BRENNEN

Industries / Mining and materials / Chemicals and petrochemicals / Refining / Lubricants

Oil analysis in service

Oil analysis takes small samples of oil from running engines, gearboxes and hydraulic systems and tests them for wear metals, dirt, water and the oil's own breakdown, to judge both the oil and the machine it runs in.

6

Milestones achieved, from 1946 to 2011.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1946

    Railroads read wear metals

    After the Second World War, US railroads began burning samples of used diesel engine oil in a spectrograph, an instrument that identifies metals by the light they give off, to spot engines that were wearing out. The Society of Tribologists and Lubrication Engineers (STLE) dates the first success to 1946 at the Denver and Rio Grande Railroad's research laboratory; Machinery Lubrication puts it between 1946 and 1948 and names aluminum, copper, iron and lead as the metals watched. By 1957, papers at an ASTM symposium on railroad lubricating oils described direct-reading spectrometers, which use phototubes instead of photographic plates to measure many elements at once, testing used oil by the "Rotrode" method at the Chicago and North Western Railway.

    stle.orgmachinerylubrication.comastm.org

    Credit

    Denver and Rio Grande Western Railroad research laboratory; V. C. Barth, Chicago and North Western Railway

  2. 1961

    Army labs follow the Navy

    STLE's history says success on locomotives led the US Navy to try spectrometric oil analysis on ships and aircraft engines in the 1950s. The Army opened its first oil analysis laboratory at Fort Rucker, Alabama, in September 1961, after a Navy laboratory had found several defective Army helicopter transmissions through their oil; on its first day the Army lab identified a defective aircraft engine. Ground equipment was added from 1975. The three services now run a single Joint Oil Analysis Program under a shared regulation (AR 700-132, OPNAVINST 4731.1, AFI 21-131) that standardizes procedures, data and instruments across their laboratories, and STLE reports that Army labs analyze more than 1.1 million samples a year.

    stle.orgwebarchive.library.unt.edu

    Credit

    US Navy; US Army Oil Analysis Program

  3. 1972

    Particles sorted by magnet

    William Seifert and Vernon Westcott published "A method for the study of wear particles in lubricating oil" in the journal Wear in 1972, the basis of ferrography. A ferrograph pulls wear particles out of an oil sample with a magnet and lays them on a glass slide by size, so an analyst can tell iron from other metals, look at each particle in transmitted and reflected light, and heat the slide to tell steel from cast iron. Machinery Lubrication's history says the US military paid Westcott's company, Trans-Sonics, to develop it because chip detectors and spectrometers warned of bearing fatigue only when failure was close, and that the British used it to predict helicopter transmission failures during the Falklands War. ASTM standardized how to read the slides in D7690 in 2011.

    doi.orgmachinerylubrication.comastm.org

    Credit

    William Seifert and Vernon Westcott (Trans-Sonics)

  4. 2000

    Rotating disc test standardized

    ASTM published D6595 in 2000, a standard test for wear metals and contaminants in used lubricating oils and hydraulic fluids by rotating disc electrode atomic emission spectrometry, in which oil is sparked and the light from each element is measured, the same family of method the railroads' "Rotrode" used in the 1950s. The standard's own scope warns that results depend on particle size and that "low results may be obtained for those elements present in used oil samples as large particles."

    LessonA spark test on a drop of oil sees fine wear and can miss the large particles a failing part sheds, so labs add particle counts and ferrography.

    astm.orgastm.org

    Credit

    ASTM International

  5. 2010

    Counting particles, not droplets

    Automatic optical particle counters, used to grade how clean hydraulic and lubricating oil is, count anything that scatters light. ASTM D7647, first issued in 2010, dilutes the sample so that fine water droplets and "soft particles", such as silicone defoamer and oxidation by-products that cause "falsely high counts", drop out of the result and only hard particles are counted.

    LessonA count of everything that scatters light can call clean oil dirty, so a cleanliness number means little unless the method says what it counted.

    astm.org

    Credit

    ASTM International

  6. 2011

    Alarm limits get statistics

    Labs grade each result against alarm limits, usually four levels from normal to critical, often set by habit or copied from other machines. ASTM's guide D7720, first issued in 2011, gives statistical ways to test whether those limits are meaningful, asking whether they produce "too many false positive or false negative results."

    LessonA limit is only as good as the data behind it, and a limit that is never checked against outcomes can flag healthy machines or miss sick ones.

    astm.org

    Credit

    ASTM International

Watch for

  1. AI drafts gearbox diagnoses

    Researchers at Roma Tre University and the gearbox maker Bonfiglioli fed formatted oil analysis results from commercial industrial gearboxes to OpenAI's o1 language model, with prompts written for lubrication, and had one experienced analyst score its diagnoses. They report that more than 70% were usable as written, about a quarter needed revision, a little over 5% were unusable, and processing time fell by 60%; the comparison covered 30 diagnoses on the company's own data.

    Achieved whenanother laboratory repeats the comparison on its own samples, with several analysts scoring blind, and publishes the agreement figures.

    iris.uniroma3.it

    Claimed by

    Diego Rigolli, Alessandro Giorgetti and colleagues (Roma Tre University, with Bonfiglioli), August 2025

  2. Forecasting when oil fails

    Spectro Scientific, an AMETEK company that makes on-site oil analyzers, says its TruVu 360 Fluid IQ uses AI trained on sample histories of viscosity, base number (TBN, the oil's remaining reserve against acids), wear metals, water and ferrous debris to forecast when each machine's oil will reach risk thresholds. It gives no accuracy figures.

    Achieved whena user or outside laboratory publishes forecasts made in advance against the oil's later measured condition, across many machines.

    spectrosci.com

    Claimed by

    Spectro Scientific, April 2026

Open

  1. 01

    Score the AI report readers

    A public test set of real oil analysis reports, stripped of owner details, each read independently by at least three experienced analysts who give a severity grade and a recommended action. Then score AI models and commercial AI tools on the same reports, using how often the analysts agree with one another as the yardstick, and count separately how often each AI misreads a number.

    The brief
    Credit · unclaimed
  2. 02

    Infrared spectra, shared openly

    An open library of FTIR spectra (mid-infrared spectra, which labs trend to track oxidation, soot, water, fuel and coolant glycol in used oil) from engine, gear and hydraulic oils in service, each paired with the spectrum of the same oil when new and with reference laboratory values for water (Karl Fischer titration), acid and base number, fuel dilution and glycol, and measured on at least two makes of spectrometer.

    The brief
    Credit · unclaimed

Open 01 · the brief

Score the AI report readers

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 oil analysis reports, stripped of owner details, each read independently by at least three experienced analysts who give a severity grade and a recommended action. Then score AI models and commercial AI tools on the same reports, using how often the analysts agree with one another as the yardstick, and count separately how often each AI misreads a number.

Why it is still open

The one published comparison (Rigolli and colleagues, 2025) scored 30 gearbox diagnoses against a single analyst on a manufacturer's private data. Blackstone Laboratories reported in May 2026 that ChatGPT misread its iron, silicon and potassium results on a sample report when the analyst's comments were left out, and called potassium a "serious red flag" when it was explained by an additive. The public DACON competition set from South Korea (2022) labels construction-machine oil samples only as normal or abnormal, with no written diagnosis. In an STLE poll of readers in November 2025, 72% expected AI to change how results are interpreted. Searches of arXiv, Hugging Face, Zenodo and GitHub in November 2026 found no public set of reports read by several analysts.

iris.uniroma3.itblackstone-labs.comgithub.comstle.org

Built with what exists

Reports donated by fleets and independent labs; Label Studio for the analysts' grading; open-weight models run on the team's own computers so customer data stay private (see the Note "A trillion open parameters"), alongside hosted models; Python and scikit-learn for agreement statistics such as Cohen's kappa, a measure of agreement beyond chance.

  1. First step

    Ask two independent oil labs and one public vehicle fleet to donate 50 anonymized reports each, with the analysts' comments split into a separate file.

  2. Done looks like

    At least 300 reports from engines, gearboxes and hydraulic systems, each graded by at least three analysts and released under an open license, with at least five AI systems scored on severity agreement (kappa) and on number-reading errors, set beside the analysts' agreement with one another.

Open 02 · the brief

Infrared spectra, shared openly

Credit · unclaimed

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

What to build

An open library of FTIR spectra (mid-infrared spectra, which labs trend to track oxidation, soot, water, fuel and coolant glycol in used oil) from engine, gear and hydraulic oils in service, each paired with the spectrum of the same oil when new and with reference laboratory values for water (Karl Fischer titration), acid and base number, fuel dilution and glycol, and measured on at least two makes of spectrometer.

Why it is still open

ASTM E2412 sets out how to trend these spectra, but says calibration to physical units is "unnecessary or impractical in many cases" and leaves alarm limits to each user, so the spectra and calibrations labs rely on stay private. Published calibrations, such as Villar and colleagues' 2011 visible and near-infrared sensor model for gas-engine oil and an RMIT University study of military aviation turbine oils (2024), do not release a reusable spectral set. Searches of Zenodo, figshare, Hugging Face and GitHub in November 2026 found no open library of in-service oil spectra with reference values.

astm.orgzenodo.orgresearch-repository.rmit.edu.au

Built with what exists

Benchtop or portable FTIR spectrometers already in university and fleet labs; samples drawn at routine oil changes; a commercial oil lab for reference values; Python with scikit-learn for partial least squares calibration; Zenodo to publish with a data card.

  1. First step

    Collect 50 samples at normal oil changes from one university or municipal vehicle fleet, run each on two spectrometers, and send splits to a commercial lab for reference values.

  2. Done looks like

    At least 1,000 spectra from at least 300 machines with reference values, released under an open license, and a published test of how well a calibration built on one spectrometer predicts water and acid number on another.