摘要
针对目前结构健康监测采集数据时间长、海量数据占用内存高、运输成本高以及数据利用率低的问题,文章提出了将压缩感知技术和深度学习网络相结合的结构损伤识别方法,对有限元模拟得到的加速度响应信号用压缩感知技术进行处理,得到了压缩信号和重构信号;针对传统重构方法重构效果不佳的问题,提出一种基于全变分正则化(TV,Total Variation)的交替方向乘子重构算法(ADMM)来改善重构性能,建立了基于卷积神经网络(CNN)的简单损伤识别模型和基于此在这个模型上同时引入了数据增强技术和多头自注意力机制(Multi-Head Attention)的MHA-CNN~+模型。结果表明:信号重构中,TV-ADMM算法能以很高的精确率重构出原始信号;在损伤识别任务中,原始信号和重构信号在CNN模型上都能够以非常高的精确率完成损伤识别任务,但对于压缩信号则表现不佳,而压缩信号在MHA-CNN+模型上则能以很高准确完成损伤识别任务。
Aiming to address the challenges of lengthy data collection, high memory consumption, elevated transportation costs, and low data utilization in current structural health monitoring, a novel approach for structural damage recognition is proposed by integrating compressive sensing and deep learning networks. Initially, acceleration response signals obtained from finite element simulations are processed using compressive sensing, yielding compressed and reconstructed signals. To address suboptimal reconstruction from conventional methods, an Alternating Direction Method of Multipliers(ADMM) with Total Variation(TV) regularization is introduced to enhance reconstruction performance. The TV-ADMM algorithm demonstrates highly accurate signal reconstruction. Subsequently, a convolutional neural network(CNN) based damage identification model is established, showing excellent accuracy for both original and reconstructed signals, but limited performance for compressed signals. To overcome this, data augmentation and Multi-Head Attention mechanisms are integrated into the model(MHA-CNN~+),resulting in effective damage recognition using improved compressed signals.
作者
赵雄鹰
任宜春
ZHAO Xiongying;REN Yichun(School of Civil Engineering,Changsha University of Scince Technology,Changsha 410114,Hunan,China)
出处
《工程建设》
2024年第8期7-14,共8页
Engineering Construction
关键词
压缩感知
深度学习
损伤识别
健康监测
compressed sensing
deep learning
damage identification
health monitorin
