Four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering

Abstract

The invention discloses a four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering, and the method comprises the steps: obtaining initial multi-parameter three-dimensional geological model data and real-time exploration data, carrying out the three-dimensional voxelization processing, and constructing a four-dimensional geological data volume through a time sliding window technology; the four-dimensional geological data volume is input into a lightweight deep learning network for spatio-temporal feature extraction and mineral product result prediction, a preliminary prediction result is obtained, and the lightweight deep learning network is one of a lightweight three-dimensional convolutional neural network, a lightweight long-short term neural network and a lightweight frequency-space neural network; and inputting the preliminary prediction result as an observation value into a time sequence fusion Kalman filter, fusing the state vector of the historical time step and the observation value of the current time step through state prediction, Kalman gain calculation and state correction processes, generating a four-dimensional prediction result, and updating the three-dimensional geologic model. According to the invention, the real-time performance and accuracy of four-dimensional mineral resource prediction evaluation can be improved.

Description

Technical Field

[0001] This application relates to the field of geoscience big data mining technology, and involves, but is not limited to, a four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering. Background Technology

[0002] Mineral resource prediction and evaluation is an important method for mineral exploration. Classic mineral resource prediction and evaluation methods mainly include the evidence-based mineral prediction and evaluation method proposed by Agterberg and Bonham-Carter of the Geological Survey of Canada, the "three-step" mineral prediction and evaluation method proposed by Singer et al. of the U.S. Geological Survey, the "three-link" mineral prediction and evaluation method proposed by Academician Zhao Pengda of my country, the "comprehensive information" mineral prediction and evaluation method proposed by Professor Wang Shicheng of my country, and the "nonlinear" mineral prediction and evaluation method proposed by Academician Cheng Qiuming of my country.

[0003] In existing technologies, mineral resource prediction methods combining big data and artificial intelligence are widely used in the field of mineral exploration, such as the proprietary big data prospecting tool—Efficacy of Information (EOI)—developed by KoBold Metals in the United States. Due to the multi-stage nature of mineral resource exploration and the real-time nature of mine production, mineral resource prediction work not only needs to determine the three-dimensional morphology and spatial location of deep prospecting target areas, but also needs to dynamically update target area information in conjunction with a fourth dimension—time—to better adapt to the actual application needs of industry. However, existing mineral resource prediction methods mostly focus on static prediction elements and lack a fourth dynamic element, resulting in weak practical applicability of the prediction models. Furthermore, existing deep learning-based prediction methods have a large computational load, making it difficult to meet the real-time requirements of the mine production environment.

[0004] Therefore, there is an urgent need for a more comprehensive prediction and evaluation method that integrates lightweight deep learning and Kalman filtering to solve the problems of existing methods, such as the difficulty in dynamically updating the prediction model and poor real-time performance. This would significantly improve the accuracy and timeliness of four-dimensional mineral resource prediction and evaluation, and meet the real-time needs of dynamic updating of multi-stage mineral resource exploration target areas and the integration of mining exploration and extraction. Summary of the Invention

[0005] This application provides a four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering.

[0006] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering. The method includes: acquiring initial multi-parameter three-dimensional geological model data and real-time exploration data; performing three-dimensional voxelization processing on the initial multi-parameter three-dimensional geological model data and the real-time exploration data; constructing a four-dimensional geological data volume using a time sliding window technique; constructing a lightweight deep learning network; inputting the four-dimensional geological data volume into the lightweight deep learning network for spatiotemporal joint feature extraction and mineral result prediction to obtain preliminary prediction results for each time step; wherein the lightweight deep learning network is one of a lightweight three-dimensional convolutional neural network, a lightweight long short-term neural network, and a lightweight frequency-space neural network; inputting the preliminary prediction results for each time step as observation values ​​into a temporal fusion Kalman filter; fusing the state vectors of historical time steps with the observation values ​​of the current time step through state prediction, Kalman gain calculation, and state correction processes to generate four-dimensional prediction results; and updating the three-dimensional geological model based on the four-dimensional prediction results.

[0007] The technical solution provided in this application acquires initial multi-parameter 3D geological model data and real-time exploration data. It then performs 3D voxelization processing on the initial multi-parameter 3D geological model data and real-time exploration data, and constructs a 4D geological data volume using a time sliding window technique, providing a reliable data foundation for subsequent prediction processes. A lightweight deep learning network is constructed, and the 4D geological data volume is input into the lightweight deep learning network for spatiotemporal joint feature extraction and mineral outcome prediction, obtaining preliminary prediction results at each time step. The lightweight deep learning network comprises a lightweight 3D convolutional neural network, a lightweight long short-term neural network, and a lightweight frequency-spatial neural network. One type of network achieves a deep separable and grouped design of the network structure, significantly reducing the number of parameters and computational load while ensuring predictive feature extraction capabilities. This effectively improves the efficiency of 3D spatial and temporal feature modeling, achieving a balance between high accuracy and high real-time performance. It is suitable for mineral resource prediction scenarios with large data volumes and complex data structures in actual production processes. Specifically, the lightweight 3D convolutional neural network structure employs 3D deep separable convolutions and lightweight fully connected layers to achieve efficient and lightweight spatial feature extraction. In the lightweight long short-term neural network, a dual-gate structure and deep separable linear layers reduce gating and parameter redundancy, improving... Efficiency of temporal feature extraction: In a lightweight frequency-space neural network, a frequency-space joint attention mechanism, combined with fast Fourier transform and spatial gating, is used to achieve multi-scale feature fusion in the frequency and spatial domains, balancing global perception and local enhancement, thereby achieving efficient parameter compression and performance balance in multi-dimensional feature extraction. The preliminary prediction results of each time step are used as observations and input into the temporal fusion Kalman filter. Through state prediction, Kalman gain calculation, and state correction processes, the state vectors of historical time steps are fused with the observations of the current time step to generate a four-dimensional prediction result. The three-dimensional geological model is then updated based on the four-dimensional prediction result. This model combines lightweight deep learning network output with Kalman filtering to suppress and correct noise and uncertainty, enabling dynamic fusion of temporal information and spatial features. This enhances the continuity and stability of prediction results in the time dimension, and maps and updates the prediction state of each time step with the 3D geological model in real time, significantly improving the continuity and spatial consistency of the model in the time dimension. Ultimately, it achieves dynamic correction of prediction results and dynamic real-time updating of the 3D geological model, completing four-dimensional real-time mineral resource prediction and dynamic reconstruction of geological bodies. It is suitable for the real-time needs of dynamic updating of multi-stage mineral resource exploration target areas and the integration of mining exploration and extraction.

[0008] Optionally, the initial multi-parameter three-dimensional geological model data includes spatial coordinates, lithological properties, geophysical properties, and geochemical properties; the real-time exploration data includes borehole data, geological profile data, radar point cloud data, geophysical data, and core spectral scanning data. The step of performing three-dimensional voxelization processing on the initial multi-parameter three-dimensional geological model data and the real-time exploration data, and constructing a four-dimensional geological data volume using time-sliding window technology, includes: uniformly converting the format of the initial multi-parameter three-dimensional geological model data and the real-time exploration data and mapping them to a preset three-dimensional voxel grid to generate a three-dimensional geological data volume with a unified spatial reference system; assigning time labels to the three-dimensional geological data volume, and integrating the three-dimensional geological data volume according to a time sequence based on time-sliding window technology to generate the four-dimensional geological data volume.

[0009] Optionally, the lightweight 3D convolutional neural network includes lightweight convolutional layers, lightweight fully connected layers, and a lightweight convolutional attention mechanism layer. When the lightweight deep learning network is the lightweight 3D convolutional neural network, the step of inputting the 4D geological data volume into the lightweight deep learning network for spatiotemporal joint feature extraction and mineral outcome prediction to obtain preliminary prediction results at each time step includes: inputting the 4D geological data volume into the lightweight convolutional layer, performing spatiotemporal joint feature extraction through 3D depthwise separable convolution to obtain a preliminary 3D feature map; and inputting the preliminary 3D feature map into the lightweight convolutional attention mechanism layer. In the control layer, lightweight convolution is used to compress the channel dimension to generate a low-dimensional feature map; lightweight fully connected processing is applied to the low-dimensional feature map to generate attention weights for each feature channel; the features of each channel of the preliminary 3D feature map are dynamically recalibrated according to the attention weights; the dimensions of the recalibrated features are restored by lightweight convolution to obtain an attention-weighted 3D feature map; the attention-weighted 3D feature map is input into the lightweight fully connected layer, and the feature channels are grouped by group convolution; after convolution operation is performed on each group of feature channels, the feature channels are fused to obtain the preliminary prediction result.

[0010] Optionally, the step of extracting spatiotemporal joint features through three-dimensional depth-separable convolution to obtain a preliminary three-dimensional feature map includes: performing three-dimensional convolution operations on each feature channel of the four-dimensional geological data volume using multiple independent three-dimensional convolution kernels to generate a depth-wise feature map corresponding to each feature channel; and performing cross-channel linear combination of the depth-wise feature maps corresponding to all feature channels using a 1×1×1 convolution kernel to obtain a preliminary three-dimensional feature map.

[0011] Optionally, the lightweight long short-term neural network includes a depthwise separable linear layer, a simplified gating mechanism, and a lightweight channel scaling layer. The simplified gating mechanism includes an update gate and a candidate gate. When the lightweight deep learning network is the lightweight long short-term neural network, the step of inputting the four-dimensional geological data volume into the lightweight deep learning network for spatiotemporal joint feature extraction and mineral result prediction to obtain preliminary prediction results for each time step includes: expanding the four-dimensional geological data volume into a three-dimensional spatial feature sequence according to the time step, inputting the three-dimensional spatial features of the current time step into the depthwise separable linear layer, and independently performing a linear transformation on each feature channel of the three-dimensional spatial features to generate a channel-wise feature representation. The channel-wise feature representations are linearly fused across channels using a small-scale point linear layer to generate a fused feature representation for the current time step. This fused feature representation, along with the hidden state from the previous time step, is input into the simplified gating mechanism. The update gate activation value and the candidate gate activation value are calculated using the update gate and the candidate gate, respectively. Based on these activation values, the cell state and hidden state at the current time step are updated. The updated hidden state is then input into the lightweight channel scaling layer. Channel scaling coefficients are generated using a low-dimensional intermediate mapping or a small fully connected layer. Based on these scaling coefficients, each channel of the hidden state is dynamically weighted to generate a preliminary prediction result for the current time step.

[0012] Optionally, when the lightweight deep learning network is the lightweight frequency-space neural network, the step of inputting the four-dimensional geological data volume into the lightweight deep learning network for spatiotemporal joint feature extraction and mineral result prediction to obtain preliminary prediction results at each time step includes: transforming the four-dimensional geological data volume into the frequency domain through fast Fourier transform to obtain amplitude and phase information; weighting the amplitude information using Sigmoid and GELU activation functions to generate frequency-gated weights; filtering the frequency domain information based on the frequency-gated weights through element-wise multiplication; performing deep feature fusion on the filtered amplitude and phase information using a bottleneck architecture to obtain a fused frequency signal; transforming the fused frequency signal into the spatial domain through inverse fast Fourier transform to obtain frequency-enhanced global features; inputting the global features into a spatial gating mechanism; generating a spatial weight map through convolution operations; and performing element-wise spatial feature weighting on the global features according to the spatial weight map to strengthen local key information and obtain the preliminary prediction results.

[0013] Optionally, the step of inputting the preliminary prediction results of each time step as observations into a temporal fusion Kalman filter, fusing the state vectors of historical time steps with the observations of the current time step through state prediction, Kalman gain calculation, and state correction processes to generate a four-dimensional prediction result, and updating the three-dimensional geological model based on the four-dimensional prediction result, includes: inputting the preliminary prediction results of each time step as observations into a temporal fusion Kalman filter; mapping the preliminary prediction results of the initial time step to the initial state vector and initial state covariance matrix of the Kalman filter; for each subsequent time step, predicting the prior state estimate of the current time step based on the state vector and state transition model of the previous time step; and based on the state covariance matrix of the previous time step, The state transition model and process noise covariance matrix are used to calculate the prior covariance estimate for the current time step. Based on the prior covariance estimate, observation model, and observation noise covariance matrix, the Kalman gain for the current time step is calculated. The preliminary prediction result for the current time step is used as the observation value, and the prior state estimate for the current time step is corrected by the Kalman gain to obtain the posterior state estimate for the current time step. The posterior state estimate is then mapped to the four-dimensional prediction result. The prior covariance estimate is updated by the Kalman gain to obtain the posterior covariance matrix for the current time step. The posterior state estimate is mapped to the corresponding position in the three-dimensional geological space, and the three-dimensional geological model is iteratively updated by combining geological constraints and mineral attribute information.

[0014] Secondly, embodiments of this application provide an electronic device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the steps in the above-mentioned four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering.

[0015] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the above-described four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering.

[0016] The beneficial effects of the technical solutions provided in this application include at least the following: This application provides a four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering. It acquires initial multi-parameter three-dimensional geological model data and real-time exploration data, performs three-dimensional voxelization on these data, and constructs a four-dimensional geological data volume using a time-sliding window technique, providing a reliable data foundation for subsequent prediction processes. A lightweight deep learning network is constructed, and the four-dimensional geological data volume is input into this network for spatiotemporal joint feature extraction and mineral result prediction, yielding preliminary prediction results at each time step. The lightweight deep learning network consists of a lightweight three-dimensional convolutional neural network and a lightweight long-range... This is a type of short-time neural network and lightweight frequency-space neural network. It achieves deep separability and grouping of the network structure, significantly reducing the number of parameters and computational load while maintaining predictive feature extraction capabilities. This effectively improves the efficiency of 3D spatial and temporal feature modeling, achieving a balance between high accuracy and high real-time performance. It is suitable for mineral resource prediction scenarios with large data volumes and complex data structures in actual production processes. Specifically, the lightweight 3D convolutional neural network structure utilizes 3D deep separable convolutions and lightweight fully connected layers to achieve efficient and lightweight spatial feature extraction. The lightweight long-short-time neural network uses a dual-gate structure and deep separable linear layers to reduce... Reduced gating and parameter redundancy improve the efficiency of temporal feature extraction. In a lightweight frequency-space neural network, a frequency-space joint attention mechanism, combined with Fast Fourier Transform and spatial gating, achieves multi-scale feature fusion in the frequency and spatial domains, balancing global perception and local enhancement, thus achieving efficient parameter compression and performance balance in multi-dimensional feature extraction. The preliminary prediction results at each time step are used as observations and input into the temporal fusion Kalman filter. Through state prediction, Kalman gain calculation, and state correction processes, the state vectors of historical time steps are fused with the observations of the current time step to generate a four-dimensional prediction result. Further optimization is then performed based on the four-dimensional prediction result. The new 3D geological model combines lightweight deep learning network output with Kalman filtering to suppress and correct noise and uncertainty, thereby achieving dynamic fusion of temporal information and spatial features. This enhances the continuity and stability of prediction results in the time dimension, and maps and updates the prediction state of each time step with the 3D geological model in real time, significantly improving the continuity and spatial consistency of the model in the time dimension. Ultimately, it realizes dynamic correction of prediction results and dynamic real-time updating of the 3D geological model, completing four-dimensional real-time mineral resource prediction and dynamic reconstruction of geological bodies. It is suitable for the real-time needs of dynamic updating of multi-stage mineral resource exploration target areas and the integration of mining exploration and extraction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A flowchart illustrating the four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering provided in this application embodiment; Figure 2 A schematic diagram of a technical route for a four-dimensional mineral resource prediction and evaluation process provided in this application embodiment; Figure 3 A schematic diagram of a software interface for selecting a lightweight deep learning network is provided for an embodiment of this application; Figure 4 A schematic diagram of a study area prediction result provided in an embodiment of this application; Figure 5 This is a schematic diagram of the hardware entity of an electronic device provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0020] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0021] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0022] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0023] In view of the current problems in the research on four-dimensional mineral resource prediction and evaluation in the field of geoscience big data mining technology, this application provides a four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering.

[0024] The technical solution of this application is described below, starting with the method embodiments.

[0025] Please refer to Figure 1 It illustrates a flowchart of the four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering provided in the embodiments of this application, such as... Figure 1 As shown, the method includes at least the following steps S110 to S130.

[0026] Step S110: Obtain initial multi-parameter three-dimensional geological model data and real-time exploration data, perform three-dimensional voxelization processing on the initial multi-parameter three-dimensional geological model data and the real-time exploration data, and construct a four-dimensional geological data volume through time sliding window technology.

[0027] In the embodiments of this application, please refer to Figure 2This document illustrates a technical route diagram of a four-dimensional mineral resource prediction and evaluation process provided in an embodiment of this application. It involves acquiring initial multi-parameter three-dimensional geological model data and importing it as basic model data. This initial multi-parameter three-dimensional geological model data includes spatial coordinates, lithological properties, geophysical properties, and geochemical properties, and all data formats are uniformly converted to .csv, .dat, and .txt formats. It also acquires various exploration data collected during real-time production, i.e., real-time exploration data, including borehole data, geological profile data, radar point cloud data, core spectral scanning data, and geophysical data. Specifically, borehole data formats are uniformly converted to .csv, .dat, and .txt; geological profile data formats are uniformly converted to .dxf and .tif; radar point cloud data formats are uniformly converted to .LAS, .PCD, .csv, .dat, and .txt; geophysical data formats are uniformly converted to .csv, .dat, .txt, and .tif; and core spectral scanning data formats are uniformly converted to .csv, .dat, and .txt. Furthermore, the initial multi-parameter 3D geological model data and real-time exploration data are processed into 3D voxels, and a 4D geological data volume is constructed using a time-sliding window technique. Specifically, the initial multi-parameter 3D geological model data and real-time exploration data are mapped onto a preset 3D voxel grid to generate a 3D geological data volume with a unified spatial reference system. A time label is assigned to this 3D geological data volume with a unified spatial reference system, and based on the time-sliding window technique, the 3D geological data volume is integrated according to the time sequence to generate a 4D geological data volume. The data in this 4D geological data volume includes four dimensions: X-coordinate, Y-coordinate, Z-coordinate, and T-time.

[0028] Step S120: Construct a lightweight deep learning network and input the four-dimensional geological data volume into the lightweight deep learning network for spatiotemporal joint feature extraction and mineral result prediction to obtain preliminary prediction results at each time step. The lightweight deep learning network is one of a lightweight three-dimensional convolutional neural network, a lightweight long short-term neural network, and a lightweight frequency-space neural network.

[0029] In this embodiment, an improvement is made to the traditional deep learning network to construct a lightweight deep learning network for spatiotemporal joint feature extraction and mineral outcome prediction of four-dimensional geological data volumes, obtaining preliminary prediction results at each time step. In practical applications, one or more lightweight deep learning networks can be flexibly selected and used for processing based on specific data characteristics and prediction objectives. Please refer to... Figure 3 It shows a schematic diagram of a software interface for lightweight deep learning network selection provided in an embodiment of this application. Figure 3 The lightweight frequency-spatial neural network was selected as the example.

[0030] In one specific embodiment, a lightweight 3D convolutional neural network is obtained by making lightweight improvements to the 3D convolutional neural network. This lightweight 3D convolutional neural network includes lightweight convolutional layers, lightweight fully connected layers, and lightweight convolutional attention mechanism layers. When the temporal sequence of the 4D geological data volume in the application scenario is not strong, but the spatial structure is relatively complex, the lightweight deep learning network is chosen to be the lightweight 3D convolutional neural network. The 4D geological data volume is input into the lightweight 3D convolutional neural network for spatiotemporal joint feature extraction and mineral result prediction, and preliminary prediction results are obtained at each time step. Specifically, the lightweight convolutional layer employs 3D depthwise separable convolution. The 4D geological data volume is input into this lightweight convolutional layer, and spatiotemporal joint feature extraction is performed through 3D depthwise separable convolution. This spatiotemporal joint feature extraction process mainly includes two stages of transformation. The first stage is channel-wise convolution, where multiple independent 3D convolution kernels are used to independently perform 3D convolution operations on each feature channel of the 4D geological data volume, generating a depthwise feature map corresponding to each feature channel, thus fully preserving local spatial and channel features. The second stage is pointwise convolution, where a 1×1×1 convolution kernel is used to linearly combine the depthwise feature maps corresponding to all feature channels across channels to obtain a preliminary 3D feature map, achieving inter-channel information fusion. This reduces the number of parameters while maintaining the network's feature representation capability. Furthermore, this 3D depthwise separable convolution can be adjusted by parameters such as kernel size, stride, number of channels, and number of groups to adapt to different computational resource constraints and feature extraction requirements in various scenarios.

[0031] The lightweight convolutional attention mechanism layer, composed of lightweight convolutional layers and lightweight fully connected layers, enhances the network's ability to perceive key regions and features. A preliminary 3D feature map is input into this layer, where lightweight convolution compresses the channel dimensions to generate a low-dimensional feature map. Further lightweight fully connected processing is applied to this low-dimensional feature map to generate attention weights for each feature channel. Based on these attention weights, the features of each channel in the preliminary 3D feature map are dynamically recalibrated. Finally, lightweight convolution restores the dimensions of the recalibrated features, resulting in an attention-weighted 3D feature map. Furthermore, by adjusting parameters such as the lightweight convolutional layer parameters, lightweight fully connected layer parameters, intermediate dimension compression rate, and kernel size, a balance can be achieved between computational overhead and feature enhancement effectiveness. The lightweight fully connected layer is implemented through grouped convolutions. Attention-weighted 3D feature maps are input into the lightweight fully connected layer, and the feature channels are divided into several groups through grouped convolutions. After performing convolution operations on each group of feature channels, the feature channels are fused to obtain a preliminary prediction result. This significantly reduces weight parameters and computational cost while maintaining feature mixing capabilities. Furthermore, the grouped convolutions can be lightweighted to varying degrees by adjusting the number of groups, the number of output channels, and the kernel size. The lightweight 3D convolutional neural network provided in this application embodiment achieves efficient parameter compression and computational optimization while ensuring 3D feature extraction and 3D feature modeling capabilities, making it particularly suitable for 4D real-time mineral resource prediction.

[0032] In one specific embodiment, a lightweight improvement is made to the Long Short-Term Neural Network (LSTN) to obtain a lightweight LSTN. This lightweight LSTN includes a depthwise separable linear layer, a simplified gating mechanism, and a lightweight channel scaling layer. The simplified gating mechanism is determined by simplifying the input gate, forget gate, output gate, and candidate gate in the traditional LSTN, retaining only two core gate structures. That is, the simplified gating mechanism only includes the update gate and the candidate gate, thus achieving a lightweight network structure. When the four-dimensional geological data volume in the application scenario has a long time series, the lightweight deep learning network chosen is the lightweight LSTN. The four-dimensional geological data volume is input into the lightweight LSTN for spatiotemporal joint feature extraction and mineral result prediction, obtaining preliminary prediction results for each time step. Specifically, a depthwise separable linear layer replaces the standard fully connected layer. This depthwise separable linear layer borrows the design concept of depthwise separable convolution in convolutional neural networks, improving the traditional linear transformation into a lightweight linear transformation. This lightweight linear transformation mainly includes two stages: channel-wise linear mapping and channel fusion mapping. The four-dimensional geological data volume is unfolded into a three-dimensional spatial feature sequence according to time steps. The three-dimensional spatial features of the current time step are input into a depth-separable linear layer. Each feature channel of the three-dimensional spatial features is independently linearly transformed to generate a channel-wise feature representation, which effectively preserves the local features of each channel. The channel-wise feature representations are linearly fused across channels through a small-scale point linear layer to generate a fused feature representation for the current time step. This significantly reduces the number of parameters and computational load while maintaining the feature expressive power. Compared with traditional fully connected layers, this depth-separable linear layer achieves a significant reduction in the number of parameters under the same input and output dimensions, making the feature transformation between the input and hidden states more efficient.

[0033] Furthermore, the fused feature representation of the current time step and the hidden state of the previous time step are input into a simplified gating mechanism. The update gate activation value and the candidate gate activation value are calculated through the update gate and candidate gate, respectively. Then, the cell state and hidden state at the current time step are updated based on the update gate activation value and the candidate gate activation value. To enhance the network's adaptability to temporal features, a lightweight channel scaling layer is introduced in the network output stage. The updated hidden state is input into this lightweight channel scaling layer, and channel scaling coefficients are generated through low-dimensional intermediate mapping or small fully connected layers. Based on these channel scaling coefficients, each channel of the hidden state is dynamically weighted to highlight key feature channels and suppress redundant information. This significantly reduces the number of parameters and computational overhead while improving the network's sensitivity to and expressive ability of temporal information, ultimately generating preliminary prediction results. The lightweight long short-term neural network provided in this application embodiment achieves efficient parameter compression and computational optimization while ensuring feature modeling capabilities, making it particularly suitable for four-dimensional real-time mineral resource prediction.

[0034] In one specific embodiment, a lightweight improvement is made to the frequency-spatial neural network to obtain a lightweight frequency-spatial neural network. When the data volume in the application scenario is large, the lightweight deep learning network is chosen as the lightweight frequency-spatial neural network. The four-dimensional geological data volume is input into the lightweight frequency-spatial neural network for spatiotemporal joint feature extraction and mineral outcome prediction, obtaining preliminary prediction results at each time step. Specifically, the lightweight frequency-spatial neural network achieves a lightweight network structure through a frequency-spatial joint attention mechanism, and its global modeling capability is close to that of the Transformer network based on a multi-head self-attention mechanism. In the frequency domain, the real and imaginary parts of the input data are extracted using Fast Fourier Transform (FFT). This transforms the four-dimensional geological data volume into the frequency domain, obtaining amplitude and phase information. The amplitude information is then weighted using Sigmoid and GELU activation functions to generate frequency-gated weights. Further, the frequency domain information is filtered based on these weights through element-wise multiplication. A bottleneck architecture is then used to further mine the deeper features of the filtered amplitude and phase information and perform deep feature fusion to obtain a fused frequency signal, achieving effective frequency feature fusion. After feature fusion, the fused frequency signal is transformed into the spatial domain using Inverse Fast Fourier Transform (IFFT) to obtain the global spatial domain information after frequency domain processing, i.e., the frequency-enhanced global features. In the spatial domain, a spatial gating mechanism is used to acquire locally important information and weight spatial features. Specifically, based on the spatial gating mechanism, global features are generated into a spatial weight map through convolution operations. Then, element-wise spatial feature weighting is applied to the global features according to the spatial weight map, strengthening key local information. The frequency-spatial joint attention mechanism achieves effective fusion of global and local features through joint modeling of frequency and spatial domain features. This reduces network parameters and computational load while enhancing sensitivity to key features, ultimately yielding preliminary prediction results. The lightweight frequency-spatial neural network provided in this application achieves efficient parameter compression and computational optimization while ensuring global modeling capabilities, making it particularly suitable for four-dimensional real-time mineral resource prediction.

[0035] Step S130: The preliminary prediction results of each time step are input as observation values ​​into the time-series fusion Kalman filter. The state vectors of historical time steps and the observation values ​​of the current time step are fused through state prediction, Kalman gain calculation and state correction process to generate four-dimensional prediction results. The three-dimensional geological model is then updated based on the four-dimensional prediction results.

[0036] In this embodiment, a temporal fusion Kalman filter is used to achieve real-time four-dimensional updates of the 3D geological model and prediction results based on a lightweight deep learning network. The preliminary prediction results at each time step are input as observations into the temporal fusion Kalman filter. Through state prediction, covariance update, Kalman gain calculation, and state correction processes, the state vectors and state covariances of historical time steps are fused with the observations of the current time step to generate four-dimensional prediction results. The 3D geological model is then updated based on these four-dimensional prediction results. Specifically, the preliminary prediction results at each time step are input as observations into the temporal fusion Kalman filter, mapping the preliminary prediction results of the initial time step to the initial state vector and initial state covariance matrix of the Kalman filter. For each subsequent time step, the prior state estimate of the current time step is predicted based on the state vector and state transition model of the previous time step, completing state prediction. The prior covariance estimate of the current time step is calculated based on the state covariance matrix, state transition model, and process noise covariance matrix of the previous time step, completing covariance prediction. The Kalman gain of the current time step is calculated based on the prior covariance estimate, observation model, and observation noise covariance matrix, completing Kalman gain calculation. The preliminary prediction result of the current time step is used as the observation value, and the prior state estimate of the current time step is corrected by the Kalman gain to obtain the posterior state estimate of the current time step. The posterior state estimate is then mapped to a four-dimensional prediction result, completing state correction. The prior covariance estimate is updated by the Kalman gain to obtain the posterior covariance matrix of the current time step, completing covariance update. This process achieves the suppression of noise and uncertainty in the time series of the input time-series fusion Kalman filter and the enhancement of key features.

[0037] Furthermore, the three-dimensional geological model is updated based on the fused four-dimensional prediction results. Specifically, the predicted state (i.e., posterior state estimate) at each time step is mapped to the corresponding position in the three-dimensional geological space. Combined with geological constraints and mineral attribute information, the three-dimensional geological model is iteratively updated to correct uncertainties in the original model, enhancing its consistency and accuracy in both spatial and temporal dimensions. This update process reflects the real-time evolution trend of ore body distribution and attributes, achieving dynamic reconstruction of geological bodies over continuous time series. The method provided in this application fuses the output state at each time step with spatial features, feature channel information, and historical time information using Kalman filtering to generate a fused four-dimensional prediction result. While maintaining the efficient feature extraction capabilities of a lightweight network, it achieves dynamic fusion of temporal information and real-time updates of the three-dimensional geological model and prediction results, ultimately completing a four-dimensional real-time mineral resource prediction and evaluation.

[0038] In one specific embodiment, to verify the feasibility of the lightweight deep learning method provided in this application, a three-dimensional mineral resource prediction experiment was conducted in the Wulong-Sidaogou gold mining area with a capacity of thousands of tons in the Liaodong Peninsula. Based on the comprehensive determination of the mineralization model of the Wulong gold mining area by the previous mineral deposit geology, geochemistry, and geophysics of the test area, the metallogenic model of the Wulong gold mining area is an IRGS system with Early Cretaceous granodiorite as its core, controlled by a N-N-W trending fault. The ore body is spatially closely related to diorite and other vein rocks. Therefore, the three-dimensional prediction elements include a three-dimensional buffer model of Early Cretaceous granodiorite, a three-dimensional buffer model of the NE-trending fault, a three-dimensional buffer model of the NW-trending fault, a three-dimensional buffer model of diorite and other vein rocks, and density, magnetic susceptibility, and resistivity models obtained by geophysical inversion. Without activating the temporal fusion Kalman filter module, the accuracy, area under the receiver operating curve (AUC), and F1 score of two mineral resource prediction methods—traditional deep learning and lightweight deep learning—were compared. Specifically, tests were conducted in a 100×100×100m grid with a 13×13×13 window size. Test examples included a traditional 3D convolutional neural network (3 layers of convolution), the lightweight 3D convolutional neural network (3 layers of lightweight convolution) provided in this application, a traditional long short-term neural network (3 layers of gating mechanism), the lightweight long short-term neural network (3 layers of lightweight gating mechanism) provided in this application, a traditional Transformer network (3 layers of feedforward network), and the lightweight frequency-spatial joint neural network (3 layers of frequency-spatial joint attention mechanism) provided in this application. Each model was run under the same training conditions, and the accuracy, AUC, F1 score, and parameter count were calculated on the test set. The test results are shown in Table 1. The test results show that the lightweight deep learning network provided in this application achieves higher prediction accuracy (Accuracy, AUC, F1 score, etc.). In terms of score, it reaches or surpasses that of traditional deep learning networks, and the number of parameters is also significantly reduced. That is, while maintaining prediction accuracy, it can achieve a significant reduction in computational complexity, thus verifying the efficiency and feasibility of the lightweight deep learning method provided in this application.

[0039] Table 1 (Test Results Table) To further verify the dynamic prediction and model updating capabilities of this invention, geological and exploration update data were collected in 2020 and 2025 from the same study area. The lightweight frequency-spatial joint neural network from the previous specific embodiment was used as the prediction backbone, combined with a temporal fusion Kalman filter module for dynamic 3D prediction. In the experiment, an initial 3D orebody probability model was first constructed based on the 2020 data. After data collection in 2025, the new data was input into the Kalman filter module and fused with the 2020 prediction results. The fused results were then updated to generate the 2025 dynamic prediction model. Please refer to [reference needed]. Figure 4 The diagram illustrates a prediction result of a research area provided by an embodiment of this application. The left image shows the prediction result in 2020, and the right image shows the prediction result after the update in 2025. The redder the color bar in the image, the higher the mineralization probability. It can be seen that after time-series fusion correction, the prediction result in 2025 is more refined in the delineated deep mineralization target area, the distribution area of ​​deep ore bodies is more concentrated, and the spatial boundary is clearer. This result shows that the lightweight deep learning and Kalman filter fusion method proposed in this application can fully integrate historical and current prediction information to achieve dynamic updating of three-dimensional geological bodies and high-precision, efficient four-dimensional real-time prediction.

[0040] In summary, the four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering provided in this application acquires initial multi-parameter three-dimensional geological model data and real-time exploration data. The initial multi-parameter three-dimensional geological model data and real-time exploration data are processed into three-dimensional voxels, and a four-dimensional geological data volume is constructed using a time sliding window technique, providing a reliable data foundation for subsequent prediction processes. A lightweight deep learning network is constructed, and the four-dimensional geological data volume is input into the lightweight deep learning network for spatiotemporal joint feature extraction and mineral result prediction, obtaining preliminary prediction results at each time step. The lightweight deep learning network is a lightweight three-dimensional convolutional neural network. This invention relates to three types of neural networks: lightweight long short-term neural networks (LSNs) and lightweight frequency-space neural networks. It achieves depth-separable and grouped network structure design, significantly reducing the number of parameters and computational load while maintaining predictive feature extraction capabilities. This effectively improves the efficiency of 3D spatial and temporal feature modeling, achieving a balance between high accuracy and high real-time performance. It is suitable for mineral resource prediction scenarios with large data volumes and complex data structures in actual production processes. Specifically, the lightweight 3D convolutional neural network structure utilizes 3D depth-separable convolutions and lightweight fully connected layers to achieve efficient and lightweight spatial feature extraction. The lightweight LSN is further enhanced through a dual-gate structure and depth-separable... Linear layers reduce gating and parameter redundancy, improving the efficiency of temporal feature extraction. In a lightweight frequency-space neural network, a frequency-space joint attention mechanism, combined with fast Fourier transform and spatial gating, achieves multi-scale feature fusion in the frequency and spatial domains, balancing global perception and local enhancement, thus achieving efficient parameter compression and performance balance in multi-dimensional feature extraction. The preliminary prediction results at each time step are used as observations and input into a temporal fusion Kalman filter. Through state prediction, Kalman gain calculation, and state correction processes, the state vectors of historical time steps are fused with the observations of the current time step to generate a four-dimensional prediction result. Based on the four-dimensional prediction result... The method updates the 3D geological model by combining the output of a lightweight deep learning network with the Kalman filter method to suppress and correct noise and uncertainty, thereby achieving dynamic fusion of temporal information and spatial features. This enhances the continuity and stability of the prediction results in the time dimension. The prediction state at each time step is mapped and updated in real time with the 3D geological model, significantly improving the continuity and spatial consistency of the model in the time dimension. Ultimately, it realizes dynamic correction of prediction results and dynamic real-time updating of the 3D geological model, completing four-dimensional real-time mineral resource prediction and dynamic reconstruction of geological bodies. It is suitable for the real-time needs of dynamic updating of multi-stage mineral resource exploration target areas and the integration of mining exploration and extraction.

[0041] It should be noted that, in the embodiments of this application, if the above-mentioned four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0042] Correspondingly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program implements the steps in any of the four-dimensional mineral resource prediction and evaluation methods based on deep learning and Kalman filtering described in the above embodiments. Correspondingly, embodiments of this application also provide a computer program product. When executed by a processor of an electronic device, this computer program product is used to implement the steps in any of the four-dimensional mineral resource prediction and evaluation methods based on deep learning and Kalman filtering described in the above embodiments.

[0043] Based on the same technical concept, this application provides an electronic device for implementing the four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering described in the above method embodiments. Figure 5 This is a hardware entity diagram of an electronic device provided in an embodiment of this application, such as... Figure 5 As shown, the electronic device 500 includes a memory 510 and a processor 520. The memory 510 stores a computer program that can run on the processor 520. When the processor 520 executes the program, it implements the steps in any of the four-dimensional mineral resource prediction and evaluation methods based on deep learning and Kalman filtering described in the embodiments of this application.

[0044] The memory 510 is configured to store instructions and applications executable by the processor 520, and can also cache data to be processed or already processed by the processor 520 and various modules in the electronic device (e.g., image data, audio data, voice communication data and video communication data), which can be implemented by flash memory or random access memory (RAM).

[0045] When the processor 520 executes the program, it implements any of the steps of the four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering mentioned above. The processor 520 typically controls the overall operation of the electronic device 500.

[0046] The aforementioned processor can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Controller, Microcontroller, and Microprocessor. It is understood that other electronic devices can also implement the functions of the aforementioned processor, and this application does not specifically limit the specific implementation.

[0047] The aforementioned computer storage media / memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM), etc.; or it can be various electronic devices that include one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0048] It should be noted that the descriptions of the storage medium and device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0049] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0050] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0051] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0052] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0053] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0054] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0055] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0056] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0057] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A four-dimensional mineral resource prediction and evaluation method based on deep learning and Kalman filtering, characterized in that, The method includes: Acquire initial multi-parameter three-dimensional geological model data and real-time exploration data, perform three-dimensional voxelization processing on the initial multi-parameter three-dimensional geological model data and the real-time exploration data, and construct a four-dimensional geological data volume through time sliding window technology; A lightweight deep learning network is constructed, and the four-dimensional geological data volume is input into the lightweight deep learning network for spatiotemporal joint feature extraction and mineral result prediction to obtain preliminary prediction results at each time step. The lightweight deep learning network is one of a lightweight three-dimensional convolutional neural network, a lightweight long short-term neural network, and a lightweight frequency-space neural network. The preliminary prediction results of each time step are input as observation values ​​into the time-series fusion Kalman filter. The state vectors of historical time steps and the observation values ​​of the current time step are fused through state prediction, Kalman gain calculation and state correction processes to generate four-dimensional prediction results. The three-dimensional geological model is then updated based on the four-dimensional prediction results.

2. The method according to claim 1, characterized in that, The initial multi-parameter three-dimensional geological model data includes spatial coordinates, lithological properties, geophysical properties, and geochemical properties. The real-time exploration data includes borehole data, geological profile data, radar point cloud data, geophysical data, and core spectral scanning data. The process of performing three-dimensional voxelization on the initial multi-parameter three-dimensional geological model data and the real-time exploration data, and constructing a four-dimensional geological data volume using a time-sliding window technique, includes: The initial multi-parameter three-dimensional geological model data and the real-time exploration data are uniformly converted into a format and mapped to a preset three-dimensional voxel grid to generate a three-dimensional geological data volume with a unified spatial reference system. The three-dimensional geological data volume is assigned a time label, and based on the time sliding window technique, the three-dimensional geological data volume is integrated according to the time sequence to generate the four-dimensional geological data volume.

3. The method according to claim 1, characterized in that, The lightweight 3D convolutional neural network includes lightweight convolutional layers, lightweight fully connected layers, and a lightweight convolutional attention mechanism layer. When the lightweight deep learning network is the lightweight 3D convolutional neural network, the step of inputting the four-dimensional geological data volume into the lightweight deep learning network for spatiotemporal joint feature extraction and mineral outcome prediction, obtaining preliminary prediction results at each time step, includes: The four-dimensional geological data volume is input into the lightweight convolutional layer, and spatiotemporal joint feature extraction is performed through three-dimensional depth separable convolution to obtain a preliminary three-dimensional feature map. The preliminary 3D feature map is input into the lightweight convolutional attention mechanism layer. The channel dimension is compressed through lightweight convolution processing to generate a low-dimensional feature map. The low-dimensional feature map is then processed by lightweight fully connected processing to generate attention weights for each feature channel. The features of each channel of the preliminary 3D feature map are dynamically recalibrated based on the attention weights. The recalibrated features are then restored in dimension through lightweight convolution processing to obtain an attention-weighted 3D feature map. The attention-weighted 3D feature map is input into the lightweight fully connected layer. The feature channels are grouped by grouped convolution. After convolution operation is performed on each group of feature channels, the feature channels are fused to obtain the preliminary prediction result.

4. The method according to claim 3, characterized in that, The process of extracting spatiotemporal joint features through 3D depthwise separable convolution to obtain a preliminary 3D feature map includes: By using multiple independent 3D convolution kernels, 3D convolution operations are performed on each feature channel of the 4D geological data volume to generate a depthwise feature map corresponding to each feature channel. By using a 1×1×1 convolution kernel, the depthwise feature maps corresponding to all feature channels are linearly combined across channels to obtain a preliminary three-dimensional feature map.

5. The method according to claim 1, characterized in that, The lightweight long short-term neural network includes a deep separable linear layer, a simplified gating mechanism, and a lightweight channel scaling layer. The simplified gating mechanism includes an update gate and a candidate gate. When the lightweight deep learning network is the lightweight long short-term neural network, the process of inputting the four-dimensional geological data volume into the lightweight deep learning network for spatiotemporal joint feature extraction and mineral outcome prediction yields preliminary prediction results for each time step, including: The four-dimensional geological data volume is unfolded into a three-dimensional spatial feature sequence according to time steps, and the three-dimensional spatial features of the current time step are input into the depth separable linear layer. Each feature channel of the three-dimensional spatial features is independently linearly transformed to generate a channel-wise feature representation. The channel-wise feature representation is then linearly fused across channels through a small-scale point linear layer to generate a fused feature representation for the current time step. The fused feature representation of the current time step and the hidden state of the previous time step are input into the simplified gating mechanism. The update gate activation value and the candidate gate activation value are calculated through the update gate and the candidate gate, respectively. The cell state and hidden state at the current time step are updated based on the update gate activation value and the candidate gate activation value. The updated hidden state is input into the lightweight channel scaling layer. Channel scaling coefficients are generated through low-dimensional intermediate mapping or a small fully connected layer. Based on the channel scaling coefficients, each channel of the hidden state is dynamically weighted to generate a preliminary prediction result for the current time step.

6. The method according to claim 1, characterized in that, When the lightweight deep learning network is the lightweight frequency-spatial neural network, the step of inputting the four-dimensional geological data volume into the lightweight deep learning network for spatiotemporal joint feature extraction and mineral outcome prediction, to obtain preliminary prediction results for each time step, includes: The four-dimensional geological data volume is transformed to the frequency domain through fast Fourier transform to obtain amplitude and phase information. The amplitude information is then weighted by combining the Sigmoid and GELU activation functions to generate frequency-gated weights. By multiplying element by element, the frequency domain information is filtered based on the frequency gating weight, and the filtered amplitude and phase information are fused in a deep feature manner in combination with the bottleneck architecture to obtain a fused frequency signal. The fused frequency signal is then converted to the spatial domain by the inverse fast Fourier transform to obtain the global features after frequency domain enhancement. The global features are input into the spatial gating mechanism, a spatial weight map is generated through convolution, and the global features are weighted element by element according to the spatial weight map to strengthen local key information and obtain the preliminary prediction result.

7. The method according to claim 1, characterized in that, The process involves inputting the preliminary prediction results of each time step as observation values ​​into a temporal fusion Kalman filter. Through state prediction, Kalman gain calculation, and state correction, the state vectors of historical time steps are fused with the observation values ​​of the current time step to generate a four-dimensional prediction result. The three-dimensional geological model is then updated based on the four-dimensional prediction result, including: The preliminary prediction results at each time step are used as observations and input into the temporal fusion Kalman filter. The preliminary prediction results at the initial time step are mapped to the initial state vector and initial state covariance matrix of the Kalman filter. For each subsequent time step, the prior state estimate for the current time step is predicted based on the state vector and state transition model of the previous time step; the prior covariance estimate for the current time step is calculated based on the state covariance matrix, state transition model, and process noise covariance matrix of the previous time step; the Kalman gain for the current time step is calculated based on the prior covariance estimate, observation model, and observation noise covariance matrix; the preliminary prediction result for the current time step is used as the observation value, and the prior state estimate for the current time step is corrected by the Kalman gain to obtain the posterior state estimate for the current time step, and the posterior state estimate is mapped to the four-dimensional prediction result; the prior covariance estimate is updated by the Kalman gain to obtain the posterior covariance matrix for the current time step. The posterior state estimate is mapped to the corresponding location in the three-dimensional geological space, and the three-dimensional geological model is iteratively updated by combining geological constraints and mineral attribute information.

8. An electronic device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 7.

Images