Structural damage diagnosis of a cable-stayed bridge based on VGG-19 networks and Markov transition field: numerical and experimental study

Abstract Traditional physical-driven modal methods are inappropriate for damage diagnosis of long-span flexible structures with complex mechanical behaviour. This study develops a deep Convolutional Neural Network-based damage diagnosis method for in-service bridges by using dynamic responses under moving loads. The dynamic responses were collected from the critical points on the girders of a cable-stayed bridge specimen under vehicle loading. These collected data was transformed into images based on Gramian Angular Field and Markov Transition Field (MTF). A deep learning algorithm based on VGG-19 was used to extract the damage feature from the data images associated with the structural responses. Finally, the unlabelled vibration data were input into the VGG-19 model for structural damage diagnosis. An experimental study was conducted for the damage diagnosis of a scale specimen of a cable-stayed bridge under moving loads. The acceleration signals of the main girder of the cable-stayed bridge under several damage conditions were monitored. The numerical results show the training accuracy of the deep learning method based on VGG-19 with MTF is up to 88%, and the average accuracy of the test dataset is 86.46% for each classification label. The transfer learning method can increase the classification accuracy up to 97.89%, indicating the advantage of intergrating transfer learning and VGG-19 network for structural damage diagnosis. The combination of VGG-19 and MTF algorithm provides a better solution for structural damage diagnosis of in-service infrastructures with long-term monitoring data.