Vibration-based deep learning approach for crack detection in stepped beams

Structural Health Monitoring (SHM) is an extremely important research area for maintaining the safety of engineering structures, extending their service life, and detecting potential damage early. In this context, damage occurring in load-bearing elements reduces the local rigidity of the cross-section, leading to changes in vibration behavior. In vibration-based damage detection methods, these changes are monitored through differences in natural frequencies and mode shapes. This thesis proposes a deep learning-based damage detection method using vibration data in stepped elastic beams. The data required for training the model was obtained from the analytical model of a cracked beam. In the analytical model, cracks on the beam were represented by a massless rotational spring model, and local flexibility coefficients were obtained. The frequency equation was derived using the Transfer Matrix Method (TMM). A comprehensive synthetic modal dataset containing the natural frequency and mode shape curvature of the beam was created for different damage scenarios and boundary conditions. This synthetic data was used with a 1D Convolutional Neural Network (1D-CNN) model to predict crack location and intensity on the beam. This study combines the accuracy provided by analytical models with the generalization capabilities of AI-based models to develop a fast and reliable vibration-based damage detection approach adaptable to different boundary conditions for stepped beams.