Deep Learning-Based Imputation Framework in Bridge Health Monitoring using Generative Adversarial Networks

Saha, Sumit (2023) Deep Learning-Based Imputation Framework in Bridge Health Monitoring using Generative Adversarial Networks. Masters thesis, Indian Institute of Technology Hyderabad.

Text
CE20MTECH12002.pdf - Submitted Version
Restricted to Registered users only until 30 June 2028.
Download (3MB) | Request a copy

Abstract

Structural Health Monitoring (SHM) has emerged as an important area of structural engineering that evaluates existing civil structures as they age over time. With aging, these structures become increasingly susceptible to various human activities, unexpected traffic loads, natural disasters such as earthquakes, and wind effects. These factors can lead to reduced performance, functionality, and residual life. Recent advancements in data science have paved the way for using Machine Learning (ML) and Deep Learning (DL) techniques in SHM. Wireless sensors are crucial for structural health monitoring systems as they collect structural responses to assess structures' load-carrying capacity and serviceability. However, sensor failure during signal transmission can result in data loss, leading to an incorrect diagnosis of the structural health status. Both continuously missing data and random missing data pose critical challenges for damage detection and condition assessment and require significant attention. Conventional methods for lost data imputation, such as correlation techniques, have low efficiency due to their incompact neural network structures. To address this issue, this paper proposes a modified generative adversarial imputation network (GAIN) to recover missing acceleration data collected during ambient vibration of the bridge. The proposed framework utilizes neural networks with a generator-discriminator architecture to extract valuable information from the non-missing portions of faulty sensors and impute the data for the missing portions. The proposed method demonstrates efficient and accurate imputation performance under retrofitting and non-retrofitting stages through an analysis of measured acceleration signals under ambient excitation of the steel bowstring railway bridge in Belgium, KW 51. Comparative results based on performance measure parameters such as root mean squared error (RMSE), mean average error (MAE), correlation factor (R2 ), and accuracy indicate that the GAIN framework performs well for different sensors and six different monitoring periods daily. Following successful imputation, the completed datasets are utilized as inputs for a generative adversarial network (GAN)-based framework to augment the vibration dataset for further assessment during several stages of SHM.

[error in script]

IITH Creators: