom.fluke.com

    Claimed by

    Fluke, March 2026

  2. Failure codes from work orders

    Akash Raj reports that Plant Memory, software whose code is on GitHub and which reads work orders exported from SAP's maintenance module into a record of components, failure modes and causes aligned with ISO 14224, found the right failure mode 88.9% of the time and the right component 96.2% of the time. The test used a made-up year of 400 work orders from a rice plant, and an optional AI model acts only as a second reader whose disagreements push a record up the technicians' review queue.

    Achieved whensomeone other than the author runs it on real work orders from a working plant and publishes its accuracy against engineers' coding.

    arxiv.org

    Claimed by

    Akash Raj, in an arXiv preprint, October 2026

Open

  1. 01

    Code real pump work orders

    A public test set of real pump and valve work orders from several plants, each coded by two reliability engineers with the ISO 14224 failure mode and cause, on which AI models are then scored.

    The brief
    Credit · unclaimed
  2. 02

    Public pump repair intervals

    A public table of how long pumps run between repairs, the mean time between repairs or MTBR, at several water and wastewater utilities, built from work orders obtained by public records requests and broken down by pump type and size, so that utilities can be compared.

    The brief
    Credit · unclaimed

Open 01 · the brief

Code real pump work orders

Credit · unclaimed

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

What to build

A public test set of real pump and valve work orders from several plants, each coded by two reliability engineers with the ISO 14224 failure mode and cause, on which AI models are then scored.

Why it is still open

The test sets published so far are small, from other equipment or made up. Michael Stewart, Melinda Hodkiewicz and Sirui Li scored GPT-3.5 on 62 observations from work orders labeled with 22 ISO 14224 failure modes, from an open set of 626 (F1 0.80 after fine-tuning, 0.46 without, 2023). The same University of Western Australia group's MaintIE corpus (2024) has two experts' labels on 1,076 texts sampled from real work orders, mostly on mining assets, but it marks objects, states and activities, not ISO 14224 failure modes. Marc-Alexander Lutz and colleagues coded wind turbine work orders (2023); ARID was scored on 300 hand-labeled maintenance records from the OMIn set (September 2026); Plant Memory was scored on 400 made-up work orders (October 2026); WO2Data, an open converter from work order text to ISO 14224 records begun on GitHub in October 2026, has so far been tried only on made-up data. No public set of real pump and valve work orders coded by two engineers was found.

arxiv.orgaclanthology.orgarxiv.orgarxiv.orgarxiv.orggithub.com

Built with what exists

Anonymized work order exports from SAP Plant Maintenance or IBM Maximo, shared with each plant's permission; the failure mode and cause tables of ISO 14224:2016; Label Studio for double labeling; a model that gives probabilities for each answer on a fixed list, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"); the open Plant Memory code as a baseline.

  1. First step

    Ask one plant's reliability engineer for a year of anonymized pump work orders, and code 100 of them by hand to settle the code list.

  2. Done looks like

    At least 2,000 pump and valve work orders from at least three plants, each coded by two engineers with their agreement reported as Cohen's kappa; at least three models scored by F1 per failure mode on a held-out 500; the set released with plant and person names removed.

Open 02 · the brief

Public pump repair intervals

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 how long pumps run between repairs, the mean time between repairs or MTBR, at several water and wastewater utilities, built from work orders obtained by public records requests and broken down by pump type and size, so that utilities can be compared.

Why it is still open

Nuclear plants pool pump failure reports through INPO and Idaho National Laboratory publishes the rates, and offshore operators pool data in OREDA, whose database is open to members and paying subscribers. For other plants, the MTBR figures in circulation are consultants' estimates, such as a 2016 post on the consultancy Becht's blog that puts top plants above 100 months and warns that comparing pump MTBR between plants needs care. Single utilities have been studied: Q. Zhang's 2023 master's thesis at TU Delft modeled the time between failures of 447 sewage pumps from Rotterdam's public works records for 2016 to 2020, and recommended going beyond Rotterdam. A search of the Socrata open data catalog in November 2026 found lists of pump station locations, such as Seattle's and Edmonton's, but no public pump work orders, and no comparison across utilities by pump type was found.

becht.comosti.govoreda.comrepository.tudelft.nlcos-data.seattle.gov

Built with what exists

Public records requests to utilities for pump asset lists and five years of pump work orders; pandas; the lifelines library for survival curves; ISO 14224's definitions to decide what counts as a repair; a model that sorts free-text work orders into repair or not repair, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"), checked by hand on a sample.

  1. First step

    Send records requests to three utilities, such as Seattle Public Utilities, whose pump station lists are already on open data portals, for five years of pump work orders.

  2. Done looks like

    At least 10 utilities and 3,000 pumps, with MTBR and survival curves by pump type and size; a random 200 work orders coded by a second person, with at least 95% agreement on which ones were repairs; the method, code and anonymized data published.