基于时频分析与深度学习的结构震后损伤评估

为评估地震后钢筋混凝土(RC)框架结构的损伤状态，提高损伤评估的效率和精度，文章提出一种基于时频分析和一维卷积神经网络(1D-CNN)的地震损伤评估方法。首先利用增量动力时程分析对一个6层RC框架结构进行地震损伤模拟，并根据最大层间位移角对加速度信号进行损伤程度的标定，以此来获取数据样本，随后应用五种不同的时频分析方法对原始信号进行处理；然后建立基于1D-CNN的地震损伤评估模型，并利用贝叶斯优化算法寻找模型中的最优参数组合；最后评估所提出模型方法在噪声情况下的泛化能力。研究结果表明：五种时频分析方法中，小波散射变换方法的准确率最高，达92.5%,且计算速度也最快，仅需144 s;另外在噪声下该方法仍可以保持较高的损伤评估准确率，具有较好的鲁棒性和泛化能力。

Structural damage assessment after earthquakes using time-frequency analysis and deep learning

To assess the damage state of reinforced concrete (RC) frame structures after earthquakes and improve the efficiency and accuracy of damage assessment, this study proposes an earthquake damage assessment method based on time-frequency analysis and one-dimensional convolutional neural network (1D-CNN). First, the earthquake damage to a six-story RC frame structure was simulated using incremental dynamic analysis. Based on the maximum story drift ratio, the degree of damage was calibrated to obtain data samples. Second, four different time-frequency analysis methods were applied to process the original signals. Third, an earthquake damage assessment model based on a 1D-CNN was established, and the optimal parameter combination in the model was determined using the Bayesian optimization algorithm. Finally, the generalization ability of the proposed model under noise was evaluated. The results show that among five time-frequency analysis methods, the wavelet-scattering transform method has the highest accuracy, reaching 92.5%, and the fastest calculation speed, taking only 144 s. In addition, the proposed method can maintain a high level of damage assessment accuracy under noise conditions, indicating good robustness and generalization ability.