Institute of Solid State Physics
|
|
SS23 WS23 SS24 WS24 SS25 Guidelines for Master Students
Hyperparameter optimisation of a neural network for classification optimisation of waste paper in near-infrared spectroscopy As circular economy gains importance in the face of factors such as global warming or limited natural resources, finding ways to re-use or recycle our everyday products becomes a larger factor in many industries. The sorting of wastepaper to separate recyclable from non-recyclable materials is done via ballistic separators and near-infrared spectroscopy. However, experience shows that the sorting of cellulose-containing materials such as paper or wood using NIR spectroscopy gets increasingly more difficult with increasing humidity and sample wetness. A suitable machine learning (ML) model applied to wastepaper data may be able to achieve the classification results needed for improved recycling. As the optimization off the parameters and architecture of neural networks is a lengthy and complicated process and requires extensive knowledge of neural networks when done manually, libraries to automate the process have been devised. This thesis examines an automated optimization of neural networks to optimize the wastepaper sorting results. Ideally, based on three different sorting applications, a part of the neural network may be found that can be fixed for future applications to be able to further reduce computational efforts. |
