CN114254566B - Neural network geological exploration inversion method based on 2.5-dimensional mixed spectral element method - Google Patents
Neural network geological exploration inversion method based on 2.5-dimensional mixed spectral element method Download PDFInfo
- Publication number
- CN114254566B CN114254566B CN202111618427.4A CN202111618427A CN114254566B CN 114254566 B CN114254566 B CN 114254566B CN 202111618427 A CN202111618427 A CN 202111618427A CN 114254566 B CN114254566 B CN 114254566B
- Authority
- CN
- China
- Prior art keywords
- resistivity
- neural network
- inversion
- model
- geological
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 82
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 54
- 230000003595 spectral effect Effects 0.000 title claims description 31
- 238000012549 training Methods 0.000 claims abstract description 23
- 238000010801 machine learning Methods 0.000 claims abstract description 18
- 238000004364 calculation method Methods 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 8
- 239000011435 rock Substances 0.000 claims description 7
- 239000000126 substance Substances 0.000 claims description 6
- 230000001788 irregular Effects 0.000 claims description 5
- 238000012360 testing method Methods 0.000 claims description 5
- 239000003086 colorant Substances 0.000 claims description 4
- 238000001125 extrusion Methods 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 4
- 230000005684 electric field Effects 0.000 claims description 3
- 229910052751 metal Inorganic materials 0.000 claims description 3
- 239000002184 metal Substances 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 238000012876 topography Methods 0.000 abstract description 14
- 238000001228 spectrum Methods 0.000 abstract description 6
- 238000005516 engineering process Methods 0.000 abstract description 5
- 238000005259 measurement Methods 0.000 abstract description 5
- 230000005672 electromagnetic field Effects 0.000 abstract description 4
- 239000011159 matrix material Substances 0.000 description 20
- 238000009826 distribution Methods 0.000 description 18
- 238000013527 convolutional neural network Methods 0.000 description 9
- 229910052500 inorganic mineral Inorganic materials 0.000 description 9
- 239000011707 mineral Substances 0.000 description 9
- 238000010586 diagram Methods 0.000 description 8
- 238000005065 mining Methods 0.000 description 8
- 238000004422 calculation algorithm Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 6
- 238000004088 simulation Methods 0.000 description 4
- 238000012795 verification Methods 0.000 description 4
- 238000001514 detection method Methods 0.000 description 3
- 239000000463 material Substances 0.000 description 3
- 230000000052 comparative effect Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- MGRWKWACZDFZJT-UHFFFAOYSA-N molybdenum tungsten Chemical compound [Mo].[W] MGRWKWACZDFZJT-UHFFFAOYSA-N 0.000 description 1
- 238000013139 quantization Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V3/00—Electric or magnetic prospecting or detecting; Measuring magnetic field characteristics of the earth, e.g. declination, deviation
- G01V3/38—Processing data, e.g. for analysis, for interpretation, for correction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/02—Agriculture; Fishing; Forestry; Mining
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Business, Economics & Management (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Environmental & Geological Engineering (AREA)
- Geology (AREA)
- Remote Sensing (AREA)
- General Life Sciences & Earth Sciences (AREA)
- Geophysics (AREA)
- Agronomy & Crop Science (AREA)
- Animal Husbandry (AREA)
- Marine Sciences & Fisheries (AREA)
- Mining & Mineral Resources (AREA)
- Geometry (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Geophysics And Detection Of Objects (AREA)
Abstract
A neural network geological exploration inversion method based on a 2.5-dimensional mixed spectrum element method relates to the fields of seismic oil and gas exploration technology and electromagnetic field joint inversion. Firstly, simulating a data acquisition scene in a wide-area electromagnetic method, carrying out 2-dimensional geologic model modeling, obtaining a random model used for machine learning by realizing random change of a geologic structure through twisting and extruding an underground medium, carrying out forward modeling on each group of models by adopting a 2.5-dimensional mixed spectrum element method, obtaining a machine learning training data set, and inverting a neural network: training the dataset using a pattern of two UNET neural networks; and then verifying the neural network inversion method provided by the invention through the actual measurement data of the ore forming model (theoretical model) and the ore area respectively, and judging the effectiveness and accuracy of the neural network inversion method. The large-scale size model with the relief topography structure and the complex underground medium structure is realized, and the geological resource exploration of about 1000-3000 m of the underground deep part is realized.
Description
Technical Field
The invention relates to the field of seismic oil and gas exploration technology and electromagnetic field joint inversion, in particular to a neural network geological exploration inversion method based on a 2.5-dimensional mixed spectral element method, which is used for acquiring an electromagnetic field data set through the mixed spectral element method and carrying out geological inversion modeling by utilizing a neural network.
Background
Mineral resources are an important material basis upon which humans survive and develop. With the rapid development of economy and industry, the consumption of mineral resources in China is increased, the deep mining technology is developed rapidly in recent years, people are more concerned about the ore body distribution in the range of 1000 meters to 3000 meters below the earth surface, and particularly the development of electromagnetic prospecting methods and the continuous perfection of the mining theory in recent years provide more powerful conditions for deep mining, so that the deep mining technology can develop rapidly.
Electromagnetic exploration is often utilized as a common means of geophysical exploration to learn the electrical properties of subsurface media to infer other important geological information (e.g., the location of a mineral deposit). The inversion is an important step in electromagnetic exploration, and the electromagnetic properties of the stratum can be obtained by inversion, and the efficiency and accuracy of inversion depend on forward modeling.
Therefore, it is important to establish a frequency domain numerical calculation method and a corresponding theory capable of efficiently and accurately solving the forward modeling problem of the multi-scale and multi-mode electromagnetic method. In the forward aspect, the mixed spectral element method has been verified by the predecessor for multiple times to have high efficiency and accuracy in electromagnetic field calculation in complex heterogeneous medium, can provide effective calculation for accurate and rapid simulation of multi-scale problems and complex topography and topography, and has verified the advantages of the method in terms of accuracy and efficiency compared with the traditional method. In the aspect of inversion, starting from various exploration methods, different sensitivities of different ore body distributions to various measurement data are researched, an electromagnetic inversion method needs to be found, a reasonable and accurate frequency domain joint inversion objective function and stratum electrical structure constraint relation can be established on the basis of a multi-scale electromagnetic measurement equation, and efficient and stable multi-scale multi-mode electromagnetic measurement data joint inversion is realized.
Disclosure of Invention
The invention aims to provide a neural network geological exploration inversion method based on a 2.5-dimensional mixed spectral element method, which combines a machine learning method with the field of electromagnetic geological exploration, adopts a convolutional neural network to replace a traditional full-wave inversion method to invert underground mineral deposit distribution, and realizes accurate exploration of shallow mineral deposit resources within 3000 meters underground through joint inversion of multi-frequency electromagnetic exploration data (apparent resistivity).
The invention comprises the following steps:
simulating a data acquisition scene in a wide area electromagnetic method, and carrying out 2-dimensional ore-forming geological model modeling;
taking the 2-dimensional ore-forming geological model as a blue book, and randomly changing the geological structure and medium parameters to obtain more random models for machine learning;
performing forward modeling on each group of models by adopting a 2.5-dimensional mixed spectral element method to obtain a machine learning training data set, namely a apparent resistivity-resistivity data pair;
Step four, inverting the neural network: training the dataset by adopting two UNET neural networks, wherein the first network input is apparent resistivity and the output is resistivity; the input of the second network is the predicted resistivity of the first network, and the output is the resistivity;
Fifthly, inputting the apparent resistivity data of the test set into a trained network to obtain a predicted resistivity value; drawing a resistivity image by using colorbar functions in Matlab;
And step six, substituting the predicted resistivity back into the 2.5-dimensional mixed spectral element method to forward calculate the apparent resistivity, and then comparing the apparent resistivity with the actual apparent resistivity to calculate DATA MISFIT, so as to judge the effectiveness and accuracy of the neural network inversion method.
In the first step, the specific method for modeling the 2-dimensional ore-forming geological model is as follows:
(1) Carrying out resistivity mapping on each region according to geometric physical parameters of stratum and rock (ore);
(2) Carrying out irregular quadrilateral mesh subdivision on the model so as to enable a data set obtained in forward calculation of a spectral element method to meet the precision requirement; because the electromagnetic parameters among different substances are greatly changed, in order to ensure enough sampling rate within 1 Hz-10 KHz and simultaneously consider solving degree of freedom, different substances are required to be split respectively by adopting different grid sizes, a metal ore layer with lower resistivity (or smaller skin depth) is split by adopting a denser grid, and a common rock layer with higher resistivity is split by adopting a larger grid.
In the second step, the specific method for randomly changing the geological structure and the medium parameters by taking the 2-dimensional ore-forming geological model as the blue book is as follows:
(1) Dividing an initial 2-dimensional ore-forming geological model into a certain number of areas by adopting equal-resistance lines, representing different area numbers by using different colors, and numbering;
(2) The new geological structure is obtained by simulating the geological movement through twisting and extruding each medium, the twisting force and the position of extrusion are random in a certain range, and the area number is not changed although the geological structure is twisted and extruded in a random size;
(3) Filling each numbered geology with resistivity, wherein the filled resistivity randomly changes within +/-10% of the resistivity of a corresponding region in the initial model; the apparent resistivity and the resistivity used for inversion are changed after taking the logarithm based on 10. Thus, the post-log + -10% change is a large dynamic range.
In the third step, each group of models is forward-modeled at least 8 times, and 8 frequencies are used for forward modeling to calculate the corresponding apparent resistivity under the condition that the resistivity is the same when the models are forward-modeled 8 times; the 8 frequencies are 1,4, 16, 64, 256, 512, 1024, 2048Hz.
In the fourth step, training the data set by adopting two UNET neural networks, wherein the input dimension of the first network is 96×8, each model is provided with 96 identical receiving points, and the electric field values at 96 receiving points under 8 different frequencies are processed into apparent resistivity data; the output dimension is 128 x 256, which is the conductivity of the model, which is processed into resistivity data; the second network has an input dimension of 128 x 256, which is the resistivity data predicted by the first network; the output dimension is 128 x 256, which is the true resistivity data.
In step six, the calculation formula of DATA MISFIT is as follows:
Wherein:
test_ true apparent resistivity;
test_, the predicted resistivity of the neural network inverse is substituted back into the calculated apparent resistivity.
The invention provides a neural network electromagnetic inversion method based on a 2.5-dimensional mixed spectrum element method, which combines a machine learning method with the field of electromagnetic geological exploration, adopts a convolutional neural network to replace the traditional full-wave inversion method to invert underground mineral deposit distribution, realizes the joint inversion of multi-frequency electromagnetic exploration data of a 'apparent resistivity-resistivity' model, realizes a large-scale size model with a relief topography structure and a complex underground medium structure, and realizes geological resource exploration of about 1000-3000 m in the deep part of the underground.
The invention firstly adopts a 2.5-dimensional mixed spectral element method to forward and obtain a batch of data sets, then trains the batch of data sets by using a convolutional neural network, and finally uses the network to reverse the inversion result of resistivity distribution corresponding to measured data. In the inversion of measured data based on wide area apparent resistivity, the neural network inversion algorithm can rapidly characterize the distribution of underground ore bodies, and rapid inversion between an observation field and a physical model is realized.
Compared with the prior art, the invention has the outstanding advantages and technical effects that:
1. The split grid is regular in structure relative to the inversion of conventional planar layered terrain, but for subsurface geological structures with complex undulating terrain, the undulating terrain produces scattering effects on electromagnetic waves near the ground, so noise increases multiple times. Many algorithms are to apply grid generating software to generate unstructured grids when processing complex terrains, but the unstructured grids have strong randomness, and when inversion is performed by using a machine learning method, all training models need to share one set of grids, and the conventional method is difficult to meet the condition. Therefore, in order to adapt to the machine learning electromagnetic joint inversion under complex topography lines, a set of unstructured grids with better adaptability to broadband and undulating topography is researched and designed. The inversion model of the undulating surface can be accurately simulated, quantization errors caused by a regular grid in the traditional inversion method based on machine learning are effectively reduced, and a foundation is laid for rapid modeling and accurate forward computation in electromagnetic joint inversion based on a convolutional neural network;
2. Adopting two UNET neural networks, wherein the first network input is apparent resistivity and the output is resistivity; the second network input is the predicted resistivity of the first network and the output is the resistivity. The first network input is a low-dimensional space, the output is a high-dimensional space, and the network realizes the mapping from the low-dimensional space to the high-dimensional space;
3. For the relief topography structure and the complex underground medium structure, the joint inversion of the multi-frequency electromagnetic detection data of the apparent resistivity-resistivity model is realized. Compared with the prior detection technology and method which are mostly concentrated at the depth of 500-800 m, the inversion method provided by the invention can realize the inversion of a large-scale size model, and the following case results can prove that the inversion is effective for the model detection of about 1000-3000 m of the underground deep part.
Drawings
FIG. 1 is a schematic diagram of a two-dimensional resistivity map of the model of the present invention.
FIG. 2 is a schematic diagram of a two-dimensional irregular quadrilateral mesh subdivision of the model of the present invention.
FIG. 3 is a schematic view of the material area numbering of each cell of the inventive model.
FIG. 4 is a schematic representation of subsurface resistivity distribution for four stochastic models of the invention.
Fig. 5 is a schematic diagram of a convolutional neural network UNET network for inversion according to the present invention.
FIG. 6 is a comparative schematic of resistivity profiles obtained from the neural network inversion of the initial model of the present invention.
FIG. 7 is a schematic diagram showing the contrast of apparent resistivity distribution calculated forward by the method of inverting the obtained resistivity based on the neural network and substituting the obtained resistivity into 2.5-dimensional mixed spectral elements.
FIG. 8 is a schematic representation of resistivity distribution obtained from the inversion of measured data from a mine based on scs2d software in accordance with the present invention.
FIG. 9 is a schematic diagram of resistivity distribution obtained from inversion of measured data of a mining area based on a neural network according to the present invention.
FIG. 10 is a graph showing the contrast of apparent resistivity distribution calculated by forward calculation of a 2.5-dimensional mixed spectral element method based on resistivity obtained by inversion of scs2d software and a neural network for actual measurement data of a certain mining area.
Detailed Description
In order that the objects, aspects and advantages of the application will become apparent, a more particular description of the application will be rendered by reference to specific embodiments thereof which are illustrated in the appended drawings. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the application.
The invention provides a neural network electromagnetic inversion method based on a 2.5-dimensional mixed spectrum element method, which combines a machine learning method with the field of electromagnetic geological exploration, adopts a convolutional neural network to replace the traditional full wave inversion method to invert underground mineral deposit distribution, and realizes accurate exploration of underground 3000 meters with shallow mineral deposit resources through joint inversion of multi-frequency electromagnetic exploration data (apparent resistivity). According to the invention, a data acquisition scene in a wide-area electromagnetic method is simulated at first, 2-dimensional geologic model modeling is carried out, random change of a geologic structure is realized by twisting and extruding an underground medium to obtain a random model used for machine learning, and a machine learning training dataset ('apparent resistivity-resistivity') is obtained by utilizing simulation of a 2.5-dimensional mixed spectral element forward algorithm to train a convolutional neural network. And finally, verifying the neural network inversion method provided by the invention through an ore forming model (theoretical model) and measured data of a certain mining area, and verifying the feasibility and accuracy of the neural network through an inversion result.
The neural network electromagnetic inversion method based on the 2.5-dimensional mixed spectrum element method provided by the embodiment of the invention specifically comprises the following steps:
Step one, modeling a 2-dimensional ore-forming geological model;
taking the initial model (ideal model) as a blue book, and randomly changing the geological structure and medium parameters of the model to obtain 2000 groups of models which are different from each other;
and thirdly, forward modeling is carried out on the 2000 groups of models by adopting a 2.5-dimensional mixed spectral element method, so that 2000 groups of 'apparent resistivity-resistivity' data pairs are obtained. In particular, each model group is forward-developed 8 times, and the corresponding apparent resistivity is calculated by forward-developed 8 frequencies under the condition that the resistivity is the same. Wherein 8 frequencies are 1, 4, 16, 64, 256, 512, 1024, 2048Hz;
and fourthly, inverting the neural network. Training the data set by adopting a mode of two UNET neural networks, wherein the input dimension of the first network is 96 x 8 and the output dimension is 128 x 256; the second network input dimension 128 x 256 and output dimension 128 x 256;
fifthly, inputting the apparent resistivity data of the test set into a network with the best training to obtain a predicted resistivity value. Drawing a resistivity image by using colorbar functions in Matlab;
And step six, substituting the predicted resistivity back into the 2.5-dimensional mixed spectral element method to forward calculate the apparent resistivity, and then comparing the apparent resistivity with the actual apparent resistivity to calculate DATA MISFIT. And judging the effectiveness and accuracy of the neural network inversion method.
Further, the specific method for establishing the geologic model in the first step is as follows:
carrying out resistivity mapping on each region according to geometric physical parameters of stratum and rock (ore);
In order to enable the obtained data set to meet the precision requirement in the forward computation of the spectral element method, the model is subjected to irregular quadrilateral mesh subdivision. Because the electromagnetic parameters among different substances are greatly changed, in order to ensure enough sampling rate within 1 Hz-10 KHz and simultaneously consider solving degree of freedom, different substances are required to be split respectively by adopting different grid sizes, a metal ore layer with lower resistivity (or smaller skin depth) is split by adopting a denser grid, and a common rock layer with higher resistivity is split by adopting a larger grid.
Further, the specific method of the second step is as follows:
firstly, an initial model is divided into a certain number of areas by adopting an equal resistance line. Wherein different colors represent different area numbers and are numbered;
The new geological structure is obtained by simulating the geological movement through twisting and extruding each medium, the twisting force and the position of extrusion are random in a certain range, and the area number is not changed although the geological structure is twisted and extruded in a random size;
each numbered geology is filled with resistivity, which is randomly varied within + -10% of the resistivity of the corresponding region in the initial model. It is noted that the apparent resistivity and resistivity used for inversion are both changed after taking the logarithm based on 10. Thus, the post-log + -10% change is a large dynamic range.
Further, the dimension description in the fourth step is specifically as follows:
the input dimension of the first network is 96 x 8, each model is provided with 96 identical receiving points, and the electric field values at the 96 receiving points under 8 different frequencies are processed into apparent resistivity data; the output dimension is 128 x 256, which is the conductivity of the model, which is processed into resistivity data;
The second network input dimension is 128 x 256, which is the resistivity data predicted by the first network; the output dimension is 128 x 256, which is the true resistivity data.
Further, the calculation formula in the sixth step DATA MISFIT is as follows:
Wherein:
test_ true apparent resistivity;
test_, the predicted resistivity of the neural network inverse is substituted back into the calculated apparent resistivity.
Specific application examples are given below in conjunction with the drawings.
A schematic of the model two-dimensional resistivity map is shown in fig. 1. The model is based on the ore forming mode of tungsten-molybdenum polymetallic ore deposit in certain ore area, and the resistivity of each area is filled according to the stratum and the geometric physical parameters of rock (ore) of the ore deposit. The low-resistance target body comprises 10 sheet-shaped geological anomalies with dark purple filling color, 5 block anomalies with green filling color and 2 large area anomalies with cyan filling color. It is noted that the model contains a horizontally extending overburden with a resistivity of 20Ω, which is not considered an anomaly in the forward and reverse calculations, but is considered a normal background formation.
The model two-dimensional irregular quadrilateral mesh subdivision scheme is shown in fig. 2, and inversion of a subsurface geological structure with complex undulating terrain relative to planar layered terrain is more complex. Because undulating terrain causes scattering effects to occur near the ground, noise can be more than in the case of voltaic terrain. Moreover, the traditional inversion method is based on horizontal topography, and a regular structural grid is applied, so that the details of the relief topography are difficult to characterize. Many algorithms are to apply grid generating software (such as cubit) to generate unstructured grids when processing complex terrains, but the unstructured grids generated by the grid generating software also have a certain problem, and when inversion is performed by a machine learning method, all training models need to share one set of grids, and the unstructured grids have strong randomness, so that the condition is difficult to meet. In order to adapt to the electromagnetic inversion problem under a complex topography line, a set of unstructured grids with better adaptability to broadband and undulating topography is researched and designed.
Firstly, determining the relief topography, dividing a calculation area into an upper part and a lower part, wherein the upper part is air, and the lower part is an underground medium, namely read topography data; the air portion of the calculation area and the cross direction of the subsurface medium are then evenly divided into N parts, for air, the resistivity is very small, so that the skin depth is very large, and therefore coarse grids can be used, whereas the conductivity of the various subsurface mediums is generally larger than that of air and the size and area cannot be determined, but generally the higher the frequency is, the smaller the skin depth is, the shallower the detectable depth, the finer grid is needed, and conversely, the lower the frequency is, the greater the skin depth is, the deeper the detectable depth is, and the coarser grid can be used. Therefore, the underground grid height adopts equal proportion subdivision, the grid height is smaller as the ground is closer, and the grid height is increased in equal proportion as the ground is deeper. In a developed software module, the proportion can be customized by a user, and meanwhile, the transverse grid number N x, the longitudinal grid number N a in air and the longitudinal grid number N y of underground media can be defined by the user, so that the whole grid N x×(Na+Ny) is adopted in forward modeling, and the underground media grid N x×Ny of interest only needs to participate in inversion, thereby improving inversion precision and efficiency.
A schematic diagram of the material area numbering of each cell of the model is shown in fig. 3. The training set data for machine learning inversion needs to have certain randomness to ensure that an inversion system obtained by training has more generalization and an inversion result is more accurate, so that the geological structure needs certain random change, and corresponding medium parameters also need certain randomness. Firstly, an initial model is divided into a certain number of areas by adopting an equal resistance line. Different colors are used for representing different area numbers, and numbering is carried out. Thus, the geological structure of the initial model can be obtained, and the resistivity mean value of the corresponding region is used as the resistivity model of the initial model. Then generating training labels by taking the initial model as a template, simulating the geological movement by twisting and extruding each medium in the aspect of geological structure to obtain a new geological structure, wherein the twisting force and the position of extrusion are random within a certain range, although the geological structure is twisted and extruded in a random size, the region numbers are not changed, then filling the geological structure with the resistivity of each number, and the filled resistivity is changed randomly within +/-10% of the resistivity of the corresponding region in the initial model. It is noted that the apparent resistivity and resistivity used for inversion are both changed after taking the logarithm based on 10. Thus, the post-log + -10% change is a large dynamic range. With the above method, 2000 sets of random models with widely different geologic structures and resistivity matrices were created from the initial model derivation.
FIG. 4 is a schematic representation of subsurface resistivity distribution for four stochastic models of the invention. It can be seen from fig. 4 that there is a large change in both the geologic structure and the resistivity parameters.
The scale of this model is 1500m 700m. 96 receivers are each placed 0.2m deep below the undulating surface line, each equidistant 5m from 125m start to 1075m end. The point source coordinates are (500, -200, 300) m, and the distance from the measuring point is 13.3km. At 8 working frequencies of 1, 4, 16, 64, 256, 512, 1024 and 2048Hz respectively, the structure and resistivity data matrix (128 x 256) of the 2000 groups of models are put into a 2.5-dimensional mixed spectral element algorithm for forward modeling, and apparent resistivity (96 x 8) is acquired in a simulation mode to obtain 2000 groups of 'apparent resistivity-resistivity' data pairs.
Fig. 5 is a schematic diagram of a convolutional neural network UNET network for inversion according to the present invention. Training a 'apparent resistivity-resistivity' dataset using a pattern of two UNET neural networks, wherein the first network input dimension is 96 x 8 and the output dimension is 128 x 256; the second network has input dimensions 128 x 256 and output dimensions 128 x 256, wherein the input matrix is the predicted resistivity data for the first network and the output matrix is the actual resistivity data. 2000 groups of different random models are divided into a training set, a verification set and a test set to train the network, the network with the best effect is selected according to the training loss and the verification loss, and parameters such as a network convolution kernel, a deconvolution kernel and the like are stored.
FIG. 6 is a comparative schematic of resistivity profiles obtained by inversion of an initial model based on a neural network. The visual resistivity matrix of the initial model is taken as an input matrix to be brought into the neural network with the best training to obtain a predicted resistivity numerical matrix. And drawing a predicted resistivity matrix image and a real resistivity matrix image by using colorbar functions in Matlab, and comparing. It can be seen from fig. 6 that the inversion results substantially coincide with the real model subsurface structure. And the resistivity distribution of the inversion result is brought back into the forward calculation of a 2.5-dimensional mixed spectral element method to obtain the apparent resistivity of the inversion result, and the two-norm error (DATA MISFIT) of the apparent resistivity of the real model is 3.85%. And a imagesc function in Matlab is used for drawing a predicted apparent resistivity matrix image and a real apparent resistivity matrix image, and the predicted apparent resistivity matrix image and the real apparent resistivity matrix image are compared, and a comparison chart is shown in figure 7. The inversion result shows that the apparent resistivity-resistivity model can be accurately inverted based on the neural network.
The present invention will be described in further detail below using measured data (apparent resistivity) of a mine area as another example.
FIG. 8 is a schematic representation of resistivity profiles obtained from the inversion of measured data from a mine based on scs2d software. Taking the model as an initial model, and establishing the model, randomly distorting the geological structure and randomly changing the medium parameters according to the first step and the second step, thereby obtaining 3000 groups of models which are different from each other.
The scale of this model is 4500m×3500m. 105 receivers are placed 0.2m deep below the undulating surface line, each receiver spacing being around 40m, starting from 83m and ending at 4484 m. The point source coordinates are (500, -200, 300) m, and the distance from the measuring point is 13.6km. At 8 working frequencies of 1, 4, 16, 64, 256, 512, 1024 and 2048Hz respectively, a structure of 3000 groups of models and a resistivity data matrix (192 x 256) are put into a 2.5-dimensional mixed spectral element algorithm for forward modeling, and apparent resistivity (105 x 8) is acquired in a simulation mode to obtain 3000 groups of 'apparent resistivity-resistivity' data pairs.
Training the 'apparent resistivity-resistivity' data set by adopting two UNET neural networks, wherein the input dimension and the output dimension of the first network are correspondingly changed, and the input dimension of the first network is 105 x 8 and the output dimension of the first network is 192 x 256; the second network input dimension 192 x 256 and output dimension 192 x 256. 3000 groups of different random models are divided into a training set, a verification set and a test set to train the network, the network with the best effect is selected according to the training loss and the verification loss, and parameters such as a network convolution kernel, a deconvolution kernel and the like are stored.
Fig. 9 is a schematic diagram of resistivity distribution obtained by inversion of measured data of a certain mining area based on a neural network. The measured data (apparent resistivity matrix) is taken as an input matrix into the neural network with the best training to obtain a predicted resistivity numerical matrix. The predicted resistivity matrix image and the measured data (apparent resistivity matrix) were plotted using colorbar functions in Matlab and compared based on a resistivity matrix image (fig. 8) developed by scs2d software. From the graph, compared with the scs2d software inversion result, the resistivity distribution of the neural network inversion result is obviously more detailed and reasonable. Respectively bringing the resistivity distribution of the two inversion results back into forward calculation by a 2.5-dimensional mixed spectral element method to obtain apparent resistivity of the inversion results, wherein the apparent resistivity inverted by the corresponding neural network and the two norm error (DATA MISFIT) between the apparent resistivity and the actually measured apparent resistivity are 7.86%; and the apparent resistivity corresponding to the scs2d software inversion and the two-norm error (DATA MISFIT) from the measured apparent resistivity were 9.62%. And a imagesc function in Matlab is used for drawing a visual resistivity matrix image, and comparison is carried out, and a comparison chart is shown in figure 10. The inversion result shows that the apparent resistivity-resistivity model can be accurately inverted based on the neural network. And the inversion effect is better than that of the scs2d traditional software.
The invention provides a neural network electromagnetic inversion method based on a 2.5-dimensional mixed spectrum element method, which combines a machine learning method with the field of electromagnetic geological exploration, adopts a convolutional neural network to replace the traditional full-wave inversion method to invert underground mineral deposit distribution, realizes the joint inversion of multi-frequency electromagnetic exploration data of a 'apparent resistivity-resistivity' model, realizes a large-scale size model with a relief topography structure and a complex underground medium structure, and realizes geological resource exploration of about 1000-3000 m in the deep part of the underground.
The above-described embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. All equivalent changes and modifications within the scope of the present invention are intended to be covered by the present invention.
Claims (8)
1. A neural network geological exploration inversion method based on a 2.5-dimensional mixed spectral element method is characterized by comprising the following steps:
simulating a data acquisition scene in a wide area electromagnetic method, and carrying out 2-dimensional ore-forming geological model modeling;
taking the 2-dimensional ore-forming geological model as a blue book, and randomly changing the geological structure and medium parameters to obtain a plurality of groups of random models for machine learning;
performing forward modeling on each group of models by adopting a 2.5-dimensional mixed spectral element method to obtain a machine learning training data set, namely a apparent resistivity-resistivity data pair;
Step four, inverting the neural network: training the dataset by adopting two UNET neural networks, wherein the first network input is apparent resistivity and the output is resistivity; the input of the second network is the predicted resistivity of the first network, and the output is the resistivity;
Fifthly, inputting the apparent resistivity data of the test set into a trained network to obtain a predicted resistivity value; drawing a resistivity image by using colorbar functions in Matlab;
And step six, substituting the predicted resistivity back into the 2.5-dimensional mixed spectral element method to forward calculate the apparent resistivity, and then comparing the apparent resistivity with the actual apparent resistivity to calculate DATA MISFIT, so as to judge the effectiveness and accuracy of the neural network inversion method.
2. The neural network geological exploration inversion method based on the 2.5-dimensional mixed spectral element method as claimed in claim 1, wherein in the step one, the specific method for modeling the 2-dimensional mineralized geological model is as follows:
(1) Carrying out resistivity mapping on each region according to geometrical physical parameters of stratum and rock or ore;
(2) Carrying out irregular quadrilateral mesh subdivision on the model so as to enable a data set obtained in forward calculation of a spectral element method to meet the precision requirement; because the electromagnetic parameters among different substances have larger change, in order to ensure enough sampling rate within 1 Hz-10 KHz and simultaneously consider solving degree of freedom, different substances are required to be split respectively by adopting different grid sizes, a metal ore layer with lower resistivity or smaller skin depth is split by adopting a denser grid, and a common rock layer with higher resistivity is split by adopting a bigger grid.
3. The neural network geological exploration inversion method based on the 2.5-dimensional mixed spectral element method according to claim 1, wherein in the second step, the plurality of groups of machine learning use random models, at least 1000 groups.
4. The neural network geological exploration inversion method based on the 2.5-dimensional mixed spectral element method as claimed in claim 1, wherein in the second step, the specific method for randomly changing the geological structure and the medium parameters by taking the 2-dimensional ore-forming geological model as a blue book is as follows:
(1) Dividing an initial 2-dimensional ore-forming geological model into a certain number of areas by adopting equal-resistance lines, representing different area numbers by using different colors, and numbering;
(2) The new geological structure is obtained by simulating the geological movement through twisting and extruding each medium, the twisting force and the position of extrusion are random in a certain range, and the area number is not changed although the geological structure is twisted and extruded in a random size;
(3) Filling each numbered geology with resistivity, wherein the filled resistivity randomly changes within +/-10% of the resistivity of a corresponding region in the initial model; the apparent resistivity and the resistivity used for inversion are changed after taking the logarithm based on 10.
5. A neural network geological exploration inversion method based on a 2.5-dimensional mixed spectral element method according to claim 1, wherein in step three, said forward modeling is performed at least 8 times for each set of models.
6. The neural network geological exploration inversion method based on the 2.5-dimensional mixed spectral element method according to claim 5, wherein 8 frequencies are used for forward modeling to calculate corresponding apparent resistivity under the condition of the same resistivity when the 8 frequencies are used for forward modeling; 8 frequencies are 1,4, 16, 64, 256, 512, 1024, 2048Hz.
7. The neural network geological exploration inversion method based on the 2.5-dimensional mixed spectral element method as claimed in claim 1, wherein in the fourth step, the data set is trained by adopting two modes of UNET neural networks, wherein the input dimension of the first network is 96 x 8, each model is provided with 96 identical receiving points, and electric field values at 96 receiving points under 8 different frequencies are processed into apparent resistivity data; the output dimension is 128 x 256, which is the conductivity of the model, which is processed into resistivity data; the second network has an input dimension of 128 x 256, which is the resistivity data predicted by the first network; the output dimension is 128 x 256, which is the true resistivity data.
8. The neural network geological exploration inversion method based on the 2.5-dimensional mixed spectral element method as claimed in claim 1, wherein in the sixth step, the calculation formula of DATA MISFIT is as follows:
Wherein:
test_Y, true apparent resistivity; test_y the predicted resistivity of the neural network inverse is substituted back into the calculated apparent resistivity.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111618427.4A CN114254566B (en) | 2021-12-27 | 2021-12-27 | Neural network geological exploration inversion method based on 2.5-dimensional mixed spectral element method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111618427.4A CN114254566B (en) | 2021-12-27 | 2021-12-27 | Neural network geological exploration inversion method based on 2.5-dimensional mixed spectral element method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114254566A CN114254566A (en) | 2022-03-29 |
| CN114254566B true CN114254566B (en) | 2024-06-04 |
Family
ID=80795342
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111618427.4A Active CN114254566B (en) | 2021-12-27 | 2021-12-27 | Neural network geological exploration inversion method based on 2.5-dimensional mixed spectral element method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114254566B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114740529B (en) * | 2022-04-08 | 2024-08-13 | 北京大学深圳研究院 | Method for inverting pore medium permeability coefficient by using active source seismoelectric spectrum ratio |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104866653A (en) * | 2015-04-29 | 2015-08-26 | 中国地质科学院矿产资源研究所 | Method for acquiring underground three-dimensional density structure |
| CN113807020A (en) * | 2021-09-28 | 2021-12-17 | 成都理工大学 | Magnetotelluric inversion method based on deep learning constraint |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| AU2020325168A1 (en) * | 2019-08-06 | 2022-03-03 | ExxonMobil Technology and Engineering Company | Petrophysical inversion with machine learning-based geologic priors |
-
2021
- 2021-12-27 CN CN202111618427.4A patent/CN114254566B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104866653A (en) * | 2015-04-29 | 2015-08-26 | 中国地质科学院矿产资源研究所 | Method for acquiring underground three-dimensional density structure |
| CN113807020A (en) * | 2021-09-28 | 2021-12-17 | 成都理工大学 | Magnetotelluric inversion method based on deep learning constraint |
Non-Patent Citations (2)
| Title |
|---|
| 2.5D重磁震联合反演在南黄海地质研究中的应用;杨金玉;海洋地质动态;20080828;全文 * |
| The Efficient Hybrid Mixed Spectral Element Method With Surface Current Boundary Condition for Modeling 2.5-D Fractures and Faults;Xuanying Hou;IEEE Xplore;20200720;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114254566A (en) | 2022-03-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107742015B (en) | DC induced polarization method three-dimensional numerical simulation method based on arbitrary dipole-dipole device | |
| CN105158808B (en) | A kind of shallow sea transient electromagnetic air-sea detection and its means of interpretation | |
| CN111815773B (en) | Three-dimensional complex geologic model label manufacturing method suitable for machine learning algorithm | |
| CN106932819A (en) | Pre-stack seismic parameter inversion method based on anisotropy Markov random field | |
| Huggenberger et al. | A sedimentological model to characterize braided river deposits for hydrogeological applications | |
| CN104422969B (en) | A kind of method for reducing electromagnetic sounding inversion result nonuniqueness | |
| CN111899338B (en) | Method, device and system for three-dimensional modeling of stratum lithology of coverage area | |
| Attwa et al. | Structure-based geoelectrical models derived from genetic algorithms: a case study for hydrogeological investigations along Elbe River coastal area, Germany | |
| CN114254566B (en) | Neural network geological exploration inversion method based on 2.5-dimensional mixed spectral element method | |
| CN115097534A (en) | Method and system for identifying space morphology of dry hot rock fracturing transformation | |
| CN117852416B (en) | Multimode grouting precontrolled analysis method and system based on digital geological model | |
| Ferhat et al. | A geoelectric approach for karst groundwater analysis | |
| Persova et al. | Geometric 3-D inversion of airborne time-domain electromagnetic data with applications to kimberlite pipes prospecting in a complex medium | |
| Auken et al. | Model recognition by 1-D laterally constrained inversion of resistivity data | |
| CN107748393B (en) | A kind of bearing calibration based on the stratigraphic dip of numerical simulation to resistivity effects | |
| CN113406707A (en) | Magnetotelluric multi-scale and multi-time-period detection method | |
| CN103901475B (en) | Attribute contour map drawing method and device | |
| Ma et al. | Upscaling equivalent hydrofacies simulation based on borehole data generalization and transition probability geostatistics | |
| CN207673329U (en) | A kind of electric imaging logging image enhancement display processing system | |
| CN114236624B (en) | Method and system for estimating fracturing modification space volume based on electromagnetic method | |
| Aizebeokhai et al. | Anomaly effects of orthogonal paired-arrays for 3D geoelectrical resistivity imaging | |
| CN111812729A (en) | Multi-dimensional geological background model construction method for optimizing electrical prospecting inversion | |
| Castillo-Reyes et al. | Electromagnetic Subsurface Imaging in the Presence of Metallic Structures: A Review of Numerical Strategies | |
| Liu et al. | 3D Structurally Constrained Inversion of the Controlled-Source Electromagnetic Data Using Octree Meshes | |
| Bakshevskaya et al. | Methods of modeling hydraulic heterogeneity of sedimentary formations |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |