联合卷积神经网络与长短期记忆深度网络的桥梁损伤识别

Bridge Damage Identification Based on Joint CNN and LSTM Deep Network

为准确评估桥梁结构状态,提升损伤识别效率,提出基于联合卷积神经网络(CNN)与长短期记忆(LSTM)深度网络的桥梁损伤识别方法,并用振动台试验数据进行验证。结合CNN空间特征和LSTM时间特征提取能力,构建桥梁结构损伤识别架构;提取5类时频域损伤特征,经对比分析后,采用结合平均频率和平均能量的组合特征进行损伤识别;基于振动台试验数据及其有限元模型数据识别了斜拉桥模型的损伤,并将识别结果分别与CNN、LSTM的识别结果对比。结果表明:采用联合CNN与LSTM深度网络建立的损伤识别方法可有效识别出桥梁的损伤位置和损伤程度,且偏差小,识别结果优于CNN、LSTM;未布置传感器的位置损伤识别精度较低;轻微损伤识别准确率相对较低。

A bridge damage identification method based on the joint convolutional neural network(CNN)and long short-term memory(LSTM)deep network is proposed,aiming to improve the accuracy of bridge condition evaluation and efficiency of bridge damage identification.And the method is verified by the shaking table test.Based on the spatial characteristics extracting capacity of CNN and the temporal characteristics extracting capacity of LSTM,a bridge structure damage identification framework is developed.Five types of time-frequency domain damage characteristics are extracted,compared and analyzed,as a result,the joint characteristics combining mean frequency and energy are drawn for damage identification.The data of shaking table test supplemented by numerical calculations are taken to identify the damages of the model of a cable-stayed bridge,and the identification results are compared with the results gained by CNN and LSTM,respectively.It is shown that the damage identification method based on joint CNN and LSTM can effectively identify the locations and extent of damages,with minimal bias,and the identification effect is superior to CNN and LSTM.The identification accuracy at the locations without sensors is low,and the identification efficiency of minor damages is relatively low.