Machine Learning and Metaheuristics within the Built Environment: From Design to Maintenance

Abstract

Despite significant technological development throughout academia and comparable sectors of the economy, our built environment is increasingly burdened by a lack of automation, data silos, disjointed workflows, slow technological adoption, and uninformed decision making. Each phase of the asset life cycle suffers from unique inefficiencies that have contributed to a growing gap between published research outcomes and the experience of practitioners who still rely on traditional approaches. Sensitivity to initialised hyperparameter sets, inherent implementation challenges, and the ‘a priori’ expertise required to effectively deploy metaheuristic algorithms have remained significant barriers impeding their broader adoption across practical design projects. While a broad variety of compelling machine learning models have been developed within the field of predictive 3D reconstruction, these solutions largely fail to meet the complexity and scale of modern design tasks. Analogously, while the academic field of structural health monitoring has transformed in recent years by adopting cutting-edge sensors and deep learning models, the industry itself has not been mobilised towards a similar level of technological adoption.

The contributions of this thesis are as follows. (1) The development of hybrid algorithms that utilise complementary aspects of metaheuristics and machine learning to reduce the sensitivity of algorithmic design tools to manually tuned hyperparameter sets. (2) The creation of novel methods to scale foundational 3D reconstruction models to meet the complexity of modern design tasks. (3) The demonstration of metaheuristics as viable tools to parametrically build training datasets of optimal structures for predictive design tasks. (4) An unprecedented improvement in prediction capacity, resolution, and inference speed for sensor-based asset monitoring systems, driven by the introduction of hypernetworks within the field of structural health monitoring.