Hiding in plain sight: What can interpretable unsupervised machine learning and clustering analysis tell us about the fire behavior of reinforced concrete columns?

Abstract

The role of machine learning (ML) continues to rise in the structural fire engineering area. Noting the widespread of supervised ML approaches, such methods are being heavily utilized nowadays. On the other hand, little interest has been dedicated to unsupervised ML. Unlike supervised learning, unsupervised learning algorithms are trained using data that is neither classified nor labeled, thus, allowing the algorithm, and without guidance, to identify unseen patterns and decode hidden structures of complex phenomena residing in data. In this pursuit, this study presents findings obtained from one of the earliest investigations aimed to explore the potential of unsupervised and interpretable machine learning (ML) clustering analysis to investigate the response of reinforced concrete (RC) columns under fire conditions. We used four algorithms (namely, K-Means, Hierarchical, Fuzzy C-Means, and DBSCAN) to cluster over 140 fire-exposed RC columns. Results from our clustering analysis show that the performance of such columns can be distinctly grouped into unique clusters. Our findings allow structural fire engineers to 1) identify and avoid RC columns with poor fire performance and, when possible, 2) upgrade the features/characteristics of columns of poor fire response to achieve an improved behavior. Overall, our analysis indicates that fire-exposed RC columns naturally fit into four clusters – each with unique properties and response – that are governed by the material, geometric and loading features.