BRENNEN

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

Diagnosis

Diagnosis is the work of finding out what is wrong with a pump or valve while it is still installed, from its vibration, its sound, the current its motor draws or the way it moves, before anyone opens it.

6

Milestones achieved, from 1967 to 2020.

2

To watch for: announced, not yet confirmed outside.

3

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1967

    Probes watch the shaft

    Before electronic sensors, operators judged a machine's vibration by ear and by touch, through a fingernail, a screwdriver or a broom handle held against the bearing case, according to the brochure for an ASME engineering landmark. ASME says Bently Nevada, in Minden, Nevada, since 1961, was the first to make a commercially successful eddy-current proximity probe, a sensor that measures the gap to a spinning shaft without touching it, and calls its standard 5000 series monitors of 1967 a watershed. The American Petroleum Institute adopted such probes for vibration acceptance tests of compressors in 1970 (API 617), and its API 670 standard of 1976 set out machine protection systems built on them.

    asme.orgasme.org

    Credit

    Donald E. Bently and Bently Nevada; named an ASME landmark on 20 June 2024

  2. 1985

    Two valves would not reopen

    On 9 June 1985 the Davis-Besse nuclear plant in Ohio lost all feedwater to its steam generators after a reactor trip. Its two auxiliary feedwater isolation valves, AF-599 and AF-608, are motor-operated valves, opened and closed by an electric motor through a gearbox; they had shut and could not be reopened from the control room. The utility traced the fault to the torque switch bypass contacts, which keep the motor's torque cutoff out of the circuit while a valve first breaks free of its seat. On AF-608 they were set close to the procedure's value, which was itself believed to be too early in the stroke, and on AF-599 they were "grossly misadjusted" under a procedure the utility called "bulky and difficult". The NRC's investigation team noted that this root cause rested on analysis, not on a test that reproduced the failure, and the utility then tested the valves with about 1,000 psi of differential pressure, the difference in pressure across the closed valve.

    LessonSettings worked out on paper and checked by counting handwheel turns could not show whether the valves would open against full pressure.

    osti.gov

    Credit

    the NRC incident investigation team, report NUREG-1154, July 1985

  3. 1986

    The motor becomes the sensor

    Howard D. Haynes and David M. Eissenberg of Oak Ridge National Laboratory (ORNL), working on NRC-funded research into aging nuclear plant equipment, applied in 1986 for a patent on motor current signature analysis. The method reads the small, fast changes in the current drawn by the motor of a motor-operated valve to see what its gears, stem and valve are doing, and the current can be measured at a distant control cabinet without fitting any sensor to the valve. The patent, US 4,965,513, was granted in 1990. By 1989 ORNL reported using the method on vacuum pumps, water pumps and air blowers, and at EPRI's monitoring and diagnostics center at the Eddystone power plant it was recording eight extraction valves on line.

    osti.govosti.govosti.govosti.gov

    Credit

    Howard D. Haynes and David M. Eissenberg, ORNL

  4. 1989

    Valves tested at full pressure

    In Generic Letter 89-10 of 28 June 1989, the US Nuclear Regulatory Commission asked every US nuclear plant to set the switches of each safety-related motor-operated valve from a documented design basis, to prove where practical that the valve works against its design-basis differential pressure or flow, to stroke-test every one, to trend failures, and to recheck the settings at least every five years or three refueling outages. Blowdown tests at the Idaho National Engineering Laboratory had shown that some valves met loads not accounted for before. The letter said that diagnostic systems measuring both stem thrust and motor torque were best suited to the job.

    LessonThe letter warned that test results at low pressure "cannot be directly extrapolated" to the pressure a valve must work against in an accident.

    nrc.gov

    Credit

    US NRC

  5. 2015

    The standard test set, checked

    Case Western Reserve University's Bearing Data Center publishes vibration recordings from a 2 hp electric motor whose bearings had faults 0.007 to 0.040 inch across cut into them by spark erosion, run at loads of 0 to 3 hp. The tests began as a way to check a Rockwell bearing monitor, and the data became the standard set for testing new bearing diagnosis methods. In 2015 Wade A. Smith and R. Randall ran three established diagnostic techniques over every record, sorted the records by what could be diagnosed, and set out how the data should be used and how future benchmark data should be made.

    LessonThe authors wrote that, without a recognized benchmark, it is difficult to judge any method tested on the data.

    engineering.case.eduapi.semanticscholar.org

    Credit

    Case Western Reserve University; Wade A. Smith and R. Randall, Mechanical Systems and Signal Processing

  6. 2020

    A contest to hear faults

    In 2019 seven researchers including Harsh Purohit and Yohei Kawaguchi released MIMII, a free set of sound recordings of valves, pumps, fans and slide rails in factory conditions, with faulty sounds such as contamination, leakage and rotating unbalance. The DCASE 2020 challenge, an open contest in machine listening, used it to score systems that learn only from normal sounds and must flag abnormal ones. On the test recordings, the organizers' baseline scored an AUC of 82.4 on pumps and 57.4 on valves, and the winning system scored 93.7 and 96.1. AUC is a score on which 50 means no better than chance and 100 means every faulty recording was ranked above every normal one.

    arxiv.orgdcase.community

    Credit

    the MIMII authors; the DCASE 2020 Task 2 organizers; the winning entry from a team at Amazon (Giri and others)

Watch for

  1. Current watches sewage pumps

    Samotics says its SAM4 system, which reads motor current and voltage at the motor control cabinets of Yorkshire Water's pumping stations and uses AI to sort changes into fault types, has found 726 developing faults, such as blockages, airlocks and misalignment, and prevented 40 pollution events since the partnership began in 2022. Yorkshire Water is adding about 1,500 more pumps over 18 months. The article says the platform is "independently validated at 97% accuracy with fewer than 1% false positives" but does not say who did the validation or how.

    Achieved whenthe validation is published with who ran it, how many flagged faults were confirmed when the pumps were lifted or opened, and how false alarms were counted.

    watermagazine.co.uk

    Claimed by

    Samotics and Yorkshire Water, in Water Magazine, July 2026

  2. Stiction forecast four hours ahead

    Natthapong Promsricha and three co-authors report a machine learning model that predicts stiction, the static friction that makes a control valve stick and then jump, up to four hours ahead, using only the controller output and flow signals that plants already record, trained on data from an oil and gas refinery. Their labels came from a rule applied to the same signals, not from tests or inspection of the valves, and the authors call it, to their knowledge, the first such prediction from real industrial data.

    Achieved whenanother team reproduces the forecast on a different plant's data, with stiction confirmed by valve tests or inspection.

    arxiv.org

    Claimed by

    Promsricha, Chatpattanasiri, Kerdgongsup and Balabani, in an arXiv preprint, January 2026

Open

  1. 01

    Check the call at teardown

    A public table that pairs what was diagnosed before a pump or control valve was removed (the vibration report, motor current reading or valve test, with the fault named and its severity) with what the repair shop found when it opened the machine, and scores how often the call was right.

    The brief
    Credit · unclaimed
  2. 02

    Blind trial of pump monitors

    A test in which a third party seeds known faults into a small water pump loop, such as unbalance, misalignment, a damaged bearing, a chipped impeller and a throttled suction that makes the pump cavitate, while several wireless monitors sold for pumps run on it. The sellers do not see the schedule, and the results say which monitor flagged what, and when.

    The brief
    Credit · unclaimed
  3. 03

    AI analysts against certified people

    A test that gives AI models and certified vibration analysts the same pump cases (spectra, waveforms, running speed, bearing numbers and a sketch of the machine) and scores both against the fault that was actually seeded.

    The brief
    Credit · unclaimed

Open 01 · the brief

Check the call at teardown

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 that pairs what was diagnosed before a pump or control valve was removed (the vibration report, motor current reading or valve test, with the fault named and its severity) with what the repair shop found when it opened the machine, and scores how often the call was right.

Why it is still open

Published accuracy figures are scored against other references, not against what was found inside. SymphonyAI's 2021 study scored its automated diagnostic system and three engineers on 332 machines, including centrifugal and screw pumps, against a senior engineer's reading of the same data, and that reading started from the system's own output. Samotics says its pump system is independently validated at 97% accuracy without naming the validator. Open data sets such as the Netherlands Navy's NLN-EMP use faults seeded on purpose on a test rig. A search for studies that score field diagnoses against teardown findings found only single case stories.

symphonyai.comwatermagazine.co.ukebi.ac.uk

Built with what exists

Anonymized diagnosis reports, work orders and teardown inspection sheets with photos from cooperating plants and repair shops; a model that answers the same fixed questions about each report, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"); DuckDB or a spreadsheet; a second person to check the coding.

  1. First step

    Ask one pump repair shop for 20 recent jobs that arrived with a customer's vibration report, and code them by hand to settle the fields.

  2. Done looks like

    At least 200 pumps and valves from at least three plants or shops, each with a diagnosis made before removal and a teardown finding; a second person checks 40 of them; published agreement rates by fault type (bearing, impeller, misalignment, cavitation damage, seal, valve packing and seat) and the share of removed equipment found with no fault.

Open 02 · the brief

Blind trial of pump monitors

Credit · unclaimed

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

What to build

A test in which a third party seeds known faults into a small water pump loop, such as unbalance, misalignment, a damaged bearing, a chipped impeller and a throttled suction that makes the pump cavitate, while several wireless monitors sold for pumps run on it. The sellers do not see the schedule, and the results say which monitor flagged what, and when.

Why it is still open

Labeled pump data exists. NLN-EMP, published by Sietze Bruinsma and colleagues under a CC0 license (version 4, August 2024, 20.8 GB), holds vibration, current and voltage from two centrifugal pumps with bearing defects, loose and soft foot, impeller damage, misalignment, unbalance, cavitation and other faults at several severities. The nearest test of the sensors themselves is by a seller: in January 2025 KCF Technologies, which makes wireless sensors, reported shaker-table tests of a stand-in built to match common wireless sensor designs, from which it projected errors of 60% to 410% in sideways vibration, without naming the products or seeding any machine faults. A search for an independent side-by-side test of commercial monitors on seeded faults found only vendor articles and university prototypes, and the accuracy figures quoted for products, such as Samotics' 97%, come from the companies selling them.

data.4tu.nlkcftech.comwatermagazine.co.uk

Built with what exists

A small end-suction water pump and motor on a baseplate, with a tank, valves, pressure gauges and a guard over the coupling; seeded parts (bearings with notched races, an impeller with a ground-off vane tip, shims for misalignment, trim weights); a reference accelerometer and an open data logger; three to five battery-powered wireless vibration monitors bought at list price, with their apps; a fault schedule held by someone outside the test; a camera to film each change.

  1. First step

    Build the loop, log a week of healthy running with the reference sensor, and publish the test protocol with an invitation to vendors.

  2. Done looks like

    At least five fault types at two severities, each run three times in random order over at least six weeks; a published table of detections, time to first alert, fault named and false alarms for each product; the reference data released under an open license.

Open 03 · the brief

AI analysts against certified people

Credit · unclaimed

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

What to build

A test that gives AI models and certified vibration analysts the same pump cases (spectra, waveforms, running speed, bearing numbers and a sketch of the machine) and scores both against the fault that was actually seeded.

Why it is still open

Studies so far score AI models against other models or simple baselines, mostly on bearing data alone. Mehadi Hasan Shawon and Prashant Kumar (September 2026) tested eight small language models that run on a local computer on three classes of bearing fault from the Paderborn, Case Western and HUST data sets, and found that logistic regression and even an untrained physics rule beat all eight. SymphonyAI's 2021 test compared its expert system with engineers but used a senior engineer's reading as the answer key. No test was found of AI models against certified analysts on labeled pump data with many fault types.

api.crossref.orgsymphonyai.com

Built with what exists

The NLN-EMP data, whose faults are documented case by case; Python with SciPy to draw spectra; frontier AI models through their APIs and an open-weight model such as Mistral Large 4 (see the Note "A trillion open parameters"), run offline; volunteer analysts who hold a vibration analysis certificate.

  1. First step

    Download NLN-EMP, cut 20 cases into a one-page report with spectra and machine data, and try them on two analysts and one model.

  2. Done looks like

    100 cases drawn at random from NLN-EMP across at least eight fault types; each case scored by at least five certified analysts and three AI models; published accuracy by fault type with confidence intervals, and the time taken per case.