IMPACT IDENTIFICATION IN FRAMED STRUCTURES USING DEEP LEARNING: A CNN-BASED APPROACH OPTIMIZED BY BAYESIAN OPTIMIZATION

Accurate detection and localization of impacts in structural systems are crucial for safety and enabling effective structural health monitoring (SHM). This paper aims to identify multiple consecutive impacts in framed structures with unknown dynamic properties, using time-domain acceleration data. Traditional methods often struggle under complex conditions such as noisy environments and multiple impacts. To overcome these limitations, we propose a deep learning-based framework utilizing Convolutional Neural Networks (CNNs) to extract intricate patterns from acceleration signals. Input data are generated through high-fidelity numerical simulations based on the Finite Element Method (FEM), allowing precise control over impact characteristics and their spatial distribution. A fixed-length sliding window is employed to segment the acceleration time series, enabling the model to perform localized and near-real-time impact detection. To further improve model performance, Bayesian optimization is utilized for hyperparameter tuning, enhancing accuracy and efficiency over traditional grid search. The proposed model is numerically evaluated on two-dimensional structures: a steel pin-jointed camel-back truss and a shear frame. The results reveal that the proposed strategy achieves high accuracy in estimating the location, timing, and magnitude of impacts, even under noisy conditions. The key novelty of this research lies in combining deep learning with advanced optimization techniques to solve the impact detection problem in structures with unknown parameters. These findings establish a robust framework for advancing intelligent, data-driven SHM systems, with direct applications in real-world infrastructure. The proposed methodology demonstrates significant potential to mitigate economic costs and safety risks associated with structural failures under impact loading.

Keywords:

Impact identification , Impact localization , Framed structures , Deep learning , CNN , CNN-LSTM , Bayesian algorithms , Time domain