Research On Monitoring Acceleration Record-based Rapid Structural Seismic Damage Assessment Method

At present,the earthquake disaster risk in China is becoming more and more serious,which has seriously affected the social stability and sustainable economic development.Since the Ms8.0 Wenchuan Earthquake in 2008,a large number of earthquake monitoring arrays and a small number of structural monitoring arrays have been established in areas with high earthquake disaster risk to obtain local monitoring seismic records and corresponding structural response data in post-earthquake.Carrying out rapid seismic damage assessment of building structures based on these monitoring data can not only improve the use efficiency of monitoring data,but also improve the accuracy and efficiency of seismic damage assessment.However,most of the current research on assessing structural seismic damage based on monitoring data focuses on the identification and detection of weak nonlinear damage in the field of structural health monitoring,rather than assessing the strong nonlinear damage state of structures in postearthquake.In this thesis,several building structures with different types are adopted as the research object.Based on data collected from the earthquake and building monitoring array,using the deep learning algorithm to carry out research on the rapid assessment of structural seismic damage.The thesis mainly completed the research as following:1.Since conducting extensive nonlinear time-history response analysis of building structures requires a lot of time,this thesis proposes a method for rapidly predicting the structural seismic response under earthquake excitation based on the improved Gated Recurrent Unit(GRU)model.The proposed method could accurately predict the structural roof acceleration response from the ground motion,with higher prediction accuracy compared to other deep learning methods.The method also improves the efficiency in predicting the roof acceleration response of a 12-story framed shear wall structure by more than 750 times compared with conventional nonlinear time history analysis.Even for high-rise buildings,the goal of second-level prediction can be achieved,which provides the data basis for the training of deep learning models in the following research.2.Since the strong noise data are mixed in the structural response records collected by the array in actual engineering contain,which brings a very serious impact on the prediction accuracy of structural seismic damage.A denoising method for structural floor acceleration records based on improved generative adversarial network(GAN)model-is proposed to suppress the strong noise from the original acceleration records effectively.The proposed denoising method has higher denoising accuracy compared to traditional denoising methods and other deep learning methods.3.Since the cumbersome process,complicated parameter adjustment,and series baseline drift involved in the traditional integration method,a sequence-based displacement prediction method is proposed based on an improved long short-term memory neural network model(LSTM).The proposed method has higher prediction accuracy and stronger noise immunity compared to other deep learning methods.In addition,a shaking table test also proves that the method could process arbitrary length data.The seismic damage level assessment of multi-story buildings is achieved by combining the peak roof displacement with the first mode of vibration.4.Since it is difficult to directly invert structural dynamic characteristic parameters based on sparse array data and use the previous research to evaluate the seismic damage state of high-rise structures,a novel structural damage assessment method based on residual dilated convolutional neural network(ResNet_Di)is proposed.The proposed method achieves an average prediction accuracy of more than 85%,which is higher than the existing methods that only use ground shaking as input.The prediction performance of the proposed ResNet_Di model outperforms other deep learning models.5.Since it is hard to accurately assess the seismic damage of undeployed array structures in the area,a transfer learning method based on domain adaptive is proposed.The knowledge of seismic damage assessment on deployed array structures will be transferred to undeployed array structures by using generative adversarial and dual classifier optimization.The results showed that,compared with directly using the model trained by deployed array structures to predict the seismic damage state of undeployed array structures,the model after transfer learning can predict the seismic damage state of undeployed array structures with higher accuracy.The proposed dual classifier training method significantly improves the prediction accuracy of the model compared to the traditional domain adaptive method.