To meet the demand of efficient and near-real-time condition assessment of high-speed railway (HSR) bridge bearings, this study proposes a deep learning-based methodology utilizing train-induced bridge displacements. Firstly, a quantitative analysis is performed to validate the sensitivity of bridge displacements to bearing damage. Then, the bidirectional long short-term memory (BiLSTM) network is trained as a baseline model under healthy bearing conditions, using the midspan displacement as input and the quarter-span displacements as outputs, to capture the intrinsic dynamic correlations among bridge displacements. Subsequently, the true-to-predicted signal ratio (TPSR) is formulated based on the measured and predicted displacements under damaged bearing conditions to enable both qualitative identification and quantitative assessment. Moreover, the mean-based resampling strategy and support vector machine (SVM) are incorporated to enhance the identification of minor damage and to achieve simultaneous multi-bearing assessment. The proposed methodology is numerically validated under both single- and multiple-bearing damage conditions with varying severities. The results demonstrate its capability to accurately identify damage types and quantify severities. Furthermore, performance evaluations considering measurement noise, train speed, and train configurations confirm the robustness and reliability of the proposed approach under complex service conditions.
Jiang et al. (Fri,) studied this question.