基于迁移学习和Bi-LSTM神经网络的桥梁温度-应变映射建模方法

A method of modeling temperature-strain mapping relationship for long-span cable-stayed bridges using transfer learning and bi-directional long short-term memory neural network

为快速构建并准确预测温度作用引起的斜拉桥主梁应变用于结构状态评估,基于某大跨度斜拉桥主梁超过1年的温度和应变监测数据,提出了一种基于迁移学习和双向长短时记忆(bi-directional long short-term memory,Bi-LSTM)神经网络的斜拉桥温度-应变映射模型建立方法。首先,利用解析模态分解(analytical mode decomposition,AMD)去噪应变数据,得到仅由温度引起的应变响应;其次,选择温度和某一测点应变数据构成数据集,采用Bi-LSTM神经网络训练该数据集,并通过网络结构和超参数优化建立温度-应变Bi-LSTM基准模型;最后,利用迁移学习方法,将已训练好的基准模型中部分参数迁移到其他温度-应变数据集,建立相应的温度-应变映射被迁移模型,并与未采用迁移学习的神经网络训练方法进行对比。研究结果表明,相比直接建立的温度-应变Bi-LSTM神经网络映射模型,采用迁移学习方法建立的被迁移模型,其拟合精度均高于所用的基准模型,且训练时间短,预测误差小。

To rapidly construct and accurately predict the strain responses of amain girder induced by temperature in along-span cable-stayed bridge for structural condition assessment,based on the measured temperature and strain data on the main girder of a long-span cable-stayed bridge over 1 year,a method of constructing the temperature-strain mapping model by using the transfer learning technique and the bidirectional long short-term memory(Bi-LSTM)neural networks was proposed in this study.Firstly,the analytical mode decomposition(AMD)was adopted to denoise the strain data to obtain the temperature-induced strain.Secondly,the temperature and the strain data at a particular measurement point were selected to form a dataset,and were fed to a Bi-LSTM neural network.Then a well-fitting neural network baseline model was constructed by optimizing the network structure and hyperparameters.Finally,using the transfer learning method,some parameters from the trained Bi-LSTM neural network model were transferred to other temperature-strain datasets to construct the transferred temperature-strain mapping models.Compared with the temperature-strain Bi-LSTM neural network models constructed directly from the datasets,the transferred temperature-strain Bi-LSTM neural network models built by using the transfer learning technique have higher fitting accuracy,shorter training time,and smaller prediction error.