Frequency-based deep learning model for beam damage detection

Abstract

Damages in beam-like structures are a critical concern, necessitating early detection to prevent catastrophic structural failure. Detecting such damage is essential for preserving structural integrity, extending operational lifespan, and reducing maintenance expenditures. Despite advancements in damage quantification and localization methods using vibration measurements, many challenges remain in accurately determining damage characteristics, including location, width, and depth. This study proposes an integrated deep learning (DL) approach leveraging frequency response data for precise damage characterization. The methodology combines experimental modal analysis and finite element simulations (FEA) to generate reliable training data for DL models. Five potential DL architectures—Feedforward Neural Networks (FNN), Multilayer Perceptron (MLP), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) networks, and 1D CNN-GRU-ResNet model—have been optimized using the Electric Eel Foraging Optimization (EEFO) algorithm for enhanced accuracy and computational efficiency. The experimental modal analysis provided vibration responses, which were transformed into the frequency domain using Fourier Transform techniques. A comparison between numerical (FEA) and experimental frequencies showed a maximum deviation of <9%, validating the accuracy of the extracted modal parameters. Among the evaluated models, the EEFO-optimized LSTM and 1D CNN-GRU-ResNet demonstrated superior predictive performance, with LSTM achieving the lowest errors—MSE of 0.378 (train) and 0.446 (test), and MAE of 0.494 (train) and 0.572 (test)—while the 1D CNN-GRU-ResNet closely followed, recording MSE values of 0.504 (train) and 0.528 (test), and MAE values of 0.518 (train) and 0.531 (test). Model calibration analysis indicated that 1D CNN-GRU-ResNet showed better reliability in predicting width and depth, while LSTM was more accurate for location. Additionally, SHAP (Shapley Additive Explanations) analysis revealed that lower-order modal frequencies mainly influenced location prediction, while higher-order frequencies were more sensitive to depth variations. The proposed frequency-based DL framework, especially the EEFO-optimized LSTM and 1D CNN-GRU-ResNet models, demonstrates strong potential for reliable, scalable, and efficient structural health monitoring of beam structures. © 2025 Informa UK Limited, trading as Taylor & Francis Group.