Deep Learning Inversion Method For Seismic Velocity In Tunnel Seismic Forwardprospecting Based On Multimodal Fusion

Tunnel construction process often faces complex geological conditions,resulting in frequent occurrence of geological disasters such as collapse,water and mud inrush,which seriously affects the safe and efficient construction of tunnels.In the face of faults and fracture zones and other adverse geological structures,tunnel seismic wave method has become one of the commonly used methods of tunnel forward prospecting due to its advantages of long detection distance and interface sensitivity.Among them,the accurate estimation of velocity distribution in front of the tunnel is the key prerequisite for the accurate positioning and imaging of the adverse geological structures,which has become a difficult problem to be solved by tunnel seismic forward-prospecting at present.Due to the narrow observation space and construction noise inside the tunnel,the effective information extracted from the tunnel seismic forward-prospecting data is often insufficient,which makes the velocity inversion have serious nonlinearity and discomfort.In recent years,with the rapid development of deep learning algorithms,deep neural networks have been used to solve the geophysical inversion optimisation problem by virtue of their strong high-dimensional optimisation and simulation of nonlinear mapping,which provides a new solution for seismic velocity inversion in tunnels.At present,a deep learning inversion method for velocity,represented by the dual-driven inversion of wave propagation physical laws and data mining,has been formed,and it has achieved better results on the tunnel seismic synthetic data.However,for real tunnel engineering applications,the existing deep learning inversion methods still have the following problems that need to be solved:1)the deep learning velocity inversion method usually relies on the synthetic data to train the network,and the synthetic data are so different from the real observed data that it is difficult to be used for the real observed data directly;2)due to the influence of the tunnel’s narrow observation,there are aberrations in the gradient of the velocity inversion,which leads to an inaccurate inversion of the tunnel velocity structure;3)Insufficient valid information in the observed tunnel data,which makes the velocity inversion more unsuitable,resulting in false anomalies for velocity inversion.In summary,there is an urgent need to improve the deep learning tunnel seismic velocity inversion method and network architecture to improve the generalization and applicability of the deep learning velocity inversion method to the real observed data,so as to achieve the application of the method in real projects.In order to solve the problem of deep learning wave velocity inversion of tunnel seismic observed data,this paper proposes the chain research idea of " observed-synthetic data feature uniformity and transfers→gradient from velocity inversion dynamic correction→multimodal data fusion and inversion" by using theoretical analysis,numerical simulation,and field test.With the starting point of alleviating the difference between observed and synthetic data,this paper improves the effect of velocity structure inversion by dynamically correcting the gradient of velocity inversion,alleviate the discomfort of inversion by fusing the multimodal information constructed from seismic data,and then optimize the effect of inversion from real observed data by using the transfer learning to form a velocity inversion method of tunnel seismic forward-prospecting.The feasibility and effectiveness of the method are verified by numerical simulation and field test.r.The main research work and results of this paper are as follows:(1)Unified method of real observed-synthetic data features for tunnel seismic velocity inversion.Existing deep learning inversion methods are mostly based on synthetic data for network training,however,there are often large differences between observed data and synthetic data,which can easily lead to poor results in practical engineering applications.In this regard,a feature unification method for tunnel seismic data based on the attention convolutional autoencoder is proposed,which uses the attention mechanism to enhance the extraction of seismic data features,realizes the unification of the observed data and synthetic data in the high-dimensional parameter space,and builds a bridge between this two;and then constructs the synthetic and observed data sets for tunnel seismic velocity inversion,and verifies the feature extraction effect and unity between the synthetic data and observed data,and laid a high-quality data foundation for the inversion of the observed data by deep learning.(2)Dynamic optimisation of inversion gradient for structural modification of seismic velocity in tunnels.Due to the restriction of small observation space in tunnels,the gradient of seismic wave velocity inversion in tunnels is prone to arc-like distortion,which affects the accuracy of the velocity inversion structure.In this regard,a dynamic optimization strategy of the velocity inversion gradient based on the guidance of the true velocity model is proposed,and the difference between the velocity model and the real model in each round of inversion iteration is used as the label for the inversion gradient correction,so as to realize the construction of the gradient optimization operator with dual inputs and the dual-loop training strategy,which effectively improves the seismic data-driven inversion gradient morphology and amplitude,and provides a more reliable velocity inversion for deep learning inversion on the observed data.(3)Layer-by-layer tunnel seismic velocity inversion method based on multimodal fusion.Existing deep learning inversion methods mostly use seismic data as the network input,and the velocity trend and structural information in the data are implicitly reconstructed through data mining,which makes the network learning difficult,resulting in false anomalies in the tunnel velocity inversion results.In this regard,based on the seismic inversion method,the low-frequency data are transformed into large-scale velocity information,the high-frequency data are transformed into velocity structural information,and the features of the observation data are jointly used as the input to the network,and the attention mechanism is used to weight the three types of modal information for cross-fertilisation,which reduces the network’s reliance on the seismic data and the difficulty in mapping the velocity inversion.Then,dynamic mask-based loss function for the layer-by-layer velocity inversion is designed to achieve the results based on the multimodal velocity mapping.The loss function based on dynamic mask is designed to achieve the layer-by-layer inversion of tunnel seismic velocity based on multimodal fusion,which provides a feasible means for more accurate inversion of observed data.(4)The Improvement of tunnel seismic velocity inversion based on transfer learning.Based on the above training results of the tunnel seismic velocity multimodal fusion inversion method on synthetic data,the use of transfer learning to improve the inversion network parameters is an effective way to improve the effectiveness of the application of the observed data.However,the number of parameters of the velocity inversion network is large,and it is difficult to adequately train and update the whole inversion network parameters using a small amount of observed data.In contrast,through the visual display and interpretation analysis of the features of the middle layer of the inversion network,combined with the difference in the distribution of the features of the middle layer of the inversion network between the synthetic data and the observed data,the key layer that determines the inversion results of the observed data is located and updated,and the deep learning inversion method of the tunnel seismic velocity applicable to the observed data is finally formed.On the basis of the above research results,the effectiveness and applicability of the method of this paper are verified by carrying out the overpass detection tests at the actual engineering sites,such as Central Yunnan Water Diversion Project,the Pearl River Delta Water Resources Allocation Project,and the Shantou Bay Submarine Tunnel.