基于BP-LSTM混合模型的钢-混组合桥面系空间温度场及温度效应实时评估及预测

考虑到传统健康监测系统与有限元方法的不足，为实现钢-混组合桥面系温度及温度效应的实时评估及预测，提出基于多层前馈——长短期记忆(BP-LSTM)混合模型的实时评估及预测方法。研究过程中，以某桥为工程背景，采用其健康监测系统数据，并利用有限元法热力耦合分析扩充数据，构建包含“结构特征、时间特征、环境特征——温度、温度效应”映射的样本库，并通过实测数据验证有限元计算的准确性；设计BP-LSTM混合模型结构，对LSTM网络以考虑时间权重的均方误差(MSE)损失函数加以改进，读入样本库数据进行训练、验证、测试，从而获得输出精度较高的混合模型；最后，分别给出利用BP-LSTM混合模型进行温度及温度效应实时评估与预测的方法。结果表明:基于气象数据的结构温度场有限元计算结果与实测基本一致，“结构特征、时间特征、环境特征——温度、温度效应”映射具有高度非线性，单独BP网络或LSTM网络效果不佳，而BP-LSTM混合模型精度较高；根据本文提出的实时评估及预测方法，混合模型输出值与目标值的决定系数最低为0.930，最高为0.993，精度较高，可用于钢-混组合桥面系空间温度场及温度效应实时评估及预测。

Real-time evaluation and prediction of spatial temperature field and temperature effect of steel-concrete composite bridge deck system based on BP-LSTM hybrid model

Considering the deficiency of traditional health monitoring system and finite element method, a real-time evaluation and prediction method based on multi-layer feedforward—long and short term memory (BP-LSTM) hybrid model was proposed to realize the real-time evaluation and prediction of steel-concrete composite bridge deck system temperature and temperature effect. The health monitoring system data of a bridge was were analyzed, the thermo-mechanical finite element method analysis was used to expand the data, and a sample database containing the mapping of “structural characteristics, time characteristics, environmental characteristics—temperature and temperature effect” was constructed. The accuracy of the finite element calculation was verified by the measured data. The structure of BP-LSTM hybrid model was designed, and the LSTM network was improved by considering the MSE loss function of the time weight. The hybrid model with high output accuracy was obtained by reading the data into the sample database for training, verification and testing. Finally, a method for real-time evaluation and prediction of temperature and temperature effect by using BP-LSTM hybrid model was proposed. The results show that the finite element calculation results of structural temperature field based on the meteorological data agree well with the measured results. The mapping of “structural characteristics, time characteristics, environmental characteristics—temperature, temperature effect” is highly nonlinear. The single BP network or LSTM network has poor effect, while the mixed model of BP-LSTM has high precision. By the the proposed real-time evaluation and prediction method, the determination coefficients between the output value and the target value of the mixed model are determined as 0.930 at the lowest and 0.993 at the highest. The research results can provide reference for real-time evaluation and prediction of spatial temperature field and temperature effect of steel-concrete composite bridge deck system.