Method and system for microwave-mechanical synergistic activation of lithium and gallium in coal gangue for gradient enrichment and extraction

Abstract

The application relates to the technical field of coal gangue resource utilization and artificial intelligence, and discloses a method and system for microwave-mechanical synergistic activation of lithium and gallium in coal gangue and gradient enrichment and extraction, which comprises the following steps: placing coal gangue raw materials in a microwave reactor, irradiating the coal gangue raw materials according to optimal microwave parameters output by an intelligent optimization system, and obtaining coal gangue materials by inducing directional microcracks; constructing a high-precision proxy model by using a reinforced random forest model, coupling a multi-objective optimization algorithm, obtaining optimal ball milling parameters, adding a composite grinding aid for wet ball milling activation, inducing lattice distortion, dislocation, defect proliferation and amorphous transformation of mineral crystals, and obtaining activated slag by solid-liquid separation; and according to optimal leaching parameters, performing two-stage differential gradient leaching to obtain lithium-containing leaching liquor and gallium-containing leaching liquor. Thus, the full-process technological parameters can be rapidly and globally intelligently optimized, and the lithium and gallium leaching efficiency is greatly improved.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of coal gangue resource utilization and artificial intelligence, and in particular to a microwave-mechanical synergistic activation method and system for the graded enrichment and extraction of lithium gallium from coal gangue. Background Technology

[0002] Coal gangue is a large-scale solid waste generated during coal mining, washing, and processing. Its stockpile is enormous and its annual increase is continuously rising. Its accumulation not only occupies land resources but also easily causes environmental problems such as dust and leaching pollution. Coal gangue contains rare and dispersed metals such as lithium and gallium, which are core raw materials for strategic emerging industries such as new energy and semiconductors, possessing extremely high recycling value. Therefore, the resource utilization of rare and dispersed metals in coal gangue has become a research hotspot in the field of solid waste disposal and resource regeneration.

[0003] Currently, the extraction of lithium and gallium from coal gangue mainly relies on hydrometallurgical processes, with mineral activation and acid leaching being the core steps. However, existing technologies still suffer from several critical shortcomings in this field: First, the mineral activation effect is poor. Lithium and gallium are mostly found in stable aluminosilicate crystal lattices. Traditional high-temperature roasting activation methods are energy-intensive and prone to flue gas pollution, while conventional mechanical activation can only physically break down mineral particles and cannot deeply disrupt the aluminosilicate crystal structure, resulting in low lithium and gallium leaching rates and poor separation selectivity. Second, the efficiency of process parameter optimization is low. Hydrometallurgy involves dozens of process parameters across multiple stages, including microwave, ball milling, and leaching. There are strong coupling relationships between the parameters. Existing technologies rely heavily on human experience for parameter control, resulting in high trial-and-error costs and long debugging cycles. They cannot achieve rapid adaptation to coal gangue raw materials of different grades. Thirdly, the support capability of machine learning models is insufficient. Traditional machine learning models such as random forest (RF) have been tried for process parameter prediction. However, these models adopt the criterion of "randomly extracting feature subsets + minimizing local mean square error" when splitting decision tree nodes. They have insufficient fitting accuracy for high-dimensional and nonlinear metallurgical process data and weak generalization ability. They cannot serve as high-precision surrogate models to support global intelligent optimization of process parameters.

[0004] Existing technologies lack an integrated process that takes into account deep mineral activation, intelligent parameter optimization, and efficient separation of lithium and gallium, which severely limits the large-scale and high-value recovery of lithium and gallium rare metals from coal gangue.

[0005] The lack of an integrated process that takes into account deep mineral activation, intelligent parameter optimization, and efficient separation of lithium and gallium extraction has severely limited the large-scale and high-value recovery of lithium and gallium rare metals from coal gangue, and no effective solution has yet been proposed. Summary of the Invention

[0006] Therefore, this application provides a microwave-mechanical synergistic activation method and system for the graded enrichment and extraction of lithium and gallium from coal gangue, in order to overcome the problem that the related technologies lack an integrated process that takes into account the deep activation of minerals, intelligent parameter optimization, and efficient separation of lithium and gallium extraction, which seriously limits the large-scale and high-value recovery of lithium and gallium rare metals from coal gangue.

[0007] To achieve the above objectives, according to a first aspect of the embodiments of this application, a method for the graded enrichment and extraction of lithium and gallium from coal gangue through microwave-mechanical synergistic activation is provided, comprising: placing crushed coal gangue raw material, screened to below 200 mesh, into a microwave reactor, and subjecting it to irradiation treatment according to the optimal microwave parameters output by an intelligent optimization system, thereby forming a non-uniform temperature field inside the particles by utilizing the differences in dielectric constants and microwave absorption capabilities of different minerals in the coal gangue raw material, inducing directional microcracks to obtain coal gangue material; feeding the coal gangue material into a planetary ball mill, and adding a composite grinding aid according to the optimal ball milling parameters output by the intelligent optimization system for wet ball milling activation, using the microcracks as stress concentration weak surfaces to induce lattice distortion, dislocation, defect proliferation, and amorphization transformation of mineral crystals, and obtaining activated slag after solid-liquid separation after ball milling; and subjecting the activated slag to two-stage differentiated graded leaching according to the optimal leaching parameters output by the intelligent optimization system to obtain lithium-containing leachate and gallium-containing leachate respectively.

[0008] In an optional embodiment, before performing irradiation treatment based on the optimal microwave parameters output by the intelligent optimization system, the method further includes: integrating N reinforcement decision trees to form a reinforcement random forest model, where N is a positive integer; using the trained reinforcement random forest model as a surrogate model, coupling it with a non-dominated sorting genetic algorithm to construct a two-layer optimization framework, thereby generating the intelligent optimization system.

[0009] In one optional embodiment, a reinforced random forest model is constructed by integrating N reinforced decision trees, including: integrating orthogonal experimental data and pilot-scale verification data from the laboratory, valid experimental data of lithium gallium extraction processes from similar coal gangue in publicly available literature, and online data collected from industrial pilot production to construct a sample dataset; normalizing multiple features in the sample dataset, and using the three sigma criterion to remove outliers from the obtained normalized features to obtain multiple normalized features after filtering; dividing the multiple normalized features after filtering into a training dataset and a test dataset according to a preset ratio, wherein the number of features in the training dataset is greater than that in the test dataset; extracting N sets of feature samples from the training dataset using sampling with replacement, and constructing N reinforced decision trees from the N sets of feature samples respectively, wherein the N sets of feature samples and the N sets of test trees are integrated into a random forest model. The N reinforcement decision trees are configured in a one-to-one correspondence. Each set of feature samples includes input features and output features. The input features include combinations of process parameters, and the output features include target variables. The combinations of process parameters include: raw material characteristic parameters, microwave pretreatment parameters, mechanical activation parameters, and leaching process parameters. The target variables include: lithium leaching rate, gallium leaching rate, overall energy consumption, and reagent consumption per unit product. The reinforcement random forest model is constructed based on the N reinforcement decision trees. The prediction result of the reinforcement random forest model for the target input sample is the arithmetic mean of the prediction results of the N reinforcement decision trees for the target input sample. The reinforcement random forest model is tested using a five-fold forward chain cross-validation method and the test dataset. If the reinforcement random forest model passes the test, the training of the reinforcement random forest model is considered complete.

[0010] In an optional embodiment, the method further includes: when constructing any internal node of each reinforcement decision tree, performing the following steps: fitting an embedded random forest sub-model using the dataset corresponding to the arbitrary internal node; and calculating the variable importance V(f) of all input features under the arbitrary internal node based on the out-of-bag data permutation method, using the following formula: Among them, MSE n PMSE is the baseline mean squared error calculated based on out-of-bag data for the embedded random forest sub-model. n (f) represents the mean squared error calculated by the embedded submodel after randomly permuting the out-of-bag data of feature f. The number of decision trees in the embedded random forest sub-model is given; the feature with the highest variable importance is selected as the optimal splitting feature for any internal node, and the splitting threshold is determined based on the minimum sum of the mean square errors of the two child nodes after splitting, thus completing the node partitioning.

[0011] In an optional embodiment, a trained reinforced random forest model is used as a surrogate model and coupled with a non-dominated sorting genetic algorithm to construct a two-layer optimization framework, generating the intelligent optimization system. This includes: establishing a mapping relationship between the process parameter combination and the lithium leaching rate, the gallium leaching rate, the overall energy consumption, and the reagent consumption per unit product; using this mapping relationship as the fitness evaluation function of the non-dominated sorting genetic algorithm; setting multi-objective optimization objectives as: maximizing lithium leaching rate, maximizing gallium leaching rate, minimizing overall energy consumption, and minimizing reagent consumption per unit product; initializing the population and iteratively optimizing within the parameter constraint space using the non-dominated sorting genetic algorithm; outputting a Pareto optimal solution set after iterative convergence; wherein the parameter constraint space is a reasonable range of values ​​for multiple process parameters; and selecting the optimal process parameter combination from the Pareto optimal solution set according to actual production priority requirements, wherein the optimal process parameter combination includes: the optimal microwave parameter, the optimal ball milling parameter, and the optimal leaching parameter.

[0012] In an optional embodiment, the activated slag is subjected to a two-stage differentiated leaching process based on the optimal leaching parameters output by the intelligent optimization system to obtain a lithium-containing leaching solution and a gallium-containing leaching solution, respectively. This process includes: mixing the activated slag with a first leaching agent at a liquid-to-solid ratio of 3:1 to 10:1, wherein the first leaching agent is a sulfuric acid solution or hydrochloric acid solution with a concentration of 0.5-2 mol / L; leaching is carried out under the conditions of a temperature of 25-60℃, a leaching time of 30-120 min, and a stirring speed of 200-400 r / min; and solid-liquid separation is performed after leaching to obtain the lithium-containing leaching solution and the lithium-extracted leaching slag.

[0013] In an optional embodiment, the activated slag is subjected to two-stage differentiated leaching according to the optimal leaching parameters output by the intelligent optimization system to obtain a lithium-containing leaching solution and a gallium-containing leaching solution, respectively. This includes: mixing the lithium-extracted leaching slag with a second leaching agent at a liquid-solid ratio of 4:1-12:1, wherein the second leaching agent is a sulfuric acid solution with a concentration of 2-4 mol / L and a sodium fluoride leaching aid with a mass fraction of 1%-5%; leaching is carried out at a temperature of 80-150℃, a leaching time of 60-180 min, and a stirring speed of 300-500 r / min; after leaching, solid-liquid separation is performed to obtain the gallium-containing leaching solution and leaching tailings.

[0014] According to a second aspect of the embodiments of this application, a microwave-mechanical synergistic activation system for the graded enrichment and extraction of lithium-gallium from coal gangue is also provided, comprising: a microwave irradiation module for placing crushed coal gangue raw material, screened to below 200 mesh, into a microwave reactor, and performing irradiation treatment according to the optimal microwave parameters output by an intelligent optimization system, utilizing the differences in dielectric constants and microwave absorption capabilities of different minerals in the coal gangue raw material to form a non-uniform temperature field inside the particles, inducing directional microcracks to obtain coal gangue material; and a ball milling activation module. The system is used to feed the coal gangue material into a planetary ball mill, add composite grinding aids to perform wet ball milling activation according to the optimal ball milling parameters output by the intelligent optimization system, and use the microcracks as the stress concentration weak surface to induce lattice distortion, dislocation, defect proliferation and amorphization transformation of mineral crystals. After ball milling, solid-liquid separation is performed to obtain activated slag; the leaching module is used to perform two-stage differentiated step leaching of the activated slag according to the optimal leaching parameters output by the intelligent optimization system to obtain lithium-containing leaching solution and gallium-containing leaching solution respectively.

[0015] According to a third aspect of the embodiments of this application, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the microwave-mechanical synergistic activation method for lithium gallium cascade enrichment and extraction from coal gangue as described in the first aspect through the computer program.

[0016] According to a fourth aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the microwave-mechanical synergistic activation method for the graded enrichment and extraction of lithium gallium from coal gangue as described in the first aspect.

[0017] According to this application, coal gangue raw materials, after being crushed and screened to below 200 mesh, enter a microwave reactor. The intelligent optimization system outputs the optimal microwave parameters to guide the irradiation treatment. The carbonaceous and iron-containing minerals in the coal gangue differ from aluminosilicate minerals such as kaolinite and quartz. The different microwave absorption capacities lead to a non-uniform temperature field within the particles, and the thermal expansion mismatch induces directional microcracks. The lithium-gallium mineral phase is exposed from the stable lattice, resulting in pretreated coal gangue material. The pretreated coal gangue material is then fed into a planetary ball mill. The intelligent optimization system outputs the optimal ball milling parameters, and a composite grinding aid is added for wet ball milling activation. The microcracks formed by microwave pretreatment become stress concentration weak points. Mechanical impact and shearing propagate along the microcracks, causing lattice distortion, dislocation multiplication, defect proliferation, and even amorphization transformation of the mineral crystals. The stable aluminosilicate lattice structure is deeply destroyed, and the lattice-bound lithium-gallium elements are transformed into a leaching active state. After ball milling, the slurry undergoes solid-liquid separation to obtain activated slag. The activated slag then enters a two-stage differentiated leaching process. The intelligent optimization system outputs the optimal leaching parameters to achieve selective separation of lithium and gallium, resulting in lithium-containing leachate and gallium-containing leachate. This application achieves deep destruction of the mineral lattice and active transformation of elements through microwave-mechanical synergistic activation. The intelligent optimization system outputs the optimal process parameters for the entire process. The two-stage leaching achieves selective dissolution and preliminary separation based on the difference in lithium and gallium binding energies, with the comprehensive lithium and gallium leaching rates both exceeding 85%. This solves the problem of the lack of an integrated process that takes into account deep mineral activation, intelligent parameter optimization, and efficient separation of lithium and gallium in related technologies, which seriously limits the large-scale and high-value recovery of lithium and gallium rare metals in coal gangue.

[0018] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0019] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of the application. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0020] Figure 1 This is a flowchart of an optional microwave-mechanical synergistic activation method for the graded enrichment and extraction of lithium gallium from coal gangue, according to an embodiment of this application.

[0021] Figure 2 This is a flowchart of an optional microwave-mechanical synergistic activation-step leaching process according to an embodiment of this application;

[0022] Figure 3This is a flowchart illustrating the construction of an intelligent optimization system for process parameters based on reinforced random forest, according to an embodiment of this application.

[0023] Figure 4 This is a structural block diagram of an optional microwave-mechanical synergistic activation lithium gallium cascade enrichment and extraction system for coal gangue, according to an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0026] To address the technical problems existing in related technologies, this embodiment provides a microwave-mechanical synergistic activation method for the graded enrichment and extraction of lithium and gallium from coal gangue. Figure 1 This is a flowchart of a microwave-mechanical synergistic activation method for the graded enrichment and extraction of lithium gallium from coal gangue according to an embodiment of this application. The process includes the following steps:

[0027] Step S102: The crushed coal gangue raw material, screened to below 200 mesh, is placed in a microwave reactor and irradiated according to the optimal microwave parameters output by the intelligent optimization system. The difference in dielectric constant and microwave absorption capacity of different minerals in the coal gangue raw material is used to form a non-uniform temperature field inside the particles, inducing directional microcracks to obtain coal gangue material.

[0028] Step S104: The coal gangue material is fed into a planetary ball mill. According to the optimal ball milling parameters output by the intelligent optimization system, a composite grinding aid is added for wet ball milling activation. The microcracks are used as the stress concentration weak surface to induce lattice distortion, dislocation, defect proliferation and amorphization transformation of mineral crystals. After ball milling, solid-liquid separation is performed to obtain activated slag.

[0029] Step S106: The activated residue is subjected to two-stage differentiated leaching according to the optimal leaching parameters output by the intelligent optimization system to obtain lithium-containing leaching solution and gallium-containing leaching solution respectively.

[0030] In this embodiment, coal gangue raw material is crushed and screened to below 200 mesh before entering a microwave reactor. The intelligent optimization system outputs optimal microwave parameters to guide irradiation treatment. The carbonaceous and iron-containing minerals in the coal gangue differ from aluminosilicate minerals such as kaolinite and quartz. These differences in microwave absorption capacity lead to a non-uniform temperature field within the particles, and the thermal expansion mismatch induces directional microcracks. The lithium-gallium mineral phase is exposed from the stable lattice, resulting in pretreated coal gangue material. The pretreated coal gangue material is then fed into a planetary ball mill. The intelligent optimization system outputs optimal ball milling parameters, and a composite grinding aid is added for wet ball milling activation. The microcracks formed by microwave pretreatment become stress concentration weak points. Mechanical impact and shearing propagate along the microcracks, causing lattice distortion, dislocation multiplication, defect proliferation, and even amorphization transformation of the mineral crystals. The stable aluminosilicate lattice structure is deeply destroyed, and the lattice-bound lithium-gallium elements are transformed into a leaching active state. After ball milling, the slurry undergoes solid-liquid separation to obtain activated slag. The activated slag then enters a two-stage differentiated leaching process. The intelligent optimization system outputs the optimal leaching parameters to achieve selective separation of lithium and gallium, resulting in lithium-containing leachate and gallium-containing leachate. This application achieves deep destruction of the mineral lattice and active transformation of elements through microwave-mechanical synergistic activation. The intelligent optimization system outputs the optimal process parameters for the entire process. The two-stage leaching achieves selective dissolution and preliminary separation based on the difference in lithium and gallium binding energies, with the comprehensive lithium and gallium leaching rates both exceeding 85%. This solves the problem of the lack of an integrated process that takes into account deep mineral activation, intelligent parameter optimization, and efficient separation of lithium and gallium in related technologies, which seriously limits the large-scale and high-value recovery of lithium and gallium rare metals in coal gangue.

[0031] Optionally, before performing irradiation treatment based on the optimal microwave parameters output by the intelligent optimization system, the method further includes: integrating N reinforcement decision trees to form a reinforcement random forest model, where N is a positive integer; using the trained reinforcement random forest model as a surrogate model, coupling it with a non-dominated sorting genetic algorithm to construct a two-layer optimization framework, thereby generating the intelligent optimization system.

[0032] The construction of the intelligent optimization system begins with the development of a reinforced random forest model. This model consists of an ensemble of N reinforced decision trees, where N is a positive integer, and 70 trees are selected in the technical solution. When constructing the internal nodes of each reinforced decision tree, the traditional random forest's random feature selection method is abandoned. Instead, an embedded random forest sub-model is first fitted using the node dataset, and then the variable importance of all input features is calculated using the out-of-bag permutation method. The feature with the highest variable importance is selected as the optimal splitting feature to complete the node partitioning. The model training data covers multi-source data including laboratory orthogonal experiments, pilot-scale verification, publicly available literature, and online data collected during industrial pilot production. The total number of effective samples is no less than 1000. The 38-dimensional input features cover raw material characteristics, microwave pretreatment, mechanical activation, and leaching process parameters. The output targets include lithium leaching rate, gallium leaching rate, comprehensive energy consumption, and reagent consumption per unit product. After training, the model's prediction coefficient of determination for lithium-gallium leaching rate is no less than 0.87, and the prediction accuracy reaches over 92%.

[0033] The trained reinforcement random forest model is coupled with a non-dominated sorting genetic algorithm as a surrogate model to construct a two-layer optimization framework. The upper layer establishes a nonlinear mapping relationship between the combination of process parameters and four output objectives for the reinforcement random forest model. The lower layer is the non-dominated sorting genetic algorithm that performs multi-objective global optimization within the parameter constraint space. The optimization objectives are set as maximizing lithium leaching rate, maximizing gallium leaching rate, minimizing overall energy consumption, and minimizing reagent consumption. The algorithm initializes the population size to 100, and after 200 iterations, it outputs a Pareto optimal solution set after convergence. The optimal parameter combination is selected according to actual production needs and distributed to the execution equipment. Actual production data is fed back into the training set every 3 months to complete incremental training of the model.

[0034] Taking the processing of high-lithium coal gangue from Shanxi as an example, a strengthened random forest model was used to learn the mapping relationship between the 38-dimensional parameters of the raw material and the historical leaching rate. A non-dominated sorting genetic algorithm was used to optimize within constraints such as power density of 1-5 W / g and irradiation time of 5-30 min, outputting the optimal combination of microwave power density of 3 W / g and irradiation time of 15 min, predicting a lithium leaching rate of 87.3% with the lowest overall energy consumption. When switching to low-grade raw materials from Inner Mongolia, the model quickly adapted to the new data distribution through incremental training, re-optimizing the output parameter combination of power density of 4.5 W / g and ball milling time of 120 min, maintaining a leaching rate of over 85%, and shortening the debugging cycle from two weeks of traditional manual trial and error to two hours.

[0035] This embodiment improves prediction accuracy and generalization ability by enhancing the node splitting rules of the random forest model, and couples it with the non-dominated sorting genetic algorithm to achieve multi-objective global optimization. It eliminates reliance on human experience, reduces trial and error costs, shortens the debugging cycle, and adapts to the parameters of raw materials of different grades to quickly match the requirements, providing reliable intelligent decision support for industrial continuous production.

[0036] Optionally, a reinforced random forest model is constructed by integrating N reinforced decision trees, including: integrating orthogonal experimental data and pilot-scale verification data from the laboratory, valid experimental data of lithium gallium extraction processes from similar coal gangue in publicly available literature, and online data collected from industrial pilot production to construct a sample dataset; normalizing multiple features in the sample dataset, and using the three sigma criterion to remove outliers from the obtained normalized features to obtain multiple normalized features after filtering; dividing the multiple normalized features after filtering into a training dataset and a test dataset according to a preset ratio, wherein the number of features in the training dataset is greater than that in the test dataset; extracting N sets of feature samples from the training dataset using sampling with replacement, and constructing N reinforced decision trees from the N sets of feature samples respectively, wherein the N sets of feature samples and the N trees are integrated... The reinforced decision trees are matched one-to-one. Each set of feature samples includes input features and output features. The input features include combinations of process parameters, and the output features include target variables. The combinations of process parameters include: raw material characteristic parameters, microwave pretreatment parameters, mechanical activation parameters, and leaching process parameters. The target variables include: lithium leaching rate, gallium leaching rate, overall energy consumption, and reagent consumption per unit product. The reinforced random forest model is constructed based on the N reinforced decision trees. The prediction result of the reinforced random forest model for the target input sample is the arithmetic mean of the prediction results of the N reinforced decision trees for the target input sample. The reinforced random forest model is tested using a five-fold forward chain cross-validation method and the test dataset. If the reinforced random forest model passes the test, the training of the reinforced random forest model is considered complete.

[0037] Strengthening the construction of the random forest model begins with the integration of multi-source data. Laboratory orthogonal experimental data, pilot-scale verification data, valid experimental data from similar processes in publicly available literature, and online data collected during industrial-scale pilot production together constitute the sample dataset, with a total of no less than 1000 valid samples. The sample dataset contains 38-dimensional input features and 4-dimensional output features. The input features cover four main categories: raw material characteristic parameters, microwave pretreatment parameters, mechanical activation parameters, and leaching process parameters. The output features correspond to four core indicators: lithium leaching rate, gallium leaching rate, overall energy consumption, and reagent consumption per unit product.

[0038] The sample dataset enters the preprocessing stage. The min-max normalization method maps all features to the 0 to 1 interval to eliminate the influence of dimensions. The three sigma criterion identifies and removes outliers that deviate from the mean by more than three times the standard deviation. The normalized features after screening are divided into training and testing datasets in a 4:1 ratio. The training dataset accounts for 80% for model learning, and the testing dataset accounts for 20% for performance evaluation.

[0039] Model training employs bootstrap sampling with replacement to extract N sets of feature samples from the training dataset. Each set of samples is used to construct a reinforcement decision tree, with a one-to-one correspondence between the N sets of samples and the N trees. When constructing internal nodes of a single reinforcement decision tree, an embedded random forest submodel is first fitted to calculate feature importance. The feature with the highest importance is selected for splitting, and this process is recursively repeated until the number of samples in a node is less than 5 or the depth reaches 12. The output of the leaf nodes is the arithmetic mean of the target variable for the sample. The N reinforcement decision trees are integrated to form a reinforcement random forest model. When a new sample is input, the N trees predict independently, and the final output is the arithmetic mean.

[0040] Model validation employs a five-fold forward chain cross-validation method. The dataset is divided into five equal parts according to time sequence. In the first round, the first part is used for training and the second part for validation; in the second round, the first and second parts are used for training and the third part for validation, and so on until the fourth round, where the first four parts are used for training and the fifth part for validation. The validation set in each round always consists of subsequent data in the time sequence of the training set to avoid future information leakage. The test dataset is used for simultaneous validation. The model is considered to have passed the test when the coefficient of determination for predicting lithium gallium leaching rate reaches 0.87 or higher and the accuracy exceeds 92%.

[0041] Taking high-lithium coal gangue from Shanxi as an example, the sample dataset included 50 sets of laboratory pilot-scale data, 30 sets of pilot-scale verification data, and 20 sets of six-month industrial online data from the mining area. The model learned the nonlinear relationship between parameters such as power density and irradiation time and lithium leaching rate. Five-fold validation showed that the deviation between predicted and measured values ​​was controlled within 8%. When switching to low-grade raw materials from Inner Mongolia, 100 newly collected data sets were added to the sample dataset through incremental training. The model automatically adjusted the parameter weight distribution, and the validation accuracy remained above 91%, without the need for retraining from scratch.

[0042] This embodiment strengthens decision tree integration by fusing multi-source heterogeneous data and constructing a rigorous preprocessing sequence. Five-fold forward chain cross-validation ensures the reliability assessment of time-series process data. The model's prediction accuracy and generalization ability meet the needs of industrial applications, providing a high-precision surrogate model foundation for subsequent coupling with optimization algorithms.

[0043] Optionally, the method further includes: when constructing any internal node of each reinforcement decision tree, performing the following steps: fitting an embedded random forest sub-model using the dataset corresponding to the arbitrary internal node; and calculating the variable importance V(f) of all input features under the arbitrary internal node based on the out-of-bag data permutation method, using the following formula:

[0044] ;

[0045] in, The baseline mean squared error calculated based on out-of-bag data for the embedded random forest sub-model. The mean squared error is calculated by the embedded submodel after randomly permuting the out-of-bag data of feature f. The number of decision trees in the embedded random forest sub-model is given; the feature with the highest variable importance is selected as the optimal splitting feature for any internal node, and the splitting threshold is determined based on the minimum sum of the mean square errors of the two child nodes after splitting, thus completing the node partitioning.

[0046] The core innovation of the reinforced random forest model lies in the reconstruction of the splitting rules of nodes within a single reinforced decision tree. Traditional random forests randomly extract a subset of features when splitting nodes and select splitting features based on minimizing local mean square error. This approach is insufficient for fitting high-dimensional, strongly coupled metallurgical process data. This scheme introduces an embedded random forest sub-model to enhance the quality of feature selection when constructing any internal node of each reinforced decision tree.

[0047] The specific implementation consists of three steps. The first step is to fit an embedded random forest sub-model using the dataset corresponding to the internal node. The sub-model contains N decision trees, with N preferably being 20 in this technical solution. The second step is to calculate the variable importance V(f) of all input features under this node based on the out-of-bag data permutation method. The calculation formula is as follows: .

[0048] The calculation process is performed independently for each feature. The random permutation operation breaks the original correspondence between the feature and the target variable. The greater the increase in prediction error after the permutation, the more significant the contribution of the feature to the model prediction, and the value of the variable importance V(f) increases accordingly. In the third step, the feature with the highest variable importance is selected as the optimal splitting feature for this node. The optimal splitting threshold is determined based on the criterion of minimizing the sum of the mean square errors of the left and right child nodes after the split, thus completing the division of the internal node.

[0049] Taking lithium leaching rate prediction as an example, a dataset of an internal node contains 38 features including microwave power density, irradiation time, ball-to-material ratio, and ball milling speed. After fitting an embedded sub-model, the importance of each feature variable was calculated. Power density V(f) value was 0.35, irradiation time 0.28, ball-to-material ratio 0.15, and ball milling speed 0.12. Power density was selected as the optimal splitting feature, with a threshold of 3.2 W / g. The samples of this node were divided into high-power and low-power groups according to this threshold, and the variance of lithium leaching rate within the two groups was significantly reduced. Taking gallium leaching rate prediction as another example, calculations of another node showed that sodium fluoride addition V(f) value was 0.42, leaching temperature 0.31, and sulfuric acid concentration 0.18. Sodium fluoride addition became the optimal splitting feature, with a threshold of 2.8%. This splitting method effectively distinguished between high gallium extraction efficiency samples and low-efficiency samples.

[0050] This embodiment replaces random feature selection with embedded sub-models and out-of-bag data permutation, and establishes feature importance assessment on the quantitative basis of prediction error changes. Node splitting decisions are more targeted, solving the information loss problem caused by random feature selection in traditional random forests, and improving the model's fitting accuracy and generalization ability for high-dimensional nonlinear metallurgical data.

[0051] In an exemplary embodiment, a trained reinforced random forest model is used as a surrogate model and coupled with a non-dominated sorting genetic algorithm to construct a two-layer optimization framework, generating the intelligent optimization system. This includes: establishing a mapping relationship between the process parameter combination and the lithium leaching rate, the gallium leaching rate, the overall energy consumption, and the reagent consumption per unit product; using this mapping relationship as the fitness evaluation function of the non-dominated sorting genetic algorithm; setting multi-objective optimization objectives as: maximizing lithium leaching rate, maximizing gallium leaching rate, minimizing overall energy consumption, and minimizing reagent consumption per unit product; initializing the population and iteratively optimizing within the parameter constraint space using the non-dominated sorting genetic algorithm; outputting a Pareto optimal solution set after iterative convergence; wherein the parameter constraint space is a reasonable range of values ​​for multiple process parameters; and selecting the optimal process parameter combination from the Pareto optimal solution set according to actual production priority requirements, wherein the optimal process parameter combination includes: the optimal microwave parameter, the optimal ball milling parameter, and the optimal leaching parameter.

[0052] The generation of the intelligent optimization system relies on the deep coupling of a reinforced random forest model and a non-dominated sorting genetic algorithm. The trained reinforced random forest model serves as a surrogate model to establish a high-precision nonlinear mapping relationship between process parameter combinations and lithium leaching rate, gallium leaching rate, overall energy consumption, and reagent consumption per unit product. This mapping relationship replaces traditional numerical simulation or experimental verification as the fitness evaluation function of the non-dominated sorting genetic algorithm. New parameter combinations input into the surrogate model can output predicted values ​​of the four indicators in milliseconds, without the need for actual material feeding experiments.

[0053] The multi-objective optimization simultaneously sets the goals of maximizing lithium leaching rate, maximizing gallium leaching rate, minimizing overall energy consumption, and minimizing reagent consumption per unit product. These four objectives are inherently conflicting, as increasing leaching rate often accompanies increased energy and reagent consumption. A non-dominated sorting genetic algorithm performs optimization within a parameter constraint space, defined by the reasonable value ranges of parameters for each process: microwave pretreatment, mechanical activation, and tiered leaching. The algorithm initializes a population of 100 and iteratively evolves through selection, crossover, and mutation operations. After 200 iterations, it converges and outputs a Pareto optimal solution set, where each solution in the set cannot be simultaneously dominated by other solutions for a given objective. In actual production, the optimal combination of process parameters is selected from the Pareto optimal solution set based on priority requirements. For high-grade lithium raw materials, solutions with high leaching rate weights are prioritized; in energy-constrained scenarios, low-energy-consumption solutions are prioritized. The optimal combination of process parameters includes optimal microwave parameters, optimal ball milling parameters, and optimal leaching parameters, and is directly distributed to the execution equipment.

[0054] Taking the treatment of high-lithium coal gangue in Shanxi as an example, the Pareto optimal solution set contains 50 non-dominated solutions. Solution A has a lithium leaching rate of 88.2%, a gallium leaching rate of 82.5%, a comprehensive energy consumption of 125 kWh / t, and a reagent consumption of 280 kg / t. Solution B has a lithium leaching rate of 85.6%, a gallium leaching rate of 86.8%, a comprehensive energy consumption of 142 kWh / t, and a reagent consumption of 310 kg / t. Solution A is chosen when the production target is battery-grade lithium carbonate, and solution B is chosen when the target is metallic gallium. Taking low-grade raw materials in Inner Mongolia as another example, the Pareto front shifts overall towards lower leaching rates. Solution C, with a lithium leaching rate of 84.3% and a comprehensive energy consumption of 98 kWh / t, becomes the optimal choice, saving 23% energy and increasing the leaching rate by 6 percentage points compared to the manual trial-and-error approach.

[0055] This embodiment achieves rapid fitness calculation by replacing experimental evaluation with a proxy model. The non-dominated sorting genetic algorithm processes multi-objective conflicts and outputs a Pareto optimal solution set for flexible decision-making. The intelligent optimization system gets rid of dependence on human experience, and the time for a single optimization is reduced from several weeks to several minutes, adapting to the dynamic adjustment needs of different production goals and raw material conditions.

[0056] Optionally, the activated slag is subjected to two-stage differentiated leaching according to the optimal leaching parameters output by the intelligent optimization system to obtain a lithium-containing leaching solution and a gallium-containing leaching solution, respectively. This includes: mixing the activated slag with a first leaching agent at a liquid-solid ratio of 3:1-10:1, wherein the first leaching agent is a sulfuric acid solution or hydrochloric acid solution with a concentration of 0.5-2 mol / L; leaching is carried out under the conditions of a temperature of 25-60℃, a leaching time of 30-120 min, and a stirring speed of 200-400 r / min; after leaching, solid-liquid separation is performed to obtain the lithium-containing leaching solution and the lithium-extracted leaching slag.

[0057] The first stage of the two-stage differentiated leaching process is a selective lithium extraction step. The activated residue is mixed with the first leaching agent at a liquid-to-solid ratio of 3:1 to 10:1. The first leaching agent is a dilute sulfuric acid or hydrochloric acid solution with a concentration of 0.5-2 mol / L. This concentration range is lower than that of conventional acid leaching processes, specifically designed to match the degree of active lithium conversion after microwave-mechanical synergistic activation. The leaching process is controlled under mild conditions: a low-temperature range of 25-60℃, a leaching time of 30-120 min, and a stirring speed of 200-400 r / min. Selective dissolution is achieved by utilizing the difference in binding energy between lithium and gallium in the activated minerals, with lithium preferentially entering the solution while gallium remains in the solid phase. After leaching, solid-liquid separation yields a lithium-containing leachate and a lithium-extracted leaching residue. The lithium-containing leachate then enters a subsequent purification and concentration process to produce lithium carbonate or lithium hydroxide products.

[0058] Taking high-lithium coal gangue from Shanxi as an example, the intelligent optimization system outputs the optimal parameter combination of a liquid-to-solid ratio of 6:1, sulfuric acid concentration of 1.2 mol / L, temperature of 45℃, time of 60 min, and rotation speed of 300 r / min. The lithium leaching rate reaches 87.3%, while the gallium dissolution rate is only 8.5%. After solid-liquid separation, the lithium concentration in the lithium-containing solution is 3.2 g / L, and the gallium enrichment in the lithium extraction residue is 2.1 times higher than that of the original ore. Taking low-grade raw materials from Inner Mongolia as another example, the system is adjusted to a liquid-to-solid ratio of 8:1, hydrochloric acid concentration of 1.5 mol / L, temperature of 50℃, and time of 90 min. The lithium leaching rate is maintained at 85.2%, while the gallium dissolution rate is controlled at 9.8%. The low-temperature dilute acid conditions effectively suppress the large-scale dissolution of impurities such as aluminum and iron, and the impurity content of the leachate is reduced by more than 40% compared with single-stage strong acid leaching.

[0059] This embodiment achieves selective extraction of lithium and preliminary enrichment and separation of gallium through low-temperature dilute acid leaching. The lithium leaching rate in the first stage exceeds 85% while the gallium loss is less than 10%, which greatly reduces the difficulty of separation and purification and reagent consumption in subsequent gallium extraction. The two-stage differentiated design replaces the traditional one-stage mixed leaching process, reducing the load on the extraction and purification process and the amount of wastewater discharged.

[0060] Optionally, the activated slag is subjected to two-stage differentiated leaching according to the optimal leaching parameters output by the intelligent optimization system to obtain lithium-containing leaching solution and gallium-containing leaching solution, respectively. This includes: mixing the lithium-extracted leaching slag with a second leaching agent at a liquid-solid ratio of 4:1-12:1, wherein the second leaching agent is a sulfuric acid solution with a concentration of 2-4 mol / L and a sodium fluoride leaching aid with a mass fraction of 1%-5%; leaching is carried out at a temperature of 80-150℃, a leaching time of 60-180 min, and a stirring speed of 300-500 r / min; after leaching, solid-liquid separation is performed to obtain the gallium-containing leaching solution and leaching tailings.

[0061] The second stage of the two-stage differentiated leaching process is an enhanced gallium extraction step. After lithium extraction, the leaching residue is mixed with a second leaching agent at a liquid-to-solid ratio of 4:1 to 12:1. The second leaching agent is a 2-4 mol / L sulfuric acid solution with 1%-5% sodium fluoride as a leaching aid. Fluoride ions form a stable complex with gallium, breaking the aluminosilicate lattice's binding of gallium. The leaching process is controlled under enhanced conditions: a high temperature range of 80-150℃, a leaching time of 60-180 min, and a stirring speed of 300-500 r / min, to deeply extract gallium that was not dissolved in the first stage. After leaching, solid-liquid separation yields a gallium-containing leachate and leaching tailings. The gallium-containing leachate is then subjected to subsequent solvent extraction, back-extraction, and purification processes to prepare metallic gallium or gallium salt products. The leaching tailings meet general industrial solid waste standards and can be disposed of harmlessly or used in building materials.

[0062] Taking high-lithium coal gangue from Shanxi as an example, the intelligent optimization system outputs the optimal parameter combination of 8:1 liquid-to-solid ratio, 3 mol / L sulfuric acid concentration, 3% sodium fluoride addition, 120℃ temperature, 120 min time, and 400 r / min rotation speed, achieving a gallium leaching rate of 86.5%, which is 23 percentage points higher than conventional acid leaching without sodium fluoride. The gallium-containing solution concentration of 0.15 g / L meets the feed requirements of the extraction stage. Taking low-grade raw materials from Inner Mongolia as another example, the system was adjusted to a liquid-to-solid ratio of 10:1, 3.5 mol / L sulfuric acid concentration, 4.5% sodium fluoride addition, 135℃ temperature, and 150 min time, achieving a gallium leaching rate of 84.7%. The high temperature and high acid combined with the fluoride ion complexation effectively broke the tight symbiotic structure between gallium and quartz in this raw material.

[0063] This embodiment achieves deep gallium extraction by using high temperature and high acid conditions combined with sodium fluoride leaching aid. The gallium leaching rate in the second stage exceeds 85%, and the overall recovery rate of the two stages is more than 15% higher than that of the traditional single-stage mixed leaching process. The introduction of sodium fluoride clearly requires the matching titanium material or corrosion-resistant equipment to ensure the equipment safety and process stability of industrial implementation.

[0064] In an optional embodiment, this application proposes a microwave-mechanical synergistic activation method for the graded enrichment and extraction of lithium gallium from coal gangue based on a reinforced random forest surrogate model. On the one hand, the microwave-mechanical synergistic activation achieves deep destruction of the mineral lattice, significantly improving the lithium gallium leaching efficiency. On the other hand, the proposed reinforced random forest (RRF) model is used to construct a high-precision surrogate model, coupled with a multi-objective optimization algorithm, to achieve rapid and global intelligent optimization of the entire process parameters, taking into account extraction rate, energy consumption, and production cost, and is suitable for large-scale industrial applications.

[0065] To achieve the above objectives, the technical solution of this application is divided into two main parts: the entire process of microwave-mechanical synergistic activation-step leaching and intelligent optimization of process parameters based on reinforced random forest. The two are deeply coupled, and the key parameters for process execution are all output by the intelligent optimization system. The specific solution is as follows:

[0066] First, the entire process steps of the microwave-mechanical synergistic activation-step leaching process are as follows: Figure 2 As shown, it includes the following steps:

[0067] Step S1: Microwave-induced thermal stress pretreatment:

[0068] Crushed and screened coal gangue raw materials to below 200 mesh are placed in a microwave reactor and irradiated according to the optimal microwave parameters output by an intelligent optimization system. Utilizing the differences in dielectric constants and microwave absorption capabilities between carbonaceous and iron-bearing minerals and aluminosilicate minerals such as kaolinite and quartz in the coal gangue, a non-uniform temperature field is formed within the particles, inducing severe thermal expansion mismatch and directional microcracks. This preferentially disrupts the symbiotic interfaces between minerals, fully exposing lithium- and gallium-bearing mineral phases encased within stable mineral phases, thus constructing interconnected reaction channels for subsequent activation.

[0069] The microwave parameters are defined as follows: microwave frequency 2.45 GHz, power density 1~5 W / g, irradiation time 5~30 min, and reaction atmosphere is air or inert atmosphere. The specific parameters are optimized and output by the intelligent optimization system.

[0070] Step S2: Mechanochemical synergistic activation:

[0071] Microwave-pretreated coal gangue is fed into a planetary ball mill. Based on the optimal ball milling parameters output by the intelligent optimization system, a composite grinding aid is added for wet ball milling activation. The mechanical impact, shearing, and grinding effects during the ball milling process, using the microcracks pre-generated by the microwave as stress concentration weak points, induce lattice distortion, dislocation, defect proliferation, and even amorphization transformation in the mineral crystals. This deeply disrupts the stable lattice structure of aluminosilicate minerals, converting lattice-bound lithium and gallium elements into leached active states.

[0072] The ball milling parameters are as follows: ball-to-material ratio 5:1 to 20:1, ball milling speed 200 to 600 r / min, and ball milling time 30 to 180 min. The composite grinding aid is a mixture of triethanolamine and ethylene glycol in a mass ratio of 1:1 to 3, and the amount added is 0.5% to 3% of the mass of coal gangue. The specific parameters are optimized and output by the intelligent optimization system.

[0073] After ball milling, the slurry is subjected to solid-liquid separation to obtain activated slag, which is then sent to the subsequent cascade leaching process.

[0074] Step S3: Lithium gallium cascade leaching and separation:

[0075] The activated residue and the first leaching agent are mixed at a liquid-solid ratio of 3:1 to 10:1. The first leaching agent is a sulfuric acid or hydrochloric acid solution with a concentration of 0.5 to 2 mol / L. Leaching is carried out under optimized leaching conditions (temperature 25 to 60℃, leaching time 30 to 120 min, stirring speed 200 to 400 r / min). After leaching, solid and liquid are separated to obtain a lithium-containing leachate and a lithium-extracted leaching residue. The lithium-containing leachate is purified and concentrated to prepare lithium carbonate or lithium hydroxide products.

[0076] Step S32: Second stage of enhanced gallium extraction:

[0077] The lithium extraction residue is mixed with a second leaching agent at a liquid-solid ratio of 4:1 to 12:1. The second leaching agent is a 2-4 mol / L sulfuric acid solution with 1%-5% sodium fluoride as a leaching aid. Leaching is carried out under optimized leaching conditions (temperature 80-150℃, leaching time 60-180 min, stirring speed 300-500 r / min). It should be noted that when using a fluorine-containing leaching aid, the leaching equipment should preferably be made of titanium or lined with corrosion-resistant materials to cope with the corrosiveness of fluoride ions. After leaching, solid and liquid are separated to obtain gallium-containing leaching solution and leaching tailings. The gallium-containing leaching solution is then solvent-extracted, back-extracted, and purified to prepare metallic gallium or gallium salt products.

[0078] On the other hand, this application proposes a reinforced random forest high-precision surrogate model and constructs a coupled optimization framework with it and a multi-objective optimization algorithm. The RRF model accurately characterizes the nonlinear mapping relationship between process parameters and target indicators, and then global parameter optimization is achieved based on the multi-objective optimization algorithm. The construction process of this intelligent process parameter optimization system based on reinforced random forest is as follows: Figure 3 As shown, it includes the following steps:

[0079] Step S4: Construction and Principle Design of the RRF Proxy Model:

[0080] The RRF model is an improved ensemble learning regression model. Its core improvement lies in the splitting rules of the decision tree nodes. By embedding sub-models, it enhances the rationality of feature selection, improving the model's fitting accuracy and generalization ability to metallurgical process data. Specific design features include:

[0081] Step S41: Core Structure Definition:

[0082] The RRF model is composed of The ensemble consists of several reinforcement decision trees, and the final prediction result is the arithmetic mean of the predictions from all the reinforcement decision trees. This represents the total number of decision trees in the RRF model, with an optimal value of 70.

[0083] Step S42: "Strengthening" the splitting principle:

[0084] When constructing each internal node of each reinforced decision tree, the traditional random forest splitting method of "randomly extracting a subset of features + minimizing local mean square error" is abandoned. Instead, the following core logic is executed: first, an embedded random forest sub-model is fitted using the dataset of the node, and then the optimal splitting features are selected through feature importance evaluation. This strengthens the feature selection quality of a single decision tree and avoids information loss caused by random feature selection.

[0085] Step S5: Definition of the core calculation formula for the RRF model:

[0086] To ensure the reproducibility of the model, this embodiment provides explicit formula definitions for the core calculation process of the RRF model, and all symbols have been defined, specifically including:

[0087] Step S51: Marginal function and model strength calculation:

[0088] For any internal node of the decision tree, the fitted embedded random forest submodel is denoted as . ,in This represents the index of the decision tree in the sub-model. , The number of decision trees for the sub-model is 20, with an optimal value of 20.

[0089] (1) Marginal function : Used to characterize the prediction confidence of a sub-model for a single set of input data, defined as:

[0090] ;

[0091] In the formula, : An embedded random forest sub-model fitted within the nodes of a decision tree. The index of the decision tree in the embedded random forest sub-model, with a value range of [value missing]. . The number of decision trees contained in the embedded random forest sub-model is a preferred model structure parameter in this embodiment, with a preferred value of 20. Marginal function, used to characterize the prediction confidence of the embedded random forest sub-model for a single set of process parameter input data. The higher the value, the higher the confidence level of the model's prediction for that sample. : A single set of process parameter combinations is used as the input of independent variables for the model. The actual target value corresponding to a single set of process parameter input samples, specifically the lithium leaching rate or gallium leaching rate of coal gangue, etc. Embedded Random Forest Submodel For input samples The predicted output results. Embedded Random Forest Submodel Predicted value vs. actual target value The probability of a match. Predicted labels that are not the true values, i.e., those other than the true target values. Other prediction results besides those mentioned above. Model strength is used to measure the overall predictive power of the embedded random forest sub-model, and is the expected value of the marginal function of all input samples.

[0092] (2) Model strength s: used to measure the overall predictive power of the embedded sub-model, defined as the expected value of the marginal function of all input samples:

[0093] ;

[0094] In the formula, The total number of input samples for the current node within the decision tree. For the first Group input samples and their corresponding true target values. The first node inside the decision tree Input samples of process parameter combinations. The first node inside the decision tree The actual target value (lithium leaching rate or gallium leaching rate) corresponding to the input sample. : In the internal nodes of the decision tree, the first Predicted labels for non-true values ​​corresponding to the input samples.

[0095] Step S52: Calculate the importance of variables: For each input feature variable Define the importance of variables As the sole criterion for selecting node splitting features, the calculation process is based on the out-of-bag data permutation method, and the formula is:

[0096] ;

[0097] In the formula, : A single input feature variable in the lithium gallium extraction process from coal gangue. : Features After randomly permuting the out-of-bag data, the first... The mean squared error is calculated from an embedded random forest sub-model. : No. An embedded random forest sub-model The baseline mean square error is calculated based on the original out-of-bag data.

[0098] It should be noted that, The importance of variables is the sole basis for selecting the splitting features of decision tree nodes in this invention, and is used to characterize the degree of influence of process parameter features on the model prediction results. The larger the value, the stronger the influence of the feature on the model's prediction results, and the higher the feature's importance.

[0099] Step S6: Training process of RRF model.

[0100] This invention establishes a standardized RRF model training process, clarifying the dataset construction, training steps, and validation methods to ensure the feasibility and reliability of the model. Specifically, it includes:

[0101] Step S61: Training dataset construction:

[0102] (1) Data sources: integrate laboratory orthogonal experimental data, pilot-scale verification data, effective experimental data of similar coal gangue lithium gallium extraction processes in the literature, and online data collected during industrial pilot production. The total number of effective samples is not less than 1,000.

[0103] (2) Feature and label definition: The input features are 38-dimensional process and raw material parameters, which are divided into four categories: raw material characteristic parameters, microwave pretreatment parameters, mechanical activation parameters, and leaching process parameters; the output labels are four target variables, namely lithium leaching rate, gallium leaching rate, comprehensive energy consumption, and reagent consumption per unit product.

[0104] (3) Data preprocessing: The min-max normalization method is used to map all input features to the [0,1] interval to eliminate the influence of the unit; after removing outliers using the 3σ criterion, the data is randomly divided into training set and test set in a 4:1 ratio.

[0105] Step S62: Core steps of model training:

[0106] (1) Bootstrap sampling: from the training dataset In this case, sampling with replacement is used. For each set of bootstrap samples, a reinforcement decision tree is constructed. The preferred value is 70.

[0107] (2) Strengthen decision tree construction: targeting the first For each bootstrap sample, construct a reinforcement decision tree using the following sub-steps:

[0108] a. For each internal node of the tree, extract the corresponding dataset and fit an embedded random forest sub-model. Number of decision trees in sub-models The preferred value is 20.

[0109] b. Based on the out-of-bag data permutation method, calculate the variable importance of all input features at this node. .

[0110] c. Select the feature with the highest variable importance. As a splitting feature, the splitting threshold is determined by minimizing the sum of the mean square errors of the two child nodes after the split, thus completing the node division.

[0111] d. Repeat the above sub-steps until the number of node samples is less than the preset threshold (preferably 5) or the depth of the tree reaches the preset maximum depth (preferably 12), stop branch growth, set the node as a leaf node, and the output of the leaf node is the arithmetic mean of the target variable of the samples in the node.

[0112] (3) Model Integration: Model integration: complete all After constructing the reinforcement decision trees, the arithmetic mean of the prediction results of all trees for the same input sample is taken to obtain the final prediction output of the RRF model, as shown in the formula:

[0113] ;

[0114] In the formula, For the first The predicted output of a reinforced decision tree; This is the final prediction output of the RRF model.

[0115] Step S63: Model Validation and Optimization: This invention employs the Five-fold Forward Chained Cross Validation (FRF) method for model training and validation. The core principle of this validation method is as follows: the dataset is divided into five equal subsets according to time or batch order. One subset is selected sequentially as the validation set, and the remaining four subsets are used as the training set. The training set always includes samples from previous validation sets to avoid future data leakage. Simultaneously, after each validation iteration, the validation set is incorporated into the training set for the next round of validation, effectively adapting to the temporal characteristics of the process data and avoiding model overfitting. The coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) are used as model performance evaluation metrics. After training, the RRF model achieves an R² prediction of lithium and gallium leaching rates of no less than 0.87 and a prediction accuracy of no less than 92% on the test set.

[0116] Step S7: Coupled optimization process of RRF model and multi-objective optimization algorithm: This invention uses the trained RRF model as a surrogate model and couples it with the non-dominated sorting genetic algorithm (NSGA-II) to construct a two-layer optimization framework, realizing multi-objective global optimization of process parameters throughout the entire process. The specific steps are as follows:

[0117] Step S71: Proxy model loading:

[0118] Load the trained RRF model and establish a high-precision mapping relationship between "process parameter combination → lithium leaching rate, gallium leaching rate, comprehensive energy consumption, and reagent consumption", which will be used as the fitness evaluation function of the NSGA-II algorithm.

[0119] Step S72: Optimization objective and constraint setting:

[0120] The multi-objective optimization objectives are set as follows: ① Maximize lithium leaching rate; ② Maximize gallium leaching rate; ③ Minimize overall energy consumption; ④ Minimize reagent consumption per unit product. The constraints are reasonable value ranges for each process parameter, consistent with the parameter ranges in steps S1, S2, and S3.

[0121] Step S73: Global optimization using the NSGA-II algorithm:

[0122] A. Initialize the population: Randomly generate an initial population within the parameter constraint space, with the population size preferably being 100.

[0123] B. Iterative optimization: New parameter combinations are generated iteratively through selection, crossover, and mutation operations. Each parameter combination is input into the RRF model, which predicts the corresponding target index value as the basis for fitness evaluation.

[0124] C. Termination condition: Stop the search for optimization when the preset maximum number of iterations (preferably 200) is reached or the population converges, and output the Pareto optimal solution set.

[0125] Step S74; Application of optimal parameters and model iteration:

[0126] A. Parameter distribution: Based on the priority requirements of actual production, the optimal combination of process parameters is selected from the Pareto optimal solution set and distributed to the process execution equipment to guide actual production.

[0127] B. Incremental Model Training: The actual production process operation data and detection results are added to the model dataset. The RRF model is incrementally trained every 3 months to continuously improve the model's raw material adaptability and prediction accuracy.

[0128] Step S8: Comparative Experiment to Verify the Invention's Effectiveness: To verify the fitting accuracy, generalization ability, and stability of the Reinforced Random Forest (RRF) model proposed in this invention in predicting the mapping relationship between process parameters and target indicators in lithium gallium extraction from coal gangue, five representative algorithms—covering single-tree models, ensemble learning, deep learning, and continuous action space reinforcement learning—were selected as baseline comparison models for a comparative experiment. The experimental design, process, and results are as follows:

[0129] Step S82: Basic experimental conditions:

[0130] A. Hardware and software environment: CPU is Intel Xeon Gold 6330, memory is 128GB, GPU is NVIDIA RTXA4000; the running environment is Python 3.8, and the machine learning frameworks used are Scikit-learn 1.0.2 and PyTorch 1.12.0.

[0131] B. Experimental Dataset: The complete process dataset constructed using step S61 of this invention has 1260 effective samples. The input features are 38-dimensional raw materials and process parameters. The output prediction targets are four core indicators: lithium leaching rate, gallium leaching rate, comprehensive energy consumption, and reagent consumption per unit product. This experiment focuses on comparing and verifying the prediction effects of lithium and gallium leaching rates.

[0132] C. Data preprocessing: All models use uniform min-max normalization to map input features to the [0,1] interval to eliminate the influence of dimensions; after removing outliers using the 3σ criterion, the training set and test set are fixedly divided in a 4:1 ratio.

[0133] D. Validation method: All models adopt a unified five-fold forward chained cross validation to avoid leakage of future information in time series process data and ensure the reliability of validation results.

[0134] Step S83: Baseline Contrast Model and Parameter Settings:

[0135] This experiment selected five mainstream algorithms as baselines. The complexity and basic parameters of all models were on the same order of magnitude as the RRF model of this invention, with only the core algorithms differing. The specific settings are as follows:

[0136] A. Decision Tree (DT): A basic single-tree model, serving as the baseline for tree models and used to validate the necessity of ensemble learning strategies. Parameter settings: maximum depth 12, minimum number of split samples 5, minimum number of leaf node samples 2, split criterion uses squared error, completely consistent with the basic parameters of a single tree in the RRF model.

[0137] B. Random Forest (RF): A traditional Bagging ensemble tree model, used as a direct baseline for the RRF model in this invention to verify the performance improvement effect of the improved node splitting rules. Parameter settings: 70 decision trees, maximum depth of a single tree 12, minimum number of split samples 5, minimum number of leaf node samples 2, the number of features randomly selected during node splitting is the square root of the total number of features, and the splitting criterion uses the squared error; the ensemble size and basic tree parameters are completely consistent with the RRF model, only the node splitting rules are different.

[0138] C. Gradient Boosting Decision Tree (GBDT): A mainstream boosting ensemble tree model used to compare the predictive performance of different ensemble strategies. Parameter settings: 70 base learners, maximum tree depth 10, learning rate 0.1, loss function is squared error, minimum number of split samples 5, minimum number of leaf node samples 2, ensuring that the model complexity is on the same order of magnitude as the RRF model.

[0139] D. Artificial Neural Network (ANN): A fully connected deep learning foundational model used to compare the fitting capabilities of tree models and connectionist models on high-dimensional metallurgical process data. Parameter settings: 38 neurons in the input layer (corresponding to 38-dimensional input features), 20 neurons in a single hidden layer, and 4 neurons in the output layer (corresponding to 4 prediction targets); the hidden layer activation function is ReLU, and the output layer uses a linear activation function; the optimizer is Stochastic Gradient Descent (SGD), with a learning rate of 0.1, 200 iterations, and a batch size of 32.

[0140] E. Deep Deterministic Policy Gradient (DDPG): A representative model of deep reinforcement learning in continuous action space, used to verify the accuracy advantage of the RRF as a supervised regression model in the task of predicting process parameters. Parameter settings: experience replay pool size 10000, batch size 1024, Actor network learning rate 0.001, Critic network learning rate 0.002, discount factor γ=0.99, soft update coefficient τ=0.005, target network update frequency 10 steps, iteration rounds 200, backbone network structure consistent with the ANN model.

[0141] The parameters of the RRF model in this embodiment are consistent with those in steps S62 and S63: total number of decision trees 70, maximum depth of a single tree 12, minimum number of split samples 5, number of embedded sub-model decision trees 20, and node splitting rule is the optimal selection of variable importance calculated based on the embedded sub-model.

[0142] Step S84: Definition of evaluation indicators:

[0143] This experiment uses three industry-standard regression model evaluation metrics, and supplements them with a prediction accuracy metric, defined as follows:

[0144] A. Coefficient of Determination (R²): Characterizes the model's ability to explain data variation. Its value ranges from -∞ to 1. The closer the value is to 1, the stronger the model's fitting accuracy and generalization ability.

[0145] B. Root Mean Squared Error (RMSE): Characterizes the overall deviation between the predicted value and the true value. The smaller the value, the smaller the prediction error of the model.

[0146] C. Mean Absolute Error (MAE): Characterizes the average absolute deviation between the predicted value and the actual value. The smaller the value, the better the prediction stability of the model.

[0147] Step S85: Experimental Results and Data Analysis:

[0148] All results in this experiment are the mean of five-fold forward chain cross-validation. The core performance comparisons are shown in Table 1 and Table 2:

[0149] Table 1 Comparison of the predictive performance of each model for lithium leaching rate

[0150]

[0151] Table 2 Comparison of the prediction performance of each model for gallium leaching rate

[0152]

[0153] Analysis of experimental results:

[0154] A. The RRF model proposed in this invention outperforms all five baseline models in all evaluation metrics of both the training and test sets in the heavy task of predicting lithium and gallium leaching rates. Specifically, the R² for lithium leaching rate prediction on the test set reaches 0.87, with a prediction accuracy of 92.20%; the R² for gallium leaching rate prediction on the test set reaches 0.88, with a prediction accuracy of 91.80%, fully meeting the accuracy requirements of the surrogate model for process parameter optimization.

[0155] B. The training set performance of each model is slightly better than that of the test set, which is consistent with the basic principles of machine learning. Among them, the RRF model of this invention has the smallest performance gap between the training set and the test set (the difference in lithium leaching rate R² is only 0.02, and the difference in MAE is only 0.01), indicating that it has the strongest generalization ability and effectively avoids the overfitting problem.

[0156] C. Compared with the traditional RF model, the RRF model of this invention improves the lithium leaching rate prediction R² by 33.8% and the prediction accuracy by 7.95% on the test set. This verifies the "node splitting rule based on the importance of embedded sub-model variables" proposed in this invention, which can significantly improve the fitting accuracy and generalization ability of the tree model and solve the information loss problem caused by the random feature selection of traditional RF.

[0157] D. Compared with the single-tree DT model, the RRF model significantly reduces the prediction error and significantly improves R², verifying that the ensemble learning strategy can effectively solve the overfitting problem of the single-tree model and improve the model stability.

[0158] E. Compared with deep learning models ANN and continuous action space reinforcement learning models DDPG, RRF models have better prediction accuracy and generalization ability on the test set, while also having stronger interpretability and lower parameter tuning costs. They have more significant application advantages in small-sample, high-dimensional, and strongly coupled data in metallurgical process scenarios.

[0159] This embodiment utilizes microwave-induced thermal stress to form a directional microcrack network within coal gangue particles, providing a stress concentration weak surface for subsequent mechanical ball milling. The mechanical impact and shearing action of ball milling achieves deep lattice distortion and amorphization transformation along the microcracks, completely breaking the lattice constraints of lithium and gallium. This synergistic activation method reduces ball milling energy consumption by more than 30% while achieving deep activation of the minerals. Combined with a two-stage step-leaching process, the overall leaching rate of lithium can reach over 88%, and the overall leaching rate of gallium can reach over 85%, far exceeding the extraction level of the traditional roasting-acid leaching process.

[0160] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0161] Embodiments of this application also provide a microwave-mechanical synergistic activation system for the graded enrichment and extraction of lithium gallium from coal gangue, such as... Figure 4 As shown, the system includes:

[0162] The microwave irradiation module 42 is used to place crushed coal gangue raw materials, screened to below 200 mesh, into a microwave reactor, and perform irradiation treatment according to the optimal microwave parameters output by the intelligent optimization system. The difference in dielectric constant and microwave absorption capacity of different minerals in the coal gangue raw materials is used to form a non-uniform temperature field inside the particles, induce directional microcracks, and obtain coal gangue material.

[0163] The ball mill activation module 44 is used to feed the coal gangue material into a planetary ball mill, add composite grinding aids to perform wet ball mill activation according to the optimal ball mill parameters output by the intelligent optimization system, and use the microcracks as the stress concentration weak surface to induce lattice distortion, dislocation, defect proliferation and amorphization transformation of mineral crystals. After ball milling, solid-liquid separation is performed to obtain activated slag.

[0164] The leaching module 46 is used to perform two-stage differentiated leaching of the activated residue according to the optimal leaching parameters output by the intelligent optimization system, so as to obtain lithium-containing leaching solution and gallium-containing leaching solution respectively.

[0165] In the aforementioned system, coal gangue raw material is crushed and screened to below 200 mesh before entering a microwave reactor. The intelligent optimization system outputs optimal microwave parameters to guide irradiation treatment. The carbonaceous and iron-containing minerals in the coal gangue differ from aluminosilicate minerals such as kaolinite and quartz. These differences in microwave absorption capacity lead to a non-uniform temperature field within the particles, and the thermal expansion mismatch induces directional microcracks. The lithium-gallium mineral phase is exposed from the stable lattice, resulting in pretreated coal gangue material. This pretreated coal gangue material is then fed into a planetary ball mill. The intelligent optimization system outputs optimal ball milling parameters, and a composite grinding aid is added for wet ball milling activation. The microcracks formed during microwave pretreatment become stress concentration weak points. Mechanical impact and shear forces propagate along these microcracks, causing lattice distortion, dislocation multiplication, defect proliferation, and even amorphization transformation of the mineral crystals. The stable aluminosilicate lattice structure is deeply destroyed, and the lattice-bound lithium-gallium elements are transformed into a leaching active state. After ball milling, the slurry undergoes solid-liquid separation to obtain activated slag. The activated slag then enters a two-stage differentiated leaching process. The intelligent optimization system outputs the optimal leaching parameters to achieve selective separation of lithium and gallium, resulting in lithium-containing leachate and gallium-containing leachate. This application achieves deep destruction of the mineral lattice and active transformation of elements through microwave-mechanical synergistic activation. The intelligent optimization system outputs the optimal process parameters for the entire process. The two-stage leaching achieves selective dissolution and preliminary separation based on the difference in lithium and gallium binding energies, with the comprehensive lithium and gallium leaching rates both exceeding 85%. This solves the problem of the lack of an integrated process that takes into account deep mineral activation, intelligent parameter optimization, and efficient separation of lithium and gallium in related technologies, which seriously limits the large-scale and high-value recovery of lithium and gallium rare metals in coal gangue.

[0166] Optionally, the microwave irradiation module 42 is further used to integrate N reinforcement decision trees to form a reinforcement random forest model, where N is a positive integer; the trained reinforcement random forest model is used as a surrogate model and coupled with a non-dominated sorting genetic algorithm to construct a two-layer optimization framework to generate the intelligent optimization system.

[0167] Optionally, the aforementioned microwave irradiation module 42 is further used to integrate orthogonal experimental data and pilot-scale verification data from the laboratory, effective experimental data of lithium gallium extraction processes from similar coal gangue in publicly available literature, and online data collected during industrial pilot production to construct a sample dataset; normalize multiple features in the sample dataset, and use the three sigma criterion to remove outliers from the obtained normalized features to obtain multiple normalized features after filtering; divide the multiple normalized features after filtering into a training dataset and a test dataset according to a preset ratio, wherein the number of features in the training dataset is greater than that in the test dataset; extract N sets of feature samples from the training dataset using sampling with replacement, and construct N reinforcement decision trees using the N sets of feature samples respectively, wherein the N sets of feature samples and the N reinforcement decision trees are used to construct N reinforcement decision trees. Each set of feature samples is configured in a one-to-one correspondence, containing input features and output features. The input features include combinations of process parameters, and the output features include target variables. The combinations of process parameters include: raw material characteristic parameters, microwave pretreatment parameters, mechanical activation parameters, and leaching process parameters. The target variables include: lithium leaching rate, gallium leaching rate, overall energy consumption, and reagent consumption per unit product. The reinforced random forest model is constructed based on the N reinforced decision trees, wherein the prediction result of the reinforced random forest model for the target input sample is the arithmetic mean of the prediction results of the N reinforced decision trees for the target input sample. The reinforced random forest model is tested using a five-fold forward chain cross-validation method and the test dataset. If the reinforced random forest model passes the test, the training of the reinforced random forest model is considered complete.

[0168] Optionally, the microwave irradiation module 42 described above is further configured to perform the following steps when constructing any internal node of each reinforcement decision tree: fitting an embedded random forest sub-model using the dataset corresponding to the arbitrary internal node; and calculating the variable importance V(f) of all input features under the arbitrary internal node based on the out-of-bag data permutation method, using the following formula: Among them, MSE n PMSE is the baseline mean squared error calculated based on out-of-bag data for the embedded random forest sub-model. n (f) represents the mean squared error calculated by the embedded submodel after randomly permuting the out-of-bag data of feature f. The number of decision trees in the embedded random forest sub-model is given; the feature with the highest variable importance is selected as the optimal splitting feature for any internal node, and the splitting threshold is determined based on the minimum sum of the mean square errors of the two child nodes after splitting, thus completing the node partitioning.

[0169] Optionally, the microwave irradiation module 42 is further configured to establish a mapping relationship between the process parameter combination and the lithium leaching rate, the gallium leaching rate, the overall energy consumption, and the reagent consumption per unit product, and use the mapping relationship as the fitness evaluation function of the non-dominated sorting genetic algorithm; the multi-objective optimization objectives are set as: maximizing lithium leaching rate, maximizing gallium leaching rate, minimizing overall energy consumption, and minimizing reagent consumption per unit product; the non-dominated sorting genetic algorithm initializes the population and iteratively optimizes within the parameter constraint space, and outputs the Pareto optimal solution set after iterative convergence, wherein the parameter constraint space is the reasonable value range of multiple process parameters; the optimal process parameter combination is selected from the Pareto optimal solution set according to the actual production priority requirements, wherein the optimal process parameter combination includes: the optimal microwave parameter, the optimal ball milling parameter, and the optimal leaching parameter.

[0170] Optionally, the leaching module 46 is further configured to mix the activated residue with a first leaching agent at a liquid-solid ratio of 3:1 to 10:1, wherein the first leaching agent is a sulfuric acid solution or hydrochloric acid solution with a concentration of 0.5-2 mol / L; leaching is carried out at a temperature of 25-60℃, a leaching time of 30-120 min, and a stirring speed of 200-400 r / min; after leaching, solid-liquid separation is performed to obtain the lithium-containing leaching solution and the lithium-extracted leaching residue.

[0171] Optionally, the leaching module 46 is further configured to mix the lithium-extracted leaching residue with a second leaching agent at a liquid-solid ratio of 4:1 to 12:1, wherein the second leaching agent is a sulfuric acid solution with a concentration of 2-4 mol / L and a sodium fluoride leaching aid with a mass fraction of 1%-5%; leaching is carried out at a temperature of 80-150℃, a leaching time of 60-180 min, and a stirring speed of 300-500 r / min; after leaching, solid-liquid separation is performed to obtain the gallium-containing leaching solution and leaching tail residue.

[0172] Embodiments of this application also provide a storage medium including a stored program, wherein the program, when executed, performs any of the methods described above. Optionally, in this embodiment, the storage medium may be configured to store program code for performing the steps of the methods described above.

[0173] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0174] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0175] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0176] Embodiments of this application also provide a computer program product, including a non-volatile computer-readable storage medium storing the computer program product, wherein the computer program, when executed by a processor, implements the steps of the methods described in various embodiments of this application.

[0177] Optionally, in this embodiment, the computer program described above can be configured to implement the steps in any of the above method embodiments when executed by a processor.

[0178] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0179] The scope of this application is defined by the appended claims rather than the foregoing description, and is therefore intended to encompass all variations falling within the meaning and scope of equivalents of the claims. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a device claim may also be implemented by a single unit or device in software or hardware. Terms such as "first," "second," etc., are used only for distinguishing descriptions and do not indicate any particular order, nor should they be construed as indicating or implying relative importance.

[0180] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily made by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims, and the above embodiments should be regarded as exemplary and non-limiting.

Claims

1. A method for the graded enrichment and extraction of lithium and gallium from coal gangue through microwave-mechanical synergistic activation, characterized in that, include: Crushed coal gangue raw material, screened to below 200 mesh, is placed in a microwave reactor and irradiated according to the optimal microwave parameters output by the intelligent optimization system. The difference in dielectric constant and microwave absorption capacity of different minerals in the coal gangue raw material is used to form a non-uniform temperature field inside the particles, inducing directional microcracks to obtain coal gangue material. The coal gangue material is fed into a planetary ball mill. According to the optimal ball milling parameters output by the intelligent optimization system, a composite grinding aid is added for wet ball milling activation. The microcracks are used as the stress concentration weak surface to induce lattice distortion, dislocation, defect proliferation and amorphization transformation of mineral crystals. After ball milling, solid-liquid separation is performed to obtain activated slag. The activated residue is subjected to two-stage differentiated leaching according to the optimal leaching parameters output by the intelligent optimization system to obtain lithium-containing leachate and gallium-containing leachate respectively.

2. The method according to claim 1, characterized in that, Before performing irradiation treatment based on the optimal microwave parameters output by the intelligent optimization system, the method further includes: A reinforced random forest model is constructed by ensembles of N reinforced decision trees, where N is a positive integer; The trained reinforcement random forest model is used as a surrogate model and coupled with the non-dominated sorting genetic algorithm to construct a two-layer optimization framework, thereby generating the intelligent optimization system.

3. The method according to claim 2, characterized in that, A reinforced random forest model is constructed by ensembles of N reinforced decision trees, including: A sample dataset was constructed by integrating orthogonal experimental data and pilot-scale verification data from the laboratory, effective experimental data of lithium gallium extraction processes from similar coal gangue in publicly available literature, and online data collected from industrial pilot production. Multiple features in the sample dataset are normalized, and outliers are removed from the normalized features using the three sigma criterion to obtain the filtered normalized features. The filtered normalized features are divided into a training dataset and a test dataset according to a preset ratio, wherein the number of features in the training dataset is greater than that in the test dataset. N sets of feature samples are extracted from the training dataset using sampling with replacement, and N reinforcement decision trees are constructed from the N sets of feature samples. Each set of feature samples corresponds one-to-one with the N reinforcement decision trees. Each set of feature samples contains input features and output features. The input features include combinations of process parameters, and the output features include target variables. The combinations of process parameters include: raw material characteristic parameters, microwave pretreatment parameters, mechanical activation parameters, and leaching process parameters. The target variables include: lithium leaching rate, gallium leaching rate, overall energy consumption, and reagent consumption per unit product. The reinforced random forest model is constructed based on the N reinforced decision trees, wherein the prediction result of the reinforced random forest model for the target input sample is the arithmetic mean of the prediction results of the N reinforced decision trees for the target input sample. The reinforced random forest model is tested using a five-fold forward chain cross-validation method and the test dataset. If the reinforced random forest model passes the test, the training of the reinforced random forest model is considered complete.

4. The method according to claim 3, characterized in that, The method further includes: When constructing any internal node of each reinforcement decision tree, perform the following steps: An embedded random forest sub-model is fitted using the dataset corresponding to any of the internal nodes; Based on the out-of-bag data permutation method, the variable importance V(f) of all input features under any given internal node is calculated using the following formula: Among them, MSE n PMSE is the baseline mean squared error calculated based on out-of-bag data for the embedded random forest sub-model. n (f) represents the mean squared error calculated by the embedded submodel after randomly permuting the out-of-bag data of feature f. The number of decision trees in the embedded random forest sub-model; The feature with the highest variable importance is selected as the optimal splitting feature for any internal node. The splitting threshold is determined based on minimizing the sum of the mean square errors of the two child nodes after splitting, thus completing the node partitioning.

5. The method according to claim 3, characterized in that, The trained reinforcement random forest model is used as a surrogate model and coupled with a non-dominated sorting genetic algorithm to construct a two-layer optimization framework, generating the intelligent optimization system, including: Establish a mapping relationship between the process parameter combination and the lithium leaching rate, the gallium leaching rate, the comprehensive energy consumption, and the reagent consumption per unit product, and use the mapping relationship as the fitness evaluation function of the non-dominated sorting genetic algorithm; The multi-objective optimization objectives are set as follows: maximizing lithium leaching rate, maximizing gallium leaching rate, minimizing overall energy consumption, and minimizing reagent consumption per unit product. The non-dominated sorting genetic algorithm is used to initialize the population within the parameter constraint space, iterate and optimize, and output the Pareto optimal solution set after iterative convergence. The parameter constraint space is the reasonable value range of multiple process parameters. The optimal combination of process parameters is selected from the Pareto optimal solution set based on actual production priority requirements. The optimal combination of process parameters includes: the optimal microwave parameters, the optimal ball milling parameters, and the optimal leaching parameters.

6. The method according to claim 1, characterized in that, The activated slag is subjected to a two-stage differentiated leaching process based on the optimal leaching parameters output by the intelligent optimization system to obtain a lithium-containing leaching solution and a gallium-containing leaching solution, respectively, including: The activated residue is mixed with the first leaching agent at a liquid-solid ratio of 3:1 to 10:1, wherein the first leaching agent is a sulfuric acid solution or hydrochloric acid solution with a concentration of 0.5-2 mol / L. The leaching was carried out at a temperature of 25-60℃, a leaching time of 30-120 min, and a stirring speed of 200-400 r / min. After leaching, solid-liquid separation is performed to obtain the lithium-containing leachate and the lithium-extracted leaching residue.

7. The method according to claim 6, characterized in that, The activated slag is subjected to a two-stage differentiated leaching process based on the optimal leaching parameters output by the intelligent optimization system to obtain a lithium-containing leaching solution and a gallium-containing leaching solution, respectively, including: The lithium extraction residue is mixed with a second leaching agent at a liquid-solid ratio of 4:1-12:1, wherein the second leaching agent is a sulfuric acid solution with a concentration of 2-4 mol / L and a sodium fluoride leaching aid with a mass fraction of 1%-5%. The leaching was carried out at a temperature of 80-150℃, a leaching time of 60-180 min, and a stirring speed of 300-500 r / min. After leaching, solid-liquid separation is performed to obtain the gallium-containing leaching solution and leaching tailings.

8. A microwave-mechanical synergistic activation system for the graded enrichment and extraction of lithium and gallium from coal gangue, characterized in that, include: The microwave irradiation module is used to place crushed coal gangue raw materials, screened to below 200 mesh, into a microwave reactor and perform irradiation treatment according to the optimal microwave parameters output by the intelligent optimization system. The difference in dielectric constant and microwave absorption capacity of different minerals in the coal gangue raw materials is used to form a non-uniform temperature field inside the particles, inducing directional microcracks to obtain coal gangue material. The ball mill activation module is used to feed the coal gangue material into a planetary ball mill, add composite grinding aids according to the optimal ball milling parameters output by the intelligent optimization system, and perform wet ball mill activation. The microcracks are used as the stress concentration weak surface to induce lattice distortion, dislocation, defect proliferation and amorphization transformation of mineral crystals. After ball milling, solid-liquid separation is performed to obtain activated slag. The leaching module is used to perform two-stage differentiated leaching of the activated residue according to the optimal leaching parameters output by the intelligent optimization system, to obtain lithium-containing leaching solution and gallium-containing leaching solution respectively.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program, when executed, performs the method of any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images