A three-dimensional mineralization prediction method, medium, equipment, and product

Abstract

This invention proposes a three-dimensional mineralization prediction method, medium, equipment, and product, belonging to the field of mineralization prediction technology. The method includes: collecting geophysical, geochemical, and remote sensing data in the mineralization area and preprocessing them to construct a three-dimensional multi-source attribute dataset, which is divided into a training set, a validation set, and a test set; using the training set with added noise to initially train a denoising autoencoder; adding a discriminator after the initially trained encoder to train both the denoising autoencoder and the discriminator; using the validation and test sets to validate and test the trained encoder, decoder, and discriminator; adding a fully connected classification layer after the trained encoder, freezing the parameters of the trained encoder, training the fully connected classification layer with labeled mineralization samples, and inputting the noisy data of the area to be predicted into the finally trained encoder and fully connected classification layer to obtain the mineralization prediction result. This invention has strong noise resistance.

Description

Technical Field

[0001] This invention relates to the field of mineralization prediction technology, and in particular to a three-dimensional mineralization prediction method, medium, equipment, and product. Background Technology

[0002] As mineral exploration technology develops towards intelligence and automation, the close integration of digital technology and geological exploration has driven rapid changes in the industry. The explosive growth of geological big data has promoted the application of data-driven machine learning methods in mineral prediction. Mahboob et al. (2022) published a review on the application of machine learning in mineral resource exploration, which compared the characteristics and application effects of various machine learning methods in detail, discussed the core problems faced by intelligent mineral exploration, and demonstrated the advantages of machine learning methods over traditional methods. Studies have shown that machine learning has broad application prospects in the field of mineralization prediction and can significantly improve the accuracy and efficiency of exploration. Liu Yanpeng et al. (2020) studied the Dongzhi Zhaojikou lead-zinc deposit in Anhui Province and trained a convolutional neural network mineralization prediction model using geochemical element distribution data in the study area. The model achieved an accuracy of over 90% on the validation set. In unsupervised machine learning, Li Yuwei et al. (2022) took Ulanhot, Inner Mongolia as an example and used fuzzy clustering to conduct spatial mineralization analysis on geological, geophysical, and geochemical data to identify mineralization patterns. Ghezelbash et al. (2023) developed a conceptual model of porphyry copper deposits in the Varzaghan region of northwestern Iran. They extracted six layers of evidence from geochemical, geological, tectonic, and remote sensing data and constructed a GIS-based mineral prospect map using traditional K-means clustering and supervised SVM (Support Vector Machine).

[0003] However, data noise interference leading to feature extraction bias is prevalent in mineralization prediction tasks, posing a serious obstacle to machine learning-based mineralization prediction. Previous research models have shown insufficient robustness to noise and lack the ability to fuse multi-source heterogeneous data, thus failing to effectively extract key mineralization features. Summary of the Invention

[0004] The purpose of this invention is to address the problem of feature extraction bias caused by data noise interference in existing three-dimensional mineralization prediction methods, and to propose a three-dimensional mineralization prediction method comprising the following steps:

[0005] S1. Collect geophysical, geochemical and remote sensing data in the metallogenic area and preprocess them to construct a three-dimensional multi-source attribute dataset, which is divided into training set, validation set and test set;

[0006] S2. Add noise to the training set and use the noise-added training set to train the denoising autoencoder. The denoising autoencoder includes an encoder and a decoder, resulting in a pre-trained encoder and decoder.

[0007] S3. After the trained encoder, add a discriminator to train the initially trained encoder, decoder and discriminator to obtain the trained encoder, decoder and discriminator.

[0008] The trained encoder, decoder, and discriminator are validated and tested using validation and test sets;

[0009] S4. Add a fully connected classification layer after the trained encoder, freeze the parameters of the trained encoder, train the fully connected classification layer with the labeled mineralized samples, and obtain the trained fully connected classification layer. Input the noisy data of the region to be predicted into the trained encoder and the fully connected classification layer to obtain the mineralized prediction result.

[0010] Furthermore, the preprocessing includes mesh generation, one-hot encoding, and normalization.

[0011] Furthermore, the encoding process is represented as:

[0012]

[0013] in, This represents the latent representation of the i-th sample output by the encoder. This represents the non-linear activation function in the encoder. Let W represent the i-th sample with added noise in the encoder input, and let W and b be the weight matrix and bias vector of the function in the encoder, respectively.

[0014] The decoding process is represented as follows:

[0015]

[0016] in, This represents the reconstructed output of the i-th sample from the decoder. Let V represent the nonlinear activation function in the decoder, and V and c represent the weight matrix and bias vector of the function in the decoder, respectively.

[0017] Furthermore, the reconstruction error loss function of the denoising autoencoder is expressed as:

[0018]

[0019] in, This represents the reconstruction error loss function of the denoising autoencoder, where N represents the number of samples. This represents the reconstructed output of the i-th sample from the decoder. Let |i| represent the i-th input sample, and ||˙|| represent the Euclidean norm.

[0020] Furthermore, the discriminator is a binary classification network.

[0021] Furthermore, the initially trained encoder, decoder, and discriminator are trained as follows:

[0022] The initially trained encoder and decoder are used as intermediate encoders and decoders; the discriminator is used as an intermediate discriminator.

[0023] Step 1: Input the training set into the frozen intermediate encoder to obtain the true latent representation; input the training set with added noise into the frozen intermediate encoder to obtain the first fake latent representation; input the true latent representation and the first fake latent representation into the intermediate discriminator, calculate the discrimination loss, and update the intermediate discriminator to the intermediate discriminator trained in Step 1.

[0024] Step 2: Input the training set with added noise into the intermediate encoder to obtain the second fake latent representation; input the second fake latent representation into the intermediate decoder and the frozen intermediate discriminator. The intermediate discriminator obtains the probability of the second fake latent representation, calculates the loss of the encoder based on the probability of the second fake latent representation, calculates the reconstruction error loss function of the denoising autoencoder, and updates the intermediate encoder and intermediate decoder to the intermediate encoder and intermediate decoder trained in Step 2.

[0025] Repeat the above two steps until the accuracy of the intermediate discriminator no longer improves; you will then obtain the trained encoder, decoder, and discriminator.

[0026] Furthermore, the formula for calculating the discriminant loss is as follows:

[0027]

[0028] in, This represents the discriminative loss, where N represents the number of samples. Representing the true potential representation, Indicates the discriminator pair The probability of the output. This indicates the first potential forgery representation. Indicates the discriminator pair The probability of the output;

[0029] The encoder loss is calculated using the following formula:

[0030]

[0031] in, This represents the encoder loss, where N represents the number of samples. This indicates a potential forgery of the second representation. Indicates the discriminator pair The probability of the output;

[0032] The total loss of the denoising autoencoder is calculated as follows:

[0033]

[0034] in, Indicates the total loss. This represents the reconstruction error loss function of the denoising autoencoder. This represents the weighting parameter.

[0035] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described three-dimensional mineralization prediction method.

[0036] The present invention also proposes an electronic device, including a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to execute the above-described three-dimensional mineralization prediction method.

[0037] The present invention also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described three-dimensional mineralization prediction method.

[0038] The beneficial effects of the technical solution provided by this invention are:

[0039] This invention constructs a denoising autoencoder and performs initial training. A discriminator is added after the initially trained encoder to form a generative adversarial network (GAN) for adversarial training. The discriminator constrains the "authenticity" of the latent representations generated by the encoder, enabling the model to learn features "directly related to mineralization," reducing redundant information interference and improving the "specificity" of predictions. This invention has strong noise resistance, effectively resisting unknown noise interference, and solves the problem of feature extraction bias caused by data noise interference in 3D mineralization prediction in existing technologies, improving prediction stability. The model structure of this invention is simple and easy to deploy quickly. Attached Figure Description

[0040] Figure 1 This is a flowchart of a three-dimensional mineralization prediction method according to an embodiment of the present invention;

[0041] Figure 2 This is a block diagram of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0043] The flowchart of the three-dimensional mineralization prediction method of this invention is as follows: Figure 1 Specifically, it includes the following steps:

[0044] S1. Geophysical, geochemical, and remote sensing data are collected in the metallogenic area and preprocessed. Preprocessing methods include grid partitioning, one-hot encoding, and normalization. The preprocessed data is used to construct a three-dimensional multi-source attribute dataset, which is then divided into training, validation, and test sets.

[0045] S2. Add noise to the training set and use the noisy training set to initialize the training of the denoising autoencoder, obtaining a pre-trained encoder and decoder. This initial training enables the encoder-decoder to reconstruct the original data from the noisy data, avoiding instability caused by the conflict between "adversarial loss" and "reconstruction loss" in later steps during the initial adversarial training phase. The denoising autoencoder consists of an encoder and a decoder, which adopt a symmetrical mirror structure.

[0046] The encoding process is represented as follows:

[0047]

[0048] in, This represents the latent representation of the i-th sample output by the encoder. This invention uses the ELU activation function to represent the nonlinear activation function in the encoder. Let W represent the i-th sample with added noise in the encoder input, and let W and b be the weight matrix and bias vector of the function in the encoder, respectively.

[0049] The decoding process is represented as follows:

[0050]

[0051] in, This represents the reconstructed output of the i-th sample from the decoder. V represents the nonlinear activation function in the decoder. In this invention, the ELU activation function is used. V and c are the weight matrix and bias vector of the function in the decoder, respectively.

[0052] The loss function of a denoising autoencoder is typically the reconstruction error, which is the sum of the noise-free input x and the reconstructed output. The difference between them. Here, MSE (mean squared error) is used as the loss function.

[0053]

[0054] in, This represents the reconstruction error loss function, where N represents the number of samples. This represents the reconstructed output of the i-th sample from the decoder. Let |i| represent the i-th input sample, and ||˙|| represent the Euclidean norm.

[0055] S3. After training the encoder, add a discriminator. Using the encoder as a generator, the discriminator and the discriminator form a generative adversarial network (GAN) to construct an adversarial training framework. This allows the encoder to generate a latent representation that closely approximates the real data distribution, enabling the discriminator to distinguish between real and fake data as accurately as possible. The initially trained encoder, decoder, and discriminator are then trained to obtain the trained encoder, decoder, and discriminator.

[0056] The discriminator is designed as a binary classification network. The input is the latent representation z, and the output is the probability value p∈[0,1] of z (p=1 represents "true latent representation", p=0 represents "fake latent representation").

[0057] The encoder, decoder, and discriminator, which have undergone initial training, are then trained. The training process is as follows:

[0058] First, the initially trained encoder and decoder are used as intermediate encoders and decoders; the discriminator is used as an intermediate discriminator.

[0059] Then, perform the following cyclical training steps:

[0060] Step 1: Freeze the parameters of the intermediate encoder, input the training set into the frozen intermediate encoder to obtain the true latent representation; input the training set with added noise into the frozen intermediate encoder to obtain the first fake latent representation; input the true latent representation and the first fake latent representation into the intermediate discriminator, calculate the discrimination loss, and update the intermediate discriminator to the intermediate discriminator trained in Step 1.

[0061] The formula for calculating discriminant loss is as follows:

[0062]

[0063] in, This represents the discriminative loss, where N represents the number of samples. Representing the true potential representation, Indicates the discriminator pair The output probability (should be close to 1). This indicates the first potential forgery representation. Indicates the discriminator pair The output probability (should be close to 0). The discriminator parameters are optimized through backpropagation; in this step, only the discriminator weights are updated.

[0064] Step 2: Freeze the parameters of the intermediate discriminator, input the training set with added noise into the intermediate encoder to obtain the second spoofed latent representation; input the second spoofed latent representation into the intermediate decoder and the frozen intermediate discriminator, the intermediate discriminator obtains the probability of the second spoofed latent representation, calculates the encoder loss based on the probability of the second spoofed latent representation, calculates the reconstruction error loss function of the denoising autoencoder, and updates the intermediate encoder and intermediate decoder to the intermediate encoder and intermediate decoder trained in Step 2.

[0065] The encoder loss is calculated using the following formula:

[0066]

[0067] in, This represents the encoder loss, where N represents the number of samples. This indicates a potential forgery of the second representation. Indicates the discriminator pair The probability of the output.

[0068] The formula for calculating the reconstruction error loss function of a denoising autoencoder is as follows:

[0069]

[0070] in, This represents the reconstruction error loss function, where N represents the number of samples. This represents the reconstructed output of the i-th sample from the decoder. Let |i| represent the i-th input sample, and ||˙|| represent the Euclidean norm.

[0071] The total loss of the denoising autoencoder is calculated as follows:

[0072]

[0073] in, Indicates the total loss. This represents the weight parameters, used to balance reconstruction and adversarial objectives. The encoder and decoder parameters are optimized through backpropagation.

[0074] Repeat the above two steps until the accuracy of the intermediate discriminator no longer improves (i.e., it can no longer effectively distinguish between real and fake latent representations). Evaluate different combinations of hyperparameters using a validation set, select the optimal configuration, and obtain the trained encoder, decoder, and discriminator. At this point, the z generated by the encoder has strong robustness.

[0075] S4. Add noise to the test set and input it into the trained encoder to obtain the latent representation. Potential representation Input the trained decoder to obtain reconstructed data. Potential representation Input the trained discriminator to obtain the latent representation. The probability of [the probability]. Model testing is used to evaluate the model's denoising ability and the robustness of the encoder in generating latent representations. The denoising ability is measured by the reconstruction error, and the model's denoising ability is measured by the discriminator's accuracy. Robustness.

[0076] S5. After the trained encoder, add a fully connected classification layer with the Softmax activation function. Freeze the parameters of the trained encoder. Train the fully connected classification layer using labeled mineralized samples, using cross-entropy loss as the loss function. The optimization objective is to enable the classification layer to determine whether mineralization has occurred based on the latent representation z output by the encoder. The trained fully connected classification layer is then obtained. Noisy data from the region to be predicted is input into the trained encoder and the fully connected classification layer to obtain the mineralization prediction result.

[0077] The invention will be further illustrated below with specific examples.

[0078] Multi-source attribute data, including geophysical, geochemical, and remote sensing data, were collected from the study area to perform three-dimensional mineralization prediction. The raw dataset underwent preprocessing, including inverse distance spatial interpolation, separation of quantitative and categorical data, normalization of the quantitative data, one-hot encoding of the categorical data, and addition of Gaussian noise.

[0079] Data is divided into quantitative and descriptive data. Quantitative data includes geochemical data and geofeature data. Geochemical data specifically includes fluorine, acid-insoluble matter, iodine, and phosphorus; geofeature data specifically includes density, magnetic susceptibility, and resistivity. Descriptive data includes stratigraphic data, lithological data, and fault data.

[0080] The quantitative data were normalized, as shown in Table 1.

[0081] Table 1. Normalized quantitative data

[0082]

[0083] One-hot encoding is performed on the categorized data, as shown in Table 2.

[0084] Table 2 lists the one-hot encodings of several categorical data.

[0085]

[0086] A dataset was constructed based on known labeled positive and negative samples, containing 20,839 positive samples and 11,483 negative samples.

[0087] First, the preprocessed samples were divided into a training set (70%), a validation set (15%), and a test set (15%). Table 3 shows the comparison of the mineralization prediction accuracy between the ATER-DAE model of this invention and the traditional DAE model.

[0088] Table 3

[0089]

[0090] The present invention employs the above-mentioned three-dimensional mineralization prediction method based on the ATER-DAE model, which can solve the problem of feature extraction deviation caused by data noise interference in the existing three-dimensional mineralization prediction. By constructing a robust latent representation space, the accuracy and reliability of three-dimensional mineralization prediction can be improved.

[0091] In one exemplary embodiment, a computer-readable storage medium is included, which stores a computer program that, when executed by a processor, implements the above-described three-dimensional mineralization prediction method.

[0092] Please see Figure 2 In one exemplary embodiment, the device further includes an electronic device including at least one processor, at least one memory, and at least one communication bus.

[0093] The memory stores a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the aforementioned three-dimensional mineralization prediction method.

[0094] In one exemplary embodiment, a computer program product is proposed, including a computer program / instructions that, when executed by a processor, implement the steps of the three-dimensional mineralization prediction method described above.

[0095] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method of three-dimensional ore prediction, characterized by, The method comprises the following steps: S1, collecting geophysical, geochemical and remote sensing data in a mineralization area and preprocessing, constructing a three-dimensional multi-source attribute data set, and dividing into a training set, a validation set and a test set; S2, adding noise to the training set, training the denoising autoencoder using the training set with added noise, the denoising autoencoder comprising an encoder and a decoder, obtaining a preliminarily trained encoder and a decoder; S3, adding a discriminator after the preliminarily trained encoder, training the preliminarily trained encoder, decoder and discriminator, and obtaining a trained encoder, decoder and discriminator; The training process is as follows: The preliminarily trained encoder and decoder are used as intermediate encoder and intermediate decoder; the discriminator is used as intermediate discriminator; Step one: input the training set into the frozen intermediate encoder to obtain a real latent representation; input the training set with added noise into the frozen intermediate encoder to obtain a first fake latent representation; input the real latent representation and the first fake latent representation into the intermediate discriminator, and calculate the discrimination loss to update the intermediate discriminator to the intermediate discriminator after step one training; Step two: input the training set with added noise into the intermediate encoder to obtain a second fake latent representation; input the second fake latent representation into the intermediate decoder and the frozen intermediate discriminator, and the intermediate discriminator obtains the probability of the second fake latent representation; calculate the loss of the encoder according to the probability of the second fake latent representation, and calculate the reconstruction error loss function of the denoising autoencoder; update the intermediate encoder and the intermediate decoder to the intermediate encoder and the intermediate decoder after step two training; Repeat training steps one and two until the accuracy of the intermediate discriminator no longer improves; obtain the trained encoder, decoder and discriminator; The loss calculation formula of the encoder is as follows: wherein, represents a loss of the encoder, N represents a number of samples, represents a second fake latent representation, represents a probability of the output by the discriminator; represents a probability of the output by the discriminator; The total loss of the denoising autoencoder is calculated as follows: wherein, represents the total loss, represents the reconstruction error loss function of the denoising autoencoder, represents the weight parameter; The trained encoder, decoder and discriminator are verified and tested using the validation set and test set; S4, adding a fully connected classification layer after the trained encoder, freezing the parameters of the trained encoder, training the fully connected classification layer with labeled mineralization samples, obtaining a trained fully connected classification layer, and inputting noisy data of the area to be predicted into the trained encoder and the fully connected classification layer to obtain a mineralization prediction result.

2. The method according to claim 1, characterized in that, The preprocessing includes grid subdivision, one-hot encoding and normalization processing.

3. The method according to claim 1, wherein, The encoding process is represented as: wherein, represents a latent representation of the i-th sample of the encoder output, represents a non-linear activation function in the encoder, represents the i-th added noised sample of the encoder input, W and b are a weight matrix and a bias vector of the function in the encoder, respectively; The decoding process is represented as: wherein, denotes the reconstructed output of the i-th sample output by the decoder, denotes a non-linear activation function in the decoder, V and c are a weight matrix and a bias vector of the function in the decoder, respectively.

4. The method according to claim 1, wherein, The reconstruction error loss function of the denoising autoencoder is represented as: wherein, represents a reconstruction error loss function of the denoising autoencoder, N represents the number of samples, represents the reconstructed output of the i-th sample output by the decoder, represents the i-th input sample, ||·|| represents the Euclidean norm.

5. The method according to claim 1, wherein, The discriminator is a binary classification network.

6. The method according to claim 1, wherein, The discrimination loss calculation formula is as follows: wherein, represents a discriminative loss, N represents the number of samples, represents a real latent representation, represents a probability output by the discriminator, represents a first fake latent representation, represents a probability output by the discriminator, represents a second fake latent representation, represents a probability output by the discriminator.

7. A computer readable storage medium storing a computer program, characterized in that: The computer program is executed by the processor to realize the method of any one of claims 1-6.

8. An electronic device, comprising: The computer program is executed by the processor to realize the method of any one of claims 1-6.

9. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to realize the steps of the method of any one of claims 1-6.

Images