Research On Post-Earthquake Safety Assesment Method Of Reinforced Concrete Buildings Using Deep Learning

It is extremely important to quickly carry out the safety assessment of buildings in the post-earthquake emergency.However,the current used one-stage assessment method in China has many problems in assessment time and workload,which is difficult to meet the requirements of cities for rapid post-earthquake safety assessment of structures.This study summarizes the post-earthquake safety assessment methods of structures,analyzes the latest relevant research progress,puts forward a new post-earthquake safety assessment method of RC structures based on deep learning,and carries out a series of exploratory research work on structural dynamic system simulation,long-series seismic signals,small sample data,transfer learning,multi-modal information fusion and so on.The main study contents and conclusions are as follows:1.The equivalent conditions between recurrent neural network and structural dynamic system are proposed,and the theoretical derivation and example analysis are carried out.The results show that for the linear elastic system,the recurrent neural network can be equivalent to the structural dynamic system under the equivalent condition.For the elastic-plastic system,the structural dynamic system is time-varying and the recurrent neural network is time-invariant,so they should not be directly equivalent.However,according to the general approximation theorem of the recurrent neural network and the analysis of examples,the recurrent neural network still has the ability to simulate the elastic-plastic dynamic system.2.The post-earthquake safety assessment method of structure based on ground motion records is proposed.The time-domain information and frequency-domain information of ground motions are studied in detail.The prediction model of recurrent neural network and fully connected neural network are established respectively,and the example analysis is carried out.The results show that the gradient vanishing problem is easy to occur when the recurrent neural network processes the long-term ground motion records.The fully connected neural network using frequency-domain information as input can effectively avoid the gradient disappearance problem,and achieve better classification performance in the data set.3.To avoid a large number of training sets,a structural safety assessment method based on transfer learning under small dataset is proposed.Based on the proposed prediction model of fully connected neural network,the transfer learning method is adopted to transfer and reuse the classification knowledge of the original structure in the target classification task.The results of example analysis show that the transfer method can be effectively applied to different target structures,effectively reduce the number of training sets and enhance the generalization ability of prediction model.4.The post-earthquake safety grade assessment method of RC structures based on multi-modal input is proposed,and a neural network prediction model based on multimodal input is built.Using the proposed structural information index and the acceleration response spectrum of ground motions,the data set required for training the model is established based on a large number of building analysis reports;The results show that the proposed assessment method can achieve 90.0% classification accuracy in the test dataset,which meets the expected accuracy requirements of structural post earthquake safety assessment.