Unsupervised Deep-learning Inversion Method For Seismic Velocity And Its Application In Tunnel Forward-prospecting

With the construction of major infrastructure projects in China gradually shifting to deep underground with complex geological structures,geo-hazards such as water and mud inrush and collapse occur frequently during tunnel construction.Carrying out the geological forward-prospecting during tunnel construction,and identifying the adverse geology ahead of the tunneling face in advance,are effective means to ensure the safety of tunnel construction.The seismic method has become one of the most commonly used tunnel geological prospecting methods due to its interface sensitivity and long detection distance.The accurate computation of seismic velocity is a key challenge to the positioning accuracy of seismic forward-prospecting in tunnels,and research is needed to develop an efficient method for calculating seismic velocity distribution applicable to the tunnel environment.Full waveform inversion(FWI)is a representative of traditional seismic velocity inversion methods,but it has problems such as high computation demend and strong inversion non-uniqueness.Especially in the tunnel environment,it faces problems such as small offset and little detection data,which limit its application effect in tunnels.Deep-learning-based seismic velocity inversion methods,which have emerged in recent years,have the advantage of good inversion effect and high computation speed,and have attracted widespread attention from scholars.The existing deep-learning velocity inversion methods rely on a large number of real velocity models and belong to the category of supervised learning.However,it is often difficult to obtain the actual velocity distribution in the real world,limiting its application in practical engineering.Therefore,there is an urgent need to study effective means to improve the inversion effect of neural networks,break the dependence of the seismic velocity inversion network on real velocity models,and realize unsupervised deep-learning inversion of seismic velocity and its application in practical engineering.In order to address the problem that the existing seismic velocity deep-learning inversion methods rely on the real velocity models and are difficult to be applied in practical engineering,this paper uses theoretical analysis,numerical experiments and field tests,and proposes the idea of "integrating background velocity information to improve the inversion effect of the network and introducing physicas to replace the real velocity models".This seismic velocity unsupervised deep-learning inversion is driven by both data mining and physical laws.The research was carried out on the construction method for large samples of seismic background velocity models,the integration network of large samples of seismic background velocity models,and the unsupervised learning inversion method for seismic velocity driven by physical laws,and finally resulted in the seismic velocity unsupervised deep-learning inversion method and proposed a network named ResIFNet.This method is further applied to geological forward-prospecting in tunnels to realize the unsupervised deep-learning inversion of seismic data.Finally,the feasibility and effectiveness of the proposed method in this paper are verified by numerical simulations and field tests.The main research work and results of this paper are as follows.(1)The generation method for large samples of seismic background velocity models.In response to the demand for background velocity models in seismic velocity inversion,on the basis of Bayesian inversion theory and the hierarchical invertible neural transport architecture,a background velocity building network named BVNet is designed based on seismic migration imaging result.This network has the function of randomly sampling the posterior probability distribution of the velocity model to obtain a sample of seismic background velocity model from seismic data.On this basis,the model samples are further smoothed to give a reliable seismic background velocity model,which provides reliable training data for the subsequent unsupervised deep-learning inversion of seismic velocities.(2)Network integration method for large samples of seismic background velocity model.In order to introduce background velocity information into the deep-learning inversion network and reduce the learning difficulty of the network for velocity inversion mapping,this paper designs ResInvNet on the basis of the background velocity model generated by the invertible neural networks.ResInvNet is designed based on the idea of linearization of non-linear problems and it is a velocity inversion network for multi-scale feature extraction and integration of seismic observation data residuals and background velocity models.This network can effectively increase the size of the training dataset and improve the effectiveness of velocity inversion,laying a network foundation for the subsequent research on unsupervised deep-learning inversion of seismic velocities.(3)A physics-driven unsupervised deep-learning inversion method for seismic velocity.In order to address the dependence issue of the existing deep-learning velocity inversion methods on real velocity models,a physics-driven velocity inversion strategy based on seismic wavefield propagation is proposed,and a seismic forwarding network named FwdNet,is studied and constructed.Combined with the ResInvNet,a physics-driven unsupervised deep-learning inversion network for seismic velocity,ResIFNet,is constructed.This method is developed to provide a feasible way for the application of deep-learning inversion of seismic velocity in the real data situation.(4)Implementation and characteristics of unsupervised learning inversion in tunnel seismic forward-prospecting.The above-mentioned unsupervised learning inversion method for seismic velocity is introduced into geological forward-propsecting in tunnels.In order to address the difficulty of applying the conventional deep neural network effectively due to the special observation mode and small offset distance in tunnel seismic detection,the velocity inversion network is optimized and the loss function is improved based on the idea of multi-scale inversion,resulting in an observation-adaptive unsupervised deep-learning inversion method for seismic detection in tunnels.This method has good applicability to more complex seismic data with different observation mode and noise contamination,and effectively improves the effect of seismic velocity inversion in tunnels.On the basis of the above research,a field test was carried out in the Zhujiang Delta Water Resources Allocation project and Shandong Binlai highway tunnels.The results reveal the geological structures and velocity distribution in front of the tunnel accurately,which verifies the feasibility and effectiveness of the method proposed in this paper.