Enhancing Structural Health Monitoring with Machine Learning and Data Surrogates - ATCA-Based Approach for Damage Detection and Localisation

Aug 4, 2026

Structural health monitoring (SHM) continuously assesses structures to detect emerging damage or deterioration over time, yet machine learning performance is often constrained by limited labeled data for many damage cases. This work applies transfer component analysis (TCA) with domain adaptation together with high-fidelity numerical models to generate surrogate data for damage identification. The method is validated on a laboratory nonlinear Brake-Reuß beam, where damage states correspond to different torque settings. In a three-class scenario, numerical-trained models successfully classify experimental measurements.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

AccessFree to download

Abstract

Introduction

Research Framework

Feature Selection and Transfer Learning Workflow

TCA with Domain Adaptation and Data Segregation

Training and Validation of Multiple Classifiers

Enhancing Structural Health Monitoring with Machine Learning and Data Surrogates: ATCA-Based Approach for Damage Detection and Localisation R. S. BATTU, K. AGATHOS and E. PAPATHEOU ABSTRACT Structural health monitoring (SHM) involves constantly monitoring the condition of structures to detect any damage or deterioration that might develop over time. Machine learning methods have been successfully used in SHM, however, their effectiveness is often limited by the availability of data for various damage cases. Such data can be especially hard to obtain from high-value structures. In this paper, transfer component analysis (TCA) with domain adaptation is utilised in conjunction with high-fidelity numerical models to generate surrogates for damage identification without the requirement for high volumes of data from various damaged states of the structure. The approach is demonstrated on a laboratory structure, a nonlinear Brake-Reuß beam, where damage scenarios correspond to different torque settings on a lap joint. It is shown that, in a three-class scenario, machine learning algorithms can be trained using numerical data and tested successfully on experimental data. INTRODUCTION Efficient Structural Health Monitoring (SHM) systems can improve the safety, reliability and service life of structures. Additionally, they aid in lowering the cost and time required for maintenance procedures. There are diverse advancements in recent research in SHM, particularly employing machine learning (ML) algorithms. Typically, ML algorithms which are used in SHM can be categorised into supervised, unsupervised and semi-supervised. Supervised ML algorithms require data from all possible damaged states for successful damage identification in SHM. While it is extremely simple to damage and subsequently obtain data from inexpensive structures, damage to high-value structures, is nearly unattainable. It is a challenging task to solve this problem of data insufficiency. Experimental approaches to use ‘damage proxies’ in the form of added masses have been shown in [1, 2], but they are generally limited to accessible areas of a structure. Another possible solution is to create surrogates with the assistance of high-fidelity physical models. However, Raja Sekhar Battu, PhD Student, Email: [rb756@exeter.ac.uk](rb756@exeter.ac.uk). Dr Konstantinos Agathos, Dr Evangelos Papatheou. Department of Engineering, University of Exeter, North Park Road, EX4 4QF, Exeter, UK damage modelling for the purposes of generating surrogate models is also challenging. Here, a novel technique is employed with the integration of high-fidelity physical models and Transfer Component Analysis (TCA) with domain adaptation. TCA with domain adaptation was initially introduced by Pan et al. [3], who provided a unique feature extraction approach as well as an extensive understanding of the theoretical foundationsand mathematical equations of TCA with domain adaptation. Chakraborty et al. [4] demonstrated a transfer learning strategy in SHM employing time-frequency features on a classification problem. In one of the applications of domain adaptation in SHM, Gardner et al. [5] provided a simple implementation with feature sets of damped natural frequencies and damping ratios in the application of population-based SHM. The results outperformed conventional supervised learning algorithms in terms of classification rates. In another investigation, Ozdagli and Koutsoukos [6] used a domain-adversarial neural network to achieve domain adaptation, which provided evidence for improved prediction accuracy. Poole et al. [7] conducted domain adaptation with statistical alignment as the initial step and proved that the normal correlation alignment is robust to solve the problem of class imbalance. In this work, the core objective is to acquire data surrogates for diverse structural damage classes, which are then used to train efficient multiple classifiers. The validation of the cl

FAQs

Why is data insufficiency a challenge in machine learning-based SHM?

Supervised ML requires data from all possible damaged states, but damage data from high-value structures is difficult to obtain. This scarcity limits effective training and identification performance.

How does the paper generate data surrogates for damage identification?

It combines high-fidelity numerical models with transfer component analysis (TCA) using domain adaptation to transfer knowledge from simulated source domains to target domains, reducing the need for large datasets across all damaged states.

What experimental system is used to demonstrate the approach?

The method is demonstrated on a laboratory nonlinear Brake-Reuß beam. Different damage scenarios are represented by different torque settings on a lap joint.

How are the machine learning classifiers trained and validated?

Multiple classifiers are trained on finite element (FE) data using the surrogate framework and validated using actual damage data collected from laboratory experiments.

Aug 4, 2026

Upload by Evangeline