Research On Seismic Signal Classification Based On Improved Deep Residual Network

The accurate classification of seismic waveforms is of great importance for the cleaning of seismic catalogs,real-time monitoring of earthquakes and earthquake early warning,as well as contributing to further research in the field of seismology.Traditional seismic signal classification is implemented by selecting representative and suitable seismic feature information,focusing only on the selected feature information and ignoring more possible information implied in the original recorded seismic waveform.Deep neural networks,with the advantages of automatic feature extraction and self-learning,have been remarkably successful in many classification problems,especially in image classification and recognition.Besides,it might be very probable to get higher precision accuracy in classifying seismic waveforms by mining the rich information contained in the original recorded seismic waveforms and studying the differences in seismic waveform signals from different sources.In this paper,some deep neural networks are extensively explored to build a classification model for natural earthquake and artificial explosion waveforms.The main research contents are as follows:(1)An attentional residual network model for seismic signal classification and recognition is designed.Traditional algorithms for seismic classification rely on waveform feature extraction,which has the disadvantage of being time-consuming,complex,and incorporating more human subjectivity in the selection of suitable features.To address these drawbacks,firstly,QRes Net(Quake Residual Network),which is suitable for seismic signal recognition,has been designed.This network model requires a small number of parameters and can be used for autonomous feature extraction.Secondly,to improve the ability of the network model to focus on key information,an attention residual network suitable for classification and recognition of seismic signals has been designed,which is a combination of Effective Channel Attention in QRes Net.This network model further enhances the ability of the network model to learn features by taking advantage of the attention mechanism to focus on key information and constitutes a simpler and more complete method for seismic signal classification.Through experimental comparisons between our method and traditional machine learning methods,the results show that our network model is more capable of portraying the features and demonstrates better classification results.(2)Based on the QRes Net,an improved deep residual shrinkage network(RSNN)is proposed.Firstly,to address the problem of the low signal-to-noise ratio of seismic signals,a shrinkage block is introduced in QRes Net,which can remove a feature that is outside the range of values according to the weight distribution to achieve noise reduction processing of the signal.Secondly,a one-dimensional convolution with an adaptively determined convolution kernel size is used to replace the two fully-connected layers in the original systolic module that require dimensionality reduction.It enables cross-channel information interaction without dimensionality reduction,thus solving the problem that the original shrinkage block requires a dimensionality reduction operation resulting in partial information loss.Then,the global average pooling in the original shrinkage block is replaced by a simple non-local block that can obtain relationships between global contexts,and the relationships obtained can be incorporated into subsequent cross-channel information interactions.This solves the problem that the original shrinkage block only focuses on information in the channel dimension and ignores the information in the spatial dimension.Finally,for the problem of limited or even small sample size of seismic waveform data corresponding to the same event,the global maximum pooling in the original network model is replaced by spatial pyramid pooling,which allows the network to obtain as much information as possible for classification by pooling the feature maps at different scales.Through ablation experiments and comparison with other neural network-based seismic classification network models,the results show that RSNN can further improve the accuracy of recognition of seismic series-time signals and obtain higher classification accuracy while simplifying the seismic signal classification step by eliminating the need to extract features from the vibration waveform before classification and recognition.