Predicting moisture pockets on a veneer drying line using machine learning
Nissinen, Aleksi (2025)
2025
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025120432520
https://urn.fi/URN:NBN:fi:amk-2025120432520
Tiivistelmä
In the veneer processing industry, veneer drying is one of the most important processes, but also one of the most energy intensive. Changes in the process have effects on multiple other phases after it, and incorrect drying results may require repeating the process. One common problem on drying lines are moisture pockets, spots of moisture remaining within the veneer after drying, which must be handled to prevent problems and possible loss of material down the line. Spotting them is not possible for a human, and industrial moisture analyzers are both costly and less effective on wet sheets.
In this thesis research was executed to find out if probable moisture pockets could be identified before drying from images taken of the pre-dried veneer sheets. The project was carried out by collecting and processing colour images of the same sheet before and after drying as well as the grid-like moisture maps measured from the dried sheets by an industrial analyzer. Then, a Fast RCNN model was trained to look for visual features within each grid cell after it was mapped to wet sheet pair and predict a moisture content percentage for it. Several models were trained as the configuration was iterated and improved on, until a final version was trained with the combined best improvements. Finally, data from each trained model was compared and analyzed.
The results showed that it is difficult for a neural network to identify the moisture pockets from visual data alone. The accuracy reached by the final model version was ~20.8% with an R² value of ~-0.7, indicating that the model could not generalize and only tried to learn the mean value of the data. Visual inspections of the results corroborated this conclusion, as the prediction errors seemed to be larger in both extremes of the scale.
The outcome reinforces the baseline that moisture pockets are difficult to detect from the outside of wet veneer. The model trained in this project could not reach the level of accuracy required for it to be usable in production, but the results did provide valuable insight into the problem and multiple ways how it could be improved, as well as other approaches that could be utilized with it.
In this thesis research was executed to find out if probable moisture pockets could be identified before drying from images taken of the pre-dried veneer sheets. The project was carried out by collecting and processing colour images of the same sheet before and after drying as well as the grid-like moisture maps measured from the dried sheets by an industrial analyzer. Then, a Fast RCNN model was trained to look for visual features within each grid cell after it was mapped to wet sheet pair and predict a moisture content percentage for it. Several models were trained as the configuration was iterated and improved on, until a final version was trained with the combined best improvements. Finally, data from each trained model was compared and analyzed.
The results showed that it is difficult for a neural network to identify the moisture pockets from visual data alone. The accuracy reached by the final model version was ~20.8% with an R² value of ~-0.7, indicating that the model could not generalize and only tried to learn the mean value of the data. Visual inspections of the results corroborated this conclusion, as the prediction errors seemed to be larger in both extremes of the scale.
The outcome reinforces the baseline that moisture pockets are difficult to detect from the outside of wet veneer. The model trained in this project could not reach the level of accuracy required for it to be usable in production, but the results did provide valuable insight into the problem and multiple ways how it could be improved, as well as other approaches that could be utilized with it.