| In large hydropower projects,3D geological modeling and visualization has become an essential technology to reveal geological conditions,which is of considerable significance to improve engineering design level and ensure engineering safety and stability.The establishment of three-dimensional geological models involves geological big data,and Machine Learning(ML)can provide theoretic and technical support for 3D refined geomodelling based on geological big data analysis.Meanwhile,Mixed Reality(MR)can realize virtual-real fusion visualization of 3D geological information.Current research is still faced with following problems:(1)In the stage of data interpretation,existing geological data interpreting methods are mainly based on feature engineering,which limits the interpreting precision;(2)In the stage of geological modeling,the challenges in geomodelling research include rapid construction of complex nonlinear geological topology and density estimation and sampling from the high dimensional joint distribution of geological parameters;(3)In the stage of visualization,the virtual reality technology cannot realize the fusion of virtual geoinformation and real geological environment,and the 3D interaction process is less immersive.Machine learning models have the advantages of automatic extracting features,fitting highly nonlinear mapping relationships,and modeling high dimensional joint distributions.Meanwhile,MR 3D registration and interaction technology can seamlessly blend the hologram of virtual geoinformation with geoobjects in the physical world,and it can create a more natural 3D interaction experience.So,this paper studies the 3D refined modeling and mixed reality visualization of hydropower engineering geology based on machine learning and mixed reality,and the main contributions are as follows:(1)Multi-source geological data interpretation based on discriminative machine learningIn hydropower engineering,current geological data interpretation methods mainly rely on feature engineering.Based on automatic extraction of discriminative features using discriminative machine learning models,the interpretation methods of multisource geological data are proposed.First,on account of borehole image data,a novel Borehole Image Structural Plane-Region Proposal Network(BISP-RPN)is proposed to optimize the anchor design of the object detection model,and the precision of fracture recognition is improved.Meanwhile,the deep convolutional neural network is adopted to automatically extract the high-level semantic feature of fracture images,which can improve the anti-noise capability of fracture recognition.Second,on account of rock mass point cloud data,a point cloud segmentation model is established using density-based clustering,which can exclude outlier points during structural plane extraction.Third,on account of geological drill data,a geological boundary curve interpolation and imputation model among multiple geological profiles is established using recurrent neural network.The engineering application results show that,as for borehole image data,the proposed BISP-RPN method can increase the average accuracy of fracture object detection by around 7.28 percentage points,and avoid the misidentification of traditional image processing methods;as for rock mass point cloud data,the outlier points are reduced by 86.3% in the point cloud segmentation results;as for geological drill data,accuracy of the geological boundary curve is increased by about 18.02%.(2)Refined modeling of large-scale geological structures based on VGAE generative machine learning and improved T-splinesIn terms of the 3D modeling of large-scale geological structures,including overlaps,folds,faults and lenses,the fast construction of complex geological topologies is difficult.To solve this problem,a method of 3D refined modeling of geo-bodies using Variational Graph Auto Encoder(VGAE)generative machine learning and improved T-splines is proposed,in which,a VGAE link prediction model is established to realize the fast construction of adjacency topological relation among geological curves,and a Inhomogeneous Boundary Adaptive Local Refinement(IBALR)algorithm is proposed to improve the constructing efficiency of complex T-mesh topology structures.The engineering application results show that based on 30% known adjacency topologies,the VGAE link prediction accuracy can remain higher than 90%.And the improved Tspline modeling method can perform adaptive local refinement from NURBS surfaces and increase the average surface precision by around 48%.(3)Refined modeling of small-scale fracture networks based on DPCAF and CapsuleGAN generative machine learningIn terms of the small-scale fracture network modeling,current methods are faced with the challenge of estimating the multi-dimensional joint distribution among fracture parameters,and cannot simulate the realistic shape of fracture planes.Considering those problems,based on the generative machine learning of Density Peak Clustering Autoregressive Flow(DPCAF)model and Capsule Generative Adversarial Network(CapsuleGAN)model,an improved discrete fracture network(DFN)modeling method is proposed.The proposed DPCAF model improves the prior distribution of autoregressive flow model using Gaussian Mixture distribution and Density Peak clustering,and can realize accurate density estimation of the multi-modal and multidimensional joint distribution among fracture parameters.At the same time,a polygon structural plane image generation model is established based on CapsuleGAN,which can realize the simulation of the real shape of structural planes.The engineering application results show that the proposed DPCAF model can reduce the Wasserstein distance error of the multi-parameter(dip,dip direction and aperture)simulation by7.84%,which verified that the simulation results are closer to the distribution of real data.Besides,benefiting from the capsule network discriminator,CapsuleGAN can improve the quality of generated fracture polygon on small samples.(4)Mixed reality interactive visualization of geological information based on3 D registrationBased on virtual reality technology,the current approach for 3D geological information visualization cannot realize virtual-real fusion,and the 3D interaction is less immersive.To overcome those shortcomings,a mixed reality interactive visualization method for 3D geological information is proposed based on Simultaneous Localization and Mapping and Model-based 3D registration.Especially,considering the problem of MR recognition interaction with the fractured rock mass,based on the Holistically-Nested Edge Detection(HED)model,a novel method of Prominent Structural Plane(Pro-SP)is proposed,which can extract the visually prominent edge features of fractured rock mass to improve the stability of edge tracking in MR recognition interactions.The engineering application results show that the proposed Pro-SP method improved inlier ratios of edge tracking by around 11 percentage points,which indicates good stability of MR recognition interaction. |
