适用于区域建筑群实时震害模拟的LSTM-FC组合深度网络模型研究
Combined LSTM-FC deep network modeling for real-time earthquake hazard simulation of regional building complexes
孙海 1徐晓君 2邢启航 2张孝伟 2姜慧 3阮雪景4
作者信息
- 1. 中国海洋大学 工程学院,山东 青岛 266100;中国海洋大学 海洋生态与环境教育部重点实验室,山东 青岛 266100
- 2. 中国海洋大学 工程学院,山东 青岛 266100
- 3. 广东省地震局 地震监测和减灾技术重点实验室,广东 广州 510070
- 4. 青岛农业大学 建筑工程学院,山东 青岛 266109
- 折叠
摘要
建筑物破坏在地震灾害中往往会导致巨大损失,对城市建筑群进行灾前灾时的震害预测具有重要意义.传统BP(back propagation)网络和CNN(convolutional neural networks)网络等人工智能方法在进行震害预测时多集中于提取建筑物信息.然而,这些方法在处理地震波的时序数据方面有所不足,导致其在整合和分析对地震灾害预测至关重要的时序相关因素时效果有限.因此,本文提出一种耦合LSTM(long short-term memory)和FC(fully connected)神经网络的震害预测方法.LSTM网络擅长处理具有时间序列特性的地震波信息,能够捕捉和分析随时间变化的地震波动模式.同时,全连接网络可用于综合分析所有相关的震害因子.通过对云浮地区 265 栋典型钢混建筑进行指标量化并确定输入指标(震害影响因子)和输出指标(震害指数),利用LSTM-FC组合深度网络、CNN网络和BP网络模型对数据进行训练并优化.通过将LSTM-FC网络模型的预测结果与弹塑性时程分析比较,发现该模型在拟合效果和精度方面优于传统的BP和CNN模型.拟合效果提升了 36.8%和10.6%,精度分别提升了 77.6%和 91.7%,表明LSTM-FC网络在地震损害预测上更为有效.同时,将该方法应用于广东省云浮市钢混结构群震害预测,构建的易损性矩阵与华南地区的易损性矩阵均值进行了对比,显示误差相对较小,说明该模型不仅理论上可行,在实际应用中也能表现出较高的准确性和有效性.
Abstract
Building damage often leads to huge losses in earthquake disasters,and it is of great significance to predict the earthquake damage of urban building complexes before and during disasters.Artificial intelligence methods such as traditional BP(back propagation)networks and CNN(convolutional neural networks)networks mostly focus on extracting building information when performing earthquake damage prediction.However,these methods are deficient in processing time-series data of seismic waves,which leads to their limited effectiveness in integrating and analyzing time-series correlations that are crucial for seismic hazard prediction.Therefore,this paper proposes a coupled LSTM(long short-term memory)and FC(fully connected)neural network for earthquake hazard prediction.The LSTM network excels in processing seismic wave information with time-series characteristics,and is able to capture and analyze seismic fluctuation patterns over time.Meanwhile,the Fully Connected network can be used to synthesize and analyze all relevant seismic hazard factors.By quantifying the metrics and determining the input metrics(seismic impact factors)and output metrics(seismic indices)for 265 typical steel-concrete buildings in Yunfu area,the data was trained and optimized by using the LSTM-FC combined deep network,CNN network and BP network model.By comparing the prediction results of the LSTM-FC network model with the elasto-plastic time-range analysis,it is found that the model outperforms the traditional BP and CNN models in terms of fitting effect and accuracy.The fitting effect was improved by 36.8%and 10.6%,and the accuracy improved by 77.6%and 91.7%,respectively,indicating that the LSTM-FC network is more effective in earthquake damage prediction.Meanwhile,the method is applied to the earthquake damage prediction of steel-concrete structural clusters in Yunfu City,Guangdong Province,and the constructed susceptibility matrix is compared with the mean value of the susceptibility matrix in South China,which shows that the error is relatively small,indicating that the model is not only theoretically feasible,but also can show high accuracy and effectiveness in practical applications.
关键词
震害预测/LSTM网络/全连接网络/钢混建筑物Key words
earthquake damage prediction/LSTM network/fully connected network/steel and concrete building引用本文复制引用
基金项目
国家自然科学基金项目(41906185)
国家自然科学基金项目(52071307)
国家自然科学基金项目(U1901602-05)
青岛市自然科学基金项目(23-2-1-61-zyydjch)
山东省科技重大专项(2020CXGC010702)
&&()
出版年
2024
NETL
NSTL
万方数据