Metallic ore logging characterization method and system based on multi-scale fusion and PINN

Abstract

The application belongs to the technical field of metal mine logging data processing, and in particular to a metal mine logging characterization method and system based on multi-scale fusion and PINN, which comprises generating a standardized data set; constructing a physical information neural network model; using a transfer learning strategy to transfer the common mapping knowledge of the physical information neural network model trained in a mature mining area to a new mining area and fine-tuning; inputting the multi-scale fusion feature curve of the new mining area to the fine-tuned physical information neural network model; establishing a three-dimensional ore body parameter attribute model using Bayesian sequential Gaussian simulation; verifying the three-dimensional ore body parameter attribute model using mining production data, and correcting the parameters of the physical information neural network model through a back propagation mechanism. The defects of poor data quality, weak model generalization and difficulty in integrating multi-dimensional and multi-scale information in the prior art are overcome, and the accuracy, efficiency and adaptability of logging interpretation are improved.

Description

Technical Field

[0001] This application belongs to the field of metal ore logging data processing technology, specifically a metal ore logging characterization method and system based on multi-scale fusion and PINN. Background Technology

[0002] As shallow metallic mineral resources are gradually depleted, exploration and development are shifting towards deep, concealed, and complex structural areas, posing a serious challenge to traditional orebody interpretation methods based on single well logging curves or simple statistical regression.

[0003] The main defects in well logging characterization of metal ores are as follows: The data issues are prominent: complex drilling conditions lead to distortion or loss of key curves such as density and resistivity, and the core chemical analysis samples are scarce, which cannot support detailed modeling. Poor model generalization ability: Traditional grade estimation models based on linear regression or empirical formulas are prone to failure in different mining areas or different mineralization types. Weak ability to integrate multi-dimensional and multi-scale information: The scale differences between borehole logging (centimeter-meter level), ground geophysical exploration (tens of meters level), and mining data (meter to 10-meter level) are huge, and there is a lack of effective fusion methods, making it difficult to construct an accurate three-dimensional ore body model.

[0004] For example, Chinese Patent Publication No. CN104793263A discloses an automatic stratification and evaluation method for well logging curves. This method mainly achieves automatic stratification and evaluation through curve synthesis, inflection point detection, and morphological recognition, which improves interpretation efficiency to some extent. However, this method mainly relies on the mathematical and statistical characteristics of the curves and lacks the ability to integrate physical constraints on complex lithologies and cross-scale information, making it difficult to solve the problems of poor model generalization and multi-scale integration in well logging evaluation of complex metal mines. Summary of the Invention

[0005] This application provides a method and system for characterizing metal ore logging based on multi-scale fusion and PINN, which solves the problems of complex ore body morphology, drastic grade changes, and difficulty in integrating multi-scale geophysical information.

[0006] The first aspect of this application provides a method for characterizing metal ore logging based on multi-scale fusion and PINN, including: Generative adversarial networks are used to augment and repair curves in raw well logging data, generating a standardized dataset. A physical information neural network model is constructed, and the density-volume fraction linear hybrid equation is used as a prior physical constraint and added to the loss function of the physical information neural network model as a physical loss term. The physical information neural network model is trained with a standardized dataset as input. By adopting a transfer learning strategy, the common mapping knowledge of the physical information neural network model trained in mature mining areas is transferred to new mining areas, and only a small amount of core grade analysis data from the new mining areas is used to fine-tune the high-level parameters of the physical information neural network model. Wavelet multi-scale decomposition was performed on various well logging curves that had been repaired in the new mining area. The coefficients of different scales were fused using a preset fusion rule. A multi-scale fusion feature curve containing geological information was generated by a reconstruction algorithm. The multi-scale fusion feature curve was then input into a finely tuned physical information neural network model to calculate the one-dimensional orebody parameters. Using one-dimensional orebody parameters as hard data and three-dimensional physical property data volume obtained by inversion from ground or well geophysical methods as soft data, a three-dimensional orebody parameter attribute model is established using Bayesian sequential Gaussian simulation. The three-dimensional ore body parameter attribute model was verified using mining production data, and the parameters of the physical information neural network model were corrected through backpropagation mechanism.

[0007] Furthermore, the generative adversarial network (GAN) takes undistorted logging curves as input and distorted or missing logging curves as reconstruction targets. The loss function of the GAN is derived from... The norm reconstruction loss and the discriminant adversarial loss are weighted and summed, and the aforementioned The norm reconstruction loss is used to constrain the point-by-point magnitude approximation of the generated curve to the true curve, while the discriminative adversarial loss is used to constrain the consistency between the overall probability distribution of the generated curve and the true curve.

[0008] Furthermore, the loss function of the physical information neural network model is composed of a weighted sum of a data loss term and a physical loss term. The data loss term is the mean square error between the predicted metal mineral volume fraction and the actual metal mineral volume fraction obtained from core chemical analysis, used to constrain the consistency between the physical information neural network model output and the measured label. The physical loss term is constructed based on a density-volume fraction linear hybrid equation, specifically representing the difference between the measured density logging value and the model-predicted density value. The norm, the model predicts density values ​​by linearly weighting the gangue mineral density, the metallic mineral density, and the metallic mineral volume fraction; the gangue mineral density and the metallic mineral density are used as trainable variables during training.

[0009] Furthermore, the transfer learning strategy specifically involves: using a physical information neural network model trained in a mature mining area as a source domain pre-trained model, and using network parameters as a carrier to transfer the common mapping knowledge between well logging response and ore body parameters to the new mining area; during the transfer execution, the parameters of the bottom feature extraction layer of the fixed source domain pre-trained model remain unchanged, and only the grade analysis data from a small number of core boreholes in the new mining area are used to fine-tune the high-level parameters, so that the bottom feature extraction layer retains the common mapping law between well logging response and mineralization characteristics learned in the source domain, and the high-level parameters are adapted to the ore body parameter response characteristics of the new mining area through fine-tuning.

[0010] Furthermore, when fine-tuning the high-level parameters in the new mining area, a small learning rate lower than the initial learning rate during training in the source domain is used to update the high-level parameters.

[0011] Furthermore, during the fine-tuning process, the similarity of the characteristic distribution of logging data from new mining areas and logging data from mature mining areas in the target layer is calculated. When the similarity is higher than a preset threshold, fine-tuning is performed. The similarity is quantitatively evaluated using the Pearson correlation coefficient, and migration weights are assigned to each source domain sample based on the similarity. The higher the similarity, the greater the assigned migration weight.

[0012] Furthermore, the wavelet multi-scale decomposition adopts the Daubechies series wavelet basis, with a decomposition layer of 3 to 5 layers, decomposing each logging curve layer by layer into low-frequency approximation coefficients and multi-level high-frequency detail coefficients; the fusion rule is as follows: the high-frequency detail coefficients at each level are fused using the rule of taking the maximum absolute value, and the low-frequency approximation coefficients are fused using the weighted average rule. The fusion weights are allocated according to the sensitivity of each logging curve to the mineralized interval or the contribution rate of principal component analysis.

[0013] Furthermore, a three-dimensional orebody parameter property model is established using Bayesian sequential Gaussian simulation, including: The original grade values ​​of the hard data points are transformed into standard normal distribution data by performing a normal distribution transformation. Variation function analysis based on standard normal distribution data: Experimental variation functions were calculated along the three main directions of the ore body strike, dip and vertical, and theoretical fitting was performed using a three-dimensional spherical model to determine the range, nugget effect and sill value in each direction; A three-dimensional regular grid is established based on the mining area and exploration accuracy. The search neighborhood of a three-dimensional ellipsoid is defined with each grid node to be estimated as the center. The radii of the three principal axes of the three-dimensional ellipsoid are equal to the range values ​​in each direction. A random access path is generated. At each grid node: the prior mean and prior variance are obtained by solving the simple Kriging equation based on the hard data and simulated values ​​in the neighborhood; soft data is extracted to calculate the soft predicted value and soft data variance to form a likelihood function; the prior mean, prior variance, and likelihood function are fused using Bayes' theorem to obtain a posterior distribution including the posterior mean and posterior variance; simulated values ​​are randomly sampled from the posterior distribution and added to the known dataset. Continue visiting the next grid node in the random path until all grid nodes have obtained simulated values; Perform an inverse Gaussian transform on the simulated values ​​of all grid nodes to restore the simulated values ​​back to the original grade scale and output a three-dimensional ore body parameter attribute model.

[0014] A second aspect of this application provides a metal ore logging characterization system based on multi-scale fusion and PINN, comprising: The data augmentation and repair module is used to perform data augmentation and curve repair on the original well logging data based on generative adversarial networks, and generate a standardized dataset. The physical information neural network evaluation module is used to construct a physical information neural network model. The physical equations of the ore body are embedded as prior knowledge into the loss function of the physical information neural network model. The physical information neural network model is trained with a standardized dataset as input. The transfer learning module is used to transfer the common mapping knowledge of the physical information neural network model trained in the mature mining area to the new mining area using a transfer learning strategy, and to fine-tune the high-level parameters of the physical information neural network model using only a small amount of core grade analysis data from the new mining area. The multi-scale logging fusion module is used to perform wavelet multi-scale decomposition on various logging curves that have been repaired in the new mining area. It uses preset fusion rules to fuse coefficients at different scales and generates a multi-scale fusion feature curve containing geological information through a reconstruction algorithm. The multi-scale fusion feature curve is then input into a fine-tuned physical information neural network model to calculate one-dimensional orebody parameters. The 3D joint inversion modeling module is used to establish a 3D ore body parameter attribute model using one-dimensional ore body parameters as hard data and three-dimensional physical property data volume obtained by ground or well geophysical methods as soft data, and Bayesian sequential Gaussian simulation. The closed-loop verification and update module is used to verify the three-dimensional ore body parameter attribute model using mining production data, and to correct the parameters of the physical information neural network model through a backpropagation mechanism.

[0015] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: By using generative adversarial networks to repair distorted curves and expand small sample data, the problems of missing and poor data quality under complex well conditions were solved. A physical information neural network model was introduced, and the rock physics equations of the ore body were embedded into the physical information neural network model as prior knowledge to ensure the physical rationality of the calculation results. Combined with the transfer learning strategy, the physical information neural network model was able to quickly adapt to new exploration areas. This application achieves multi-dimensional and multi-scale information integration, fuses multiple well logging curve features through wavelet transform, and utilizes geostatistical methods to realize joint inversion of well logging and geophysical exploration, thus establishing a three-dimensional orebody parameter attribute model with well logging resolution.

[0016] To overcome the shortcomings of existing technologies, such as poor data quality, weak model generalization, and difficulty in integrating multi-dimensional and multi-scale information, and to improve the accuracy, efficiency, and adaptability of well logging interpretation. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a metal ore logging characterization method based on multi-scale fusion and PINN, provided for embodiments of this application; Figure 2 A schematic diagram of the network structure of the physical information neural network model provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] See Figure 1 As shown in the embodiment of this application, a method for characterizing metal ore logging based on multi-scale fusion and PINN (Physical Information Neural Network) is provided, including: S101, based on generative adversarial networks, performs data augmentation and curve repair on the original well logging data to generate a standardized dataset; S102, Construct a physical information neural network model, take the density-volume fraction linear mixed equation as a prior physical constraint, add it to the loss function of the physical information neural network model as a physical loss term, and train the physical information neural network model with a standardized dataset as input; S103, using a transfer learning strategy, the common mapping knowledge of the physical information neural network model trained in the mature mining area is transferred to the new mining area, and only a small amount of core grade analysis data from the new mining area is used to fine-tune the high-level parameters of the physical information neural network model. S104 performs wavelet multi-scale decomposition on various well logging curves that have been repaired in the new mining area, uses preset fusion rules to fuse coefficients at different scales, and generates a multi-scale fusion feature curve containing geological information through a reconstruction algorithm. S105 uses one-dimensional orebody parameters as hard data and three-dimensional physical property data volume obtained by inversion from ground or well geophysical methods as soft data. A three-dimensional orebody parameter attribute model is established using Bayesian sequential Gaussian simulation. S106 uses mining production data to verify the three-dimensional ore body parameter attribute model and corrects the parameters of the physical information neural network model through backpropagation mechanism.

[0020] In step S101, the generative adversarial network is a conditional generative adversarial network. Its generator adopts the U-Net architecture. U-Net serves as the generator skeleton of the conditional generative adversarial network (CGAN), including an encoder and a decoder. The encoder extracts the variation features of the logging curve at different scales (such as lithological boundaries and mineralization layer thickness) through downsampling. The decoder gradually recovers the curve length through upsampling. Skip connections transmit the high-frequency detail features of each layer of the encoder to the corresponding layer of the decoder, so that the repaired logging curve approximates the true value at both the amplitude accuracy (local details) and the variation trend (global context).

[0021] Conditional input refers to the information source used by the generator for inference and repair, namely, high-quality, undistorted logging curve data within the target depth range. Specifically, it includes the following two categories: The first category includes other logging curves within the same well section. When a logging curve is distorted or missing in a specific depth range, other undistorted logging curves within the same depth range are selected as input conditions. For example, if the density logging curve is distorted in the 500m to 510m range due to wellbore enlargement, then the undistorted natural gamma ray curve, resistivity curve, and sonic transit time curve within that range are selected as input conditions. Since different logging curves have complementary responses to lithology and mineralization, by combining multiple undistorted curves, the generator can obtain sufficient geological information to reconstruct the target curve.

[0022] The second category is the undistorted segments of distorted curves. When a logging curve is distorted only in a portion of its range, the undistorted segments of that logging curve itself are selected as part of the conditional input. For example, if a density logging curve is distorted only in the 500m to 510m range, then the undistorted segments of that curve shallower than 500m and deeper than 510m are used as conditional input, together with other undistorted curves in the same well section to form complete conditional information.

[0023] The loss function of generative adversarial networks is derived from... The weighted sum of norm reconstruction loss and discriminative adversarial loss is as follows: , in, For the input condition information, For the true curve, For generator, For discriminator, For balance coefficient, To generate the loss function values ​​for the adversarial network, For expectation operators.

[0024] In one example, the encoder path consists of multiple downsampling modules connected in series. Each downsampling module contains a convolutional layer, a batch normalization layer, and a LeakyReLU activation function to extract multi-scale features of the logging curve layer by layer. As the number of downsampling layers increases, the size of the feature map is halved at each level, while the number of channels doubles at each level, enabling the network to capture the variation patterns of the logging curve at different scales (such as abrupt changes in thin layers and trends in thick layers). The decoder path consists of multiple upsampling modules connected in series. Each upsampling module contains a transposed convolutional layer, a batch normalization layer, and a ReLU activation function to progressively restore the spatial resolution of the feature map.

[0025] A skip connection is provided between the corresponding layers of the encoder and decoder. The skip connection concatenates the feature maps output by each layer of the encoder to the input of the corresponding layer of the decoder in the channel dimension. This allows the decoder to simultaneously utilize the high-resolution detail information (such as curve abrupt change points, mineralization interfaces, etc.) retained by the shallow layer of the encoder and the global semantic information (such as the overall lithological variation trend) extracted from the deep layer when reconstructing the curve, thereby taking into account both local accuracy and global consistency in the restoration process.

[0026] The discriminator employs a PatchGAN (Patch Generative Adversarial Network) structure, consisting of multiple stacked convolutional layers. It distinguishes between real and fake curves for each local region of the input curve, outputting an N×N probability matrix. Each element in the matrix represents the probability that a curve segment within its receptive field is a real curve. The PatchGAN structure allows the discriminator to focus on the local texture and morphological features of the curve, effectively identifying discontinuities and unnatural local anomalies in the repaired curve. This guides the generator to maintain the continuity and geological plausibility of the logging curve in the depth direction during the repair process.

[0027] Assume there are multiple logging curves in the target area, including natural gamma (GR), resistivity (RT), density (DEN), and acoustic transit time (AC). Select the logging curve that has not experienced distortion as the input condition. The target output y is a complete and true logging curve at the corresponding depth within the same well section. The generator G takes x as input and outputs the repaired logging curve. The discriminator takes the logging curve as input and outputs the probability that the curve belongs to the true curve.

[0028] For the distortion curve repair task: the logging response value of the undistorted segment in the curve to be repaired is used as the conditional input, and the real logging value corresponding to the distorted segment is used as the supervision target. The generator is trained to learn the mapping relationship from the conditional input to the complete curve.

[0029] For missing curve repair tasks: use combinations of other logging curves in the same well section as input conditions, and use the actual logging value of the curve to be repaired at the corresponding depth as the supervision target to train the generator to learn the cross-curve mapping relationship.

[0030] In the loss function of the above generative adversarial network, the first term for Norm reconstruction loss is calculated by averaging the absolute differences between the generated curve and the true curve at each depth point, and is used to constrain the point-by-point magnitude approximation between the generated curve and the true curve. norm compared to Norms are insensitive to outliers, producing reconstruction results with clearer boundaries and avoiding the loss of thin-layer mineralization information due to over-smoothing of the repair curve.

[0031] Second item and the third item Together, they constitute the discriminative adversarial loss, used to constrain the consistency between the overall probability distribution of the generated curve and the true curve. The former makes the discriminator maximize the discrimination probability of the true curve, while the latter makes the discriminator maximize the misclassification probability of the generated curve. Through the adversarial game between the two, the generated curve becomes indistinguishable from the true curve in terms of overall probability distribution.

[0032] Balance coefficient Used for control The relative weights between norm reconstruction loss and discriminative adversarial loss.

[0033] In the target mining area, complete logging curve segments with good logging quality and no distortion were selected as training samples. To simulate distortion scenarios, a certain depth range was randomly masked on the complete curve (the masking length was set to 2m to 20m). The masked curve segments were used as the generator input, and the original complete curve segments before masking were used as the supervision targets to construct training sample pairs.

[0034] To enhance the model's generalization ability, the following data augmentation strategies are adopted: random noise (simulating instrument error), random scaling (simulating changes in logging response amplitude), and random time shift (simulating depth correction error) are applied to the training samples to enable the model to adapt to different types and degrees of distortion.

[0035] In one example, an alternating optimization strategy is used to train the generator and discriminator: Discriminator training phase: The generator parameters are fixed, and a batch of paired samples is randomly drawn from the training dataset. The input generator produces the generated curve, which is then compared to the real curve. and generating curve Input the data into the discriminator, calculate the discriminant loss, and update the discriminator network parameters through backpropagation so that the discriminator can distinguish between the real curve and the generated curve as accurately as possible.

[0036] Generator training phase: Fix discriminator parameters, and The input generator gets ,Will The discriminator is input to obtain the discrimination probability. The total loss is calculated by combining the reconstruction loss and the adversarial loss. The generator network parameters are updated through backpropagation so that the generator can generate as realistic a repair curve as possible to deceive the discriminator.

[0037] To address the scarcity of core physical property analysis data in mining areas, a pre-trained generative adversarial network (GAN) is employed for data augmentation. Specifically, a conditional GAN ​​is constructed using the combination of well logging responses at various depth points as input conditions and the corresponding core analysis data (such as density, porosity, and metal grade) as the supervised objective. After training, high-fidelity pseudo-samples of core physical properties at all depth points are generated using the combination of well logging responses from all depth points as input conditions, thereby expanding the physical property dataset under small sample conditions.

[0038] After repair and enhancement, the repaired logging curves are stitched together with the undistorted portions of the original logging data according to depth index to form a complete multi-logging curve dataset. Subsequently, all logging curves are normalized to ensure they are within the same numerical dimension range. Min-max normalization is used as the normalization method. After normalization, all logging curves are aligned by depth to form a standardized multi-logging dataset with consistent spatial stratigraphy and uniform dimensions.

[0039] In step S102, a physical information neural network model is constructed. The density-volume fraction linear hybrid equation is used as a priori physical constraint and added to the loss function of the physical information neural network model as a physical loss term. Using a standardized dataset as input, the physical information neural network model is trained to output one-dimensional orebody parameters. The loss function of the physical information neural network model... It consists of two weighted components: data loss and physical loss. , in, These are the physical constraint weighting coefficients. The mean square error between the predicted volume fraction or grade of metallic minerals and the actual volume fraction of metallic minerals obtained from core chemical analysis; This refers to the physical constraint term constructed based on the density-volume fraction linear hybrid equation, specifically the difference between the measured density logging value and the model-predicted density value. Norm, denoted as: , in, The value is the measured density from the well logging, and C is the volume fraction of the metallic minerals. Density of gangue minerals, Density of metallic minerals The volume fraction of the metal mineral is the predicted value, while the latter two are automatically optimized as trainable variables during the training process. Grade refers to the proportion of the mass of useful metal elements or useful components (target minerals) in the total mass of the ore. By introducing density, the conversion between grade and volume fraction can be performed, which is an existing method and will not be elaborated here.

[0040] Represented as , representing the predicted volume fraction of metallic minerals. Or the grade and the true volume fraction of metallic minerals obtained from core chemical analysis at that depth. The mean square error between them.

[0041] The density-volume fraction linear mixing equation is expressed as: = .

[0042] The model-predicted density values ​​are calculated by linearly weighting the gangue mineral density, metallic mineral density, and metallic mineral volume fraction, i.e.: , Predict density values ​​for the model.

[0043] See Figure 2 As shown, the Physics-Informed Neural Network (PINN) model employs a multi-layer fully connected feedforward network structure, including: The input layer receives multi-channel logging curves from the standardized dataset generated in step S101. The input features are arranged point by point according to depth, and each depth point corresponds to a multi-dimensional feature vector. The hidden layers consist of multiple stacked fully connected layers, each containing a number of neurons. Feature mapping between adjacent layers is achieved through linear transformations and non-linear activation functions. The depth and width of the hidden layers are determined based on the amount of training data and the complexity of the ore body parameters. In this embodiment, the number of hidden layers is set to 4 to 8, and the number of neurons in each layer is set to 64 to 256. The network employs residual connections to mitigate the vanishing gradient problem during deep network training.

[0044] The output layer is a linear fully connected layer, with no activation function applied, directly outputting three one-dimensional orebody parameters. These one-dimensional orebody parameters include the following three items: The volume fraction of metallic minerals refers to the percentage of volume of metallic minerals in an ore, expressed as a dimensionless percentage. This parameter is used to calculate the physical loss term and establishes a constraint relationship between the density linear mixing equation and the measured density logging values. It is a core variable for physical consistency constraints.

[0045] Grade refers to the percentage of the mass of useful metallic elements in the total mass of the ore, expressed as a percentage. This parameter is used to calculate data loss items, for monitoring and comparison with the grade label from core chemical analysis, and is the ultimate indicator for evaluating the industrial value of the ore.

[0046] Mineral density refers to the density of metallic minerals, measured in grams per cubic centimeter. This parameter, as a trainable variable, is automatically and iteratively optimized during training, along with gangue mineral density. After training convergence, it is fixed as the representative density value of metallic minerals in that mining area.

[0047] Before training, the physical information neural network model randomly divides labeled samples into training, validation, and test sets according to a preset ratio, while all unlabeled samples are included in the training set for physical constraints. During training, each training batch contains both labeled and unlabeled samples, ensuring the network is simultaneously subject to data supervision and physical constraints in each batch. In each training epoch, the sample data of the current batch is input into the physical information neural network model. For the multi-channel logging feature vector of each depth point received by the input layer, the network maps the input features layer by layer to higher-level abstract features through successive linear transformations and nonlinear activations of multiple fully connected layers. Finally, the output layer outputs three one-dimensional orebody parameters for that depth point: metallic mineral volume fraction, grade, and ore density. This process is called a forward propagation, yielding the network's predicted values ​​for all samples in that batch under the current parameter conditions.

[0048] The optimization algorithm uses the Adaptive Moment Estimation Optimization Algorithm (Adam) or its variants to simultaneously optimize network parameters and trainable variables, and employs an early stopping strategy to determine whether the training has converged.

[0049] In step S103, in actual metal ore exploration, there are significant differences in geological conditions, mineralization types, and surrounding rock properties among different mining areas. Although the physical information neural network model trained in the mature mining area (source domain) in step S102 can achieve high-precision one-dimensional orebody parameter evaluation in that mining area, when the physical information neural network model is directly applied to a new mining area (target domain), the prediction effect often drops significantly due to the following reasons: The gangue mineral composition (such as the different contents of quartz, calcite, feldspar, etc.) in the new mining area leads to differences in logging response and source domain; The different types or enrichment levels of metallic minerals in new mining areas (such as changing from magnetite to chalcopyrite) lead to changes in physical parameters such as density, grade, and volume fraction (e.g., density of metallic minerals). The geological structure of the new mining area is different, and the correspondence between the logging curve morphology and the mineralized strata is not completely consistent with that of the source area.

[0050] Meanwhile, new mining areas often only have a small amount of grade analysis data from core drilling (e.g., only 2-3 core holes and dozens of grade analysis samples in the entire mining area), which is insufficient to train a high-precision physical information neural network model from scratch. If traditional supervised learning methods are used, the small number of samples can easily lead to overfitting, causing the physical information neural network model to fail in unlabeled well sections. Core drilling is a standard professional term in the field of geological exploration, referring to drilling holes that use diamond or alloy drill bits to extract cylindrical rock samples (called "cores") from underground rock strata.

[0051] To this end, a transfer learning strategy is introduced in step S103 to transfer the common mapping knowledge of the physical information neural network model trained in the mature mining area to the new mining area. Only a small amount of core grade analysis data from the new mining area is used to fine-tune the high-level parameters of the model, so as to realize the rapid adaptation of the physical information neural network model in the new geological environment.

[0052] The transfer learning strategy is as follows: a physical information neural network model trained in a mature mining area is used as the source domain pre-training model. The network parameters are used as the carrier to transfer the common mapping knowledge between well logging response and ore body parameters to the new mining area. During the transfer, the parameters of the bottom feature extraction layer of the source domain pre-training model remain unchanged. Only the grade analysis data of a small number of core boreholes in the new mining area are used to fine-tune the high-level parameters. This allows the bottom feature extraction layer to retain the common mapping law between well logging response and mineralization characteristics learned in the source domain. The high-level parameters are then fine-tuned to adapt to the specific ore body parameter response characteristics of the new mining area.

[0053] The idea behind transfer learning is that although the source domain (mature mining area) and the target domain (new mining area) differ in geological conditions, they share certain common patterns. Specifically, in the well logging evaluation scenario: Common patterns (transferable knowledge) include: the morphological characteristics of well logging response as lithology changes (e.g., mineralized layers show low values ​​on the natural gamma curve, high values ​​on the resistivity curve, and high values ​​on the density curve), the nonlinear mapping law between the combination of multiple well logging curves and the degree of mineralization enrichment, and the density-volume fraction-grade physical constraint mechanism embedded in the network loss function.

[0054] Individual characteristics (required to be relearned) include: the absolute range of grade, the specific density values ​​of gangue and metallic minerals, the response amplitude of the mineralized layer on the logging curve, and the curve combination characteristics corresponding to different mineralization types.

[0055] Based on the above understanding, the transfer learning strategy takes the common rules already learned by the pre-trained model in the source domain as the foundation, and only makes a small number of data-driven adjustments for the unique characteristics of the new mining area, thus avoiding training from scratch.

[0056] The common mapping knowledge contained in the well-trained physical information neural network model of mature mining areas is reflected in the following three levels: The network's bottom-layer weight parameters encode the well-logging-mineralization feature recognition capability. The bottom-layer feature extraction layer of the physical information neural network model, through extensive data learning, has mastered the common variation patterns of well-logging curves (such as natural gamma, resistivity, density, and sonic transit time) with lithology and mineralization. For example, regardless of the mining area, mineralized layers typically exhibit relatively low values ​​on the natural gamma curve, relatively high values ​​on the resistivity curve, and relatively high values ​​on the density curve. This curve combination pattern and its correspondence with mineralization is a common pattern found in different mining areas and possesses good transferability.

[0057] The density-volume fraction-grade physical constraint mechanism is embedded in the loss function. During the training of the source domain pre-trained model (physical information neural network model), the two trainable variables, gangue mineral density and metallic mineral density, have converged to the optimal values ​​of the source domain ore area. Simultaneously, the network itself has learned the physical conversion relationship between metallic mineral volume fraction and grade. Although the absolute value of the density parameter may change in a new ore area, the physical law of the linear mixing relationship between density and volume fraction is universal. By inheriting the physical constraint mechanism, density logging data can continue to be used to apply physical constraints to grade prediction during fine-tuning in new ore areas, thus maintaining physical consistency with a small number of grade labels.

[0058] The high-level parameters of the physical information neural network model have been preliminarily learned to identify grade grading and ore body boundaries. The high-level regression layer of the source domain pre-trained model has established a mapping framework from the feature space to the ore body parameter space. Although the specific grade numerical mapping relationship needs to be readjusted based on the data from the new mining area, the basic structure and direction of the mapping have been formed, and only correction is needed with a small amount of labeled data from the new mining area.

[0059] During transfer learning, the network structure of the source domain pre-trained model (physical information neural network model) remains unchanged, including the input layer dimension, the number of hidden layers and the number of neurons per layer, and the output layer dimension. The network structure itself does not change due to transfer, but the trainable network parameters are altered.

[0060] During transfer learning, the network parameters of the physical information neural network model are divided into the following two categories and processed using different strategies: The first type involves freezing layer parameters. This involves fixing the weight matrix and bias terms of the bottom feature extraction layers (usually the first two to three fully connected layers) of the source domain pre-trained model, preventing them from being updated during fine-tuning. The purpose of freezing the bottom layer parameters is to preserve the well-logging-mineralization common feature extraction capabilities already learned by the source domain pre-trained model, preventing the loss of this universal knowledge through a small number of samples from new mining areas during fine-tuning.

[0061] The second category involves fine-tuning layer parameters. The weight matrices and bias terms of the high-level regression layers (usually the last two to three fully connected layers) of the pre-trained model in the source domain are used as trainable parameters and updated under the guidance of a small amount of core sample analysis data from the new mining area. These high-level parameters are responsible for mapping the common features extracted from the lower layers to specific orebody parameter values. These values ​​are closely related to the specific mineral composition, grade distribution range, and other individual characteristics of the mining area, and therefore need to be adjusted based on the measured data from the new mining area.

[0062] The fine-tuning process employs supervised retraining, specifically as follows: The first step is to organize the grade analysis data of a small number of core boreholes in the new mining area and pair them with the standardized logging feature vectors of the corresponding depths to form a small sample training set for the target domain.

[0063] The second step involves using the standardized well logging feature vector of the new mining area as input, the corresponding grade analysis value as the supervision label, and fine-tuning the physical information neural network model using the same loss function form (including data loss term and physical loss term) as in step S102.

[0064] Third, during the fine-tuning process, the parameters of the frozen layer remain fixed, and only the gradient of the fine-tuning layer parameters is calculated and updated via backpropagation. Simultaneously, the two trainable variables, gangue mineral density and metallic mineral density, are reinitialized during fine-tuning in the new mining area, participating in the synchronous optimization of the fine-tuning process to adapt to the actual gangue mineral composition and metallic mineral type of the new mining area.

[0065] The fourth step is fine-tuning, which uses a small learning rate lower than the initial learning rate during source domain training to update parameters. The use of a small learning rate is to prevent a small number of samples from new mining areas from causing drastic perturbations to the model, and to avoid destroying the common knowledge already learned by the pre-trained model in the source domain during fine-tuning.

[0066] To further improve the migration effect, this embodiment preferably uses a loss function based on similarity weighting for fine-tuning and optimization.

[0067] During the fine-tuning process, the similarity of the characteristic distribution of well logging data from new mining areas and mature mining areas in the target layer is calculated. When the similarity is higher than a preset threshold, fine-tuning is performed. The similarity is quantitatively evaluated using the Pearson correlation coefficient, and migration weights are assigned to each source domain sample based on the similarity. The higher the similarity, the greater the assigned migration weight.

[0068] The similarity of well logging data distribution between the new mining area (target domain) and the mature mining area (source domain) in the target interval was calculated, and the similarity evaluation used the Pearson correlation coefficient: , in, The Pearson correlation coefficient is used. The total number of sampling points. For the source region (reference borehole in mature mining area) at the first Well logging response values ​​at each sampling point The mean value of the source domain logging response. Let k be the logging response value of the target domain (target borehole in the new mining area) at the kth sampling point. This represents the mean value of the logging response values ​​in the target domain.

[0069] For source domain samples and target domain samples The Pearson correlation coefficient between the point-by-point response values ​​of the logging curves of the two samples in the target layer is calculated. The Pearson correlation coefficient ranges from -1 to 1. The closer the value is to 1, the more similar the two samples are in terms of logging response characteristics. The closer the value is to 0, the greater the difference. The closer the value is to -1, the negative correlation is indicated.

[0070] Transfer weights are assigned to each source domain sample based on similarity; the higher the similarity, the greater the transfer weight. Specifically, source domain samples with a Pearson correlation coefficient greater than 0.8 are given high weights, playing a greater role in the loss function during fine-tuning; source domain samples with a Pearson correlation coefficient between 0.5 and 0.8 are given medium weights; and source domain samples with a Pearson correlation coefficient less than 0.5 are given low or zero weights.

[0071] The optimization objective of the fine-tuning process is a weighted loss function: The weighted loss function is equal to the sum of the loss functions of each source domain sample. The loss function of each sample is determined by the deviation between its predicted value and the true value. This deviation is multiplied by the similarity weight corresponding to the sample and then summed to obtain the total weighted loss function.

[0072] After transfer learning is completed, the prediction accuracy of the fine-tuned physical information neural network model is verified using the reserved verification grade data (i.e., depth points with grade analysis but not involved in fine-tuning) in the new mining area. The verification indicators include: the mean absolute error between the predicted grade and the measured grade, the coefficient of determination, and the Pearson correlation coefficient between the predicted grade and the measured grade.

[0073] If the verification results show that the prediction accuracy of the fine-tuned physical information neural network model is significantly higher than that of the model pre-trained directly from the source domain (without fine-tuning), and comparable to the accuracy of training an ideal model from scratch in the new mining area (assuming sufficient data), then the transfer learning is considered successful. At this point, the fine-tuned physical information neural network model can continuously evaluate the one-dimensional orebody parameters across the entire well section in the new mining area, outputting three parameters: metallic mineral volume fraction, grade, and ore density.

[0074] In step S105, different logging curves exhibit varying sensitivities to geological features. For example, density curves are sensitive to changes in mineralized layer density, resistivity curves are sensitive to mineralization fluids and metallic mineral enrichment, and natural gamma curves are sensitive to changes in clay content and lithology. A single logging curve often only highlights one type of geological feature and cannot simultaneously address multiple needs such as identifying mineralized layer interfaces, differentiating lithologies, and depicting detailed internal features of the ore body. Furthermore, ore body information is distributed across different scales in logging signals; thick ore bodies exhibit low-frequency macroscopic trends, while thin mineralization and ore body boundaries show high-frequency local abrupt changes. Traditional single-scale analysis methods struggle to extract these multi-scale features simultaneously.

[0075] Therefore, in step S105, the wavelet multi-scale decomposition method is used to perform multi-resolution decomposition on the various repaired logging curves. At different scales, mineralized layer interface information (high-frequency details) and lithological sensitive information (low-frequency trends) are extracted respectively. The coefficients at different scales are fused using preset fusion rules, and then a multi-scale fusion feature curve containing rich geological information is generated through a reconstruction algorithm.

[0076] Wavelet transform is a time-frequency localization analysis method that can decompose one-dimensional logging signals into different frequency components, accurately locating the depth of abrupt change points while preserving the overall trend of the signal.

[0077] In one example, wavelet multi-scale decomposition uses the Daubechies wavelet basis to perform multi-scale decomposition on various logging curves, including density, natural gamma, resistivity, and magnetic susceptibility. Each logging curve is decomposed layer by layer into low-frequency approximation coefficients and multi-level high-frequency detail coefficients to separate macroscopic trend information and local abrupt change information in the logging signal. The number of decomposition layers is set according to the logging curve sampling interval and the minimum thickness of the ore body to be identified, with 3 to 5 layers being preferred.

[0078] During the fusion phase, differentiated fusion rules are applied to decomposition coefficients of different properties. For high-frequency detail coefficients, an absolute value maxima fusion rule is used: at each scale layer and each depth location, the absolute values ​​of the high-frequency detail coefficients of each logging curve at the same location are compared, and the coefficient with the largest absolute value is selected as the fused high-frequency detail coefficient for that location. The theoretical basis of this rule is that the magnitude of the absolute value of the high-frequency detail coefficient reflects the intensity of the signal abrupt change at that depth point. The larger the absolute value, the more sensitive the logging curve is to the geological interface at that location. Therefore, the curve information with the strongest abrupt change is selected to preserve key geological boundary features such as the top and bottom interfaces of the mineralized layer.

[0079] For the low-frequency approximation coefficient, a weighted average fusion rule is adopted: the low-frequency approximation coefficients of each logging curve at the lowest decomposition layer are weighted and summed according to a preset weight to obtain the fused low-frequency approximation coefficient.

[0080] In the weighted average fusion of the low-frequency approximation coefficients, the fusion weight of each logging curve is allocated according to its sensitivity to the mineralized interval, specifically as follows: Typical mineralized and non-mineralized zones within the target work area are pre-selected as standard sample intervals. The mean and standard deviation of the response of each logging curve within the mineralized zone, and the mean and standard deviation of the response within the non-mineralized zone, are statistically analyzed. Based on the above statistical results, a mineralization sensitivity index for each logging curve is calculated. This index is used to quantify the ability of each curve to distinguish between mineralized and non-mineralized zones.

[0081] The mineralization sensitivity index is calculated as follows: the absolute value of the difference between the mean response of the logging curve in the mineralized zone and the mean response in the non-mineralized zone is divided by the standard deviation of the logging curve's response in the mineralized zone. The resulting ratio is the mineralization sensitivity index of the logging curve. The larger the mineralization sensitivity index value, the stronger the logging curve's ability to identify mineralized zones and the more sensitive it is to mineralization information.

[0082] Based on this, the mineralization sensitivity index of each logging curve is normalized so that the sum of the sensitivity indices of all logging curves equals 1. The normalized sensitivity index value of each curve is then used as the fusion weight in the low-frequency approximation coefficient weighted average fusion. Logging curves with higher mineralization sensitivity receive higher fusion weights in the fusion, thus dominating the low-frequency trend of the multi-scale fusion characteristic curves; logging curves with lower mineralization sensitivity receive lower fusion weights and play a supporting role in the fusion.

[0083] For new mining areas or regions lacking prior statistical data on mineralized layers, principal component analysis (PCA) is used as an alternative to determine the fusion weights of each curve. PCA is performed on standardized data from multiple well logging curves in the target layer. Principal components with cumulative contribution rates reaching a preset threshold are extracted. The contribution rate of each well logging curve to the principal components is calculated, and the normalized contribution rate is used as the fusion weight for each curve. Curves with higher contribution rates are assigned higher fusion weights. The fused low-frequency coefficients reflect the average trend information of multiple curves, embodying the main variation patterns of thick ore bodies.

[0084] The fused low-frequency approximation coefficients and the fused high-frequency detail coefficients from each layer are subjected to inverse wavelet transform using the Mallat reconstruction algorithm to generate a fused logging curve containing rich geological information. This multi-scale fused feature curve retains the curve abrupt change characteristics most sensitive to mineralization response at the mineralized layer interface, and preserves the comprehensive lithological variation trend of multiple logging curves within thick ore bodies, thus balancing interface identification accuracy and lithological differentiation capability.

[0085] The aforementioned fine-tuned physical information neural network model and multi-scale fusion feature curves containing rich geological information have been obtained for the new mining area. However, these results are all limited to one-dimensional borehole profiles and cannot reflect the overall distribution morphology, grade variation law, and resource distribution characteristics of the ore body in three-dimensional space. Traditional three-dimensional modeling methods rely solely on discrete borehole grade data, and the spatial interpolation accuracy is limited by borehole density, resulting in insufficient model reliability in sparsely drilled areas.

[0086] Therefore, in step 105, the multi-scale fusion feature curve is input into the fine-tuned physical information neural network model to calculate high-precision one-dimensional orebody parameters, which are used as hard data. Combined with the three-dimensional physical property data volume obtained by ground or well geophysical methods, the three-dimensional orebody parameter attribute model with well logging resolution is established using the Bayesian sequential Gaussian simulation method.

[0087] Hard data refers to a one-dimensional orebody parameter profile calculated using a two-physical-information neural network model, specifically including quality grade profiles and ore density profiles. This data is continuously distributed along the borehole depth direction, with sampling intervals consistent with well logging curves, typically 0.125m. Each borehole can provide grade and density values ​​at thousands of depth points, forming a precise spatially constrained point set.

[0088] Soft data refers to three-dimensional physical property data volumes obtained by inversion using surface or well geophysical methods. These geophysical methods include high-precision surface gravity exploration, magnetic exploration, electromagnetic exploration, and in-well induced polarization (IPC) methods.

[0089] Taking high-precision ground gravity exploration as an example, gravity measuring points are laid out in a regular grid within the mining area. After collecting gravity data, the data undergoes processing such as topographic correction, Bouguer correction, and separation of regional and residual fields to obtain a three-dimensional residual density anomaly data volume. This three-dimensional residual density anomaly data volume covers the entire mining area horizontally, with a grid spacing typically between 50m and 100m; vertically, it is layered according to a certain layer spacing (e.g., 10m to 20m) to form a three-dimensional regular grid data. High anomaly areas in the residual density anomaly volume indicate the possible presence of high-density mineralization concentration areas, while low anomaly areas correspond to surrounding rocks or alteration zones.

[0090] Soft data does not directly participate in grade assignment, but participates in grade estimation in the co-kriging estimation in the form of trend correction terms or covariate weights through cross-variance functions, controlling the macroscopic spatial distribution direction of the model in sparse borehole regions.

[0091] A three-dimensional orebody parameter and property model was established using Bayesian sequential Gaussian simulation, including: The raw grade values ​​of the hard data points are transformed into standard normal distribution data by performing a normal distribution transformation. Variation function analysis based on standard normal distribution data: Experimental variation functions were calculated along the three main directions of the ore body strike, dip and vertical, and theoretical fitting was performed using a three-dimensional spherical model to determine the range, nugget effect and sill value in each direction; A three-dimensional regular grid is established based on the mining area and exploration accuracy. A three-dimensional ellipsoid search neighborhood is defined with each grid node to be estimated as the center. The radii of the three principal axes of the three-dimensional ellipsoid are equal to the range values ​​in each direction. A random access path is generated. At each grid node: the prior mean and prior variance are obtained by solving the simple Kriging equation based on the hard data and simulated values ​​in the neighborhood; soft data is extracted to calculate the soft predicted value and soft data variance to form a likelihood function; the prior mean, prior variance, and likelihood function are fused using Bayes' theorem to obtain a posterior distribution including the posterior mean and posterior variance; simulated values ​​are randomly sampled from the posterior distribution and added to the known dataset. Continue visiting the next grid node in the random path until all grid nodes have obtained simulated values; Perform an inverse Gaussian transform on the simulated values ​​of all grid nodes to restore the simulated values ​​back to the original grade scale and output a three-dimensional ore body parameter attribute model.

[0092] This involves performing a normal distribution transformation on the original grade values ​​of the hard data points, converting them into standard normal distribution data; specifically, this includes: The raw grade values ​​of all hard data points are sorted in ascending order of numerical value. The purpose of sorting is to map the raw data to an ordered rank space.

[0093] For each hard data point in the sorted sequence, its empirical cumulative probability value is calculated. The basic calculation method is to divide the sort number of the hard data point by the total number of samples; the result represents the proportion of samples in the original data that are less than or equal to that value. When multiple samples have the same original grade value, the average rank method is used: first, determine the minimum and maximum rank of the same value in the sorted sequence; then calculate the average rank; finally, use the ratio of the average rank to the total number of samples as the empirical cumulative probability value corresponding to that value. The average rank method ensures that all hard data points with the same value obtain the same cumulative probability value, avoiding inconsistencies caused by random selection of sorting positions. After this step, each original grade value is mapped to an empirical cumulative probability value in the interval between 0 and 1, which represents the quantile position of the grade value in the entire data distribution.

[0094] Normal score mapping. Each empirical cumulative probability value is substituted into the inverse function of the standard normal cumulative distribution function (i.e., the standard normal quantile function) to calculate the corresponding standard normal score value. The standard normal cumulative distribution function is a monotonically increasing function that converts standard normal scores into empirical cumulative probabilities. It can inversely solve for the corresponding standard normal quantiles of empirical cumulative probability values ​​in the interval from 0 to 1. After this step, each empirical cumulative probability value is mapped to a standard normal score value, completing the transformation from the probability space to the standard normal score space.

[0095] After the normal distribution transformation of the hard data was completed, the original grade values ​​of all hard data points were converted into standard normal scores with a mean of zero and a variance of one. These standard normal scores serve as the basis for variogram analysis and are involved in all calculations. Variation analysis is the bridge connecting hard data and the three-dimensional orebody parameter property model in Bayesian sequential Gaussian simulation.

[0096] The specific execution process of variation function analysis is as follows: The experimental variability function is calculated along the three principal directions of the ore body: strike, dip, and vertical. In practice, the spatial vectors of the three principal directions must first be determined: the strike direction is the direction in which the ore body extends the longest; the dip direction is perpendicular to the strike and inclined along the ore body; and the vertical direction is perpendicular to both the strike and dip. Along each direction, the lag distance is increased successively by a certain increment (usually determined based on the average borehole spacing; in this embodiment, it is set to an integer multiple of the grid spacing, i.e., a multiple of 25 meters). For each pair of known hard data points satisfying the lag distance condition, the square of the difference between their standard normal scores is calculated. The sum of the squares of the differences of all hard data pairs is divided by twice the total number of data pairs to obtain the experimental variability function value for that direction at that lag distance. The experimental variability function calculations for all three directions are based on the same set of standard normal score hard datasets; the only difference is the direction of the lag distance vector when selecting hard data pairs. After the calculations are completed, a set of discrete experimental variability function scatter plots is obtained for each direction.

[0097] After calculating the experimental variogram, a three-dimensional spherical model was used to theoretically fit the scatter points of the experimental variogram in three directions. The spherical model is one of the most commonly used theoretical variogram models in geostatistics.

[0098] Its mathematical expression is: when hour, This means that the grade values ​​are exactly the same at zero distance, with a variation of zero. The lag distance, This is the value of the variation function when the lag distance is zero; when hour, This indicates that within the range of variation, the variogram value increases with increasing hysteresis, but the rate of increase gradually slows down, and the curve rises in a parabolic shape to the sill value. Indicates a change in range. When the lag distance is The value of the variation function at time, For the nugget effect, This is the value of the partial sill. when hour, This indicates that when the lag distance exceeds the range, the hard data points no longer have spatial correlation, and the variogram value stabilizes at the sill value level.

[0099] in, The nugget effect reflects micro-variability or measurement error. It is determined by the intercept of the experimental variability function when the hysteresis approaches zero, and the unit is the square of the grade value. The partial sill value reflects spatial structure variation and is equal to the sill value minus the nugget effect. The unit is the square of the grade value. For variable range, the effective range reflecting spatial correlation is expressed in meters, when the lag distance is... Greater than range At that time, the data no longer have spatial correlation.

[0100] nugget effect The sources include two aspects: first, the inherent variability of the ore body itself at the microscale (short-distance grade fluctuations caused by uneven mineralization); and second, measurement errors (accuracy of logging instruments and random errors in core sampling). A large value indicates strong short-scale noise in the data, and the measurement accuracy should be checked. The magnitude of the nugget effect directly affects the smoothness of the simple Kriging estimate; the greater the smoothness, the smoother the estimate, and the easier it is to suppress local details. Offset sill value. It reflects the intensity of large-scale structural variations in the ore body in space. A larger value indicates a more significant spatial structural variation in the ore body's grade. (Platform value) The structure ratio represents the upper limit to which the variance function approaches as the lag distance increases, and is equal to the population variance of the hard data. This reflects the structural strength of the space.

[0101] Variable range The variogram is one of the most important parameters in variogram analysis, directly determining the influence range of hard data in Kriging estimation. In this embodiment, the variogram analysis results determined the spatial structure parameters in three directions: In one example, the primary range in the strike direction is 500 meters, indicating that the ore body has the best continuity along the strike direction and the largest spatial correlation influence range; the secondary range in the dip direction is 300 meters, indicating that the ore body has the second best continuity along the dip direction; and the vertical range in the vertical direction is 30 meters, indicating that the ore body has the most drastic vertical changes and the smallest spatial correlation influence range. The huge differences in the ranges in the three directions reflect the typical morphological characteristics of skarn-type copper-iron ore bodies, that is, the ore body extends in a layered or quasi-layered manner in the horizontal direction (strike and dip), and is thinner and more drastically varied in the vertical direction.

[0102] Theoretical fitting is performed using the least squares method, which aims to minimize the sum of squared residuals between the experimental variogram scatter points and the theoretical model curve, and iteratively adjusts the model. , and The optimal values ​​for the three parameters are determined. During the fitting process, the initial parameters are first visually estimated based on the scatter plot, and then a nonlinear optimization algorithm is used for precise fitting. The goodness of fit is determined using the coefficient of determination. Evaluation, in one example, requires The value should not be lower than 0.85; otherwise, it is necessary to check whether there are any abnormalities in the calculation of the experimental variation function, or to consider using other models (such as exponential or Gaussian models) for refitting.

[0103] After the variogram analysis is completed, the covariance function is derived from the variogram. The relationship between the covariance function and the variogram is as follows: , in, Lag distance The covariance function value at that location; The covariance function value at zero lag distance is equal to the total variance of the data, i.e., the sill value. This relationship shows that: when the lag distance is zero, the covariance is the largest and equal to the total variance; as the lag distance increases, the covariance gradually decreases; when the lag distance exceeds the range, the covariance approaches zero, indicating that there is no longer spatial correlation between the data.

[0104] The covariance function will be directly used in the subsequent steps to solve the simple Kriging equation. This is crucial in the Kriging estimation at each point to be estimated. Therefore, the accuracy of the variogram analysis directly affects the precision of the Kriging estimate.

[0105] After the variogram analysis is completed, the range values ​​in the three directions are used to define the search neighborhood of the 3D ellipsoid. Centered on each grid point to be estimated, the radii of the three principal axes of the 3D ellipsoid are equal to the primary range of the strike, the secondary range of the dip, and the vertical range, respectively. Simultaneously, the nugget effect and the sill value participate in the calculation of the Kriging variance, determining the degree of dispersion of the prior distribution at each point to be estimated. All results of the variogram analysis are recorded in parametric form, serving as the basic input for the entire Bayesian sequential Gaussian simulation.

[0106] After completing the variogram analysis and obtaining the range parameters in each direction, a three-dimensional regular grid is established and the search neighborhood is set. The establishment of the three-dimensional regular grid is determined based on the mining area and exploration accuracy. The grid origin is set at the lowest point in the southwest corner of the mining area. Along the strike direction, with the strike direction corresponding to the main range as the axis, it extends from the westernmost end of the ore body to the easternmost end, covering the entire ore body area and extending outwards a certain distance to ensure the integrity of the boundary simulation. Along the dip direction, it extends from the northernmost end of the ore body to the southernmost end. Vertically, it extends from the surface to a certain depth below the bottom plate of the ore body. The grid spacing is reasonably set according to the exploration density. After the grid is established, the hard data is assigned to the corresponding grid nodes. If multiple borehole data exist in the same grid node, the weighted average value is taken as the hard data value of that grid node.

[0107] A search neighborhood of a three-dimensional ellipsoid is defined centered on each grid node to be estimated. The shape and size of this search neighborhood are controlled by spatial structure parameters (range, nugget effect, and sill value) determined by variogram analysis: the three principal axes of the three-dimensional ellipsoid correspond to the strike, dip, and vertical directions of the ore body, respectively, and the radii of the three principal axes are equal to the range values ​​in each direction. This setting ensures that the search neighborhood matches the actual spatial structure characteristics of the ore body, with the largest search range in the strike direction to fully utilize hard data information along the ore body's extension direction; and the smallest search range in the vertical direction to avoid including distant data points that are not representative due to drastic vertical variations.

[0108] After completing the variation function analysis, establishing the 3D regular mesh, and setting the search neighborhood, the sequential simulation main loop begins. First, random access paths are generated for all mesh nodes to be simulated. A random number generation algorithm is used to randomly arrange all mesh nodes, producing a non-repeating access sequence. This random path ensures that each mesh node is visited only once, and the access order does not introduce any directional bias. By randomizing the access order, the simulated value of each mesh node is constrained by known hard data from all directions, resulting in more objective simulation results that accurately reflect spatial uncertainty.

[0109] Select the grid nodes to be simulated sequentially according to a random path, and perform the following sequential simulation operation for each grid node: The first step is the extraction of known hard data points. Centered on the current grid node to be estimated, a pre-defined 3D ellipsoid is used to search the neighborhood, extracting all known data points falling within this ellipsoid from the dataset. This known dataset consists of two parts: first, the initial borehole hard data points, containing all original borehole grade values ​​after Gaussian score-normal transformation. These data points remain constant throughout the simulation and form the foundation of Kriging estimation; second, the values ​​of grid nodes simulated before the current grid node. These values ​​act as carriers of hard data information, gradually transferring information from the simulated area to the unsimulated area. As the simulation progresses, the size of the known dataset gradually increases, allowing subsequent Kriging estimations of grid nodes to obtain more neighborhood information.

[0110] The second step is prior distribution calculation. Based on the known data points extracted in the first step, the simple Kriging equations are solved. For Given a set of data points, the basic form of the simple Kriging equation is: For each known data point... All weight coefficients With known data points and known data points The sum of the products of the covariances of the known data points is equal to the sum of the products of the known data points. With the data points to be estimated The covariance between the data points is calculated, where the covariance function is derived from the spherical model parameters determined during the variation function analysis phase. The weighting coefficients for each known data point are obtained by solving the simple Kriging equation. These weighting coefficients satisfy the condition of minimizing the estimated variance, and nearby known data points are given greater weights, while redundant information is automatically compressed.

[0111] Then the simple Kriging estimate was calculated. (Prior mean), which is equal to the weight coefficients. The sum of products of the standard normal scores of the corresponding known data points. This simple kriging estimate is the statistically best estimate of the grade of the point to be estimated, considering only the hard data. The kriging variance is also calculated. It is equal to the value of the covariance function at zero lag. Subtract each weight coefficient Kriging variance is the sum of the covariance products between the corresponding known data points and the grid nodes to be estimated. Kriging variance reflects the degree of uncertainty in estimation based on hard data; it is small when hard data is dense in the neighborhood and close to the point to be estimated, and large when hard data is sparse or located at the edge of the ore body. Simple Kriging estimate. (Prior mean) and Kriging variance (Prior variance) together constitute the prior normal distribution.

[0112] The third step is the construction of the likelihood function. In the example, the residual density anomaly at the current node location to be estimated is extracted from the gravity inversion 3D data volume. This anomaly is then converted into a grade prediction value, i.e., a soft prediction value, using a density-grade linear mixture model. The density-grade linear mixture model is derived from the density-volume fraction linear mixture equation by converting between volume fraction and grade, and can be expressed as: , in, F represents the mass percentage of useful metallic elements in a metallic mineral, and F represents the grade.

[0113] Simultaneously, based on the cross-validation results of the model fitting residuals and gravity inversion accuracy (using conventional calculation methods), the soft data variance is determined. The soft predicted values ​​and the soft data variance constitute a likelihood function, which follows a normal distribution. Given the true grade values, the observed soft data follows a normal distribution with the soft predicted values ​​as the mean and the soft data variance as the variance.

[0114] The fourth step is Bayesian update fusion. The prior distribution obtained in the second step is fused with the likelihood function obtained in the third step using Bayes' theorem. Since both the prior distribution and the likelihood function are normal distributions, the posterior distribution is also a normal distribution.

[0115] The formula for calculating the posterior mean is: the posterior mean equals the product of the prior mean and the quotient of the prior variance, plus the product of the soft prediction and the quotient of the soft data variance, divided by the sum of the reciprocals of the prior variance and the soft data variance. Physically, this formula represents a precision-weighted average of the prior mean and the soft prediction, with the weights being the reciprocals of their respective variances. When hard data density is high and Kriging variance is low, the prior variance is small and the weights are large, with the prior mean dominating the posterior result. When hard data is sparse and the prior variance is large, the weight of the soft data increases, playing a supplementary constraint role.

[0116] The formula for calculating the posterior variance is: the posterior variance equals the reciprocal of the sum of the reciprocals of the prior variance and the soft data variance. Physically, this formula represents a harmonic combination of the prior and soft data variances, whose value is always less than either the prior or soft data variance, reflecting the objective fact that uncertainty is reduced after fusing two independent information sources. The posterior mean and posterior variance together constitute a posterior normal distribution, which represents the comprehensive probability distribution of the grade values ​​of the grid nodes to be estimated under the condition of simultaneously considering both hard and soft data.

[0117] Step 5: Posterior random sampling. A value is randomly selected from the posterior normal distribution obtained in Step 4 as the simulation representation for that grid node. Specifically, a standard normal random variable is generated using a standard normal random number generator. Then calculate the simulated values. It equals the posterior mean plus the posterior standard deviation (the positive square root of the posterior variance) multiplied by... The purpose of random sampling is not to output a deterministic estimate, but to generate a random realization that reflects the characteristics of the posterior distribution probability. This realization is controlled by both the central location of the posterior mean and the discrete range specified by the posterior standard deviation, so that the differences between the generated realizations can reflect the spatial uncertainty of the model.

[0118] Step 6, data update. Replace the simulated values ​​obtained in step 5. and their corresponding three-dimensional spatial coordinates The known dataset is added. The purpose of this data update operation is to pass the simulation results of the current grid node as an information source to subsequent nodes. The simulated values ​​added will become one of the known data points for subsequent grid nodes to be estimated, participating in Kriging estimation. The hard data constraint is gradually diffused outward from the borehole location through this sequential information propagation mechanism, eventually covering the entire 3D regular grid. The closer to the borehole, the stronger the direct hard data constraint, and the closer the simulated values ​​are to the actual observations; the farther away from the borehole, the weaker the indirect hard data constraint, and the more significant the supplementary role of soft data.

[0119] Steps one through six are repeated sequentially for each grid node in the random path until all grid nodes to be simulated have been assigned values, completing a full equal-probability implementation. The simulation at each grid node utilizes all currently available information, including initial hard data, simulated node values, and soft data constraints.

[0120] Create a lookup table containing two columns: the first column represents the sorted raw grade values, and the second column represents the corresponding standard normal score. The data is sorted in ascending order by raw grade values, and each row records the correspondence between a raw grade value and a standard normal score. This lookup table is bidirectional; it allows you to find the standard normal score based on the raw grade value, and vice versa.

[0121] The lookup table is used to perform an inverse Gaussian transform on the simulated values ​​of all grid nodes, restoring the simulated values ​​back to the original grade scale, and outputting the three-dimensional ore body parameter attribute model. Specifically, this includes: For each grid node in a 3D regular grid, the mean of that grid node in the standard normal score space is first obtained. Then, the two closest adjacent records to the standard normal score are located in the lookup table; that is, a standard normal score less than or equal to the value to be restored is found, and the next standard normal score greater than or equal to the value to be restored is found. After determining these two adjacent records, the corresponding original grade value is calculated using linear interpolation. The basic principle of linear interpolation is to interpolate between the two corresponding original grade values ​​according to the relative position of the value to be restored between two adjacent standard normal scores, using the same proportion. For example, for a mean field of 50 equally probable realizations, the input of each grid node is the mean of the standard normal score, and the output is the original grade mean corresponding to the mean. For cases where the complete data volume of all realizations needs to be output, the above inverse transformation is performed separately for each realization, and each grid node finally obtains 50 restored original grade values.

[0122] Furthermore, in practical applications, there may be cases where the standard normal score to be restored falls outside the range of the lookup table. When the value to be restored is less than the minimum standard normal score in the lookup table, linear extrapolation is used, that is, extrapolation is performed outward based on the trend determined by the first two records of the lookup table. When the value to be restored is greater than the maximum standard normal score in the lookup table, linear extrapolation is also used. However, the extrapolation result must be checked for physical constraints. If the original grade value after extrapolation is less than zero, the grade value is forcibly set to zero; if the original grade value after extrapolation is greater than 1, the grade value is forcibly set to 1. The physical range of the grade value is 0 to 1. Values ​​outside this range have no geological significance and must be truncated to ensure the physical rationality of the model output.

[0123] Finally, in step S106, the three-dimensional ore body parameter attribute model is verified using mining production data, and the parameters of the physical information neural network model are corrected through a backpropagation mechanism, including: Collect actual ore grade data from the mine for three consecutive months (or at least one complete mining cycle). This data comes from the actual production indicators of the concentrator, including the ore quantity, grade, and metal content of each mining block. The production data has temporal continuity and spatial representativeness, serving as the benchmark for the entire closed-loop verification. Match the production data with the 3D model block by block according to spatial location to ensure spatial consistency in the comparative analysis.

[0124] The predicted average grade of the 3D orebody parameter attribute model is compared with the actual mined ore grade data block by block. The relative error between the predicted average grade and the actual mined ore grade is calculated as follows: the error equals the absolute value of the difference between the predicted average grade and the actual mined ore grade divided by the actual mined ore grade, then multiplied by 100%. The comparative analysis covers all mined blocks, calculating the overall average error, maximum error, and error distribution characteristics of the ore grade prediction, and comprehensively evaluating the prediction accuracy of the 3D orebody parameter attribute model.

[0125] When the calculated prediction error is lower than or equal to a preset threshold, the 3D orebody parameter attribute model is considered to meet the mine production requirements. The 3D orebody parameter attribute model has been verified and can be directly used for subsequent resource estimation and mining design optimization. When the prediction error exceeds the preset threshold, a closed-loop update mechanism is triggered. In this embodiment, the preset threshold is set to 5%. This threshold is set based on the actual requirements of mine production for grade control accuracy. Different mines can adjust this threshold according to their actual circumstances.

[0126] After triggering the closed-loop update, a backpropagation mechanism is used to systematically trace the source of the prediction error and pinpoint its root cause. The backpropagation analysis follows this path: First, starting from the production verification stage, the difference patterns between the model prediction and actual production data are compared to determine whether the error is systematically high or low, or simply a spatial mismatch. Second, the modeling parameters, such as the variogram parameter settings, search neighborhood parameters, and the number of equally probable implementations, are checked for rationality. Next, the analysis traces back to the multi-scale fusion stage to check whether the fused feature curves fully reflect the boundary characteristics of the mineralized layer. Finally, the analysis traces back to the physical information neural network model to analyze whether there are deficiencies in the network structure, physical constraint settings, or training data. Common sources of error include: insufficient utilization of certain types of logging curve information leading to blurred orebody boundary predictions; improper setting of physical constraint weights causing the model to deviate from physical laws; biased selection of variogram parameters leading to spatial structure distortion; and insufficient representativeness of training data resulting in poor generalization ability of the physical information neural network model.

[0127] Based on the error sources identified by backpropagation analysis, the corresponding parameters of the physical information neural network model are adjusted accordingly.

[0128] Another aspect of this application provides a metal ore logging characterization system based on multi-scale fusion and PINN, which can be interpreted in correspondence with the above-mentioned metal ore logging characterization method based on multi-scale fusion and PINN. It includes: a data augmentation and repair module, which is used to perform data augmentation and curve repair on the original logging data based on generative adversarial network to generate a standardized dataset. The physical information neural network evaluation module is used to construct a physical information neural network model. The physical equations of the ore body are embedded as prior knowledge into the loss function of the physical information neural network model. The physical information neural network model is trained with a standardized dataset as input. The transfer learning module is used to transfer the common mapping knowledge of the physical information neural network model trained in the mature mining area to the new mining area using a transfer learning strategy, and to fine-tune the high-level parameters of the physical information neural network model using only a small amount of core grade analysis data from the new mining area. The multi-scale logging fusion module is used to perform wavelet multi-scale decomposition on various logging curves that have been repaired in the new mining area. It uses preset fusion rules to fuse coefficients at different scales and generates a multi-scale fusion feature curve containing geological information through a reconstruction algorithm. The multi-scale fusion feature curve is then input into a fine-tuned physical information neural network model to calculate one-dimensional orebody parameters. The 3D joint inversion modeling module is used to establish a 3D ore body parameter attribute model using one-dimensional ore body parameters as hard data and three-dimensional physical property data volume obtained by ground or well geophysical methods as soft data, and Bayesian sequential Gaussian simulation. The closed-loop verification and update module is used to verify the three-dimensional ore body parameter attribute model using mining production data, and to correct the parameters of the physical information neural network model through a backpropagation mechanism.

[0129] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for characterizing metal ore logging based on multi-scale fusion and PINN, characterized in that, include: Generative adversarial networks are used to augment and repair curves in raw well logging data, generating a standardized dataset. A physical information neural network model is constructed, and the density-volume fraction linear hybrid equation is used as a prior physical constraint and added to the loss function of the physical information neural network model as a physical loss term. The physical information neural network model is trained with a standardized dataset as input. By adopting a transfer learning strategy, the common mapping knowledge of the physical information neural network model trained in mature mining areas is transferred to new mining areas, and only a small amount of core grade analysis data from the new mining areas is used to fine-tune the high-level parameters of the physical information neural network model. Wavelet multi-scale decomposition was performed on various well logging curves that had been repaired in the new mining area. The coefficients of different scales were fused using a preset fusion rule. A multi-scale fusion feature curve containing geological information was generated by a reconstruction algorithm. The multi-scale fusion feature curve was then input into a finely tuned physical information neural network model to calculate the one-dimensional orebody parameters. Using one-dimensional orebody parameters as hard data and three-dimensional physical property data volume obtained by inversion from ground or well geophysical methods as soft data, a three-dimensional orebody parameter attribute model is established using Bayesian sequential Gaussian simulation. The three-dimensional ore body parameter attribute model was verified using mining production data, and the parameters of the physical information neural network model were corrected through backpropagation mechanism.

2. The method for characterizing metal ore logging based on multi-scale fusion and PINN according to claim 1, characterized in that, Generative Adversarial Networks (GANs) take undistorted well logging curves as input and distorted or missing well logging curves as reconstruction targets. The loss function of the GAN is... The norm reconstruction loss and the discriminant adversarial loss are weighted and summed. The norm reconstruction loss is used to constrain the point-by-point magnitude approximation of the generated curve to the true curve, while the discriminative adversarial loss is used to constrain the consistency between the overall probability distribution of the generated curve and the true curve.

3. The method for characterizing metal ore logging based on multi-scale fusion and PINN according to claim 1, characterized in that, The loss function of the physical information neural network model is a weighted sum of a data loss term and a physical loss term. The data loss term is the mean squared error between the predicted metal mineral volume fraction and the actual metal mineral volume fraction obtained from core chemical analysis, used to constrain the consistency between the physical information neural network model output and the measured label. The physical loss term is constructed based on a density-volume fraction linear hybrid equation, specifically the difference between the measured density logging value and the model-predicted density value. The norm, the model predicts density values ​​by linearly weighting the gangue mineral density, the metallic mineral density, and the metallic mineral volume fraction; the gangue mineral density and the metallic mineral density are used as trainable variables during training.

4. The method for characterizing metal ore logging based on multi-scale fusion and PINN according to claim 1, characterized in that, The transfer learning strategy is as follows: a physical information neural network model trained in a mature mining area is used as a source domain pre-training model. The network parameters are used as a carrier to transfer the common mapping knowledge between well logging response and ore body parameters to the new mining area. During the transfer, the parameters of the bottom feature extraction layer of the source domain pre-training model remain unchanged. Only the grade analysis data of a small number of core boreholes in the new mining area are used to fine-tune the high-level parameters. This allows the bottom feature extraction layer to retain the common mapping law between well logging response and mineralization characteristics learned in the source domain. The high-level parameters are then fine-tuned to adapt to the ore body parameter response characteristics of the new mining area.

5. The method for characterizing metal ore logging based on multi-scale fusion and PINN according to claim 4, characterized in that, When fine-tuning the high-level parameters in the new mining area, a small learning rate lower than the initial learning rate during training in the source domain is used to update the high-level parameters.

6. The method for characterizing metal ore logging based on multi-scale fusion and PINN according to claim 1, characterized in that, During the fine-tuning process, the similarity of the characteristic distribution of logging data from new mining areas and logging data from mature mining areas in the target layer is calculated. When the similarity is higher than a preset threshold, fine-tuning is performed. The similarity is quantitatively evaluated using the Pearson correlation coefficient, and migration weights are assigned to each source domain sample based on the similarity. The higher the similarity, the greater the assigned migration weight.

7. The method for characterizing metal ore logging based on multi-scale fusion and PINN according to claim 1, characterized in that, The wavelet multi-scale decomposition uses the Daubechies series wavelet basis, with 3 to 5 decomposition layers. Each logging curve is decomposed layer by layer into low-frequency approximation coefficients and multi-level high-frequency detail coefficients. The fusion rules are as follows: the high-frequency detail coefficients at each level are fused using the rule of taking the maximum absolute value, and the low-frequency approximation coefficients are fused using the weighted average rule. The fusion weights are allocated according to the sensitivity of each logging curve to the mineralized interval or the contribution rate of principal component analysis.

8. The method for characterizing metal ore logging based on multi-scale fusion and PINN according to claim 1, characterized in that, A three-dimensional orebody parameter and property model was established using Bayesian sequential Gaussian simulation, including: The original grade values ​​of the hard data points are transformed into standard normal distribution data by performing a normal distribution transformation. Variation function analysis based on standard normal distribution data: Experimental variation functions were calculated along the three main directions of the ore body strike, dip and vertical, and theoretical fitting was performed using a three-dimensional spherical model to determine the range, nugget effect and sill value in each direction; A three-dimensional regular grid is established based on the mining area and exploration accuracy. The search neighborhood of a three-dimensional ellipsoid is defined with each grid node to be estimated as the center. The radii of the three principal axes of the three-dimensional ellipsoid are equal to the range values ​​in each direction. A random access path is generated. At each grid node: the prior mean and prior variance are obtained by solving the simple Kriging equation based on the hard data and simulated values ​​in the neighborhood; soft data is extracted to calculate the soft predicted value and soft data variance to form a likelihood function; the prior mean, prior variance, and likelihood function are fused using Bayes' theorem to obtain a posterior distribution including the posterior mean and posterior variance; simulated values ​​are randomly sampled from the posterior distribution and added to the known dataset. Continue visiting the next grid node in the random path until all grid nodes have obtained simulated values; Perform an inverse Gaussian transform on the simulated values ​​of all grid nodes to restore the simulated values ​​back to the original grade scale and output a three-dimensional ore body parameter attribute model.

9. A metal ore logging characterization system based on multi-scale fusion and PINN, characterized in that, include: The data augmentation and repair module is used to perform data augmentation and curve repair on the original well logging data based on generative adversarial networks, and generate a standardized dataset. The physical information neural network evaluation module is used to construct a physical information neural network model. The physical equations of the ore body are embedded as prior knowledge into the loss function of the physical information neural network model. The physical information neural network model is trained with a standardized dataset as input. The transfer learning module is used to transfer the common mapping knowledge of the physical information neural network model trained in the mature mining area to the new mining area using a transfer learning strategy, and to fine-tune the high-level parameters of the physical information neural network model using only a small amount of core grade analysis data from the new mining area. The multi-scale logging fusion module is used to perform wavelet multi-scale decomposition on various logging curves that have been repaired in the new mining area. It uses preset fusion rules to fuse coefficients at different scales and generates a multi-scale fusion feature curve containing geological information through a reconstruction algorithm. The multi-scale fusion feature curve is then input into a fine-tuned physical information neural network model to calculate one-dimensional orebody parameters. The 3D joint inversion modeling module is used to establish a 3D ore body parameter attribute model using one-dimensional ore body parameters as hard data and three-dimensional physical property data volume obtained by ground or well geophysical methods as soft data, and Bayesian sequential Gaussian simulation. The closed-loop verification and update module is used to verify the three-dimensional ore body parameter attribute model using mining production data, and to correct the parameters of the physical information neural network model through a backpropagation mechanism.

Images