Deconstructability Prediction using Artificial Intelligence Models

Abstract

This thesis introduces artificial intelligence (AI) predictive modelling techniques for evaluating building deconstructability. It is the first research to create a deconstructability prediction model that includes variables from technical, economic, legal, social, environmental, and scheduling perspectives. The model uses advanced AI predictive models such as gradient boosting, artificial neural network, support vector machine, and random forest, among others, and can provide deconstructability prediction for different building types, including those designed for deconstruction (DfD) and those not designed for deconstruction, as well as BIM-compliant and non-BIM buildings, nearing or at the end of life. The research uses a positivist paradigm, focusing on objectivity and quantitative methods. The research employs a systematic literature review to identify variables influencing deconstructability. This review aids the development of a deconstructability construct-based conceptual framework, guiding the creation of a questionnaire. Deconstruction professionals, such as demolition engineers, civil/structural engineers and others with deconstruction expertise, complete this questionnaire based on a single past deconstruction project. After scrutinising all the returned questionnaires, 263 were discovered to be relevant. Since each professional responds based on a single past deconstruction project, each of the 263 questionnaires is assumed to represent a deconstruction project. These questionnaires help form two feature sets: all identified variables/features; the feature set is reduced to 22 variables using eight feature selection techniques. For consistency, the research experiments with and uses the two feature sets to develop twelve AI predictive models. The data is divided into 75% for training and 25% for testing across all feature sets. To address the imbalance in class, the research uses an oversampling technique (i.e., the synthetic minority over-sampling technique (SMOTE)) on the training data, ensuring a balanced representation of classes for model training across different predictive models. Additionally, the research employs a 5-fold Cross- Validation (CV) to rigorously assess each model’s performance. The research trains on the balanced training data and tests on the untouched 25% test set for all the AI predictive models. This provided robust and unbiased performance estimates. Importantly, this step ensured the oversampling process did not artificially inflate the models' performance metrics. The research finds that support vector machines with the polynomial kernel (SVM-P) using all features and Artificial neural networks with multilayer perceptron (MLP) using the features deducted from the FS techniques are the two high-performing models. Among the two, the SVM-P shows the highest predictive capabilities because of its higher accuracy and AUC, even as it uses all features. These findings made it known that, though researchers have proved the use of FS for enhancing predictive capabilities in AI predictive models, their uses and advantages may depend on the problem and scenario; as such, their uses may not apply to all kinds of issues/scenario where AI predictive model is used. Additionally, the predictive modelling performance of SVM-P suggests and supports the idea that deconstructability is a multifaceted concept. This is evidenced by the fact that the highest performance was achieved when all the diverse set of variables was used. A significant achievement of this research is the successful implementation of a generalisable and explainable AI-deconstructability predictive model that assesses building deconstructability. Another achievement is the establishment of various variables and perspectives, which provide a holistic view of the deconstructability of the building. Lastly, the AI-deconstructability predictive model developed in this research is the first AI-predictive model for deconstructability applicable to different types of buildings: DfD and non-DfD, as well as BIM-compliant and non-BIM buildings nearing or at the end of their useful life. The findings show that AI enables deconstructability decision-making for buildings near or at the end of their useful life, leading to innovative solutions for real-world challenges. The research implications are threefold: first, it enriches the knowledge base on AI applications for deconstruction, promoting collaboration between AI researchers and deconstruction professionals. Second, deconstruction professionals leverage AI's predictive capabilities to enhance decision-making processes, with potential applications extending to industries like manufacturing, thereby contributing to sustainability across multiple domains. Finally, the research further emphasises the need to explore AI model scalability, incorporating larger samples, diverse data sources and larger industrial validation by experts. Additionally, it suggests integrating emerging technologies such as IoT and capturing technologies to enhance real-time deconstructability predictions.