| Author: | Zeng, Yanfu |
| Title: | Intelligent building fire safety design driven by deep learning methods |
| Advisors: | Huang, Xinyan (BEEE) Fu, Xiao (BEEE) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Subject: | Building -- Design and construction Fire prevention Fire protection engineering Machine learning Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Building Environment and Energy Engineering |
| Pages: | ix, 97 pages : color illustrations |
| Language: | English |
| Abstract: | Buildings are supposed to be carefully designed in terms of fire safety to safeguard the lives and properties under the occurrence of fire incidents. However, current fire safety design process suffers from time-consuming and computational-expensive analysis of fire-smoke behaviours, as well as intensive workloads and inevitable human errors due to manual preparation of the engineering drawings. This thesis aims to automate the design process by introducing deep learning engines that can complete design tasks including fire engineering analysis and drawing preparation. Results show that the pre-trained deep learning models can learn the complex dynamic relations of fire-smoke behaviour from numerical data and thus provide real-time engineering assessment of building fire safety. They can also capture design patterns from historical drawings and then generate sprinkler layout designs comparable to engineers within seconds. Compared to the traditional simulation-based analysis and manual drawing preparation, deep learning methods can offer significant design efficiency improvement and cost reduction, encouraging more cost-effective solutions for building fire safety. This thesis is presented in a manuscript style: the first chapter introduces the research background and motivations. The following chapters take the form of independent papers, which have been published or submitted to a journal publication. The final chapter summarizes the conclusions and points out the potential areas of future research. Chapter 1 introduces the research background and motivations. Emerging innovative architectural designs and new material inventions continue challenging building fire safety and calling for new design approaches. Today, fire engineers play a core while iterative role during the whole design process, including drawing, auditing, reviewing, and approving. Although computer-aided tools can help conduct engineering analysis such as smoke management and evacuation, the current design process is very time-consuming with many repetitive works and inevitable human errors. Consequently, there is a pressing necessity for smart innovation in design methodologies. In Chapter 2, the Intelligent Fire Engineering Tool (IFETool) is introduced to expedite fire engineering analyses and swiftly pinpoint design constraints. An extensive dataset encompassing 1,080 fire simulations conducted in a basic cubic atrium was compiled, taking into account crucial parameters like building dimensions, fire positioning, severity, ventilation aspects, and post-ignition timing. Utilizing a transposed convolutional neural network (TCNN), a deep learning model was created to forecast the evolution of smoke visibility, gas temperature, and CO concentration—essential metrics for assessing building tenability. In contrast to traditional numerical modeling approaches that require two days for results, the AI model can deliver precise predictions in just one second with a 97% accuracy rate. Additionally, an accessible interface has been developed to offer pertinent design data such as Available Safe Egress Time (ASET), tenability profiles, and smoke evolution trends. This software showcases its effectiveness across diverse fire engineering scenarios, aiding in safety evaluations, constraint identification, and the provision of swift and practical recommendations to streamline architectural design reviews. Chapter 3 presents advancements on AI model to enable the extension of predicted building geometries from simple cubic boxes to structures with varied floor plans. A comprehensive numerical fire database was established, comprising 136 fire simulations featuring diverse building layouts, random fire placements, varying Heat Release Rate (HRR) patterns, ceiling heights, and floor enclosure conditions. Leveraging the pix2pix GAN algorithms, a deep learning model was developed, showcasing the GAN model's adeptness in capturing overall spatial patterns. It successfully replicates the region with the highest temperature based on provided fire location data and accurately identifies structural boundaries and their impact on gas flow dynamics. The AI-generated representations of ceiling temperature, velocity, and soot density fields achieve an impressive 88% accuracy within a second. These outcomes can be further utilized to estimate detection times for thermal-based devices like heat detectors and sprinklers with a 95% accuracy rate. When juxtaposed with conventional methods reliant on empirical models or CFD simulations, this GAN-powered intelligent approach stands out for its accessibility, affordability, and high accuracy, establishing itself as the most cost-effective solution to date. In Chapter 4, the complexity of the anticipated building geometries was heightened to three dimensions to cater to the practical necessities of cutting-edge architectural designs. A comprehensive numerical fire database was crafted, detailing the progression of smoke layers in various atrium setups and fire locations. This dataset served as the foundation for training two deep learning models: the pix2pix GAN algorithm and the image-guided diffusion model. By inputting data on the building's structure and the fire's placement, the proposed GAN models can swiftly generate three-dimensional soot visibility representations from X, Y, and Z perspectives, with a slight margin of error of 15% for novel scenarios. While the diffusion model excels in replicating intricate flow dynamics and localized smoke patterns, it necessitates a longer processing time of 20 minutes. These outcomes can further aid in estimating the Available Safe Egress Time by analyzing smoke descent profiles, boasting an accuracy of up to 92% for irregular building shapes not explicitly covered in the training set. When contrasted with traditional CFD modeling methodologies, this AI-driven intelligent approach amplifies design efficacy, diminishes computational expenditures, and cultivates a sustainable design environment. Chapter 5 shifts the automation task from fire-smoke analysis to design drawing preparation. Deep learning model that can complete design tasks of sprinkler layout was proposed. A total of 80 sprinkler design drawings were collected as the database. The drawings were pre-processed to keep the main information of the building plan, room functions, and the pre-defined hazard levels which affect the maximum spacing requirement of sprinkler heads. Following the preprocessing of the database, it was employed to train the GAN model with pix2pixHD framework. By inputting the information with building plan and hazard level, the proposed GAN model can generate the placement of the sprinkler nozzles within seconds for new architectural inputs, which is also comparable to the manual design by engineers. The sprinkler coverage rate can reach over 99% which means the output drawings are also code compliant. With AI design models, engineers can free themselves from the labour-intensive drawing workloads and invest their time in more valuable works such as design concept exploration and scheme optimization. Chapter 6 summarizes the overall outcomes of the AI-powered intelligent building fire safety design in its current stage. Challenges and potential areas for future research are also discussed. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13278