Machine learning for defect condition rating of wall wooden columns in ancient buildings

This study proposes a Machine Learning-based classification method for rating the defect conditions of wooden columns in ancient building walls using Nondestructive Testing data, including defect images, micro-drilling resistance data, and material loss depths. Data processing algorithms were introduced to process the source data, fuse their features, and augment the datasets. Using traditional rating results as labels, Machine Learning classification models for rating the defect conditions of wooden columns were trained and evaluated. The results indicated that the accuracies of neural network models, including Convolutional Neural Network, Long Short-Term Memory, Radial Basis Function (RBF) neural network, and Extreme Learning Machine, exceeded 84.24 %, with Kappa coefficients greater than 0.79. All four models are suitable for the general assessment of defect conditions in wall wooden columns of ancient buildings. The RBF neural network model achieved the highest accuracy (94.57 %) on the feature fusion dataset, while Grey Wolf Optimizer (GWO) optimization further improved accuracy to 96.74 %. After re-splitting the training and test datasets, the rating accuracy of GWO-RBF neural network exceeded 95 %, with a Kappa coefficient of 0.94, demonstrating that the algorithms can effectively, conveniently, accurately, and stably assess the defect conditions of wall wooden columns. The automatic rating method lays the groundwork for developing a convenient onsite defect condition rating system for wall wooden columns in ancient buildings. • Defect images, micro-drilling resistance data, and material loss depths. • Data processing algorithms and dataset augmentation of wall wooden columns. • Machine Learning classification models performed well on the feature fusion dataset. • The GWO-RBF neural network model achieved a rating accuracy exceeding 95 %.