Machine learning-based non-destructive testing model for high precision and stable evaluation of mechanical properties in bamboo-wood composites

Abstract

The efficient evaluation of mechanical performance of bamboo-wood composites (BWCs) is an important part for their development and application. To address the issues of low efficiency, high consumables usage, and low accuracy in traditional BWC mechanical performance testing, a non-destructive testing method for BWC mechanical performance was proposed based on machine learning. First, the images of the cross-section and longitudinal sections of the BWCs were collected. Then, a UNet-based image-segmentation model was used to segment the bamboo, wood, and holes in the cross-section. Additionally, the image features, including texture, frequency, and spatial characteristics of the BWC were extracted using the gray-level co-occurrence matrix (GLCM), db wavelet, Fast Fourier Transform (FFT), and Gabor filtering. Finally, the results of image segmentation and feature extraction served as inputs, and the corresponding mechanical performance parameters as outputs to build the dataset that informs the artificial neural networks (ANNs) model predicting the mechanical properties of BWCs. The research results show that the accuracy, mean intersection-over-union (MIoU), and Kappa coefficient of the image segmentation model are 0.9586, 0.8242, and 0.9125, respectively. In predicting the elastic modulus (MOE) and static bending strength (MOR) of the BWC using ANNs, the coefficient of determination (R) values were found to be 0.85 and 0.89, respectively. Besides, the mean absolute percentage error (MAPE) of the ANNs was 11.6% and 7.4% for MOE and MOR, respectively. These results indicate that machine learning methods demonstrated superior precision, accuracy, and stability for predicting the mechanical properties of BWCs.