A method and system for identifying underground lithology based on machine learning

Abstract

The present invention discloses a method and system for identifying underground lithology based on machine learning, and relates to the technical field of machine learning. The method comprises gridding a study area to obtain a plurality of underground grid units, obtaining the spatial physical property characteristics of each physical property of underground rocks and ores in each underground grid unit, and for each underground grid unit, inputting the spatial physical property characteristics of all physical properties of underground rocks and ores in the underground grid unit into a trained underground lithology identification model to obtain the underground lithology of the underground grid unit. The present invention uses three physical properties of density, magnetic susceptibility and resistivity of the study area to identify underground lithology, thereby improving the accuracy of lithology identification. The present invention uses a machine learning algorithm to identify underground lithology, and through machine learning of the laws of the corresponding relationship between multiple physical properties and different lithologies, it has strong objectivity, further improving the accuracy of lithology identification and mapping.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to an underground lithology identification method and system based on machine learning. Background Art

[0002] Underground lithology identification is an important task in the field of earth science and engineering. It involves the understanding and research of underground rock properties, which is of great significance to geological mapping, mineral exploration, oil and gas exploration, geological engineering and other fields. For mineral exploration, lithology identification can help determine the location and distribution of ore, as well as the properties of ore, which is very important for the development and management of mineral resources downstream of exploration; for oil and gas exploration, lithology identification plays a vital role in oil and gas exploration. By analyzing the properties of underground rocks, explorers can determine the location, thickness and properties of oil and gas reservoirs, thereby guiding drilling and development work; for geological engineering, the field of geological engineering needs to accurately identify the properties of underground rocks to guide engineering design and construction processes. Lithology identification can help engineers understand the stability, strength and permeability of underground rocks, thereby ensuring the safety and feasibility of the project.

[0003] Lithology identification mainly includes direct identification methods such as hand specimen identification, thin section identification, element testing and mineral testing, as well as indirect identification methods such as remote sensing and geophysical inversion. For rocks deep underground that are difficult to reach by rock sample collection methods, geophysical inversion as an indirect method is a key way to identify lithology. It can determine the physical properties of underground rocks, and then determine the type and composition of underground rocks, which is crucial for geological exploration and resource development.

[0004] Geophysical methods measure, analyze and invert the physical fields (gravity field, geomagnetic field, geoelectric field) generated by underground materials on or above the ground, so as to achieve the transformation from the geophysical potential field on or above the ground to the underground physical properties. This type of geophysical methods includes gravity, geomagnetism, electromagnetism, seismology, etc. Geophysical inversion is a mathematical method that uses geophysical measurement data to infer the properties of underground rocks. It is an important method to help understand the physical properties of underground rocks. It is usually based on physical equations and numerical simulations, and can obtain information related to rock physical properties (density, magnetic susceptibility, resistivity). The indirect way of lithology identification - geophysical means has a very obvious effect on mineral exploration, oil and gas exploration, geological engineering, etc.: gravity, magnetic and electrical methods are usually used in mineral exploration, and geophysical inversion technology can provide key information; geophysical inversion methods are widely used in oil and gas exploration to help explorers accurately understand the underground situation; magnetic and electrical methods are widely used in geological engineering.

[0005] However, geophysical inversion usually provides information on underground physical properties and structures, rather than direct lithological information. The process of converting the rock physical properties obtained by inversion into rock types and compositions is very arduous and critical. Therefore, it is increasingly important to combine efficient and accurate methods to extract lithological information from inversion results.

[0006] For a long time, the comprehensive interpretation of gravity and magnetic data has been one of the effective methods for characterizing and identifying underground targets. Initially, people only directly compared the anomaly maps of gravity and magnetic data, conducted qualitative analysis on the plane, and interpreted the geological units that may correspond to different combinations of gravity and magnetic data anomalies to achieve lithology identification of "surface geological mapping". In addition to directly interpreting the anomaly maps, some researchers have started from the correspondence between gravity and magnetic data to conduct research on extracting more information for lithology identification. Kanasewich and Agarwal (1970) implemented a method for obtaining the ratio of magnetic susceptibility to density in the wavenumber domain, and verified the effect of this ratio in lithology identification with examples. Dransfield et al (1994) directly solved the correspondence between gravity and magnetic data gradients through Poisson's equation. Price and Dransfield (1995) used this method for visual lithology mapping and achieved certain results. In view of the development of gravity and magnetic data measurement technology, it is now possible to directly measure the gradient of gravity and magnetic methods. High-precision gradient data has brought convenience to lithology identification (Braga et al., 2009; Martinez et al., 2011). In addition to directly using gravity and magnetic data to calculate apparent magnetic susceptibility and apparent density maps, obtaining magnetic susceptibility and density through inversion to study lithology distribution characteristics is an important direction of current gravity and magnetic technology development. Lane and Guillen (2005) explored the method of using the inverted magnetic susceptibility and density as auxiliary information for lithology classification and identification. Williams and Dipple (2007) used prior geological information as a reference inversion model to perform three-dimensional physical property inversion, obtained three-dimensional density and magnetic susceptibility models, and combined with drilling data to explore the identification method of mineralized alteration lithology. Kowalczyk et al. (2010) used three-dimensional inversion of gravity and magnetic data to obtain a regional magnetic susceptibility and density difference model. Based on the scatter plot of density and magnetic susceptibility, they determined the correspondence between lithology and physical properties, and used this model to distinguish and identify the distribution of different lithological units.

[0007] Yan Jiayong, Lü Qingtian, et al. (2014) took the Lusong mining area in Anhui as an example, proposed a lithology identification process based on three-dimensional inversion of gravity and magnetic data, and carried out mapping experiments. Although the existing mapping results reflect the changes in deep lithology and make up for the shortcomings of surface geological mapping, the accuracy of underground lithology identification is low. Summary of the invention

[0008] The purpose of the present invention is to provide a method and system for underground lithology identification based on machine learning, which can improve the accuracy of underground lithology identification.

[0009] To achieve the above object, the present invention provides the following solution.

[0010] The present invention provides an underground lithology identification method based on machine learning, comprising the following steps.

[0011] The study area is divided into grids to obtain a plurality of underground grid units in the study area.

[0012] The physical properties of underground rocks and minerals in each of the underground grid cells are obtained; the physical properties include density, magnetic susceptibility and resistivity.

[0013] For each of the physical properties of underground rocks and minerals in each of the underground grid units, feature extraction is performed on the physical property according to the spatial position of the underground grid unit corresponding to the physical property to obtain the spatial physical property feature of the physical property.

[0014] For each of the underground grid units, the spatial physical property characteristics of all physical attributes of underground rocks and ores in the underground grid unit are input into a trained underground lithology recognition model to obtain the underground lithology of the underground grid unit.

[0015] The present invention also provides an underground lithology identification system based on machine learning, comprising the following modules.

[0016] The grid division module is used to divide the study area into grids to obtain multiple underground grid units in the study area.

[0017] The physical property acquisition module is used to acquire the physical properties of underground rocks and ores in each of the underground grid units; the physical properties include density, magnetic susceptibility and resistivity.

[0018] The spatial physical property feature extraction module is used to extract the feature of each physical property of the underground rock ores in each underground grid unit according to the spatial position of the underground grid unit corresponding to the physical property, so as to obtain the spatial physical property feature of the physical property.

[0019] The underground lithology prediction module is used to input the spatial physical property characteristics of all physical properties of underground rocks and ores in each underground grid unit into a trained underground lithology recognition model to obtain the underground lithology of the underground grid unit.

[0020] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: The present invention provides an underground lithology identification method based on machine learning, firstly, the study area is gridded to obtain multiple underground grid units, and the spatial physical property characteristics of each physical attribute of underground rocks and minerals in each underground grid unit are obtained. For each underground grid unit, the spatial physical property characteristics of all physical attributes of underground rocks and minerals in the underground grid unit are input into the trained underground lithology identification model to obtain the underground lithology of the underground grid unit. The present invention uses the three physical attributes of density, magnetic susceptibility and resistivity of the study area to identify underground lithology, thereby improving the accuracy of lithology identification. The present invention uses a machine learning algorithm to identify underground lithology, and by learning the laws of the corresponding relationship between multiple physical properties and different lithologies by machine learning, not only the efficiency of lithology identification and mapping is improved, but also it has a strong objectivity, further improving the accuracy of lithology identification and mapping. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0022] Figure 1 A schematic flow chart of a method for identifying underground lithology based on machine learning provided in an embodiment of the present invention.

[0023] Figure 2 A schematic diagram of density obtained by gravity inversion provided in an embodiment of the present invention.

[0024] Figure 3 A schematic diagram of magnetic susceptibility obtained by magnetic inversion according to an embodiment of the present invention.

[0025] Figure 4 A schematic diagram of resistivity obtained by electrical inversion according to an embodiment of the present invention.

[0026] Figure 5 A corresponding relationship diagram between the magnetic susceptibility and density of rocks (ores) in the northeastern Jiangxi mining area provided in an embodiment of the present invention.

[0027] Figure 6 A corresponding relationship diagram between the resistivity, magnetic susceptibility and density of rocks (ores) in the northeastern Jiangxi mining area provided in an embodiment of the present invention.

[0028] Figure 7 The embodiments of the present invention provide the lithology identification and mapping results of rocks (ores) in the northeastern Jiangxi mining area.

[0029] Figure 8A separable feature space is provided for an embodiment of the present invention.

[0030] Fig. 9 An inseparable feature space is provided for an embodiment of the present invention.

[0031] Fig.10 A schematic diagram of a prior art process provided for an embodiment of the present invention.

[0032] Fig.11 A corresponding relationship diagram between the magnetic susceptibility and density of rocks (ores) in the Lusong ore cluster area provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] Yan Jiayong, Lü Qingtian, et al. (2014) proposed a lithology identification and mapping technology based on three-dimensional inversion of gravity and magnetic data. The technical process of the solution is as follows: Fig.10 As shown. However, this technical solution has the following disadvantages: 1. Due to the lack of vertical resolution of gravity and magnetic data, the ability to distinguish vertical geological structures is poor. Only under the premise of increasing the vertical prior information constraints can the vertical resolution be effectively improved. 2. The properties of underground rocks are usually affected by multiple physical factors, not just gravity and magnetic methods, so it is necessary to integrate multiple geophysical methods for comprehensive identification. 3. Geophysical inversion results are usually accompanied by a certain degree of uncertainty. Therefore, how to deal with inversion uncertainty and obtain accurate inversion results is an important research topic. 4. The analysis of the relationship between multiple geophysical inversion results and lithology involves the coordinated processing of different data sources, and the conventional comparison relationship based on rectangular coordinates is difficult to meet the requirements. 5. Lithology identification is a complex task that requires comprehensive analysis of multiple geophysical data. Different rock types have different physical properties, and the classification of the corresponding relationship between multiple physical properties and different lithologies is very cumbersome and difficult. The reasons for these deficiencies are, on the one hand, the deficiencies in the design of the above-mentioned existing schemes, and on the other hand, the defects in the methods and technologies adopted by the above-mentioned existing schemes. Ultimately, the main disadvantage is the inaccurate lithology identification caused by insufficient data information and backward technical means.

[0035] The purpose of the present invention is to provide an underground lithology identification method based on machine learning, which improves the accuracy of lithology identification by using three physical properties of density, magnetic susceptibility and resistivity of the study area to identify underground lithology. The present invention uses a machine learning algorithm to identify underground lithology, and learns the laws of the corresponding relationship between multiple physical properties (i.e., physical properties, which can also be called physical property information, physical property data) and different lithologies through machine learning, which not only improves the efficiency of lithology identification and mapping, but also has strong objectivity, further improving the accuracy of lithology identification and mapping.

[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Example 1.

[0038] like Figure 1 As shown, this embodiment provides an underground lithology identification method based on machine learning, including the following steps.

[0039] S1: Divide the study area into grids to obtain multiple underground grid units in the study area.

[0040] S2: Obtain the physical properties of underground rocks and minerals in each of the underground grid units; the physical properties include density, magnetic susceptibility and resistivity.

[0041] S3: For each of the physical properties of the underground rock ores in each of the underground grid units, feature extraction is performed on the physical property according to the spatial position of the underground grid unit corresponding to the physical property to obtain the spatial physical property feature of the physical property.

[0042] S4: For each of the underground grid units, the spatial physical property characteristics of all physical properties of underground rocks and minerals in the underground grid unit are input into a trained underground lithology recognition model to obtain the underground lithology of the underground grid unit.

[0043] The underground lithology identification method based on machine learning in this embodiment is mainly divided into: prior information collection and reference inversion model formation, constrained inversion of gravity, magnetic and electrical data, physical property and lithology relationship analysis based on triangular coordinate system, lithology identification and filling based on machine learning, and the like. Figure 4 The method comprises the following steps.

[0044] 1. Collection of prior information and formation of reference inversion model.

[0045] S2 obtains the physical properties of underground rocks and minerals in each of the underground grid cells, specifically including: obtaining underground drilling information of drilling locations in the study area; the underground drilling information includes the real physical properties of underground rocks and minerals in the underground grid cells passed by the drilling; based on the underground drilling information, interpolation processing is performed on the underground grid cells of the undrilled locations in the study area to obtain the preliminary physical properties of the underground rocks and minerals in each of the underground grid cells; for each of the underground grid cells, based on the surface data of the study area, the preliminary physical properties of the underground rocks and minerals in each of the underground grid cells are used as initial values, and the underground drilling information is used as a constraint, and the final physical properties of the underground rocks and minerals in each of the underground grid cells are calculated through constraint inversion; the final physical properties are the physical properties of the underground rocks and minerals in each of the underground grid cells; the surface data include gravity, magnetic and electrical data observed above the surface.

[0046] Based on the underground drilling information, interpolation processing is performed on the underground grid cells at the undrilled positions in the study area, specifically including: taking the center point coordinates of several underground grid cells in the underground drilling information as interpolation nodes, taking the real physical properties of underground rocks and ores in the several underground grid cells in the underground drilling information as the accurate values ​​of the interpolated functions on the interpolation nodes, and using a linear interpolation method in which the interpolation function is a first-order polynomial to interpolate the underground grid cells at the undrilled positions in the study area.

[0047] Based on the surface data of the study area, taking the preliminary physical properties of the underground rocks and minerals in each of the underground grid units as initial values ​​and taking the underground drilling information as constraints, the final physical properties of the underground rocks and minerals in each of the underground grid units are calculated through constraint inversion, specifically including: based on the known mathematical mapping between the surface data of the study area and the physical properties of the underground rocks and minerals in the underground grid units of the study area, using the surface data of the study area and the underground drilling information, the final physical properties of the underground rocks and minerals in each of the underground grid units are calculated using discrete physical property grid inversion.

[0048] Prior information is in various forms and types. When collecting information, the main information is physical property information (physical property information, i.e., physical properties, including density, magnetic susceptibility and resistivity) or information that may be converted into physical property information. In the actual geophysical exploration process, the collected prior information is often limited and unevenly distributed. Therefore, the collection of prior information is mostly sparse prior information, which can be divided into two categories: one is the data that directly obtains physical properties, such as well logging, ground physical property samples, drill hole samples, etc.; the other is the data that needs to be converted to obtain physical properties, including geological mapping, seismic wave velocity obtained by inversion, etc. This embodiment mainly takes the surface geological map as the reference object, converts the surface geology into density, magnetic susceptibility and resistivity, and constructs the corresponding reference inversion model.

[0049] Among them, surface geology includes: rock lithology, geological structure, rock density, magnetic susceptibility and resistivity. Based on the geological laws such as rock lithology and geological structure observed on the surface, the initial model is established by interpolating the density, magnetic susceptibility and resistivity of the surface rock below the surface using artificial (geologist) experience. This initial model is the reference inversion model.

[0050] 2. Constrained inversion of gravity, magnetic and electrical data.

[0051] 1. Gravity data acquisition and preprocessing: Gravity is a commonly used geophysical method that infers underground properties and structures by measuring the gravity field at and above the surface of the study area. In terms of data acquisition, it is necessary to collect high-precision gravity measurement data, usually using a gravimeter and a global positioning system (GPS) to obtain location information. In order to improve data quality, preprocessing steps such as data smoothing, denoising, and instrument calibration are required.

[0052] 2. Magnetic data acquisition and preprocessing: Magnetic method is another commonly used geophysical method that obtains underground information by measuring the geomagnetic field. Data acquisition requires the use of magnetometers and GPS for measurement, as well as magnetic field drift correction and noise removal. In addition, geomagnetic data usually requires magnetic susceptibility conversion to obtain underground magnetic distribution.

[0053] 3. Electrical data acquisition and preprocessing: For electrical methods, select appropriate electromagnetic instruments (geoelectric, geomagnetic, electromagnetic sensors) and deploy them in the study area according to the research purpose. Determine the location, direction and parameter settings of the measurement, such as the layout of the measurement line grid, frequency range, etc. Measure at the appropriate time and conditions to ensure the quality and accuracy of the data. Preprocess the data through denoising, correction, data format conversion, etc. Check the integrity and accuracy of the data, eliminate potential errors or outliers, and other data quality assessments.

[0054] 4. Inversion of gravity, magnetic and electrical data: The three-dimensional inversion algorithms of gravity and magnetic methods are relatively mature, and there is a lot of room for choice. This embodiment uses the physical property inversion software UBCMag3D and Grav3D developed by the University of British Columbia, Canada. The core algorithm of the software is based on the gravity and magnetic inversion algorithm proposed by Li Yaoguo et al. (Li Yaoguo, 1996, 1998). The software provides an interface for carrying out prior information constraints, which can facilitate inversion with prior constraints.

[0055] The electrical inversion algorithm is an optimization process, so a large number of optimization algorithms are widely used in various geophysical inversion practices. Currently, the commonly used inversion algorithms can be summarized into two categories: nonlinear inversion algorithms and linear iterative algorithms. In general, for 2D and 3D problems, the currently feasible inversion algorithm is the linear iterative algorithm. This embodiment uses a linear iterative algorithm for electrical inversion.

[0056] 5. Constrained inversion of gravity, magnetic and electrical data: Constrained inversion uses the reference inversion model constructed in step 1, “Collecting prior information and forming a reference inversion model”, as the initial condition for inversion. Under the premise of sparsely limiting the average value of physical properties, boundaries, and smoothing factors of the inversion model, the density, magnetic susceptibility, and resistivity of the grid cells with determined physical property data are fixed and remain unchanged during the inversion iteration process. The physical property (density, magnetic susceptibility, and resistivity) data of the grid cells with determined physical property data are obtained through core physical property measurements obtained through drilling. The physical properties of the “grid cells with determined physical property data” are interpolated to the “grid cells without determined physical properties and upper and lower boundaries” to obtain the initial model for inversion.

[0057] For underground grid cells whose upper and lower boundaries of physical properties can be obtained, the inverted physical properties are restricted to not exceed the upper and lower boundaries during the iteration process. For underground grid cells without definite physical properties and upper and lower boundaries, their physical property boundaries are extrapolated through the nearest underground grid cells with definite physical properties to limit the range of physical property changes during the inversion iteration process, so as to add mathematical constraints to the inversion formula. At the same time, according to the prior information constructed in step 1 "prior information collection and reference inversion model formation", for grid cells that may have sudden changes in physical properties such as faults, their smoothing coefficients are reduced, thereby realizing the three-dimensional inversion of gravity, magnetic and electrical data with prior geological information constraints, and obtaining the three-dimensional results of the three physical properties of underground density, magnetic susceptibility and resistivity, that is, obtaining the physical properties of each grid in the study area.

[0058] It should be noted that whether there are sudden changes in physical properties such as faults in the underground grid unit is determined by prior information, which refers to the sudden changes in the physical properties in the vertical direction obtained by measuring the physical properties of the core obtained through drilling.

[0059] The basic principles of gravity and magnetic inversion are as follows: Gravity and magnetic inversion is also called grid inversion, and its goal is to find a three-dimensional function that can describe the distribution of physical properties. During the inversion calculation, the underground half-space is divided into a large number of grid cells, and each underground grid cell has an unknown fixed physical property value. The divided cells should be small enough until the same inversion result is obtained by reducing the cell size for inversion. From a mathematical point of view, this type of inversion problem has a greater number of unknowns than the number of data, and is an underdetermined problem. The general method to solve the underdetermined problem is to use a set of solutions to well-posed problems that are "adjacent" to the original ill-posed problem to approximate the solution to the original problem, which is the regularization method. In order to obtain an effective result, it is necessary to incorporate some forms of regularization. In this process, the choice of regularization method is very important because it is also the main way to incorporate geological information in this type of inversion. In order to obtain a meaningful result, a suitable model objective function is designed. To obtain, The general formula is formula (1).

[0060] (1).

[0061] Among them, m is the model to be sought; The coefficients indicate the degree of smoothness in different directions; is a weighting function; Represents the minimum contribution of the difference between the target function model and the reference model; A weighting function representing the degree of closeness of the physical characteristics between the target function model and the reference model; is the reference model; Represents the smoothness contribution of the difference between the target function model and the reference model in the x direction; A weighting function representing the smoothness of the physical characteristics between the target function model and the reference model in the x direction; Represents the smoothness contribution of the difference between the target function model and the reference model in the y direction; A weighted function representing the smoothness of the physical characteristics between the target function model and the reference model in the y direction; Represents the smoothness contribution of the difference between the target function model and the reference model in the z direction; A smooth weighting function that represents the degree of closeness of the physical characteristics between the target function model and the reference model in the z direction.

[0062] During inversion, the parameters in the above formula need to be set, and the understanding of the initial reference model determines the complexity of the final objective function. For example, if the study area is an area where geological and geophysical work has never been conducted, and there is little or no prior information, the reference model is considered to be the most basic underground uniform half-space. However, if the study area is a known mining area and the researchers have a certain understanding of the geological conditions of the area, the reference model can have a more complex geological structure, that is, the reference model function will be more complicated, which will inevitably lead to the objective function is also more complicated. The coefficients can vary greatly depending on the specific situation. In addition, different weighting functions (such as depth weighting and distance weighting) can be used to help integrate prior information.

[0063] The construction of the model objective function is an important and arduous task. It is the most important part of this type of inversion problem because the characteristics and structures shown in the final model are derived from the objective function. In short, the inversion problem is an optimization problem, that is, to minimize the following equation (2).

[0064] (2).

[0065] In formula (2), is a trade-off parameter or Tikhonov parameter that is adjusted during the inversion process. , so that a model with satisfactory fitting error can be obtained in the end. When solving this problem, the underground half-space is divided into a large number of cell units (underground grid units) with unknown attribute values. This gridded underground half-space method is used to discretize the model objective function and forward equation, and the minimized objective function is formula (3).

[0066] (3).

[0067] Where W d and W m are matrices, W d is the data covariance matrix; W m is the model covariance matrix; F is the forward simulation operator; is the observed data. At this point, the objective function is decomposed into a gradient equation. There are many ways to solve it. The most typical method is the Gauss-Newton iteration method. This method is used to solve the equation to obtain equation (4). The inversion result is obtained by iterating equation (4). In each iteration, a solution is obtained. .

[0068] (4).

[0069] In the above formula, J is the sensitivity, where , d i is the i-th observation data, m j is the j-th model element; is the model residual; is the gradient.

[0070] The Gauss-Newton iteration method is widely used. Its basic idea is to use Taylor series expansion to approximate the nonlinear regression model, and then through multiple iterations and multiple corrections of the regression coefficient, the regression coefficient is constantly approaching the optimal regression coefficient of the nonlinear regression model, and finally the residual sum of squares of the original model is minimized. This method can be used for the inversion calculation of different geophysical data, and can also be applied to discrete body inversion and lithology inversion to estimate its parameters.

[0071] The basic principle of electrical inversion is as follows: The inverse problem of geophysical electrical method is essentially a mapping problem from data space to model space. Since the observation data exists objectively, the solution to the inverse problem must exist. Using finite-dimensional data space to solve infinite-dimensional model space, the solution is non-unique, and due to the existence of observation data errors, the solution is unstable. Therefore, the geophysical inverse problem is an ill-posed problem. In order to solve this ill-posed problem, according to the Tikhonov regularization idea, constraints can be imposed on the inversion objective function to obtain the optimal solution in a certain sense. The inverted objective equation Written as formula (5).

[0072] (5).

[0073] in, is the data fitting term, is the model fitting term, It is a regularization factor, which plays a trade-off between the data fitting term and the model fitting term. is an N-dimensional data vector, is an M-dimensional model vector; is the reference model or a priori model; For forecast data. is the data covariance matrix, expressed as an N×N diagonal matrix, with the inverse of the data standard deviation on the diagonal; is the model covariance matrix. In the electromagnetic field, the data space and the model space are nonlinear mappings. By taking the derivative of the objective function, we can get equation (6).

[0074] (6).

[0075] in, The target equation The derivative of .

[0076] The gradient of the model fitting term It can be expressed as formula (7).

[0077] (7).

[0078] The gradient of the data fitting term can be expressed as formula (8).

[0079] (8).

[0080] in, express .

[0081] Electrical inversion is the process of optimizing the objective function (5).

[0082] Taking the gravity, magnetic and electrical data inversion of the northeastern Jiangxi mining area as an example, the gravity inversion results show that the density of the northeastern Jiangxi mining area is as follows: Figure 2 As shown in the magnetic inversion results, the magnetic susceptibility of the northeastern Jiangxi mining area is as follows Figure 3 As shown in the electrical inversion results, the resistivity of the northeastern Jiangxi mining area is as follows Figure 4 shown.

[0083] This embodiment, in addition to inputting the spatial physical property characteristics of all physical properties of underground rocks and ores in the underground grid unit into the trained underground lithology identification model, also includes: training the underground lithology identification model, which specifically includes the following steps.

[0084] (1) Obtaining a data set for training an underground lithology recognition model; the data set includes input features of N sample areas and sample underground lithologies of all sample underground grid cells in each of the sample areas; the input features include sample spatial physical property features of all sample physical properties of underground rocks and ores in all sample underground grid cells; N is greater than a set value.

[0085] (2) Taking the input features of each sample area in the data set as input and the sample underground lithology of all sample underground grid cells in each sample area in the data set as output labels, an underground lithology recognition model is trained to obtain a trained underground lithology recognition model.

[0086] The first underground lithology identification model is obtained by training using the SVM machine learning algorithm and the data set; the second underground lithology identification model is obtained by training using the RF machine learning algorithm and the data set; the first underground lithology identification model and the second underground lithology identification model are weightedly fused to obtain a trained underground lithology identification model.

[0087] The method for obtaining the sample underground lithology of all sample underground grid units in the sample area in the data set specifically includes the following steps.

[0088] The physical properties of underground rocks and minerals in all sample underground grid units in the sample area and the lithology of rock and mineral samples in the sample area are obtained; the rock and mineral samples include rock and mineral samples collected from the surface and rock and mineral samples obtained by drilling.

[0089] Based on a triangular coordinate system, the physical properties of the rock and mineral specimens in the sample area are used as the coordinate axes of the triangular coordinate system, the x-axis range of the triangular coordinate system is the density value, the y-axis range of the triangular coordinate system is the resistivity value, the z-axis range of the triangular coordinate system is the magnetic susceptibility value, and the lithology of the sample area and the rock and mineral specimens are used as symbolic identifiers to obtain the triangular coordinate system correspondence between the physical properties of the rock and mineral specimens and the lithology of the rock and mineral specimens.

[0090] Based on the triangular coordinate system correspondence between the physical properties of the rock and mineral samples and the lithology of the rock and mineral samples, the physical properties of the underground rock and mineral in the sample underground grid cells of the sample area are converted into the sample underground lithology of the sample underground grid cells of the sample area.

[0091] 3. Analysis of the relationship between physical properties and lithology based on triangular coordinate system.

[0092] Physical properties are the link and bridge between geophysics and geology, and detailed physical property research is a key step in distinguishing lithology. The density, magnetic susceptibility and resistivity combination characteristics and logical relationships corresponding to different rock samples obtained from ground collection and cores are analyzed. Based on the triangular coordinate system, this embodiment identifies the corresponding underground lithology through the scatter plot of the density, magnetic susceptibility and resistivity triangular coordinate system, that is, based on the relationship between the physical properties of the triangular coordinate system and the underground lithology, the label of the spatial physical property characteristics of the physical attributes of each underground grid unit can be obtained.

[0093] Taking the physical property data of rocks and minerals in the northeastern Jiangxi mining area as an example, the density, magnetic susceptibility and resistivity of the main rock types were statistically analyzed. The qualitative summary rule is: different underground lithologies correspond to different combinations of density, magnetic susceptibility and resistivity.

[0094] (1) Quaternary system, mainly low-resistance, low-density and weak magnetic susceptibility layers.

[0095] (2) Volcanic rocks, generally of medium resistivity (n×10 2 Ω·m), the resistivity value is lower than that of intrusive rock. Medium density, the average density of surface samples is 2.50×10 3 kg / m 3 , among which the density of tuff is relatively low, which is 2.44×10 3 kg / m3 , andesite and trachyandesite have a higher density of 2.54×10 3 kg / m 3 ; In the borehole, the density of volcanic rock increased significantly, with an average value of 2.70×10 3 kg / m 3 The density of trachyandesite is 2.73×10 3 kg / m 3 , which can cause significant gravity anomalies. The magnetic susceptibility is generally low to medium, but varies greatly.

[0096] (3) Sandstone is a layer with low resistivity, medium-low density and weak magnetic susceptibility.

[0097] (4) Limestone is a high-resistance layer with an average resistivity of 5×10 3 Ω·m, distributed in layers, is the electrical marker layer of the area; high density, the average density of surface samples is 2.70×10 3 kg / m 3 ; Weakly magnetic layer.

[0098] (5) Shale, the resistivity value is (1~2)×10 3 Ω·m, is the low-resistance marker layer in the study area; medium-low density, weak magnetic layer.

[0099] (6) The magnetism of intrusive rocks and subvolcanic rocks increases from acidic to neutral to basic. Among them, syenite and diorite have medium magnetism, while diorite porphyry has slightly lower magnetism.

[0100] Magnetite, density is 3.83×10 3 kg / m 3 , the magnetic susceptibility is 95300×10 -5 SI.

[0101] According to the rectangular coordinate system, a scatter plot of the relationship between density and magnetic susceptibility and lithology is made, such as Figure 5 As shown, a scatter plot of the relationship between density, magnetic susceptibility and resistivity and lithology is made based on triangular coordinates according to this embodiment (as shown in FIG. Figure 6 As shown in the figure, it can be found that the result of the "physical property-lithology" relationship in this embodiment is more intuitive, the classification is clearer and more accurate.

[0102] 4. Lithology identification and mapping based on machine learning.

[0103] 1. Feature extraction.

[0104] In S3, the physical attribute is feature extracted according to the spatial position of the underground grid unit corresponding to the physical attribute to obtain the spatial physical property feature of the physical attribute, which specifically includes: considering the spatial distribution of the physical attribute of the underground rock and ore, taking the spatial position of the underground grid unit as a three-dimensional coordinate system (x, y, z), and taking the physical attribute of the underground rock and ore in the underground grid unit as a function value W (x, y, z); W (x, y, z) includes W_den (x, y, z), W_sus (x, y, z) and W_res (x, y, z); W_den (x, y, z) is density, W_sus(x,y,z) is magnetic susceptibility, and W_res(x,y,z) is resistivity; feature extraction is performed on the function value W(x,y,z) of the physical property of underground rock and ore in the underground grid unit to obtain local and global features of the three-dimensional coordinate system (x,y,z) of the spatial position of the underground grid unit, and the local and global features are the spatial physical property features of the physical property; the local and global features include the function value of the physical property, the minimum value of the function value, the maximum value of the function value, the mean value of the function value, the median value of the function value, the variance of the function value and the gradient of the function value.

[0105] Spatial physical property characteristics: Considering the spatial distribution of underground rock properties, local and global physical property (density, magnetic susceptibility and resistivity) characteristics can be extracted, including the mean, median, variance, minimum, maximum and spatial gradient of the physical properties.

[0106] Taking the physical property of density W_den (x, y, z) as an example, the mean value of its density function value is The calculation method of is as follows.

[0107] .

[0108] in, Indicates the position of the underground grid unit at the ith x-axis, j-axis, and k-th z-axis. There are a total of m grid divisions on the y-axis, where i represents its serial number; there are a total of n grid divisions on the z-axis, where k represents its serial number.

[0109] In the data set, feature extraction is performed on the physical properties of the training sample set (rock samples with measured physical properties) and the test set to be applied (underground rock three-dimensional results of physical properties obtained by inversion) to obtain the sample spatial physical property characteristics of the sample physical properties of underground rocks and ores in the sample grid units.

[0110] 2. Data representation.

[0111] Feature vector: The sample spatial physical property characteristics of the extracted sample physical properties are combined into a feature vector and put into the triangular coordinate system to obtain the label corresponding to the sample spatial physical property characteristics (sample underground lithology), which is used to input the machine learning model and process. Underground lithology includes siltstone, fine sandstone, dolomite, etc. That is, for each physical property (density, magnetic susceptibility and resistivity) of each underground grid unit, its mean, median, variance, minimum, maximum and spatial gradient are calculated, and then the physical property itself and the mean, median, variance, minimum, maximum and spatial gradient corresponding to the physical property are combined into a feature vector (that is, the spatial physical property characteristics of the physical property are obtained), that is, density corresponds to a feature vector, magnetic susceptibility corresponds to a feature vector, and resistivity corresponds to a feature vector.

[0112] When training the model, the labels (underground lithology) are numerically represented, such as -1, -0.5, 0.5, etc. "-1" represents siltstone, "-0.5" represents fine sandstone, and "0.5" represents dolomite.

[0113] 3. Machine learning model selection.

[0114] When performing underground lithology identification, it is critical to select a suitable machine learning model. This embodiment selects the following two supervised machine learning methods that specialize in classification and identification, and performs weighted fusion on the machine model results to obtain more accurate lithology identification and mapping results.

[0115] (1) Support vector machine (SVM).

[0116] SVM is a supervised learning model for classification and regression, which can be used for lithology classification by finding the optimal hyperplane to separate different lithology categories.

[0117] Support vector machine (SVM) is a representative algorithm in machine learning. It is based on binary classification algorithm, and its core idea is dimensionality increase and linearization (Zhou Yongzhang et al., 2018). Support vector machine is often used to solve linear inseparable problems. It plots each data item as a point in n-dimensional space (where n refers to the number of features), and the value of each feature is the value of a specific coordinate. Then, a nonlinear function (kernel function) is used to map the original feature space to a higher-dimensional Hilbert space, thereby converting the nonlinear separable problem into a linear separable problem. The decision boundary for performing the classification task is the maximum margin hyperplane for solving the learning sample, which can also be extended to the regression problem to obtain support vector regression (SVR). Figure 8 and Fig. 9 As shown, Figure 8 is a separable feature space; Fig. 9 is an inseparable feature space. Figure 8 and Fig. 9 middle, That is .

[0118] The specific working principle of SVM is to find the points closest to the separating hyperplane and ensure that they are as far away from the separating plane as possible. The support vectors are the points closest to the separating hyperplane. In the sample space, the dividing hyperplane can be described in the following form.

[0119] (9).

[0120] in, is the normal vector, which determines the direction of the hyperplane; is the displacement term, which determines the distance between the hyperplane and the origin. It is recorded as . Any point in the sample space To the hyperplane Distance It can be written as formula (10).

[0121] (10).

[0122] The constant in the above formula (10) is Similar to the intercept in logistic regression With the hyperplane, we can discuss the problem of classifiers and determine a classifier. For any real number, when , the underground lithology identification result of the i-th sample ; , With a classifier, the goal is to find the data point with the smallest interval, and then maximize it and find the parameters , This can be written as equation (11).

[0123] (11).

[0124] Obviously, , , m is the total number of samples, in order to maximize the interval, we only need to maximize , which is equivalent to minimizing Therefore, the basic form of the support vector machine is as follows (12) and (13).

[0125] (12).

[0126] (13).

[0127] The above equation itself is a convex quadratic programming problem, which can be solved using an existing optimization computing package, or using a more efficient method. Using the Lagrange multiplier method on the above equation can yield its dual problem. The Lagrangian function of the problem can be written as equation (14).

[0128] (14).

[0129] make right and The partial derivative of is 0, and we can get the following formula.

[0130] (15).

[0131] (16).

[0132] Substitute the above formula into The final optimization objective function is obtained, which is formula (17).

[0133] (17).

[0134] The constraint condition is formula (18).

[0135] (18).

[0136] Solve Then, find the model parameters and The first underground lithology identification model can be obtained.

[0137] When dealing with linearly inseparable problems, SVM usually transforms data from one feature space to another. In the new feature space, there are often clearer test results. Therefore, if the original space is finite-dimensional, that is, the attributes are finite, then there must be a high-dimensional feature space that makes the samples separable.

[0138] The advantages of the SVM algorithm are: (1) It is an applicable small sample learning method with a solid theoretical foundation. It basically does not involve probability measurement and the law of large numbers, and simplifies common classification and regression problems; (2) The final decision function of the SVM is determined by only a few support vectors. The complexity of the calculation depends on the number of support vectors, not the dimension of the sample space, which in a sense avoids the "curse of dimensionality"; (3) A few support vectors determine the final result, which not only captures the key samples and "eliminates" a large number of redundant samples, but also determines that the method is not only simple in algorithm, but also has good "robustness"; (4) The SVM learning problem can be expressed as a convex optimization problem, so the global minimum of the objective function can be found using known effective algorithms; (5) It has excellent generalization ability. Disadvantages are: (1) The SVM algorithm is difficult to implement for large-scale training samples; (2) It is difficult to use SVM to solve multi-classification problems; (3) It is sensitive to the selection of parameters and kernel functions.

[0139] (2) Random Forest (RF machine learning algorithm).

[0140] Random forest is an ensemble learning model that can handle high-dimensional data and complex features. It can be used for lithology classification and feature importance assessment.

[0141] The basis of random forest is decision tree. Decision tree uses tree structure to build classification model. Each node represents an attribute. According to the division of this attribute, it enters the child node of this node until the leaf node. Each leaf node represents a certain category, so as to achieve the purpose of classification. Commonly used decision trees are ID3, C4.5, and CART. The ID3 algorithm uses information gain g (D, A) for feature selection; the C4.5 algorithm goes a step further and uses information gain rate g (D, A) / H (A) for feature selection; and the CART algorithm uses the Gini coefficient to select the partition attribute. In the process of generating a tree, it is necessary to choose which feature to use for splitting. Generally speaking, the principle of selection is to improve the purity as much as possible after separation, and this purity can be measured by indicators such as information gain, information gain rate, and Gini coefficient. If it is a tree, in order to avoid overfitting, pruning is also required to remove those nodes that may cause the error of the validation set to increase.

[0142] A decision tree is a process of classifying data through a series of rules. It provides a rule-like method for what value will be obtained under what conditions. Decision trees are divided into classification trees and regression trees. Classification trees are used to make decision trees for discrete variables, while regression trees are used to make decision trees for continuous variables. One of the important reasons why decision tree algorithms are so popular is that users basically do not need to understand machine learning algorithms or delve into how they work. Intuitively, a decision tree classifier is like a flowchart consisting of a judgment module and a termination block. The termination block represents the classification result (that is, the leaves of the tree). The judgment module represents the judgment of the value of a feature (the judgment module has several branches if there are several values ​​for the feature). If efficiency is not considered, the judgment of all features of the sample will eventually classify a sample into a class termination block. In fact, some of the features of the sample play a decisive role in classification. The construction process of a decision tree is to find these decisive features, construct an inverted tree according to their degree of decisiveness, and use the feature with the greatest decisiveness as the root node. Then, recursively find the second largest decisive feature in the sub-dataset under each branch until all data in the sub-dataset belong to the same class. Therefore, the process of constructing a decision tree is essentially a recursive process of classifying data sets according to data features. The first problem that needs to be solved is which feature of the current data set plays a decisive role in dividing data into categories.

[0143] The generation process of a decision tree is mainly divided into the following three parts.

[0144] (1) Feature selection: Feature selection refers to selecting a feature from the numerous features in the training data as the splitting criterion for the current node. There are many different quantitative evaluation criteria for how to select features, which lead to different decision tree algorithms.

[0145] (2) Decision tree generation: According to the selected feature evaluation criteria, child nodes are recursively generated from top to bottom until the data set is indivisible, at which point the decision tree stops growing. For tree structures, recursive structures are the easiest to understand.

[0146] (3) Pruning: Decision trees are prone to overfitting and generally require pruning to reduce the size of the tree structure and alleviate overfitting. There are two types of pruning techniques: pre-pruning and post-pruning.

[0147] Random forest generally refers to a classifier that uses multiple decision trees to train and predict samples. Decision trees use tree structures to build classification models. Each node represents an attribute. According to the division of this attribute, the child nodes of this node are entered until the leaf node. Each leaf node represents a certain category, thereby achieving the purpose of classification. The order of each feature item of the decision tree model can be determined according to the actual situation. Usually, the closer to the root node, the stronger the classification ability. The prediction decision tree model is arranged from high to low according to the resolution of data collection. It represents the threshold of segmentation. The principle of feature splitting is to improve the purity as much as possible after separation. It can be measured by information gain, gain rate, and Gini coefficient. During training, the number of feature items, order of priority, and threshold size are continuously tried until the sample purity on the leaf node is the highest, and the segmentation is terminated. The segmentation path and threshold are recorded for prediction.

[0148] Random forest belongs to the bagging method in ensemble learning. It achieves the optimal solution by integrating the advantages of multiple decision trees. The detailed process is as follows: (1) N training samples are randomly selected with replacement from the original sample set, and k rounds of extraction are performed to obtain k training sets (the k training sets are independent of each other and elements can be repeated); (2) For the k training sets, k decision tree models are constructed and trained. The number and order of feature items of each decision tree are random. The split threshold with the largest information gain (information entropy before split minus information entropy after split) is found for the extracted features and applied to the node split until the tree grows to the maximum extent; (3) The trained model is used to predict data in the unknown area of ​​the result. According to the nature of the output result (discrete or continuous), it can be divided into classification problem and regression problem. For classification problem, the classification result is generated by voting. For regression problem, the average of the prediction results of k models is used as the final prediction result. All models have the same weight.

[0149] Through the above process, the second underground lithology identification model can be obtained.

[0150] Weighted fusion of machine model results.

[0151] In order to improve the performance of integrated learning as much as possible, this embodiment proposes a weighted fusion method based on the known sample prediction performance size and global data prediction result analysis. The specific approach is: if a certain algorithm has the best prediction performance, it is given the highest weight to participate in model fusion. When dealing with classification problems, the weight is quantified according to its prediction accuracy on known samples and the percentage of predicted values ​​of 1 in global prediction.

[0152] (19).

[0153] (20).

[0154] in, Represents the weight of the nth classification model (underground lithology identification model) participating in the fusion (the weight assigned by ensemble learning), A measure representing the classification performance of the nth subsurface lithology identification model, Represents the accuracy of the model on the test set, Represents the percentage of the model's results being 1 during global prediction.

[0155] (twenty one).

[0156] in, is the model fusion result; V n is the output result of the nth model.

[0157] 4. Machine learning model selection and evaluation.

[0158] After obtaining the first underground lithology identification model, it also includes: using performance evaluation indicators to perform performance evaluation on the first underground lithology identification model; the performance evaluation indicators include accuracy, recall rate, precision, ROC curve and area under AUC curve.

[0159] (1) Based on the "physical property and lithology relationship analysis based on the triangular coordinate system" in step 3, the obtained "physical property and lithology relationship based on the triangular coordinate system" is used as a known training sample set and label for machine learning to judge and identify lithology. The selected "machine learning method" is trained using the training sample set.

[0160] (2) The machine learning model for lithology identification needs to be evaluated and optimized. The performance evaluation indicators used in this embodiment are as follows: Accuracy: The classification accuracy of the classification model is an important indicator used to measure the model's ability to correctly classify lithology. Recall and precision: The recall rate measures the proportion of positive examples correctly identified by the model, and the precision measures the proportion of true positive examples in the model's prediction results. ROC curve and AUC: The ROC curve and AUC (area under the curve) are used to evaluate the performance of the classification model.

[0161] After the model performance evaluation, the model parameters can be adjusted and optimized as needed to improve the accuracy and robustness of lithology identification.

[0162] 5. Lithology identification and mapping based on machine learning.

[0163] (1) In order to convert the physical property information of underground rocks into features that can be used for lithology identification, physical property feature extraction and data representation are required. The three-dimensional density, magnetic susceptibility and resistivity of underground rocks obtained by the "constrained inversion of gravity, magnetic and electrical data" in step 2 are grid-unitized, and the physical property features of underground rocks obtained by the "constrained inversion of gravity, magnetic and electrical data" are extracted and represented.

[0164] (2) Use the trained machine learning model to find the correspondence between the density, magnetic susceptibility and resistivity of underground rocks obtained by inversion of geophysical data (gravity, magnetic and electrical methods) and the lithology, and display them through slices and equipotential surfaces to determine the three-dimensional spatial morphology and mutual relationship of different lithologies, thereby realizing the classification, identification and mapping of underground lithologies based on machine learning methods.

[0165] Taking the lithology identification and mapping of the northeastern Jiangxi mining area as an example, Figure 7 These are the results of lithology identification and mapping in the northeastern Jiangxi mining area.

[0166] This embodiment has the following beneficial effects: (1) In view of the insufficient vertical resolution of gravity and magnetism, the influence of multiple physical factors and the uncertainty of inversion, the present invention adds the electrical inversion results to the lithology identification and mapping process based on gravity and magnetic inversion, so as to improve the vertical resolution of the inversion results, improve the uncertainty of inversion, improve the accuracy of inversion, and thus improve the accuracy of lithology identification and mapping. In this process, the complementarity of the underground rock physical property information obtained by various geophysical data to the understanding of underground rock lithology is fully explored. (2) In view of the problem of coordinated processing of lithology analysis of various geophysical inversion results, and the unclear and overlapping classification of "physical property-lithology relationship corresponding to lithology based on rectangular coordinate system of density and magnetic susceptibility" compared with the existing solution, the present invention introduces a triangular coordinate system to map the physical property-lithology relationship of gravity, magnetic and electrical inversion results, and further focuses on the relationship between the three physical property parameters and lithology parameters. If the rectangular coordinate system method of the existing scheme is used, three sets of physical property-lithology relationship mapping are required, namely, "scatter plots of the relationship between the two physical properties of density and magnetic susceptibility and lithology", "scatter plots of the relationship between the two physical properties of density and resistivity and lithology", and "scatter plots of the relationship between the two physical properties of magnetic susceptibility and resistivity and lithology". The classification results are not focused, not easy to understand, and the actual operation capabilities of underground lithology identification and mapping are poor. The present invention is based on a triangular coordinate system. Through the scatter plots of the triangular coordinate system of density, magnetic susceptibility and resistivity, in the process of determining the correspondence between physical properties and lithology, not only the physical property information of the resistivity dimension is added, but also the three physical property information is used to correspond to the core, which also improves the coordination of the physical property-lithology relationship, reduces the "overlapping" area of ​​the physical property-lithology relationship diagram, further clarifies the correspondence and classification of physical properties and lithology, and improves the accuracy of lithology identification and mapping. (3) It is very cumbersome and difficult to classify the correspondence between multiple physical properties and different lithologies. Even if a triangular coordinate system is introduced, there is still an "overlapping" area between the correspondence between physical properties and lithologies, resulting in the inability to complete underground lithology identification and mapping simply relying on the physical property-lithology relationship. This embodiment proposes to use two machine learning methods for training, classification and identification. These two types of machine learning methods are important and key means in the field of "classification" and are very similar to the lithology identification problem to be solved by the present invention. They not only improve the efficiency of lithology identification and mapping, but also further improve the accuracy of lithology identification and mapping.

[0167] The solution of this embodiment is to perform more accurate lithology identification and mapping using the underground physical property results obtained by inversion of geophysical gravity, magnetic and electrical methods based on machine learning methods, which can obtain more accurate underground lithology identification and mapping.

[0168] Example 2.

[0169] In order to execute the method corresponding to the above-mentioned embodiment 1 to achieve the corresponding functions and technical effects, a machine learning-based underground lithology identification system is provided below, including the following modules.

[0170] The grid division module is used to divide the study area into grids to obtain multiple underground grid units in the study area.

[0171] The physical property acquisition module is used to acquire the physical properties of underground rocks and ores in each of the underground grid units; the physical properties include density, magnetic susceptibility and resistivity.

[0172] The spatial physical property feature extraction module is used to extract the feature of each physical property of the underground rock ores in each underground grid unit according to the spatial position of the underground grid unit corresponding to the physical property, so as to obtain the spatial physical property feature of the physical property.

[0173] The underground lithology prediction module is used to input the spatial physical property characteristics of all physical properties of underground rocks and ores in each underground grid unit into a trained underground lithology recognition model to obtain the underground lithology of the underground grid unit.

[0174] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0175] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for identifying underground lithology based on machine learning, characterized in that: include: Gridding the study area to obtain a plurality of underground grid units in the study area; Obtaining the physical properties of underground rocks and minerals in each of the underground grid cells, specifically including: Acquire underground drilling information of drilling locations in the study area; the underground drilling information includes real physical properties of underground rocks and minerals in underground grid cells passed by the drilling; Based on the underground drilling information, interpolation processing is performed on underground grid cells at undrilled locations in the study area to obtain preliminary physical properties of underground rocks and minerals in each of the underground grid cells; For each of the underground grid cells, based on the surface data of the study area, taking the preliminary physical properties of the underground rocks and minerals in each of the underground grid cells as initial values, taking the underground drilling information as constraints, the final physical properties of the underground rocks and minerals in each of the underground grid cells are calculated through constraint inversion; the final physical properties are the physical properties of the underground rocks and minerals in each of the underground grid cells; the surface data include gravity, magnetic and electrical data observed above the surface; the physical properties include density, magnetic susceptibility and resistivity; For each of the physical properties of the underground rock ores in each of the underground grid units, feature extraction is performed on the physical property according to the spatial position of the underground grid unit corresponding to the physical property to obtain the spatial physical property feature of the physical property, specifically including: Considering the spatial distribution of the physical properties of underground rocks and minerals, the spatial position of the underground grid unit is taken as a three-dimensional coordinate system (x, y, z), and the physical properties of the underground rocks and minerals in the underground grid unit are taken as function values ​​W(x, y, z); W(x, y, z) includes W_den(x, y, z), W_sus(x, y, z) and W_res(x, y, z); W_den(x, y, z) is density, W_sus(x, y, z) is magnetic susceptibility, and W_res(x, y, z) is resistivity; Perform feature extraction on the function value W(x, y, z) of the physical property of the underground rock ores in the underground grid unit to obtain local and global features of the three-dimensional coordinate system (x, y, z) of the spatial position of the underground grid unit, wherein the local and global features are spatial physical property features of the physical property; the local and global features include the function value of the physical property, the minimum value of the function value, the maximum value of the function value, the mean value of the function value, the median value of the function value, the variance of the function value and the gradient of the function value; For each of the underground grid units, the spatial physical property characteristics of all physical attributes of underground rocks and ores in the underground grid unit are input into a trained underground lithology recognition model to obtain the underground lithology of the underground grid unit.

2. The underground lithology identification method based on machine learning according to claim 1, characterized in that: Inputting the spatial physical property characteristics of all physical properties of underground rocks and ores in the underground grid unit into the trained underground lithology recognition model also includes: training the underground lithology recognition model, specifically including: Acquire a data set for training an underground lithology recognition model; the data set includes input features of N sample areas and sample underground lithologies of all sample underground grid cells in each of the sample areas; the input features include sample spatial physical property features of all sample physical properties of underground rocks and ores in all sample underground grid cells; N is greater than a set value; Taking the input features of each sample area in the data set as input and the sample underground lithology of all sample underground grid units in each sample area in the data set as output labels, the underground lithology recognition model is trained to obtain a trained underground lithology recognition model.

3. The underground lithology identification method based on machine learning according to claim 1, characterized in that: Based on the underground drilling information, interpolation processing is performed on underground grid cells at undrilled locations in the study area, specifically including: The center point coordinates of the underground grid cells passed by the borehole in the underground drilling information are used as interpolation nodes, and the real physical properties of the underground rock and ore in the underground grid cells passed by the borehole in the underground drilling information are used as the accurate values ​​of the interpolated functions on the interpolation nodes. The underground grid cells at the undrilled positions in the study area are interpolated using a linear interpolation method in which the interpolation function is a first-order polynomial.

4. The underground lithology identification method based on machine learning according to claim 1, characterized in that: Based on the surface data of the study area, taking the preliminary physical properties of the underground rocks and minerals in each of the underground grid cells as initial values ​​and taking the underground drilling information as constraints, the final physical properties of the underground rocks and minerals in each of the underground grid cells are calculated through constraint inversion, specifically including: Based on the known mathematical mapping between the surface data of the study area and the physical properties of underground rocks and minerals in the underground grid units of the study area, the surface data of the study area and the underground drilling information are used to calculate the final physical properties of the underground rocks and minerals in each of the underground grid units using discrete physical property grid inversion.

5. The underground lithology identification method based on machine learning according to claim 2, characterized in that: The method for obtaining the sample underground lithology of all sample underground grid units in the sample area in the data set specifically includes: Obtaining the physical properties of underground rocks and minerals in all sample underground grid cells in the sample area and the lithology of rock and mineral samples in the sample area; the rock and mineral samples include rock and mineral samples collected from the surface and rock and mineral samples obtained by drilling; Based on a triangular coordinate system, the physical properties of the rock and mineral specimens in the sample area are used as coordinate axes of the triangular coordinate system, the x-axis range of the triangular coordinate system is the density value, the y-axis range of the triangular coordinate system is the resistivity value, the z-axis range of the triangular coordinate system is the magnetic susceptibility value, and the lithology of the sample area and the rock and mineral specimens are used as symbolic identifiers to obtain a triangular coordinate system correspondence between the physical properties of the rock and mineral specimens and the lithology of the rock and mineral specimens; Based on the triangular coordinate system correspondence between the physical properties of the rock and mineral samples and the lithology of the rock and mineral samples, the physical properties of the underground rock and mineral in the sample underground grid cells of the sample area are converted into the sample underground lithology of the sample underground grid cells of the sample area.

6. The underground lithology identification method based on machine learning according to claim 2, characterized in that: The underground lithology recognition model is trained to obtain a trained underground lithology recognition model, which specifically includes: The first underground lithology identification model is obtained by training using the SVM machine learning algorithm and the data set; the second underground lithology identification model is obtained by training using the RF machine learning algorithm and the data set; the first underground lithology identification model and the second underground lithology identification model are weightedly fused to obtain a trained underground lithology identification model.

7. The underground lithology identification method based on machine learning according to claim 6, characterized in that: After obtaining the first underground lithology identification model, the method further includes: The performance of the first underground lithology identification model is evaluated by using performance evaluation indicators; the performance evaluation indicators include accuracy, recall rate, precision, ROC curve and area under AUC curve.

8. An underground lithology identification system based on machine learning, characterized in that: include: A grid division module, used for performing grid division on the study area to obtain a plurality of underground grid units in the study area; The physical property acquisition module is used to acquire the physical properties of underground rocks and ores in each of the underground grid units, specifically including: Acquire underground drilling information of drilling locations in the study area; the underground drilling information includes real physical properties of underground rocks and minerals in underground grid cells passed by the drilling; Based on the underground drilling information, interpolation processing is performed on underground grid cells at undrilled locations in the study area to obtain preliminary physical properties of underground rocks and minerals in each of the underground grid cells; For each of the underground grid cells, based on the surface data of the study area, taking the preliminary physical properties of the underground rocks and minerals in each of the underground grid cells as initial values, taking the underground drilling information as constraints, the final physical properties of the underground rocks and minerals in each of the underground grid cells are calculated through constraint inversion; the final physical properties are the physical properties of the underground rocks and minerals in each of the underground grid cells; the surface data include gravity, magnetic and electrical data observed above the surface; the physical properties include density, magnetic susceptibility and resistivity; The spatial physical property feature extraction module is used to extract the feature of each physical property of the underground rock ores in each underground grid unit according to the spatial position of the underground grid unit corresponding to the physical property, so as to obtain the spatial physical property feature of the physical property, specifically including: Considering the spatial distribution of the physical properties of underground rocks and minerals, the spatial position of the underground grid unit is taken as a three-dimensional coordinate system (x, y, z), and the physical properties of the underground rocks and minerals in the underground grid unit are taken as function values ​​W(x, y, z); W(x, y, z) includes W_den(x, y, z), W_sus(x, y, z) and W_res(x, y, z); W_den(x, y, z) is density, W_sus(x, y, z) is magnetic susceptibility, and W_res(x, y, z) is resistivity; Perform feature extraction on the function value W(x, y, z) of the physical property of the underground rock ores in the underground grid unit to obtain local and global features of the three-dimensional coordinate system (x, y, z) of the spatial position of the underground grid unit, wherein the local and global features are spatial physical property features of the physical property; the local and global features include the function value of the physical property, the minimum value of the function value, the maximum value of the function value, the mean value of the function value, the median value of the function value, the variance of the function value and the gradient of the function value; The underground lithology prediction module is used to input the spatial physical property characteristics of all physical properties of underground rocks and ores in each underground grid unit into a trained underground lithology recognition model to obtain the underground lithology of the underground grid unit.

Images