Reconstructing the complete force-deformation response envelope for unreinforced masonry walls using self-supervised learning

Abstract

This article introduces an interpretable data-driven approach to reconstructing the complete force-deformation response envelope (backbone curve) of unreinforced masonry (URM) walls using the structural properties as inputs. The framework utilizes supervised and unsupervised learning combined with autoencoder neural networks. The dataset used to develop the predictive model includes crack texture images, design, and mechanical properties, as well as the backbone curves associated with 165 brick and block URM walls subjected to quasi-static cyclic loading. Most of the walls in the experimental database experienced hybrid failure modes. Therefore, autoencoder networks are used to extract informative features from the backbone curves, which are incorporated into an unsupervised learning model to identify potential failure modes. A representative backbone curve (or baseline backbone) is then formulated for each failure mode. The recorded force-deformation response for each test specimen is then decomposed into a combination of the scaled baseline backbones. Finally, the supervised learning model is used to establish a relationship between the participation weights of each baseline backbone (the output) and the wall properties (the input). Manual labeling is also conducted to validate the results obtained from the clustering models, ensuring that the most likely failure mode is assigned to each URM wall. Unlike conventional codes and standards which estimate specific points along the backbone curve, the proposed model reconstructs the complete backbone curve given the mechanical and design properties of the wall with an accuracy level that exceeds 90% in both the test and train datasets.