基于深度学习代理模型和改进CMAES方法的非对称框架损伤识别
摘要
将基于深度学习的代理模型与改进协方差矩阵自适应进化策略(CMAES)算法相结合,本文提出一种用于损伤识别的时域变分贝叶斯框架.所构建的代理模型由并行卷积神经网络与编码器-解码器长短期记忆网络组成,并引入残差连接、自注意力机制以及综合损失函数.该模型能够实现对结构动力响应的高精度预测,均方误差为0.000 12,R2达到0.998 00.此外,提出一种由深度学习代理模型梯度信息引导的改进CMAES算法,以加速局部收敛.基于非对称框架结构的实验验证表明,该方法能够同时识别损伤位置与损伤程度,并对结果不确定性进行量化.同时,该方法显著提升了计算效率,为结构健康监测提供了可行的技术基础.
This study proposes a time-domain variational Bayesian framework for damage detection that integrates a deep learning-based surrogate model with an improved cova-riance matrix adaptation evolution strategy(CMAES)algo-rithm.The surrogate model comprises parallel convolu-tional neural networks followed by an encoder-decoder long short-term memory module with residual connections,self-attention,and a comprehensive loss function.The proposed approach can achieve accurate dynamic response prediction with a remarkable mean squared error of 0.000 12 and R2 of 0.998 00.Furthermore,an improved CMAES algorithm guided by gradient information derived from the deep-learn-ing surrogate model is developed to accelerate local conver-gence.The proposed method enables the detection of the lo-cation and level of damage while quantifying the associated uncertainties,as validated by experimental case studies on an asymmetrical frame.Furthermore,it significantly en-hances computational efficiency,offering a viable technical foundation for structural health monitoring.
年,卷(期):
2026, 42(3)
分类号:
TU317
O212.8
机标分类号:
TP391
TU375.4
TP18
在线出版日期:
2026年9月24日