A deep learning three-dimensional quantitative prediction method for concealed ore body

Abstract

The embodiment of the present application provides a kind of concealed ore body depth learning three-dimensional quantitative prediction method, belong to data processing technical field, specifically include: the sequence sliding window of three-dimensional geologic body element ore-controlling index is constructed;The original ore-controlling feature in the serialized sample is mapped to high-dimensional representation space;Initial feature sequence is input into selective state space model, and deep feature sequence is output;Selective state space model is trained using sample weighting strategy, and model parameters are optimized;The model trained is applied to all body elements, the multiple prediction probabilities obtained by different sliding windows for the same body element are spatially weighted and fused, the final mineralization probability of the body element is obtained, and a three-dimensional mineralization probability body is constructed;The ore-controlling index corresponding to the body element of the region to be predicted is input into the three-dimensional mineralization probability body, and the concealed ore body prediction target area is delineated according to the preset probability threshold value.The scheme of the present application improves the prediction accuracy and interpretability.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a deep learning-based three-dimensional quantitative prediction method for concealed ore bodies. Background Technology

[0002] Currently, in the quantitative prediction of concealed ore bodies, due to the significant differences in deep geological patterns and the obvious asymmetry in observable information, prediction models often struggle to accurately depict the true mineralization control mechanisms. While traditional geostatistical methods possess strong geological interpretability, they are ill-suited to effectively handle the nonlinear relationships arising from changes in deep geological patterns. Conversely, machine learning-based methods, despite their strong nonlinear modeling capabilities, lack reliable geological interpretability when faced with insufficient deep information and inadequate geological constraints.

[0003] It is evident that there is an urgent need for a deep learning-based three-dimensional quantitative prediction method for concealed ore bodies with high accuracy and interpretability. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a deep learning-based three-dimensional quantitative prediction method for concealed ore bodies, which at least partially solves the problems of poor prediction accuracy and interpretability in the prior art.

[0005] This invention provides a deep learning-based three-dimensional quantitative prediction method for concealed ore bodies, comprising:

[0006] Step 1: Construct a serialized sliding window for the mineralization control indicators of three-dimensional geological elements to obtain serialized samples;

[0007] Step 2: Map the original ore-controlling features in the serialized samples to a high-dimensional representation space to obtain the initial feature sequence;

[0008] Step 3: Input the initial feature sequence into the selective state space model, perform dependency modeling and feature extraction through stacked Mamba modules, and output the deep feature sequence;

[0009] Step 4: Based on the features representing the center position of the window in the deep feature sequence, calculate the mineralization probability through a classification network, and train a selective state space model using a sample weighting strategy to optimize the model parameters;

[0010] Step 5: Apply the trained model to all voxels, perform spatial weighted fusion of multiple predicted probabilities obtained from different sliding windows for the same voxel to obtain the final mineralization probability of the voxel, and construct a three-dimensional mineralization probability volume.

[0011] Step 6: Input the ore-controlling indicators corresponding to the volume elements of the area to be predicted into the three-dimensional mineralization probability volume, and delineate the target area for predicting the hidden ore body according to the preset probability threshold.

[0012] According to a specific implementation of an embodiment of the present invention, step 1 specifically includes:

[0013] Define the window length on the 3D voxel set. and step length For each central element Constructing local feature sequences:

[0014] ;

[0015] in, For volume element The vector formed by the dimensional control ore characteristics, This represents the transpose of a vector, and a sliding window is used to generate serialized samples and label the central volume elements. , Let L be the dimension of the real matrix, L be the sequence length, and F be the feature dimension.

[0016] According to a specific implementation of an embodiment of the present invention, step 2 specifically includes:

[0017] For each ore-controlling feature in the local feature sequence After performing a linear transformation, a nonlinear activation function is used to process the data to obtain the corresponding initial latent representation:

[0018] ;

[0019] in, Represents the initial potential representation. For the input weight matrix, For the input bias vector, For ReLU function, For tensor shape, For model dimensions.

[0020] According to a specific implementation of an embodiment of the present invention, step 3 specifically includes:

[0021] Step 3.1: Input the initial feature sequence into the model consisting of N sequentially connected Mamba modules;

[0022] Step 3.2, each Mamba module is selectively gated. State update matrix and control items For each sequence position Use selective discrete SSM to update and obtain the hidden state. and output the state as Based on this, a deep representation is formed, wherein the expression for the selective discrete SSM update is:

[0023] ;

[0024] ;

[0025] ;

[0026] in, Indicates the step size parameter. Represents the state transition matrix. For the input matrix, For the output matrix, To discretize matrix A, To discretize the B matrix, For the state dimension, For gated bias. Let be the hidden state at time t. Let be the hidden state at time t+1. To perform element-wise multiplication, update the hidden state point by point along the sequence direction. and output the state as to form a deep expression , A lightweight multilayer perceptron is used to determine the current input features. Adaptive generation;

[0027] Step 3.3: Stack the deep representations of each Mamba module to form a deep feature sequence.

[0028] According to a specific implementation of an embodiment of the present invention, step 4 specifically includes:

[0029] Step 4.1: Extract the representation of the center position of the window in the deep feature sequence, input it into a classification head that contains at least one fully connected layer, and obtain the mineralization probability prediction value.

[0030] Step 4.2: Based on the predicted mineralization probability, train the selective state-space model using weighted binary cross-entropy loss as the loss function, and optimize the model parameters. The expression for the weighted binary cross-entropy loss is:

[0031] ;

[0032] in, This is the loss value. For real labels Category weights, For the first The sequence number is... The predicted probability of each point.

[0033] According to a specific implementation of an embodiment of the present invention, step 5 specifically includes:

[0034] Step 5.1: Apply the trained model to all voxels;

[0035] Step 5.2, for each volume element Collect its in The set of predicted probabilities obtained from different sliding windows:

[0036] ;

[0037] in It is the number of times the window is covered. In the first In each window, the volume element The probability of being predicted as positive.

[0038] Step 5.3, using weights For the set of prediction probabilities Weighted averages are performed in the logit space to obtain the volume elements. The final probability of mineralization;

[0039] Step 5.4: Integrate the final mineralization probabilities of all elements to construct a three-dimensional mineralization probability volume.

[0040] The deep learning-based three-dimensional quantitative prediction scheme for concealed ore bodies in this embodiment of the invention includes: Step 1, constructing a serialized sliding window of ore-controlling indicators for three-dimensional geological elements to obtain serialized samples; Step 2, mapping the original ore-controlling features in the serialized samples to a high-dimensional representation space to obtain an initial feature sequence; Step 3, inputting the initial feature sequence into a selective state-space model, performing dependency modeling and feature extraction through stacked Mamba modules, and outputting a deep feature sequence; Step 4, calculating the mineralization probability through a classification network based on the features representing the center position of the window in the deep feature sequence, and training the selective state-space model using a sample weighting strategy to optimize the model parameters; Step 5, applying the trained model to all elements, performing spatial weighted fusion of multiple prediction probabilities obtained from different sliding windows for the same element to obtain the final mineralization probability of the element, and constructing a three-dimensional mineralization probability volume; Step 6, inputting the ore-controlling indicators corresponding to the elements in the area to be predicted into the three-dimensional mineralization probability volume, and delineating the target area for concealed ore body prediction according to a preset probability threshold.

[0041] The beneficial effects of the embodiments of the present invention are as follows: By using the structured state space model (SSM) to capture the correlation between mineralization information indicators and mineralization distribution, the present invention not only has the ability of deep learning models to handle nonlinear relationships, but also avoids the computational cost problem of Transformer-type models under deep and dense volume data. It can directly learn deep-level mineralization control features and mineral exploration indicators from the three-dimensional structural data of the ore deposit, and efficiently realize the fusion and expression of deep mineralization features. It provides a more suitable model structure for the prediction of deep concealed ore bodies, and improves the prediction accuracy and interpretability. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart illustrating a deep learning-based three-dimensional quantitative prediction method for concealed ore bodies provided in an embodiment of the present invention;

[0044] Figure 2 This invention provides a schematic diagram of the decrease curve of the loss value of the training set and the validation set during model training and a schematic diagram of the increase curve of the accuracy of the training set and the validation set during training. (a) is a schematic diagram of the decrease curve of the loss value of the training set and the validation set during model training, and (b) is a schematic diagram of the increase curve of the accuracy of the training set and the validation set during training.

[0045] Figure 3 A ROC curve for model training provided in an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of volumetric elements in the known and unknown areas of a mineral deposit in a certain region, provided as an embodiment of the present invention.

[0047] Figure 5 This is a schematic diagram illustrating the prediction results of existing ore bodies and unknown areas in a certain location, provided as an embodiment of the present invention. Detailed Implementation

[0048] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0049] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0050] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this invention, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0051] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0052] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0053] This invention provides a deep learning-based three-dimensional quantitative prediction method for concealed ore bodies, which can be applied to the ore body exploration process in geological scenarios.

[0054] See Figure 1 This is a flowchart illustrating a deep learning-based three-dimensional quantitative prediction method for concealed ore bodies provided in an embodiment of the present invention. Figure 1 As shown, the method mainly includes the following steps:

[0055] Step 1: Construct a serialized sliding window for the mineralization control indicators of three-dimensional geological elements to obtain serialized samples;

[0056] The purpose of this invention is to provide a deep learning-based three-dimensional quantitative prediction method for concealed ore bodies based on the Mamba model. By utilizing the structured state space model (SSM) to capture the correlation between mineralization information indicators and mineralization distribution, this method not only possesses the ability of deep learning models to handle nonlinear relationships but also avoids the computational cost problem of Transformer-type models under deep and dense volume data. It can directly learn deep-level ore-controlling features and prospecting indicators from the three-dimensional structural data of the ore deposit, and efficiently achieve the fusion and expression of deep mineralization features, providing a more suitable model structure for the prediction of deep concealed ore bodies.

[0057] In practice, a fixed window length can be defined on the three-dimensional voxel point set. and step length For each central body element Constructing local feature sequences:

[0058] (1)

[0059] in, For volume element The vector formed by the mineral control indicators, where T represents the transpose of the vector. It is the sequence length. The input is the number of features; a sliding window is used to generate all samples and label the central volume elements. .

[0060] Step 2: Map the original ore-controlling features in the serialized samples to a high-dimensional representation space to obtain the initial feature sequence;

[0061] In practice, to enhance the feature representation capability, the input ore control index first undergoes two layers of fully connected feature linear mapping to obtain the initial hidden state, completing the transformation from the geological feature space to the high-dimensional semantic space and enhancing the representation capability of deep nonlinear patterns.

[0062] (2)

[0063] The output dimension is:

[0064] (3)

[0065] in, This is the initial hidden state. For the first layer of weights, For the second layer of weights, As input features, , , It uses ReLU activation, where B is the batch size and L is the sequence length. It is the model dimension. It is the number of input features.

[0066] Mineral control characteristics First, after a linear transformation, the ReLU activation function is used to avoid gradient vanishing.

[0067] (4)

[0068] The definition of ReLU is as follows:

[0069] (5)

[0070] in, It is in a hidden state. For the first layer of weights, Let be the input features of the t-th individual element. This is a bias term.

[0071] Step 3: Input the initial feature sequence into the selective state space model, perform dependency modeling and feature extraction through stacked Mamba modules, and output the deep feature sequence;

[0072] In practical implementation, the ore control indicators after linear mapping Entering the main body of the Mamba model, by Each MambaBlock consists of:

[0073] (6)

[0074] in, This represents the hidden state at level k. This represents the (k-1)th hidden state.

[0075] Each MambaBlock contains: Selective State Space Model (SSM) and Input / Output Gating;

[0076] Dynamic convolutional representation of SSM:

[0077] (7)

[0078] in The output at time t, For the convolution kernel at the k-th time step, This is the input from the k-th time in the past. The state equations are discretized to obtain:

[0079] (8a)

[0080] (8b)

[0081] in, The state at time t, This represents the state at time t+1. The output state at time t. Let A be the input at time t, B be the state matrix, and C be the input matrix.

[0082] Selective State-Space Model (SSM):

[0083] (9a)

[0084] (9b)

[0085] in, For selective input matrices, For selective output matrix, The input features at time t are... For the input matrix weights, To output the matrix weights, This is the normalization function.

[0086] The selective state-space model uses an input-dependent selection matrix to enable the SSM parameters to be dynamically adjusted according to the ore-controlling characteristics, thereby taking into account the differences in deep patterns, nonlinear mineralization processes, and three-dimensional spatial continuity.

[0087] Step 4: Based on the features representing the center position of the window in the deep feature sequence, calculate the mineralization probability through a classification network, and train a selective state space model using a sample weighting strategy to optimize the model parameters;

[0088] In practice, mineral control indicators The output layer, consisting of two fully connected layers, completes the final classification probability prediction.

[0089] ;

[0090] (10)

[0091] in, For unnormalized category scores, For the final layer features, For the weights of the first fully connected layer, For the weights of the second fully connected layer, For the first layer bias, For the second layer bias, For the activation function, the output size is B is the batch size, and L is the sequence length. It is the number of target categories.

[0092] Applying softmax activation to the final linear output yields the probability matrix:

[0093] (11)

[0094] in, For each point, the predicted probability is given for each category, and softmax is the normalization function.

[0095] The probability of a single point forming mineral deposits is the second type:

[0096] (12)

[0097] in, Let be the mineralization probability at point t. Let be the probability of the second class at point t in all batches.

[0098] Weighted cross-entropy:

[0099] (13)

[0100] in, This is the loss value. For real labels Category weights, For the first The sequence number is... The predicted probability of each point. The weights are set according to the sample proportion:

[0101] (14)

[0102] in Negative class weights For positive class weights, The number of negative class samples. This represents the number of positive class samples.

[0103] Step 5: Apply the trained model to all voxels, perform spatial weighted fusion of multiple predicted probabilities obtained from different sliding windows for the same voxel to obtain the final mineralization probability of the voxel, and construct a three-dimensional mineralization probability volume.

[0104] In practice, to improve the accuracy of the mineralization probability of all volumes in the prediction area, repeated predictions of the same volume from multiple windows are weighted and integrated using a probability fusion strategy.

[0105] For any volume element It may be predicted multiple times as the center of different sliding windows, denoted as . Assume that a certain body element During the sliding window prediction phase Each window covers a point (either as the center or within the window), and the predicted probability of that volume element for each covered window is denoted as:

[0106] (15)

[0107] in It is the number of times the window is covered. In the first In each window, the volume element The probability of being predicted as positive (mineralized).

[0108] A distance-weighted averaging strategy is used to combine these probabilities into a final mineralization probability. The fusion process effectively reduces the impact of local noise in single-window predictions, improving the consistency and stability of 3D spatial predictions. For the... Volumetrics repeatedly predicted in a window Using weights Predicting windows Weighted average in logit space:

[0109] (16)

[0110] in, For the final probability, Let j be the weight of the j-th sliding window. This is for predicting the mineralization probability of window j for volume element v. Weights It can be determined by both the distance from the window center to the voxel and the model confidence level:

[0111] (17)

[0112] Wherein, the center coordinates of the window are Body element The coordinates are , The spatial distance from the center of the window to the volume element. To control the Gaussian bandwidth of the distance decay rate, the fusion probabilities of all voxels are ultimately combined into a three-dimensional probability volume.

[0113] Step 6: Input the ore-controlling indicators corresponding to the volume elements of the area to be predicted into the three-dimensional mineralization probability volume, and delineate the target area for predicting the hidden ore body according to the preset probability threshold.

[0114] In practice, after estimating the model parameters, for a certain volume element in the prediction region... Given its mineral control indicators Based on the existing prediction model, Algorithm 1 is used to solve for its corresponding mineralization index.

[0115] ① The algorithm uses the ore-controlling feature vector of the spatial volume element. As input, after linear feature projection, it enters a depth-selective state update structure to model the nonlinear, long-sequence dependencies between ore-controlling factors.

[0116] ② The initial hidden state is obtained by linear transformation of the input features. This step is equivalent to constructing a representation of the ore-controlling features in a high-dimensional space, making it more suitable for the sequence modeling mechanism of the subsequent state update module. Then, the model in the... In each depth layer, based on the hidden state of the previous layer Adaptive generation of three sets of selective weights It is used to control the degree of state retention, the intensity of input information absorption, and the adjustment of output salience.

[0117] ③ Model execution state update: previous time step state A new state is formed by the interaction of selective weights and the current input driving term. This enables the dynamic integration of multi-scale, long-dependency information from ore-controlling characteristics. This is achieved through output channel weights. Mapping the internal state to a new implicit representation This provides input for the next layer.

[0118] ④ In all After the layer iterations are completed, the final hidden state will be input into the classification head to output the predicted mineralization probability value. .

[0119] The method of this disclosure embodiment can be used to control the ore-controlling characteristics of spatial volume elements. Given, the predicted probability of mineralization To predict the probability of mineralization under unknown conditions. The prediction.

[0120] The deep learning-based three-dimensional quantitative prediction method for concealed ore bodies provided in this embodiment captures the correlation between mineralization information indicators and mineralization distribution by utilizing a structured state-space model (SSM). This method not only possesses the ability of deep learning models to handle nonlinear relationships but also avoids the computational cost problem that occurs with Transformer-type models under deep and dense volume data. It can directly learn deep-level ore-controlling characteristics and prospecting indicators from the three-dimensional structural data of the ore deposit, efficiently achieving the fusion and expression of deep mineralization characteristics. This provides a more suitable model structure for the prediction of deep concealed ore bodies, improving prediction accuracy and interpretability.

[0121] The method of the present invention will be further described below with reference to a specific embodiment, taking the three-dimensional quantitative prediction of concealed ore bodies in the Jinchuan copper-nickel sulfide deposit as an example. Based on the geological background and metallogenic conceptual model of the Jinchuan deposit, ore-forming structural morphology and magma channel-related indicators are selected as the ore-controlling factors of the three-dimensional metallogenic prediction model. In this example, as shown in Table 1,

[0122] Table 1

[0123]

[0124] The ore-controlling factors are: (1) dF—fault distance field; (2) Ratio—relative distance ratio between the top and bottom plates; (3) toFloordis—absolute distance to the bottom plate; (4) toTopdis—absolute distance to the top plate; (5) WaFtrend—first-order undulation of the bottom plate; (6) WaTtrend—first-order undulation of the top plate; (7) dMC_flux—magma channel benefit; a total of 7 ore-controlling indicators are used to obtain the basic dataset of feature indicators for model training. The method of this invention is used to realize the three-dimensional prediction modeling of concealed ore bodies based on Mamba. The specific implementation steps are described as follows:

[0125] Step 1: Transfer the ore control information of the original 3D structure to the set window length. and step length Construct a sliding window sequence and uniformly convert it into a continuous sequence input to form sequence samples. .

[0126] Step 2: By analyzing the ore-controlling features at each position in the sliding window sequence Linear transformations and nonlinear activation function calculations are performed to map the original feature sequence to a high-dimensional representation space, resulting in the latent representation as shown in the formula.

[0127] Step 3: Given the known set of ore-controlling indices dF, Ratio, toFloordis, toTopdis, WaFtrend, WaTtrend, dMC_flux, and the known mineralization index Iore, a Mamba model is constructed and the optimal parameters are solved to accurately express the deep relationship and geological characteristics between ore-controlling factors and mineralization indices. The model training parameters are shown in Table 2, and the loss curves of the training and validation sets during model training are shown in Table 2. Figure 2 As shown, the ROC curve of the model training is as follows: Figure 3 As shown.

[0128] Table 2

[0129]

[0130] Step 4: Based on the Mamba prediction model trained on the known area data and the ore-controlling factors dF, Ratio, toFloordis, toTopdis, WaFtrend, WaTtrend, and dMC_flux in the unknown area, use the Mamba prediction model to predict the ore-bearing properties of the unknown area. The prediction results are as follows: Figure 4 and Figure 5 As shown, it enables continuous quantitative prediction of concealed ore bodies in three-dimensional space, and can automatically delineate target areas according to set thresholds, thus providing a scientific basis for deep exploration layout.

[0131] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof.

[0132] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for three-dimensional quantitative prediction of concealed ore bodies using deep learning, characterized in that, The application relates to a method for predicting a concealed ore body, and belongs to the technical field of mineral resources prediction. Step 1, constructing a three-dimensional geological body ore-controlling index sequence sliding window to obtain a sequence sample; Step 2, mapping the original ore-controlling characteristics in the sequence sample to a high-dimensional representation space to obtain an initial characteristic sequence; The step 2 specifically comprises: each ore-controlling feature in the local feature sequence After linear transformation, a nonlinear activation function is used for processing to obtain the corresponding initial latent representation: wherein, represents an initial latent representation, is an input weight matrix, is an input bias vector, is a ReLU function, is a tensor shape, is a model dimension; Step 3, inputting the initial characteristic sequence into a selective state space model to perform dependence modeling and feature extraction through stacked Mamba modules, and outputting a deep feature sequence; The step 3 specifically comprises: Step 3.1, inputting the initial characteristic sequence into a model composed of N sequentially connected Mamba modules; Step 3.2, each Mamba module updates its hidden state by selectively gating , the state update matrix , and the control term for each sequence position using a selective discrete SSM update, resulting in a hidden state and outputs the state as , forming a deep representation, where the selective discrete SSM update is expressed as: wherein, denotes a step parameter, denotes a state transition matrix, is an input matrix, is an output matrix, is a discretized A matrix, is a discretized B matrix, is a state dimension, is a gating bias, is a hidden state at time t, is a hidden state at time t+1, is an element-wise multiplication, the hidden state is updated point by point in the sequence direction and the state is output as , forming a deep representation , is adaptively generated by a lightweight multi-layer perceptron according to the current input features . Step 3.3, stacking the deep expressions of each Mamba module to form a deep feature sequence; Step 4, calculating a mineralization probability through a classification network based on the characteristics representing the window center position in the deep feature sequence, and training the selective state space model through a sample weighting strategy to optimize the model parameters; Step 5, applying the trained model to all the body elements, spatially weighting and fusing multiple prediction probabilities of the same body element obtained from different sliding windows to obtain the final mineralization probability of the body element, and constructing a three-dimensional mineralization probability body; Step 6, inputting the ore-controlling index corresponding to the body element of a to-be-predicted area into the three-dimensional mineralization probability body, and delineating a concealed ore body prediction target area according to a preset probability threshold.

2. The method of claim 1, wherein, The step 1 specifically comprises: Defining window length on a three-dimensional voxel point set and step size for each center voxel constructing local feature sequence: wherein, is a vector of body elements controlling the characteristics of the mine, denotes the transpose of the vector, the sliding window generates a serialized sample and labels the center body element , is the dimension of the real matrix, L is the sequence length, and F is the feature dimension.

3. The method of claim 2, wherein, The step 4 specifically comprises: Step 4.1, extracting the representation corresponding to the window center position in the deep feature sequence and inputting the representation into a classification head comprising at least one fully connected layer to obtain a mineralization probability prediction value; Step 4.2, training the selective state space model using a weighted binary cross-entropy loss as a loss function according to the mineralization probability prediction value to optimize the model parameters, wherein the expression of the weighted binary cross-entropy loss is: wherein, is a loss value, is a true label is a class weight, is a prediction probability for the th point of the th sequence.

4. The method of claim 3, wherein, The step 5 specifically comprises: Step 5.1, applying the trained model to all the body elements; Step 5.2, for each cell , collect the set of prediction probabilities it gets in different sliding windows: ； wherein is the number of window covers, is the probability that the voxel is predicted as positive class in the th window. Step 5.3, using weights On the set of prediction probabilities In the logit space, a weighted average is made to obtain a voxel Final ore-forming probability; Step 5.4, integrating the final mineralization probabilities of all the body elements to construct a three-dimensional mineralization probability body.

Images