Mugumya, Kevin Luwemba
(2025)
A framework for knowledge representation learning-based building control.
PhD thesis, University of Nottingham.
Abstract
Current Building Automation Systems (BASs) have crucial context-awareness limitations that must be addressed before they can reach human-like levels and better adapt to the dynamic needs of modern buildings. Among other limitations, our buildings still lack sensors, actuators, and control agents that can learn reliable models of the environment and plan complex action sequences. Moreover, modern Machine Learning (ML)-backed BASs, though trained on massive datasets, are usually overly specialised (trained for one task) and brittle (prone to errors). In contrast, human learning is very efficient, and with only a few examples, we can find intuitive ways to complete a task while generalising our knowledge to other tasks. To address the above limitations, this thesis proposes a foundational framework that aims to advance the context-awareness capabilities of BASs using knowledge graphs and Knowledge Representation Learning (KRL). At the framework’s core is the notion of using Semantic Web Technologies (SWT) to model the semantic relationships between different building components. These relationships are then packaged inside a network-like data structure called a Building Information Modeling (BIM)-based Knowledge Graph (BIM-KG), and KRL is applied to learn the hidden patterns within the BIM-KG. During the learning phase, KRL utilises message-passing to propagate the learnt information throughout all nodes/entities in the BIM-KG. This research hypothesises that building automation agents can leverage this notion of message-passing to aggregate contextual information from all entities in the graph and use it to continuously update their understanding of a building’s systems and components. The perception is that imbuing building automation agents with holistic information about the buildings they control can presumably support context-aware decision-making during downstream automation tasks.
To test the research hypothesis, a three-phase investigation was carried out: literature review, framework development, and framework applicability. Phase one focused on situating the research within the scholarly discourse of BIM, BIM-KGs, building automation, and KRL. The results show that since 2010, SWTs have been a driving force advancing BIM research in the Architecture, Engineering, Construction and Facility Management (AEC/FM) fields by providing the mechanics to represent complex relationships within the built environment. Concurrently, KRL has seen significant development in domains such as bioinformatics, where it has been used to understand complex biological relationships and processes. However, despite the apparent suitability of applying KRL to the BIM field, such integration has not materialised and remains largely unexplored. To get around these research shortcomings, the next phase of this thesis was to develop a framework for applying KRL to BIM-KGs using performance analysis experiments. Five baseline KRL models were chosen for this. The chosen models are well-regarded techniques from existing studies, cover a wide range of methodologies, and have been extensively investigated in the context of drug discovery, whose data structures closely mirror those of BIM-KGs. Two publicly available BIM-KGs datasets were used in these experiments. The overall goal was not to identify the best KRL model configurations. Instead, the study examined more closely how model performance can be affected by modifications to the training step, selection of hyperparameters and their optimisation. The experimental results were used to define the prerequisites for integrating KRL with BIM-KGs in a domain-independent framework. This means that although a building automation use case is used to formulate the framework, it can assumingly be applied to other AEC/FM domains such as heritage, quantity-takeoff and energy analysis. The experimental findings show that RotatE and TransE consistently outperform other models across both datasets, establishing themselves as robust baselines when integrating KRL with BIM-KGs. It is also interesting to see that older models like TransE can still be competitive with optimised training and Hyper-parameter Optimization (HPO) configurations. Adam and NSSA emerged as favourable training setup choices, suggesting their potential as initial benchmarks for future evaluations. Despite extensive hyperparameter searches, there was considerable variance among top-performing model configurations, indicating the need for nuanced parameter combinations. This complexity suggests that manual tuning may not yield optimal results, advocating for the adoption of HPO strategies. Furthermore, the disparity in hyperparameters between the two datasets underscores the influence of dataset-specific parameters. Finally, random search methods, when repeated sufficiently, yield configurations closely comparable to more systematic approaches, albeit in less time.
To illustrate the applicability of the framework, phase three lays out a high-level system architecture consisting of a BIM model, Internet of Things (IoT) devices, and a prototype program of the framework wrapped inside an Application Programming Interface (API). The API consists of a server-side module and a client-side module. The server-side module demonstrates how a building automation system can communicate with KRL configurators, external services such as BIM-KG databases, sensor data stores, and Message Queuing Telemetry Transport (MQTT) brokers. The client-side module consists of a Graphical User Interface (GUI) with a Construction Operations Building Information Exchange (COBie) handler service that facilitates the curation of BIM-KGs from COBie files and an interrogation service that facilitates declarative interrogation of the server-side module using SPARQL Protocol and RDF Query Language (SPARQL) and Graph Query Language (GraphQL).
In conclusion, for KRL to impact the AEC/FM domain, this work emphasizes the critical importance of comprehensively reporting model architectures, training setups, and hyperparameters to enhance trust, reproducibility, and understanding of KRL-based methods among AEC/FM stakeholders and researchers. This insight highlights a prevalent issue in the AEC/FM field where results are often difficult to replicate due to incomplete documentation.
| Item Type: |
Thesis (University of Nottingham only)
(PhD)
|
| Supervisors: |
Wong, Jing-Ying
Chan, Andy Tak Yee |
| Keywords: |
building information modelling; knowledge representation; knowledge engineering; data representation; knowledge representation learning |
| Subjects: |
T Technology > TH Building construction |
| Faculties/Schools: |
University of Nottingham, Malaysia > Faculty of Science and Engineering — Engineering > Department of Civil Engineering |
| Related URLs: |
|
| Item ID: |
81266 |
| Depositing User: |
MUGUMYA, Dr Kevin
|
| Date Deposited: |
26 Jul 2025 04:40 |
| Last Modified: |
26 Jul 2025 04:40 |
| URI: |
https://eprints.nottingham.ac.uk/id/eprint/81266 |
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
