Predicting fire-induced spalling in concrete tunnel linings using machine learning techniques

Abstract

Fire-induced spalling is the phenomenon where the outer cracked or delaminated layer of a concrete element detaches due to the exposure to high temperatures during a fire. Spalling is a phenomenon that has raised concerns in the research community since the 19th century. Since then, many experimental, analytical, numerical, and other studies have been conducted around the world to explain this phenomenon. However, an accurate model to predict the occurrence of spalling remains elusive, particularly for tunnel linings. Tunnel fires have drawn increasing attention and raised more concerns in recent decades. The rapid growth of freight transportation, particularly flammable ones such as fuel, increases the potential to cause a rapid-fire spread. When compared to building fires, tunnel fires can be more destructive due to their high temperatures, quick heating rates, prolonged duration, and uneven temperature distribution inside the tunnel. Spalling is a complex phenomenon with a high degree of randomness that interdepends on too many factors. The occurrence of spalling phenomena is significantly influenced by various microstructural properties of concrete. Internal factors such as concrete permeability, moisture content, water-cement ratio, and aggregate type have a significant impact on spalling. Furthermore, temperature, heating rate, humidity, and loading conditions are some of the external factors that affect spalling. To gain a comprehensive understanding of this phenomenon, it is crucial to consider the interdependencies among these various factors and their combined effects. The current method used in industry to evaluate the performance of a concrete tunnel lining is to test the specimen in large-scale furnaces. However, this method has several limitations. It requires the use of large-scale furnaces, which is time-consuming, expensive, and difficult to replicate due to their dependence on specific concrete mixtures and test setups. Alternative approaches, such as Machine Learning (ML), can be considered to overcome these challenges. Recent advancements in data analytics & ML have demonstrated their capability to solve such complex problems. This study aims to create a framework for predicting fireinduced spalling in tunnel linings using several ML techniques. Python programming language was utilised to develop this framework and Jupyter Notebook was used as the web based interactive platform. Using the previously published fire test data, a new dataset was created, and after performing the appropriate preprocessing, it was fed into 10 distinct ML techniques. These includes 7 ensemble techniques and 3 traditional ML techniques. Then the developed model was further refined using hyperparameter tuning & kfold cross-validation techniques. The results of this model revealed that it is possible to forecast the occurrence of spalling with an accuracy of more than 90% using ensemble ML techniques.