Novel digital-twin modelling for smart building inspection

• Main Author: Mui, Jia Duan
• Other Authors: Fu Yuguang
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Neural networks; Structural health monitoring; Building responses; Digital twin; Machine learning
• Online Access: https://hdl.handle.net/10356/177523
• Institution: Nanyang Technological University

Buildings are everywhere in modern society. They make up the bulk of the structures in modern civilization. As time goes on, buildings get increasingly taller. Since buildings take up most of the space in our cities, there is a necessity to maintain the structural integrity of these buildings. Structural Health Monitoring is a process where various key data points are used to evaluate the overall health of the structure. These data points such as strain, acceleration and displacement are measured and recorded through sensors placed throughout the structure. With these data points, Structural Health Monitoring aims to detect and make decisions about the structure’s condition either in real time or over a certain duration. This decisions can help improve the safety of the structure, prevent any potential minor or major failures and help prolong the structure’s lifespan. This Final Year Project aims to explore combining this aspect of Structural Health Monitoring with deep learning neural networks in hopes to provide real time predictions to a building’s structural responses. Three different architectures of neural networks will be used in this Final Year Project. The three different model architectures used are MLP, CNN and LSTM. Each of these models were trained using actual recorded building responses as data point inputs and evaluated by comparing the predicted data point outputs against the actual measured data points. The data points recorded were strain, acceleration and displacement on various floors and columns of a proposed physical laboratory model of a building. A total of 6 batches of data points were collected. In the first batch, the 3 different data points of each floor of a 6 story laboratory model were recorded. In subsequent rounds, the top most floor was removed from the model. In the final batch, the data points of a 1 story laboratory model were recorded. The strain and acceleration data points were used as the training inputs while all 3 data points were used as the predictive output of the different neural network models. The MLP, CNN and LSTM models were trained and evaluated for 4 different prediction rounds. After a comparison of the different neural network architectures, the CNN model was selected as most adept at predicting any of the 3 datasets given different strain and acceleration inputs from various floors. The results of this project hope to be combined with a real time Structural Health Monitoring program where an actual life sized building may become the intended proposed structure. The predictive prowess of the neural networks may be used to create and update the parameters for a digital structural model with the end goal of creating a digital twin of the life sized building.