Field Notes
Lessons from the plant floor
Written by the engineers who commission our systems, train our models and take the night-shift calls. What works on real lines, what fails, and why.
About Field Notes
Written by the people who do the work
Every article is written by a Perflens engineer, analyst or manager about a problem we have met in the field: a model that drifted with the seasons, a lighting angle that finally made a defect visible, a vibration spectrum that gave a bearing away weeks before it failed. Each one names the physics, the trade-offs and the limits, and none of them asks for your email address before you can read it.
Filter by topic below: computer vision, edge AI, maintenance, operations, security, facilities and strategy. New to the vocabulary? The glossary explains the terms we use.
All articles
Browse Field Notes by topic
Vision-language models on the production line: where they help and where they hallucinate
A VLM can tell an operator why a pouch was rejected and find every similar reject from last month. It should never be the thing that decides to reject it.
What a minute of downtime really costs: a calculation you can defend
Most downtime cost figures fall apart at the first question from finance. A conservative method built on contribution margin, idled labor and restart scrap, worked through with hypothetical numbers.
An IEC 62443 hardening checklist for edge devices
IEC 62443 is a long series with a short practical core. This is the checklist we work from when hardening an edge device, mapped to the standard, plus what the EU Cyber Resilience Act now expects of device makers.
Outcome-based pricing for industrial AI: how it works and when it fits
Paying for results sounds like the vendor finally has skin in the game. It can be, if the baseline is measured, the metric resists gaming and the outcome is one the vendor can actually move.
Choosing edge AI compute in 2026: Jetson, Hailo or Qualcomm?
Peak TOPS is the least useful number on the datasheet. Pick edge compute by workload, thermal envelope, toolchain and the engineering time it will cost you.
Continuous thermal monitoring vs periodic IR surveys for electrical assets
A scheduled infrared survey shows you one afternoon in the life of your switchgear. What that snapshot misses, what fixed thermal sensing adds, and how to combine the two into one program.
Why your inspection model got worse in July
The part didn't change. The light, the humidity and the resin supplier did. A field guide to the quiet ways vision models decay — and the monitoring that catches it.
Hyperspectral imaging for foreign-material detection in food
X-ray finds what is dense. Spectral imaging finds what is chemically different. An optical engineer's guide to when spectral imaging earns its place on a food line, and how to build one that keeps working.
Seeing without surveillance: privacy-preserving sensing in buildings
Facilities teams need to know how many people use a space, not who they are. How to design occupancy sensing that answers the first question and cannot answer the second.
Synthetic defect data: when generated images help and when they hurt
Generated defect images can rescue a model starved of rare examples, or quietly teach it to spot the generator instead of the defect. The difference is in how you validate.
Compressed-air leaks: the utility nobody is watching
Most plants meter electricity, water and gas closely, then let their most expensive utility run unmeasured. A practical guide to finding leaks, fixing them and keeping them fixed.
Air-gapped AI: running inspection and monitoring without the cloud
Plenty of plants will never let a camera feed touch the internet. Here is what has to live on site, how models still get updated and retrained, and what you give up.
Anomaly detection or supervised classification? Picking the right defect model
One model learns what good looks like; the other learns what each defect looks like. Most production lines need both, wired together with care.
Microstops and the hidden factory your OEE number hides
A 40-second jam cleared by hand never reaches the downtime log, but forty of them in a shift add up to half an hour. Where microstops hide inside OEE, and how to make them visible.
How to write an inspection spec your integrator can't misread
'Detect all scratches' is not a specification; limit samples, numbers and an agreed acceptance test are. Here is how to write one, with a template outline.
Retrofit or replace? Adding AI vision to a 20-year-old line
On an old line the AI is rarely the hard part. The camera interface, the bracket, the encoder and the PLC usually decide whether a retrofit works.
Encoders, handshakes and reject gates: integrating vision with PLCs
A correct verdict applied to the wrong part is worse than no inspection at all. A field guide to triggers, encoder tracking, handshakes, fail-safe states and the timing budget that ties them together.
Lighting is the lens: choosing illumination geometry for machine vision
Most inspection problems are decided before the model sees a pixel. A practical guide to matching light geometry, wavelength and timing to the defect you need to find.
From pilot to fleet: why industrial AI pilots stall
The model worked. The pilot still went nowhere. Six reasons industrial AI pilots stall between the first line and the tenth, and how to design a pilot that is really a first deployment.
Five vibration spectra every maintenance manager should recognize
You don't need to be a certified analyst to read a vibration spectrum. Five patterns explain many of the faults you will meet on motors, fans and pumps, and each points to a different fix.
Seeing something on your line that no article explains?
Describe it to us. A 45-minute discovery call costs nothing, and we'll tell you plainly whether it's a problem we can help with.