Digital twin and AI enabled predictive maintenance in building industry

• Main Author: Hu, Wei
• Other Authors: Cai Yiyu
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Digital twin; Artificial intelligence; Predictive maintenance; Building industry
• Online Access: https://hdl.handle.net/10356/180365
• Institution: Nanyang Technological University

The rapid advancement of information and communication technologies (ICT) and artificial intelligence (AI) has catalysed a significant shift in maintenance practices within the building industry, paving the way for a data-driven paradigm. Predictive maintenance (PdM) has emerged as a critical approach to anticipating failures and reducing unscheduled maintenance tasks. However, the surge in ICT implementations within building-related infrastructure presents several challenges for PdM research and development. Current frameworks are often constrained to specific facilities and lack scalability and generality. Additionally, existing studies focus on condition monitoring and fault detection, with insufficient attention to failure prediction. The effectiveness of PdM has also been hindered by the dependence on labelled datasets, which are expensive and time-consuming to generate. Furthermore, indoor climate management is crucial to building performance. However, it has received less attention than facilities maintenance in PdM research despite its integral role in occupant comfort and environmental sustainability. In response to these challenges, this thesis introduces a unified framework that integrates Industry 4.0 technologies within a digital twin (DT) structure, grounded in the innovative Six M methodology—Machine, Manpower, Material, Measurement, Milieu, and Method. This approach emphasises the entire building lifecycle, enabling stakeholders to optimise operations, resource allocation, and decision-making across multiple facets of building management. The 6M methodology is crucial for transforming PdM by providing a structured and holistic approach to integrating diverse building assets and operational processes within the DT environment, thereby enhancing system scalability and operational efficiency. The research also employs a Three-by-Three M analysis methodology and a keywords network analysis to identify key research clusters and critical factors in existing DT-enabled PdM-related studies. This analysis underscores the transformative potential of DTs in revolutionising PdM applications across the building industry. These fundamental studies pave the way for future PdM applications in the building and construction (B&C) industry, significantly advancing maintenance strategies. Following a comprehensive investigation, this thesis introduces a pioneering failure prediction methodology that addresses the challenge of limited labelled datasets by leveraging semi-supervised generative adversarial networks (GANs). This innovative approach enables the model to utilise labelled and unlabelled data, reducing the reliance on costly manual labelling while improving prediction accuracy. Based on publicly available datasets from building facilities, empirical results demonstrate the model's superior performance in predicting failures, enhancing the system’s proactive maintenance capabilities. An online platform was also developed to integrate real-time monitoring with predictive alarms, allowing for efficient, data-driven decision-making in building maintenance. In addition to traditional facilities maintenance, this thesis extends PdM applications to include indoor climate management, addressing the gap in existing research. Indoor climate, particularly air quality, directly impacts occupant well-being, comfort, and productivity and can be an essential indicator of system failures or inefficiencies. Proper PdM ensures optimal air quality by predicting and preventing system malfunctions that could lead to poor ventilation, temperature control, or humidity levels. The proposed framework incorporates remaining useful life (RUL) and time shift (TS) methods, dividing the prediction tasks into supervised and unsupervised subtasks. A parallel prediction model, combining long short-term memory (LSTM) networks and autoencoders (AE), is developed to handle this complex task. An innovative DT-enabled PdM framework for indoor climates is validated through an online platform that reconstructs 3D building models and provides real-time monitoring and alerts. Experimental results demonstrate the framework’s ability to accurately predict faults at varying warning times and severities using practical datasets from buildings in Singapore. In conclusion, this thesis explores the integration of DT and PdM in the building industry through a comprehensive analysis of four academic papers, leading to significant contributions in system scalability, efficiency, and sustainability. The proposed methodologies enhance building maintenance practices and extend PdM’s impact to encompass critical indoor climate factors, paving the way for more resilient, cost-effective, and sustainable building operations.