A data-driven framework for structural health monitoring using reinforcement learning and deep autoencoders

The importance of structural health monitoring, especially for bridges that have exceeded their design life, has increased significantly to reduce maintenance costs, improve safety, and optimize performance. This article proposes a combination of machine learning tools for automatic detection of anomalies in bridge structures from accelerometer data. The Z-24 bridge, which has a comprehensive database of accelerometer data, was considered as a case study. By extracting modal frequencies from accelerometer data using operational modal analysis, the proposed framework, which combines machine learning methods, identifies damage-sensitive features and provides a coherent framework for structural health monitoring. One of the main innovations of the proposed method is the integration of reinforcement learning with autoencoders. The proposed method, Reinforcement Learning Autoencoder Neural Network (RLAENN), is validated using real accelerometer data, and its performance is compared with the classical benchmark algorithms, including Principal Component Analysis (PCA), Kernel PCA (KPCA), Gaussian Mixture Model (GMM), and Mahalanobis Square Distance (MSD). By eliminating the need to tune sensitive parameters, this method achieves significant improvements in accuracy and F1 score over conventional methods, confirming its performance and effectiveness.