Hållfasthetsklassning av konstruktionsvirke med maskininlärning utgående från datortomografidata från stockar

The wood industry’s importance for construction has been emphasized in recent years due to increased demand for environmentally sustainable materials. This, together with the ever-growing interest in using machine learning to solve practical problems, gave rise to this thesis. Several machine learning approaches were used to predict the strength of sawn structural timber without knowledge of the exact sawing pattern. This was made possible by a large dataset collected with a CT Log, an industrial computed tomography (CT) scanner, and a Goldeneye scanner calculating board strength, as modulus of rupture (MOR), in both green and dry stages. As well as by demonstrating that the strength of boards from the same log is highly correlated. Three machine learning architectures were explored: a fully connected neural network (FCNN), a one-dimensional convolutional neural network (1D CNN), and a transformer neural network. The FCNN and 1D CNN were also combined to predict the strength of pine, as these individual models showed the greatest promise. The results showed that the FCNN model was promising, particularly for pine strength grading. The 1D CNN demonstrated good performance, though not sufficient for industrial application. In contrast, the transformer model underperformed for both pine and spruce. Lastly, the combined FCNN and 1D CNN model outperformed the individual models, suggesting this hybrid technique should be investigated further. The FCNN model, produced in this thesis, is good enough to be considered for use in an industrial context. The combination of a FCNN and 1D CNN model was also shown to be promising and is presented together with insights regarding how this technique should be further developed.