A method, device and medium for online identification of gold ore grade based on multispectral imaging

Abstract

The application discloses a kind of gold ore grade online identification method, equipment and medium based on multispectral imaging, it is related to mineral processing intelligent detection technical field, comprising, using moisture interference index to carry out compensation calculation to original multispectral image data, reconstruct the dry basis spectral data of each pixel point;Spectral image formed by dry basis spectral data is carried out space domain segmentation, obtain a plurality of regions corresponding to different physical ore objects, and for the pixel point in each region, extract the spectral feature vector related to gold ore grade from dry basis spectral data;Spectral feature vector is input into physical constraint neural network model, and the predicted gold ore grade value of each pixel point is output;According to the spatial coordinates and predicted gold ore grade value of all pixel points, generate pixel-level gold ore grade distribution map, and real-time analysis is carried out to gold ore grade distribution map, and obtain gold ore grade identification report.The application realizes high-precision correction to spectral distortion caused by complex optical environment in industrial field.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology in mineral processing, and in particular to a method, equipment and medium for online identification of gold ore grade based on multispectral imaging. Background Technology

[0002] In the field of mineral processing technology, the real-time optimization and control of gold ore beneficiation processes highly depends on the rapid and accurate analysis of the raw ore grade. Traditional gold ore grade analysis mainly relies on offline sampling and laboratory testing. Although this method has high accuracy, it suffers from significant time lag and cannot meet the needs of modern beneficiation plants for real-time control of production processes. To overcome this bottleneck, online detection methods based on spectral analysis technology have been gradually developed. X-ray fluorescence technology has high sensitivity for detecting heavy metal elements, but it is difficult to directly detect low-content elements such as gold; laser-induced breakdown spectroscopy can achieve rapid multi-element detection, but the equipment is expensive and lacks stability in slurry environments; while visible-near-infrared imaging spectroscopy, combined with high-resolution image sensors, shows great application potential due to its advantages such as speed, non-destructive nature, and ability to acquire spatial distribution information.

[0003] When multispectral imaging technology is applied to complex minerals like gold deposits with low content and inclusions, current techniques still fall short. First, existing methods generally lack effective compensation mechanisms for the complex optical environment of industrial sites. In wet processing steps such as crushing and grinding, the water film or slurry environment adhering to the ore surface severely distorts its true spectral reflectance characteristics, especially in the strong absorption band of water molecules, where the spectral signal becomes severely distorted. Existing technologies mostly rely on spectral data acquired by a single image sensor, making it difficult to effectively distinguish whether spectral changes originate from differences in mineral composition or environmental interference. Second, the depth of utilization of spectral information is insufficient, failing to fully explore its deep characteristics related to gold ore grade. Gold itself has weak spectral characteristics; its grade information is mainly indirectly reflected through its symbiotic relationship with specific carrier minerals (such as pyrite and arsenopyrite), and existing methods struggle to accurately extract these indirect characteristics from the massive amounts of data acquired by image sensors. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an online gold ore grade identification method based on multispectral imaging to solve the problems of insufficient accuracy in gold ore grade prediction caused by spectral distortion due to moisture interference and the failure to effectively mine the spectral characteristics of coexisting minerals in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an online gold ore grade identification method based on multispectral imaging, comprising: scanning gold ore on a conveyor belt to acquire original multispectral image data, three-dimensional point cloud data of the ore surface, and real-time incident light spectral data, and performing radiometric calibration to obtain reflectance data of each pixel in each band; extracting reflectance information of each pixel in the water molecule characteristic absorption band range based on the reflectance data of each band, and calculating the moisture interference index of each pixel; using the moisture interference index to compensate the original multispectral image data and reconstructing the dry-base spectral data of each pixel; performing spatial domain segmentation on the spectral image composed of dry-base spectral data to obtain multiple regions corresponding to different physical ore objects, and extracting spectral feature vectors related to gold ore grade from the dry-base spectral data for each pixel in each region; inputting the spectral feature vectors into a physically constrained neural network model to output the predicted gold ore grade value of each pixel; generating a pixel-level gold ore grade distribution map based on the spatial coordinates of all pixels and the predicted gold ore grade value, and performing real-time analysis on the gold ore grade distribution map to obtain a gold ore grade identification report.

[0007] As a preferred embodiment of the online gold ore grade identification method based on multispectral imaging described in this invention, the specific steps for obtaining the reflectance data of each pixel in each band are as follows: The system acquires raw multispectral image data using a multispectral imaging camera, while simultaneously acquiring three-dimensional point cloud data of the ore surface using a three-dimensional laser profilometer and real-time incident light spectral data using a spectrometer. Extract the surface normal vector and microscopic relative height difference of the ore from the 3D point cloud data, which correspond to the spatial position of each pixel in the original multispectral image data. The original multispectral image data of each pixel, the surface normal vector of the ore, the microscopic relative height difference, and the real-time incident light spectrum data are input into the physical information neural network. By minimizing the spectral-spatial consistency constraint function, the reflectance data of each pixel in each band is obtained by inversion.

[0008] As a preferred embodiment of the online gold ore grade identification method based on multispectral imaging described in this invention, the specific steps for calculating the moisture interference index of each pixel are as follows: Based on the reflectance data of each pixel in each band, the reflectance information of each pixel in the preset water molecule characteristic absorption band range is extracted. Based on reflectance information, select multiple characteristic wavelength points within the preset water molecule characteristic absorption band range, and obtain the reflectance time series data of each pixel point at the characteristic wavelength points. Dynamic moisture interference index is calculated on reflectance time series data to obtain the moisture interference index of each pixel.

[0009] As a preferred embodiment of the online gold ore grade identification method based on multispectral imaging described in this invention, the specific steps for using the moisture interference index to compensate for the original multispectral image data and reconstruct the dry-basis spectral data of each pixel are as follows. Based on the moisture interference index and the pre-stored moisture absorption reference spectrum, a spectral deformation vector field is constructed for each pixel. Based on the spectral deformation vector field, the dry basis reflectance of each pixel at each wavelength is calculated by the inverse variational integral equation; Based on the compensated reflectance values ​​of each pixel at each wavelength, the dry-base spectral data of each pixel are reconstructed.

[0010] As a preferred embodiment of the online gold ore grade identification method based on multispectral imaging described in this invention, the specific steps for spatial domain segmentation of the spectral image composed of dry-basis spectral data to obtain multiple regions corresponding to different physical ore objects are as follows. Based on dry-based spectral data, low-dimensional manifold coordinates of each pixel are constructed using a spectral-spatial feature fusion method. Adaptive multi-scale clustering analysis is performed on low-dimensional manifold coordinates to obtain the initial segmentation regions; Obtain the region affiliation probability of each pixel from the initial segmentation region; Hierarchical region merging is performed on the region affiliation probability to obtain each region corresponding to different physical ore objects.

[0011] As a preferred embodiment of the online gold ore grade identification method based on multispectral imaging described in this invention, the specific steps for extracting spectral feature vectors related to gold ore grade from dry-basis spectral data for each pixel within a region are as follows: A multi-scale feature pyramid is constructed based on the dry-base spectral data of each pixel in each region. Perform feature fusion analysis on the multi-scale feature pyramid to obtain the fused feature vector; The fused feature vector is transformed into a spectral feature vector related to the gold ore grade by nonlinear kernel function mapping.

[0012] As a preferred embodiment of the online gold ore grade identification method based on multispectral imaging described in this invention, the specific steps for outputting the predicted gold ore grade value for each pixel are as follows: An enhanced feature tensor is constructed based on spectral feature vectors and multi-source auxiliary features. The enhanced feature tensor is input into the physical constraint neural network model, and the data fitting term and physical constraint term are fused to obtain the initial predicted value of each pixel. The initial predicted values ​​are post-processed and uncertainty is quantified to obtain the predicted gold ore grade value for each pixel.

[0013] As a preferred embodiment of the online gold ore grade identification method based on multispectral imaging described in this invention, the specific steps for obtaining the gold ore grade identification report are as follows: Based on the spatial coordinates of each pixel and the predicted gold ore grade, obtain the adaptive length scale of each pixel; Based on an adaptive length scale, a local Gaussian regression relationship is established for each pixel. Based on the local Gaussian process regression relationship, the corrected gold ore grade value of each pixel is calculated, and the corrected gold ore grade values ​​of all pixels are integrated to generate a pixel-level gold ore grade distribution map. Based on the current gold ore grade distribution map and the historical gold ore grade distribution map sequence, a spatiotemporal dynamic feature map is constructed. Calculate the dynamic quality comprehensive score based on the spatiotemporal dynamic feature map; Based on the dynamic quality comprehensive score, the ore quality is diagnosed and analyzed in real time, and a gold ore grade identification report is generated.

[0014] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the online gold ore grade identification method based on multispectral imaging as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the online gold ore grade identification method based on multispectral imaging as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by using physical information fusion for reflectivity inversion and moisture compensation, high-precision correction of spectral distortion caused by complex optical environments in industrial sites is achieved; by using spatiotemporal dynamic intelligent decision-making that integrates physical constraints and utilizing spatial distribution and sequence information provided by image sensors, the physical rationality of grade prediction and deep insight into the production process are realized. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for online identification of gold ore grades based on multispectral imaging.

[0019] Figure 2 This is a flowchart for outputting reflectivity data for each band.

[0020] Figure 3 A flowchart for generating a moisture disturbance index distribution map.

[0021] Figure 4 This is a comparison chart of multiple curve spectra.

[0022] Figure 5 This is a curve showing the prediction error of gold ore grade. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Reference Figures 1-5 This is one embodiment of the present invention, which provides an online identification method for gold ore grade based on multispectral imaging, comprising the following steps: S1. Scan the gold ore on the conveyor belt to obtain the original multispectral image data, the three-dimensional point cloud data of the ore surface and the real-time incident light spectrum data, and perform radiometric calibration to obtain the reflectivity data of each pixel in each band.

[0027] Raw multispectral image data is acquired using a multispectral imaging camera, while three-dimensional point cloud data of the ore surface is acquired using a three-dimensional laser profilometer, and real-time incident light spectral data is acquired using a spectrometer.

[0028] The specific process includes: a multispectral imaging camera is used to record the reflection information of the ore in multiple continuous bands, specifically by acquiring the spectral response values ​​of hundreds of narrow bands in the visible to near-infrared range for each spatial pixel, forming multispectral image data containing both spectral and spatial dimensions; a three-dimensional laser profilometer is used to capture the spatial geometry of the ore surface, obtaining the three-dimensional coordinates of each point on the ore surface through laser scanning; and a spectrometer is used to simultaneously measure the intensity distribution of incident light irradiated onto the ore surface at various wavelengths, recording the irradiance changes of ambient lighting conditions at different wavelengths in real time. The three components work together to support the subsequent accurate inversion of ore reflectance and compensation for environmental interference factors.

[0029] Extract the surface normal vector and microscopic relative height difference of the ore from the 3D point cloud data, which correspond to the spatial position of each pixel in the original multispectral image data.

[0030] The specific process includes spatially registering the 3D point cloud data with the original multispectral image data, aligning each 3D coordinate point in the 3D point cloud with the corresponding pixel in the original multispectral image data in spatial position, fitting a differential geometric plane or surface based on the 3D points in the local neighborhood to obtain the normal vector direction of the ore surface at that location, and obtaining the micro-relative height difference by comparing the elevation differences between neighboring points, providing each pixel with an ore surface normal vector and micro-relative height difference that strictly correspond to its spatial position.

[0031] The original multispectral image data of each pixel, the surface normal vector of the ore, the microscopic relative height difference, and the real-time incident light spectrum data are input into the physical information neural network. By minimizing the spectral-spatial consistency constraint function, the reflectance data of each pixel in each band is obtained by inversion.

[0032] The specific process involves inputting the original multispectral image data of each pixel, the surface normal vector of the ore, the microscopic relative height difference, and the real-time incident light spectral data into a physical information neural network. The original multispectral image data provides the observed response values ​​of each pixel in multiple continuous bands. The surface normal vector of the ore describes the local orientation of the corresponding position of the pixel. The microscopic relative height difference reflects the geometric shading and shadowing effects caused by the surface microstructure. The real-time incident light spectral data characterizes the incident irradiance of each wavelength under the current illumination conditions. Based on these inputs, the physical information neural network minimizes the spectral-spatial consistency constraint function during the training process, so that the output result simultaneously satisfies the spectral reflectance law and spatial geometric consistency, thereby inverting the reflectance data of each pixel in each band.

[0033] It should be noted that minimizing the spectral-spatial consistency constraint function is an optimization objective. By jointly constraining the physical reflectance characteristics of the spectral dimension and the geometric continuity of the spatial dimension, it ensures that the inversion results conform to spectral laws across bands and maintain spatial smoothness between adjacent pixels. Its function is to improve the physical rationality and spatial structure consistency of reflectance inversion and reduce estimation bias caused by changes in illumination or surface roughness.

[0034] Furthermore, the training process of the Physical Information Neural Network (PIN) involves taking the original multispectral image data of each pixel, the surface normal vector of the ore, the microscopic relative height difference, and the real-time incident light spectrum data as inputs, and using the reflectance data of each pixel in each band obtained by real measurement or high-precision simulation as supervision labels. A combination of fully connected layers and convolutional layers containing weight matrices and bias vectors is used to construct the network backbone. During training, the loss function is defined as a spectral-spatial consistency constraint function, which consists of the residual term between the predicted reflectance and the physical radiative transfer result in the spectral dimension and the consistency term between the reflectance gradient of adjacent pixels and the surface geometry in the spatial dimension. The gradient of the loss function with respect to all learnable parameters of the network is obtained through the backpropagation algorithm, and these parameters are iteratively updated using an adaptive moment estimation optimizer until the loss function converges, thereby completing the training of the PIN.

[0035] Spectral reflectance law refers to the physical relationship between an object's ability to reflect electromagnetic waves at different wavelengths and its material composition, surface properties, and incident and observation angles. It is established based on radiometry, Fresnel reflection theory, and measured spectral data. Spatial geometric consistency refers to the local continuity and geometric rationality that adjacent pixels should satisfy in terms of surface orientation, height variation, and illumination response in three-dimensional space. It is derived by combining the surface normal vector and microscopic relative height difference of the ore obtained by a three-dimensional laser profilometer with spectral observation data.

[0036] S2. Based on the reflectance data of each band, extract the reflectance information of each pixel in the range of water molecule characteristic absorption bands, and calculate the water interference index of each pixel.

[0037] Based on the reflectance data of each pixel in each band, the reflectance information of each pixel in the preset water molecule characteristic absorption band range is extracted.

[0038] The specific process includes selecting reflectance values ​​from the reflectance data of each pixel in each band, corresponding to the specific wavelength range covered by the known absorption characteristics of water molecules in the electromagnetic spectrum. These wavelength ranges are the pre-determined characteristic absorption band ranges of water molecules. The extraction process is to retain all reflectance information of each pixel in this band range, forming a local spectral response that is sensitive to water content.

[0039] It should be noted that the preset characteristic absorption band range of water molecules is determined based on the inherent absorption spectral characteristics of water molecules in the near-infrared and short-wave infrared bands.

[0040] Based on reflectance information, multiple characteristic wavelength points within a preset water molecule characteristic absorption band range are selected, and time series data of reflectance of each pixel point at the characteristic wavelength points are obtained.

[0041] The specific process includes identifying several representative wavelength positions that are sensitive to changes in moisture within a preset water molecule characteristic absorption band range based on reflectance information. These characteristic wavelength points usually correspond to the center of the water molecule absorption peak or its adjacent region. For each pixel, the reflectance value of the pixel at each characteristic wavelength point is extracted from the reflectance data of each band collected at multiple consecutive time points, and arranged in chronological order to construct the reflectance time series data of each pixel at each characteristic wavelength point.

[0042] Dynamic moisture interference index is calculated from reflectance time series data to obtain the moisture interference index for each pixel. The expression is as follows: ; in, Indicates the pixel as Moisture interference index. Represents the index of a pixel. Indices representing characteristic wavelength points. This represents the total number of characteristic wavelength points. Indicates the first The wavelength of each characteristic wavelength point Indicates at wavelength The intensity of water absorption at that location. Indicates the pixel as At wavelength The standard deviation of the reflectance time series at a given location.

[0043] It should be noted that water absorption intensity is a physical quantity that characterizes the ability of water molecules to absorb incident light at a specific wavelength. It reflects the sensitivity of that wavelength to the presence of water and is derived from the absorption depth or absorption coefficient at the corresponding wavelength position in a pre-stored water absorption reference spectrum.

[0044] The specific process includes calculating the dynamic moisture interference index of reflectance time series data, multiplying the standard deviation of the reflectance time series at each characteristic wavelength point by the moisture absorption intensity at the corresponding wavelength, where the standard deviation of the reflectance time series characterizes the degree of spectral response fluctuation of the pixel point at that wavelength due to moisture changes, and the moisture absorption intensity reflects the sensitivity of the wavelength to moisture. The product results of all characteristic wavelength points are accumulated and normalized to the sum of the moisture absorption intensities of all characteristic wavelength points to obtain the moisture interference index of each pixel point.

[0045] S3. The moisture interference index is used to compensate the original multispectral image data and reconstruct the dry basis spectral data of each pixel.

[0046] Based on the moisture interference index and the pre-stored moisture absorption reference spectrum, a spectral deformation vector field is constructed for each pixel.

[0047] The specific process includes multiplying the moisture interference index of each pixel with the absorption intensity of the pre-stored moisture absorption reference spectrum in each band to obtain the direction and amplitude of the spectral shift caused by moisture in the entire band range. The direction and amplitude of the spectral shift form a vector component at each wavelength, and the vector components at all wavelengths together constitute the spectral deformation vector field of the pixel.

[0048] It should be noted that the pre-stored moisture absorption reference spectrum is the moisture characteristic absorption curve extracted from the reflectance spectrum of pure water or water-bearing flat mineral samples obtained by a spectrometer under controlled conditions, and is stored in digital form in the storage medium as a reference.

[0049] Based on the spectral deformation vector field, the dry-base reflectance of each pixel at each wavelength is calculated using the inverse variational integral equation, expressed as: ; in, Indicates coordinates as The pixels at wavelength Dry substrate reflectance at that location Represents the spatial coordinates of a pixel in an image. Indicates spectral wavelength, Indicates coordinates as The pixels at wavelength The wet substrate reflectance at that location Represents the global compensation strength coefficient. Represents the hyperbolic tangent function. Indicates coordinates as The pixels at wavelength The spectral deformation vector field caused by moisture interference, The normalized parameter representing the sensitivity to the intensity of the deformation field. Indicates wavelength Gradient operator for partial derivatives This represents the gradient normalization parameter.

[0050] It should be noted that the global compensation intensity coefficient is an empirical parameter used to adjust the overall correction amplitude of moisture interference. It is determined by comparing the difference between wet and dry reflectance on ore samples with known water content and dry conditions and fitting the spectral deformation relationship. The normalization parameter for the sensitivity to deformation field intensity is an adjustment factor used to balance the response differences at different wavelengths in the spectral deformation vector field. It includes the moisture absorption intensity weights and local spectral gradient amplitudes corresponding to each wavelength. It is obtained by normalizing the pre-stored moisture absorption reference spectrum and combining the differential changes in reflectance between adjacent bands. The gradient normalization parameter is a normalization factor used to adjust the influence of the local rate of change of the spectral deformation vector field. It includes the differential gradient amplitude of reflectance between adjacent bands at each wavelength and the moisture absorption intensity weights at the corresponding wavelengths. It is obtained by calculating the first or second order difference of the wet reflectance sequence of each pixel at each wavelength to obtain the spectral gradient, and then combining it with the absorption intensity of the corresponding wavelength in the pre-stored moisture absorption reference spectrum for weighted normalization.

[0051] The specific process includes taking the wet substrate reflectance observed at each wavelength of each pixel as the initial input, combining it with the vector components at each wavelength in the spectral deformation vector field corresponding to that pixel, and using the inverse variational integral equation to establish the mapping relationship between wet substrate reflectance and dry substrate reflectance. This equation includes a global compensation intensity coefficient to control the overall correction amplitude, and a gradient normalization parameter to balance the local spectral variation rate. At the same time, the hyperbolic tangent function is used to nonlinearly modulate the modulus and spatial gradient of the spectral deformation vector field to suppress overcompensation and maintain the physical rationality of the spectral structure. Finally, the dry substrate reflectance of each pixel after eliminating moisture interference is retrieved wavelength by wavelength.

[0052] Based on the compensated reflectance values ​​of each pixel at each wavelength, the dry-base spectral data of each pixel are reconstructed.

[0053] The specific process includes, based on the compensated reflectance values ​​of each pixel at each wavelength, organizing the reflectance values ​​of each pixel at all wavelength positions after correction by the inverse variational integral equation in order of wavelength from shortest to longest, forming a continuous, smooth spectral curve that eliminates the influence of moisture. This curve fully preserves the inherent spectral characteristics of the ore in the visible to near-infrared range, including the absorption peak position, absorption depth and overall spectral morphology, thereby reconstructing the dry-base spectral data of each pixel.

[0054] like Figure 4The diagram illustrates the spectral shape changes of "reflectance data (inversion)" and "dry-based spectral data (reconstruction)" relative to the "true dry-based reflectance value" under different moisture content conditions. The upper overview diagram uses a red dashed box to enlarge the window related to the characteristic absorption of water molecules, and connects this window to the corresponding enlarged local view in the lower section with a red dashed line. The enlarged local view further marks the feature points and the points of greatest difference, and provides scientific notation values ​​to intuitively illustrate that after "compensating the original multispectral image data using the moisture interference index", the reconstructed dry-based spectrum is closer to the true dry-based value, thus demonstrating the high-precision correction capability of this invention for spectral distortion caused by on-site interference such as humidity / uniform lighting.

[0055] S4. Spatial domain segmentation is performed on the spectral image composed of dry-based spectral data to obtain multiple regions corresponding to different physical ore objects. For each pixel in the region, spectral feature vectors related to gold ore grade are extracted from the dry-based spectral data.

[0056] Based on dry-base spectral data, low-dimensional manifold coordinates of each pixel are constructed using a spectral-spatial feature fusion method.

[0057] The specific process includes jointly encoding the dry-base spectral data of each pixel with the surface normal vector and microscopic relative height difference of the corresponding ore. The local absorption features in the spectral dimension and the geometric neighborhood relationship in the spatial dimension are extracted using a multi-scale feature pyramid. Then, the high-dimensional spectral-spatial joint features are mapped to a low-dimensional embedding space while maintaining the local adjacency structure through a nonlinear dimensionality reduction algorithm. This generates low-dimensional manifold coordinates for each pixel that can simultaneously characterize its spectral composition characteristics and surface geometric context information.

[0058] Adaptive multi-scale clustering analysis is performed on low-dimensional manifold coordinates to obtain the initial segmentation regions.

[0059] The specific process includes dynamically adjusting the neighborhood radius and clustering threshold at multiple scales based on the distribution density and local geometric structure of each pixel in the low-dimensional manifold coordinate space. By detecting high-density connected regions on the manifold, a set of pixels with similar spectral-spatial features is identified. Each set corresponds to a local region with high homogeneity, thereby forming an initial segmentation region covering the entire image.

[0060] Obtain the region affiliation probability of each pixel from the initial segmentation region.

[0061] The specific process includes obtaining the confidence level of each pixel belonging to the initial segmentation region based on the distance distribution and local density relationship between each pixel and other pixels in the low-dimensional manifold coordinate space and the initial segmentation region. Pixels that are closer to the initial segmentation region and are located in the high-density core region have a higher probability of region assignment, while pixels located in the boundary or sparse region have a lower probability of region assignment. Finally, a region assignment probability for each pixel to the assigned initial segmentation region is generated.

[0062] Hierarchical region merging is performed on the region affiliation probability to obtain each region corresponding to different physical ore objects.

[0063] The specific process includes iteratively merging adjacent regions from bottom to top at multiple levels based on the similarity of the boundary pixel region attribution probabilities between each initial segmentation region and the spectral-spatial feature distance in the low-dimensional manifold coordinate space. Each time, the pair of adjacent regions with the closest region attribution probability distribution and the smallest feature difference are selected and gradually aggregated to form a larger region with semantic consistency and clear boundaries until the stopping criterion is met, thus obtaining each region that corresponds one-to-one with different physical mineral objects in the image.

[0064] A multi-scale feature pyramid is constructed based on the dry-base spectral data of each pixel in each region.

[0065] The specific process includes performing spectral smoothing and downsampling operations at different scales on the dry-based spectral data of all pixels in each region to form multiple levels of representation from fine to coarse. Each level retains the spectral absorption features and local variation trends at a specific scale. The higher levels capture the wide-area spectral contour and overall composition information, while the lower levels retain narrowband absorption details and weak feature structures, thus forming a hierarchical multi-scale feature pyramid.

[0066] Perform feature fusion analysis on the multi-scale feature pyramid to obtain the fused feature vector.

[0067] The specific process involves associating the spectral features of each level in the multi-scale feature pyramid layer by layer in order from coarse to fine or from fine to coarse. Then, using a cross-scale attention mechanism or a weighted splicing strategy, the spectral absorption position, depth, and morphological information extracted at different scales are integrated. This allows the high-level global component features and the low-level local detail features to be co-expressed in a unified representation, forming a fused feature vector that contains multi-scale spectral context information.

[0068] The fused feature vector is transformed into a spectral feature vector related to the gold ore grade by nonlinear kernel function mapping.

[0069] The specific process includes inputting the fused feature vector into a preset nonlinear kernel function, using the nonlinear kernel function to implicitly map the fused feature vector in the high-dimensional regenerating kernel Hilbert space, so that the complex relationship between the spectral features that were originally linearly inseparable in the low-dimensional space and the gold ore grade presents a stronger correlation in the high-dimensional space, thereby generating a spectral feature vector that can effectively characterize the gold ore grade information.

[0070] It should be noted that the preset nonlinear kernel function is selected and fixed based on the nonlinear statistical relationship between the dry basis spectral data of the gold ore sample and the corresponding gold ore grade label, after cross-validation and performance evaluation of various candidate kernel functions (such as radial basis function, polynomial kernel, etc.) during the training phase.

[0071] S5. Input the spectral feature vector into the physical constraint neural network model and output the predicted gold ore grade value of each pixel.

[0072] An enhanced feature tensor is constructed based on spectral feature vectors and combined with multi-source auxiliary features.

[0073] The specific process involves concatenating or performing tensor product operations on the feature dimension of the spectral feature vector with the ore surface normal vector, the microscopic relative height difference, and the region attribution probability, etc., to form a high-dimensional joint representation that includes spectral response characteristics, spatial geometric information, and regional semantic context. This high-dimensional joint representation is organized in tensor form, preserving the structural correlation of each feature source, and constructing an enhanced feature tensor.

[0074] It should be noted that the multi-source auxiliary features refer to the ore surface normal vector, microscopic relative height difference, and region attribution probability, which are obtained through three-dimensional topography reconstruction, high-resolution topographic survey, and similarity calculation of the initial segmented region, respectively. They are used to supplement spectral information and provide spatial geometric and regional semantic context.

[0075] The enhanced feature tensor is input into the physical constraint neural network model, and the data fitting term and physical constraint term are fused to obtain the initial predicted value of each pixel.

[0076] The specific process includes the following steps during the training and inference of the physical constraint neural network model: on the one hand, the enhanced feature tensor is nonlinearly mapped through the backbone structure of the neural network to minimize the data fitting term between the predicted gold ore grade and the actual gold ore grade; on the other hand, physical constraint terms based on mineral symbiosis laws and spectral reflectance physical mechanisms are embedded in the loss function. By weighted combination of data fitting terms and physical constraint terms, the network parameters are updated to ensure that the output results conform to both the statistical laws of the observed data and the prior physical laws, thereby generating the initial predicted value of each pixel.

[0077] The construction process of the physically constrained neural network model is as follows: The network input is determined to be the enhanced feature tensor, and the output is the predicted gold grade value of each pixel. A deep neural network structure composed of convolutional layers, fully connected layers, and nonlinear activation functions is used as the backbone. The network loss function includes both a data fitting term and a physical constraint term. The data fitting term is the mean square error between the predicted gold grade and the actual gold grade. The physical constraint term is constructed based on the mineral co-occurrence law and the physical mechanism of spectral reflectance, and is used to limit the physical rationality of the prediction results. By combining the data fitting term and the physical constraint term with preset weights to form a joint loss function, and configuring learnable weight matrices and bias vectors as network parameters, the construction of the physically constrained neural network model is completed.

[0078] The training process of the physically constrained neural network model involves taking the enhanced feature tensor as input and the actual gold ore grade of the corresponding pixel as the supervision label. A data fitting term is calculated through forward propagation of the neural network, which measures the error between the predicted and actual gold ore grades. Simultaneously, a physical constraint term is constructed based on mineral co-occurrence patterns and spectral reflectance physical mechanisms. This physical constraint term quantifies the deviation of the prediction results in terms of physical reasonableness. The data fitting term and the physical constraint term are added together with preset weights to form a joint loss function. The gradient of the joint loss function with respect to the network parameters is obtained through backpropagation, and the parameters of the physically constrained neural network model are iteratively updated using an optimizer until the joint loss function converges, thus completing the training of the physically constrained neural network model.

[0079] The initial predicted values ​​are post-processed and uncertainty is quantified to obtain the predicted gold ore grade value for each pixel.

[0080] The specific process includes applying spatial smoothing filtering to the initial predicted values ​​to suppress isolated noise and enhance regional consistency, and using Monte Carlo Dropout or ensemble prediction methods to estimate the variance or confidence interval of the prediction results for each pixel based on the local density of the enhanced feature tensor in the feature space, the probability of regional attribution, and the residual size of the physical constraint term. This uncertainty measure is then combined with the initial predicted values, and the predicted gold ore grade is generated through weighted correction or threshold truncation.

[0081] like Figure 5 The curve structure of the average absolute error of the predicted gold ore grade value as a function of the actual gold ore grade is given under different moisture content conditions (each curve corresponds to a moisture content condition). The fluctuation shape of the curve is used to reflect the distribution characteristics and stability of the error in different grade ranges, thereby supporting the "physically constrained neural network model" to maintain good generalization ability and interpretable error behavior under multiple conditions, rather than just giving a single average grade value.

[0082] S6. Based on the spatial coordinates of all pixels and the predicted gold ore grade value, generate a pixel-level gold ore grade distribution map, and perform real-time analysis on the gold ore grade distribution map to obtain a gold ore grade identification report.

[0083] Based on the spatial coordinates of each pixel and the predicted gold ore grade, an adaptive length scale for each pixel is obtained.

[0084] The specific process includes combining the position of each pixel in three-dimensional space and the spatial variation gradient of the predicted gold grade value in its neighborhood to obtain the degree of heterogeneity of the local gold grade distribution. When the difference in the predicted gold grade value between neighboring pixels is large, the corresponding region is identified as a high gradient boundary region and a smaller length scale is assigned to preserve details. When the change in the predicted gold grade value between neighboring pixels is gradual, the corresponding region is identified as a homogeneous region and a larger length scale is assigned to enhance smoothness. Thus, an adaptive length scale matching its local spatial structure and grade distribution characteristics is generated for each pixel.

[0085] Based on an adaptive length scale, a local Gaussian regression relationship is established for each pixel.

[0086] The specific process includes defining a neighborhood range centered on each pixel using a corresponding adaptive length scale, collecting pixels with similar spatial coordinates and related predicted gold ore grade values ​​within the neighborhood as training samples, and using Gaussian process regression to construct a probabilistic mapping relationship between the input spatial coordinates and the output predicted gold ore grade value. The length parameter of the covariance function is determined by the adaptive length scale of the pixel, thereby dynamically adjusting the smoothness and generalization ability of the regression model according to the local gold ore grade variation characteristics in different regions, forming a local Gaussian process regression relationship for each pixel.

[0087] Based on the local Gaussian process regression relationship, the corrected gold grade value of each pixel is calculated, and the corrected gold grade values ​​of all pixels are integrated to generate a pixel-level gold grade distribution map. The expression is: ; in, Indicates the pixel as The corrected gold ore grade value, Indicates the index of the neighboring pixels. Indicates the pixel as The local neighborhood set, This represents the natural exponential function. Indicates the pixel as spatial coordinate vector, Indicates that the neighboring pixels are spatial coordinate vector, This represents the length scale parameter.

[0088] It should be noted that the length scale parameter is a hyperparameter of the covariance function in Gaussian process regression, used to control the decay rate of spatial correlation, including a measure of the smoothness of local gold ore grade changes, and is directly determined by the adaptive length scale of each pixel.

[0089] The specific process includes using the covariance function defined in the local Gaussian process regression relationship corresponding to each pixel, taking the Euclidean distance between spatial coordinate vectors as input, and calculating the weight contribution of each pixel in the neighborhood to the current pixel through the natural exponential function. This weight contribution decreases as the spatial distance increases. The length scale parameter is determined by an adaptive length scale. The predicted gold grade values ​​of all pixels in the neighborhood are weighted and summed according to their corresponding weights to obtain the corrected gold grade value of the current pixel. This process is repeated for all pixels in the image to form a pixel-level gold grade distribution map covering the entire imaging area.

[0090] Based on the current gold ore grade distribution map and the historical gold ore grade distribution map sequence, a spatiotemporal dynamic characteristic map is constructed.

[0091] The specific process includes aligning the current gold ore grade distribution map with multiple historical gold ore grade distribution maps arranged in chronological order in spatial location and stacking them along the time dimension to form a four-dimensional tensor structure. Each spatial location contains a sequence of gold ore grade values ​​that evolve over time. Then, the rate of change of gold ore grade, the local trend slope, and the autocorrelation of the time series between adjacent time steps are obtained. Temporal features reflecting the evolution of grade are extracted and fused with the current spatial distribution features to generate a spatiotemporal dynamic feature map that simultaneously encodes spatial heterogeneity and temporal dynamics.

[0092] Based on the spatiotemporal dynamic feature map, the dynamic quality comprehensive score is calculated, and the expression is: ; in, This represents the dynamic quality comprehensive score. This represents the mean weighting coefficient. This represents the average value of the current gold ore grade distribution map. This represents the average reference grade. This represents the standard deviation weighting coefficient. This represents the standard deviation of the current gold ore grade distribution map. Indicates the reference standard deviation. This represents the total variation weighting coefficient. This represents the total variation in the current gold ore grade distribution map. This represents the maximum possible total variation in the current gold ore grade distribution map.

[0093] It should be noted that the mean weighting coefficient is a preset parameter used to adjust the contribution of grade level indicators to the dynamic quality comprehensive score, obtained through regression analysis between historical gold mine grade distribution map sequences and actual beneficiation recovery rates or economic values; the standard deviation weighting coefficient is a preset parameter used to adjust the contribution of grade uniformity indicators to the dynamic quality comprehensive score, determined by analyzing the statistical relationship between standard deviation and beneficiation stability or processing costs in historical gold mine grade distribution map sequences; the reference standard deviation is a preset value used to measure the current gold mine grade distribution uniformity benchmark, based on gold mine grades obtained under stable production conditions in historical gold mine grade distribution map sequences. The standard deviation of the gold grade is determined by the statistical mean; the total variation weighting coefficient is a preset parameter used to adjust the contribution of the grade variation index to the dynamic quality comprehensive score. It is determined by analyzing the correlation between the total variation and ore beneficiation or crushing and grinding energy consumption in the historical gold mine grade distribution map series; the reference grade mean is a benchmark value used to measure the overall grade level of the current gold mine. It is obtained by statistical analysis of gold grade data collected in the historical gold grade distribution map series under normal production conditions for multiple time periods. This value reflects the average gold content level of the mining area under typical mining and processing conditions and serves as a comparison benchmark for grade level indicators in the dynamic quality comprehensive score.

[0094] The specific process includes: using the ratio of the average value of the current gold ore grade distribution map to the reference average value as a grade level indicator; combining the ratio of the standard deviation of the current gold ore grade distribution map to the reference standard deviation to convert it into a grade uniformity indicator through a natural exponential function; and then using the ratio of the total variation of the current gold ore grade distribution map to the maximum possible total variation as a grade variation range indicator. Each indicator is weighted and combined with its corresponding mean weight coefficient, standard deviation weight coefficient, and total variation weight coefficient, and then summed to obtain a dynamic quality comprehensive score that reflects the overall grade level, uniformity, and dynamic variation characteristics of the gold ore.

[0095] Based on the dynamic quality comprehensive score, the ore quality is diagnosed and analyzed in real time, and a gold ore grade identification report is generated.

[0096] The specific process includes, based on multiple preset condition-conclusion rules, defining the ore quality grade, ore washability judgment, and processing suggestions corresponding to the dynamic quality comprehensive score within different threshold ranges for each rule (e.g., when the dynamic quality comprehensive score is greater than 0.85, it is judged as a high-grade stable ore with good washability, and it is recommended to directly enter the main beneficiation process; when the score is between 0.65 and 0.85, it is judged as a medium-grade fluctuating ore with average washability, and it is recommended to perform pre-enrichment treatment before beneficiation; when the score is less than 0.65, it is judged as a low-grade or highly heterogeneous ore with poor washability, and it is recommended to temporarily store it or use it for blending). The currently obtained dynamic quality comprehensive score is matched to the corresponding rule conditions one by one, triggering the rule conclusions that meet the conditions. The output results of all activated rules are combined to form a structured gold ore grade identification report that includes grade level evaluation, spatial uniformity assessment, time stability judgment, and process treatment suggestions.

[0097] It should be noted that the multiple condition-conclusion rules are pre-set based on the statistical correlation between the historical gold ore grade distribution sequence and the corresponding beneficiation effect, process parameters and economic indicators, combined with the process experience accumulated in mining production practice, through inductive analysis and threshold division.

[0098] This embodiment also provides a computer device applicable to the online gold ore grade identification method based on multispectral imaging, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the online gold ore grade identification method based on multispectral imaging as proposed in the above embodiment.

[0099] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0100] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the online identification method for gold ore grade based on multispectral imaging as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0101] In summary, this invention achieves high-precision correction of spectral distortion caused by complex optical environments in industrial settings through reflectivity inversion and moisture compensation via physical information fusion; and achieves physical rationality in grade prediction and deep insight into the production process by incorporating spatiotemporal dynamic intelligent decision-making based on physical constraints and utilizing spatial distribution and sequence information provided by image sensors.

[0102] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for online identification of gold ore grade based on multispectral imaging, characterized in that: include, The gold ore on the conveyor belt is scanned to obtain raw multispectral image data, three-dimensional point cloud data of the ore surface and real-time incident light spectrum data, and radiometric calibration is performed to obtain the reflectivity data of each pixel in each band. Based on the reflectance data of each band, the reflectance information of each pixel in the range of water molecule characteristic absorption band is extracted, and the water interference index of each pixel is calculated. The moisture interference index was used to compensate for the original multispectral image data, and the dry basis spectral data of each pixel was reconstructed. Spatial domain segmentation is performed on the spectral image composed of dry-based spectral data to obtain multiple regions corresponding to different physical ore objects. For each pixel in the region, spectral feature vectors related to gold ore grade are extracted from the dry-based spectral data. The spectral feature vector is input into the physical constraint neural network model, and the predicted gold ore grade value of each pixel is output. Based on the spatial coordinates of all pixels and the predicted gold ore grade value, a pixel-level gold ore grade distribution map is generated, and the gold ore grade distribution map is analyzed in real time to obtain a gold ore grade identification report.

2. The online gold ore grade identification method based on multispectral imaging as described in claim 1, characterized in that: The specific steps for obtaining the reflectance data of each pixel in each wavelength band are as follows: The system acquires raw multispectral image data using a multispectral imaging camera, while simultaneously acquiring three-dimensional point cloud data of the ore surface using a three-dimensional laser profilometer and real-time incident light spectral data using a spectrometer. Extract the surface normal vector and microscopic relative height difference of the ore from the 3D point cloud data, which correspond to the spatial position of each pixel in the original multispectral image data. The original multispectral image data of each pixel, the surface normal vector of the ore, the microscopic relative height difference, and the real-time incident light spectrum data are input into the physical information neural network. By minimizing the spectral-spatial consistency constraint function, the reflectance data of each pixel in each band is obtained by inversion.

3. The online gold ore grade identification method based on multispectral imaging as described in claim 2, characterized in that: The specific steps for calculating the moisture interference index of each pixel are as follows. Based on the reflectance data of each pixel in each band, the reflectance information of each pixel in the preset water molecule characteristic absorption band range is extracted. Based on reflectance information, select multiple characteristic wavelength points within the preset water molecule characteristic absorption band range, and obtain the reflectance time series data of each pixel point at the characteristic wavelength points. Dynamic moisture interference index is calculated on reflectance time series data to obtain the moisture interference index of each pixel.

4. The online gold ore grade identification method based on multispectral imaging as described in claim 3, characterized in that: The method of using the moisture interference index to compensate for the original multispectral image data and reconstructing the dry basis spectral data of each pixel is as follows: Based on the moisture interference index and the pre-stored moisture absorption reference spectrum, a spectral deformation vector field for each pixel is constructed. Based on the spectral deformation vector field, the dry basis reflectance of each pixel at each wavelength is calculated by the inverse variational integral equation; Based on the compensated reflectance values ​​of each pixel at each wavelength, the dry-base spectral data of each pixel are reconstructed.

5. The online gold ore grade identification method based on multispectral imaging as described in claim 4, characterized in that: The specific steps for spatial domain segmentation of the spectral image composed of dry-based spectral data to obtain multiple regions corresponding to different physical ore objects are as follows. Based on dry-based spectral data, low-dimensional manifold coordinates of each pixel are constructed using a spectral-spatial feature fusion method. Adaptive multi-scale clustering analysis is performed on low-dimensional manifold coordinates to obtain the initial segmentation regions; Obtain the region affiliation probability of each pixel from the initial segmentation region; Hierarchical region merging is performed on the region affiliation probability to obtain each region corresponding to different physical ore objects.

6. The online gold ore grade identification method based on multispectral imaging as described in claim 5, characterized in that: For each pixel within a region, the spectral feature vector related to the gold ore grade is extracted from the dry-basis spectral data. The specific steps are as follows: A multi-scale feature pyramid is constructed based on the dry-base spectral data of each pixel in each region. Perform feature fusion analysis on the multi-scale feature pyramid to obtain the fused feature vector; The fused feature vector is transformed into a spectral feature vector related to the gold ore grade by nonlinear kernel function mapping.

7. The online gold ore grade identification method based on multispectral imaging as described in claim 6, characterized in that: The specific steps for outputting the predicted gold ore grade value for each pixel are as follows. An enhanced feature tensor is constructed based on spectral feature vectors and multi-source auxiliary features. The enhanced feature tensor is input into the physical constraint neural network model, and the data fitting term and physical constraint term are fused to obtain the initial predicted value of each pixel. The initial predicted values ​​are post-processed and uncertainty is quantified to obtain the predicted gold ore grade value for each pixel.

8. The online gold ore grade identification method based on multispectral imaging as described in claim 7, characterized in that: The specific steps for obtaining the gold ore grade identification report are as follows. Based on the spatial coordinates of each pixel and the predicted gold ore grade, obtain the adaptive length scale of each pixel; Based on an adaptive length scale, a local Gaussian regression relationship is established for each pixel. Based on the local Gaussian process regression relationship, the corrected gold ore grade value of each pixel is calculated, and the corrected gold ore grade values ​​of all pixels are integrated to generate a pixel-level gold ore grade distribution map. Based on the current gold ore grade distribution map and the historical gold ore grade distribution map sequence, a spatiotemporal dynamic feature map is constructed. Calculate the dynamic quality comprehensive score based on the spatiotemporal dynamic feature map; Based on the dynamic quality comprehensive score, the ore quality is diagnosed and analyzed in real time, and a gold ore grade identification report is generated.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the online gold ore grade identification method based on multispectral imaging as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the online gold ore grade identification method based on multispectral imaging as described in any one of claims 1 to 8.

Images