Low-story damage detection of buildings using deep neural network from frequency phase angle differences within a low-frequency band

Abstract

Lower stories of buildings are much more vulnerable due to the critically dead load from high stories. Vibration-based damage detection methods are commonly used to quantify damage with indicators and early assess structural deterioration. However, these indicators are usually not directly related to structural performance and can be distorted by the modeling errors or sensing uncertainties. In addition, damage detection of structures can be realized from sensor data by the remaining useful life estimation framework in machine learning. Meanwhile, a machine learning model, which detects structural damage, can be trained by the generated data from a numerical/identified model, e.g., phase angle differences of structural transfer functions. Moreover, the bounded nature of phase angles coincides with the needs of machine learning algorithms, and the combined stiffness and damping changes involved in damaged structures can also be considered in a learning model. Therefore, this study proposes a neural network damage detection method based on frequency response phase angle differences for low-story damage of buildings. In this proposed method, a simplified model is established from identified modal parameters of a structure and optimized using the least-squared stiffness approach, and the transfer function phase angle differences are generated from the model considering the stiffness and damping changes in various damage scenarios. The generated data are exploited to train a deep learning model which predicts the residual stiffness percentages as the condition parameters in the remaining useful life estimation framework. In this study, a numerical example provides detailed procedure to develop the neural network damage detection model and evaluate the effectiveness of the learning model. The proposed method is also employed in a scaled six-story steel-frame building for experimental verification. As seen in the results, the neural network damage detection method can estimate the residual stiffness percentages in the lower stories of a building based on frequency response phase angle differences.