Research was performed on the best approaches to tackle said algorithmic challenges and which algorithm architectures to use. Also, time was spent on the design, creation, testing and evaluating of prototypes for algorithms. Additionally, the prototyping infrastructure and pipelines used to acquire and process data were created, prototypes tested and deployed.
It turned out that random sampling with artificial neural networks, is effective at differentiating material that is known from materials that the algorithm does not know, which is considered a complex problem as such algorithms are generally deterministic and forced to choose between the available options.
Optimization of the above approach for dealing with unknown classes, such as minimizing false positives and false negatives, has been a challenge throughout the action. It has been found that there is no one setting that always delivers the best results hence it seems to be required to optimize settings per input, limiting the scalability of the technology.
After the described methods show which materials are unknown, clustering is used to map the properties of the material onto vectors which correlate to physiochemical properties of plastics, such as monomer, molecular weight, viscosity, tensile strength, and additive presence, all of which influence a plastic’s mechanical and rheological properties. Caveat is that a certain degree of uncertainty in the mapping of said properties remains, which is not easy to quantify.
The aforementioned clustering methods have also been used to train algorithms on existing datasets while omitting the data’s labelling, meaning the algorithm does not know up front which objects belong to the same classes, nor does it know which classes should exist.
The work performed as part of the action has produced some very useful and valuable results, the most noteworthy (and nonconfidential) of which are:
- A semi-selfsupervised learning algorithm has been created by enabling algorithms to show what they are uncertain about, which makes improving the algorithms a more well directed effort.
- The above has led the company’s algorithms to improve their classification accuracy from 98.7% by 99.5%, essentially reducing the errors made by 60%.
- Correlating certain rheological and mechanical properties of plastic objects to their near-infrared spectral reflectance data has been successful.
- A method has been created that is able to train algorithms without requiring upfront labelled data.
At the time of writing the action’s application, it was said to be expected that during the action, merely a prototype version of the algorithm was to be expected and that further development to reach a market ready state was to be expected for several years. Reflecting on this, it can be said that progress has been more swift than anticipated, as a part of the technology created during the action has already contributed to the improvement of market ready systems, while the remaining developments are likely to become ready for an introduction onto the market in the first half of 2021.
The results of the action are thus already being exploited in the field, by plastic recyclers, to improve their production processes by doing quality control with the technology. In the future, the technology is expected to be exploited in the form of sorting of plastic waste as well.
The results of the action shall be disseminated to a limited extent, due to their confidential nature. This very report is likely to be the only form of dissemination of the technical results. Technically speaking, however, is the process of exploiting the technology also a form of dissemination, whereas the technology interprets information, which is then redistributed to other technologies and/or people (plastic recyclers).