Quantification of Damage Using Indirect Structural Health Monitoring

TL;DR

This paper presents a method for damage quantification in bridges using indirect accelerometer-based monitoring combined with machine learning models, demonstrating promising results in controlled tests.

Contribution

It introduces a novel approach combining FFT, PCA, and regression models for damage quantification via indirect structural health monitoring.

Findings

Support Vector Regression outperformed Gaussian Process Regression in MSE.

The methodology successfully distinguished different damage levels in tests.

Normalized FFT data improved model accuracy.

Abstract

Structural health monitoring is important to make sure bridges do not fail. Since direct monitoring can be complicated and expensive, indirect methods have been a focus on research. Indirect monitoring can be much cheaper and easier to conduct, however there are challenges with getting accurate results. This work focuses on damage quantification by using accelerometers. Tests were conducted on a model bridge and car with four accelerometers attached to to the vehicle. Different weights were placed on the bridge to simulate different levels of damage, and 31 tests were run for 20 different damage levels. The acceleration data collected was normalized and a Fast-Fourier Transform (FFT) was performed on that data. Both the normalized acceleration data and the normalized FFT data were inputted into a Non-Linear Principal Component Analysis (separately) and three principal components were extracted for each data set. Support Vector Regression (SVR) and Gaussian Process Regression (GPR) were used as the supervised machine learning methods to develop models. Multiple models were created so that the best one could be selected, and the models were compared by looking at their Mean Squared Errors (MSE). This methodology should be applied in the field to measure how effective it can be in real world applications.