Preparing the published article.
Inspection Robot Data: Fields That Drive Maintenance Work
How to record inspection runs and findings so they become dated, owned work orders instead of an unreadable folder of images.
Record the Run Before You Record the Finding
An inspection finding is only useful if someone can act on it, and that depends on the record of the run, not just the image. Two runs can produce the same photograph with completely different value: one has a location, a condition at the time, and a comparison point; the other has a timestamp and a folder name. Design the capture so the second case cannot happen.
Start with a stable asset and location reference that a maintenance planner already recognises. For a pipeline, that is distance from a defined origin plus a weld or feature reference; for a vessel, a plate or course reference; for a well or a plant room, the tag number your maintenance system uses. Test the reference by asking a technician who did not run the robot to navigate to the reported point from the record alone, and note how long it took.
- Asset and location reference, plus the reference used in the maintenance system.
- Run ID, operator, date, and the inspection route or coverage achieved.
- Method and settings used, so two runs can be compared on the same basis.
- Environmental conditions during the run: surface condition, lighting, temperature, product level.
- Raw files retained, in an uncompressed format where the equipment offers it.
Ask the supplier before purchase where the raw data lives, who owns it, how long it is kept, and what happens to it if the contract ends. Those answers belong in the procurement file, not in a remembered conversation.
Turn Findings Into Work Orders, Not Folders
The maintenance decision is where inspection value is won or lost. Record a severity judgement next to every finding, and record who made it. A finding without a severity cannot be prioritised, and a folder of images with no owner is a liability rather than an inspection.
Define the severity scale before the first run, and keep the scale short enough that two reviewers reach the same answer independently. Then link each finding to the maintenance workflow with four fields.
- Owner: the person or role accountable for the next step, with a date.
- Action decision: monitor, repair at next opportunity, or repair now, with the stated reason.
- Planned window: the outage, shutdown or opportunity in which the work can be done.
- Completion: the date the work was recorded as done, and the evidence attached.
Ask a narrow question of any inspection system: will a finding that nobody acts on generate a reminder, remain open, or disappear? Then decide how you will treat repeat findings. A defect seen at the same location on three consecutive inspections is a reliability signal, even if each individual report was classified as monitor; record the inspection count alongside the finding so the trend is visible without manual comparison.
Test the Record Against a Later Failure
The honest test of an inspection record is whether it predicted a failure that later happened. Plan one review per quarter in which you take the failures and unplanned repairs from the period and ask a simple question for each one: was there a prior inspection of that location, what did it report, and would the recorded severity have led to action under your current rules?
Run that review whether or not a robot is involved, because the same record supports manual inspection. Score the outcome with a short list of failure modes of the record itself.
- Missed: the location was covered, and the defect was present but not reported.
- Misclassified: a defect was reported at the location but described in a way that led to no action.
- Uncovered: the location was never inspected, which is a coverage problem rather than a detection problem.
- Acted-on: the finding led to a planned repair, and record what the repair confirmed.
Feed the result back into the inspection plan: extend coverage to the uncovered locations, tighten the severity definitions behind the misclassified cases, and adjust the interval for the classes that keep appearing. Record one change at a time and re-run the same review in the next quarter.
Then ask a second question that is easy to skip: does the record survive without the person who made it? Open a report from a year ago and check whether a reviewer who has never seen the asset can understand the location, the comparison, and the action taken. If not, the recording format needs work before the automation does.
Photo credits: the images in this article are licensed reference photographs of the relevant robot class, credited with author and licence in each caption; they are not photographs of the specific equipment discussed. Licence and author details are recorded with each image.