Transfer Learning Based Bridge Damage Detection: Leveraging Time-Frequency Features

Bridges are a crucial part of the transport infrastructure network, and their safety and operational conditions need to be ensured. An early warning of damage is valuable for condition-based maintenance to avoid costly consequence including structural collapse. Dynamic behaviour of bridge structures can be used as indicators for their health status. Machine Learning techniques allow high-dimensional connections between the structures' vibrational responses and their state of health. In this research, a novel transfer learning-based approach is presented for identifying the location of damage in concrete bridges utilising the time-frequency characteristics from dynamic responses of the bridge under moving vehicles. Convolutional neural networks (CNNs) are used to extract discriminative features from the time-varying input data of vehicle-bridge interaction. Pre-trained CNNs are then fine-tuned for multiple damage classification tasks. The performance of the proposed method is evaluated by comparing it with a variety of pre-trained networks and optimized classification algorithms. Effects of the noise level, vehicle speed, and sensor location on the predicted results are also studied. The numerical results show that the proposed method can precisely locate the damage on concrete bridges using only a single sensor on the bridge deck. The method has valuable potential for practical application for localising the bridge structural damage with further fine-tuning using the field measurement data.