Deep Learning Methods for Wood Composites Failure Predication

Abstract

For glulam bonding performance assessment, the traditional method of manually measuring the wood failure percentage (WFP) is insufficient. In this paper, we developed a rapid assessment approach to predicate the WFP based on deep-learning (DL) techniques. bamboo/Larch laminated wood composites bonded with either phenolic resin (PF) or methylene diphenyl diisocyanate (MDI) were used for this sample analysis. Scanning of bamboo/larch laminated wood composites that have completed shear failure tests using an electronic scanner allows a digital image of the failure surface to be obtained, and this image is used in the training process of a deep convolutional neural networks (DCNNs).The result shows that the DL technique can predict the accurately localized failures of wood composites. The findings further indicate that the UNet model has the highest values of MIou, Accuracy, and F1 with 98.87%, 97.13%, and 94.88, respectively, compared to the values predicted by the PSPNet and DeepLab_v3+ models for wood composite failure predication. In addition, the test conditions of the materials, adhesives, and loadings affect the predication accuracy, and the optimal conditions were identified. The predicted value from training images assessed by DL techniques with the optimal conditions is 4.3%, which is the same as the experimental value measured through the traditional manual method. Overall, this advanced DL method could significantly facilitate the quality identification process of the wood composites, particularly in terms of measurement accuracy, speed, and stability, through the UNet model.