A method for intelligent classification of magnesite ore grade based on machine learning

Abstract

The application discloses a kind of based on machine learning's magnesite ore grade intelligent classification method, it is related to ore classification technical field, the method includes: collecting different grade magnesite ore sample, and the real grade label of each sample is calibrated by chemical analysis;Raw hyperspectral image data of ore sample is collected using hyperspectral imager;Wavelet transform algorithm is used to carry out denoising processing to the collected spectral data, and the training sample set is amplified by data enhancement technology;Machine learning model is constructed to extract deep nonlinear features, and multiscale feature fusion is carried out in combination with mineral spectral diagnostic absorption characteristics;Learning machine classification model is constructed, and the model is trained;Noise robustness test is used to evaluate the trained model;The optimized model is deployed to edge computing device, and the automatic classification of magnesite ore grade is realized.The application solves the problem of low efficiency and insufficient accuracy in the prior art for grading magnesite ore grade.

Description

Technical Field

[0001] This invention relates to the field of ore classification technology, and in particular to an intelligent classification method for magnesite ore grades based on machine learning. Background Technology

[0002] Magnesite is a carbonate mineral primarily composed of magnesium carbonate, with a theoretical chemical composition of 47.81% magnesium oxide and 52.19% carbon dioxide. However, in nature, it often contains substitute elements such as iron, manganese, and calcium. It is typically white, grayish-white, or pale yellow, with a vitreous luster, a hardness of 3.5-4.5, and a density of 2.9-3.1 g / cm³. Based on its crystallinity, it can be divided into crystalline magnesite and cryptocrystalline magnesite. Magnesium oxide obtained from calcined magnesite possesses excellent refractory and binding properties, and is mainly used in refractory materials in the metallurgical industry, magnesium refining, building materials, chemicals, and environmental protection. According to the Chinese metallurgical industry standard YB / T 5208-2004, magnesite is classified into different grades based on its chemical composition. The core classification criteria are the content of magnesium oxide and the limits on the content of calcium oxide and silicon dioxide, which are harmful to smelting. The higher the grade, the higher the MgO content and the stricter the impurity requirements, directly determining the industrial use and economic value of the ore: premium grade is used for high-purity magnesia and special refractories; grade one is used for high-quality refractories; grade two is used for conventional refractories; grade three is used for magnesia-silica sand or thermal beneficiation raw materials; grade four is used for metallurgical magnesia; and magnesite powder is mainly used as a sintering raw material. In actual production, in addition to the chemical composition content, the occurrence state and dissemination characteristics of impurities also affect the processing difficulty and usability of the ore. For example, although the MgO content of high-silica, high-calcium grade one ore meets the standard, the extremely fine dissemination of impurities increases the difficulty of beneficiation.

[0003] While traditional classification methods based on chemical analysis for grade determination are accurate and reliable, they are lengthy, time-consuming, and labor-intensive, requiring multiple complex steps such as sample collection, preparation, chemical reagent preparation, titration, or gravimetric determination. This fails to meet the demands of modern industrial production for real-time and rapid testing. Furthermore, traditional methods such as manual visual inspection and sorting are heavily influenced by subjective factors, making it difficult to accurately determine the grade of complex ores with uneven mineral content distribution. From the perspective of primary separation processes, existing gravity separation methods are ineffective due to the small density difference between magnesite and gangue minerals, and are only suitable for coarse-grained ores larger than 3 mm. While flotation is widely used, magnesite and gangue minerals such as dolomite have similar chemical compositions and crystal structures, with indistinct surface properties, requiring the selection of appropriate flotation reagent combinations and making process control difficult. Thermal separation methods are energy-intensive and have specific requirements regarding the thermal properties of the ore. Simultaneously, the long-standing predatory mining model of extracting high-grade magnesite while discarding lower-grade ore has led to a decline in my country's high-grade magnesite resources, while a large amount of low-grade ore remains unutilized, resulting in significant resource waste. Furthermore, the differences in the degree of crystallization of magnesite crystals (fine-grained, medium-grained, coarse-grained, and giant-grained) directly affect its physicochemical behavior and sintering characteristics during high-temperature calcination. However, traditional classification methods often overlook this important factor, making it difficult to accurately match subsequent thermal treatment processes.

[0004] Therefore, there is an urgent need to develop a fast, accurate, and intelligent method for classifying magnesite ore grades. Summary of the Invention

[0005] This invention provides a machine learning-based intelligent classification method for magnesite ore grades. By constructing a machine learning model, it solves the problems of low efficiency and insufficient accuracy in the existing technology for magnesite ore grade classification.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a machine learning-based intelligent classification method for magnesite ore grades, comprising:

[0007] S1. Collect magnesite ore samples of different grades and perform chemical analysis on each sample to determine its true grade label.

[0008] S2, using a hyperspectral imager to acquire raw hyperspectral image data of ore samples;

[0009] S3 uses wavelet transform algorithm to denoise the collected spectral data and expands the training sample set through data augmentation technology;

[0010] S4. Construct a machine learning model to extract deep nonlinear features and combine it with diagnostic absorption features of mineral spectra for multi-scale feature fusion.

[0011] S5. Construct a learning machine classification model, using the fused multi-scale features as input and the ore grade as output, to train the model;

[0012] S6. The trained model is evaluated using a noise robustness test, and the model parameters are optimized based on the evaluation results. The S7 deploys the optimized model to edge computing devices and interfaces with an industrial IoT platform to achieve automatic classification of magnesite ore grades.

[0013] The beneficial effects of the technical solution provided by this invention include at least the following: The method of this invention integrates deep features extracted by deep learning with the unique diagnostic absorption features of magnesite mineral spectra at multiple scales, so that the classification model has both powerful data-driven representation capabilities and incorporates the physical mechanism constraints of mineral spectra, effectively avoiding the overfitting and lack of physical meaning problems that may occur in purely data-driven models.

[0014] This invention designs a method for testing and optimizing the noise robustness of a system. By adding Gaussian white noise, impulse noise, and baseline drift noise to the test samples, it simulates various interferences that may occur in an industrial environment, thereby minimizing the performance degradation of the model under various noise interferences.

[0015] This invention employs hyperspectral imaging technology, which can quickly acquire full-band spectral information of ore samples through push-broom scanning. Combined with a lightweight classification model on an edge computing device, it can efficiently complete the entire process from data acquisition to grade output. At the same time, this invention adopts a non-contact measurement method, which does not require any preprocessing of the ore samples and does not damage the original state of the ore, thus achieving non-destructive testing. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0017] Figure 1 This is a flowchart of a machine learning-based intelligent classification method for magnesite ore grades provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0019] Example A machine learning-based intelligent classification method for magnesite ore grades.

[0020] Please refer to Figure 1 This is a flowchart of the intelligent classification method for magnesite ore grades based on machine learning provided in an embodiment of the present invention.

[0021] S1. Collect magnesite ore samples of different grades and perform chemical analysis on each sample to determine its true grade label. S101, collect magnesite ore samples covering the range of special grade, first grade, second grade and third grade; It should be noted that the determination of the sample grade label is based on the requirements of magnesite grade and chemical composition specified in the metallurgical industry standard YB / T5208-2004, including six grades: M47, M46, M45, M44, M41, and M33. The corresponding grade level is determined according to the content of magnesium oxide, calcium oxide, and silicon dioxide.

[0022] The samples included blocky ore samples from different mining areas and powdery samples of different particle sizes; It should be noted that manual sampling was used for sample collection, and the sampling location, sampling layer, collection date, and geological feature description of each sample must be recorded.

[0023] S102. For each collected sample, chemical analysis is performed according to national standards to determine the content of magnesium oxide, calcium oxide, and silicon dioxide, and the true grade label of each sample is determined by referring to metallurgical industry standards.

[0024] It should be noted that the samples must be chemically analyzed in accordance with the provisions of the national standard GB / T34332-2017 "Chemical Analysis Methods for Magnesite and Dolomite Refractory Products"; the contents of magnesium oxide and calcium oxide shall be determined by EDTA titration or atomic absorption spectrometry; and the contents of silicon dioxide shall be determined by gravimetric method or molybdenum blue spectrophotometry.

[0025] S2, using a hyperspectral imager to acquire raw hyperspectral image data of ore samples; S201, using a hyperspectral imager to perform push-broom scanning on magnesite ore to obtain raw hyperspectral image data of the ore sample; It should be noted that the imaging spectrometer has a spectral range of 400nm to 2500nm, covering the visible to near-infrared bands; the spectral resolution is better than 5nm, and the spatial resolution is selected from 0.1mm to 1mm / pixel depending on the sample particle size.

[0026] S202, dark current correction and reflectance conversion are performed on the acquired hyperspectral image data to obtain corrected reflectance hyperspectral image data; It should be noted that relative reflectance is calculated using the following formula: In the formula, Indicates the wavelength of the band; Indicates the sample image in the band The original grayscale value at that location; Indicates the dark reference image in the band The average gray value at the location (if multiple images are collected, take the average). Indicates the whiteboard image in the band The average gray value at that location; This represents the known reflectivity of a standard whiteboard, usually taken as 1, assuming the whiteboard is an ideal diffuse reflector.

[0027] By normalizing the calculated reflectance values ​​to the range of 0 to 1 and recombining them according to the original spatial and spectral dimensions, corrected hyperspectral reflectance image data is generated.

[0028] Dark current correction is achieved by acquiring a dark reference image, and reflectivity conversion is achieved by acquiring a standard white board image and calculating the relative reflectivity.

[0029] It should be noted that the dark reference image is obtained by turning off the illumination source or completely covering the camera lens with an opaque object while keeping the imaging system parameters (integration time, gain, scanning speed, etc.) exactly the same as those of the sample acquisition, and then starting the hyperspectral imaging system to acquire the image; the white board reference image is obtained by placing a standard white board with known high reflectivity at the sample imaging position and acquiring it under the same illumination conditions and system parameters.

[0030] S3 uses wavelet transform algorithm to denoise the collected spectral data and expands the training sample set through data augmentation technology; S301, the acquired hyperspectral image data cube is subjected to pixel-by-pixel extraction of the spectral dimension to obtain the original spectral curve of each pixel; S302 performs discrete wavelet transform decomposition on each original spectral curve, decomposing the spectral signal into approximation coefficients and detail coefficients at different scales; It should be noted that the wavelet basis functions are selected from the Daubechies series db4, db6, and db8, with db6 being the preferred option; the number of decomposition levels L is adaptively determined based on the spectral data length N and the noise level, satisfying the following conditions: The preferred number of decomposition layers is 3-5.

[0031] S303 performs thresholding on the decomposed detail coefficients to remove noise components; Thresholding uses a soft threshold function, and the threshold selection method is BayesShrink adaptive threshold. It should be noted that the detail coefficient thresholding process is as follows: the detail coefficients of layers 1 to L are thresholded respectively, the coefficients greater than the threshold are retained, and the coefficients less than the threshold are set to zero or shrunk to zero to remove noise components; the approximation coefficient AL remains unchanged to preserve the overall trend of the spectrum and the main absorption characteristics.

[0032] S304, perform inverse wavelet transform on the coefficients after thresholding to reconstruct the denoised spectral curve; S305. Data augmentation techniques are used to augment the training sample set on the denoised spectral curves to obtain the augmented spectral dataset.

[0033] Data augmentation includes noise enhancement, translation enhancement, scaling enhancement, and baseline change enhancement; Noise enhancement is achieved by adding Gaussian white noise with different signal-to-noise ratios to the denoised spectral curves to generate noise-enhanced samples. It should be noted that the signal-to-noise ratio ranges from 20dB to 50dB, preferably 30dB; the step size is 5dB.

[0034] Translation enhancement generates translation-enhanced samples by performing a small translation transformation on the denoised spectral curve, randomly shifting it by 1-3 bands along the wavelength direction. It should be noted that the translation range is 1-5 bands, and cyclic filling or edge filling is used to keep the spectral length unchanged.

[0035] Scaling enhancement generates scaled-enhanced samples by multiplicatively scaling the denoised spectral curve and multiplying it by random coefficients. It should be noted that the random coefficient is in the range of 0.95-1.05, preferably 1.05.

[0036] Baseline variation enhancement generates baseline variation enhanced samples by simulating baseline drift on the denoised spectral curve and adding a quadratic polynomial trend term.

[0037] S4. Construct a machine learning model to extract deep nonlinear features and combine it with diagnostic absorption features of mineral spectra for multi-scale feature fusion. S401 takes the magnesite hyperspectral image data after wavelet transform denoising and data enhancement as input, and performs normalization processing on the spectral curve of each pixel to obtain a normalized spectral feature vector. S402 constructs a deep neural network consisting of multiple sparse autoencoders stacked together, and learns deep nonlinear features of spectral data through layer-by-layer unsupervised pre-training. It should be noted that sparse autoencoders need to add a sparsity constraint term to the loss function so that the average activation of the hidden layer neurons is close to the preset sparse target parameters, thereby learning a sparser and more effective feature representation.

[0038] A sparse autoencoder consists of an encoding layer and a decoding layer; S403, the normalized spectral feature vector is input into the first sparse autoencoder, the first hidden layer features are learned by minimizing the reconstruction error, and the first hidden layer features are used as the input of the second sparse autoencoder to learn the second hidden layer features. This process is repeated until the deep abstract features are obtained.

[0039] It should be noted that the maximum number of layers in the autoencoder is 5, preferably 3; the number of hidden nodes in each layer decreases progressively, forming a bottleneck structure, thereby achieving progressive dimensionality reduction and nonlinear feature extraction of high-dimensional spectral data.

[0040] S404 integrates the deep nonlinear features extracted by the stacked autoencoder with the diagnostic feature parameters of mineral spectra at multiple scales to construct a joint feature vector. It should be noted that the process of constructing the joint feature vector is as follows: In the formula, Represents deep abstract features; This indicates that the m-th diagnostic feature parameter is denoted as .

[0041] The diagnostic parameters of the mineral spectrum are extracted from the original spectral curve based on the mineral spectrum formation mechanism of magnesite.

[0042] Diagnostic spectroscopic features include carbonate ion diagnostic features, hydroxyl ion diagnostic features, water molecule diagnostic features, and iron ion diagnostic features; The diagnostic characteristics of carbonate ions are: the position, depth and symmetry of the absorption peaks in the range of 2300-2350 nm. It should be noted that the diagnostic characteristics of carbonate ions may also include the asymmetry coefficient of the absorption peak and the absorption area parameter, which are used to distinguish magnesite from other carbonate minerals.

[0043] The diagnostic features of hydroxyl groups are: extracting the position and depth of the absorption peaks near 1400 nm and 1900 nm; It should be noted that the diagnostic features of hydroxyl groups may also include the full width at half maximum (FWHM) and slope change rate of the 1400 nm absorption peak, reflecting the hydroxyl content and crystallinity information in magnesite.

[0044] The diagnostic features for water molecules are: extracting absorption characteristic parameters around 1900 nm; The diagnostic features of iron ions are: extracting electronic transition absorption characteristic parameters in the range of 400-1200 nm.

[0045] S5. Construct a learning machine classification model, using the fused multi-scale features as input and the ore grade as output, to train the model; S501, construct a single hidden layer feedforward neural network as a classification model, set the number of input layer neurons to be equal to the fusion feature dimension, the number of output layer neurons to be equal to the number of grade categories, and select an activation function for the hidden layer neurons; The number of neurons in the hidden layer is determined by cross-validation, and the value ranges from 50 to 500. It should be noted that the activation function is selected from one of the following: Sigmoid function, Sine function, radial basis function, Hardlim function, ReLU function, or LeakyReLU function, with Sigmoid function being preferred.

[0046] S502, randomly generate the connection weight matrix between the input layer and the hidden layer and the bias vector of the hidden layer; The weights and biases remain unchanged throughout the training process after initialization; S503 inputs the feature matrix after normalizing the joint feature vector into the network and calculates the hidden layer output matrix through the activation function. It should be noted that the feature matrix is ​​obtained by using Z-score normalization to obtain the fused multi-scale feature vectors and their corresponding magnesite grade labels after preprocessing and feature extraction.

[0047] The formula for calculating the hidden layer output matrix is ​​as follows: In the formula, Indicates the activation function; This represents the normalized feature matrix; This represents the weight matrix of the hidden layer; This represents the hidden layer bias vector.

[0048] S504 obtains the output layer weight matrix by solving the least squares problem, and then combines the obtained output weights with the randomly generated hidden layer weight matrix and hidden layer bias vector to form a complete classification model.

[0049] It should be noted that the formula for calculating the output layer weight matrix is ​​as follows: In the formula, This represents the uniquely encoded matrix converted from ore grade labels; This represents the regularization coefficient, used to prevent overfitting, and its preferred value range is... to ; Represents the identity matrix.

[0050] S6. The trained model is evaluated using a noise robustness test, and the model parameters are optimized based on the evaluation results. S601 extracts a subset of test samples independent of the training set from the magnesite sample dataset obtained after feature extraction and fusion, and adds noise simulating the industrial field environment to the spectral feature data of the test samples. It should be noted that the sample subset for testing was selected using a stratified random sampling method to ensure that there are no fewer than 20 samples in each grade category and that the total number of test samples is no fewer than 100.

[0051] Noise types include Gaussian white noise, impulse noise, and baseline drift noise; It should be noted that the Gaussian white noise is added as follows: a Gaussian white noise sequence is generated and superimposed on the original spectral features according to the preset signal-to-noise ratio range (10dB to 40dB, with a step size of 5dB and a total of 7 noise intensity levels); the impulse noise is added as follows: a portion of the spectral features is randomly selected, and its values ​​are replaced with preset maxima or minima (the maxima is 1.5 times the maximum value of the original spectral features, and the minima is 0.5 times the minimum value of the original spectral features); the baseline drift noise is added as follows: a linear function with a random slope is generated as the baseline drift term and superimposed on the original spectral features.

[0052] S602: Set multiple intensity levels for each type of noise, input the spectral features after adding noise of different intensities into the trained extreme learning machine classification model, and calculate the classification accuracy at each intensity level. It should be noted that when calculating the classification accuracy, multiple independent repeated tests are performed for each noise intensity level. A new noise sequence is randomly generated for each test, and the average accuracy of multiple tests is taken as the final accuracy at that intensity. The number of repeated tests is no less than 10.

[0053] S603 calculates the noise resistance performance index of the model based on the test results of classification accuracy, and evaluates the results based on the noise resistance performance index in order to adjust the hyperparameters of the extreme learning machine model. It should be noted that during the adjustment process, the optimized model parameters are saved until the average accuracy of the model within the preset noise intensity range reaches the threshold (preferably 85%) or the performance no longer improves.

[0054] The metrics include: the attenuation curve of accuracy as a function of noise intensity for each noise type, the accuracy retention rate at specific key noise intensities, and the area under the noise intensity versus accuracy curve. It should be noted that the accuracy retention rate is obtained by comparing the classification accuracy with added noise with the classification accuracy under noise-free conditions.

[0055] Hyperparameters include the number of neurons in the hidden layer, the regularization coefficient, and the type of activation function.

[0056] S7 deploys the optimized model to edge computing devices and connects with the industrial IoT platform to achieve automatic classification of magnesite ore grades; S701 performs lightweight processing on the optimized Extreme Learning Machine classification model and loads it into the edge computing device, configuring the model's runtime environment; It should be noted that configuring the model's runtime environment, including the operating system, deep learning inference framework, and dependency libraries, and performing model loading tests are necessary to ensure that the model can run normally on edge devices.

[0057] Lightweighting includes model pruning, weight quantization, and structure optimization; S702 inputs the generated fusion feature vector into the deployed extreme learning machine model to calculate the ore grade classification result, records the classification confidence, and stores the classification result and corresponding spectral data locally.

[0058] It should be noted that the fused feature vector is generated by receiving magnesite ore sample data collected by the hyperspectral imaging system in real time through edge computing devices, and automatically performing spectral extraction, wavelet transform denoising and feature extraction processes.

[0059] Edge computing devices connect to the industrial IoT platform via MQTT, CoAP, or HTTP / HTTPS protocols, uploading real-time classification results, device operating status, and anomaly information to the cloud platform.

[0060] Furthermore, it should be noted that the present invention can be provided as a method, apparatus, or computer program product. Therefore, embodiments of the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0061] The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0062] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, including a defined element by a statement does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0064] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A machine learning-based intelligent classification method for magnesite ore grades, characterized in that, include: S1. Collect magnesite ore samples of different grades and perform chemical analysis on each sample to determine its true grade label. S2, using a hyperspectral imager to acquire raw hyperspectral image data of ore samples; S3 uses wavelet transform algorithm to denoise the collected spectral data and expands the training sample set through data augmentation technology; S4. Construct a machine learning model to extract deep nonlinear features and combine it with diagnostic absorption features of mineral spectra for multi-scale feature fusion. S5. Construct a learning machine classification model, using the fused multi-scale features as input and the ore grade as output, to train the model; S6. The trained model is evaluated using a noise robustness test, and the model parameters are optimized based on the evaluation results. The S7 deploys the optimized model to edge computing devices and interfaces with an industrial IoT platform to achieve automatic classification of magnesite ore grades.

2. The intelligent classification method for magnesite ore grades based on machine learning as described in claim 1, characterized in that, S1 collects magnesite ore samples of different grades and performs chemical analysis on each sample to determine its true grade label, wherein: S101, collect magnesite ore samples covering the range of special grade, first grade, second grade and third grade; The samples include blocky ore samples from different mining areas and powdery samples of different particle sizes; S102. For each collected sample, chemical analysis is performed according to national standards to determine the content of magnesium oxide, calcium oxide, and silicon dioxide, and the true grade label of each sample is determined by referring to metallurgical industry standards.

3. The intelligent classification method for magnesite ore grades based on machine learning as described in claim 1, characterized in that, S2 utilizes a hyperspectral imager to acquire raw hyperspectral image data of the ore sample, wherein: S201, using a hyperspectral imager to perform push-broom scanning on magnesite ore to obtain raw hyperspectral image data of the ore sample; S202, dark current correction and reflectance conversion are performed on the acquired hyperspectral image data to obtain corrected reflectance hyperspectral image data; The dark current correction is achieved by acquiring a dark reference image, and the reflectivity conversion is achieved by acquiring a standard white board image and calculating the relative reflectivity.

4. The intelligent classification method for magnesite ore grades based on machine learning as described in claim 1, characterized in that, S3 employs a wavelet transform algorithm to denoise the acquired spectral data and augments the training sample set using data augmentation techniques, wherein: S301, the acquired hyperspectral image data cube is subjected to pixel-by-pixel extraction of the spectral dimension to obtain the original spectral curve of each pixel; S302 performs discrete wavelet transform decomposition on each original spectral curve, decomposing the spectral signal into approximation coefficients and detail coefficients at different scales; S303 performs thresholding on the decomposed detail coefficients to remove noise components; The threshold processing uses a soft threshold function, and the threshold selection method uses the BayesShrink adaptive threshold. S304, perform inverse wavelet transform on the coefficients after thresholding to reconstruct the denoised spectral curve; S305. Data augmentation techniques are used to augment the training sample set on the denoised spectral curves to obtain the augmented spectral dataset.

5. The intelligent classification method for magnesite ore grades based on machine learning as described in claim 4, characterized in that, The denoised spectral curves are augmented using data augmentation techniques to increase the training sample set, resulting in an augmented spectral dataset, wherein: The data enhancement includes noise enhancement, translation enhancement, scaling enhancement, and baseline change enhancement; The noise enhancement is achieved by adding Gaussian white noise with different signal-to-noise ratios to the denoised spectral curve to generate noise-enhanced samples. The translation enhancement is achieved by performing a small translation transformation on the denoised spectral curve, randomly shifting it by 1-3 bands along the wavelength direction to generate a translation enhancement sample. The scaling enhancement is achieved by multiplicatively scaling the denoised spectral curve and multiplying it by a random coefficient to generate a scaled enhanced sample. The baseline change enhancement is achieved by simulating baseline drift on the denoised spectral curve and adding a quadratic polynomial trend term to generate a baseline change enhancement sample.

6. The intelligent classification method for magnesite ore grades based on machine learning as described in claim 1, characterized in that, The S4 method constructs a machine learning model to extract deep nonlinear features and combines them with diagnostic absorption features from mineral spectra for multi-scale feature fusion, wherein: S401 takes the magnesite hyperspectral image data after wavelet transform denoising and data enhancement as input, and performs normalization processing on the spectral curve of each pixel to obtain a normalized spectral feature vector. S402 constructs a deep neural network consisting of multiple sparse autoencoders stacked together, and learns deep nonlinear features of spectral data through layer-by-layer unsupervised pre-training. S403, the normalized spectral feature vector is input into the first sparse autoencoder, the first hidden layer features are learned by minimizing the reconstruction error, and the first hidden layer features are used as the input of the second sparse autoencoder to learn the second hidden layer features. This process is repeated until the deep abstract features are obtained. S404 integrates the deep nonlinear features extracted by the stacked autoencoder with the diagnostic feature parameters of mineral spectra at multiple scales to construct a joint feature vector. The mineral spectral diagnostic characteristic parameters are extracted from the original spectral curves based on the mineral spectral formation mechanism of magnesite.

7. The intelligent classification method for magnesite ore grades based on machine learning as described in claim 6, characterized in that, The mineral spectral diagnostic characteristic parameters are extracted from the original spectral curve based on the mineral spectral formation mechanism of magnesite, wherein: The diagnostic spectral features include carbonate ion diagnostic features, hydroxyl ion diagnostic features, water molecule diagnostic features, and iron ion diagnostic features. The diagnostic features of carbonate ions are: extracting the position, absorption depth and absorption peak symmetry in the range of 2300-2350 nm; The diagnostic features of the hydroxyl group are: extracting the position and absorption depth of the absorption peaks near 1400 nm and 1900 nm; The water molecule diagnostic features are: extracting absorption characteristic parameters near 1900 nm; The diagnostic features for iron ions are: extracting electronic transition absorption characteristic parameters in the range of 400-1200 nm.

8. The intelligent classification method for magnesite ore grades based on machine learning as described in claim 1, characterized in that, The S5 constructs a learning machine classification model, using the fused multi-scale features as input and the ore grade as output, to train the model, wherein: S501, construct a single hidden layer feedforward neural network as a classification model, set the number of input layer neurons to be equal to the fusion feature dimension, the number of output layer neurons to be equal to the number of grade categories, and select an activation function for the hidden layer neurons; The number of hidden layer neurons is determined by cross-validation, and the value ranges from 50 to 500. S502, randomly generate the connection weight matrix between the input layer and the hidden layer and the bias vector of the hidden layer; The weights and biases remain unchanged throughout the training process after initialization; S503 inputs the feature matrix after normalizing the joint feature vector into the network and calculates the hidden layer output matrix through the activation function. S504 obtains the output layer weight matrix by solving the least squares problem, and then combines the obtained output weights with the randomly generated hidden layer weight matrix and hidden layer bias vector to form a complete classification model.

9. The intelligent classification method for magnesite ore grades based on machine learning as described in claim 1, characterized in that, S6 uses a noise robustness test to evaluate the trained model and optimizes the model parameters based on the evaluation results, wherein: S601 extracts a subset of test samples independent of the training set from the magnesite sample dataset obtained after feature extraction and fusion, and adds noise simulating the industrial field environment to the spectral feature data of the test samples. The noise types include Gaussian white noise, impulse noise, and baseline drift noise; S602: Set multiple intensity levels for each type of noise, input the spectral features after adding noise of different intensities into the trained extreme learning machine classification model, and calculate the classification accuracy at each intensity level. S603 calculates the noise resistance performance index of the model based on the test results of classification accuracy, and evaluates the results based on the noise resistance performance index in order to adjust the hyperparameters of the extreme learning machine model. The metrics include: the attenuation curve of accuracy as a function of noise intensity for each noise type, the accuracy retention rate at specific key noise intensities, and the area under the noise intensity versus accuracy curve. The hyperparameters include the number of neurons in the hidden layer, the regularization coefficient, and the activation function type.

10. The intelligent classification method for magnesite ore grades based on machine learning as described in claim 1, characterized in that, The S7 deploys the optimized model to an edge computing device and interfaces with an industrial IoT platform to achieve automatic classification of magnesite ore grades, wherein: S701 performs lightweight processing on the optimized Extreme Learning Machine classification model and loads it into the edge computing device, configuring the model's runtime environment; The lightweighting process includes model pruning, weight quantization, and structure optimization. S702 inputs the generated fusion feature vector into the deployed extreme learning machine model to calculate the ore grade classification result, records the classification confidence, and stores the classification result and corresponding spectral data locally.

Images