基于SegRNN的运营环境下大跨拱桥结构响应预测研究

Analysis on Structural Response Prediction of Long-span Arch Bridges in Operation Environment Based on SegRNN

摘要:

针对桥梁结构响应的预测问题，基于分段递归神经网络（segment recurrent neural network，SegRNN）提出了一种响应预测方法，可以对未来24、48 h内的响应进行预测。SegRNN将输入数据分段，减小了递归神经网络（recurrent neural network，RNN）的迭代次数，在解码阶段，使用并行多步预测的方法，代替了RNN的迭代循环多步预测的方法，提高了预测精度。将过去的桥梁响应数据和温度、交通荷载数据作为输入，使用SegRNN预测未来响应。对一座大跨度钢管混凝土拱桥为期1年的应变、梁端位移和GPS数据进行了分析。选择了10条典型测点位置的监测数据作为数据集，测试了提出方法的有效性。结果表明，SegRNN在所有测试中的平均绝对误差和平均均方误差指标的平均值分别为0.254和0.387，分别取得了12个和11个最佳结果，相比于的LSTM（long short-term memory）、Transformer网络具有显著的优势。提出的方法可以应用于实际桥梁监测中，为桥梁的风险预测提供支持。

Abstract:

Aiming at the prediction problem of bridge structural response, and based on the segment recurrent neural network (segment recurrent neural network, SegRNN), a response prediction method is proposed able to predict the response within the next 24h and 48h. The data segment will be input into SegRNN to reduce the number of iterations of the recurrent neural network (recurrent neural network, RNN). In the decoding phase, a parallel multi-step prediction method is used to replace the iterative loop multi-step prediction method of RNN, which enhances the prediction accuracy. Taking the past bridge response data and temperature and traffic load data as input, SegRNN is used to predict the future responses. The strain, end-beam displacement and GPS data of a long-span concrete-filled steel tube arch bridge over a one-year period are analyzed. The monitoring data from 10 typical measuring point locations are selected as the dataset to test the effectiveness of the proposed method. The results show that the average values of the average absolute error and average mean square error indicators of SegRNN in all tests are 0.254 and 0.387 respectively, achieving 12 and 11 best results respectively, which has the significant advantage compared with LSTM（long short-term memory） and Transformer networks. The proposed method can be applied to actual bridge monitoring and provide the support for the risk prediction of bridges.