Grading gangue selection method for coal gangue

Abstract

The invention relates to the technical field of grading gangue selection of coal gangue, in particular to a grading gangue selection method of coal gangue, which comprises the following steps: acquiring mineral texture data through a multispectral imaging device, analyzing the spatial distribution of aluminum and silicon elements in combination with an X-ray diffractometer, and generating a three-dimensional component map through spatial registration and fusion with infrared heat value data; impurities are dynamically removed through sulfur element spectrum detection, and the air pressure and the material flow of winnowing equipment are adjusted by matching roasting process parameters; components of the sorted waste are analyzed through an X-ray fluorescence spectrometer, particle size data are obtained through a laser particle analyzer, a genetic algorithm is input to iteratively optimize detection channel combination and sampling density, a resource distribution instruction is generated, and equipment parameter correction is fed back. According to the method, aluminum-silicon ratio critical value accurate identification, sorting parameter dynamic adaptation and waste recycling path optimization are realized, and the problem of overburning caused by insufficient quantitative analysis of element distribution in the prior art is solved.

Description

Technical Field

[0001] The invention relates to the technical field of coal gangue classification and separation, and in particular to a coal gangue classification and separation method. Background Art

[0002] Gangue is a solid waste generated during the coal mining and washing process. The accuracy of its composition detection directly affects the effectiveness of gangue classification and selection. Existing sorting processes rely on surface characteristics to screen kaolin, lacking quantitative detection methods for the distribution of aluminum and silicon elements and calorific value characteristics within the gangue. This leads to significant deviations between the raw material purity test results and the kiln roasting process requirements. Especially for kaolin with a complex layered structure, existing detection methods are unable to accurately identify the critical aluminum-silicon ratio. As a result, the sorted gangue raw materials may be overburned during the roasting process due to inaccurate calorific value detection, which restricts the quality control of gangue in alumina production. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the present invention provides a method for grading and selecting coal gangue, which is used to solve the problem that the existing detection method cannot quantitatively analyze the distribution characteristics of aluminum and silicon elements inside the coal gangue, resulting in inaccurate identification of the critical value of the aluminum-silicon ratio, and causing overburning due to insufficient compatibility between the calorific value of the sorted raw materials and the roasting process.

[0004] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows: The present invention provides a method for classifying and separating coal gangue, comprising: Step S1, obtaining multispectral image data, X-ray diffraction data, and calorific value distribution data representing the surface mineral texture of coal gangue, spatially registering and fusing the multispectral image data with the X-ray diffraction data to generate a three-dimensional composition map, and associating the three-dimensional composition map with the calorific value distribution data and storing them in a dynamic database; Step S2: Based on the three-dimensional component map in the dynamic database, extract the characteristics of the aluminum-silicon concentration gradient change, combine the layered structure boundary recognition results in the multispectral image data, train the aluminum-silicon ratio classification model, output a classification label representing the critical value of the aluminum-silicon ratio, and represent the critical value of the aluminum-silicon ratio as aluminum-silicon ratio ≥ 0.9. The classification label is associated with the calorific value distribution data stored in the dynamic database to generate a sorting instruction including a material sorting priority and a calorific value adaptation threshold, with a target calorific value ≤ 1000 kcal / kg, and a grading threshold of the calorific value adaptation threshold of less than 400 kcal / kg, 400-600 kcal / kg, and 400-800 kcal / kg, and the instruction is transmitted to the sorting equipment control terminal; Step S3: Controlling the sorting equipment to screen materials with an aluminum-silicon ratio higher than a threshold value of 0.9 according to the sorting instruction, and dynamically removing sulfur-containing impurities based on the detection results output by the sulfur element spectrum detection module; matching the infrared calorific value distribution data stored in the dynamic database with the preset roasting kiln process parameter threshold, with a target calorific value of ≤1000kcal / kg, and graded adaptation thresholds of the roasting kiln process parameter threshold of less than 400kcal / kg, 400-600kcal / kg, and 400-800kcal / kg; calculating the sorting ratio control parameters and sending them to the air separation equipment; associating the abnormal calorific value data detected during the sorting process with the correction instruction of the sorting parameters, and updating them to the process parameter threshold in the dynamic database; Step S4, performing secondary testing on the sorted waste, said secondary testing comprising: analyzing the iron, silicon, and carbonate components by an X-ray fluorescence spectrometer, obtaining particle size distribution data by a laser particle size analyzer, inputting the component analysis data and particle size data into a genetic algorithm, iteratively optimizing the detection channel combination of the X-ray fluorescence spectrometer and the sampling density parameters of the laser particle size analyzer, generating instructions for waste resource allocation, and feeding the optimized detection channel combination and sampling density parameters back to the X-ray fluorescence spectrometer and the laser particle size analyzer.

[0005] Furthermore, in the method for classifying and separating coal gangue of the present invention, step S1 comprises: Multispectral image data of coal gangue surface mineral texture is collected through a multispectral imaging device, X-ray diffraction data is obtained and the spatial distribution of aluminum and silicon elements is analyzed through an X-ray diffractometer, and calorific value distribution data is obtained through an infrared calorific value detection module; spatially registering the multispectral image data with the X-ray diffraction data, and fusing them to generate a three-dimensional composition map including aluminum and silicon element concentration gradients; The calorific value distribution data is associated with the time series of the three-dimensional composition map to construct input features of the dynamic database.

[0006] Furthermore, in the coal gangue classification and selection method of the present invention, step S2 includes: extracting the concentration gradient variation characteristics of aluminum and silicon elements based on X-ray diffraction data obtained by an X-ray diffractometer, combining the layered structure boundary recognition results in the multispectral image data, training a deep learning model to output an aluminum-silicon ratio classification label; The classification label is associated with the calorific value distribution data obtained by the infrared calorific value detection module to generate a sorting equipment execution instruction, which includes the material sorting priority and the calorific value adaptation threshold.

[0007] Furthermore, in the method for classifying and separating coal gangue of the present invention, step S3 comprises: According to the sorting instructions, materials with aluminum-silicon ratios higher than the set threshold are screened, and the removal ratio of sulfur-containing impurities is dynamically adjusted based on the sulfur element spectrum detection results; Adjust the air pressure and material flow of the air separation equipment based on the matching calculation results of the infrared calorific value distribution data and the preset roasting kiln process parameter thresholds stored in the dynamic database; The abnormal calorific value data and the correction instructions of the sorting parameters are synchronously updated to the dynamic database, triggering the iterative update of the parameters of the classification model.

[0008] Furthermore, in the method for classifying and separating coal gangue of the present invention, step S4 comprises: The distribution characteristics of iron, silicon and carbonate components were analyzed by X-ray fluorescence spectrometer, and the particle size distribution data of the waste was obtained by laser particle size analyzer; The component distribution characteristics and particle size data are input into the genetic algorithm to generate detection parameter optimization instructions and resource path allocation plans.

[0009] Furthermore, in the coal gangue classification and separation method of the present invention, the detection parameter optimization instruction includes: Iteratively optimize the detection channel combination of the X-ray fluorescence spectrometer and adjust the channel selection weight based on the goal of minimizing the detection error of iron and silicon components; Optimize the sampling density parameters of the laser particle size analyzer and dynamically adjust the sampling interval based on the confidence threshold of the particle size distribution data.

[0010] Furthermore, in the coal gangue classification and separation method of the present invention, the resource path allocation scheme includes: The waste composition data analyzed by X-ray fluorescence spectrometer and the downstream production line raw material demand database are input into the genetic algorithm to calculate the optimal allocation path; During the detection parameter optimization process, an initial population is generated based on the historical error data of the detection channel combination of the X-ray fluorescence spectrometer; In the resource path matching process, the fitness function is set based on the raw material demand fluctuation characteristics of the downstream production line; The optimized detection parameters and the coordinated results of the allocation path are used as constraints to iteratively update the crossover and mutation probability of the genetic algorithm. The distribution paths are dynamically sorted based on the matching scores of iron content, silicon purity and carbonate concentration with the downstream production lines, generating directional delivery instructions for building materials, cement or chemical fillers.

[0011] Furthermore, the method for classifying and separating coal gangue of the present invention further comprises: Feedback the detection parameters optimized by the genetic algorithm to the X-ray fluorescence spectrometer and laser particle size analyzer to adjust the sampling frequency and resolution of multimodal data acquisition; The execution results of the resource allocation path are input into the dynamic database to drive the update of the aluminum-silicon ratio classification threshold of the classification model.

[0012] Furthermore, the method for classifying and separating coal gangue of the present invention further comprises: High-silicon waste is classified into coarse and fine particles based on the particle size data optimized by the laser particle size analyzer, and allocated to the building materials or ceramic production line according to the threshold value; The iron-containing waste is combined with the magnetic separation results and the matching score of the genetic algorithm, and is preferentially transported to the high-portland cement production line; Carbonate materials are matched with building material additive standards through multi-spectral characteristics and distributed to the concrete additive production line.

[0013] Furthermore, the method for classifying and separating coal gangue of the present invention further comprises: The gangue is classified into three grades according to particle size: 0-10mm, 10-50mm and 50-300mm by a preset vibrating screening device; For materials with a diameter of 10-50mm and 50-300mm, multispectral image data of the gangue surface mineral texture is obtained, and a gangue particle size sorting model is established based on the aluminum-silicon ratio; Identify the mineral composition of coal gangue according to the coal gangue particle size sorting model and separate the gangue with aluminum-silicon ratio ≥ 0.9; The sorted gangue is air-jet-pulverized to a particle size of ≤200μm; Conduct secondary testing and resource processing on the sorted waste; Among them, materials with a size of 0-10 mm are separated by gravity to separate useful materials with an aluminum-silicon ratio ≥ 0.9 and coal powder by-products.

[0014] Beneficial effects of the present invention: The present invention generates a three-dimensional composition map by fusing multi-spectral imaging and X-ray diffraction data, combines the extraction of aluminum-silicon concentration gradient features and layered boundary identification, and uses a deep learning model to accurately determine the critical value of the aluminum-silicon ratio and generate sorting instructions, thereby solving the defect of the existing technology in insufficient quantitative analysis of the internal element distribution of gangue; dynamically adjusts the impurity removal ratio through sulfur element spectral detection, matches the calorific value distribution with the roasting process parameter threshold to optimize the sorting air pressure and material flow, and improves the sorting accuracy and process adaptability; adopts a genetic algorithm to iteratively optimize the detection channel of the X-ray fluorescence spectrometer and the sampling density of the laser particle size analyzer, combines the waste composition and particle size data to generate resource allocation instructions, and forms a closed-loop correction path for the detection parameters; the incremental learning mechanism dynamically updates the classification model threshold based on the calorific value abnormality data, realizes the adaptive adjustment of the sorting strategy, and systematically improves the sorting efficiency, resource utilization rate and quality control level of coal gangue. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on the drawings without paying any creative labor.

[0016] Figure 1 The present invention provides a flow chart of a method for classifying and selecting coal gangue. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. In order to better understand the purpose of the present invention, the present invention is further described in detail below.

[0018] See also Figure 1 The present invention provides a method for classifying and selecting coal gangue, comprising: Step S1, obtaining multispectral image data, X-ray diffraction data, and calorific value distribution data representing the surface mineral texture of coal gangue, spatially registering and fusing the multispectral image data with the X-ray diffraction data to generate a three-dimensional composition map, and associating the three-dimensional composition map with the calorific value distribution data and storing them in a dynamic database; In step S1, a multispectral imaging device collects multispectral image data of the mineral texture on the gangue surface, covering the visible to near-infrared band (400-2500nm). Differences in reflectivity across different bands are used to characterize the mineral surface texture. The multispectral imaging device uses a high-resolution CCD sensor to synchronously collect multi-band reflectivity data, generating a mineral texture map with spatial resolution. Simultaneously, an X-ray diffractometer, based on the Bragg equation, scans the gangue, analyzing the lattice structure and spatial distribution of aluminum and silicon elements and generating a thermogram of the elemental concentration gradient distribution. The infrared calorific value detection module uses a non-contact thermal radiation sensor to collect infrared radiation intensity from different areas of the gangue surface, generating calorific value distribution data to characterize the material's energy release characteristics.

[0019] Multispectral image data and X-ray diffraction data are aligned using spatial registration technology. A feature point matching algorithm is used to extract corner features of mineral textures in the multispectral image and match them with the spatial coordinates of the X-ray diffraction data, eliminating offset errors caused by equipment installation deviations or sampling angles. The registered data are then fused using a pixel-level weighted superposition algorithm, combining the texture information of the multispectral image with the elemental concentration gradients from the X-ray diffraction data to generate a three-dimensional composition map. Each pixel in the map is associated with multispectral reflectance, aluminum and silicon concentrations, and three-dimensional spatial coordinates, forming a multidimensional data matrix.

[0020] The three-dimensional composition map and calorific value distribution data are dynamically associated through timestamps. The dynamic database adopts a distributed storage architecture, hierarchically storing the grid data of the three-dimensional composition map and the hot zone mapping relationship of the calorific value distribution, and realizing fast retrieval through an indexing mechanism. The time series data in the database records the map and calorific value information of each sampling node to construct the input feature set. The three-dimensional map after spatial registration fusion provides a spatial consistency basis for the subsequent extraction of aluminum-silicon concentration gradient features, and the calorific value associated storage provides data support for the dynamic matching of the calorific value adaptation threshold in the generation of sorting instructions. The collaborative collection and structured storage of multimodal data lays a multi-dimensional analysis foundation for the training of deep learning models and the optimization of sorting process parameters.

[0021] Step S2: Based on the three-dimensional component map in the dynamic database, extract the characteristics of the aluminum-silicon concentration gradient change, combine the layered structure boundary recognition results in the multispectral image data, train the aluminum-silicon ratio classification model, output a classification label representing the critical value of the aluminum-silicon ratio, and represent the critical value of the aluminum-silicon ratio as aluminum-silicon ratio ≥ 0.9. The classification label is associated with the calorific value distribution data stored in the dynamic database to generate a sorting instruction including a material sorting priority and a calorific value adaptation threshold, with a target calorific value ≤ 1000 kcal / kg, and a grading threshold of the calorific value adaptation threshold of less than 400 kcal / kg, 400-600 kcal / kg, and 400-800 kcal / kg, and the instruction is transmitted to the sorting equipment control terminal; In step S2, based on the three-dimensional compositional maps in the dynamic database, grayscale gradient analysis is used to extract the concentration gradient characteristics of aluminum and silicon, quantifying the concentration differences between elements in different regions. The multispectral image data uses the Canny edge detection algorithm to identify layered structure boundaries, combined with morphological filtering to remove noise interference and generate a high-precision boundary contour map. The concentration gradient characteristics and boundary identification results are spatially superimposed to form a joint input dataset. This dataset, containing multispectral reflectance, element concentration gradients, and structural boundary information, serves as the training input for the deep learning model.

[0022] Based on the three-dimensional compositional maps in the dynamic database, grayscale gradient analysis was used to extract the concentration gradient characteristics of aluminum and silicon. Combined with the layered structure boundaries identified in the multispectral image data using the Canny edge detection algorithm (combined with morphological filtering to remove noise interference), a joint input dataset (including multispectral reflectance, elemental concentration gradients, and structural boundary information) was constructed. A deep learning model (using the ResNet-50 architecture, with multi-channel data input layers, local feature extraction through convolutional layers, and fully connected layers learning the relationship between the aluminum-silicon ratio, layered structure, and calorific value distribution) was trained to output classification labels for the aluminum-silicon ratio.

[0023] Classification labels are divided into multiple discrete levels, corresponding to different aluminum-silicon ratio threshold ranges, such as Level A (aluminum-silicon ratio > 0.9) and Level B (aluminum-silicon ratio 0.7 ≤ ≤ 0.9). Classification labels are matched with infrared calorific value distribution data stored in a dynamic database through an association mapping module. This module constructs a calorific value adaptation threshold matrix based on the energy release characteristics of the calorific value distribution data (target calorific value ≤ 1000 kcal / kg, classification thresholds: low calorific value zone < 400 kcal / kg, medium-low calorific value zone 400-600 kcal / kg, medium calorific value zone 400-800 kcal / kg).

[0024] For example, high calorific value areas (>800kcal / kg but ≤1000kcal / kg) are associated with aluminum-silicon ratio level A (aluminum-silicon ratio >0.9) to generate sorting priority weights (the priority weights are dynamically adjusted based on the positive correlation between the aluminum-silicon ratio and calorific value; for example, for every 0.1 unit increase in the aluminum-silicon ratio, the priority weight increases by 5%). Sorting instructions include material sorting priorities (e.g., priority 1 for high aluminum-silicon ratio materials with an aluminum-silicon ratio >0.9) and calorific value adaptation threshold ranges (e.g., for calorific values ​​≤1000kcal / kg, the 400-800kcal / kg range is prioritized). These instructions are transmitted to the sorting equipment control terminal via the Modbus protocol, driving the photoelectric sorter to perform the screening operation.

[0025] Step S3: Controlling the sorting equipment to screen materials with an aluminum-silicon ratio higher than a threshold value of 0.9 according to the sorting instruction, and dynamically removing sulfur-containing impurities based on the detection results output by the sulfur element spectrum detection module; matching the infrared calorific value distribution data stored in the dynamic database with the preset roasting kiln process parameter threshold, with a target calorific value of ≤1000kcal / kg, and graded adaptation thresholds of the roasting kiln process parameter threshold of less than 400kcal / kg, 400-600kcal / kg, and 400-800kcal / kg; calculating the sorting ratio control parameters and sending them to the air separation equipment; associating the abnormal calorific value data detected during the sorting process with the correction instruction of the sorting parameters, and updating them to the process parameter threshold in the dynamic database; In step S3, based on the received sorting instructions, the sorting equipment uses the grayscale recognition module of the photoelectric sorter to screen materials with aluminum-silicon ratios above the threshold of 0.9. The threshold set in the sorting instructions is dynamically adjusted according to the roasting process requirements. For example, a higher aluminum-silicon ratio threshold is required for a high-temperature kiln. The synchronously operating sulfur spectral detection module uses laser-induced breakdown spectroscopy (LIBS) technology to collect characteristic spectral signals of sulfur on the material surface (e.g., wavelength range of 180-280nm) in real time and calculate the sulfur concentration using a threshold determination algorithm. When the sulfur concentration is detected to exceed a preset safety threshold (e.g., >0.5%), the pneumatic nozzle's rejection ratio is dynamically adjusted according to the deviation value, forming a closed-loop control logic for the removal of sulfur impurities.

[0026] A vector similarity algorithm is used to calculate the matching degree between infrared calorific value distribution data and the roasting kiln process parameter thresholds preset in the dynamic database (e.g., the target calorific value range of 4800-5200 kcal / kg). This matching result reflects the degree of compatibility between the current material calorific value and process requirements, driving the actuators of the air separation equipment to adjust the sorting air pressure and material flow. Air pressure regulation utilizes a proportional-integral-derivative (PID) control algorithm, dynamically adjusting the fan speed based on the calorific value matching deviation. For example, if the matching degree falls below 90%, the air pressure is increased by 5%. Material flow is controlled by a variable frequency feeder. Taking into account the regional differences in calorific value distribution, the flow rate in low calorific value areas is reduced to 80% of the baseline value, achieving refined control of the sorting ratio.

[0027] Abnormal calorific value data detected during the sorting process (such as local calorific values ​​below 4500kcal / kg or above 5500kcal / kg) is identified by the abnormality marking module and associated with the sorting timestamp and spatial coordinate information. The correction instruction generation module generates adjustment parameters for air pressure, flow rate, and rejection ratio based on the distribution characteristics of the abnormal data, and synchronously updates them to the process parameter threshold field in the dynamic database. Database updates trigger the parameter iteration mechanism of the classification model, using an incremental learning algorithm to fine-tune the weights of the fully connected layer and optimize the decision boundary of the aluminum-silicon ratio classification label. The dynamic database adopts a version control mechanism to retain historical parameter versions and correction records, support process parameter backtracking and expand the model training data set, and form a collaborative optimization closed loop between the sorting process and the classification model.

[0028] In the above process, the simultaneous implementation of aluminum-silicon ratio threshold screening and sulfur removal improves the purity control accuracy of the sorted materials. Dynamic adjustment of air separation parameters driven by calorific value matching enables real-time adaptation of the sorting process to roasting requirements. The model iteration mechanism driven by abnormal data enhances the classification model's adaptability to complex working conditions. Through a closed-loop control system, the problem of insufficient sorting accuracy and process adaptability is systematically addressed.

[0029] Step S4, performing secondary testing on the sorted waste, said secondary testing comprising: analyzing the iron, silicon, and carbonate components by an X-ray fluorescence spectrometer, obtaining particle size distribution data by a laser particle size analyzer, inputting the component analysis data and the particle size data into a genetic algorithm, iteratively optimizing the detection channel combination of the X-ray fluorescence spectrometer and the sampling density parameters of the laser particle size analyzer, generating instructions for waste resource allocation, and feeding back the optimized detection channel combination and sampling density parameters to the X-ray fluorescence spectrometer and the laser particle size analyzer, thereby forming a closed-loop correction path for the detection parameters.

[0030] In step S4, the sorted waste is analyzed for composition using an X-ray fluorescence spectrometer. The detection channels are prioritized for the Kα characteristic peak (6.40 keV) of iron and the Kβ characteristic peak (1.74 keV) of silicon. An energy spectrum analysis algorithm is used to generate a heat map of the distribution of iron, silicon, and carbonate, quantifying the concentration of each component. A laser particle size analyzer, based on the principle of dynamic light scattering, collects multi-angle scattered light intensity data from the waste particles. Using the Malvern model, the characteristic particle size parameters (D10, D50, and D90) are calculated to generate a particle size distribution curve, characterizing the particle size characteristics of the waste.

[0031] Composition analysis data and particle size distribution data were fed into a genetic algorithm. The initial population was generated from historical channel combination error data from the X-ray fluorescence spectrometer. The chromosome encoding contained the channel weights and the sampling density parameters of the laser particle size analyzer. The fitness function, which minimized the detection errors for iron and silicon components and maximized the confidence in the particle size distribution, drove the iterative optimization of the parameter combination through crossover and mutation operations. The optimized channel combination prioritized the iron Kα and silicon Kβ channels, reducing background noise. The sampling density of the laser particle size analyzer was increased to 50 points / mm in the D50±10μm range, improving resolution in the critical particle size range.

[0032] Optimization instructions are fed back to the X-ray fluorescence spectrometer and laser particle size analyzer via the industrial bus. The spectrometer's multi-channel detector integration time is adjusted to 200ms, and the particle size analyzer's laser pulse frequency is increased to 10kHz, forming a closed-loop correction path for the detection parameters. Waste resource allocation instructions are based on composition and particle size data, using a genetic algorithm to match the downstream production line demand database. For example, waste with an iron content greater than 15% is allocated to the cement production line, waste with a silicon purity greater than 80% is transported to the building materials production line, and materials with a carbonate concentration greater than 30% are directed to the concrete additive production line. The allocation results are transmitted to the conveying system via the Modbus protocol, driving the belt conveyor and pneumatic sorting valve to perform the directional conveying action.

[0033] During execution, optimized inspection data and resource allocation results are synchronously updated to a dynamic database, providing incremental training data for subsequent model threshold iterations in the sorting process. This closed-loop correction mechanism continuously optimizes inspection accuracy and resource matching efficiency, improving the reliability of waste composition analysis and resource allocation decision-making, thereby achieving refined and efficient resource utilization of coal gangue.

[0034] The coal gangue classification and separation method of the present invention comprises the following technical steps: In step S1, a multispectral imaging device collects multispectral image data of the mineral texture on the gangue surface, covering wavelengths from visible light to near-infrared. The spatial distribution of aluminum and silicon elements is analyzed using an X-ray diffractometer, and calorific value distribution data is obtained using an infrared calorific value detection module. The multispectral image data and X-ray diffraction data are aligned using spatial registration technology and fused to generate a three-dimensional compositional map containing aluminum and silicon concentration gradients. The three-dimensional compositional map and calorific value distribution data are linked in time series to form input features for a dynamic database, providing a multimodal data foundation for subsequent analysis.

[0035] Step S2 extracts the concentration gradient characteristics of aluminum and silicon elements based on the three-dimensional composition map. This is combined with the layered structure boundary information identified by edge detection algorithms in the multispectral image to construct a training dataset. A deep learning model is used to train the aluminum-silicon ratio classification labels, which output classification labels representing critical values. The classification labels are mapped to the calorific value distribution data in the dynamic database to generate sorting instructions. These instructions, which include material priority and calorific value adaptation thresholds, are transmitted to the sorting equipment control terminal, enabling dynamic matching of sorting strategies with roasting processes.

[0036] Step S3 controls the sorting equipment based on the sorting instructions to screen materials with aluminum-silicon ratios exceeding a threshold. Simultaneously, the sulfur spectral detection module monitors the sulfur impurity content in real time and dynamically adjusts the rejection ratio. Infrared calorific value distribution data is matched with preset roasting kiln process parameter thresholds to generate sorting ratio control parameters, which are used to adjust the air pressure and material flow rate of the air separation equipment. Abnormal calorific value data detected during the sorting process is linked to correction instructions, and the process parameter thresholds in the dynamic database are updated in real time, triggering iterative optimization of the classification model and forming a dynamic closed-loop control of the sorting parameters.

[0037] Step S4 performs a secondary inspection on the sorted waste, analyzes the distribution characteristics of iron, silicon and carbonate components through X-ray fluorescence spectrometer, and collects the particle size distribution data of the waste in combination with the laser particle size analyzer. The component analysis data and particle size data are input into the genetic algorithm, with the goal of minimizing the detection error, and the detection channel combination weight of the X-ray fluorescence spectrometer is optimized, and the sampling density parameter of the laser particle size analyzer is adjusted based on the particle size distribution confidence threshold. The optimized detection parameters are fed back to the detection equipment to form a closed-loop correction path. The waste resource allocation instruction is generated based on the component data and the downstream production line demand database, dynamically sorted by matching scores, and directed to the building materials, cement or chemical filler production lines to achieve efficient resource utilization of waste.

[0038] In the above steps, the data fusion of step S1 provides the basis for subsequent feature extraction, the classification model output of step S2 drives the generation of sorting instructions, the real-time control and parameter update of step S3 ensure the dynamic adaptation of the sorting process and roasting requirements, and the detection parameter optimization and resource allocation of step S4 form a closed-loop feedback. The overall process improves sorting accuracy and resource utilization through multimodal data collaboration and algorithm optimization.

[0039] Specifically, the method for classifying and separating coal gangue of the present invention, step S1 comprises: Multispectral image data of coal gangue surface mineral texture is collected through a multispectral imaging device, X-ray diffraction data is obtained and the spatial distribution of aluminum and silicon elements is analyzed through an X-ray diffractometer, and calorific value distribution data is obtained through an infrared calorific value detection module; spatially registering the multispectral image data with the X-ray diffraction data, and fusing them to generate a three-dimensional composition map including aluminum and silicon element concentration gradients; The calorific value distribution data is associated with the time series of the three-dimensional composition map to construct input features of the dynamic database.

[0040] In step S1 of the present invention, multispectral image data of the mineral texture on the surface of coal gangue is collected by a multispectral imaging device, covering the visible light to near-infrared band, and the reflection characteristics of different bands are used to characterize the texture characteristics of the mineral surface. At the same time, the diffraction spectrum data of the coal gangue is collected by an X-ray diffractometer, and the lattice structure and spatial distribution information of aluminum and silicon elements are analyzed based on the Bragg equation to generate an element distribution thermogram. The infrared calorific value detection module obtains the energy release characteristics of different areas of the coal gangue through non-contact thermal radiation measurement to form calorific value distribution data. The multispectral image data and the X-ray diffraction data are aligned by spatial registration technology. Specifically, the spatial coordinate systems of the two are calibrated using a feature point matching algorithm to eliminate the offset error caused by the equipment position or sampling angle, and achieve accurate overlap of the data in three-dimensional space.

[0041] During the fusion process, a pixel-level weighted overlay algorithm combines the texture features of the multispectral image with the elemental distribution information from X-ray diffraction to generate a three-dimensional compositional map containing aluminum and silicon concentration gradients. Each pixel in the map is associated with multispectral reflectance, elemental concentration, and spatial coordinate information, forming a multidimensional data matrix. Calorific value distribution data is linked to the three-dimensional compositional map via timestamps. A dynamic database records the map data and corresponding calorific value at each time point, constructing an input feature set that includes composition, spatial distribution, and calorific value characteristics.

[0042] The dynamic database utilizes a distributed storage architecture, hierarchically storing gridded data from three-dimensional composition maps and mappings of thermal zones within calorific value distributions. This data fusion, after spatial registration, provides a spatially consistent foundation for subsequent extraction of aluminum-silicon concentration gradient features. Calorific value-associated storage supports the dynamic matching of calorific value adaptation thresholds during sorting instruction generation. The collaborative collection and structured storage of multimodal data lays the foundation for multi-dimensional analysis, enabling deep learning model training and optimization of sorting process parameters.

[0043] Specifically, the method for grading and selecting coal gangue of the present invention, step S2 comprises: extracting the concentration gradient variation characteristics of aluminum and silicon elements based on X-ray diffraction data obtained by an X-ray diffractometer, combining the layered structure boundary recognition results in the multispectral image data, training a deep learning model to output an aluminum-silicon ratio classification label; The classification label is associated with the calorific value distribution data obtained by the infrared calorific value detection module to generate a sorting equipment execution instruction, which includes the material sorting priority and the calorific value adaptation threshold.

[0044] In step S2, based on the diffraction data acquired by the X-ray diffractometer, grayscale gradient analysis is used to extract the concentration gradient characteristics of aluminum and silicon, quantifying the relative abundance differences between the elements in different regions. A multiscale edge detection algorithm is used to identify the boundary contours of the layered structure using the multispectral image data. Morphological filtering is then used to eliminate noise interference and generate boundary recognition results. The concentration gradient characteristics and boundary recognition results are spatially superimposed to form a joint input dataset for training the deep learning model.

[0045] The deep learning model utilizes a convolutional neural network architecture. Its input layer receives multi-channel data, including concentration gradient maps, boundary identification maps, and multispectral reflectance data. The model extracts local features through convolutional layers, reduces dimensionality through pooling layers, and learns the nonlinear relationship between the aluminum-silicon ratio and boundary structure through fully connected layers. The model ultimately outputs a classification label representing the critical aluminum-silicon ratio, which is defined as ≥0.9. The classification label is mapped to a probability distribution using the Softmax function, and the cross-entropy loss function is used to optimize model parameters and improve classification accuracy.

[0046] Classification labels are matched with infrared calorific value distribution data stored in a dynamic database through an association mapping module. Based on the energy release characteristics of the calorific value distribution data, the mapping module constructs a calorific value adaptation threshold matrix. Combined with the aluminum-silicon ratio threshold of the classification label, this module generates execution instructions for the sorting equipment. These instructions contain material sorting priority weights and calorific value adaptation threshold ranges. Priority weights are dynamically adjusted based on the positive correlation between the aluminum-silicon ratio and calorific value. The sorting instructions are transmitted to the sorting equipment control terminal via an industrial communication protocol, directing the equipment to execute the sorting action.

[0047] In the above process, the combined analysis of concentration gradient characteristics and lamellar boundary identification results enhances the physical relevance of aluminum-silicon ratio classification. Deep learning models enhance the robustness of critical value determination through multimodal data fusion. A dynamic mapping mechanism for calorific value adaptation thresholds enables sorting instructions to adapt to the differentiated calorific value characteristics of raw materials required by different roasting processes, achieving precise adaptation of sorting strategies to downstream processes.

[0048] Specifically, in the method for classifying and separating coal gangue of the present invention, step S3 comprises: According to the sorting instructions, materials with aluminum-silicon ratios higher than the set threshold are screened, and the removal ratio of sulfur-containing impurities is dynamically adjusted based on the sulfur element spectrum detection results; Adjust the air pressure and material flow of the air separation equipment based on the matching calculation results of the infrared calorific value distribution data and the preset roasting kiln process parameter thresholds stored in the dynamic database; The abnormal calorific value data and the correction instructions of the sorting parameters are synchronously updated to the dynamic database, triggering the iterative update of the parameters of the classification model.

[0049] In step S3, the sorting equipment uses a photoelectric sorter to screen the gangue according to the aluminum-silicon ratio threshold set in the sorting instructions. A grayscale recognition module identifies materials with aluminum-silicon ratios above the threshold. A synchronous sulfur spectral detection module, based on laser-induced breakdown spectroscopy (LIBS), collects characteristic spectral signals from the material surface in real time. Using a threshold determination algorithm, it calculates the concentration of sulfur impurities and dynamically adjusts the rejection ratio of the pneumatic nozzles in the sorting equipment. This rejection ratio is adjusted based on feedback from the deviation between the sulfur concentration and a preset safety threshold, forming a dynamic closed-loop control system for sulfur impurity removal.

[0050] A vector similarity algorithm is used to calculate the matching degree between infrared calorific value distribution data and the roasting kiln process parameter thresholds preset in the dynamic database. This matching result reflects the degree of compatibility between the calorific value characteristics of the currently sorted material and the target process, driving the actuators of the air separation equipment to adjust the sorting air pressure and material flow. Air pressure regulation utilizes a proportional-integral-derivative (PID) control algorithm, dynamically adjusting the fan speed based on the calorific value matching deviation. Material flow is controlled by a variable frequency feeder, and combined with the regional differences in calorific value distribution data, precise control of the sorting ratio is achieved.

[0051] Abnormal calorific value data detected during the sorting process is identified by the abnormal marking module and associated with the corresponding sorting timestamp and spatial coordinate information. The correction instruction generation module generates adjustment parameters for wind pressure, flow rate and rejection ratio based on the distribution characteristics of the abnormal data, and synchronously updates them to the process parameter threshold field of the dynamic database. The database update triggers the parameter iteration mechanism of the classification model, fine-tunes the weights of the fully connected layer of the model based on the incremental learning algorithm, and optimizes the decision boundary of the aluminum-silicon ratio classification label. The dynamic database adopts a version control mechanism to retain historical parameter versions, support process parameter backtracking and the expansion of the model training data set, and form a collaborative optimization closed loop of the sorting process and classification model.

[0052] In the above process, the simultaneous execution of aluminum-silicon ratio threshold screening and sulfur element removal improves the purity control accuracy of the sorted materials. The dynamic adjustment of the air separation parameters driven by calorific value matching realizes the real-time adaptation of the sorting process and roasting requirements. The model iteration mechanism driven by abnormal data enhances the adaptability of the classification model to complex working conditions, forming an overall dual optimization path for sorting efficiency and quality control.

[0053] Specifically, in the method for classifying and separating coal gangue of the present invention, step S4 comprises: The distribution characteristics of iron, silicon and carbonate components were analyzed by X-ray fluorescence spectrometer, and the particle size distribution data of the waste was obtained by laser particle size analyzer; The component distribution characteristics and particle size data are input into the genetic algorithm to generate detection parameter optimization instructions and resource path allocation plans.

[0054] In step S4, an X-ray fluorescence spectrometer analyzes the composition of the sorted waste using multi-element simultaneous detection technology. An energy-dispersive detector collects the characteristic X-ray spectra of iron, silicon, and carbonate, and an energy spectrum analysis algorithm generates an element distribution thermogram. A laser particle size analyzer, based on the principle of dynamic light scattering, inverts the particle size distribution data of the waste particles using the multi-angle scattered light intensity distribution. The Malvern model is then used to calculate the characteristic particle size parameters D10, D50, and D90, generating a particle size distribution curve.

[0055] When inputting component distribution characteristic data and particle size distribution data into the genetic algorithm, the optimization objectives are to minimize detection error and maximize resource matching. During the algorithm initialization phase, an initial population is generated based on historical channel combination error data from the X-ray fluorescence spectrometer. Channel selection weights and laser particle size analyzer sampling density parameters are set as chromosome encodings. The fitness function, based on a comprehensive score of iron and silicon component detection errors and particle size distribution confidence, drives selection, crossover, and mutation operations to iteratively optimize the detection parameter combination.

[0056] The optimized detection parameter instructions include: the X-ray fluorescence spectrometer prioritizes detection channels corresponding to the characteristic peaks of iron Kα and silicon Kβ to reduce background noise interference; the laser particle size analyzer's sampling density is dynamically adjusted based on the confidence interval of the particle size distribution curve, increasing the density of sampling points in the area near D50 to improve resolution. The resource allocation path is implemented by matching waste composition data with the downstream production line raw material demand database. The iron content, silicon purity, and carbonate concentration are scored for similarity with the process standards of building materials, cement, and chemical fillers, respectively. A Pareto optimal solution based on a genetic algorithm is used to generate a priority list for directional transportation.

[0057] Optimization instructions are fed back to the X-ray fluorescence spectrometer and laser particle size analyzer via the industrial bus, adjusting the spectral integration time and the particle size analyzer's laser power, forming a closed-loop correction loop for detection parameters. Resource allocation instructions are transmitted to the conveyor system control terminal, driving the prioritized allocation of sorted waste to the corresponding production lines. The composition and particle size data generated during this process are synchronously updated to a dynamic database, providing incremental training data for subsequent classification model threshold iterations in the sorting process, forming a two-way collaborative mechanism for detection optimization and resource allocation.

[0058] In the above process, the joint analysis of X-ray fluorescence and particle size detection data provides multi-dimensional physical property parameters for waste resource utilization. The genetic algorithm balances detection accuracy and resource matching efficiency through multi-objective optimization. The closed-loop feedback mechanism enhances the detection system's adaptability to fluctuations in waste composition, thereby achieving an overall simultaneous improvement in waste detection accuracy and resource utilization.

[0059] Specifically, in the coal gangue classification and separation method of the present invention, the detection parameter optimization instruction includes: Iteratively optimize the detection channel combination of the X-ray fluorescence spectrometer and adjust the channel selection weight based on the goal of minimizing the detection error of iron and silicon components; Optimize the sampling density parameters of the laser particle size analyzer and dynamically adjust the sampling interval based on the confidence threshold of the particle size distribution data.

[0060] During the detection parameter optimization process, the detection channel combination of the X-ray fluorescence spectrometer is adjusted through an iterative optimization mechanism. Based on the energy spectrum analysis results of the Kα characteristic peak of iron and the Kβ characteristic peak of silicon, detection channels with high signal-to-noise ratio are prioritized, and background scattering interference is reduced through a channel weight allocation algorithm. Channel weights are dynamically corrected based on historical detection error data, and the activation priority of each channel is updated using the gradient descent method, so that the detection errors of iron and silicon components gradually converge to the target range. The optimized combination of detection channels is fed back to the spectrometer control unit via the industrial bus, which adjusts the integration time and filter switching frequency of the multi-channel detector to improve the accuracy and efficiency of elemental analysis.

[0061] The laser particle size analyzer's sampling density parameter optimization is based on the confidence threshold of the particle size distribution data. The confidence score for each size interval is calculated by statistically analyzing the data dispersion between D10 and D90 in the particle size distribution curve. In areas where the confidence level falls below the set threshold, an adaptive sampling algorithm dynamically increases the sampling point density, filling data gaps through interpolation. For areas where the confidence level meets the required level, the baseline sampling interval is used to balance detection accuracy and equipment load. Sampling density adjustment commands are implemented by controlling the laser pulse frequency and the detector acquisition cycle, enabling the particle size analyzer to obtain higher-resolution distribution data in critical particle size ranges.

[0062] The optimized detection parameters are synchronized to the dynamic database via the data bus and stored in association with the waste composition and particle size data. The iterative records of the X-ray fluorescence spectrometer channel weights and the adjustment log of the laser particle size analyzer sampling density form a parameter optimization historical data set, which is used for the initial population generation of the subsequent genetic algorithm. The continuous optimization of detection errors reduces the misjudgment rate of waste composition analysis, and the dynamic adjustment of sampling density enhances the reliability of particle size distribution data. The two work together to provide a high-confidence physical parameter basis for resource path allocation. The waste resource recovery instruction is generated based on the optimized detection data, driving the sorting system to perform directional transportation according to the composition and particle size characteristics, forming a closed-loop technical path that improves detection accuracy and enhances resource utilization efficiency.

[0063] Specifically, the gangue classification and selection method of the present invention includes the following resource utilization path allocation scheme: The waste composition data analyzed by X-ray fluorescence spectrometer and the downstream production line raw material demand database are input into the genetic algorithm to calculate the optimal allocation path; The distribution paths are dynamically sorted based on the matching scores of iron content, silicon purity and carbonate concentration with the downstream production lines, generating directional delivery instructions for building materials, cement or chemical fillers.

[0064] In the resource recovery path allocation scheme, waste composition data analyzed by X-ray fluorescence spectrometry, including quantitative test results for iron, silicon, and carbonate, is standardized to generate a composition vector. The downstream production line raw material demand database integrates the upper limit requirements for silicon purity in building materials production lines, the acceptable range for iron content in cement production lines, and the process standards for carbonate concentration in chemical filler production lines to construct a multidimensional demand matrix. The input layer of the genetic algorithm receives initial parameters for the matching degree between the component vector and the demand matrix. Using chromosome encoding, the allocation path variables are mapped to gene sequences to form an initial population.

[0065] The fitness function calculates a comprehensive score for each chromosome based on the positive correlation between iron content and cement production line demand, the negative tolerance of silicon purity and building materials production lines, and the threshold matching of carbonate concentration and chemical filler production lines. Selection retains chromosomes with high scores. Crossover and mutation operations introduce random perturbations to explore the Pareto front solution set. After iterative optimization, the optimal allocation path is output. The resulting paths are prioritized using a dynamic sorting module. Priorities are ranked based on the weighted sum of each production line's matching scores, with weight coefficients dynamically adjusted based on real-time market demand data.

[0066] The module that generates directional conveying instructions converts the priority list into control signals. Building materials production lines are assigned conveying channels for high-silicon waste, cement production lines activate magnetic separation for high-iron materials, and chemical filler production lines activate vibratory screening for carbonate materials. These conveying instructions are transmitted to the logistics control system via industrial protocols, driving the coordinated operation of belt conveyors, pneumatic sorting valves, and storage and stacking cranes. During execution, waste composition data and allocation results are fed back to a dynamic database, updating the raw material inventory and demand parameters of downstream production lines, forming a closed-loop optimization mechanism of demand-allocation-feedback.

[0067] In the above process, the matching degree between compositional data and the demand matrix provides optimization targets for the genetic algorithm. Dynamic sorting and weight adjustment mechanisms enhance the allocation path's adaptability to market fluctuations, and closed-loop feedback enables continuous optimization of resource utilization paths. Through multi-objective collaboration and real-time data-driven development, the compositional characteristics of waste are precisely matched to the process requirements of downstream production lines, improving the overall efficiency of resource utilization.

[0068] Specifically, the method for classifying and separating coal gangue of the present invention further includes: Feedback the detection parameters optimized by the genetic algorithm to the X-ray fluorescence spectrometer and laser particle size analyzer to adjust the sampling frequency and resolution of multimodal data acquisition; The execution results of the resource allocation path are input into the dynamic database to drive the update of the aluminum-silicon ratio classification threshold of the classification model.

[0069] Specifically, the method for classifying and separating coal gangue of the present invention further includes: High-silicon waste is classified into coarse and fine particles based on the particle size data optimized by the laser particle size analyzer, and allocated to the building materials or ceramic production line according to the threshold value; The iron-containing waste is combined with the magnetic separation results and the matching score of the genetic algorithm, and is preferentially transported to the high-portland cement production line; Carbonate materials are matched with building material additive standards through multi-spectral characteristics and distributed to the concrete additive production line.

[0070] During the parameter feedback and model update phase, the detection parameters optimized by the genetic algorithm are transmitted via the industrial bus to the control units of the X-ray fluorescence spectrometer and the laser particle size analyzer. The X-ray fluorescence spectrometer adjusts the integration time and filter switching frequency of the multi-channel detector based on the channel weight optimization instructions, improving the detection sensitivity of iron and silicon elements. Simultaneously, the laser particle size analyzer dynamically adjusts the laser pulse frequency and the acquisition cycle of the photodetector based on the sampling density parameters, increasing the number of sampling points in the critical particle size range and optimizing the resolution of the particle size distribution curve. Coordinated adjustment of multimodal data acquisition is achieved through device firmware upgrades, synchronizing the detection parameters with the dynamic changes in the waste material's physical properties.

[0071] The results of the resource allocation path, including scrap composition, particle size data, and actual delivery records, are processed by the data standardization module and then entered into a dynamic database. The classification model update module in the database, based on an incremental learning mechanism, integrates the actual distribution data of the aluminum-silicon ratio from the allocation results with the original training set to recalculate the classification threshold boundaries. The threshold update utilizes a sliding window method to preserve the statistical characteristics of historical data. Backpropagation is used to fine-tune the weight coefficients of the fully connected layer of the deep learning model to optimize the accuracy of the aluminum-silicon ratio critical value.

[0072] A dynamic database establishes a parameter version management mechanism, recording detection parameter adjustment logs and model threshold iteration trajectories, supporting historical data backtracking and abnormal operating condition analysis. After the classification model is updated, the new threshold parameters are distributed to the sorting equipment via the control interface, adjusting the aluminum-silicon ratio screening criteria in real time. This closed-loop linkage mechanism of detection parameter optimization and model updates enables the sorting process to adapt to fluctuations in scrap composition and changes in downstream production line demand, improving the consistency of resource allocation and sorting quality control.

[0073] In this process, real-time feedback on detection parameters enhances the accuracy of multimodal data collection, while model iteration driven by resource allocation results improves the dynamic adaptability of classification thresholds. These two processes, through collaborative database management, form a bidirectional optimization path, achieving systematic improvements in detection accuracy, sorting efficiency, and resource utilization.

[0074] Specifically, the method for classifying and separating coal gangue of the present invention further includes: During the detection parameter optimization process, an initial population is generated based on the historical error data of the detection channel combination of the X-ray fluorescence spectrometer; In the resource path matching process, the fitness function is set based on the raw material demand fluctuation characteristics of the downstream production line; The optimized detection parameters and the coordinated results of the allocation paths are used as constraints to iteratively update the crossover and mutation probability of the genetic algorithm.

[0075] In the waste distribution process, high-silicon waste is classified into coarse particles (D50 > 100μm) and fine particles (D50 ≤ 100μm) using a multi-stage vibrating screening device based on particle size distribution data optimized by a laser particle size analyzer. The coarse particles are transported via a belt conveyor to the building materials production line for aggregate preparation. The fine particles are then pneumatically conveyed to the powder pre-treatment bin of the ceramic production line to meet the fineness requirements of the ceramic sintering process. The classification threshold is dynamically adjusted based on the particle size standards of the building materials and ceramic production lines, and the classification accuracy is controlled online using an electrically adjustable screen aperture mechanism.

[0076] After magnetic separation, the iron-containing waste is tested for iron content using an X-ray fluorescence spectrometer. A genetic algorithm is then used to calculate its compatibility score with the high-Portland cement production line. This score is weighted based on the iron content, the iron oxide dosage standard for cement clinker calcination, and the silicon-aluminum ratio. High-scoring materials are preferentially transported to the cement production line's pre-homogenization yard, where the flow rate is controlled by an electromagnetic feeder. Low-scoring materials are returned to the secondary sorting process. The magnetic separation intensity is dynamically adjusted based on the score, using controllable electromagnets to adjust the magnetic field gradient to optimize the separation efficiency of the iron-containing materials.

[0077] For carbonate materials, a multispectral imaging device captures spectral reflectance curves in the 400-2500nm band, extracting the positions and intensities of the characteristic absorption peaks of the carbonate minerals. The spectral signatures are then matched against a dynamic database of industry-standard spectra for building material additives, and a matching score is calculated using the Euclidean distance algorithm. Materials with scores above a threshold are allocated to the concrete additive production line and introduced into the mixing and stirring equipment via a screw conveyor. Materials with low scores enter a secondary resource recovery assessment process. The spectral matching threshold is dynamically adjusted based on process updates from the additive production line to ensure that the allocation strategy remains consistent with industry standards.

[0078] In the allocation process described above, particle size classification, magnetic separation scoring, and spectral matching form a multi-dimensional decision-making mechanism. The integration of laser particle size data and a vibrating screening device enables precise grading of physical properties; magnetic separation and genetic algorithm scoring enhance the adaptability of chemical composition to production line requirements; and multi-spectral feature matching enhances the targeted resource recovery path through optical property analysis. Test data and allocation instructions at each stage are synchronized through a dynamic database, forming a closed-loop control system that aligns waste characteristics, production line requirements, and allocation actions, systematically improving the refinement of coal gangue resource utilization.

[0079] Specifically, the method for classifying and separating coal gangue of the present invention further includes: The gangue is classified into three grades according to particle size: 0-10mm, 10-50mm and 50-300mm by a preset vibrating screening device; For materials with a diameter of 10-50mm and 50-300mm, multispectral image data of the gangue surface mineral texture is obtained, and a gangue particle size sorting model is established based on the aluminum-silicon ratio; Identify the mineral composition of coal gangue according to the coal gangue particle size sorting model and separate the gangue with aluminum-silicon ratio ≥ 0.9; The sorted gangue is air-jet-pulverized to a particle size of ≤200μm; Conduct secondary testing and resource processing on the sorted waste; Among them, materials with a size of 0-10 mm are separated by gravity to separate useful materials with an aluminum-silicon ratio ≥ 0.9 and coal powder by-products.

[0080] The vibrating screening system utilizes a multi-layered screen structure with apertures set to critical values ​​of 10mm and 50mm. After processing through the vibrating screening system, the material is physically separated into three separate streams based on particle size range: fine particles 0-10mm, medium particles 10-50mm, and coarse particles 50-300mm. The different particle sizes are then transported to separate processing sections via a diversion device, achieving size-driven physical separation pretreatment.

[0081] Materials in the 10-50mm and 50-300mm particle sizes are transported to the multispectral imaging area via a closed conveyor belt. The multispectral imaging device collects surface reflectance data in the 400-2500nm band as the material passes through, generating a mineral texture map. A quantitative calculation model for the aluminum-silicon ratio is constructed based on the reflectance difference ratio between aluminum and silicon in the multispectral feature space. This model correlates the aluminum-silicon ratio value with particle size distribution characteristics to establish a decision rule library for particle size sorting.

[0082] The gangue particle size sorting model receives multispectral image data as input, extracts texture feature vectors, and feeds them into a classifier. The classifier uses the values ​​output by the aluminum-silicon ratio calculation model to perform a binary classification: aluminum-silicon ratios ≥ 0.9 are labeled high-value gangue, while those below this threshold are labeled waste. A pneumatic sorting valve physically separates the material based on the classification results, directing high-value gangue to a separate collection bin and waste material to a waste disposal channel.

[0083] After separation, high-value gangue is introduced into a pneumatic pulverizer via a pneumatic conveying system. The pulverizer generates a supersonic jet using high-pressure airflow, causing high-speed collisions within the pulverization chamber. The pulverized material is screened by a classifying wheel. Particles with a size greater than 200 μm are returned to the pulverization area for recycling, while fine powders meeting the particle size requirement of ≤200 μm are discharged through a negative pressure collection system. The pulverized particle size control parameters are aligned with the leaching efficiency requirements of the downstream alumina extraction process.

[0084] For fine particles sized 0-10mm, a heavy medium cyclone is used for gravity separation. Based on the density difference between aluminum-silicon minerals and pulverized coal, layered separation is achieved in a centrifugal field. The separated heavy mineral components are dehydrated and dried to produce useful materials with an aluminum-silicon ratio ≥ 0.9, while the light components are recovered as a pulverized coal byproduct. The waste from the separation is combined with the waste from the medium-coarse particle separation and transported to the secondary testing and resource recovery process.

[0085] The waste generated by each particle size separation is collected and transported to the component analysis section. The waste composition data and particle size distribution characteristics are input into the resource utilization decision-making system, triggering downstream resource allocation instructions. The processing paths of different particle sizes converge in the waste resource utilization process, forming a complete closed-loop particle size classification process.

[0086] The technical features involved in the technical solution of the present invention are explained as follows: Multispectral imaging device: A device capable of collecting multi-band reflectance data from the visible to near-infrared range (400-2500nm). By analyzing the differences in reflective properties of light at different wavelengths, it characterizes the mineral texture and composition distribution on the surface of coal gangue, identifying layered structure boundaries and mineral types.

[0087] X-ray diffractometer: Utilizing the Bragg diffraction principle of X-rays and crystal structure, this instrument analyzes the lattice arrangement and spatial distribution of aluminum and silicon in coal gangue. The diffraction pattern generates a heat map of elemental concentration gradients, quantitatively describing the local abundance differences of aluminum and silicon.

[0088] Infrared calorific value detection module: This module uses a non-contact thermal radiation sensor to measure the infrared radiation intensity of different areas on the gangue surface, generating calorific value distribution data. This data reflects the energy release characteristics of the material and is used to correlate the thermal adaptation threshold of the roasting process.

[0089] Spatial registration technology: The coordinate system of the multispectral image and X-ray diffraction data is aligned through a feature point matching algorithm to eliminate spatial offset errors caused by device posture or sampling angle, achieve accurate three-dimensional spatial superposition of data, and form a three-dimensional composition map containing element concentration, reflectivity and coordinates.

[0090] Dynamic Database: This system uses a distributed architecture to store three-dimensional composition maps, calorific value distribution data, and temporal correlation information for sorting instructions. It supports rapid retrieval and updating of multimodal data, providing structured data support for deep learning model training and sorting parameter optimization.

[0091] Aluminum-silicon concentration gradient characteristics: Based on X-ray diffraction data, the grayscale gradient analysis method is used to extract the spatial concentration variation pattern of aluminum and silicon elements, and the relative abundance of elements in different regions is quantified as the key input feature for aluminum-silicon ratio classification.

[0092] Layered structure boundary identification: Using multispectral image data, the Canny edge detection algorithm is used to identify the structural boundaries of layered minerals in coal gangue. Morphological filtering is combined to remove noise and generate boundary contour maps to assist in the training of the aluminum-silicon ratio classification model.

[0093] Deep learning model (ResNet-50 architecture): A convolutional neural network whose input layer receives concentration gradient maps, boundary recognition maps, and multispectral reflectance data. It extracts local features through a residual connection structure. The fully connected layer learns the nonlinear relationship between the aluminum-silicon ratio and the calorific value distribution, and outputs classification labels to generate sorting instructions.

[0094] Sulfur element spectrum detection module (LIBS technology): Based on the principle of laser-induced breakdown spectroscopy, it collects the characteristic spectral signals of sulfur elements on the material surface in real time (such as the wavelength range of 180-280nm), calculates the sulfur concentration through a threshold judgment algorithm, and dynamically adjusts the pneumatic nozzle to remove the proportion of sulfur-containing impurities.

[0095] Genetic algorithm optimization: With the goal of minimizing detection error and maximizing resource matching, the detection channel weights of the X-ray fluorescence spectrometer (such as the priority of iron Kα and silicon Kβ channels) and the sampling density parameters of the laser particle size analyzer (such as the D50 interval sampling point density) are iteratively adjusted to optimize detection accuracy and resource allocation efficiency.

[0096] Resource Path Allocation: Based on the matching scores between waste composition (iron, silicon, carbonate) and the process standards of downstream production lines (building materials, cement, chemicals), a genetic algorithm generates a priority list, which drives the conveying system to allocate materials in a targeted manner according to the scoring results, achieving efficient resource utilization of waste.

[0097] Incremental learning mechanism: During the sorting process, abnormal calorific value data and sorting parameter correction instructions are recorded through a dynamic database, the sliding window method is used to retain the statistical characteristics of historical data, and the weights of the fully connected layer of the deep learning model are fine-tuned to achieve dynamic iterative updates of the aluminum-silicon ratio classification threshold.

[0098] Aluminum-Silicon Ratio Classification Model: This model is built on a deep learning framework. Its input data includes reflectance features from multispectral images, aluminum-silicon concentration gradients extracted from X-ray diffraction data, and layered structure boundary identification results. The model uses a convolutional neural network (such as ResNet-50) to extract local features. Residual connections address the vanishing gradient problem in deep network training. Fully connected layers learn the nonlinear relationship between aluminum-silicon concentration gradients and calorific value distribution. The output layer generates classification labels using a softmax function. These labels represent the critical aluminum-silicon ratio value, guiding sorting equipment to select materials based on the threshold. The model's training data is mapped to calorific value distribution data using a three-dimensional composition map in a dynamic database, enhancing the adaptability of the classification results to the roasting process.

[0099] Genetic Algorithm Optimization Model: This algorithm is used to optimize detection parameters and resource allocation paths. For the detection channel combination of an X-ray fluorescence spectrometer, the algorithm aims to minimize the detection errors of iron and silicon components. It characterizes channel weights through chromosome encoding, and the fitness function dynamically adjusts channel priorities based on historical error data and confidence scores. In resource allocation, the algorithm matches waste composition (iron, silicon, carbonate) with downstream production line demand, constructs a multidimensional demand matrix, explores Pareto optimal solutions through crossover and mutation operations, and generates directional transport instructions. The initial population is initialized using historical detection error data, and during the iterative process, the fluctuation characteristics of downstream production line demand are introduced as constraints to improve resource allocation efficiency.

[0100] Incremental Learning Mechanism: This mechanism dynamically updates the classification thresholds of the deep learning model. Based on abnormal calorific value data detected during the sorting process, the incremental learning module uses a sliding window method to retain the statistical characteristics of historical data. It then fine-tunes the weight coefficients of the fully connected layer through a backpropagation algorithm to optimize the decision boundary for the critical aluminum-silicon ratio. The updated model parameters are stored in a dynamic database with version control, allowing the sorting equipment to adjust the screening criteria in real time to adapt to fluctuations in gangue composition and changes in process parameters.

[0101] Sulfur spectral detection model: Based on laser-induced breakdown spectroscopy (LIBS) technology, this model collects characteristic sulfur spectral signals (e.g., wavelengths between 180 and 280 nm) and combines them with a threshold-based algorithm to calculate sulfur concentration in real time. The model's output signal drives the pneumatic nozzles of the sorting equipment to dynamically adjust the rejection rate (e.g., between 5% and 15%). The rejection rate is positively correlated with the sulfur concentration deviation, forming a closed-loop control logic for the removal of sulfur-containing impurities.

[0102] Resource Recycling Matching Scoring Model: This model uses a Euclidean distance algorithm or similarity calculation to match waste composition data (such as iron content, silicon purity, and carbonate concentration) with the process standards of downstream production lines (building materials, cement, and chemical fillers). The scoring results serve as input to the fitness function of a genetic algorithm, driving the generation of a priority list. For example, waste with an iron content greater than 15% is prioritized for cement production, while waste with a silicon purity greater than 80% is directed to building materials production. This achieves dynamic sorting and precise allocation of resource recycling paths.

[0103] During the gangue sorting process, a multispectral imaging device first collects mineral texture data from the visible to near-infrared wavelengths (400-2500nm). Simultaneously, an X-ray diffractometer is used to analyze the lattice structure distribution of aluminum and silicon, generating a thermal map of the element's spatial distribution. An infrared calorific value detection module uses non-contact thermal radiation measurement technology to capture the energy release characteristics of the gangue surface and generate calorific value distribution data. The multispectral image and X-ray diffraction data are spatially aligned using a feature point matching algorithm. After eliminating device position errors, pixel-level weighted fusion is used to generate a three-dimensional composition map. Each pixel in the map is associated with element concentration, reflectivity, and spatial coordinate information. The calorific value data is dynamically linked to the map using a timestamp and stored in a distributed database, forming a multimodal dataset containing composition, calorific value, and spatial distribution, providing high-resolution input for aluminum-silicon ratio classification.

[0104] Based on a three-dimensional composition map, grayscale gradient analysis is used to extract aluminum-silicon concentration gradient features. Combined with the layered structure boundary information identified in the multispectral image using the Canny edge detection algorithm, a joint training dataset is constructed. The deep learning model utilizes the ResNet-50 architecture. The input layer receives the concentration gradient map, boundary identification map, and multispectral reflectance data. The convolutional layer extracts local features, and the fully connected layer learns the relationship between the aluminum-silicon ratio and calorific value distribution. After the model outputs classification labels, it uses a mapping module to generate sorting instructions based on the calorific value data in the dynamic database. The instructions contain material priority weights and calorific value adaptation threshold ranges. The priority weights are dynamically adjusted based on the positive correlation between the aluminum-silicon ratio and calorific value. The sorting instructions are transmitted to the photoelectric sorter via the Modbus protocol, controlling the pneumatic nozzle to remove sulfur-containing impurities. LIBS technology is used for sulfur detection to calculate the concentration deviation value in real time, dynamically adjusting the rejection ratio to a range of 5%-15%.

[0105] The sorted waste is analyzed for iron, silicon, and carbonate content using an X-ray fluorescence spectrometer. Detection channels prioritize the characteristic peaks of iron Kα (6.40 keV) and silicon Kβ (1.74 keV). A laser particle size analyzer collects D10-D90 particle size distribution data based on dynamic light scattering. Composition and particle size data are fed into a genetic algorithm, with an initial population generated from historical detection error data. The fitness function aims to minimize detection error and maximize resource matching. After iterative optimization, the algorithm outputs detection channel weights and sampling density parameters. The optimized parameters are fed back to the equipment, adjusting the X-ray fluorescence spectrometer integration time to 200 ms and increasing the laser particle size analyzer sampling density to 50 points / mm in the D50±10μm range. Resource allocation instructions are dynamically sorted based on the genetic algorithm's matching scores based on the process standards of iron content >15%, silicon purity >80%, and carbonate concentration >30%. These instructions drive a belt conveyor to distribute the material to the building materials, cement, or chemical filler production lines. During the sorting process, abnormal calorific value data triggers an incremental learning mechanism, updates the classification model threshold, and forms a closed-loop control system of detection-sorting-feedback, thereby improving the accuracy of identifying the critical value of the aluminum-silicon ratio and systematically solving the overburning problem.

[0106] This invention addresses the problem of insufficient quantitative analysis of the distribution of aluminum and silicon elements within coal gangue by integrating multimodal data with a deep learning model. First, a multispectral imaging device and an X-ray diffractometer are used to collect data on the mineral texture and spatial distribution of aluminum and silicon elements on the coal gangue surface. Spatial registration technology is then used to generate a three-dimensional compositional map containing aluminum and silicon concentration gradients. This three-dimensional map is then associated with infrared calorific value distribution data and stored in a dynamic database. Combined with time series analysis, this quantitatively characterizes the distribution characteristics of aluminum and silicon elements and their dynamic relationship with calorific value, providing a high-precision data foundation for subsequent critical value determination.

[0107] Secondly, based on the three-dimensional compositional map, the aluminum-silicon concentration gradient characteristics are extracted. Combined with the layered structure boundary identification results from the multispectral image, a deep learning model is trained to output classification labels for the aluminum-silicon ratio. Using a convolutional neural network, the model integrates multispectral reflectance, element concentration gradients, and boundary structure characteristics to learn the nonlinear relationship between the aluminum-silicon ratio and calorific value distribution, dynamically generating classification labels. The classification labels are mapped to a calorific value adaptation threshold matrix to generate sorting instructions, instructing the sorting equipment to select materials based on the aluminum-silicon ratio threshold. Furthermore, impurities are dynamically removed through sulfur spectral detection, ensuring that sorting accuracy is aligned with the calorific value requirements of the roasting process.

[0108] Finally, a genetic algorithm is used to optimize detection parameters and resource recovery paths, forming a closed-loop feedback mechanism. The sorted waste undergoes secondary inspection using an X-ray fluorescence spectrometer and a laser particle size analyzer. Composition and particle size data are then fed into a genetic algorithm to iteratively optimize the detection channel combination and sampling density parameters, improving detection accuracy. The optimized parameters are fed back to the equipment to drive updates to the classification model threshold. Simultaneously, resource recovery allocation instructions dynamically adjust the conveying path based on the composition matching score. Multimodal data collaboration and algorithm optimization enable the sorting process to adapt to fluctuations in gangue composition and the needs of downstream production lines, systematically addressing the overburning problem caused by inaccurate critical values ​​for the aluminum-silicon ratio.

Claims

1. A method for classifying and separating coal gangue, characterized in that: include: Step S1, obtaining multispectral image data, X-ray diffraction data, and calorific value distribution data representing the surface mineral texture of coal gangue, spatially registering and fusing the multispectral image data with the X-ray diffraction data to generate a three-dimensional composition map, and associating the three-dimensional composition map with the calorific value distribution data and storing them in a dynamic database; Step S2: Based on the three-dimensional component map in the dynamic database, extract the characteristics of the aluminum-silicon concentration gradient change, combine the layered structure boundary recognition results in the multispectral image data, train the aluminum-silicon ratio classification model, output a classification label representing the critical value of the aluminum-silicon ratio, and represent the critical value of the aluminum-silicon ratio as aluminum-silicon ratio ≥ 0.

9. The classification label is associated with the calorific value distribution data stored in the dynamic database to generate a sorting instruction including a material sorting priority and a calorific value adaptation threshold, with a target calorific value ≤ 1000 kcal / kg, and a grading threshold of the calorific value adaptation threshold of less than 400 kcal / kg, 400-600 kcal / kg, and 400-800 kcal / kg, and the instruction is transmitted to the sorting equipment control terminal; Step S3: Controlling the sorting equipment to screen materials with an aluminum-silicon ratio higher than a threshold value of 0.9 according to the sorting instruction, and dynamically removing sulfur-containing impurities based on the detection results output by the sulfur element spectrum detection module; matching the infrared calorific value distribution data stored in the dynamic database with the preset roasting kiln process parameter threshold, with a target calorific value of ≤1000kcal / kg, and graded adaptation thresholds of the roasting kiln process parameter threshold of less than 400kcal / kg, 400-600kcal / kg, and 400-800kcal / kg; calculating the sorting ratio control parameters and sending them to the air separation equipment; associating the abnormal calorific value data detected during the sorting process with the correction instruction of the sorting parameters, and updating them to the process parameter threshold in the dynamic database; Step S4, performing secondary testing on the sorted waste, said secondary testing comprising: analyzing the iron, silicon, and carbonate components by an X-ray fluorescence spectrometer, obtaining particle size distribution data by a laser particle size analyzer, inputting the component analysis data and particle size data into a genetic algorithm, iteratively optimizing the detection channel combination of the X-ray fluorescence spectrometer and the sampling density parameters of the laser particle size analyzer, generating instructions for waste resource allocation, and feeding the optimized detection channel combination and sampling density parameters back to the X-ray fluorescence spectrometer and the laser particle size analyzer.

2. The method for classifying and separating coal gangue according to claim 1, characterized in that: The step S1 comprises: Multispectral image data of coal gangue surface mineral texture is collected through a multispectral imaging device, X-ray diffraction data is obtained and the spatial distribution of aluminum and silicon elements is analyzed through an X-ray diffractometer, and calorific value distribution data is obtained through an infrared calorific value detection module; spatially registering the multispectral image data with the X-ray diffraction data, and fusing them to generate a three-dimensional composition map including aluminum and silicon element concentration gradients; The calorific value distribution data is associated with the time series of the three-dimensional composition map to construct input features of the dynamic database.

3. The method for classifying and separating coal gangue according to claim 2, characterized in that: The step S2 comprises: extracting the concentration gradient variation characteristics of aluminum and silicon elements based on X-ray diffraction data obtained by the X-ray diffractometer, combining the layered structure boundary recognition results in the multispectral image data, training a deep learning model to output an aluminum-silicon ratio classification label; The classification label is associated with the calorific value distribution data obtained by the infrared calorific value detection module to generate a sorting equipment execution instruction, which includes the material sorting priority and the calorific value adaptation threshold.

4. The method for classifying and separating coal gangue according to claim 3, characterized in that: The step S3 comprises: According to the sorting instructions, materials with aluminum-silicon ratios higher than the set threshold are screened, and the removal ratio of sulfur-containing impurities is dynamically adjusted based on the sulfur element spectrum detection results; Adjust the air pressure and material flow of the air separation equipment based on the matching calculation results of the infrared calorific value distribution data and the preset roasting kiln process parameter thresholds stored in the dynamic database; The abnormal calorific value data and the correction instructions of the sorting parameters are synchronously updated to the dynamic database, triggering the iterative update of the parameters of the classification model.

5. The method for classifying and separating coal gangue according to claim 4, characterized in that: The step S4 comprises: The distribution characteristics of iron, silicon and carbonate components were analyzed by X-ray fluorescence spectrometer, and the particle size distribution data of the waste was obtained by laser particle size analyzer; The component distribution characteristics and particle size data are input into the genetic algorithm to generate detection parameter optimization instructions and resource path allocation plans.

6. The method for classifying and separating coal gangue according to claim 5, characterized in that: The detection parameter optimization instruction includes: Iteratively optimize the detection channel combination of the X-ray fluorescence spectrometer and adjust the channel selection weight based on the goal of minimizing the detection error of iron and silicon components; Optimize the sampling density parameters of the laser particle size analyzer and dynamically adjust the sampling interval based on the confidence threshold of the particle size distribution data.

7. The method for classifying and separating coal gangue according to claim 6, characterized in that: The resource path allocation scheme includes: The waste composition data analyzed by X-ray fluorescence spectrometer and the downstream production line raw material demand database are input into the genetic algorithm to calculate the optimal allocation path; During the detection parameter optimization process, an initial population is generated based on the historical error data of the detection channel combination of the X-ray fluorescence spectrometer; In the resource path matching process, the fitness function is set based on the raw material demand fluctuation characteristics of the downstream production line; The optimized detection parameters and the coordinated results of the allocation path are used as constraints to iteratively update the crossover and mutation probability of the genetic algorithm. The distribution paths are dynamically sorted based on the matching scores of iron content, silicon purity and carbonate concentration with the downstream production lines, generating directional delivery instructions for building materials, cement or chemical fillers.

8. The method for classifying and separating coal gangue according to claim 7, characterized in that: Also includes: Feedback the detection parameters optimized by the genetic algorithm to the X-ray fluorescence spectrometer and laser particle size analyzer to adjust the sampling frequency and resolution of multimodal data acquisition; The execution results of the resource allocation path are input into the dynamic database to drive the update of the aluminum-silicon ratio classification threshold of the classification model.

9. The method for classifying and separating coal gangue according to claim 8, characterized in that: Also includes: High-silicon waste is classified into coarse and fine particles based on the particle size data optimized by the laser particle size analyzer, and allocated to the building materials or ceramic production line according to the threshold value; The iron-containing waste is combined with the magnetic separation results and the matching score of the genetic algorithm, and is preferentially transported to the high-portland cement production line; Carbonate materials are matched with building material additive standards through multi-spectral characteristics and distributed to the concrete additive production line.

10. The method for classifying and separating coal gangue according to claim 9, characterized in that: Also includes: The gangue is classified into three grades according to particle size: 0-10mm, 10-50mm and 50-300mm by a preset vibrating screening device; For materials with a diameter of 10-50mm and 50-300mm, multispectral image data of the gangue surface mineral texture is obtained, and a gangue particle size sorting model is established based on the aluminum-silicon ratio; Identify the mineral composition of coal gangue according to the coal gangue particle size sorting model and separate the gangue with aluminum-silicon ratio ≥ 0.9; The sorted gangue is air-jet-pulverized to a particle size of ≤200μm; Conduct secondary testing and resource recovery on the sorted waste; Among them, materials with a size of 0-10 mm are separated by gravity to separate useful materials with an aluminum-silicon ratio ≥ 0.9 and coal powder by-products.

Images