A mineral resource intelligent prediction method based on a liquid neural network

Abstract

The application discloses a mineral resource intelligent prediction method based on a liquid neural network, relates to the cross technical field of artificial intelligence and geological exploration, and comprises the following steps: reconstructing a time series for a space-time coordinate observation data set, generating a multi-dimensional continuous space-time field function, and calculating a spatial gradient field and a time derivative field of the multi-dimensional continuous space-time field function to generate a continuous space-time data stream; fusing a geological background static feature vector and an ore-forming dynamic feature vector, inputting the fused vector into a full connection layer connected with a liquid neural network branch and a convolutional neural network branch, mapping to generate a mineral resource prediction value, and simultaneously rendering a mineral resource quantity distribution map through spatial interpolation. The application inputs the continuous space-time data stream into the liquid neural network branch to evolve the internal state, generates the ore-forming dynamic feature vector, realizes high-resolution capture of a dynamic time sequence mode in an ore-forming process, and improves the accuracy and time sequence resolution of ore-forming dynamic process prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and geological exploration, and particularly relates to a mineral resource intelligent prediction method based on a liquid neural network. BACKGROUND

[0002] In the field of mineral resource intelligent prediction, conventional methods mainly rely on statistical analysis of multi-source geological data and traditional machine learning algorithms. These conventional methods usually integrate geochemical element content, geophysical measurement values and static geological map data, establish a statistical correlation model between mineralization potential and geological characteristics, and discretize dynamic observation data (such as time-series geochemical data) or simply splice static geological data (such as lithology distribution) to realize mineral resource prediction by using classification and regression models, aiming to optimize the efficiency of delineating ore prospecting target areas through data-driven methods.

[0003] However, the traditional method usually discretizes continuous time series data into fixed time interval samples in dynamic data processing, which cannot fully capture the continuous evolution characteristics in the geological process. In the feature fusion mechanism, shallow fusion strategies (such as direct splicing and linear weighting) are often used, which cannot effectively mine the deep association between dynamic and static data, and may lose key cross-modal features. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a mineral resource intelligent prediction method based on a liquid neural network to solve the problems of being difficult to effectively capture the continuous evolution law of the geological process and mine the deep association between dynamic and static data.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The application provides a mineral resource intelligent prediction method based on a liquid neural network, which comprises the following steps: collecting dynamic observation data and static geological data of a mineral research area, giving the dynamic observation data a continuous geological time coordinate, and generating a space-time coordinate observation data set; performing time series reconstruction on the space-time coordinate observation data set, generating a multi-dimensional continuous space-time field function, and calculating a spatial gradient field and a time derivative field of the multi-dimensional continuous space-time field function to generate a continuous space-time data stream; inputting the continuous space-time data stream into a liquid neural network branch to evolve the internal state, generating an ore-forming dynamic feature vector, and inputting the static geological data into a convolutional neural network branch to extract a spatial invariance feature, generating a geological background static feature vector; fusing the geological background static feature vector and the ore-forming dynamic feature vector, and inputting them into a full connection layer connected with the liquid neural network branch and the convolutional neural network branch to generate a mineral resource prediction value, and meanwhile, rendering a mineral resource quantity distribution map through spatial interpolation; performing binary segmentation on the mineral resource quantity distribution map to delineate the spatial range of a prospecting target area, and analyzing the evolution sequence of the liquid neural network hidden state with geological time to generate an ore-forming process analysis report.

[0008] As a preferred scheme of the mineral resource intelligent prediction method based on the liquid neural network, the dynamic observation data of the mineral research area comprises geochemical element content data and geophysical measurement data.

[0009] The static geological data comprises basic geological maps, remote sensing image data and terrain relief grid data.

[0010] As a preferred scheme of the mineral resource intelligent prediction method based on the liquid neural network, the step of giving the dynamic observation data a continuous geological time coordinate to generate a space-time coordinate observation data set comprises the following steps:

[0011] Collecting stratum sequences and isotope age data of the mineral research area to establish a continuous geological time mapping framework, and giving each dynamic observation point in the dynamic observation data a time coordinate according to the continuous geological time mapping framework to generate a space-time coordinate observation point set;

[0012] Performing spatial gridding interpolation on the space-time coordinate observation point set through a Kriging interpolation algorithm to generate a space-time coordinate observation data set.

[0013] As a preferred scheme of the mineral resource intelligent prediction method based on the liquid neural network, the step of performing time series reconstruction on the space-time coordinate observation data set to generate a multi-dimensional continuous space-time field function comprises the following steps:

[0014] Performing continuous representation on the space-time coordinate observation data set in the time dimension through a Gaussian process regression method to fit and generate preliminary reconstructed space-time sequence data;

[0015] Fusing and spatial gridding the preliminary reconstructed spatiotemporal sequence data and three-dimensional spatial coordinates to form a multi-dimensional continuous spatiotemporal field function.

[0016] As a preferred scheme of the mineral resource intelligent prediction method based on the liquid neural network, the step of calculating the spatial gradient field and the time derivative field of the multi-dimensional continuous spatiotemporal field function to generate a continuous spatiotemporal data stream is as follows,

[0017] Deriving the multi-dimensional continuous spatiotemporal field function in a spatial direction to obtain spatial gradient field data;

[0018] Performing a time direction partial derivative operation on the multi-dimensional continuous spatiotemporal field function to generate time derivative field data;

[0019] Splicing and standardizing the multi-dimensional continuous spatiotemporal field function, the spatial gradient field data and the time derivative field data in a feature dimension to generate a continuous spatiotemporal data stream.

[0020] As a preferred scheme of the mineral resource intelligent prediction method based on the liquid neural network, the step of inputting the continuous spatiotemporal data stream into a liquid neural network branch to evolve an internal state to generate an ore-forming dynamic feature vector is as follows,

[0021] Inputting the continuous spatiotemporal data stream into the liquid neural network branch to continuously evolve an internal state of a neuron along a geological time axis driven by a system of ordinary differential equations to generate a dynamic evolution state sequence;

[0022] Aggregating features of the dynamic evolution state sequence to extract a time sequence pattern with geological significance to generate an ore-forming dynamic feature vector.

[0023] As a preferred scheme of the mineral resource intelligent prediction method based on the liquid neural network, the step of inputting static geological data into a convolutional neural network branch to extract a spatial invariance feature to generate a geological background static feature vector is as follows,

[0024] Inputting the static geological data into the convolutional neural network branch to extract multi-level spatial features through a convolutional layer and a pooling layer to generate a spatial feature map group;

[0025] Performing global average pooling and feature compression on the spatial feature map group to generate a geological background static feature vector.

[0026] As a preferred scheme of the mineral resource intelligent prediction method based on the liquid state neural network, the geological background static feature vector and the mineralization dynamic feature vector are fused and input into the full connection layer connected by the liquid state neural network branch and the convolutional neural network branch, and a mineral resource prediction value is generated by mapping.

[0027] Based on the geological background static feature vector and the mineralization dynamic feature vector, cross attention weight is calculated, and feature weighted fusion is performed to generate a spatio-temporal fusion feature vector.

[0028] The spatio-temporal fusion feature vector is input into the full connection layer connected by the liquid state neural network branch and the convolutional neural network branch, nonlinear transformation and feature dimension reduction are performed, and a mineral resource prediction value is output.

[0029] As a preferred scheme of the mineral resource intelligent prediction method based on the liquid state neural network, the mineral resource quantity distribution map is generated by spatial interpolation rendering, and the steps are as follows.

[0030] The mineral resource prediction value is calculated by gridding to generate mineral resource density grid data.

[0031] The mineral resource density grid data is processed by the contour tracking algorithm to generate a mineralization intensity contour, and the contour is rendered in combination with a geographic base map to generate a mineral resource quantity distribution map.

[0032] As a preferred scheme of the mineral resource intelligent prediction method based on the liquid state neural network, the mineral resource quantity distribution map is binarized and segmented to delineate the spatial range of the prospecting target area, and the evolution sequence of the liquid state neural network hidden state with geological time is analyzed to generate a mineralization process analysis report, and the steps are as follows.

[0033] The mineral resource quantity distribution map is binarized and segmented according to a preset mineralization probability threshold to generate a mineral resource quantity segmentation map.

[0034] The mineral resource quantity segmentation map is subjected to spatial clustering analysis to identify the connected region of the minimum target area, and the target area spatial range vector data is output.

[0035] The evolution sequence of the liquid state neural network hidden state with geological time is analyzed, and the time sequence characteristic parameters of the evolution trajectory are calculated to generate mineralization dynamic evolution sequence data.

[0036] The target area spatial range vector data and the mineralization dynamic evolution sequence data are integrated to generate a mineralization process analysis report.

[0037] The present application has the beneficial effects that: by inputting continuous spatiotemporal data flow into the liquid neural network branch to evolve the internal state, a mineralization dynamic feature vector is generated, high-resolution capture of dynamic time sequence patterns in the mineralization process is realized, and the accuracy and time sequence resolution of the prediction of the mineralization dynamic process are improved; by inputting static geological data into the convolutional neural network branch to extract spatial invariance features, a geological background static feature vector is generated, multi-scale spatial representation of lithology distribution and tectonic boundaries is realized, and adaptability and feature robustness of the prediction model to complex geological environments are enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0039] Fig. 1 Flowchart of the mineral resource intelligent prediction method based on liquid neural network.

[0040] Fig. 2 Flowchart for generating continuous spatiotemporal data flow.

[0041] Fig. 3 Flowchart for generating spatiotemporal fusion feature vector.

[0042] Fig. 4 Flowchart for generating mineralization process analysis report. DETAILED DESCRIPTION

[0043] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0044] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0045] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0046] REFERENCE Figs. 1-4For an embodiment of the present application, the embodiment provides a liquid neural network-based intelligent prediction method for mineral resources, comprising the following steps:

[0047] S1, collecting dynamic observation data and static geological data of the mineral research area, and giving continuous geological time coordinates to the dynamic observation data to generate a spatio-temporal coordinate observation data set.

[0048] S1.1, the dynamic observation data of the mineral research area includes geochemical element content data and geophysical measurement data.

[0049] It should be noted that the geochemical element content data is a data set formed by collecting natural substances (such as rocks, soils and stream sediments) in the mineral research area and determining the content of various chemical elements; the geophysical measurement data is a data set formed by measuring the distribution and change of the geophysical field (such as gravity field, magnetic field, electric field and radioactive field) through various instruments (such as absolute gravimeter, optical pump magnetometer and airborne gamma spectrometer) to infer the underground geological structure and mineral information.

[0050] S1.2, the static geological data of the mineral research area includes basic geological map, remote sensing image data and terrain relief grid data.

[0051] It should be noted that the basic geological map is formed by field observation, identification and delineation of stratigraphic boundaries, rock body boundaries, fault lines and fold axis geology of different ages, combined with indoor research means such as rock thin section identification, paleontological fossil analysis and isotope dating, to determine the age and nature of the geological body, and draw according to international standard legend and scale (such as 1:50,000); remote sensing image data is a digital image of different wave bands generated by receiving electromagnetic wave signals reflected and radiated by satellite and airplane-mounted multi-spectral and hyperspectral sensors; the terrain relief grid data is a high-precision digital elevation model (DEM) generated by laser radar by emitting laser pulses to the ground and receiving echoes, and combining the positioning and orientation data of the sensor to accurately calculate the distance between the sensor and the ground.

[0052] S1.3, collecting stratigraphic sequence and isotope age data of the mineral research area to establish a continuous geological time mapping framework, and giving time coordinates to each dynamic observation point in the dynamic observation data according to the continuous geological time mapping framework to generate a spatio-temporal coordinate observation point set.

[0053] Further, the stratigraphic column, profile and isotope dating report of the mineral research area are collected, and the stratigraphic sequence from new to old and the corresponding isotope absolute age data are extracted to establish a normalized continuous geological time axis from the newest stratum to the oldest stratum, that is, a continuous geological time mapping framework; the known isotope age of the sampling surrounding rock corresponding to each dynamic observation point in the dynamic observation data is matched with the continuous geological time mapping framework, a continuous geological time coordinate value ranging from 0 to 1 is given to each dynamic observation point, and the continuous geological time coordinate value is combined with the inherent geographical coordinate of the dynamic observation point to generate a set of space-time coordinate observation points containing longitude, latitude, altitude and geological time coordinate.

[0054] S1.4, the space-time coordinate observation data set is generated by spatial gridding interpolation of the space-time coordinate observation point set through the Kriging interpolation algorithm.

[0055] Further, according to the spatial range of the mineral research area, the area size of the regular grid (such as 100m x 100m) is set to cover the entire mineral research area; for each unsampled node in the regular grid, search for space-time coordinate observation points within a certain distance range (such as a radius of 500m), construct a Kriging equation set according to the spatial distance between the space-time coordinate observation points and the unsampled node, and the spatial distance and orientation relationship between the space-time coordinate observation points, combined with the spatial autocorrelation of the attribute value described by the semi-variogram (for example, within the search radius, the space-time coordinate observation points with closer distance are considered to have higher spatial correlation, and the corresponding attribute value is more similar); solve the Kriging equation set to obtain the Kriging weight of each space-time coordinate observation point, and use the Kriging weight to perform weighted average on the attribute values of the adjacent space-time coordinate observation points to calculate the attribute value of the unsampled node; integrate the spatial coordinates, geological time coordinates and attribute values of each unsampled node to form a space-time coordinate observation data set.

[0056] S2, the time series reconstruction of the space-time coordinate observation data set is performed to generate a multi-dimensional continuous space-time field function, and the spatial gradient field and the time derivative field of the multi-dimensional continuous space-time field function are calculated to generate a continuous space-time data stream.

[0057] S2.1, the space-time coordinate observation data set is continuously represented in the time dimension by using the Gaussian process regression method to fit and generate preliminary reconstructed space-time sequence data.

[0058] Furthermore, based on the geological time coordinates and corresponding attribute values ​​of each regular grid node in the spatiotemporal coordinate observation dataset, a Gaussian process is defined with geological time coordinates as input and attribute values ​​as output. A covariance function (such as a radial basis function) is selected for the Gaussian process to characterize the correlation of attribute values ​​in the geological time dimension. Using the known geological time coordinates and attribute values ​​in the spatiotemporal coordinate observation dataset, the hyperparameters of the Gaussian process are optimized by maximizing the marginal likelihood function. The optimized Gaussian process is then used to calculate the posterior mean of the attribute value corresponding to each geological time point at a series of dense and continuous geological time points. These values ​​are then arranged and combined in geological time order to form an attribute sequence that changes continuously in the time dimension. The Gaussian process regression prediction is repeatedly executed on each regular grid node in the spatiotemporal coordinate observation dataset to generate preliminary reconstructed spatiotemporal sequence data that covers all regular grid nodes in the study area and is continuously represented in the time dimension.

[0059] The mathematical expression for the covariance function of Gaussian process regression is:

[0060] ;

[0061] in, It is a geological time coordinate; It is the covariance function value, describing the time point. and The correlation of attribute values; It is a length scale parameter that controls the rate at which the covariance decays with distance, and is obtained by maximizing the marginal likelihood function. It is the signal variance parameter, representing the overall degree of variation in attribute values, and is also determined by maximizing the marginal likelihood function.

[0062] It should be noted that the radial basis function (RBF) is a kernel function based on spatial distance as a measure of similarity. The principle is that the output value of the RBF depends only on the norm of the Euclidean distance between two points. It typically exhibits a monotonically changing characteristic: the closer the distance, the larger the output value of the RBF; the farther the distance, the greater the output value of the RBF. In Gaussian process regression, the RBF calculates the normalized time distance between different geological time points, mapping discrete time coordinates to continuous correlation strength, thereby constructing a covariance matrix that describes the smooth change of attribute values ​​over time, and realizing the reconstruction from discrete observations to continuous time series.

[0063] S2.2. The preliminary reconstructed spatiotemporal sequence data and three-dimensional spatial coordinates are fused and spatially gridded to form a multidimensional continuous spatiotemporal field function.

[0064] Further, the three-dimensional spatial coordinates (longitude, latitude, and elevation) and the attribute values corresponding to the continuous time series of each regular grid node are extracted from the preliminary reconstructed spatiotemporal sequence data, and are bound one by one to ensure that each spatial position is associated with a complete continuous time series; for each regular grid node that has completed binding, a mathematical function expression is constructed, taking continuous geological time as the only input variable and outputting the attribute values at the regular grid node over time, and the mathematical function expression can obtain the attribute values corresponding to any geological time point through interpolation and function calculation; the function set established on all regular grid nodes covering the entire study area, taking three-dimensional spatial coordinates as the identifier and continuous geological time as the variable, is collectively defined as a multi-dimensional continuous spatiotemporal field function.

[0065] S2.3, spatial direction derivation is performed on the multi-dimensional continuous spatiotemporal field function to obtain spatial gradient field data.

[0066] Further, for the multi-dimensional continuous spatiotemporal field function, at each regular grid node, a specific geological time point is fixed, the attribute values of the adjacent regular grid nodes in the east-west direction and the north-south direction at the same geological time point are extracted, and the central difference method is used to calculate the attribute value change rate of the grid node in the east-west direction, i.e., the longitude direction partial derivative, and the attribute value change rate of the regular grid node in the north-south direction, i.e., the latitude direction partial derivative; the longitude direction partial derivative and the latitude direction partial derivative are combined into a two-dimensional vector, which is the spatial gradient of the regular grid node at the current geological time point; the spatial direction derivation operation is repeatedly performed on all regular grid nodes and all geological time points covered by the multi-dimensional continuous spatiotemporal field function to generate spatial gradient field data corresponding to the dimensions of the multi-dimensional continuous spatiotemporal field function.

[0067] The expression for calculating the spatial gradient is:

[0068] ;

[0069] wherein, represents the multi-dimensional continuous spatiotemporal field function; is the longitude coordinate and the latitude coordinate of the regular grid node; is the grid spacing in the east-west direction, and to ensure the self-consistency of the derivative unit, the longitude and latitude coordinates are converted into a plane rectangular coordinate system (such as UTM projection) before calculation, which is the actual distance (in meters) of the adjacent grid nodes in the east-west direction in the plane rectangular coordinate system; and respectively represent the attribute values at the two regular grid nodes adjacent to the current regular grid node in the east-west direction; represents the index number of the regular grid node in the east-west direction (longitude direction); Index number of the regular grid node in the north-south direction (latitude direction).

[0070] S2.4, time direction partial derivative operation is performed on the multi-dimensional continuous space-time field function to generate time derivative field data.

[0071] Further, a specific time point of each regular grid node in the multi-dimensional continuous space-time field function on the continuous geological time axis is selected, and the attribute values corresponding to the time points before and after the specific time point of the regular grid node are extracted. The central difference method is used to calculate the rate of change of the attribute values with time, i.e. the time direction partial derivative; the time direction partial derivative operation is repeatedly performed on all regular grid nodes and all geological time points covered by the multi-dimensional continuous space-time field function to generate time derivative field data.

[0072] S2.5, the multi-dimensional continuous space-time field function, the spatial gradient field data and the time derivative field data are spliced and standardized in feature dimensions to generate a continuous space-time data stream.

[0073] Further, the attribute values of the multi-dimensional continuous space-time field function at each regular grid node and geological time point, the gradient vectors of the spatial gradient field data at the regular grid nodes and geological time points, and the derivative values of the time derivative field data at the regular grid nodes and geological time points are extracted, and are directly spliced in feature dimensions to form a composite feature vector; the composite feature vector is subjected to Z-score standardization processing, so that the numerical value of each feature dimension is 0 and the standard deviation is 1, and is arranged and combined according to the geological time sequence to generate a continuous space-time data stream.

[0074] S3, the continuous space-time data stream is input to the liquid neural network branch to evolve the internal state, generate a metallogenic dynamic feature vector, and input the static geological data to the convolutional neural network branch to extract spatial invariance features, generate a geological background static feature vector.

[0075] S3.1, the continuous space-time data stream is input to the liquid neural network branch, the internal state of the neuron is continuously evolved along the geological time axis driven by the system of ordinary differential equations, and a dynamic evolution state sequence is generated.

[0076] Further, the continuous spatiotemporal data stream is input into the liquid neural network branch in the order of geological time step by step, the liquid neural network branch receives the corresponding input data at each geological time point, and calculates the instantaneous change rate of the neuron internal state at the current time point according to the input data and the neuron internal state at the current time point through the predefined system of ordinary differential equations; the neuron internal state value at the next geological time point is obtained by updating the neuron internal state through the Runge-Kutta method using the instantaneous change rate; the updated neuron internal state value of the liquid neural network branch at each geological time point is recorded and arranged in the order of geological time to form a dynamic evolution state sequence.

[0077] It should be noted that the principle of the Runge-Kutta method is to estimate the increment of the function at the next step by selecting multiple slope sampling points (such as four points for the fourth-order method) within the interval defined by the ordinary differential equation and performing weighted averaging, thereby improving the accuracy and stability of the numerical integration of the ordinary differential equation. The above numerical solution of the system of ordinary differential equations describing the state evolution of the liquid neural network neuron is performed by the Runge-Kutta method. In each geological time step, the instantaneous change rate of the neuron state is calculated through multiple stages and weighted integration, thereby updating the internal state of the neuron with high precision and driving the internal state of the neuron to evolve along the continuous geological time axis.

[0078] S3.2, the dynamic evolution state sequence is aggregated for feature, and a time sequence pattern with geological significance is extracted to generate a mineralization dynamic feature vector.

[0079] Further, the neuron internal state value at each geological time point in the dynamic evolution state sequence is calculated for the statistical characteristics changing with time, including the mean, variance, maximum and minimum; the dynamic evolution state sequence is subjected to principal component analysis transformation, the covariance matrix is calculated and the eigenvalues and eigenvectors are solved, and the principal component corresponding to the eigenvector with the largest eigenvalue is selected as the principal component component; the dynamic evolution state sequence is subjected to Fourier transform, the time domain signal is converted into frequency domain power spectrum, and the peaks with amplitudes higher than the background noise in the frequency domain power spectrum are screened out, and the period corresponding to the peak is calculated as the frequency feature; the mean, variance, maximum, minimum, principal component component and frequency feature are integrated to generate a mineralization dynamic feature vector.

[0080] S3.3, the static geological data is input into the convolutional neural network branch, and multi-level spatial features are extracted through convolutional layers and pooling layers to generate a spatial feature map group.

[0081] Further, the static geological data is input to the first layer of the convolutional layer of the convolutional neural network branch, a 3x3 size of the convolution kernel is used to slide window scanning on the static geological data with a step of 1, and a convolution operation is performed at each window position to generate a first layer feature map containing local spatial features; the first layer feature map is input to the pooling layer, a 2x2 maximum pooling operation is performed in the pooling layer, the maximum value in each non-overlapping 2x2 area is selected as the output, and the size of the first layer feature map is reduced by half to generate a pooled feature map; the pooled feature map is input to the subsequent convolutional layer and pooling layer, and convolution operation and pooling operation are performed layer by layer, each layer of convolutional layer uses a convolution kernel of different size to extract spatial features of different scales, and each layer of pooling layer gradually reduces the size of the feature map; after the alternating processing of multiple layers of convolutional layer and pooling layer, a group of spatial feature maps with different abstract levels and spatial scales is generated.

[0082] S3.4, global average pooling and feature compression are performed on the spatial feature map group to generate a geological background static feature vector.

[0083] Further, a global average pooling operation is performed on each spatial feature map in the spatial feature map group, and the arithmetic mean of all pixel values of each spatial feature map is calculated, so that each spatial feature map is compressed into a scalar value; the scalar values obtained by global average pooling of all spatial feature maps in the spatial feature map group are spliced into an initial feature vector in order, and linear transformation and nonlinear activation are performed through a fully connected layer to realize feature dimension reduction and information condensation, so that the high-dimensional initial feature vector is mapped to a lower-dimensional vector space, and a geological background static feature vector is output.

[0084] S4, the geological background static feature vector and the ore-forming dynamic feature vector are fused and input to a fully connected layer connected by the liquid neural network branch and the convolutional neural network branch, and a mineral resource prediction value is generated by mapping, and a mineral resource distribution map is generated by spatial interpolation rendering.

[0085] S4.1, based on the geological background static feature vector and the ore-forming dynamic feature vector, cross attention weight is calculated, and feature weighted fusion is performed to generate a spatio-temporal fusion feature vector.

[0086] Further, the geological background static feature vector is mapped into a query vector and the ore-forming dynamic feature vector is mapped into a key vector through linear transformation (linear transformation is realized through matrix multiplication); the matrix product of the query vector and the transpose of the key vector is calculated to obtain an original attention score matrix reflecting the correlation between features, and each element in the original attention score matrix is numerically scaled according to the proportional relationship between the original attention score matrix and the square root of the corresponding dimension size of the key vector; the Softmax function is applied to each row of the scaled attention score matrix for normalization processing to ensure that the sum of each row element is 1, and a cross-attention weight matrix is obtained; the ore-forming dynamic feature vector is mapped into a value vector again through linear transformation, and the cross-attention weight matrix is used for weighted summation of the value vector to obtain a weighted dynamic feature representation; the weighted dynamic feature representation is directly spliced with the geological background static feature vector in the feature dimension, and linear transformation is performed on the spliced feature vector to integrate information and adjust the dimension, thereby generating a spatio-temporal fusion feature vector.

[0087] S4.2, input the spatio-temporal fusion feature vector into the fully connected layer connected by the liquid neural network branch and the convolutional neural network branch, perform nonlinear transformation and feature dimension reduction, and output a mineral resource prediction value.

[0088] It should be noted that the embodiment adopts a hybrid neural network to predict mineral resources, and the hybrid neural network is pre-trained. The specific operation is as follows. The labeled multi-element geological data (including dynamic observation data sequence and static geological data) extracted from the historical mineral database is used as a training sample, the training sample is randomly divided into a training set and a validation set according to a fixed ratio (7:3), in the training set, the dynamic observation data of each training sample is normalized and missing value is filled, the static geological data is scaled and standardized and the format is unified, a regular multi-element geological data tensor is generated, and the mineralization probability value is converted into a binary label as the output target of the hybrid neural network; the regular dynamic observation data sequence is input into the liquid neural network branch, the dynamic characteristic evolution is calculated through numerical integration of ordinary differential equations, and the static geological data is input into the convolutional neural network branch, the spatial features are extracted through convolution and pooling operations; the output features of the two branches are spliced and then passed through a fully connected layer, the Sigmoid activation function is used to calculate the mineralization prediction probability, and the error between the mineralization prediction probability and the true label is compared based on the binary cross-entropy loss function; the Adam optimizer is used for back propagation, the differential equation parameters of the liquid neural network branch are updated using the adjoint sensitivity method, and the convolution layer parameters of the convolutional neural network branch are updated using the standard error back propagation; the forward inference is periodically performed on the validation set, the prediction accuracy and the loss value of the hybrid neural network on the validation set are recorded, the training is terminated when the validation loss does not decrease in a plurality of consecutive training cycles (such as 10 training cycles), and the pre-training of the hybrid neural network is completed.

[0089] It should be noted that the hybrid neural network specifically refers to a composite architecture integrated by the liquid neural network branch and the convolutional neural network branch through a common fully connected layer, wherein the liquid neural network branch is responsible for processing the continuous evolution features of the dynamic spatio-temporal data stream, the convolutional neural network branch is responsible for extracting the spatial invariance features of the static geological data, and finally the feature fusion and prediction output are realized through the fully connected layer.

[0090] Further, the spatio-temporal fusion feature vector is input into the fully connected layer connected by the liquid neural network branch and the convolutional neural network branch, the matrix multiplication operation is performed between the weight matrix of the fully connected layer and the spatio-temporal fusion feature vector to generate a linear transformation feature vector; the linear transformation feature vector is subjected to nonlinear transformation through the ReLU activation function, the negative values are set to zero and the positive values are retained to generate a 256-dimensional nonlinear feature vector; the 256-dimensional nonlinear feature vector is input into another fully connected layer for feature dimension reduction, the feature dimension is compressed from 256 dimensions to 64 dimensions through weight matrix multiplication and bias addition to generate a reduced dimension feature vector; the reduced dimension feature vector is subjected to output transformation through the Sigmoid activation function, the numerical value is compressed to the range of 0 to 1 to generate a mineral resource prediction value.

[0091] S4.3, grid calculation is performed on the mineral resource prediction value to generate mineral resource density grid data.

[0092] Further, based on the established rule grid covering the mineral research area, the spatial coordinates of each rule grid node are associated with the corresponding mineral resource prediction value to form a set of spatial position-prediction value data points; the Kriging interpolation algorithm is used to perform spatial interpolation calculation on the discrete data points in the set of spatial position-prediction value data points, and the mineral resource prediction value of each unsampled grid node is estimated based on the spatial autocorrelation principle; the spatial coordinates of all grid nodes and the mineral resource prediction value obtained by interpolation are integrated to generate mineral resource density grid data covering the entire mineral research area and continuously distributed in space.

[0093] S4.4, contour tracing algorithm processing is performed on the mineral resource density grid data to generate mineralization intensity contours, and the rendering is performed in combination with the geographic base map to generate a mineral resource quantity distribution map.

[0094] Further, a mineralization intensity threshold (such as 0.3, 0.5 and 0.7) is set, and each grid region is scanned from left to right and from bottom to top starting from the lower left corner grid of the mineral resource density grid data, and the intersection coordinates equal to the mineralization intensity threshold on the grid edge are calculated by linear interpolation; when the intersection point is found, the intersection point is taken as the starting point of the contour, the contour direction is determined according to the density values of adjacent grids, and the continuous curve is connected by connecting the intersection points, until the contour is closed or reaches the data boundary, to generate the mineralization intensity contour; after the contour tracing of all mineralization intensity thresholds is completed, the generated mineralization intensity contours are matched with the geographic base map in coordinates to ensure accurate spatial position correspondence, and the mineralization intensity contours generated for different mineralization intensity thresholds are set with colors and line types (such as 0.3 with light yellow dashed line, 0.5 with orange solid line, and 0.7 with red thick solid line); all the mineralization intensity contours are superimposed and drawn on the geographic base map, and a legend and a scale are added for explanation and annotation to generate a mineral resource quantity distribution map.

[0095] It should be noted that the mineralization intensity threshold is a numerical value used to divide the critical points of different levels of mineral resource prediction value, which is based on the statistical distribution characteristics of the mineral resource prediction value of the known deposits in the mineral research area, the cumulative frequency distribution curve is calculated, the inflection point positions with geological significance are selected on the cumulative frequency curve, such as the mineral resource prediction values corresponding to the 25%, 50% and 75% percentile points, and the numerical values of the initially selected inflection point positions are adjusted by referring to the regional metallogenic regularity and deposit industrial index requirements to determine a set of threshold sequence with clear geological meaning and reasonable interval, such as 0.3, 0.5 and 0.7, which correspond to low, medium and high mineralization intensity levels, respectively.

[0096] S5.1, binarize the mineral resource quantity distribution map according to the preset mineralization probability threshold to generate a mineral resource quantity segmentation map.

[0097] S5.1, binarize the mineral resource quantity distribution map according to the preset mineralization probability threshold to generate a mineral resource quantity segmentation map.

[0098] Further, each pixel point in the mineral resource quantity distribution map is traversed, and the mineral resource prediction value corresponding to the pixel point is compared with the preset mineralization probability threshold: when the mineral resource prediction value of the pixel point is greater than or equal to the mineralization probability threshold, the value of the pixel point is set to 1; when the mineral resource prediction value of the pixel point is less than the mineralization probability threshold, the value of the pixel point is set to 0, and a mineral resource quantity segmentation map is generated.

[0099] It should be noted that the mineralization probability threshold is obtained by collecting the mineral resource prediction values of all known ore deposit locations to form an ore deposit sample set, and collecting the mineral resource prediction values away from the known ore deposit area to form a background sample set. The mean and standard deviation of the ore deposit sample set and the mean and standard deviation of the background sample set are calculated, a hypothesis test (such as t-test) is used to determine the confidence interval of the significant difference between the two distributions, and the value corresponding to the cumulative frequency of the mineral resource prediction value in the ore deposit sample set is selected at a high confidence level (such as 90%). The example value range is 0.5 to 0.7. If the value is higher than 0.7, a large number of real mineralization anomalies will be missed, significantly reducing the success rate of prospecting. If it is lower than 0.5, too many barren areas will be selected as target areas, significantly increasing the exploration cost and reducing the work efficiency.

[0100] S5.2, perform spatial clustering analysis on the mineral resource quantity segmentation map to identify the minimum target area connected region, and output the target area spatial range vector data.

[0101] Further, all connected regions with pixel value 1 in the mineral resource quantity segmentation map are labeled, and an eight-neighbor connectivity criterion is used to traverse the image, and spatially adjacent pixel points are merged into the same connected region. The number of pixels in each connected region is calculated, the actual area is converted from the number of pixels, and the area of each connected region is compared with the preset minimum target area threshold to select the connected regions with area greater than or equal to the minimum target area threshold. For each connected region meeting the area requirement, the vertex coordinates of the outer contour are extracted, and the target area spatial range vector data is generated in clockwise order.

[0102] It should be noted that the minimum target area threshold is determined by collecting the planar projection area data of all known economic deposits in the mineral resource area, forming a deposit area sample set, and calculating the low percentile area value as a reference benchmark by statistical analysis of the area distribution characteristics of the deposit area sample set. Combined with the economic and technical conditions of exploration engineering deployment (such as single drill hole exploration radius and exploration line spacing engineering parameters), the minimum economic area region that can be implemented for exploration operation is determined. Finally, the reference benchmark and the minimum economic area region are comprehensively weighed, and the larger value of the two is set. The exemplary value range is 0.5-2 square kilometers. If the value is higher than 2 square kilometers, it is easy to miss small ore bodies with economic value, reducing the success rate of prospecting. If it is lower than 0.5 square kilometers, it is easy to enclose a large number of isolated anomalies without economic value, increasing the exploration cost and reducing the efficiency of prospecting.

[0103] S5.3, analyze the evolution sequence of the liquid neural network hidden state with geological time, and calculate the time sequence characteristic parameters of the evolution trajectory to generate the ore-forming dynamic evolution sequence data.

[0104] Further, the hidden state vector of the liquid neural network at each geological time point is extracted from the dynamic evolution state sequence, the first-order difference of each hidden state vector is calculated to obtain the state change quantity sequence, and the mean, variance, skewness and kurtosis statistics of the hidden state vector are calculated to generate the hidden state statistical feature vector. The autocorrelation of the hidden state under different time lags is analyzed, and the autocorrelation coefficients from lag 1 to lag 5 are calculated. The hidden state vector is subjected to fast Fourier transform, the frequencies corresponding to the top three peak values in the frequency spectrum are identified, and the corresponding amplitudes are recorded. The state change quantity sequence, hidden state statistical feature vector, autocorrelation coefficient, three main frequencies and amplitudes are integrated in chronological order to generate the ore-forming dynamic evolution sequence data.

[0105] S5.4, integrate the target area spatial range vector data and the ore-forming dynamic evolution sequence data to generate the ore-forming process analysis report.

[0106] Further, the polygon boundary vertex coordinates recorded in the target area spatial range vector data are read, and the hidden state statistical feature vector, autocorrelation coefficient, three main frequencies and amplitudes recorded in the ore-forming dynamic evolution sequence data are read. The polygon vertex coordinates of each target area spatial range are matched with the corresponding ore-forming dynamic evolution sequence data through geographic coordinates to establish the correspondence between the target area location and the evolution characteristics, and the report is organized according to the standard report format to generate the ore-forming process analysis report containing the target area positioning map and the feature data table.

[0107] In summary, the application realizes high-resolution capture of dynamic time sequence patterns in the mineralization process and improves the accuracy and time sequence resolution of the prediction of the mineralization dynamic process by inputting continuous spatiotemporal data flow into the liquid state neural network branch to evolve the internal state and generate a mineralization dynamic feature vector.

[0108] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A liquid neural network-based intelligent prediction method for mineral resources, characterized by: The application relates to a mineral resource prediction method based on liquid-state neural network and convolutional neural network. Dynamic observation data and static geological data of a mineral research area are collected, and continuous geological time coordinates are given to the dynamic observation data to generate a time-space coordinate observation data set; A time series reconstruction is performed on the time-space coordinate observation data set to generate a multi-dimensional continuous time-space field function, and a spatial gradient field and a time derivative field of the multi-dimensional continuous time-space field function are calculated to generate a continuous time-space data stream; The continuous time-space data stream is input into a liquid-state neural network branch to perform internal state evolution, and a mineralization dynamic feature vector is generated; and the static geological data is input into a convolutional neural network branch to perform spatial invariance feature extraction, and a geological background static feature vector is generated; The geological background static feature vector and the mineralization dynamic feature vector are fused and input into a full connection layer connected with the liquid-state neural network branch and the convolutional neural network branch to generate a mineral resource prediction value, and a mineral resource quantity distribution map is generated through spatial interpolation rendering; The mineral resource quantity distribution map is binarized and segmented to determine the spatial range of a prospecting target area, and an evolution sequence of a liquid-state neural network hidden state along with geological time is analyzed to generate a mineralization process analysis report.

2. The liquid neural network-based intelligent prediction method for mineral resources according to claim 1, characterized in that: The dynamic observation data of the mineral research area include geochemical element content data and geophysical measurement data; The static geological data include basic geological maps, remote sensing image data and terrain relief grid data.

3. The liquid neural network-based intelligent prediction method for mineral resources according to claim 1, characterized in that: The continuous geological time coordinates are given to the dynamic observation data to generate the time-space coordinate observation data set in the following steps, Stratigraphic sequences and isotope age data of the mineral research area are collected to establish a continuous geological time mapping framework, and time coordinates are given to each dynamic observation point in the dynamic observation data according to the continuous geological time mapping framework to generate a time-space coordinate observation point set; The time-space coordinate observation point set is subjected to spatial gridding interpolation through a Kriging interpolation algorithm to generate a time-space coordinate observation data set.

4. The liquid neural network-based intelligent prediction method for mineral resources according to claim 1, characterized in that: The time series reconstruction is performed on the time-space coordinate observation data set to generate a multi-dimensional continuous time-space field function in the following steps, A Gaussian process regression method is adopted to continuously represent the time dimension of the time-space coordinate observation data set to generate preliminary reconstructed time-space sequence data; The preliminary reconstructed time-space sequence data are fused with three-dimensional space coordinates and subjected to spatial gridding processing to form a multi-dimensional continuous time-space field function.

5. The liquid neural network-based intelligent prediction method of mineral resources according to claim 1, characterized in that: The spatial gradient field and the time derivative field of the multi-dimensional continuous time-space field function are calculated to generate a continuous time-space data stream in the following steps, The multi-dimensional continuous time-space field function is subjected to spatial direction derivation to obtain spatial gradient field data; The multi-dimensional continuous time-space field function is subjected to time direction partial derivation operation to generate time derivative field data; The multi-dimensional continuous time-space field function, the spatial gradient field data and the time derivative field data are spliced and standardized in feature dimensions to generate a continuous time-space data stream.

6. The liquid neural network based intelligent forecasting method of mineral resources as claimed in claim 1, wherein: The continuous time-space data stream is input into a liquid-state neural network branch to perform internal state evolution to generate a mineralization dynamic feature vector in the following steps, The continuous time-space data stream is input into a liquid-state neural network branch, and the internal state of a neuron is continuously evolved along a geological time axis driven by a system of ordinary differential equations to generate a dynamic evolution state sequence; The dynamic evolution state sequence is aggregated to extract a time sequence pattern with geological significance, and a mineralization dynamic feature vector is generated.

7. The liquid neural network based intelligent forecasting of mineral resources method as claimed in claim 1, wherein: The static geological data is input into the convolutional neural network branch to extract spatial invariance features and generate a geological background static feature vector, and the steps are as follows, The static geological data is input into the convolutional neural network branch, multi-level spatial features are extracted through convolutional layers and pooling layers, and a spatial feature map group is generated; The spatial feature map group is globally averaged and compressed to generate a geological background static feature vector.

8. The liquid neural network based intelligent forecasting of mineral resources method as claimed in claim 1, wherein: The geological background static feature vector and the mineralization dynamic feature vector are fused and input into the fully connected layer of the liquid neural network branch and the convolutional neural network branch, and the mineral resource prediction value is generated, and the steps are as follows, Based on the geological background static feature vector and the mineralization dynamic feature vector, cross-attention weights are calculated and feature weighted fusion is performed to generate a spatio-temporal fusion feature vector; The spatio-temporal fusion feature vector is input into the fully connected layer of the liquid neural network branch and the convolutional neural network branch, nonlinear transformation and feature dimension reduction are performed, and the mineral resource prediction value is output.

9. The liquid neural network based intelligent forecasting method of mineral resources according to claim 1, wherein: The mineral resource quantity distribution map is generated by spatial interpolation rendering, and the steps are as follows, The mineral resource prediction value is calculated by gridding to generate mineral resource density grid data; The mineral resource density grid data is processed by contour tracing algorithm to generate mineralization intensity contour lines, and combined with the geographic base map to generate the mineral resource quantity distribution map.

10. The liquid neural network based intelligent forecasting of mineral resources method as claimed in claim 1, wherein: The mineral resource quantity distribution map is binarized and segmented to delineate the spatial range of the prospecting target area, and the evolution sequence of the liquid neural network hidden state with geological time is analyzed to generate a mineralization process analysis report, and the steps are as follows, The mineral resource quantity distribution map is binarized and segmented according to the preset mineralization probability threshold to generate a mineral resource quantity segmentation map; The mineral resource quantity segmentation map is subjected to spatial clustering analysis to identify the connected region of the minimum target area, and the target area spatial range vector data is output; The evolution sequence of the liquid neural network hidden state with geological time is analyzed, and the time sequence feature parameters of the evolution trajectory are calculated to generate mineralization dynamic evolution sequence data; The target area spatial range vector data and the mineralization dynamic evolution sequence data are integrated to generate a mineralization process analysis report.

Images