CN110060173B

CN110060173B - Deep gold deposit forming and prospecting method

Info

Publication number: CN110060173B
Application number: CN201910348879.1A
Authority: CN
Inventors: 李滨; 王永超; 苏维范
Original assignee: YANTAI MOUJIN MINING CO Ltd
Current assignee: YANTAI MOUJIN MINING CO Ltd
Priority date: 2019-04-27
Filing date: 2019-04-27
Publication date: 2021-05-18
Anticipated expiration: 2039-04-27
Also published as: CN113488117A; CN110060173A; CN113488117B

Abstract

The invention discloses a deep gold deposit mining and prospecting method, which is characterized in that a hierarchical knowledge system is introduced, analysis results including indication types, index types, comprehensive characteristic types and the like are rapidly and comprehensively obtained according to actual sounding analysis data, and the utilization rate of previous work and research results is enhanced; meanwhile, when the detection data are analyzed and researched based on the hierarchical knowledge system module, a deep learning method is introduced, the requirement on the relevance analysis of the detection data is reduced, and the efficiency of the detection analysis and the positioning of the target area is greatly improved.

Description

Deep gold deposit forming and prospecting method

Technical Field

The invention belongs to the technical field of mineral resource exploration, and particularly relates to a deep gold deposit mining and prospecting method.

Background

Along with the continuous development of the earth surface, mineral resources of shallow and surface underground parts are greatly consumed, most shallow ores which are easy to discover are discovered, deep ore exploration is more and more difficult, the structure superposition blind ore exploration method and the research method of Lihui et al mainly adopt a geochemical method to explore deep ores and blind ores, other deep metal ore exploration methods and crisis mine deep ore exploration methods mainly adopt an ore exploration idea and various geophysical combined methods, and the deep ore exploration is still in the exploration and development stages at present and does not have a very mature method technology.

Meanwhile, in the prior art, the ore searching and prospecting excessively depends on the professional and experience of personnel, and the utilization degree of the work and research results of the predecessors is seriously insufficient, so that the ore searching efficiency and the ore finding rate are influenced;

especially, when a fracture zone vertical to the trend of the main ore body appears in the main ore body, due to the staggered fracture and expansion structure, the direction of ore finding is easy to lose and the target of ore finding is lost according to the conventional exploration ore finding, so that the problems of engineering waste, low underground delayed detection and analysis and low efficiency of positioning the target area of the ore are caused.

Disclosure of Invention

The invention provides an ore-finding method for deep gold deposit, which solves the problems that in the prior art, ore-finding and ore-finding excessively depend on the professional and experience of personnel, the utilization degree of previous work and research results is seriously insufficient, the ore-finding direction is easy to lose when a staggered breaking and unfolding structure is subjected to ore-finding, an ore-finding target is lost, engineering waste and underground delay are caused, and the efficiency of detection analysis and ore target area positioning is low.

The specific technical scheme is as follows:

an ore-forming and prospecting method for deep gold deposit comprises the following steps:

s100: establishing a gold deposit classification knowledge system base according to previous work and research results;

step 101: the method comprises the steps of summarizing geochemical background of mineral deposit mineralization, mineral deposit conditions, mineral deposit characteristics, mineral deposit formation rules, mineral control factors, mineral occurrence rules, structural properties, mineral body distribution rules in structure, ore-free intervals and laterals, mineral deposit mineral combination, alteration characteristics, axial zoning and predecessor work and research result materials including a mineral deposit mineralization stage;

step 102: classifying and grading the materials, constructing a gold deposit grading knowledge system base, and filling the materials of a mineral dressing and prospecting basic principle and previous work and research results into the knowledge system base to serve as a machine learning base; and, the record elements in the knowledge system base are divided into the following categories: (1) the indication element (2) the index element (3) the dynamic element (4) the comprehensive characteristic element (5) the instance element (6) and the mining element (7) are self-defined elements;

s200: dynamically quantizing all types of record elements in the knowledge system base established in the step S100;

s300: training a machine learning or deep learning network model to generate an inference library on the basis of a quantitative knowledge system library;

s400: on site, aiming at mining areas with different characteristics, corresponding technical methods are adopted to obtain the physical and chemical characteristics of the target mining area;

s500: and (5) field drilling verification of the deep gold mine.

Preferably, the step S200 is implemented as follows:

step 201: performing quantization processing mainly based on binarization on the record elements; step 202: and performing adaptive adjustment and optimization on the adjustment coefficient of the index element in the step 201.

Preferably, the step 202 is implemented as follows:

step 2021: initializing an index element adjusting coefficient;

step 2022: the quantitative results of the index element of the initialization state and other various record elements are used as input elements and input into a machine learning model for network training and learning, and a test error is output, and if the test error is delta, an objective function of the genetic algorithm can be established as delta-f (k)₁，k₂...k_m) Adjusting the optimization problem of the coefficient, and converting into solving the objective function delta f (k)₁，k₂...k_m) The fitness function F of the genetic algorithm can be determined, and the genetic algorithm for solving the minimum optimization is selected, so that the fitness function F is changed into an objective function delta;

step 2023: after the group optimization in step 2022, the index element adjustment coefficients obtained by the current optimization are transmitted to the dynamic quantitative knowledge system library in step 201 to form a new index element quantitative result;

step 2024: inputting the optimized quantized data into the machine learning or deep learning network established in the step S300 again to obtain a new test error, and performing the step 2022 and the step 2023 in a recycling manner until a stop condition of the group optimization algorithm is satisfied to obtain an optimal machine learning or deep learning network model.

Preferably, the step S300 is implemented as follows:

step 301: dividing 70% of data in a knowledge system base into a training set, and dividing 30% of data in the knowledge system base into a test set;

step 302: carrying out deep fusion on the convolutional neural network to construct a convolutional neural network model, wherein an input layer consists of 6 inputs and is used for each quantization element X ═ X in each quantization element library in a knowledge system library₁...X₆]The hidden layer comprises 2 convolutional layers, 2 pooling layers and 1 full-connection layer, and finally is an output layer;

step 303: extracting the characteristics of the network input, fusing the characteristics, and training the network;

step 304: after the network training is finished, inputting a test set into the network to obtain a test error delta; and feeding back the test error delta to the step 202, and optimizing the index element adjustment coefficient to obtain an optimal machine learning or deep learning model.

Preferably, the step S400 is implemented as follows:

step 401A: for the deep prospecting of old mining areas, firstly carrying out the geoacoustic electromagnetic depth measurement or EH4 system measurement of a controllable source, determining the extension condition of a mine control structure, and inputting the graphs measured by the electromagnetic depth measurement or EH4 system into an inference library;

step 401B: for a new mining area with serious coverage and poor earth surface outcrop, firstly carrying out live ladder measurement, and inputting a measurement result into an inference library;

step 402: and determining whether blind ore bodies exist in the predicted target area by adopting a method of constructing superposition halo measurement, and further delineating the target area of the ore.

Has the advantages that:

the invention provides a deep gold mine mineralization finding method, which is characterized in that a hierarchical knowledge system is introduced, analysis results including indication types, index types, comprehensive characteristic types and the like are rapidly and comprehensively obtained according to actual sounding analysis data, and the utilization rate of previous work and research results is enhanced; meanwhile, when the detection data are analyzed and researched based on the hierarchical knowledge system module, the deep learning method is introduced, the requirement on the correlation analysis of the detection data is reduced, and the efficiency of the detection analysis and the positioning of the target area is greatly improved.

Description of the drawings:

FIG. 1: the invention relates to a flow chart of a deep gold mine mineralization prospecting method;

FIG. 2: a flow chart of adaptive adjustment and optimization of index element adjustment coefficients;

FIG. 3: calculating a model diagram by a deep fusion convolutional neural network;

FIG. 4: the invention discloses a schematic block diagram of a deep gold mine mineralization finding method;

FIG. 5-1: 230 linear induced polarization sounding measurement visual polarizability eta a isoline cross-section drawing;

FIG. 5-2: 230, simulating a section diagram by using a contour line of apparent resistivity rho a for measuring the line-induced electrical sounding;

FIG. 6-1: 230-2, simulating a sectional view by using a contour line of the apparent polarizability eta a for measuring the linear induced polarization sounding;

FIG. 6-2: 230-2, simulating a cross-sectional view by using a contour line of apparent resistivity rho a for measuring the line induced electrical sounding;

FIG. 7-1: 260 linear induced electrical sounding measurement visual polarizability eta a isoline cross-section drawing;

FIG. 7-2: 260, simulating a section diagram by using a contour line of apparent resistivity rho a for measuring the line-induced electrical sounding;

FIG. 8: the deep mineralization favorable part and the prediction target position of the Deng Guzhuang I1-1 ore body;

FIG. 9: the deep mineralization favorable part and the prediction target position of the Deng Gezhuang I2-2 ore body;

FIG. 10: the deep mineralization favorable part and the prediction target position of the Deng Zhuang II-1 ore body;

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In the embodiment of the invention, a deep gold mine stepped prospecting method is provided, a DengGezhu is selected as a mineralizing area, the mineralizing area is positioned in the east half part of a gold mineralizing area in Jiaodong, the structure position of the earth belongs to the southeast part of the east China towards the quasi-terrace (level I) Jiaodong Tailong (level II) Jiaodong Beilong (level III), and the stratum in the area is only sporadically exposed from ancient and ancient Jingshan group metamorphic rocks; the fourth series sediments of the new growing field are distributed along the valley and the two sides of the river. The ancient and ancient wattle groups are distributed sporadically in a bag shape in the region. The lithology mainly comprises the following components of nephrite, inclined long diopside, transparent marble, inclined long-angle amphibole and the like; the main ore body is controlled by a north-south fracture zone, has an EW (engineering) direction fracture and expansion structure, performs dislocation and translation on the main ore body, finds ores according to conventional exploration, and easily loses the ore finding direction and loses the ore finding target when the underground excavation meets the EW direction complex geological fracture structure, thereby causing the problems of engineering waste, underground delay and the like.

The method comprises the following specific steps:

referring to fig. 1, the method for mining and prospecting deep gold ore according to the present invention includes:

s100, establishing a gold deposit classification knowledge system base according to previous work and research results.

Step 101: the method is characterized by summarizing former work and research result materials comprising a mineral deposit mineralizing geochemical background, mineralizing conditions, mineral deposit characteristics, a mineral deposit forming rule, mineral control factors, mineral occurrence rules, structural properties, mineral body distribution rules in a structure, ore-free intervals and lateralization rules, mineral deposit mineral combination, alteration characteristics, axial zoning, a mineral forming period-a mineral forming period and the like.

Step 102: classifying and grading the materials, constructing a gold deposit grading knowledge system base, and filling the basic principles of ore dressing and prospecting and the materials of previous work and research results into the knowledge system base to serve as a machine learning base.

It should be noted that the classification method of the knowledge system base is not constant, and any classification method that can provide a learning base for a subsequent machine learning model is within the scope of the present invention. In this embodiment, the classification method of the knowledge system base is as follows:

record elements in the knowledge system base are divided into the following categories:

(1) and indicating the element. The indication element is capable of directly judging to obtain a certain result according to the content of the record element. For example, in the deep blind mine prospecting process of the old mine area, the indicating elements of leading edge halo, near mine halo and tail halo of the current ore body can be found according to the first-stage mine forming axial zone data of the old ore deposit, and the element M is assumed to be₁Is one of the indicating elements of the leading edge halo of the current ore body, and the M appears in the tail halo of the current ore body₁If the middle and inner zones of the element are abnormal, blind ore bodies are found in the deep part of the current ore body. Such a record element is divided into indication elements.

(2) And (5) index elements. The index element means that a certain result can be judged when a certain quantitative index is reached. For example, in the process of positioning the deep blind ore target position of the old gold mining area, when leading edge halo and near ore halo indicating elements of the current ore body appear to be mostly abnormal with weak outer zone in the forming axial zonal, if the content of the gold element is less than the index value alpha (percentage index), the blind ore body head is deeper, and is more than 200 meters. Such record elements are divided into index elements.

(3) And (4) dynamic elements. The dynamic element means that some index data cannot reflect the problem to be researched, and the result is judged by the dynamic change of the index data. For example, the geochemical parameters of a known gold ore body are dynamically changed in an axial direction to indicate the stacking

If gamma appears high → low → high dynamic change along with the geochemical parameters of the known gold ore body, the old ore deep part is indicated to have blind ore body. Such record elements are divided into dynamic elements.

(4) And (6) integrating the feature elements. The comprehensive characteristic element refers to common characteristics which are obtained by summarizing the same type of characteristics with similar geographic positions or ore control structures without specific index values. For example, according to the historical data of mineral exploration of Chinese gold mines, the axial zonal sequence characteristics of the connate geochemistry parameters of the same type of gold deposits are obtained through statistics and serve as common characteristics, and when the axial zonal sequence of the connate geochemistry parameters of the target gold deposits is compared with the common zonal sequence characteristics, abnormal and reverse zonal phenomena occur, such as leading edge halo indicating elements appear at the lower part of the axial zonal sequence of the connate, and blind ore bodies are shown in the deep part. Such record elements are divided into comprehensive feature elements.

(5) Instance elements. The example element refers to the summarized real deep gold mine prospecting examples, and the method and the representation adopted in the example process. For example, in 1965, a delbrudge deposit was discovered, the trace element content from the fracture front of the known blind deposit was analyzed, surface and core samples were taken near the known blind deposit and analyzed, and the abnormal amount of mercury was 150300ppb (ppb is the concentration unit, representing 10 ppb)^-9) Detailed petro-geochemical investigations of the blind deposits of the zone are carried out. Such record elements are divided into instance elements.

(6) And (5) excavating the element. The solution mining element is a recording element which is formed in the process of finding the mine and can judge a result only by experience. For example, in the operation of detecting the extension condition of the ore control structure, the fluctuation of the obtained electromagnetic characteristic curve judges the ore control extension condition. Such record elements are divided into mining elements.

(7) And (5) self-defining the element. In order to realize that each ore body fully reflects the personalized characteristics of the ore body, the knowledge system base is provided with a user-defined element, and the element is actually input and maintained by combining the ore-finding of the professional, so that the learning base has the adaptive evolution capability.

And S200, dynamically quantizing all types of record elements in the knowledge system base established in the step S100.

In order to provide the knowledge system base established in step S100 with the learning base required for machine learning, the record elements in the knowledge system base need to be quantized. However, the method aims at the index quantification engineering in the field of finding the ores, is widely researched from late mineral years, and has no uniform result so far. In order to avoid the problem of unreliable quantization mode, the invention introduces a dynamic quantization method of record elements of a knowledge system base, the method can comprehensively plan the diversity of the record elements in the knowledge system base and the actual ore-forming and ore-finding effect, dynamically adjust the record elements into adaptive values so as to provide a quantized learning base for a machine learning model, and simultaneously, in order to improve the learning efficiency, the invention adopts binarization quantization to the record elements, and the method comprises the following steps:

step 201: the recording element is subjected to quantization processing mainly based on binarization, and the quantization processing comprises the following steps:

(1) quantization of the indicator elements. Let X₁Representing the quantization result of the indicator. According to the nature of the indicator, this is a type of switching value element, and its quantization can be directly based on whether the indication effect is achieved, the value of which is 1, or else 0

In the deep blind mine prospecting process of the old mine area, the indicating elements of leading edge halo, near mine halo and tail halo of the current ore body can be found according to the first-stage mine forming axial zone data of the old mine deposit, and the element M is assumed to be₁Is one of the indicating elements of the leading edge halo of the current ore body, and the M appears in the tail halo of the current ore body₁If the middle and inner zones of the element are abnormal, blind ore bodies are found in the deep part of the current ore body. Then the example indicates the element

Quantization result of final indicator

n represents an element dimension of the indicator element,

the value of each dimension index is represented, the value is 0 or 1, i belongs to [1, n ]]。

(2) And (5) quantizing the index element. Let X₂Indicating the quantization result of the index element. The index elements may be varied according to the rock stratum properties, ore control structure and geographical position of the target mining areaAnd the universal applicability is difficult to achieve. If dynamic quantization is adopted, the problem can be well solved. Therefore, in this embodiment, the dynamic quantization is implemented by adjusting the coefficients of the elements of the index elements, and the method includes:

index element recording:

in the process of positioning deep blind ore target positions in old gold mining areas, when leading edge halo and near ore halo indicating elements of the current ore body appear to be mostly in weak and abnormal outer zone in the mining axial zonal mode, if the content of the gold elements is smaller than an index value alpha (percentage index), the blind ore body heads are deep and are more than 200 meters.

After dynamic quantization, namely the content of the gold element is less than the index value k alpha, k belongs to (0, infinity) as an adjustment coefficient, when the measured value meets the condition that the content of the gold element is less than the index value k alpha, the index value is 1, otherwise the index value is 0.

In practical application, there are multiple index element records, and assuming that the dimension of the index element record is m, the index element will have m adjustment coefficients (k)₁，k₂...k_m) Further, the number of the dynamic index values is m, (k)₁α₁，k₂α₂...k_mα_m)。

Then

The value of each dimension index is represented, the value is 0 or 1, i belongs to [1, m ]]。

(3) And (5) quantizing the dynamic elements. Let X₃Representing the quantization result of the dynamic element. And (4) quantizing the dynamic elements by judging whether the measured values meet the characteristic trend of the dynamic elements. For example, the geochemical parameters of a known gold ore body are dynamically changed in an axial direction to indicate the stacking

If gamma appears high → low → high dynamic change along with the geochemical parameters of the known gold ore body, the old ore has blind ore body deep. If trueThe geochemical parameter of the gold-measuring ore body has the characteristic trend of high → low → high in the axial direction, and then the index value is 1, otherwise the index value is 0. Finally, the product is processed

p represents the element dimension of the dynamic element,

the value of each dimension index is represented, the value is 0 or 1, i belongs to [1, p ]]。

(4) And (5) quantizing the comprehensive characteristic elements. Let X₄And representing the quantization result of the comprehensive characteristic element. And quantifying the value according to the characteristic whether the common characteristic element is conformed or not. If the measured value accords with the common characteristic, the index value is 1, otherwise, the index value is 0, and finally

q represents the element dimension of the dynamic element,

the value of each dimension index is represented, the value is 0 or 1, i belongs to [1, q ]]。

(5) And (5) quantization of example elements. Let X₅Representing the quantization result of the example element according to the coincidence rate of the measured element and the example element

To be quantized. Example meta-records in this embodiment: in 1965, a DelibulOre deposit was discovered, the trace element content from the fracture front of the known blind deposit was analyzed, surface and core samples were collected near the known blind deposit and analyzed, and the abnormal amount of mercury was 150-300ppb (ppb is the concentration unit, representing 10 ppb)^-9) Detailed petro-geochemical investigations of the blind deposits of the zone are carried out. The example elements of the record include time, place, method 1, method 2, method 3, and method 4, and if there are 4 items that the measured element matches, then there are 6 elements in total

Finally, the product is processed

If a plurality of records with matching elements can be found in an example element by a certain measured element, the record with the largest number of matching items between the measured element and the example element is taken as the standard.

(6) And (5) quantizing the mining elements. The quantification of the mining element cannot be calculated by an intuitive formula because the nature of the mining element determines that the mining element belongs to a regular characteristic, and the acquisition of the mining element usually comes from an electromagnetic echo curve, a structural superposed halo profile, different section plane diagrams, a vertical longitudinal projection diagram and the like. In the embodiment, the remarkable characteristics of the mining elements are directly used as image elements, for example, a magnetic field, an electric field, apparent resistivity and a phase curve generated by controllable source geodetic audio electromagnetic sounding are adopted. Let the quantization result of the mining element be

s represents the element dimension of the mining element,

set of image objects representing each dimension, i ∈ [1, q ]]。

(7) And (5) processing the custom element. The self-defined element is maintained by an explorationist autonomously, and the definition of the element does not have a uniform standard, so that the processing of the self-defined element adopts an expert experience method which does not directly participate in machine learning and can be judged by the explorationist by experience.

Step 202: and in step 201, performing adaptive adjustment and optimization on the adjustment coefficient of the index element.

In practical application, the target mine areas with the same geophysical and chemical structure have obvious index values due to different extending conditions or layout conditions of fracture zones of the ore control structure, and directly influence blind ore body prediction and even ore finding results. This is also the biggest problem faced by the current index quantification research engineering.

The group optimization method is introduced, manual research is not relied on, batch optimization is carried out on the adjustment coefficients in use, the optimized index element quantification result is used as an input element for machine learning, the accuracy of a machine learning model is continuously enhanced, and the general applicability of the method is greatly improved. Referring to fig. 2, the present embodiment adopts a genetic algorithm to automatically optimize the adjustment coefficient of the index element by counting the test error of the machine learning model, and includes the steps of:

step 2021: and initializing index element adjusting coefficients.

In the initialization state, the adjustment coefficient (k) of the m-dimensional index element is set₁，k₂...k_m) When the unified initialization is 1, the corresponding m-dimensional dynamic index value is (alpha)₁，α₂...α_m) And finally, the binary quantization result of the index element

Step 2022: the quantitative results of the index element of the initialization state and other various record elements are used as input elements and input into a machine learning model for network training and learning, and a test error is output, and if the test error is delta, an objective function of the genetic algorithm can be established as delta-f (k)₁，k₂...k_m) Adjusting the optimization problem of the coefficient, and converting into solving the objective function delta f (k)₁，k₂...k_m) The fitness function F of the genetic algorithm can be determined, and the genetic algorithm for solving the minimum optimization is selected, so that the fitness function F is the objective function delta. The technicians in the field can use the genetic algorithm for solving the minimum optimization according to the prior art, such as the genetic algorithm tool gaot of the university of north carolina or the genetic algorithm tool gadst embedded in matlab tool software, to complete the multivariate optimization of the index element adjustment coefficient and realize the group optimization effect.

Step 2023: after the group optimization in step 2022, the index element adjustment coefficients obtained by the current optimization are transmitted to the dynamic quantization knowledge system library in step 201, so as to form a new index element quantization result.

The m-dimensional adjustment coefficient of the optimized index element is (k)₁，k₂...k_m) Further, the dynamic index value is (k)₁α₁，k₂α₂...k_mα_m). Then

Step 2024: inputting the optimized quantization data into the machine learning or deep learning network established in step S300 again to obtain a new test error, and performing step 2022 and step 2023 in a recycling manner until a stop condition of the group optimization algorithm is satisfied. And obtaining the optimal machine learning or deep learning network model.

S300, training a machine learning or deep learning network model to generate an inference base based on the quantitative knowledge system base.

After each element of the knowledge system base established by the invention is quantized, the element data comprises an array and an image type, and in order to improve the processing efficiency, the embodiment of the invention adopts a deep fusion convolution neural network. The method comprises the following steps:

step 301: and dividing 70% of data in the knowledge system base into a training set and 30% into a test set.

Step 302: the convolutional neural network is deeply fused to construct a convolutional neural network model as shown in fig. 3, an input layer is composed of 6 inputs and is used for each quantization element X ═ X in a knowledge system base₁...X₆]The hidden layer comprises 2 convolutional layers, 2 pooling layers and 1 full-link layer, and finally the output layer.

Step 303: and extracting the characteristics of the network input, fusing the characteristics, and training the network.

The input features are extracted, and for the features extracted from the training set, the embodiment is not particularly limited, and those skilled in the art may refer to the prior art. Then the input features are fused:

wherein x is₀Is the input characteristic, K is the number of the basic convolutional neural network, Xⁱ(x₀) Is the input characteristic of each underlying convolutional network.

Partitioning each underlying network into B blocks

There are B summation fusion processes:

wherein, i is 1, 2., B,

output of sufficient i fusions. The convolution neural network adopts a local filter to carry out convolution process, namely, the local submatrix of the input item and the local filter carry out inner product operation:

where 1 is the 1 st convolutional layer, i is the value of the ith convolutional output matrix, j is the number of the output matrix, the size of the output matrix is N-m +1, N is the number of convolutional output matrices, f is the activation function, ω is the weight value, b is the offset, and a is the number of channels of the feature map.

The output of the convolutional layer serves as an input to the pooling layer, which can further reduce the dimension of the matrix without destroying the intrinsic connection of the data. Dimensionality reduction by mean:

wherein the content of the first and second substances,

is the output item of the pooling layer, and is obtained by averaging the matrix with the size of n multiplied by n of the previous layer. The output item of the pooling layer is used as the input item of the full-connection layer, the full-connection layer is equivalent to a hidden layer in a feedforward neural network, the characteristic diagram loses a three-dimensional structure on the full-connection layer, the characteristic diagram is expanded into a vector and is transferred to the output layer through an excitation function, and the excitation function is as follows:

wherein, I_cIs an entry of the fully connected layer. The fused gradient block in the depth fused gradient backpropagation is:

wherein the content of the first and second substances,

is the gradient of the b +1 th block in the ith convolutional neural network. Establishing an error function e (y ', y), y' and y being the desired output and the actual output, the gradient is calculated as:

and (4) performing network calculation in a mode of descending the gradient fastest, and setting the maximum iteration times until network training is stopped to obtain an inference library.

Step 304: and after the network training is finished, inputting a test set into the network to obtain a test error delta. The test error Δ is fed back to step 202 to optimize the index adjustment coefficient. And further obtaining an optimal machine learning or deep learning model.

Thus, the actual physical and chemical measurement parameters are input into the trained inference base calculation model, whether the target mining area has favorable positions for mineralization or not can be rapidly classified, and the predicted target area is marked according to the data of the favorable positions for mineralization.

S400, aiming at mining areas with different characteristics, the scene adopts a corresponding technical method to obtain the physical and chemical characteristics of a target mining area, and the method comprises the following steps:

step 401A: for the deep prospecting of old mining areas, firstly, the geoacoustic electromagnetic depth measurement or EH4 system measurement of a controllable source is carried out, the extension condition of a mine control structure is determined, and the graph measured by the electromagnetic depth measurement or EH4 system is input into an inference library.

Step 401B: for a new mining area with serious coverage and poor surface outcrop, firstly, carrying out live ladder measurement, and inputting a measurement result into an inference library.

The measurement results for dunge are described below: the system quality inspection 2 sections are carried out in the whole area, the total inspection points are 72, the total workload accounts for 6.13%, and the precision is as follows: the visual polarization rate mean square error is: ε η s is 0.055; the resistivity mean square relative error M rho a is +/-2.277%, and the quality inspection precision meets the requirements of design and specification.

According to the principle of delineation of abnormal excitation, 4 abnormal bands are delineated in the work and are numbered as DJH-1, DJH-2, DJH-3 and DJH-4. Now specifically described as follows:

1. DJH-1 abnormality

The DJH-1 anomaly is positioned in the middle of a work area, 2.0% is taken as the lower anomaly limit coil contour line and is in a strip shape, the length is about 600m, the width is about 120m, the trend is nearly north and south, the anomaly center point is positioned near the 112 point of the 230 line, and the maximum value is 2.25%.

2. DJH-2 abnormality

The DJH-2 anomaly is located eastern part of DJH-1, the anomaly is parallel to DJH-1 anomaly, 2.0% is taken as anomaly lower limit coil contour line form and is in strip shape, the length is about 300m, the width is about 100m, the strip shape is from north to south, the anomaly center point is located near 139 point of 230 line, and the maximum value is 2.2%.

3. DJH-3 abnormality

The DJH-3 anomaly is positioned in the middle of a western region, 3.20% of the anomaly lower limit delineating isoline is in a strip shape, the length is about 900m, the width is about 100m, the trend is near north and south, the anomaly center point is positioned near a point 247 of a line 260, and the maximum value is 2.60%.

4. DJH-4 abnormality

DJH-4 abnormity is positioned at the south of a work area, 2.0% is used as the lower abnormity limit coil contour line form to be a strip shape, the length is about 500m, the width is about 100m, the strip shape extends to the south and the north, the abnormity center point is positioned near the 100 line 145 point, and the maximum value is 2.20%.

Explanation for induced electrical sounding

1. DJH-1 abnormality

In order to study the abnormity, excitation sounding work is carried out at points 104-120 of a 230 line, and a 230 line comprehensive section diagram shows that the curve of the corresponding eta abnormity center rho a of the excitation elevator measurement near the point 112/230 reflects low-resistance characteristics. The induced electrical sounding reflects a high polarizer inclined to the east at 112/230 points, and from 110-116 points, the rho a section diagram is reflected in a V-shaped low-resistance manner, a fracture structure is inferred to exist at the rho a section diagram, an east inclination abnormal zone exists between 106-116 points, the eta a section diagram, the abnormal center is located at 112 points, and the estimated depth is 50 m. The proposed verification location 112 has coordinates (X-4118300Y-373120) and a depth of about 50 m. (see FIG. 5-1, FIG. 5-2)

2. DJH-2 abnormality

In order to research the abnormity, excitation sounding work is carried out at 135-147 points of a 230 line, and a 230 line comprehensive section diagram shows that the corresponding eta abnormity center rho curve of the excitation elevator measurement near 141/230 points reflects low-resistance characteristics. The induced electrical sounding reflects a high polarizer inclined to the east at the point 112/230, the rho a section diagram is reflected in a V shape with low resistance from the point 141-145, the fracture structure is inferred to exist at the rho a section diagram, the east inclination abnormal zone exists between the point 139-143, the eta a section diagram, the abnormal center is located at the point 141, and the estimated depth is 30 m. It is proposed to verify the position 141 point with coordinates (X-4118300Y-373410) and a depth of about 30 m. (see FIG. 6-1, FIG. 6-2)

3. DJH-3 abnormality

In order to study the abnormity, excitation sounding work is carried out at the No. 234-254 point of the line 260, and the comprehensive section diagram of the line 260 shows that the excitation sounding reflects a high polarizer inclined to the east, and a V-shaped low stop band is reflected on the section diagram corresponding to the high polarizer rho a. Eta a section 245 shows a vertical anomalous zone, and AB/2 at the center of 245 is 220 m. It is proposed to verify the location 245 point with coordinates (X-4118600Y-374450) at a depth of about 200 m. (see FIG. 7-1, FIG. 7-2)

4. DJH-4 abnormality

The abnormality was located in the south of the work area, and it was estimated that the abnormality was in the same structure as DJH-1, so no research work was done on it.

The work completes the measurement of 2.10Km of the power-exciting escalator²Total 1126 physical points, 3 pieces of excitation sounding section and 18 measurement points. Through early-stage ground geology and geophysical prospecting work, four excitation anomalies are defined in a work area, and are numbered as DJH-1, DJH-2, DJH-3 and DJH-4. It is inferred to be governed by the gold bullseye fracture zone or secondary configurations parallel thereto, caused by metal mineralizers, and therefore has favorable conditions for mineralization. Due to the limitation of workload, only the excitation sounding research work is carried out on DJH-1, DJH-2 and DJH-3, which shows low-resistance and high-polarization characteristics, and the conclusion that the drilling work reveals abnormality at the inferred abnormality is provided, so that favorable conditions are provided for the next work.

The step of constructing a superimposed halo measurement includes:

(1) and (6) sampling. The sampling method for the superimposed halo of the quartz vein type gold ore structure comprises the following steps: collecting the smoke gray quartz and sulfide vein or network vein at the 2 nd and 3 rd stages without collecting the large quartz vein at the 1 st stage; sampling geological analysis auxiliary samples, and selecting the auxiliary samples with relatively high mineralization; sampling along the vein or along the structure of the earth surface, wherein the point distance is 5-10 m, the width of the structure belt is more than 5m, the vertical alteration belt is 2-5m, the line distance is 10-30 m, and the sampling weight is 400-500 g.

The embodiment collects and constructs 2595 pieces of superposed halo samples, wherein the superposed halo samples comprise 395 pieces of gallery samples, 204 pieces of surface samples, 1350 pieces of 163 drilling samples, 610 pieces of soil geochemistry and mercury adsorption research samples, 30 pieces of background samples and 6 pieces of single ore. The following table is a structural superposition halo study sampling statistical table of the gold deposit of mupingeng gangzhuang.

(2) After inspection, drawing different cross-sectional views, different middle-section plane contrast views and vertical longitudinal projection views of the structural superposed halo;

(3) calculating the geochemical parameters of the sample: determining the ore deposit element combination by including the geometric mean value, the lining value and the correlation coefficient among elements of the ore deposit; the average value and the contrast value of element sets in different mineralization stages; different elevation element content ratios.

(4) Identifying characteristic element sets of leading edge corona, near-ore corona and tail corona of primary corona formed by single-stage secondary ore formation, and identifying an ore body-corona superposed structure formed by different stages.

(5) Inputting the inspection and calculation results and the physical parameters measured in the step 401A or the step 401B into the inference library obtained in the step S300, and obtaining the favorable positions of the mineral formation and predicting the target area of the mineral through the deep fusion of the inference library and the calculation of the convolutional neural network.

(6) The length and the width of the target position of the predicted blind mine can be determined by a person skilled in the art according to the length and the width of the structure superposition halo by utilizing the prior art; and (4) calculating blind mine depth according to the concentrations of leading edge halo indicating elements (arsenic, antimony, mercury and boron) and ore forming elements gold in the superimposed halo, the upper known ore body form, the lateral-lying rule and the non-ore interval qualitative mode, and defining the deep blind mine prediction target position.

Comprehensively analyzing all the predicted target areas obtained in the steps (5) and (6) to determine the final high mineral percentage predicted target area.

And S500, field drilling verification of the deep gold mine.

Step 501A: the drilling verification can be directly carried out on the old mining area.

Step 501B: and (4) carrying out groove exploration exposure and then drilling verification on the new mining area with poor surface outcrop.

Through the steps, the prediction verification is carried out on the deep parts of the gold ore bodies I1-1, I2-2 and II-1 of the mu ping dung zhuang gold ore deposit, the deep parts of the main gold ore bodies or the mineralization zones of six ore sections such as the island mountain and the rear banker at the periphery, and the prediction target positions 23 (wherein, the deep part of the dung zhuang ore zone 10, the island mountain ore zone 3, the rear banker ore zone 1, the black cattle table ore zone 4, the north water channel ore zone 3, the No. XI mineralization zone 1 and the Xuge banker ore zone 1) are defined on the section, and the deep mineralization favorable positions and the prediction target positions of the ore bodies I1-1, I2-2 and II-1 are shown in figures 8-10.

On the basis of obtaining a large amount of first geochemical data and geochemical test data, the deep gold deposit mining and prospecting method disclosed by the invention is completed through researching the geochemical characteristics and the structural superposition halo overall characteristics of the Deng Zhuang gold deposit, the method reduces the dependence on expert experience, greatly improves the mining and prospecting efficiency of the deep gold deposit, has remarkable economic and practical benefits, quickly and comprehensively obtains analysis results including indication types, index types, comprehensive characteristic types and the like according to actual probing analysis data by introducing a hierarchical knowledge system, and enhances the utilization rate of the work research results of predecessors; meanwhile, when the detection data are analyzed and researched based on the hierarchical knowledge system module, a deep learning method is introduced, the requirement on the correlation analysis of the detection data is lowered, the accumulated ore quantity of the deep resource quantity of the Dungen is predicted to be 478.38 ten thousand tons, the gold metal quantity is predicted to be 23.92 tons, and the efficiency of the detection analysis and the ore target area positioning is greatly improved.

Claims

1. An ore-forming and prospecting method for deep gold deposit is characterized by comprising the following steps:

step 101: the method comprises the steps of summarizing geochemical background of mineral deposit mineralization, mineral deposit conditions, mineral deposit characteristics, mineral deposit formation rules, mineral control factors, mineral occurrence rules, structural properties, mineral body distribution rules in structure, ore-free intervals and laterals, mineral deposit mineral combination, alteration characteristics, axial zoning and predecessor work and research result materials including a mineral deposit mineralization stage and a mineral deposit mineralization stage;

step 102: classifying and grading the materials, constructing a gold deposit grading knowledge system base, and filling mineral dressing and prospecting basic principles and predecessor work and research result materials into the gold deposit grading knowledge system base to serve as a machine learning base; and dividing record elements in the gold deposit classification knowledge system base into the following classes: (1) the indication element (2) the index element (3) the dynamic element (4) the comprehensive characteristic element (5) the instance element (6) and the mining element (7) are self-defined elements;

s200: dynamically quantizing all types of record elements in the gold deposit classification knowledge system base established in the step S100 to obtain a dynamic quantization knowledge system base;

step 201: performing quantization processing mainly based on binarization on the record elements;

step 202: performing adaptive adjustment and optimization on the adjustment coefficient of the index element in step 201;

s300: training a machine learning or deep learning network model to generate an inference library on the basis of a dynamic quantitative knowledge system library;

s500: on-site drilling verification of deep gold ores;

the specific implementation manner of step 202 is as follows:

step 2021: initializing an index element adjusting coefficient; in the initialization state, the adjustment coefficient (k) of the m-dimensional index element is set₁，k₂...k_m) When the unified initialization is 1, the corresponding m-dimensional dynamic index value is (alpha)₁，α₂...α_m) And finally, the binary quantization result of the index element

The value of each dimension index is represented, the value is 0 or 1, i belongs to [1, m ]]；

Step 2022: inputting the quantized results of the index element in the initialized state and other various record elements into a machine learning model as input elements, performing deep learning network training and learning, and outputting a test error, wherein if the test error is delta, an objective function of the genetic algorithm can be established as delta-f (k)₁，k₂…k_m) Adjusting the optimization problem of the coefficient, and converting into solving the objective function delta f (k)₁，k₂…k_m) The fitness function F of the genetic algorithm can be determined, and the genetic algorithm for solving the minimum optimization is selected, so that the fitness function F is changed into an objective function delta; by utilizing a genetic algorithm, the multivariate simultaneous optimization of m-dimensional element adjustment coefficients is completed, and the group optimization effect, namely 'group optimization';

step 2023: after the group optimization in step 2022, the index element adjustment coefficients obtained by the current optimization are transmitted to the dynamic quantization knowledge system library described in step 200 to form new index element quantization results, and optimized quantization elements are obtained;

step 2024: inputting the quantization element optimized in the step 2023 into the machine learning or deep learning network established in the step S300 again to obtain a new test error, and performing the step 2022 and the step 2023 in a circulating manner until a stop condition of the group optimization algorithm is satisfied to obtain an optimal machine learning or deep learning network model.

2. The deep gold deposit mineralization and prospecting method of claim 1, wherein the step S300 is implemented as follows:

step 302: carrying out deep fusion on the convolutional neural network to construct a convolutional neural network model, wherein an input layer consists of 6 inputs and is used for each quantization element X in each quantization element library in a knowledge system library₁...X₆]The hidden layer comprises 2 convolutional layers and 2 hidden layersThe system comprises a pooling layer, 1 full-connection layer and an output layer;

3. The deep gold deposit mineralization prospecting method of claim 1, wherein the step S400 is implemented as follows: