MaintenanceVerified Real deployment at a named or precisely described company, with a measured result from a primary source.

AI reads oil condition from sight-glass photos and lab reports, so lubricant changes follow the machine, not the calendar (vacuum pump maker, railcar component maker)

Added 2026-11-12 · Last updated 2026-11-12

Role

Maintenance and reliability (lubrication program and oil analysis) · Typical users: reliability engineer, maintenance planner, lubrication technician, maintenance manager, service engineer at an equipment manufacturer

The problem

Many failures of gearboxes, pumps and compressors start with degraded or contaminated oil, but oil checks are often a quick look through a sight glass or a lab report that nobody reads closely. At the largest global maker of industrial vacuum pumps (client of Saviant, not named), oil health was judged by technicians looking through pump sight glasses during maintenance: decisions differed between people, early degradation was missed, oil changes followed a fixed schedule and the company paid for oil-related failures under warranty. At a heavy haul railcar component maker (pilot with KCF Technologies and FUCHS), the customer later confirmed that 38% of the issues flagged by machine monitoring were linked to lubrication, mainly degradation and contamination. Lab oil analysis itself takes expert time: a 2025 peer-reviewed study on industrial gearboxes describes long analysis times and a strong dependence on the analyst's interpretation.

How AI helps

Two complementary approaches. At the vacuum pump maker, a convolutional neural network (Azure Custom Vision) trained on about 1,000 labelled sight-glass photos classifies oil as healthy, degrading or critical with a confidence score; pump ID, operating hours and service history are added, degradation rates are tracked over time to estimate remaining useful life, and engineers use a web portal to upload images and see the trend. For lab results, the gearbox study combined spectrometry, particle counts, viscosity and FT-IR data with an AI model that drafts the diagnosis (wear, lubricant properties, contamination) for an expert to confirm. At the railcar component maker, vibration monitoring flagged a grinder gearbox; oil sampling then showed wear metals and large particles, and the leaking cover was repaired before failure. In each case people decide when to change oil or open the machine.

Step by step

  1. List the assets where oil matters most (gearboxes, hydraulic units, compressors, vacuum pumps) and the cost of a failure for each.
  2. Gather what you already have: lab reports, sight-glass photos taken the same way each time, oil change history, failures and their causes.
  3. Start with lab reports: have an AI tool summarise trends per asset and flag the samples an expert should look at first.
  4. If you use photos, standardise them (same light, same distance, same sight glass) and have an expert label a few hundred before training anything.
  5. Link every alert to an action in the CMMS (resample, filter, top up, change oil, inspect seal) and record what was found.
  6. Review each quarter: alerts confirmed, oil changes avoided or brought forward, failures that the program missed.

Tools

  • Custom CNN built with Azure Custom Vision and a web portal, developed by Saviant: used at a global vacuum pump manufacturer (not named)
  • KCF Technologies SMARTdiagnostics (vibration monitoring) with FUCHS lubrication program: used at a heavy haul railcar component manufacturer (not named)
  • Other tools to evaluate: the online portal and AI comments of your oil lab, in-house analyzers such as eralytics eralab OCM, inline oil sensors, the condition module of your CMMS
  • To try it without new software: your last lab reports and any LLM (ChatGPT, Claude, Copilot, Gemini) with the prompt below

In the AI tools index: KCF Technologies SMARTdiagnostics · all AI tools used by manufacturers

Result

Vacuum pump maker (Saviant case study, proof of concept): an estimated 30% potential reduction in pump failures caused by contaminated or degraded oil and 20% warranty savings in one year; next phase combines images with lab tests, ISO particle counts, vibration and temperature. Gearbox study (Machines, MDPI, 2025): AI diagnoses agreed with expert evaluations, especially for wear and lubricant properties, more than 75% were rated sufficiently accurate, and diagnostic time dropped by about 60%. Railcar component maker (KCF Technologies case study): a critical gearbox failure avoided, oil sampling every two months added, and a long-term monitoring contract signed after the pilot; no figures on cost.

As reported by the source; most figures are vendor or self-reported and not independently verified.

Human review: cautions

  • The vacuum pump result is a proof of concept with a 'potential' failure reduction, published by the consultant; neither manufacturer is named.
  • A photo through a sight glass shows colour and clarity, not wear metals: it complements lab analysis, it does not replace it.
  • An AI summary of lab reports can misread units or limits; keep an expert review on every flagged asset, especially at the start.
  • Oil data is only useful if samples are taken the same way, from the same point, at regular intervals.

Try it yourself

You are a reliability engineer responsible for our lubrication program. I want to know which machines need attention first based on oil analysis.

I will paste: (1) the last three to six oil lab reports for each asset [asset, oil type, date, hours on oil, viscosity, water, wear metals such as Fe, Cu, Pb, Al, particle count ISO code, TAN or oxidation, lab comment]; (2) the alarm limits we use or that the lab gives.

Do this:
1. For each asset, summarise the trend (stable, rising, sudden change) for wear metals, contamination and oil degradation, in one line.
2. Flag the assets that need action, ranked by risk, with the likely cause (wear, contamination, wrong oil, overdue change) and the evidence.
3. Propose the next action for each flagged asset (resample, filter, change oil, inspect seals or breathers, vibration check) and a resampling interval.
4. List the assets where the data is too thin or inconsistent to conclude.
5. Draft a short note for the maintenance meeting.

Do not invent limits or values: use only the limits I give and say when one is missing. Use only the data I paste in; if something is missing, write 'not found' or list it as a question instead of guessing. Do not send, order, post or change anything in any system: produce drafts for my review and approval.

Want more like this? Join the newsletter

Related cases

All 14 Maintenance use cases →

Sources

Weekly: 1–2 practical AI use cases for your plant

Newsletter in

Free, one email a week. By subscribing you agree to receive the weekly newsletter (and its sponsor message) from AI Agents for Manufacturing. Every email includes an unsubscribe link, and you can unsubscribe anytime. See our Privacy Policy.