A method for rating static stability conditions of ancient masonry pagodas based on a defect dataset and machine learning

Rating static stability conditions of ancient masonry pagodas is the foremost problem to be addressed in implementing preventive conservation measures. Information on 153 ancient masonry pagodas is collected. Key features affecting static stability conditions of ancient masonry pagodas are identified as four types of defects. The criteria for feature grading and assignment have been developed. A defect dataset, composed of the normalised values of the features and the qualitative outcomes from cultural heritage authorities, is created. A method for rating the static stability conditions of ancient masonry pagodas classified into three categories is proposed based on the defect dataset and machine learning algorithms. Three models are trained using the K-means clustering, ordinal logistic regression, and multilayer perceptron (MLP) neural network algorithms, whose performance is assessed and compared. The proposed method is applied to rate the static stability conditions of four ancient masonry pagodas outside the dataset. The results indicate that the accuracy of the MLP neural network model is slightly higher than that of the ordinal logistic regression model and much higher than that of the K-means clustering model. The most conservative result from all evaluation models is suggested to be the final category for a pagoda outside the dataset.