Cerrion helps manufacturers automatically detect problems on their production lines using computer vision. You can see this in action here, for detecting issues on conveyor belts and in glass bottle production: https://youtu.be/DuSN-qJcoNQ

It’s estimated that undetected problems on production lines cost the manufacturing industry $1 trillion in lost production time per year. This is because staying on top of your production is hard and works best with trained and experienced eyes. We are working on making this easier, by automating production line monitoring with computer vision.

The basic idea is simple: our product learns how a manufacturing process looks when things are going well, then can detect and track anomalies and other problems in real time.

This has several major benefits: (1) it allows detecting subtle issues on the production line before they become big and costly, (2) it reduces the need for human monitoring, and (3) it facilitates root cause analysis remotely in a matter of minutes by showing video data of the problem(s).

We came to work on this because Nikolay previously co-founded Assaia (https://assaia.com/) and we learnt how messy and intransparent ground operations at airports are. We quickly realized that manufacturing companies suffer from similar pain points, given that manufacturing processes are highly complex. Thus, we started talking to manufacturing companies and soon recognized that computer vision could significantly increase their process transparency and thus help them better run their production lines.

We have built a video analysis pipeline using a dockerized Python stack. The pipeline processes RTSP video streams and analyzes them in real-time, using a Convolutional Neural Network, making predictions for what goes wrong where in the production process. We aggregate these predictions into events, push them to a Kafka queue and serve them back to the customer. We do this via a real-time alerting system and a detection library. The real-time alerting allows customers to take actions immediately. The detection library offers an analytics dashboard, as well as videos of the relevant problems. With this, our customers can find systematic production issues and do root-cause analysis.

The manufacturing landscape is heterogeneous and production processes are constantly changing. To be able to serve all kinds of industries, we need an adaptable product. To get there, we are working hard to make our product plug-and-play—essentially, to get to the point where it fully automatically learns how a manufacturing process looks when things are going well and automatically detect deviations based on this. In practice this means we need to build a performant model using transfer learning and self-supervision; and automatically adapt and keep it up to date from just a handful of user inputs (for which we use active learning).

BTW, we pay our bills by charging a SaaS license per production line.

Thanks for reading! We are curious to hear your thoughts!