Multi-lithology identification and prediction method based on semi-supervised learning framework

Abstract

The invention provides a multi-lithology identification and prediction method based on a semi-supervised learning framework, and the method comprises the steps: generating a data set, and carrying out the preprocessing of the data set; obtaining a basic model, and initializing parameters of the basic model; optimizing parameters of the basic model based on a Bayesian function; performing iterative training on the basic model through a variational EM algorithm; the variational EM algorithm comprises a variational expectation step and a variational maximization step; whether the variational EM algorithm converges is judged, then the classification accuracy of the model is evaluated, and a prediction result of data without labels is observed. According to the method, an innovative solution is provided for multi-lithology identification and prediction, and the method has important practical significance and application value in the field of mineral exploration.

Description

Technical Field

[0001] The invention belongs to the technical field of geophysical exploration, and in particular relates to a multi-lithology identification and prediction method based on a semi-supervised learning framework. Background Art

[0002] In modern geological exploration, due to the increasing depletion of surface mineral resources, exploring deep underground blind deposits has become a new focus. Although existing mineral exploration technologies, especially those based on data-driven machine learning methods such as decision trees, neural networks, support vector machines, and random forests, have made some progress, they often show performance limitations when processing unseen data and are overly dependent on large amounts of labeled data. Summary of the invention

[0003] The purpose of this invention is to provide a multi-lithology identification and prediction method based on a semi-supervised learning framework, providing a more efficient and accurate solution for high-risk and high-cost mineral exploration.

[0004] The technical solution of the present invention is as follows:

[0005] A multi-lithology identification and prediction method based on a semi-supervised learning framework, the method comprising:

[0006] Generate a data set and preprocess the data set;

[0007] Obtaining a basic model and initializing parameters of the basic model;

[0008] Optimizing the parameters of the basic model based on the Bayesian function;

[0009] Iteratively training the basic model by using a variational EM algorithm; the variational EM algorithm includes a variational expectation step and a variational maximization step;

[0010] Determine whether the variational EM algorithm has converged, then evaluate the classification accuracy of the model and observe the prediction results of unlabeled data, and finally perform geological interpretation on the prediction results.

[0011] Furthermore, the generated data set is specifically:

[0012] Unify geophysical, geochemical, remote sensing and hyperspectral data into the same coordinate system to form a feature set of data samples;

[0013] The data provided by the prior geological information is used as a label and matched to the same position of the feature set; corresponding labels are designed for different lithologies and represented by non-zero numbers; labels without lithology representation are defined as 0.

[0014] Furthermore, the preprocessing is Gaussian-like and standardized processing, and the data set is updated.

[0015] Furthermore, the parameters for initialization include: the mean μ of each Gaussian component k , covariance Λ k and the mixing weight π k .

[0016] Furthermore, the parameters of the basic model are optimized based on the Bayesian function as follows:

[0017] Select an optimization target; the optimization target is the mixed weight π k and the precision parameter β of the Gaussian distribution;

[0018] Set the optimization target interval, the number of function iterations, and the convergence condition;

[0019] Use Gaussian process search to iterate repeatedly until the termination condition is met and output the optimal mixing weight π k and the precision parameter β of the optimal Gaussian distribution.

[0020] Furthermore, the variational expectation step is specifically as follows:

[0021] Based on the optimal mixing weight π k , the precision parameter β of the optimal Gaussian distribution, the mean μ of each Gaussian component k and covariance Λ k Calculate for each data point x i The posterior probability of each Gaussian component is the responsibility γ ij ;

[0022] Then the “Cannot-link” label constraint is performed to obtain the model’s responsibility matrix γ 3 ij .

[0023] Furthermore, the "Cannot-link" tag constraint is specifically:

[0024] First, the responsibility of each data point r ij Normalize to ensure that the responsibility of each data point sums to 1;

[0025] Then for each labeled data point x i , find the most likely cluster j;

[0026] For each cluster j and each data point x i , if x i The actual label l(i) and the label l(j) of the representative data point of cluster j * ) is different, then reduce x iThe degree of responsibility of j:

[0027] Normalize the updated responsibility to ensure that the sum of the responsibility of each data point to all clusters is 1, and finally update the model's responsibility matrix γ 3 ij .

[0028] Furthermore, the variational maximization step is specifically as follows:

[0029] Use the responsibility γ calculated by the variational expectation step 3 ij Update the parameter mean μ of each Gaussian component k , covariance Λ k and the mixing weight π k .

[0030] Furthermore, the criterion for judging whether the variational EM algorithm converges is:

[0031] Whether the algorithm has converged is determined based on the change in the variational lower bound. When the increase in the variational lower bound is less than a preset threshold or the variational lower bound no longer increases, the algorithm is considered stable and converged.

[0032] Compared with the prior art, the present invention has the following advantages:

[0033] The present invention can effectively use implicit information in the data to train the model in the case of a small number of labeled samples and a large number of unlabeled samples, and proposes a corresponding semi-supervised learning framework under the variational Bayesian model. At the same time, in the context of semi-supervised learning, the "Cannot-link" label constraint condition is designed to solve the problem that undefined data in the training set cannot be accurately predicted. At the same time, in order to solve the problem that the model is prone to fall into the local optimal solution, we introduced the Bayesian optimization algorithm, designed the minimization objective function, and improved the parameter adjustment efficiency. It provides an innovative solution for multi-lithology identification and prediction, which has important practical significance and application value in the field of mineral exploration.

[0034] The present invention not only solves the limitations of traditional supervised learning in terms of lack of data labels and misclassification of lithology in unknown areas, but also significantly reduces the overall exploration cost by reducing the reliance on large amounts of labeled data. At the same time, this method improves the accuracy of prediction of lithology categories in unknown areas and greatly enhances the ability to finely characterize underground geological structures. Overall, the present invention has demonstrated significant innovation and application value in both theory and practice, bringing new perspectives and technical methods to the field of earth science. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings generally illustrate various embodiments by way of example and not limitation, and together with the description and claims, serve to illustrate the embodiments of the invention. Where appropriate, the same reference numerals are used throughout the drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the present apparatus or method.

[0036] Figure 1 A schematic diagram of the method flow of the present invention is shown;

[0037] Figure 2 A schematic diagram of a data set of the present invention is shown;

[0038] Figure 3 A schematic diagram of a real model of the present invention is shown;

[0039] Figure 4 The gravity and magnetic inversion profile of the present invention is shown;

[0040] Figure 5 The prediction results of the present invention are shown. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0042] The present invention proposes a lithology identification and prediction method based on a semi-supervised variational Gaussian mixture model with Bayesian optimization. The method mainly comprises the following steps:

[0043] 1) Dataset generation: Unify geophysical, geochemical, remote sensing, hyperspectral and other data into the same coordinate system to form a feature set (Features) of data samples, use the data provided by prior geological information (lithology) such as drilling holes as labels (Labels), and match them to the same position in the feature set. Design corresponding labels for different lithologies, represented by non-zero numbers; labels without lithology representation are defined as 0. Datasets such as Figure 2 After the data set is generated, the Features are Gaussian-like and standardized, and the data set is updated.

[0044] 2) Parameter initialization: Before the model training process begins, a preliminary estimate of the model parameters is required, including the mean μ of each Gaussian component. k , covariance Λ k and the mixing weight π k wait.

[0045] 3) Parameter optimization: The purpose of this step is to optimize global parameters and improve the efficiency of parameter adjustment. kand β as optimization targets, and design the Bayesian function. When the algorithm is running, first set π k The interval range of the two parameters π and β, the number of iterations of the function, and the convergence condition. k is the mixing weight, β is the precision parameter of the Gaussian distribution, and the covariance matrix Λ k In the Bayesian optimization process, the Gaussian process is used as a priori to obtain a function to guide the search. The process is iterated repeatedly until the termination condition is met and the optimal π is output. k and the optimal β.

[0046] 4) Model training: The algorithm alternately executes two steps: the variational expectation step (E step) and the variational maximization step (M step).

[0047] Variational E step: First, based on the optimal parameter π obtained in step 3) k and β and other parameters in step 2) are calculated for each data point x i The posterior probability (i.e., responsibility γ) of each Gaussian component ij ), see formula (1):

[0048]

[0049] Among them, π k , μ k and Λ k are the mixture weight, mean and covariance of the kth Gaussian component, and r ij is the responsibility of data point n to belong to distribution k.

[0050] Then the “Cannot-link” label constraint is performed: First, the responsibility r of each data point is ij Normalize (Formula (2)) to ensure that the sum of the responsibility of each data point is 1:

[0051]

[0052] Then for each labeled data point x i , find its most likely cluster j:

[0053] j * (i) = argmax j γ 1 ij (3)

[0054] Here * (i) is the data point x i The most likely cluster index. For each cluster j and each data point x i , if x iThe actual label l(i) and the label l(j) of the representative data point of cluster j * ) is different, then reduce x i The degree of responsibility of j:

[0055]

[0056] Normalize the updated responsibility to ensure that the sum of the responsibility of each data point to all clusters is 1, and finally update the model's responsibility matrix γ 3 ij .

[0057] Variational M-step: Use the responsibility γ calculated by variational E-step 3 ij , we update the parameter mean μ of each Gaussian component k , covariance Λ k and the mixing weight π k . Parameter updating can be achieved by maximizing the expected complete-data log likelihood obtained in the E step. The expected complete-data log likelihood (ECCLL) is the expectation of the log likelihood of the complete data given the observed data and the current model parameter estimates. The "complete data" here includes the observed data X and the latent variable Z, and the expectation is calculated relative to the posterior distribution of the current estimate of the latent variable Z.

[0058] The variational EM algorithm iteratively executes the variational E step and variational M step until convergence. Convergence is determined based on the change in the variational lower bound (ELBO). When the increase in ELBO is less than a preset threshold, the algorithm is considered to have converged.

[0059] 5) Model evaluation and geological interpretation: First, determine whether the algorithm has converged based on the ELBO situation. If the ELBO no longer increases, the algorithm is considered stable and converged. Secondly, calculate the classification accuracy of labeled data (data with labels other than 0) to evaluate the classification accuracy of the model; and observe the prediction results of unlabeled data (labeled 0), which is the result predicted by the model. Finally, perform geological interpretation on the results and compare them with other prior data.

[0060] Example 1

[0061] Figure 3 For the real model, there are three anomalies (A1, A2, A3) and a uniform space (A4). We sampled the entire model into 15×39=585 sample data points. A1, A2 and A3 each account for 25 data points, and A4 has a total of 510 data points. Each sample data has three characteristic attributes, namely "depth", "magnetic susceptibility" and "density". Figure 4 It is a gravity and magnetic inversion profile. We set A1 and A2 as known labels, randomly selected 25 data of A4 as known labels, and A3 and the remaining A4 as unknown labels. Figure 5 This is the result of prediction using the technology proposed in the present invention. It can be found that the model predicts the data (A4) not seen in the training set very well, indicating that this method can perform multi-lithology classification and prediction very well.

[0062] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A multi-lithology identification and prediction method based on a semi-supervised learning framework, characterized in that: The method comprises: Generate a data set and preprocess the data set; Obtaining a basic model and initializing parameters of the basic model; Optimizing the parameters of the basic model based on the Bayesian function; Iteratively training the basic model by using a variational EM algorithm; the variational EM algorithm includes a variational expectation step and a variational maximization step; Determine whether the variational EM algorithm has converged, then evaluate the classification accuracy of the model and observe the prediction results of unlabeled data, and finally perform geological interpretation on the prediction results.

2. The multi-lithology identification and prediction method based on a semi-supervised learning framework according to claim 1 is characterized in that: The generated data set is specifically: Unify geophysical, geochemical, remote sensing and hyperspectral data into the same coordinate system to form a feature set of data samples; The data provided by the prior geological information is used as a label and matched to the same position of the feature set; corresponding labels are designed for different lithologies and represented by non-zero numbers; labels without lithology representation are defined as 0.

3. The multi-lithology identification and prediction method based on a semi-supervised learning framework according to claim 1 is characterized in that: The preprocessing is Gaussian-like and standardized processing, and the data set is updated.

4. The multi-lithology identification and prediction method based on a semi-supervised learning framework according to claim 1 is characterized in that: The parameters for initialization include: the mean μ of each Gaussian component k , covariance Λ k and the mixing weight π k .

5. The multi-lithology identification and prediction method based on a semi-supervised learning framework according to claim 1 is characterized in that: The optimization of the parameters of the basic model based on the Bayesian function is specifically as follows: Select an optimization target; the optimization target is the mixed weight π k and the precision parameter β of the Gaussian distribution; Set the optimization target interval, the number of function iterations, and the convergence condition; Use Gaussian process search to iterate repeatedly until the termination condition is met and output the optimal mixing weight π k and the precision parameter β of the optimal Gaussian distribution.

6. The multi-lithology identification and prediction method based on a semi-supervised learning framework according to claim 1 is characterized in that: The variational expectation step is specifically as follows: Based on the optimal mixing weight π k , the precision parameter β of the optimal Gaussian distribution, the mean μ of each Gaussian component k and covariance Λ k Calculate for each data point x i The posterior probability of each Gaussian component is the responsibility γ ij ; Then the "Cannot-link" label constraint is performed to obtain the model's responsibility matrix γ 3 ij .

7. The multi-lithology identification and prediction method based on a semi-supervised learning framework according to claim 6 is characterized in that: The "Cannot-link" tag constraints are specifically: First, the responsibility of each data point r ij Normalize to ensure that the responsibility of each data point sums to 1; Then for each labeled data point x i , find the most likely cluster j; For each cluster j and each data point x i , if x i The actual label l(i) and the label l(j) of the representative data point of cluster j * ) is different, then reduce x i The degree of responsibility of j: Normalize the updated responsibility to ensure that the sum of the responsibility of each data point to all clusters is 1, and finally update the model's responsibility matrix γ 3 ij .

8. The multi-lithology identification and prediction method based on a semi-supervised learning framework according to claim 1 is characterized in that: The variational maximization step is specifically as follows: Use the responsibility γ calculated by the variational expectation step 3 ij Update the parameter mean μ of each Gaussian component k , covariance Λ k and the mixing weight π k .

9. The multi-lithology identification and prediction method based on a semi-supervised learning framework according to claim 1 is characterized in that: The criterion for judging whether the variational EM algorithm converges is: Whether the algorithm has converged is determined based on the change in the variational lower bound. When the increase in the variational lower bound is less than a preset threshold or the variational lower bound no longer increases, the algorithm is considered stable and converged.

Images