| China is a country prone to earthquakes.Frequent earthquake disasters have brought huge casualties and economic losses to human society.At the same time,the rapid development of the construction scale of the station has also led to a large number of collected ground motion records that need to be processed urgently.However,the quality of ground motion records varies,and the current methods are difficult to process them quickly and accurately.Ground motion data processing generally consists of two parts: abnormal waveform identification and filtering.The current abnormal waveform identification method is mainly based on the threshold to judge a single abnormal pattern,which is inefficient and depends on experience,and it is difficult to classify multiple abnormal patterns at the same time.And current filtering methods all rely on the filter,and the accuracy and stability of the filtering results depend heavily on the determination of parameters such as cut-off frequency,filter order,and ripple coefficient.In this paper,the abnormal waveform recognition task is converted into an image classification problem,and multivariate pattern recognition is performed by learning waveform features.In order to effectively separate the effective signal and noise in the noisy recording,a filtering model based on GAN is proposed.The main research contents are as follows:(1)Firstly,with summarizing the waveform characteristics of ground motion acceleration records,the abnormal waveform types and discrimination criteria are determined,and then establish a database of abnormal ground motion waveforms based on STEAD data.By normalizing and visualizing ground motion recordings,the task of identifying abnormal waveforms is transformed into an image classification problem.Using the Res Net-50 residual network learning image features,the network hyperparameters are debugged through the training results of different working conditions,and the multivariate classification problem of abnormal waveforms of ground motion records is solved.The final average overall accuracy rate reaches 94.5%.Compared with the existing abnormal waveform identification methods,the method in this paper can realize the rapid and accurate identification and classification of various abnormal patterns,and provide data basis for the filtering processing of subsequent ground motion records.(2)Firstly,screen the high signal-to-noise ratio ground motion acceleration records in the STEAD data set and the noise acceleration records in the INSTANCE data set,perform random amplitude modulation stacking,establish a noisy ground motion database,and perform preprocessing on all signals,such as resampling,zero-filling,integration,wavelet packet transformation,etc.Based on GAN and U-Net network,a filtering processing network is constructed to learn the amplitude,frequency and waveform shape of ground motion signal and noise signal in the time-frequency domain,so as to realize the separation of signal and noise.(3)In order to illustrate the generalization and stability of the filtering model in this paper,a noisy ground motion data set with different regions and different frequency distributions is established,and the model is used for filtering processing.Firstly,a synthetic noisy ground motion record dataset is established based on the Italian INSTANCE dataset,and then a Ki K-Net real noisy ground motion record dataset is established,and the same preprocessing is performed on all signals.The filtering model in this paper is applied to the above data set,and the filtering effect of the model is evaluated by multiple indicators such as the amplitude and waveform shape of acceleration,velocity and displacement time history.The filtering results are compared with the results obtained by the current filtering method,and the results show that the proposed method has better stability and robustness. |
