Domain Adaptation for Structural Health Monitoring

Abstract

In recent years, machine learning (ML) algorithms gained a lot of interest within structural health monitoring (SHM) community. Many of those approaches assume the training and test data come from similar distributions. However, real-world applications, where an ML model is trained on numerical simulation data and tested on experimental data, are deemed to fail in detecting the damage, as both domain data are collected under different conditions and they don’t share the same underlying features. This paper proposes the domain adaptation approach as a solution to particular SHM problems where the classifier has access to the labeled training (source) and unlabeled test (target) domains. The proposed domain adaptation method forms a feature space to match the latent features of both source and target domains. To evaluate the performance of this approach, we present a case study where we train three neural network-based classifiers on a three-story test structure: i) Classifier A uses labeled simulation data from the numerical model of the test structure; ii) Classifier B utilizes labeled experimental data from the test structure; and iii) Classifier C implements domain adaptation by training on labeled simulation data (source) and unlabeled experimental data (target). The performance of each classifier is evaluated by computing the accuracy of the discrimination against labeled experimental data. Overall, the results demonstrate that domain adaption can be regarded as a valid approach for SHM applications where access to labeled experimental data is limited.