Wood is an excellent building material, due to its outstanding weight-performance characteristics, its sustainable availability and its appearance, which is generally perceived as very pleasant. Hence, wood has been widely used as main building component for various engineering structures. Nevertheless, it could be used more efficiently and in more demanding applications, but the inherent heterogeneous material structure at different length scales results in a high variability of wood properties. This variability calls for large safety factors limiting material utilization efficiency. In order to fully exploit the potential of wood, wood-based products, and timber building components, the challenge of the large variability of wood material properties and its influence on mechanical behaviour needs to be addressed. Of crucial relevance for increasing the reliability of wood and the efficiency in its use in engineering applications is the development of improved processing technologies, e.g. improved machine grading of wood into more defined and higher strength classes.

Wood grading has already gone through optimization steps, as next to the established, yet inef-ficient visual inspection, machine grading systems have been developed. Visual inspection includes the detection of e.g. knots, fibre orientation or wood damage mainly in a qualitative, partly in a semi-quantitative way. With the development of machine grading systems, more parameters, including mechanical properties, can be measured at higher speed, with greater precision and in a quantitative manner. There-fore, we can now speak of big data continuously produced by the wood industry.

The grade of a wood sample is strongly related to its strength, which can be determined exactly only by a destructive test, in which case we can label the samples according to their strength. On the other hand, machine grading of wood involves non-destructive tests, so that the strength can only be predicted or estimated via a model. In those cases, we cannot know the true strength of each sample, so they remain unlabelled. Despite the availability of big data, the predictive methods currently used for grading are based on very simple linear regression models, which generally show poor results.

Our project aims at significantly improving the technique for estimating the strength of wood by using more modern tools for processing the available big data. Our approach relies on combining labelled and unlabelled data under the semi-supervised learning paradigm to develop and train a deep neural network that will be optimized for this task.

We present a proof of concept that machine learning can be used to predict the tensile shear strength of bonded beech wood as a function of the composition of polyurethane prepolymers and various pretreatments. A comprehensive experimental data set was used for training and testing the algorithms support vector machines (svm), random forest (RF), and artificial neural networks (ANN). Within the framework of the experiments, the structure-property relationships of 1C PUR prepolymers were analyzed by systematical variation of the structural parameters urea and urethane group content, cross-link density, ethylene oxide content, and the functionality via isocyanate (NCO) or polyether component.

The bonded wood joints were tested according to DIN EN 302-1 (2004). Prior to testing, the shear-test-specimens were pre-treated according to procedures A1 and A4, five temperature steps (5, 40, 70, 150 and 200°C) and two alternating climates. The complete data set (N=2840) was pre-processed and split into a training set and a test set using 10-fold cross-validation. The performance of the algorithms was evaluated with the correlation coefficient (R), the coefficient of determination (R²), root mean square error (rmse) and mean absolute percentage error (MAPE). All machine learning algorithms revealed a high accuracy, but the artificial neural network showed the best performance with R=0.96, R²=0.92, rmse=0.948 and a MAPE of 9.21. Our work paves the way for future machine learning applications in the field of adhesive bonding technology and may enable a fast and effective development of new adhesives and enhance the efficiency of their application.

In this case study, machine and process variables were extracted from the process control system (Prod-IQ) and combined with tested mechanical properties of wood fiber insulation boards according to product type and time of manufacture. The boards were taken from the production line and the internal bond strength (σ_mt) and the compressive strength at 10% deformation (σ₁₀) were determined according to the European Standard EN 826 and 1607. The complete data set was preprocessed and split into training and test sets using k-fold cross-validation. The performance of the random forest algorithm (RF) was evaluated with the correlation coefficient (R), the coefficient of determination (R²), root mean square error (RMSE) and mean absolute percentage error (MAPE) and compared with artificial neural networks (ANN), and support vector machines (SVM). Forward feature selection was used to reduce input dimensionality and improve the generalizability of the algorithms. All machine learning algorithms predicted the mechanical properties with high accuracy, but the RF algorithm revealed the best generalization performance (σ_mt: R=0.960, R²=0.916, RMSE=4.05, MAPE=12.11; σ₁₀: R=0.981, R²=0.963, RMSE=17.19, MAPE=5.64). Our work demonstrates that machine learning can be applied to predict relevant properties of wood fiber boards for an improved quality control in real time.