基于多任务学习的结构损伤识别  

Structural damage identification based on multi-task learning

作　　者：虞建 张纯[1] 李睿 江汇强 YU Jian;ZHANG Chun;LI Rui;JIANG Huiqiang(School of Infrastructure Engineering,Nanchang University,Nanchang 330031,China;School of Civil Engineering,Shandong University,Jinan 250061,China)

机构地区：[1]南昌大学工程建设学院,江西南昌330031 [2]山东大学土建与水利学院,山东济南250061

出　　处：《南昌大学学报（工科版）》2022年第4期366-372,共7页Journal of Nanchang University(Engineering & Technology)

基　　金：江西省自然科学基金项目(20202BAB204029);江西省研究生教改项目(JXYJG-2019-018)。

摘　　要：利用信号特征自动学习和提取的特性,深度神经网络已被成功应用于基于振动的结构损伤定位或程度诊断。单一的损伤位置或损伤程度诊断网络虽分别能实现一定的诊断功能,但不同任务之间的相关信息没有得到充分利用;因此,将损伤定位任务和损伤程度诊断任务相结合,基于一维空洞卷积神经网络,提出了一个具有信号特征共享与反馈特性的多任务联合学习模型。框架结构的数值模拟和实验模型研究表明,与单任务模型相比,多任务模型能有效降低定位错误和损伤程度估计误差,且具有更好的泛化性。Using the characteristics of automatic learning and extraction of signal features, deep neural network has been successfully applied to structural damage location or degree diagnosis based on vibration.Although a single damage location or damage degree assessment network can achieve certain diagnosis functions respectively, the relevant information between different tasks has not been fully utilized.Therefore, this paper combines the damage location task with the damage degree diagnosis task.Based on one-dimensional dilated convolution neural network, a multi-task joint learning model with signal feature sharing and feedback characteristics was proposed.Numerical simulations and experimental modeling studies of the frame structure showed that the multitasking model can effectively reduce localization errors and damage degree estimation errors with better generalization than the single-tasking model.

关 键 词：深度学习 多任务学习 损伤识别 一维空洞卷积 神经网络 

分 类 号：TU17[建筑科学—建筑理论]