Production-line vision for U-Scan

U-Scan is Withings' at-home urine analysis device. It reads test strips inside a cartridge, and every cartridge goes past a vision station on the production line.

The vision station

I extended the line's Raspberry Pi vision station with new checks in Python and OpenCV, to catch cartridge defects the existing station missed:

• YCbCr channel tests on the cartridge images.
• A bubble check for the integrity of the cartridge seal.

Preproduction in China

I flew to the manufacturing site in China, alone, for the preproduction run, which lasts about a week. I worked with engineers I had only met on video calls, and few people on the line spoke English or French.

• Before production started, I went through every station to anticipate failures.
• Once the line was running, I kept a schedule for improving the stations not in use that day, and reserved time every day for anything that could block production.
• I diagnosed one station's failing script on the spot: a Python package issue.
• The new checks caught defect types nobody had seen in the lab.

It taught me to design with how something will be built and tested at scale in mind.

Back in Paris

After the trip, I added a TensorFlow Lite autoencoder for anomalous strips to the Pi station.

I also built a Pi-controlled jig for battery-life stress testing. It simulated daily device use with a pump and measured current draw.