Robust Vibration Output-only Structural Health Monitoring Framework Based on Multi-modal Feature Fusion and Self-learning

Abstract

Output-only structural health monitoring is a highly active research direction because it is a promising methodology for building digital twin applications providing near-real-time monitoring results of the structure. However, one of the technical bottlenecks is how to work effectively with multiple high-dimensional vibration signals. To address this question, this study develops a two-stage data-driven framework based on various advanced techniques, such as time-series feature extractions, self-learning, graph neural network, and machine learning algorithms. At first, multiple features in statistical, time, and spectral domains, are extracted from raw vibration data; then, they subsequently enter a graph convolution network to account for the spatial correlation of sensor locations. After that, the high-performance adaptive boosting machine learning algorithm is leveraged to assess structures' health states. This method allows for learning a lower-dimensional yet informative representation of vibration data; thus, the subsequent monitoring tasks could be performed with reduced time complexity and economical computational resources. The performance of the proposed method is qualitatively and quantitatively demonstrated through two examples involving both numerical and experimental structural data. Furthermore, comparison and robustness studies are carried out, showing that the proposed approach outperforms various machine learning/deep learning-based methods in terms of accuracy and noise/missing-robustness.

Development of Structural Damage Detection Method Working with Contaminated Vibration Data via Autoencoder and Gradient Boosting

“…It is possible to include more features in the data to improve the detection accuracy; on the other hand, a feature importance study could be carried out to assess the contribution of every feature, and irrelevant ones could be discarded accordingly. A more in-depth study about feature extraction could be found in [24].…”