Problem
Heavy chemical plants have a blind spot in the running process. A conveyor belt moves tons of granulated product per hour, but people take a small sample every four hours and send it to a lab. Problems that appear between samples go unnoticed, and the lab analysis is costly. Continuous analysis needs models that run on cheap CPU devices at the plant.
Example from the article: a granulated product. Devices on the belt photograph it and output a granule size distribution, cutting analysis time to five minutes.
Solution in brief
Mount small CPU-only devices on the belt, build the segmentation dataset with an active learning loop, and measure particles with instance segmentation. To read the process state fast, recast it as tagging: each frame gets a tag (big granules, a lot of dust, belt stopped) from a MobileNet encoder plus BERT, as in named entity recognition. The same tags can check a vision-language model's output.
Questions this case answers
- How to monitor a continuous process from images between periodic manual samples, using models that run on cheap CPU-only devices
- How to replace per-object masks with a cheaper state-tagging task when only the overall process state is needed
- How to monitor product quality on a conveyor belt continuously instead of by manual sampling
- How to build an annotated dataset for dense instance segmentation with an active learning loop
Tools and methods
- Instance segmentation (U-Net with MobileNet) — measures each particle; the baseline.
- Active learning with uncertainty — picks the images that need manual annotation.
- Semantic tagging as NER (MobileNet encoder plus BERT) — per-frame process state at high FPS.
- ClearML, CVAT, FiftyOne — experiment logging, annotation, dataset selection.
- ONNX, OpenVINO — model export for the devices.
Google or a one-time email code. No password.
