Method for constructing mineral prediction deep learning model based on hard constraint of metallogenic model

Abstract

The application provides a mineral prediction deep learning model construction method based on a hard constraint of a metallogenic model, comprising the following steps: determining key ore-controlling elements according to a metallogenic model; constructing a power law function of deposit distribution density and ore-controlling element distance; assigning values to different ore-controlling element layers based on the power law function; constructing a hidden layer of a deep learning model based on the ore-controlling element layers, establishing a connection between the ore-controlling element layers and linear functions in neurons of the hidden layer, and performing operations on outputs of the linear functions by using a nonlinear activation function, so as to construct a hidden layer based on a hard constraint of a metallogenic model and add the hidden layer to the deep learning model, so that a delineated prospecting area conforms to metallogenic regularity and geological cognition; and inputting multiple source prospecting information layers into the hard constraint deep learning model, minimizing a difference between a predicted value and an actual value by using a cross-entropy loss function, and thus delineating a prospecting area. The application can improve the explainability and prediction accuracy of a mineral prediction deep learning model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of earth science, and in particular to a mineral prediction deep learning model construction method based on hard constraints of a mineralization model. BACKGROUND

[0002] Due to weak and mixed information, partial missing of prospecting information, it is difficult to identify, separate and determine deep prospecting information, and the prospecting effect is poor. Big data and deep learning provide a new research paradigm and effective tool for geological prospecting data analysis, which can effectively mine complex and nonlinear geoscience spatial data and extract unknown patterns of geological processes, and has strong classification and prediction performance.

[0003] The convolutional neural network is a commonly used supervised deep learning model, and the input data is an image, which considers the spatial structure and relationship of adjacent prediction units, and can effectively capture the spatial coupling relationship of mineralization information, and has the best prediction effect in various mineral prediction models. The convolutional neural network extracts the spatial features of mineralization by setting the convolutional layer and the pooling layer, wherein the convolution operation learns the spatial coupling relationship between the ore deposit and the prospecting information by sliding a local training filter with the same weight and bias on the image; the pooling process further reduces the feature map and simplifies the network calculation complexity. Then the extracted spatial features are flattened into a one-dimensional vector and input into the fully connected layer to map the information to the classification task (whether it has mineralization potential), and finally the SoftMax layer outputs the mineralization probability. Various mineral prediction methods based on big data and deep learning, including convolutional neural networks, are data-driven methods, which to some extent ignore the geological knowledge and established facts in the field of mineral resource prediction, and they mainly focus on the spatial correlation between prospecting information and ore deposits, and less consider the genetic relationship between them, which leads to the phenomenon that the prediction results may be unexplainable and contrary to the field knowledge, and restricts the further application of deep learning models in quantitative prediction of mineral resources. Therefore, it is urgent to propose a method to build a bridge between the structure of the deep learning model and the geological cognition, so as to improve the credibility of the mineral prediction of the deep learning model. SUMMARY

[0004] In order to solve the above problems, the present application provides a mineral prediction deep learning model construction method based on hard constraints of a mineralization model, which is used to solve the problem of lack of geological field knowledge constraints in the mineral prediction process of the deep learning model, avoid the phenomenon that the prediction result does not conform to the geological cognition, and improve the effectiveness and explainability of the mineral prediction deep learning model. A mineral prediction deep learning model construction method based on hard constraints of a mineralization model mainly includes:

[0005] S1: determining key ore-controlling elements according to the mineralization model;

[0006] S2: constructing a power-law function of deposit distribution density and distance of the key ore-controlling factor;

[0007] S3: assigning values to different ore-controlling factors based on the power-law function to obtain the weight of different ore-controlling factors;

[0008] S4: multiplying the weight values of different ore-controlling factors and the output results of the original deep learning model and then summing them up, adding trainable weight parameters and bias term parameters, constructing a linear function in the hidden layer neurons, operating the output of the linear function by using a nonlinear activation function to map neuron features, increasing the nonlinearity of the hidden layer, and then constructing a hidden layer based on the ore-forming model, adding the hidden layer based on the ore-forming model to the output layer of the original deep learning model, realizing the hard constraint of the deep learning model, and then completing the construction of the mineral prediction deep learning model based on the hard constraint of the ore-forming model;

[0009] S5: inputting multi-source prospecting information into the mineral prediction deep learning model based on the hard constraint of the ore-forming model, minimizing the difference between the predicted value and the true value by using the cross-entropy loss function, training the mineral prediction deep learning model based on the hard constraint of the ore-forming model, and obtaining the trained mineral prediction deep learning model based on the hard constraint of the ore-forming model when the predetermined accuracy is reached.

[0010] Further, in step S1, the key ore-controlling factor is determined according to the ore-forming model, and the ore-controlling factor buffer zone is made by using GIS.

[0011] Further, in step S2, the number of deposits falling on each buffer zone in each ore-controlling factor layer is counted respectively, and the cumulative deposit number n and the deposit distribution density p are calculated, where p = n / d, d represents the width of the buffer zone, and a double logarithmic scatter plot is made with the buffer zone width d as the abscissa and the deposit spatial distribution density p as the ordinate. A straight line is fitted by using the least squares method to obtain the power-law function of the deposit distribution density and the distance from the ore-controlling factor p = cd a-2 , where c is a constant and a is a singularity index.

[0012] Further, in step S3, since different buffer zones in each ore-controlling factor layer have different control degrees on the formation of deposits, the buffer zones are assigned weights s, s = cd a-2 / p max , c is a constant, d represents the width of the buffer zone, a is a singularity index, and p max is the maximum value of the power-law function of the deposit distribution density and the distance from the ore-controlling factor. Thus, different buffer zones are assigned values, that is, the closer the region is to the ore-controlling factor, the greater the weight is; the farther the region is to the ore-controlling factor, the smaller the weight is.

[0013] Further, an equation based on the metallogenic model is constructed in step S4 as a linear function in the hidden layer neurons, where x represents the metallogenic probability of the original deep learning model output layer, k i is the weight of the i-th ore-controlling element layer corresponding to each input sample of the deep learning model, ω is a trainable weight parameter, b is a trainable bias term parameter, and the output of the linear equation is operated using a Sigmoid activation function to map the neuron features, that is where μ is a constant.

[0014] The technical solution provided by the present application has the beneficial effects that: the present application constructs a hidden layer based on the metallogenic model according to the spatial nonlinear correlation between the ore-controlling elements and the deposit location, which not only improves the effectiveness of mineral prediction, but also obtains results with physical consistency, improves the interpretability of the mineral prediction results, and can more accurately delineate the prospecting areas. BRIEF DESCRIPTION OF DRAWINGS

[0015] The present application will be further described below in conjunction with the drawings and examples, and the drawings are as follows:

[0016] Figure 1 is a flowchart of a mineral prediction deep learning model construction method based on a metallogenic model hard constraint in an embodiment of the present application.

[0017] Figure 2 is a schematic diagram of ore-controlling element buffer zone analysis taking faults as an example in an embodiment of the present application.

[0018] Figure 3 is a buffer zone distance-deposit distribution density double logarithmic scatter plot taking faults as an example in an embodiment of the present application.

[0019] Figure 4 is a weighted function diagram taking faults as an example in an embodiment of the present application.

[0020] Figure 5 is a schematic diagram of a mineral prediction deep learning model structure based on a metallogenic model hard constraint taking a convolutional neural network as an example in an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to have a clearer understanding of the technical features, objectives and effects of the present application, the specific embodiments of the present application will be described in detail with reference to the drawings.

[0022] The embodiment of the application provides a mineral prediction deep learning model construction method based on a hard constraint of a metallogenic model.

[0023] Please refer to Figure 1 , Figure 1 The embodiment of the application provides a mineral prediction deep learning model construction method based on a hard constraint of a metallogenic model.

[0024] S1: According to the metallogenic model, determine the key ore-controlling elements, for example, use the GIS software to perform buffer analysis to obtain the schematic diagram of the ore-controlling element buffer analysis as shown in Figure 2

[0025] S2: Respectively count the number of deposits in each ore-controlling element layer falling on each layer buffer, and calculate the cumulative deposit number n and the deposit distribution density p, p=n / d, wherein d represents the width of the buffer. Take the buffer width d as the abscissa and the deposit spatial distribution density p as the ordinate to draw a double logarithmic scatter diagram, fit a straight line by using the least square method, and obtain the power law function p=cd a-2 of the deposit distribution density and the ore-controlling element distance, wherein c is a constant, and a is a singularity index. For example, as shown in Figure 3 , the fitting straight line expression on the double logarithmic diagram is p=52.01x -0.42 ;

[0026] S3: Normalize the formula in step S2 to calculate the weight sigma of different buffers, sigma=cd a-2 / ρ max , p max is the maximum value of the power law function of the deposit distribution density and the buffer distance, so that the closer the region to the ore-controlling element, the greater the potential for prospecting, and the greater the weight; the farther the region to the ore-controlling element, the smaller the potential for mineralization, and the smaller the weight. For example, the weighted function of the fracture is as shown in Figure 4 , and the weight distribution formula is d is in units of kilometers;

[0027] S4: Construct an equation based on the metallogenic model ​as a linear function in the hidden layer neuron, where x represents the original deep learning model output layer of the probability of mineralization, k i is the weight of the i-th ore-controlling factor layer corresponding to each input sample of the deep learning model, which is obtained according to step S3, ω is a trainable weight parameter, b is a trainable bias term parameter, and the output of the linear equation is operated by using a Sigmoid activation function to map the neuron features, that is where μ is a constant, which is set to 1 in this embodiment. The linear function and the nonlinear activation function together constitute the hidden layer of the mineral prediction deep learning model based on the hard constraint of the ore-forming model, so that the mineral prediction deep learning model based on the hard constraint of the ore-forming model can learn the spatial coupling relationship between the key ore-controlling factors and the training samples in the forward propagation path during the training process, and can adjust the trainable weight parameter ω in the reverse direction to improve the prediction accuracy of the model by calculating the difference between the sample prediction value and the target value in the reverse propagation path. After multiple training, the continuous updating of ω adjusts the correlation between the samples and the key ore-controlling factors in the ore-forming model in the forward propagation of the model, thereby building a mineral prediction deep learning model based on the hard constraint of the ore-forming model and improving the interpretability and prediction accuracy of the mineral prediction deep learning model.

[0028] S5: Please refer to Figure 5 The multi-source prospecting information is input into the mineral prediction deep learning model based on the hard constraint of the ore-forming model taking the convolutional neural network as an example. The role of the convolutional neural network is to extract deep prospecting information from the multi-source prospecting data, capture the spatial coupling relationship of the mineralization information, and serve the delineation of the ore-forming prospective area. The multi-source prospecting data is extracted by the convolution kernel in the convolution layer, and then the feature dimension is reduced by the pooling layer, and finally the feature matrix retaining the key information is obtained. Then the output of the linear equation is operated by using a Sigmoid activation function to map the neuron features, and the difference between the predicted value and the true value is minimized by using a cross-entropy loss function, so as to train the model and delineate the prospecting area, and improve the interpretability, generalization ability and prediction accuracy of the mineral prediction deep learning model.

[0029] The beneficial effects of the present application are: the present application constructs a hidden layer based on the ore-forming model according to the spatial nonlinear correlation between the ore-controlling factors and the deposit position, which not only improves the mineral prediction effect, but also obtains a result with physical consistency, improves the interpretability of the mineral prediction result, and can more accurately delineate the prospecting area.

[0030] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for constructing a deep learning model for mineral prediction based on hard constraints of ore-forming models, characterized in that: include: S1: Based on the metallogenic model, determine the key ore-controlling elements; S2: Construct a power-law function relating the ore deposit distribution density to the distance of the key ore-controlling elements; S3: Assign values ​​to different ore-controlling elements based on the power law function to obtain the weights of different ore-controlling elements; S4: Construct the hidden layer of the original deep learning model based on key ore-controlling elements. Multiply the weight values ​​of different ore-controlling elements with the output of the original deep learning model and sum them. Add trainable weight parameters and bias parameters to construct the linear function in the neurons of the hidden layer. Use a non-linear activation function to operate on the output of the linear function to map the neuron features and increase the non-linearity of the hidden layer. Then construct the hidden layer based on the ore-forming model. Add the hidden layer based on the ore-forming model after the output layer of the original deep learning model to realize the hard constraint of the deep learning model. Then build the mineral prediction deep learning model based on the hard constraint of the ore-forming model. S5: Input multi-source mineral exploration information into a mineral prediction deep learning model based on hard constraints of a metallogenic model. Minimize the difference between the predicted value and the true value through the cross-entropy loss function to train the mineral prediction deep learning model based on hard constraints of a metallogenic model. When the predetermined accuracy is reached, the trained mineral prediction deep learning model based on hard constraints of a metallogenic model is obtained. Use the trained mineral prediction deep learning model based on hard constraints of a metallogenic model to delineate prospective mineral exploration areas. In step S2, the number of ore deposits falling on each buffer layer in each ore-controlling element layer is counted, and the cumulative number of ore deposits n and the ore deposit distribution density ρ are calculated, where ρ = n / d, and d represents the width of the buffer. A double logarithmic scatter plot is plotted with the buffer width d as the abscissa and the ore deposit spatial distribution density ρ as the ordinate. A straight line is fitted using the least squares method to obtain the power law function ρ = cd for the ore deposit distribution density and its distance from the ore-controlling element. a-2 Where c is a constant and a is the singularity index; In step S3, since different buffers in each ore-controlling element layer have different degrees of control over ore deposit formation, a weight σ is assigned to the buffer, where σ = cd. a-2 / ρ max c is a constant, d represents the width of the buffer, a is the singularity exponent, and ρ max The maximum value of the power-law function of the ore deposit distribution density and the distance to the ore-controlling element is used to assign values ​​to different buffer zones. That is, the closer the region is to the ore-controlling element, the greater the weight; the farther the region is from the ore-controlling element, the smaller the weight.

2. The method for constructing a deep learning model for mineral prediction based on hard constraints of an ore-forming model as described in claim 1, characterized in that: In step S1, based on the mineralization model, key mineral control elements are determined, and a buffer zone for mineral control elements is created using GIS.

3. The method for constructing a deep learning model for mineral prediction based on hard constraints of an ore-forming model as described in claim 1, characterized in that: In step S4, an equation based on a mineralization model is constructed. As a linear function in the hidden layer neurons, where x represents the mineralization probability of the output layer of the original deep learning model, k i ω represents the weights of the i-th mineral control element layer corresponding to each input sample in the deep learning model, ω is a trainable weight parameter, and b is a trainable bias term parameter. The Sigmoid activation function is used to operate on the output of the linear equation to map neuron features. Where μ is a constant.

Images