The project started unexpectedly when our lecturer Ago Rootsi asked in the corridor if we would be interested in inventing a machine for a company. They were looking for a solution to detect defects in knitting machines, which sounded like an exciting challenge to us students. The team included me, Gregor Kokk, Ott-Kaarel Vään, Ago Rootsi, Kädi Veeroja, and Taavi Kase.

The first challenge was finding funding. We narrowly missed funding from EIS, and the project could have ended there, but a few months later we heard about the AIRE program. Since our original plan was to use cameras and machine vision, we decided to search for our project, try again once more, and WE SUCCEEDED!

The work was divided into three main stages. The first step was to clarify what exactly our solution needed to detect in the factory. It’s easy to say it has to find "defects," but how do you explain to a computer using a camera image what is a defect and what is not? To do this, we built a small model of a knitting machine at school, which allowed us to test cameras, experiment with different lighting conditions, and make quick adjustments. Working on the model gave us the opportunity to test and find solutions without having to stop the entire production line.

Once the initial architecture was set, we took our solution to the factory and mounted the cameras on an actual knitting machine. Our goal was to collect images as early as possible from the same environment where the system would eventually operate, because data obtained under real working conditions reduced the risk that the final solution would need later adjustments or fail to work as expected. We found that lighting and shadows could significantly affect image recognition accuracy and adjusted the devices accordingly. Additionally, we were able to create defects ourselves in the factory, such as broken threads or knitting needles. This gave us a better understanding of how defects look in real production and how our system should detect them. The data collected during these tests formed the basis for training the model.

In the final stage, our machine learning specialist Gregor Kokk focused on developing the machine vision model. We used hundreds of image examples of both good and defective knits to teach the model to detect different types of defects. Training the model required constant parameter adjustments and testing to achieve the best possible accuracy. We had to learn a lot on the fly because many changes occurred that were impossible to foresee in the classroom.

By the end of the project, we had built our machine learning computer that monitored production lines in the factory using two cameras and alerted employees if the machine detected something wrong. Defect detection was made especially convenient for employees because the screen also showed on which conveyor and exactly where the defect was detected.

This project was not just about developing technology – it was an opportunity to learn how to turn ideas into reality. Our team gained invaluable experience in managing a complex project from start to finish. Collaboration skills, from dividing roles within the team to communicating with company representatives, helped us understand how the real industrial world operates. Besides experience, the project opened new career opportunities, as practical experience gives much more confidence and assurance in applying one’s skills.

Projects like these show how important practical experience is for students. It is a real leap from theory to practice, where learning becomes meaningful, and it proves that collaboration between universities and companies can create solutions that benefit both students and the industry as a whole. 

The article was published in the Tallinn University of Technology magazine Mente et Manu.