A PCA-assisted machine learning framework for non-destructive multi-strength prediction of structural bamboo (Dendrocalamus sericeus Munro)

Bamboo’s inherent material variability poses a major challenge for structural standardization compared with engineered materials such as concrete or steel. Our study evaluated the capability of multiple machine learning (ML) techniques to predict key mechanical properties, namely compressive, tensile and shear capacities, using non-destructive physical measurements, and to identify the principal variables governing variability through principal component analysis (PCA). Following experimental testing and data preprocessing, eight ML algorithms, including tree-based ensemble models and regularized linear regressions, were developed and comparatively assessed. Model performance was evaluated in terms of prediction accuracy and generalization capability. Regularized linear models outperformed ensemble approaches, demonstrating superior robustness against noise inherent in natural bamboo datasets. Subsequently, PCA was applied to investigate the multivariate structure of the data, with eigenvalues and eigenvectors revealing that density and linear mass were the dominant contributors to variance. Based on these findings, we propose a practical predictive framework for on-site quality assurance: Geometric measurements are first screened using PCA and Hotelling’s T² statistic to ensure consistency with the training domain, after which the trained ML models are employed for capacity prediction. This integrated methodology enhances reliability in assessing naturally variable bamboo, thereby improving its feasibility for structural applications.