Comparative Analysis of SHM Features Leveraging Observational Bias for Ageing Damage Detection
Please login to view abstract download link
In the field of Structural Health Monitoring (SHM), accurately identifying different types of damage in structures is a fundamental challenge. This holds especially true for long-term ageing damage, which evolves gradually and may be pronounced in historic structures. Such damage can often be overlooked by traditional methods that are attuned to detecting sudden changes. Recognizing the critical need for better understanding and monitoring of such progressive damage, this study focuses on characterizing ageing damage using long-term vibration data. These advancements are essential for the effective monitoring and maintenance of structures, playing a significant role in preserving their structural integrity over time. The research methodology employs a comparative analysis of various damage-sensitive features to test their effectiveness in identifying trends indicative of ageing damage. This analysis spans across different data types, including raw vibration data, features informed by physical principles, and latent features extracted using Autoencoders (AE). A critical part of the comparison is evaluating the differences between using physically-informed features versus latent features and understanding the benefits of leveraging observational bias in a Physics-Informed Machine Learning (PIML) framework, as opposed to a conventional black-box, data-driven machine learning approach for feature extraction. The effectiveness of each feature set is evaluated using a Gaussian Process model for time-series forecasting. The model is tested on various input-output pairs to determine the relative efficacy of different feature datasets. The data used is generated from a Single Degree of Freedom (SDOF) system under dynamic loading, to provide a reliable ground truth and potentially expand the relevance of the results to a wider range of structural scenarios.
