A multimodal photoelectric fusion coal gangue purification kaolin separation method and system

Abstract

The application discloses a multi-modal photoelectric fusion coal gangue purification kaolin separation method and system, and the multi-modal photoelectric fusion coal gangue purification kaolin separation method adopts a gradient separation architecture: pretreatment is subjected to narrow particle grading and surface purification to expose mineral characteristics; a first-stage XRT+RGB multi-modal fusion removes coal and enriches, late fusion and priority determination logic in a decision layer are used, that is, XRT high density is prior to RGB black to solve black gangue same color misjudgment, and fuzzy sample flow is transferred to a second stage; a second-stage RGB+NIR spectrum selection removes iron-containing impurity gangue and non-kaolinite waste rock by using a double-flow AI model, the application has the beneficial effects that a single sensor identification blind area is eliminated, the purification quality is excellent, the method is green and environmentally friendly, a dry method is used throughout, the clean coal recovery rate is 87.1%, the Fe2O3 content of kaolin is reduced to 0.92%, the whiteness is 84.3%, the water consumption is 0.11t / t, coal and kaolin are efficiently and coordinately recovered, and the method is economical and efficient.

Description

Technical Field

[0001] This invention relates to the field of coal solid waste resource utilization technology, and in particular to a multimodal photoelectric fusion method and system for purifying and separating kaolin from coal gangue. Background Technology

[0002] In existing technologies, coal gangue is a large amount of solid waste generated during coal mining and washing. Long-term stockpiling not only occupies significant land resources but also easily causes environmental problems such as dust pollution, spontaneous combustion, and water and soil pollution. Kaolinite, an associated coal-bearing rock in most of my country's coalfields, is a high-quality raw material for ceramics, refractory materials, chemical fillers, and paper coatings, possessing high resource utilization value. Effectively purifying the kaolinite in coal gangue while simultaneously recovering coal resources is an important way to achieve high-value utilization of coal gangue. Currently, the main problems with coal gangue sorting and kaolinite purification processes are as follows: I. Traditional wet separation processes consume a lot of water and cause serious pollution in the tailwater. Heavy media separation, jigging separation and other wet processes mainly rely on the difference in material density for coarse separation. For kaolinite with similar density and ordinary impurities such as gangue, the separation accuracy is low. At the same time, the cost of washing water treatment is high and secondary pollution is easy to generate, making it difficult to meet the requirements of green and environmentally friendly production.

[0003] Second, the single X-ray transmission sorting technology has blind spots. X-ray transmission technology distinguishes coal from rocks based on the differences in internal density and atomic number of materials, but it cannot identify the yellow and brownish-red iron oxide impurities on the surface of kaolinite, nor can it determine whether the material contains kaolinite. This results in the iron content of the purified kaolin product being too high and the whiteness of the calcined product not meeting the standards.

[0004] Third, single RGB visible light visual sorting technology is prone to misjudgment. Coal-series kaolinite often has organic carbon attached to its surface, making it black. Its appearance is very similar to that of raw coal. Single RGB visual sorting can easily misjudge black kaolinite as coal, resulting in a large loss of kaolinite resources. At the same time, RGB vision is difficult to identify invalid waste rock with a surface color similar to kaolinite but without kaolinite components, such as quartz sandstone.

[0005] Fourth, existing kaolin iron removal and whitening processes are costly and pose a significant pollution risk. Conventional iron removal and whitening processes in the industry mostly employ post-process chemical bleaching and high-gradient magnetic separation, which involve large equipment investments, high operating costs, and chemical processes are prone to secondary pollution, making it impossible to achieve dry, green, and low-cost purification and production.

[0006] In summary, existing technologies suffer from bottlenecks such as low sorting accuracy, high missorting rate, poor environmental performance, high purification cost, and low resource utilization. There is a lack of a green and intelligent sorting process and system that can simultaneously achieve efficient coal recovery and precise iron removal and whitening of kaolin. Summary of the Invention

[0007] The purpose of this invention is to provide a multimodal photoelectric fusion method and system for purifying and separating kaolin from coal gangue.

[0008] To achieve the above objectives, the technical solution proposed by this invention is as follows: A multimodal photoelectric fusion method for purifying and separating kaolin from coal gangue includes the following steps: S1. Pre-treatment before selection: The raw coal gangue ore is crushed and classified into multiple independent particle size segments through narrow-level screening. Each particle size segment is then subjected to surface cleaning and desliming or dry heavy dust removal treatment to remove coal powder and mud adhering to the material surface, exposing the true surface color, texture and mineral characteristics of the material.

[0009] S2, First-stage composite photoelectric separation for coal enrichment: The pre-treated material is fed into the first-stage photoelectric separator, which simultaneously collects dual-energy XRT transmission image data and RGB visible light image data of the material. The equivalent atomic number and density characteristics of the material are extracted through XRT data, and the surface color and texture characteristics of the material are extracted through RGB data. A multi-modal data fusion algorithm is used for coupled analysis to identify low-density materials with coal texture characteristics as coal products and blow them out for recycling. The high-density mixture of kaolinite and impurity gangue is sent to the next process as a kaolin-rich mixture.

[0010] S3. Second-stage visual sorting for yellowing and iron removal: The kaolin-rich mixture is fed into the second-stage visual sorting machine, which simultaneously collects RGB image data and NIR near-infrared spectral data of the material. Through the RGB image recognition model and color threshold calibration, the machine identifies and removes iron oxide impurities such as gangue with yellow or brownish-red surfaces. The NIR near-infrared spectroscopy identifies the molecular bond characteristics of the material's minerals, and the machine removes invalid impurity rocks that do not contain kaolinite based on the hydroxyl absorption peak characteristics of kaolinite. The remaining material is the high-purity kaolinite product.

[0011] The multimodal data fusion algorithm has a priority judgment logic: when XRT detection determines that the material is high density and RGB detection determines that the material is black, the XRT density feature data is given priority and the material is judged as black kaolinite rather than coal; when the XRT and RGB output judgment results are within the preset fuzzy confidence interval, the coal blowing action is not performed and the material is sent to the second process.

[0012] In step S1, the narrow-level screening and grading particle size is divided into three grades: 10-30mm, 30-50mm, and 50-100mm. Different particle sizes are matched with corresponding photoelectric sorting equipment parameters and pneumatic spray valve parameters, and the particle size is sorted independently.

[0013] In step S2, the multimodal data fusion algorithm is a late-stage fusion algorithm at the decision layer. The features extracted from the dual-energy XRT image include the gray-level statistical features of the particle ROI, the transmission texture features inside the XRT gray-level co-occurrence matrix, the mean, variance, quantiles, statistical distribution features of the proportion of high atomic number pixels of the particle pixels, and particle morphology features. The features extracted from the RGB visible light image include the RGB three-channel statistical features, HSV color space features, hue color histogram, surface texture features of the RGB gray-level co-occurrence matrix, and particle morphology features. The XRT features and RGB features are input into independent classifiers, which output the confidence scores of coal and kaolinite / impurity gangue in the corresponding modalities.

[0014] Priority determination logic includes: setting a high confidence threshold. Low confidence threshold When the confidence levels of both XRT and RGB outputs are within a certain range ~ When the fuzzy confidence interval is reached, if the two modal discriminations are in the same direction, the basic weight weighted fusion is used; if the two modal discriminations are in opposite directions and conflict, the XRT modal weights are increased and the fusion confidence is recalculated. Particles whose fusion results are still within the fuzzy confidence interval are not subjected to coal blowing action and are sent to the second visual selection process for further identification. The basic weights are XRT weight 0.65 and RGB weight 0.35. When there is a conflict, the dynamic weights are adjusted to XRT weight 0.8 and RGB weight 0.2.

[0015] In step S3, the RGB image recognition model and the NIR spectral recognition model employ a dual-stream multimodal fusion network, including an RGB image branch, an NIR spectral branch, a feature fusion layer, and an output layer. The RGB image branch is a lightweight CNN convolutional network that extracts color and surface texture features and outputs an image feature vector. The NIR spectral branch is a one-dimensional convolutional CNN or MLP that extracts spectral feature vectors from the preprocessed spectral curves, focusing on learning the spectral fingerprint of the kaolinite hydroxyl absorption peak. The feature fusion layer concatenates the image feature vector and the spectral feature vector and then feeds them into a fully connected layer. The output layer uses... Output the three-class confidence score ,in, For qualified kaolinite confidence level, Confidence level for iron-containing gangue, The confidence level for invalid waste rock.

[0016] The NIR spectroscopy detection range is 1300–2500 nm, with a focus on covering the characteristic absorption peaks of kaolinite hydroxyl groups around 1400 nm and 2200 nm.

[0017] The training dataset of the dual-stream multimodal fusion network includes four types of particles: qualified kaolinite, iron-containing gangue, invalid waste rock, and interference samples. It also includes coal gangue material samples from at least two different origins. The training set, validation set, and test set are divided in a 7:1:2 ratio, and data augmentation is performed on the RGB images and NIR spectra respectively.

[0018] A multimodal photoelectric fusion coal gangue purification and kaolin sorting system is used to implement the above sorting method, including a pretreatment unit, a first-stage composite photoelectric sorting unit, a second-stage visual fine sorting unit, and a material collection unit connected in sequence. The pretreatment unit includes crushing equipment, grading and screening equipment, and dust removal and desliming equipment, which are used to complete coal gangue crushing, narrow particle size classification, and material surface purification treatment. The first composite photoelectric sorting unit includes a dual-energy XRT detection module, an RGB vision detection module, a multimodal data processing module, and a first pneumatic sorting execution module, which are used to fuse density and visual features to achieve precise separation of coal and kaolin-rich mixtures. The second visual selection unit includes an RGB color recognition module, a NIR near-infrared spectroscopy detection module, an AI intelligent recognition module, and a second pneumatic sorting execution module, used to remove iron-containing yellow gangue and non-kaolinite impurity rocks; The material collection unit includes a coal product collection bin, a high-purity kaolin product collection bin, and an impurity gangue discharge bin.

[0019] The multimodal data processing module has a built-in feature priority determination program. For black gangue with the same color interference, it locks the XRT high-density feature as the core determination criterion to avoid misjudgment and loss of black kaolin.

[0020] The second visual selection unit is equipped with an AI image recognition model and a spectral feature comparison database, which can adaptively identify iron-containing impurities and invalid mineral impurities in coal gangue from different origins.

[0021] The beneficial effects of this invention are: Eliminating blind spots in single-sensor identification, the clean coal recovery rate reached 87.1%, the Fe2O3 content of kaolin decreased to 0.92%, and the calcination whiteness reached 84.3%.

[0022] Excellent purification quality: Simultaneous iron and titanium removal increases Al2O3 to 38.75%, eliminating the need for downstream chemical bleaching or high-gradient magnetic separation.

[0023] Green and environmentally friendly: The entire process is dry physical separation, with a water consumption of only 0.11t per ton of ore, saving more than 96% of water compared to wet processes, and producing no wastewater or reagents.

[0024] Economical and efficient: The power consumption per ton of ore is 19.2 kWh / t, which is 41% more energy-efficient than the wet process. It can simultaneously recover coal and kaolin, realizing the full resource utilization of solid waste. Attached Figure Description

[0025] Figure 1 This is a flowchart of the sorting method of the present invention. Detailed Implementation

[0026] The present invention will now be described in further detail with reference to the accompanying drawings. Example 1: A multimodal photoelectric fusion method for purifying and separating kaolin from coal gangue includes the following steps: S1. Pre-treatment before selection: The raw coal gangue ore is crushed and classified into multiple independent particle size segments through narrow-level screening. Each particle size segment is then subjected to surface cleaning and desliming or dry heavy dust removal treatment to remove coal powder and mud adhering to the material surface, exposing the true surface color, texture and mineral characteristics of the material.

[0027] S2, First-stage composite photoelectric separation for coal enrichment: The pre-treated material is fed into the first-stage photoelectric separator, which simultaneously collects dual-energy XRT transmission image data and RGB visible light image data of the material. The equivalent atomic number and density characteristics of the material are extracted through XRT data, and the surface color and texture characteristics of the material are extracted through RGB data. A multi-modal data fusion algorithm is used for coupled analysis to identify low-density materials with coal texture characteristics as coal products and blow them out for recycling. The high-density mixture of kaolinite and impurity gangue is sent to the next process as a kaolin-rich mixture.

[0028] S3. Second-stage visual sorting for yellowing and iron removal: The kaolin-rich mixture is fed into the second-stage visual sorting machine, which simultaneously collects RGB image data and NIR near-infrared spectral data of the material. Through the RGB image recognition model and color threshold calibration, the machine identifies and removes iron oxide impurities such as gangue with yellow or brownish-red surfaces. The NIR near-infrared spectroscopy identifies the molecular bond characteristics of the material's minerals, and the machine removes invalid impurity rocks that do not contain kaolinite based on the hydroxyl absorption peak characteristics of kaolinite. The remaining material is the high-purity kaolinite product.

[0029] The multimodal data fusion algorithm includes a priority judgment logic: when XRT detection determines the material to be high-density and RGB detection determines the material to be black, the XRT density feature data is prioritized, classifying the material as black kaolinite rather than coal; when the XRT and RGB output results are within a preset fuzzy confidence interval, the coal blowing action is not performed, and the material is sent to the second stage process. The workflow diagram of the sorting method of this invention is shown below. Figure 1 As shown.

[0030] In step S1, the narrow-level screening and grading particle size is divided into three grades: 10-30mm, 30-50mm, and 50-100mm. Different particle sizes are matched with corresponding photoelectric sorting equipment parameters and pneumatic spray valve parameters, and the particle size is sorted independently.

[0031] In step S2, the multimodal data fusion algorithm is a late-stage fusion algorithm at the decision layer. The features extracted from the dual-energy XRT image include the gray-level statistical features of the particle ROI, the transmission texture features inside the XRT gray-level co-occurrence matrix, the mean, variance, quantiles, statistical distribution features of the proportion of high atomic number pixels of the particle pixels, and particle morphology features. The features extracted from the RGB visible light image include the RGB three-channel statistical features, HSV color space features, hue color histogram, surface texture features of the RGB gray-level co-occurrence matrix, and particle morphology features. The XRT features and RGB features are input into independent classifiers, which output the confidence scores of coal and kaolinite / impurity gangue in the corresponding modalities.

[0032] Priority determination logic includes: setting a high confidence threshold. Low confidence threshold When the confidence levels of both XRT and RGB outputs are within a certain range ~ When the fuzzy confidence interval is reached, if the two modal discriminations are in the same direction, the basic weight weighted fusion is used; if the two modal discriminations are in opposite directions and conflict, the XRT modal weights are increased and the fusion confidence is recalculated. Particles whose fusion results are still within the fuzzy confidence interval are not subjected to coal blowing action and are sent to the second visual selection process for further identification. The basic weights are XRT weight 0.65 and RGB weight 0.35. When there is a conflict, the dynamic weights are adjusted to XRT weight 0.8 and RGB weight 0.2.

[0033] In step S3, the RGB image recognition model and the NIR spectral recognition model employ a dual-stream multimodal fusion network, including an RGB image branch, an NIR spectral branch, a feature fusion layer, and an output layer. The RGB image branch is a lightweight CNN convolutional network that extracts color and surface texture features and outputs image feature vectors. The NIR spectral branch is a one-dimensional convolutional CNN or MLP that extracts spectral feature vectors from the preprocessed spectral curves, focusing on learning the spectral fingerprint of the kaolinite hydroxyl absorption peak. The feature fusion layer concatenates the image feature vector and the spectral feature vector and then feeds them into a fully connected layer. The output layer uses... Output the three-class confidence score ,in, For qualified kaolinite confidence level, Confidence level for iron-containing gangue, The confidence level for invalid waste rock.

[0034] The NIR spectroscopy detection range is 1300–2500 nm, with a focus on covering the characteristic absorption peaks of kaolinite hydroxyl groups around 1400 nm and 2200 nm.

[0035] The training dataset for the dual-stream multimodal fusion network includes four types of particles: qualified kaolinite, iron-containing gangue, invalid waste rock, and interference samples. It also includes coal gangue material samples from at least two different origins. The training set, validation set, and test set are divided in a 7:1:2 ratio, and data augmentation is performed on the RGB images and NIR spectra respectively.

[0036] Example 2: A multimodal photoelectric fusion coal gangue purification and kaolin sorting system is used to implement the above sorting method, including a pretreatment unit, a first-stage composite photoelectric sorting unit, a second-stage visual fine sorting unit, and a material collection unit connected in sequence. The pretreatment unit includes crushing equipment, grading and screening equipment, and dust removal and desliming equipment, which are used to complete the crushing of coal gangue, narrow particle size classification, and surface purification of materials. The first composite photoelectric sorting unit includes a dual-energy XRT detection module, an RGB vision detection module, a multimodal data processing module, and a first pneumatic sorting execution module, which are used to fuse density and visual features to achieve precise separation of coal and kaolin-rich mixtures. The second visual selection unit includes an RGB color recognition module, a NIR near-infrared spectroscopy detection module, an AI intelligent recognition module, and a second pneumatic sorting execution module, used to remove iron-containing yellow gangue and non-kaolinite impurity rocks. The material collection unit includes a coal product collection bin, a high-purity kaolin product collection bin, and an impurity gangue discharge bin.

[0037] The multimodal data processing module has a built-in feature priority determination program. For black gangue with the same color interference, it locks the XRT high-density features as the core determination basis to avoid misjudgment and loss of black kaolin.

[0038] The second visual selection unit is equipped with an AI image recognition model and a spectral feature comparison database, which can adaptively identify iron-containing impurities and invalid mineral impurities in coal gangue from different origins.

[0039] Example 3: This embodiment uses raw coal gangue ore from a certain mining area as raw material. The main mineral composition of the raw ore is kaolinite, quartz, feldspar, and a small amount of iron-bearing minerals. The initial Al2O3 content of the raw ore is 21.35%, Fe2O3 content is 4.28%, and TiO2 content is 1.86%. The whiteness of the raw ore after calcination is only 38.7% without purification. After primary crushing by a jaw crusher, the particle size of the raw ore is ≤150mm, and it enters the subsequent sorting process.

[0040] S1. Pre-processing before selection: The crushed coal gangue material is fed into a vibrating classifier screen, and a narrow-level screening process is used to separate the material into three independent particle size ranges: 10-30mm, 30-50mm, and 50-100mm. The screening efficiency is ≥90%. Each particle size is then subjected to surface purification treatment using a dry high-power dust removal device. The dust removal air pressure is controlled at 0.4-0.6MPa. High-speed airflow is used to rub and wash the particle surface to remove the coal dust and mud covering the surface, so that the true surface color, texture, and mineral characteristics of the particles are fully exposed. The surface cleanliness of the treated material is visually inspected, and the coal dust adhesion area on the particle surface is ≤5%.

[0041] S2, First stage composite photoelectric separation for coal enrichment: The pre-treated materials of each particle size are fed into the first-stage composite photoelectric sorter of the corresponding specification. The sorter integrates a dual-energy XRT detection module and an RGB vision detection module to simultaneously acquire the dual-energy XRT transmission image and RGB visible light image of each material.

[0042] XRT Feature Extraction: For particle segmentation of the masked ROI region, the following feature vectors are extracted on a per-particle basis. : Gray-scale statistical characteristics: Average gray-scale of particle ROI Gray-scale variance Gray median 25th percentile of grayscale 75th percentile of grayscale .

[0043]

[0044] Transmissive texture features: Calculate the gray-level co-occurrence matrix (GLCM) based on XRT grayscale images and extract entropy. ,energy Contrast Inverse difference moment It reflects the internal composition, interlayers, and internal transmission texture information of particles.

[0045] Equivalent atomic number statistical distribution characteristics: Solving the equivalent atomic number of each pixel using dual-energy XRT. The mean of the equivalent atomic number of all pixels within the ROI of the particle is calculated. ,variance 25th percentile 75th percentile High atomic number pixel ratio .

[0046]

[0047]

[0048] Particle morphology characteristics: Particle projected area equivalent diameter Aspect Ratio Solidity .

[0049] XRT feature vectors:

[0050]

[0051] RGB feature extraction: ROI extraction for the same particle registration It characterizes the surface color, surface texture, and shape of materials, and is easily affected by surface coal dust contamination.

[0052] Color characteristics: R / G / B three-channel mean , , Hue mean in HSV color space Mean saturation ,brightness H channel color histogram Hue color histogram … The percentage of pixels in different color tone ranges is statistically analyzed by bin.

[0053] Surface texture features: After converting the RGB image to grayscale, calculate the GLCM and extract the surface texture entropy. Contrast .

[0054] Particle morphology characteristics: contour roughness, convex hull density Aspect Ratio .

[0055] RGB feature vector: ,

[0056]

[0057] Multimodal data fusion decision: This embodiment employs a late-stage fusion algorithm at the decision level to... Input a pre-trained XRT classifier and use a lightweight gradient boosting machine. Output the confidence level of belonging to coal. And the confidence level of belonging to kaolinite / impurity gangue ,and ,Will Input pre-trained The classifier uses a support vector machine. Output and ,and .

[0058] Set process calibration threshold: High confidence threshold Low confidence threshold Coal blowing out judgment threshold Basic weight , Dynamic weights for conflict fuzzy intervals , .

[0059] The logic behind the fusion decision is as follows: (1) Hard priority triggering condition: when and When XRT determines it to be high density and RGB determines it to be black, the setting is directly forced. , The material was determined to be kaolinite-impurity gangue, and the coal blowing operation was not performed. The threshold for determining high density in XRT is set to 0.75. The threshold for determining black in RGB is set to 0.65. This logic effectively prevents black kaolinite with coal dust adhering to its surface from being misidentified as coal.

[0060] (2) Weighted fusion of normal operating conditions: When the hard priority condition is not triggered, the calculation is as follows: , .

[0061] (3) Handling when both modes are within the fuzzy confidence interval: when and All in Interval time: If the two modes are determined to be in the same direction, both leaning towards coal or both leaning towards gangue, the basic weight fusion is used, coal blowing is not performed, and the data is directly sent to the second stage. Selected units, relying on kaolinite Hydroxyl group characteristics are used for secondary screening.

[0062] If the two modes conflict in their discrimination, such as XRT slightly favoring gangue while RGB slightly favors coal, or vice versa, then increase the XRT weight and recalculate using dynamic weights. .

[0063] If the fusion result after dynamic weighting is still in the fuzzy range If the coal is not blown out, the particle will be sent to the second stage of visual sorting for further identification.

[0064] (4) Final selection decision: If If it is determined to be coal, it will be blown out and recovered by the first pneumatic sorting execution module and enter the coal product collection bin. All other cases were determined to be kaolin-rich mixtures and sent to the second stage of visual sorting process.

[0065] S3, Second Visual Selection: Removing Yellow Staphylococcus Acid and Iron Substances The kaolin-rich mixture obtained from the first stage of sorting is sent to the second stage visual sorting machine. The second stage sorting machine is equipped with an RGB color recognition module and a NIR near-infrared spectral detection module, and works with an AI intelligent recognition module to perform accurate sorting.

[0066] RGB Image Recognition: An industrial color camera is used to capture RGB images of a single particle. After ROI segmentation, the images are input into an AI recognition model. The AI ​​recognition model is based on a lightweight convolutional neural network, i.e., CNN, which recognizes the color and texture of the particle surface. Combined with color threshold calibration, it accurately identifies iron oxide impurity gangue with yellow or brownish-red surfaces. When the proportion of yellow / brownish-red pixels on the particle surface is ≥8%, it is determined to be iron impurity gangue and is removed by the second pneumatic sorting execution module and enters the impurity gangue discharge bin.

[0067] NIR spectral identification: The NIR reflectance spectra of the particles were acquired simultaneously, with a spectral detection range of 1300–2500 nm, focusing on the 1400 nm and 2200 nm range of the kaolinite hydroxyl characteristic absorption peaks. The acquired spectra were preprocessed sequentially with baseline correction, multivariate scattering correction (MSC), and spectral normalization. The effective characteristic band of 1300–2500 nm was extracted. The preprocessed spectral curves were input into a one-dimensional convolutional neural network branch, i.e., 1D-CNN, to extract spectral feature vectors, with a focus on learning the spectral fingerprint of the kaolinite hydroxyl absorption peak.

[0068] Dual-stream multimodal fusion network: The AI ​​intelligent recognition module adopts a dual-stream multimodal fusion network, including: RGB Image Branch: A lightweight CNN convolutional network takes a granular ROI color image as input and outputs an image feature vector after multiple convolutions and pooling. .

[0069] NIR spectral branch: 1D-CNN or MLP, input is the preprocessed spectral curve, output is the spectral feature vector. .

[0070] Feature fusion layer: and After splicing, it is sent into the fully connected layer.

[0071] Output layer: Output the three-class confidence score ,in For qualified kaolinite confidence level, Confidence level for iron-containing gangue, The confidence level for invalid waste rock.

[0072] Sorting decision rule: The category corresponding to the highest confidence level is taken as the final judgment result.

[0073] If it is determined to be qualified kaolinite, it will enter the high-purity kaolin product collection bin; if it is determined to be iron-containing gangue or ineffective waste rock, it will be removed to the impurity gangue discharge bin.

[0074] Experimental results: This embodiment adopts the above-mentioned complete process, with a clean coal recovery rate of 87.1%, and the Al2O3 content of kaolin concentrate is 38.75%, the Fe2O3 content is reduced to 0.92%, the TiO2 content is reduced to 0.63%, the calcination whiteness is increased to 84.3%, the water consumption per ton of ore is only 0.11t, and the power consumption per ton of ore is 19.2kW·h / t.

[0075] Example 4: The adaptive sorting of coal gangue from different origins differs from Example 3 in that the coal gangue being processed comes from two different mining areas, Mining Area A and Mining Area B. The coal gangue from the two mining areas has significant differences in mineral composition, surface iron staining, and coal content. The coal gangue from Mining Area A has more severe surface iron staining, with yellow / brownish-red impurities accounting for about 18%; the coal gangue from Mining Area B has a higher carbon content, with black kaolinite accounting for about 32%, and a larger amount of coal powder adhering to the surface.

[0076] AI model training dataset construction: Actual particulate matter was collected from mining areas A and B after crushing and narrow-segmentation, covering particle sizes of 10–30 mm, 30–50 mm, and 50–100 mm. The samples contained four types of actual particles: 1. Qualified kaolinite particles: contain kaolinite, and have no obvious yellowish-brown iron staining on the surface; 2. Iron-containing gangue: The surface is stained with yellow or brownish-red iron oxides, and the interior may or may not contain kaolinite; 3. Ineffective waste rock: Quartz sandstone and limestone, whose mineral composition does not contain kaolinite; 4. Interference samples: Interference samples in the field conditions such as surface coal dust, local mud, uneven lighting, particle stacking, and partial particle obstruction.

[0077] The total sample size is ≥8000 particles, divided into a 7:1:2 ratio: 70% training set, 10% validation set, and 20% test set. Samples of different particle sizes and categories are kept balanced. Physical mineral analysis is used, specifically X-ray diffraction (XRD) combined with manual visual annotation. Each RGB image + NIR spectrum sample is labeled with three categories: Label0—Qualified kaolinite, Label1—Iron-containing impurity gangue, and Label2—Invalid non-kaolinite waste rock.

[0078] Data preprocessing and augmentation: RGB image preprocessing: ROI cropping, illumination normalization; image enhancement includes random brightness, contrast perturbation, slight rotation, flipping, and simulating local dust occlusion.

[0079] NIR spectral preprocessing: baseline correction, multivariate scattering correction (MSC), spectral normalization, extraction of effective characteristic bands from 1300 to 2500 nm, and enhancement of the characteristic peak signals of kaolinite hydroxyl groups at 1400 nm and 2200 nm.

[0080] Model training: The dual-stream multimodal fusion network uses the cross-entropy loss function, with an initial learning rate of 0.001 and the Adam optimizer. After 50 epochs of training, the model achieves a three-class classification accuracy of 94.2% on the test set, including 95.1% accuracy for identifying qualified kaolinite, 92.8% accuracy for identifying iron-containing gangue, and 93.6% accuracy for identifying invalid waste rock.

[0081] Sorting results: Materials from mines A and B were sorted separately, and the resulting kaolin concentrate indicators were as follows: product from mine A had an Fe2O3 content of 0.88% and a calcination whiteness of 85.1%; product from mine B had an Fe2O3 content of 0.95% and a calcination whiteness of 83.7%. The results show that the AI ​​model in this embodiment has good adaptive recognition ability for coal gangue from two different origins and can run stably without recalibrating the model parameters.

[0082] Example 5: The specific implementation of the first-stage composite photoelectric sorting algorithm is described in detail below, using three typical particles as an example: Particle 1: Coal particles with clean surfaces.

[0083] XRT characteristics: mean equivalent atomic number grayscale mean Low density, uniform transmission texture, exhibiting typical characteristics of low-density organic matter. Classifier output , .

[0084] RGB characteristics: The color is black, the gloss is weak, the texture is relatively uniform, and the RGB classifier output is... , .

[0085] Fusion computing: does not trigger hard priority, because Low, using basic weights: It was determined to be coal and was blown out for recycling.

[0086] Particle 2: Black kaolinite with coal powder adhering to its surface.

[0087] XRT characteristics: mean equivalent atomic number High atomic number pixel ratio Its density characteristics are significantly higher than those of coal. Classifier output , .

[0088] Characteristics: The surface is black, and its appearance is very similar to that of coal. Classifier output , .

[0089] Fusion computing: Triggers a hard priority condition, i.e. ≥ 0.75 and ≥ 0.65, mandatory judgment: , If identified as kaolinite / impurity gangue, it is not blown out and proceeds to the second stage of the process, effectively avoiding the misidentification of black kaolinite as coal.

[0090] Particle 3: Particles with indistinguishable boundaries.

[0091] Feature: Mean of equivalent atomic number It lies between typical coal and typical gangue. Classifier output , .

[0092] Characteristics: The surface is dark gray with a slight amount of coal dust adhering to it. Classifier output , .

[0093] Fusion computation: Both modal discriminations are within the fuzzy confidence interval and exhibit reverse conflict, i.e. Slightly biased towards coal, Slight tendency towards gangue, improving Weight: , The fusion result is still in the ambiguous range, so coal blowing is not performed; instead, the coal is sent to the second stage of visual sorting. The second stage... Further identification using the hydroxyl spectrum characteristics of kaolinite: if a hydroxyl absorption peak at 1400 nm / 2200 nm is detected, it is identified as kaolinite; otherwise, it is identified as waste rock or containing iron impurities.

[0094] Example 6: This embodiment of the optimization system with closed-loop feedback control, based on Embodiment 3, adds a feedback control loop between the first-stage composite photoelectric sorting unit and the second-stage visual sorting unit. The second-stage visual sorter uses an AI model to statistically analyze the proportion of coal content in the material entering the second stage process in real time. When the proportion of incorrectly entered coal in the second stage feed exceeds 3%, the feedback control system automatically lowers the coal blowing judgment threshold of the first-stage sorter. That is, the value is reduced from 0.60 to 0.55, while the triggering time of the pneumatic spray valve is finely adjusted to blow out more boundary particles; when the coal content in the second stage feed is less than 1%, the value is automatically increased. To 0.63, reducing coal residue in the rich kaolin mixture.

[0095] The test results show that after adding closed-loop feedback control, the clean coal recovery rate is stable between 87.0% and 87.5% after 24 hours of continuous operation, the Fe2O3 content of kaolin products is stably controlled below 0.90%, and the overall sorting stability of the system is significantly improved.

[0096] Example 7: The raw ore was sorted into three particle sizes: 10–30 mm, 30–50 mm, and 50–100 mm, using a complete process. Each particle size was independently matched with photoelectric separation equipment parameters and pneumatic spray valve parameters. The specific parameters are as follows:

[0097] Sorting results:

[0098] It is evident that narrow-level grading, combined with the optimization of independent sorting parameters for each particle size, enables materials of each particle size to achieve high sorting accuracy. The overall throughput and sorting efficiency are superior to the non-screening, full-particle-size mixed sorting method.

[0099] Comparative Example 1: The traditional wet heavy medium + flotation purification process adopts the existing mainstream wet purification process for coal gangue. The specific process is as follows: the crushed coal gangue material enters the heavy medium separation system, and the coal is separated and de-coaled using a heavy suspension prepared with magnetite powder to obtain de-coaled gangue. The de-coaled gangue is further refined by flotation to remove impurities. Conventional amine reagents are used as flotation collectors. After washing, desliming, sedimentation, and drying, kaolin concentrate is obtained.

[0100] Experimental results: The clean coal recovery rate was 78.3%, the Al2O3 content of kaolin concentrate was 33.62%, the Fe2O3 content was 1.85%, the TiO2 content was 1.12%, and the calcination whiteness was 76.4%. The water consumption per ton of ore was as high as 2.85t, and the electricity consumption per ton of ore was 32.6kW·h / t. This process has the defects of large water consumption, high tailwater treatment cost, and incomplete removal of iron and titanium impurities in kaolin.

[0101] Comparative Example 2: Solo Dual-Ability The sorting process uses only dual-energy X-ray transmission technology sorts materials based on differences in internal density and equivalent atomic number. Color and texture recognition, NIR-free spectral discrimination, sorting equipment parameters and the first paragraph of Example 3 The detection modules are the same.

[0102] Experimental results: Coal recovery rate was 86.5%; kaolin concentrate Al2O3 content was 35.14%, Fe2O3 content was 2.93%, TiO2 content was 1.54%; calcination whiteness was 68.2%; water consumption per ton of ore was 0.12 t; and electricity consumption per ton of ore was 18.4 kW·h / t. Due to... It cannot identify iron-containing mineral impurities on the surface, nor can it determine whether the material contains kaolinite. The product has a high content of Fe2O3 and TiO2, and the calcined whiteness is far below the industry standard for high-whiteness kaolin, which usually requires a whiteness of ≥80%, indicating insufficient purification effect.

[0103] Comparative Example 3: alone Visual sorting processes rely solely on the color, texture, and shape features of visible light images for sorting, without... Internal physical properties, none Mineral spectral characteristics are used in the sorting equipment, which employs an industrial color camera and conventional image processing algorithms.

[0104] Experimental results: The clean coal recovery rate was 65.2%, the Al2O3 content of kaolin concentrate was 31.26%, the Fe2O3 content was 2.67%, the TiO2 content was 1.41%, the calcination whiteness was 70.5%, the water consumption per ton of ore was 0.10t, and the power consumption per ton of ore was 16.8kW·h / t. Because the surface of coal-series kaolin is covered with organic carbon and appears black, its appearance is highly similar to that of raw coal. A large amount of black kaolinite was misidentified as coal and lost with the coal products, resulting in a low clean coal recovery rate and serious loss of kaolin resources.

[0105] Summary table of test results data:

[0106] Conclusion: Example 3 outperforms the three comparative examples in all four core quality indicators: clean coal recovery rate, kaolin grade, iron and titanium removal effect, and calcination whiteness. In terms of water and electricity consumption, it takes into account the environmental protection and energy-saving advantages of dry separation, and achieves a significant improvement in product quality with a minimal increase in electricity consumption. Overall, this separation process has obvious advantages in both technical indicators and economic benefits.

[0107] Working principle: A tiered sorting architecture is adopted. Preprocessing involves narrow-particle grading and surface purification to expose the true characteristics of minerals. The first stage uses dual-energy XRT to extract internal density and atomic number features, and RGB to extract surface color and texture features. After late-stage fusion and priority judgment logic by the decision layer, XRT high density takes precedence over RGB black, solving the problem of misjudging black gangue by the same color. The fuzzy confidence samples are then transferred to the second stage. The second stage combines RGB image recognition to remove iron-containing yellow gangue and NIR near-infrared spectroscopy to identify the hydroxyl characteristic peak of kaolinite to remove invalid waste rock. The three-class classification output is obtained through a dual-stream AI model, finally obtaining a high-purity, low-iron, and high-whiteness kaolin product.

[0108] The beneficial effects of this invention are: high sorting accuracy: eliminating blind spots in single sensor identification, achieving a clean coal recovery rate of 87.1%, reducing kaolin Fe2O3 to 0.92%, and calcined whiteness to 84.3%; excellent purification quality: simultaneous iron and titanium removal, increasing Al2O3 to 38.75%, eliminating the need for downstream chemical bleaching or high-gradient magnetic separation; green and environmentally friendly: the entire process is dry physical sorting, with water consumption of only 0.11t per ton of ore, saving more than 96% of water compared to wet processes, and eliminating wastewater and reagents; economical and efficient: power consumption is 19.2kW·h / t per ton of ore, saving 41% of energy compared to wet processes, and simultaneously recovering coal and kaolin, achieving full resource utilization of solid waste.

[0109] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A multimodal photoelectric fusion method for purifying and separating kaolin from coal gangue, characterized in that, Includes the following steps: S1. Pre-treatment before selection: The raw coal gangue ore is crushed and classified into multiple independent particle size segments by narrow-level screening process. Each particle size segment is then subjected to surface cleaning and desliming or dry heavy dust removal treatment to remove coal powder and mud adhering to the surface of the material, exposing the true surface color, texture and mineral characteristics of the material. S2, First-stage composite photoelectric separation for coal enrichment: The pre-treated material is fed into the first-stage photoelectric separator, and dual-energy XRT transmission image data and RGB visible light image data of the material are collected simultaneously. The equivalent atomic number and density characteristics of the material are extracted through XRT data, and the surface color and texture characteristics of the material are extracted through RGB data. A multi-modal data fusion algorithm is used for coupled analysis to identify low-density materials with coal texture characteristics as coal products and blow them out for recycling. The high-density mixture of kaolinite and impurity gangue is sent to the next process as a kaolin-rich mixture. S3. Second-stage visual sorting for yellowing and iron removal: The kaolin-rich mixture is fed into the second-stage visual sorting machine, which simultaneously collects RGB image data and NIR near-infrared spectral data of the material. Through the RGB image recognition model and color threshold calibration, the machine identifies and removes iron oxide impurities such as gangue with yellow or brownish-red surfaces. The NIR near-infrared spectroscopy identifies the molecular bond characteristics of the material's minerals, and the machine removes invalid impurity rocks that do not contain kaolinite based on the hydroxyl absorption peak characteristics of kaolinite. The remaining material is the high-purity kaolinite product. The multimodal data fusion algorithm has a priority judgment logic: when XRT detection determines that the material is high density and RGB detection determines that the material is black, the XRT density feature data is given priority and the material is judged as black kaolinite rather than coal; when the XRT and RGB output judgment results are within the preset fuzzy confidence interval, the coal blowing action is not performed and the material is sent to the second process.

2. The multimodal photoelectric fusion method for purifying and separating kaolin from coal gangue as described in claim 1, characterized in that, In step S1, the narrow-level screening and grading particle size is divided into three grades: 10-30mm, 30-50mm, and 50-100mm. Different particle sizes are matched with corresponding photoelectric sorting equipment parameters and pneumatic spray valve parameters, and the particle size is sorted independently.

3. The multimodal photoelectric fusion method for purifying and separating kaolin from coal gangue as described in claim 1, characterized in that, In step S2, the multimodal data fusion algorithm is a late-stage fusion algorithm for the decision layer; Among them, the features extracted from dual-energy XRT images include the gray-level statistical features of particle ROI, the transmission texture features inside the XRT gray-level co-occurrence matrix, the mean, variance, quantile, statistical distribution features of the proportion of high atomic number pixels of particle pixels, and particle morphology features. Features extracted from RGB visible light images include RGB three-channel statistical features, HSV color space features, hue and color histograms, RGB grayscale co-occurrence matrix surface texture features, and particle morphology features; The XRT and RGB features are input into independent classifiers, which output the confidence scores of coal, kaolinite / extra gangue in the corresponding modes.

4. The multimodal photoelectric fusion method for purifying and separating kaolin from coal gangue as described in claim 3, characterized in that, Priority determination logic includes: setting a high confidence threshold. Low confidence threshold When the confidence levels of both XRT and RGB outputs are within a certain range ~ When the confidence interval is fuzzy, if the two modal discriminations are in the same direction, the basic weight weighted fusion is used; if the two modal discriminations are in opposite directions and conflict, the XRT modal weights are increased and the fusion confidence is recalculated. Particles whose fusion results are still within the fuzzy confidence interval will not be blown out and will be sent to the second visual sorting process for further identification. The base weights are XRT weight 0.65 and RGB weight 0.

35. In case of conflict, the dynamic weights are adjusted to XRT weight 0.8 and RGB weight 0.

2.

5. The multimodal photoelectric fusion method for purifying and separating kaolin from coal gangue as described in claim 1, characterized in that, In step S3, the RGB image recognition model and the NIR spectrum recognition model adopt a dual-stream multimodal fusion network, including an RGB image branch, an NIR spectrum branch, a feature fusion layer, and an output layer; The RGB image branch is a lightweight CNN convolutional network that extracts color and surface texture features and outputs an image feature vector. The NIR spectral branch is a one-dimensional convolutional CNN or MLP, which extracts spectral feature vectors from the preprocessed spectral curves, focusing on learning the spectral fingerprint of the kaolinite hydroxyl absorption peak. The feature fusion layer concatenates the image feature vector and the spectral feature vector before feeding them into the fully connected layer. The output layer uses... Output the three-class confidence score ,in, For qualified kaolinite confidence level, Confidence level for iron-containing gangue, The confidence level for invalid waste rock.

6. The multimodal photoelectric fusion method for purifying and separating kaolin from coal gangue as described in claim 5, characterized in that, The NIR spectroscopy detection range is 1300–2500 nm, with a focus on covering the characteristic absorption peaks of kaolinite hydroxyl groups around 1400 nm and 2200 nm.

7. The multimodal photoelectric fusion method for purifying and separating kaolin from coal gangue as described in claim 5, characterized in that, The training dataset of the dual-stream multimodal fusion network includes four types of particles: qualified kaolinite, iron-containing gangue, invalid waste rock, and interference samples. It also includes coal gangue material samples from at least two different origins. The training set, validation set, and test set are divided in a 7:1:2 ratio, and data augmentation is performed on the RGB images and NIR spectra respectively.

8. A multimodal photoelectric fusion coal gangue purification and kaolin sorting system, characterized in that, The sorting method according to any one of claims 1-7 includes a preprocessing unit, a first composite photoelectric sorting unit, a second visual sorting unit, and a material collection unit connected in sequence. The pretreatment unit includes crushing equipment, grading and screening equipment, and dust removal and desliming equipment, which are used to complete coal gangue crushing, narrow particle size classification, and material surface purification treatment. The first composite photoelectric sorting unit includes a dual-energy XRT detection module, an RGB vision detection module, a multimodal data processing module, and a first pneumatic sorting execution module, which are used to fuse density and visual features to achieve precise separation of coal and kaolin-rich mixtures. The second visual selection unit includes an RGB color recognition module, a NIR near-infrared spectroscopy detection module, an AI intelligent recognition module, and a second pneumatic sorting execution module, used to remove iron-containing yellow gangue and non-kaolinite impurity rocks; The material collection unit includes a coal product collection bin, a high-purity kaolin product collection bin, and an impurity gangue discharge bin.

9. A multimodal photoelectric fusion coal gangue purification and kaolin sorting system as described in claim 8, characterized in that, The multimodal data processing module has a built-in feature priority determination program. For black gangue with the same color interference, it locks the XRT high-density feature as the core determination criterion to avoid misjudgment and loss of black kaolin.

10. A multimodal photoelectric fusion coal gangue purification and kaolin sorting system as described in claim 8, characterized in that, The second visual selection unit is equipped with an AI image recognition model and a spectral feature comparison database, which can adaptively identify iron-containing impurities and invalid mineral impurities in coal gangue from different origins.

Images