Structural Damage Classification of Large-Scale Bridges Using Convolutional Neural Networks and Time Domain Responses

This study presents the structural damage classification of a large-scale bridge, considering several damage scenarios using One dimensional convolutional neural network (1D-CNN). Measurements obtained from the Z24 bridge in Switzerland during the short-term progressive damage tests are used for this study. Acceleration responses at 291 sensor locations are measured under forced and ambient excitations. This study considers only the measurements under ambient excitations, which has the advantage over forced excitation of not required to measure the excitations. Furthermore, to reduce the overall cost of monitoring the structure, this study aims to use fewer sensor measurements. Out of 291 sensors, only three measurements are used in this study. Each measurement contains 65,536 samples collected at a sampling frequency of 100 Hz. The measurements from three sensors are processed into shorter lengths of 150 data points, each with a 50% overlap. The processed data are inputted to the proposed 1D-CNN model. The proposed 1D-CNN consists of two 1D-CNN networks with different kernel sizes to perform better with different abstract features. The flattened outputs from these two networks using the same input are concatenated and fed into a fully connected dense network for damage classification. The labelled outputs are the different damage scenarios introduced in the progressive damage tests. The performance of the proposed approach is measured in terms of accuracy supported by a confusion matrix. The performance is measured for three cases. The result indicates that better performance is obtained compared to a previous study with the fused features as input to the deep learning models, although fewer sensors are used.

The findings from this study demonstrated that a good damage classification could be achieved using fewer sensor measurements from a large-scale bridge. The Z24 bridge benchmark data are used as an example in this study. Several damage scenarios were considered during the progressive damage test, and all tests were performed under ambient and forced excitation conditions. A multi-headed, one-dimensional convolutional neural network is proposed to classify the damages using the ambient condition data set. The performance is compared with an existing study using the same data set, but the data pre-processing techniques and model are improved. Three cases are defined by varying the size and length of the available time-series data. The proposed model has obtained better damage classification results for all the cases than an existing study. The advantage of the study is that the damage classification is performed using data obtained from a real large-scale structure under ambient conditions, eliminating the need for external force excitation. The proposed method could also be used for condition monitoring and safety evaluation of aerospace and mechanical structures.