Structural damage detection using voting ensemble of fine-tuned convolutional neural networks and time-frequency images

Investigating the recorded response of a structure to dynamic loads is an efficient method for understanding and describing its current status. In the present paper, the ability of different Convolutional Neural Network (CNN) algorithms using time-frequency images and the performance of a voting ensemble of the models have been investigated in classifying various types of structural damage. The time-frequency images fed into CNNs were generated from acceleration responses obtained from undamaged and damaged conditions of experimental and real-world structures. The structural damages considered in the case studies encompassed various types, severities, and locations, highlighting the variation in the damaged conditions. The findings indicated that employment of a soft voting ensemble learning method, with an average prediction accuracy of 98.5%, Yielded appropriate outcomes. Moreover, in the evaluation of different CNN architectures assessed, DenseNet-based models exhibited superior performance in three distinct considered structures, while VGG-based models exhibited the highest performance across all CNNs in one specific case study focused on the location of damages, respectively. Additionally, an examination was carried out to evaluate the impact of factors that could influence the prediction accuracy of the algorithms. The results showed that increasing the duration of each acceleration record led to an improvement in the final accuracy by about 4% in the investigated structure. Furthermore, the usage of Bump mother wavelet gave rise to the highest performance.