Research On Inversion Method Of High Frequency Surface Wave Based On Deep Learning

Near surface underground space plays an important role in human development.Rayleigh wave is an important technical means to obtain information on underground structure,and has received extensive attention in recent years.The investigation of nearsurface velocity structure mainly obtains the underground velocity structure by inversion of the Rayleigh wave dispersion curve,which has the advantages of high resolution and fast speed.However,in the application process of this technology,there are problems such as time-consuming manual extraction of dispersion curves and many influencing factors of nonlinear inversion.If we can use the newly developed artificial intelligence technology,especially deep learning and other technologies to realize the intelligent transformation of Rayleigh wave exploration technology,it can undoubtedly promote the further development of this technology and has important practical significance.Rayleigh wave exploration technology is mainly divided into two steps: dispersion curve extraction and dispersion curve inversion.Realizing accurate dispersion curve extraction and inversion,and obtaining reliable learning samples are important prerequisites for the intelligent transformation of this technology.But the dispersion curve inversion has problems of inversion multi-solution and stability.In addition,manually picking up the dispersion curve has the problems of low efficiency and subjective factors,time-consuming and laborious,and is not suitable for the operation of large amounts of data.This article has conducted research on the above two aspects to provide good and accurate training data for deep learning,and on this basis,the inversion method of deep learning surface wave dispersion curve is realized.The main work content and contributions of this article are:(1)Based on simulated annealing algorithm,a new dispersion curve inversion method based on differential evolution simulated annealing is proposed.The new method uses the idea of block coordinate descent to transform the multi-parameter highdimensional surface wave inversion problem into multiple single-parameter lowdimensional inverse problems,which improves the accuracy of the surface wave inversion results.At the same time,the algorithm introduces a differential evolution algorithm to replace the termination temperature of the simulated annealing with an error,which realizes the direct control of the simulated annealing to the inversion error,thereby improving the stability of the inversion.(2)Aiming at the time-consuming and labor-intensive problem of manually picking up dispersion curves,this article develops an automatic picking technique of Rayleigh wave base-order dispersion curves based on Unet.This method can accurately extract the dispersion curve from the pictures with better signal to noise.For the dispersion curve with strong noise interference,this article uses multiple Unet segmentation schemes,which effectively suppresses the noise interference and improves the accuracy of the segmentation results.(3)The intelligent inversion method of the Rayleigh wave dispersion curve is realized based on the long and short time memory network(LSTM).Using the accurate results of the above new algorithm inversion and the dispersion curve extracted by the Unet as training samples for the LSTM,fast and intelligent inversion model training of high-frequency Rayleigh waves is achieved.The algorithm does not need to manually pick up the dispersion curve multiple times,and does not need to perform multiple iterations of the nonlinear inversion objective function,so it is expected to jump out of the multi-solution misunderstanding and avoid human interference,thereby improving the practicality of the method.Through the research and application of the above methods,this article hopes to provide help for the development of key technologies in the surface wave inversion process,and effectively promote the development of surface wave inversion technology.