Quality inspection of injection molded prototypes and production parts, e.g., light-switches. Reducing waste in the injection molding process can save time and money and make the molding process more efficient.
01. The problem
A European leader in the electrical components sector (e.g. Switches, socket outlets, …) was looking for a way to solve issues related to their current quality inspection process.
Their range features over 5,000 products that are being shipped all over Europe. Of their most popular product over 200,000 pieces a week are being produced. As this would take too much time to manually inspect these pieces, they decided to implement a vision based quality control system.
However, this system yielded too many false positives and false negatives translating into a lot of good pieces being thrown away and bad pieces that are let through.
OK / NOK
Unlike other anomaly detection methods training only on OK examples, the algorithm we used takes NOT OK examples into account to further tune the anomaly detection and creating a better tuned threshold when discerning OK and NOT OK examples.
This means our system can allow small defects, and still block the larger defects, making for a higher yield. In comparison, other methods training only on OK examples would be too strict, or too lenient, depending on where the user sets the threshold. Our method fixes this by focusing on creating a sharp boundary on what defects are and aren’t allowed.
In the figure on the right, we show how we are able to use the representations of the images learned by the neural network to make clusters in a 3D space.
Robovision’s platform has both improved the outcome of our quality control process and reduced the time to adapt the system for new products.
Even though we did not explicitly label the defects, this visualization technique (and underlying mathematical approach) allows us to locate the different occurring defects.
This approach would also be able to create estimations of what percentage of objects show a certain defect. Furthermore, adding time data, could provide insights into what defects started occurring at what time.
Robovision.AI’s success shows in its numbers
Through the use of our Robovision AI platform, they were able to reduce the number of bad pieces let through by 33% and good pieces being thrown away by 72%, translating into huge savings.
Next to realising savings on the number of disqualified pieces, our platform significantly reduces the time needed to set up a deep learning quality control system for new parts. Deep learning allows to switch product types in no time. This takes just a few clicks if you have a trained model in your library.
If not, it will take only a few hours to train a new model on a new SKU or product. Their old system took around 4 weeks to be set up, while our system only takes around 3 hours to learn in a new part.
An intelligent anomaly, or error, detection system that helps manufacturers detect fabric defects during the textile production process. The model alerts manufactures on defects and helps them adjust the process. Visit * for details.