Agents Megger test records that file and flag themselves
It's late afternoon at the combiner boxes and the tester is staring at a megger display, shielding it from glare, a smudged paper form in his other hand. The next reading is good. The one after it is not — and whether anyone notices today depends on that form making it back to the trailer.
The problem
Insulation-resistance testing produces records that matter most and get kept worst. A string tests low, the reading goes on a form or a glove, and three weeks later commissioning wants the complete log — values, dates, instrument, who tested — and half of it is reconstructed from memory and meter photos.
That reconstruction is the risk. A record written after the fact looks identical to one written at the tester, and a marginal reading never flagged means a string gets energized and troubleshooting later starts from zero. The paperwork is clerical work done tired, outdoors — where readings get transposed and strings mislabeled.
How the agent workflow works
The tester still tests. The agent owns everything after the reading exists.
What it reads:
- A photo of the meter display or test form, sent as each string or feeder is completed.
- The string schedule or feeder list, so every expected test has a slot and a missing test is visible.
- The project's acceptance criteria — minimum values and test voltage — loaded once, from the spec, by a human.
What it produces: A running test log: each reading transcribed, tied to its string or feeder ID, date, and instrument, and checked against the criteria. Anything below threshold, borderline, or missing is flagged the same day, while a retest is still cheap.
Where the human checks it: The tester verifies the transcription against the photo before it counts as final — a transposed digit reads exactly like a real reading. The superintendent reviews the flags and decides what gets retested, investigated, or accepted with a note. The agent never passes or fails a string. It files, compares, and points.
What you need to set it up
- A capture habit the crew will keep: one photo, sent one place, per test. A group chat or email alias works; a new app nobody opens does not. Cost is small: this runs on the AI assistant and storage you already pay for.
- The string and feeder schedule in a readable list, matching field IDs. If drawings, labels, and schedule call the same string three things, fix that first.
- Acceptance criteria pulled from the actual spec and manufacturer requirements, entered by a person who has read them — never a guess or a default.
- A meter whose display photographs legibly. A shaded photo habit beats a smarter model.
Setup is a half day: build the log format, load the criteria, run one section in parallel with paper, and compare.
Where it breaks
Photo legibility. A decimal point lost to glare, a 7 that photographs like a 1 — transcription will misread sometimes, which is why verification is not optional. Treat the photo as the record and the transcription as a convenience until checked.
Mislabeled strings. If the field label is wrong, the agent files a perfect record against the wrong string — worse than a missing record, because it looks closed. Spot checks against the drawings are the only cure.
Criteria are project-specific. Voltages and acceptance values vary by equipment, cable, and spec. An agent carrying last project's thresholds will flag good strings and pass bad ones with total confidence. Reload the criteria, by hand, every project.
Weather context gets lost. Readings move with moisture and temperature. A low reading on a wet morning may be conditions, not damage; what a reading means stays with the humans who were standing there.
The bottom line
This does not make testing faster. It makes the records trustworthy: every reading filed the day it was taken, checked against the right threshold, with failures surfaced while a retest is still easy. Run it alongside paper for one section and compare the logs at handover.