Development of a procedure for model selection in machine learning-based methods for structural damage detection

Abstract

Output-only methods based on machine/deep-learning algorithms are highly practical approaches for timely detecting potential damages in civil structures as they directly employ measured vibration signals but do not
require exact knowledge of input loading nor the service interruption for manual inspection. However, there is
no one-size-fits-all model that is optimal for all problems in different perspectives; hence, it is necessary to
discover the advantages as well as drawbacks of different models, then leverage these understandings to select
the most appropriate model for specific structures in reality. Therefore, this study develops a framework that
facilitate the model selection by extensively comparing various machine learning-based methods ranging from
relatively simple ones such as Na¨ıve Bayes to complex ones such as the extreme boosting tree-based ensemble
model. The framework can provide comparison results include various aspects such as model complexity, training time, detection accuracy, and inference time.