| Strong motion data plays an important role in earthquake research,structural seismic design and ground deformation during earthquakes,and has always provided important reference value for post earthquake disaster assessment.With the continuous development of seismology,artificial neural network has been gradually applied in various studies.At the same time,seismology has gradually combined traditional methods with artificial neural network,and the most important application is to use it to classify seismic data.At present,the baseline correction of near-field strong motion records has always been the top priority of seismic engineering.It has important research significance for the study of ground faults after earthquakes and the collapse causes of buildings or highway structures（such as long-span bridges）.However,there are many errors in the strong motion data directly obtained,which has caused great trouble in obtaining the movement trend of near-field strong motion.However,China has many earthquake prone areas,so it is urgent to carry out further research on baseline correction.However,the current baseline correction methods are subjective and empirical.The correction results depend on the verification of GPS data,and most baseline correction methods can not achieve automatic correction.Combined with the traditional two parameter baseline correction method,this paper proposes a baseline correction method based on LSTM（long-term and short-term memory model）using artificial neural network,which can obtain more accurate permanent displacement results without GPS data verification,and realize automatic data processing and correction.The thesis mainly completed the following work:1.The research status of baseline correction methods for near-field strong motion records at home and abroad is summarized,and the problems to be solved in this paper are put forward according to the advantages and disadvantages of the existing methods.2.The basic concept of neural network,the structure,principle and corresponding formula of LSTM model are described in detail.The trained LSTM model is used to identify the strong motion data and realize the classification of near-field strong motion records,so that it can distinguish whether the baseline of strong motion records is offset.Finally,this paper uses Python program and related libraries to deploy and implement the model,and gives the model evaluation parameters,accuracy and related results according to the training results.3.This paper analyzes the error sources of baseline offset of strong motion,summarizes the existing methods for correcting low-frequency errors,and analyzes the comparison of advantages and disadvantages between various methods.In view of the disadvantages of relying on researchers’subjective selection of parameters and being unable to automate iteration,combining the traditional two parameter fitting method with the artificial neural network method,a baseline correction method based on LSTM model is proposed,which is used to judge the zeroing degree,The two parameters t1and t2are automatically adjusted to ensure the accuracy of the correction results.4.The baseline correction method in this paper is used to correct some typical near-field strong motion data obtained in the Chi Chi earthquake and Wenchuan earthquake.The permanent displacement data calculated after correction is compared with the adjacent GPS station data at 5km.The permanent displacement is calculated and compared with the existing mainstream method.The error percentage of this method with GPS station data and other mainstream methods is calculated.The results show that,This method can effectively correct strong motion records and obtain relatively accurate permanent displacement,so as to verify the effectiveness of this method. |
