| Seismology,an essential method in geophysical research,focuses on the analysis of seismic wave data.Seismic waves,generated by earthquakes or ambient noise,propagate through the Earth’s interior.The information revealed by these seismic waves can be utilized to investigate velocity structures at different scales beneath the Earth’s surface.By observing the propagation of seismic waves in the mantle,inner core,and outer core,different seismic phases provide valuable insights into the Earth’s deep structures.Additionally,seismic wave data can be used to study the mechanisms of earthquakes and lessen the impact of seismic hazards.In recent years,the advancement of seismic observation technologies,including transportable arrays(TA),intelligent seismic sensors,and distributed fiber acoustic sensors,has expanded the magnitude of seismic observation data.Simultaneously,breakthroughs in seismic ambient noise cross-correlation techniques have overcome temporal and spatial limitations in seismic events and observations,significantly improving the availability of seismic observation data.The development of these technologies and methods has not only significantly increased the quantity of seismic observation data but has also posed new challenges to traditional seismic signal data processing and analysis workflows.Thus,it is crucial to develop new automated methods matched to specific applications and requirements to enhance the efficiency and accuracy of data processing.In the field of machine learning,deep neural networks,along with increased computational performance,have gradually evolved.The introduction of convolutional neural networks has effectively improved predictive performance and training efficiency on high-dimensional data.In recent years,these methods have been widely applied in digital image and audio processing fields.Therefore,by combining the abundant data obtained from seismic observations with advanced machine learning techniques,it is possible to further develop efficient and accurate automated data processing methods,providing sufficient tools for seismological research and advancing the analysis and interpretation of seismic data,thus enhancing the understanding of seismic processes and the Earth’s internal structures.Surface wave dispersion inversion has been extensively applied for imaging subsurface shear wave velocity structures.Recently developed frequency-Bessel transform(F-J)methods enable the extraction of higher mode dispersion curves with distinct characteristics from noise cross-correlation functions and earthquake events’ signals.To further invert subsurface velocity structures based on surface wave dispersion curves,it is necessary to extract dispersion curve features from two-dimensional dispersion spectra.However,conventional manual and semi-automatic methods are time-consuming and lack unified standards,which limit the efficiency and progress of surface wave dispersion inversion and subsurface imaging.In this study,we first apply a convolutional auto-encoder to extract dispersion features and subsequently separate different modes of dispersion curves by an unsupervised clustering method.The application and validation of this method on various regions and scales demonstrate its capability in accurately extracting dispersion curves.This method can provide an efficient and precise approach for dispersion curve picking in surface wave inversion studies.Acceleration data observed from earthquake events are valuable data resources for earthquake research and seismic engineering.Due to its directly measuring ground acceleration without being limited by the recording range of seismometers,the accelerogram can accurately record ground motions associated with larger earthquake magnitudes.However,sampling errors in the observed data and the impact of strong ground motions on the observation instruments often lead to significant baseline drift during the double times of trapezoidal integrations that convert acceleration data into displacement data.In this study,we propose a new deep convolutional neural network,TraceNet,trained on a synthetic dataset generated by a newly proposed baseline simulating method,which is based on conventional multi-segment baseline correction methods.TraceNet can provide effective,stable,and rapid baseline correction for observed acceleration data.TraceNet is first tested and validated in synthetic data,including synthetic noise and synthetic baselines.The application of observed data demonstrates that the method not only effectively corrects baseline drifts in field data but also enables the tracking of co-seismic displacements from the recovered displacement data,providing a new observational reference for seismological research.Distinct boundary features exist in the deep structures of the Earth,such as the coremantle boundary(CMB)and inner-outer core boundary.Initially,these features were discovered and detected through a combination of theoretical calculations and seismic phase observations.In recent years,extensive observations based on TA have suggested the uneven spatial distributions of these boundaries,including undulations in different scales,ultra-low velocity zones,and rigid layers,and so on.The direct and reflected waves traveling through these boundaries can sample and detect the velocity structures near the boundaries.PmKP core phase(the m-1 reflection at the CMB)can provide additional sampling ranges the of CMB when the seismic observation is spatially limited by ground-based stations.In this study,we propose a novel method for seismic multi-station record sampling,which samples discrete seismic station observation data onto a uniform spatiotemporal grid.By further applying a deep convolutional neural network on the gridded seismic samples,the feature of the PmKP core seismic phase can be detected and measured in ranges of time and epicenter distance.The automatic verification of PmKP phase signals is accomplished through slowness calculations and convolutional neural networks.This method effectively improves the detection efficiency of different PmKP phase signals in a large volume of seismic event data and can expand the existing seismic phase observation inventories and CMB sampling range.Furthermore,this method demonstrates similar detection performance for other core seismic phases,such as the PKiPK phase,indicating its potential application value.This study introduces three machine learning methods applied in different seismic data processing fields.By designing corresponding machine learning methods and training set construction methods for different types and features of seismic data,automatic,efficient,and stable processing of seismic observation data can be implemented.These methods can provide effective tools for seismological research,extracting valuable feature data from a large amount of observation data.The application of these methods can promote seismic studies and may support a deeper understanding of seismic processes and the Earth’s interior. |
