Comparative study of machine learning algorithms for health monitoring of benchmark buildings using multi-domain features

Abstract

Traditional manual inspection approaches for structural health monitoring are time-consuming, unreliable, and sometimes impractical for large-scale structures, motivating the use of automated, data-driven techniques. This study compares different machine learning algorithms and multi-domain features, from simulated data to the health monitoring of an ASCE benchmark building. For that purpose, an ASCE benchmark building is modelled in the ANSYS environment, and time-history acceleration data is collected for healthy and various unhealthy cases. Three distinct features are extracted from the data. (1) statistical features, (2) frequency-domain features (3) time-frequency features, which are utilised as input to the artificial neural networks (ANN), k-nearest neighbours (kNN), and random forests (RF) algorithms. The RF and statistical features combination provides the highest classification accuracy. The findings offer helpful information about selecting the most effective ML algorithms and suitable features for SHM applications.