Structural damage identification and classification from time-frequency impact echo data using deep learning

Deep learning advancements have enabled non-destructive crack detection, yet it remains inadequate to quantitatively identify and evaluate the structural damage states leveraging deep learning with impact echo (IE). Therefore, the present study proposes a method to achieve this target on concrete structures, utilising Convolutional Neural Networks (CNNs) and IE. During the experiments, a reinforced concrete beam was loaded to simulate various damage levels. The IE test was conducted on the pure bending zone of the beam, and the obtained data were transformed into two-dimensional (2D) time-frequency data for a six-state damage dataset. Subsequently, several CNNs were used for training, testing, and analysing their performance. The findings revealed the networks’ proficiency in distinguishing various degrees of damage. Among them, GoogLeNet emerged as the most accurate classifier. Further analysis indicated that GoogLeNet, trained with datasets from at least two monitoring units, significantly outperformed those trained using datasets from a single unit, achieving a remarkable F1 score of no less than 0.778. Additionally, the study compared 1D and 2D GoogLeNet models trained on different data formats. The results showed that the model trained on 2D time-frequency data can achieve a higher accuracy of 0.984, surpassing the 1D model trained on time-series data.