【摘要】 针对桥梁钢桁架疲劳损伤识别难度大、精度低的现状，提出基于声发射信号时频分析与深度学习的钢桁架焊接节点损伤程度识别方法。对桁架节点在桥梁运营状态下产生的声发射信号进行小波变换，表征不同损伤程度信号的时频能量分布模式，然后建立卷积神经网络(convolutional neural network, CNN)模型对时频图进行损伤特征提取，并通过迁移学习思想提升模型的训练效率和学习能力，从而实现桁架焊接节点严重损伤、轻微损伤和噪声工况的准确识别。进一步对模型各卷积层激活区域进行可视化分析，解剖模型的损伤特征学习过程及分类逻辑。某悬索桥中央纵向腹板钢桁架焊接节点现场试验结果表明：相较于利用时域波形进行特征学习的一维卷积神经网络模型，时频图包含了更丰富的损伤信息，所建立的二维卷积神经网络模型对钢桁架焊接节点三种损伤程度的识别准确率超过94%,具有更强鲁棒性和实际应用价值。更多还原

【Abstract】 Here, aiming at the current situation of large difficulty and low accuracy in identifying fatigue damage of bridge steel trusses, a method for identifying damage degree of steel truss welded joints based on time-frequency maps of acoustic emission(AE) signals and deep learning was proposed. Wavelet transform was performed for AE signals generated by truss nodes under bridge operation status to characterize time-frequency energy distribution modes of signals with different degrees of damage. Then, a convolutional neural network(CNN) model was established to extract damage features from time-frequency maps, and the training efficiency and learning ability of the model were improved with transfer learning ideas to realize correct recognition for truss welded nodes’ severe damage, minor damage and noise working conditions.Further, visualization analyses were performed for activation regions of various convolutional layers of the CNN model to dissect damage feature learning process and classification logic of the model.The on-site test results for welded joints of the central longitudinal web steel truss of a certain suspension bridge showed that compared to the one-dimensional CNN model with time-domain waveform for feature learning, the time-frequency maps obtained with the proposed method contain more abundant damage information; the established two-dimensional CNN model has a recognition accuracy of over 94% for 3 damage levels of welded joints of steel truss, and has stronger robustness and practical application value.更多还原