Hybrid Modal-Machine Learning Approach for Structural Damage Diagnosis

Abstract

Many machine-learning-based structural damage diagnosis methods have been developed in the recent decade due to the advancements of sensors and cloud computing. Machine learning models are accurate in making predictions when they are trained with sufficient labelled data, but only on seen events. In the context of damage diagnosis, damage events are rare. This leads to insufficient labelled damage data for machine learning model training, which is also known as the cold-start issue. A physics-based approach such as modal-based method that identifies damage through the changes in dynamic characteristics could be implemented at early phases of damage diagnosis before the trained model is available. Therefore, a two-stage hybrid modal-machine learning approach is proposed for structural damage diagnosis to solve the cold-start issue. The first stage applies an unsupervised method to detect damage presence, while the second stage implements a combination of mode shape assessment and supervised method to locate the damage when damage is present. Results showed accuracies of 100% in detecting damage presence at the first stage, up to 100% in locating unseen single damage at the second stage, and up to 83.3% in locating unseen multiple damages at the second stage.