A comprehensive comparison study of deep-learning based methods for structural health monitoring

Vibration-based structural damage detection offers a practical method for timely and remotely identifying existing damages in structures before they grow to irrecoverable failures. Nevertheless, challenges such as handling high-dimensional vibration signals, limited data on diverse damage scenarios, and unavoidable environmental and operational confounding factors complicate its application. To tackle these challenges, advanced pattern recognition approaches, particularly, deep-learning models, have become increasingly attractive to structural engineers, thanks to their ability to learn hidden complex patterns within high-dimensional data and map them to structural states. This study conducts a comprehensive comparative analysis of seventeen machine learning/deep learning models for structural damage detection. These models are evaluated using several benchmarks, including numerical and experimental databases ranging from relatively simple beams to complex multi-story frame structures. A modular workflow is designed to ensure a fair comparison, incorporating consistent training/testing splits, hyperparameter optimization, and measurement metrics for assessing model performance, computational time, model complexity, and robustness. Results reveal that modern convolution-based models such as ResNet and FCN consistently emerge as top-performing models, whereas a 1D convolutional neural network demonstrates a notable balance across various perspectives. Transformer-based models, due to their complexity, may not be as practical as machine learning models for structural engineering applications.