基于GA-LSTM的桥梁缆索腐蚀钢丝力学性能预测模型  

Mechanical property prediction of bridge cable corroded steel wires based on GA-LSTM model

作　　者：缪长青[1,2] 吕悦凯 万春风[1,2] MIAO Changqing;LÜYuekai;WAN Chunfeng(Key Laboratory of Concrete and Prestressed Concrete Structure of Ministry of Education,Southeast University,Nanjing 211189,China;School of Civil Engineering,Southeast University,Nanjing 211189,China)

机构地区：[1]东南大学混凝土及预应力混凝土结构教育部重点实验室,南京211189 [2]东南大学土木工程学院,南京211189

出　　处：《东南大学学报(自然科学版)》2025年第1期140-145,共6页Journal of Southeast University:Natural Science Edition

基　　金：国家自然科学基金资助项目(52178119)。

摘　　要：为了精准捕捉桥梁缆索腐蚀钢丝的时变规律并预测其力学性能,开发了一种基于遗传算法(genetic algorithm, GA)优化的长短期记忆(long short-term memory, LSTM)神经网络模型。该模型利用GA依次优化LSTM模型的迭代次数、隐藏层层数、神经元数量、窗口大小4个超参数,以预测不同腐蚀特征状态下钢丝的力学性能。将其与传统LSTM和GA-反向传播模型的预测结果进行比较。结果表明,GA-LSTM模型具有更高的预测精度和鲁棒性。在屈服强度与极限强度预测效果方面,均方根误差(root mean square error, RMSE)、平均绝对误差(mean absolute error, MAE)、决定系数分别提高约44%~61%、43%~57%、35%~92%。在屈服应变与极限应变预测效果方面,RMSE、MAE、决定系数分别提高约0~46%、7%~49%、12%~229%。所建立的模型可以作为一个有用的工具支持桥梁缆索腐蚀安全性评估工作。In order to accurately capture the time-varying rules of corroded steel wires in bridge cables and pre-dict their mechanical properties,a long short-term memory(LSTM)neural network model optimized by ge-netic algorithm(GA)was developed.The four hyperparameters of the LSTM model,namely the number of it-erations,the number of hidden layers,the number of neurons,and the window size,were optimized in se-quence by GA in this model to predict the mechanical properties of steel wires under different corrosion charac-teristic states.The prediction results were compared with those of the traditional LSTM and GA-back propaga-tion models.The results show that the GA-LSTM model has higher prediction accuracy and robustness.In the aspect of prediction effects of yield strength and ultimate strength,the root mean square error(RMSE),the mean absolute error(MAE),and the coefficient of determination are improved by approximately 44%to 61%,43%to 57%,and 35%to 92%,respectively.In the aspect of prediction effects of yield strain and ulti-mate strain,RMSE,MAE,and the coefficient of determination are increased by approximately 0 to 46%,7%to 49%,and 12%to 229%,respectively.The established model can serve as a useful tool to support the safety assessment of bridge cable corrosion.

关 键 词：桥梁缆索腐蚀钢丝 力学性能预测 时序预测 神经网络 遗传算法 超参数优化 

分 类 号：U443.38[建筑科学—桥梁与隧道工程]