Two Days on a Customer Production Line with Weiyi Intelligent Manufacturing's AI Inspection System

I spent two days trying out MicroIntelligence's AI smart inspection. "Trying" is a stretch, though — I can't exactly take this thing home. It's a whole setup: a few industrial cameras plus lighting, an edge-side inference box, and software running on the production line's industrial PC. I tagged along at a client site in East China that makes precision structural parts — they just brought this line online last month — and I followed the engineer from data labeling all the way to running batches.

First, let me be clear about what it is. MicroIntelligence does industrial vision inspection: cameras shoot the product, and the algorithm decides whether the image shows a good part or a defect. What sets it apart from traditional machine vision is that the old approach has engineers writing rules — how to compute edges, what grayscale difference, what area — all hardcoded one by one. This thing trains a model on defect samples, and the model learns what the features look like on its own. For defects with fuzzy boundaries like scratches, smudges, and color deviations, this route is way more reliable than writing rules.

The data labeling step is more labor-intensive than I expected

The first day was basically all data labeling. The process isn't complicated: shoot a batch of known-good and known-bad parts, then in the software draw boxes around the defect locations and tag them with categories. The interface isn't hard to use — image on the left, label tree on the right, scroll wheel to zoom, box-select and click a category to save. But after about two hours of labeling I started getting annoyed, because defects aren't stamped from a single mold. The same type of scratch — long, short, deep, shallow — you have to box them all. Box too few and the model doesn't learn the full picture.

The engineer told me they labeled for about two days the first time on this line, several hundred images. And every time they switch product models, they have to re-label everything, because the surface material changes and so do the reflection characteristics. This matches the complaints I've heard at other factories.

After labeling, training starts. From what I saw, training time is on the order of tens of minutes. You don't need to set up your own environment — just click in the software and it runs. When it's done you get an accuracy number on a validation set, plus a confusion matrix showing which two categories tend to get mixed up. I think this is more solid than a lot of similar products — it doesn't just give you a lonely accuracy number, it tells you where it's wrong.

Where it gets stuck when running batches

The second day we ran it on the production line. The pleasant surprise was that it keeps up with the cycle time — from camera capture to result, you basically can't perceive any delay by eye, so the line doesn't have to slow down for inspection. The false detection rate looked OK to me — obvious defects get caught, and the occasional over-detection clusters around reflections and oil stains, which has a lot to do with lighting angle. The engineer adjusted the light source position twice before it got better.

Two things got stuck. One is model changeover. After switching product models, the model doesn't adapt on its own — you have to re-label and re-train, and in the meantime the line either stops or goes manual. For high-mix low-volume lines, this switching cost is very real. The other is exception handling. The model flags a defect, but occasionally it hits an image it's unsure about, and then the system pushes a low-confidence flag to human review. The mechanism itself is right, but the review interface's operation path is a bit convoluted — workers have to go through two menu layers to see the original image. I asked the operator on site, and he said he kept clicking wrong during his first few days.

One more thing: this setup depends on imaging quality more than I expected. Camera, lens, lighting, fixture — if any link isn't right, the model is useless no matter how strong. It's not something you buy, plug in, and use. That optical tuning work up front is unavoidable.

Verdict

It suits teams with relatively fixed product models, clear defect types, and spare manpower on the line for initial labeling. In those scenarios it's less hassle than the old rule-writing approach, especially when defect shapes are hard to describe with rules.

It doesn't suit high-mix low-volume lines that switch models several times a day. The switching cost eats up the benefit. And teams expecting out-of-the-box use without optical tuning capability should stay away too.

It solves the problem at the model layer, but the bottleneck in inspection often isn't the model.