MICROSHOCK MAGNITUDE PREDICTION METHOD BASED ON TRANSFER LEARNING OF RESIDUAL SHRINKAGE NETWORKS

Abstract

In view of missing seismic data markers and magnitude estimation and more accurate and effective microshock magnitude estimation, a classification model based on residual shrinkage network transfer learning is proposed to realize the prediction of microshock magnitude. First, the microshock waveform picture was preprocessed and used as the model input, and then the microshock waveform picture data set was divided. With the residual network pretraining model weight parameters trained by ImageNet, the CBATM attention module and soft thresholding were introduced to build the residual shrinkage network model. Finally, this model is used for magnitude classification research to realize microshock magnitude prediction, and it is verified and compared with other commonly used neural network models. In order to better classify the magnitude, the data and images of magnitude 0-3 in STanford EArthquake Dataset (STEAD) seismic data set were selected as the training sample. The results show that the prediction accuracy of the model with error in the range of positive or negative 0.2 can reach 90%, and the prediction accuracy of the error in the range of positive or negative 0.3 can reach 97.6%. The model is more stable and reliable, verifying that the accuracy of microshock magnitude prediction using residual shrinkage network transfer learning is higher and the effectiveness of the attention mechanism module.