Redes neurais convolucionais na avaliação de juntas coladas com resorcinol-formaldeído: comparação com a norma ASTM D5266-13

Abstract

This study evaluated the accuracy of a convolutional neural network trained to estimate the wood failure rate in glued joints bonded with resorcinol-formaldehyde adhesive, using image semantic segmentation techniques. The results obtained by artificial intelligence were compared with visually estimated values, in accordance with ASTM D5266-13 (2020). Statistical analysis included a paired t-test, Pearson correlation, and linear regression. The average wood failure rate estimated by the convolutional neural network was 89.09%, statistically lower than the visual evaluation (92.08%), although both were above the minimum normative threshold. The strong correlation between the methods (significant r, p < 0.05) and the high adjusted coefficient of determination indicate that artificial intelligence was effective in reproducing the trend of variation in failure rates. These results demonstrate the feasibility of applying convolutional neural networks in the quality control of engineered wood products, promoting greater objectivity, speed, and reliability in assessments. The integration of traditional methods and artificial intelligence algorithms represents a significant advancement for the forestry sector in the context of Industry 4.0