Reflection
Solder and metal surfaces reflect the fixed lights. I looked for cases where glare changed the result.
Industrial AI Vision Inspection
A summer internship at Solomon Technology. My main project was a PCB inspection demo for Automation Taipei 2024. I set up the camera and lighting, collected and labeled about 200 board images, and trained the inspection models in Solomon’s vision platform. I also maintained the live demo in the company showroom.
- Company
- Solomon Technology
- Role
- Field Application Engineer Intern
- Period
- July–August 2024
- Work
- Automation Taipei 2024 PCB demo; showroom live demo
- Scope
- Imaging setup, labeling, model training in the company platform, testing
Problem
A soldered PCB can be wrong in a few ways. A part can sit in the wrong position, the wrong part can be fitted, or a solder joint can be bad. Some of these are easy for a camera to see. Others depend on small details, like the color bands that give a resistor its value.
How the board is imaged matters as much as the model. Solder and metal reflect light, taller parts cast shadows, and a board that sits slightly off position looks different to the camera.
My supervisor assigned work project by project. The main one was a PCB inspection demo for Automation Taipei 2024, an industrial automation trade show. The other was keeping the company’s showroom demo working well.
My role and contributions
There were two pieces of work. The expo project was mine from setup to demonstration. In the showroom demo, my job was maintenance, testing, and retraining. Both were on the application side of Solomon’s vision platform.
Expo project (mine)
- Set up a fixed camera and lighting rig for imaging soldered PCBs.
- Collected and labeled about 200 PCB images.
- Trained and tuned the inspection models in Solomon’s vision platform. They checked component placement, wrong components (for example, a resistor’s value from its color bands), and solder condition.
- Tested the models against reflection, shadow, and changes in board position, and reported what failed to the AI engineering team.
- Demonstrated the system at Automation Taipei 2024.
Showroom demo
- Maintained and tested the inspection software used in the showroom’s live demo.
- Retrained the whole model to improve the live results.
Worked within Solomon’s vision platform, turning its detection tools into a working PCB inspection system.
System
The expo setup kept the camera and the lights fixed. That way, a change in the result could be traced to the board and not to the capture.
Soldered PCB
Fixed camera and lighting
Models trained in Solomon’s platform
OK / NG for each check
Review and report
Placement
Is each part in the position it should be?
Wrong part
Is the right part fitted? For resistors, this means reading the value from the color bands.
Solder
Does each solder joint look properly formed?
| Expo project | Showroom demo | |
|---|---|---|
| Purpose | PCB inspection demo for Automation Taipei 2024 | Live demo for showroom visitors |
| Camera | Fixed camera, fixed lighting | Tablet camera, held by hand at any angle, live video |
| My part | Rig, images and labels, model training, testing, demonstration | Maintenance, testing, retraining the whole model |
The showroom demo has to cope with more variation than the expo rig, because the camera moves and the viewing angle changes.
Validation
I tested the expo models by changing what the camera sees, then reported what failed to the AI engineering team.
Shadow
Taller components cast shadows on the parts next to them. I checked how those shadows affected the checks.
Position
I moved the board around in the frame to see whether each check still found the right part.
Both videos are internal test recordings from the expo project, not footage from the trade show.
Outcome and limitations
The expo project went from camera setup to a working demonstration at Automation Taipei 2024, and my test findings went to the AI engineering team. In the showroom, I retrained the live demo’s model and kept the demo running.
I presented the demo to visitors at the expo, and Solomon’s chairman showcased it to Taiwan’s Vice President.
Judged by eye
The showroom demo inspects live video, so the retrained model was judged by watching it work, not by counting results per trial.
One rig, about 200 images
The expo models were trained on about 200 images from one fixed setup, built for a live trade-show demonstration.