First-arrival Pickup Analysis of Seismic Waves Based on Improved U-Net Model
In order to improve the accuracy of seismic wave picking,an optimized U-Net model is designed based on analyzing the working mechanism of U-Net model.The model consists of bridging,coding and decoding.Convolutional neural network is used for classification,and pooling operation is used to extract low-level feature parameters.In selecting pooling operation mode,the picking accuracy is significantly reduced,and the convolution mode is selected as the 4-step.According to the evalu-ation indexes,the processing performance of STA/LTA,U-Net and improved U-Net are judged.The results show that the im-proved U-Net model achieves 99％accuracy by Tensorflow and Keras to train the network.The epcho is increased from 10 to 50,and the accuracy of P-wave identification is increased from 72％to 94％,and the accuracy of S-wave identification is in-creased from 63％to 94％.The initial loss is 10.25,and reduced to 0.15 after training to 100 epcho.Compared with the tradi-tional seismic signal processing method,the improved U-Net achieves better precision,recall and F1 Score index.When the convolutional neural network is used for processing,the feature yield of seismic waveform can be extracted by the convolutional operation,which has stronger adaptability.
万方数据
维普
