Stepwise flotation method for separating low-grade fluorite ore

Abstract

The present application relates to the technical field of flotation, and discloses a step-by-step flotation method for separating low-grade fluorite ore. The method comprises the following steps: performing multi-dimensional mineralogical quantitative characterization on the ore to construct a six-dimensional characteristic input vector; inputting the vector into a pre-trained deep neural network model to predict the flotation index; using the model as a fitness function and adopting a genetic algorithm to globally optimize the dosages of collectors and depressants and the pH value of the ore slurry; and finally performing a step-by-step flotation process of one roughing, two scavenging and three cleaning according to the optimal parameters. Through digital twinning and intelligent optimization technology, the present application realizes rapid, self-adaptive and efficient control of the flotation process, improves the fluorite recovery rate and concentrate grade, and reduces the experimental cost and dependence on artificial experience.

Description

Technical Field

[0001] This invention belongs to the field of flotation technology, specifically relating to a stepwise flotation method for separating low-grade fluorite ore. Background Technology

[0002] Fluorite, as an important non-metallic mineral resource, is widely used in metallurgy, chemical industry, building materials, and new energy materials. With the increasing depletion of high-grade fluorite resources, the efficient separation and comprehensive utilization of low-grade fluorite ore has become a key direction for industry development. Flotation is currently the core process for fluorite beneficiation, achieving effective separation of valuable minerals from gangue by adjusting parameters such as reagent formulation, pulp concentration, and aeration intensity.

[0003] However, low-grade fluorite ore typically has fine intercalation, complex symbiotic relationships, and high impurity content, making it difficult for traditional flotation processes to balance recovery rate and concentrate grade. This results in poor process stability and a strong dependence on operational experience.

[0004] Stepwise flotation, as a strategy to improve selectivity, uses multiple stages of roughing, scavenging, or cleaning operations to selectively suppress or activate specific minerals at different stages, thereby optimizing the overall separation effect. While this method theoretically has the potential to improve performance, its practical application heavily relies on extensive physical experiments for process debugging and parameter optimization. Each adjustment to the flotation sequence, reagent type, or dosage requires small-scale or even pilot-scale testing, which not only consumes significant time and manpower but also makes it difficult to systematically assess the coupling effects between various variables.

[0005] Especially in production scenarios where ore properties fluctuate frequently, fixed processes lack dynamic adaptability, resulting in large fluctuations in sorting efficiency and serious waste of resources.

[0006] In existing technologies, flotation process modeling is mostly focused on empirical formula fitting or static equilibrium calculation of a single stage, lacking an integrated description of multi-physics coupling mechanisms such as fluid flow, bubble-particle collision, surface chemical reaction and mineral dissociation behavior.

[0007] Therefore, it is impossible to realistically reproduce the dynamic evolution of stepwise flotation in a virtual environment, and it is even more difficult to predict the recovery rate and impurity residue levels under different operating strategies. This limitation has kept process optimization stuck in an inefficient cycle of "trial and error - feedback - more trial and error," severely restricting the economical and efficient development of low-grade fluorite ore. There is an urgent need for an intelligent optimization method that can integrate multi-scale mechanistic models and support scenario extrapolation and performance prediction to break through the bottleneck of traditional flotation process development. Summary of the Invention

[0008] The purpose of this invention is to provide a stepwise flotation method for separating low-grade fluorite ore, aiming to overcome the technical problems in the prior art where the optimization of flotation process relies too much on physical experiments, resulting in high costs, long cycles, and insufficient adaptability and predictive ability to changes in ore properties.

[0009] To achieve the above objectives, the present invention provides a stepwise flotation method for separating low-grade fluorite ore. The core of this method lies in constructing a digital twin prediction model based on ore depth characteristics and flotation process parameters, and solving this model using a global optimization algorithm to obtain the optimal combination of process parameters for specific ore properties, thereby guiding the precise execution of the physical flotation process. The method includes the following steps:

[0010] First, the selected low-grade fluorite ore is quantitatively characterized by multi-dimensional mineralogical features to generate a set of accurate ore feature input vectors.

[0011] This step is not a simple chemical composition analysis, but a deep and structured analysis of the physicochemical properties of the ore.

[0012] Specifically, X-ray diffractometer was used to perform full-spectrum scanning analysis on the ore sample. The Bragg diffraction equation was used to compare with the standard mineral diffraction database to quantitatively determine the mass fraction of key mineral phases such as fluorite, calcite, quartz and barite in the ore.

[0013] Using a laser particle size analyzer, the particle size distribution curve of the ore slurry after grinding under standard ball milling conditions was measured by Mie scattering theory, and the median particle size D50 value was calculated.

[0014] An automated mineral analysis system was used to perform surface scanning of the prepared ore sections using a combination of scanning electron microscopy and energy dispersive spectroscopy. This allowed for the identification and delineation of different mineral particles, the calculation of the degree of liberation between fluorite and gangue mineral particles, and the degree of liberation of individual fluorite minerals at different particle sizes.

[0015] The obtained fluorite mass fraction, calcite mass fraction, quartz mass fraction, barite mass fraction, median grain size, and fluorite monomer dissociation degree are combined into a six-dimensional ore feature input vector, which is used as the input for the subsequent prediction model.

[0016] Next, the ore feature input vector is input into a pre-constructed and trained multi-objective prediction deep neural network model for the flotation process.

[0017] The deep neural network model is a feedforward multilayer perceptron used to characterize the complex nonlinear mapping relationship between flotation process parameters and flotation results.

[0018] The network structure of the model was determined to include an input layer, three hidden layers, and an output layer.

[0019] The input layer has 6 neurons, corresponding to the 6 dimensions of the mineral feature input vector; the first hidden layer contains 128 neurons, the second hidden layer contains 64 neurons, and the third hidden layer contains 32 neurons. The activation function of all hidden layers is the modified linear unit function.

[0020] The output layer contains three neurons, corresponding to the three key flotation indicators of predicted fluorite concentrate grade, fluorite recovery rate, and tailings grade, respectively, with linear activation functions. This deep neural network model was trained using supervised learning on a database containing thousands of historical physical flotation test data sets.

[0021] Each set of historical data includes a complete ore feature vector, the flotation process parameters used, and the actual measured flotation results.

[0022] The training process employs an adaptive moment estimation optimizer, using the mean square error between the predicted output and the actual measured output as the loss function. The weights and biases of all neurons in the network are adjusted through the backpropagation algorithm until the model converges to the preset accuracy threshold.

[0023] Then, using the multi-objective prediction deep neural network model of the flotation process as the core fitness function, a genetic algorithm is applied to perform a global optimization search within the preset process parameter optimization space to determine an optimal combination of flotation process parameters.

[0024] The optimization process of the genetic algorithm is as follows: the key adjustable process parameters in the flotation process, namely the amount of sodium oleate collector per ton of ore, the amount of water glass inhibitor per ton of ore, and the pH value of the pulp, are encoded as real chromosomes.

[0025] Set the search range for various parameters: sodium oleate dosage range is 800g to 1500g per ton of ore, water glass dosage range is 800g to 2000g per ton of ore, and slurry pH value range is 0.8 to 10.5.

[0026] Initialize a population of 100 individuals, each representing a randomly generated combination of process parameters. Define the overall fitness function, whose mathematical expression is:

[0027] ;

[0028] G represents the fluorite concentrate grade predicted by the deep neural network model. For the predicted fluorite recovery rate, and These are preset weighting coefficients used to balance the relative importance of grade and recovery rate in production targets; their sum is 1. Penalty item. Used to ensure that the concentrate grade meets industrial requirements, it is defined as: when When it is less than 97%, Equal to a large positive value, such as 100; otherwise It equals 0.

[0029] The genetic algorithm iterates through selection, crossover, and mutation operations. The selection operation uses the roulette wheel selection method, the crossover operation uses the single-point crossover method with a crossover probability set to 0.8, and the mutation operation uses the Gaussian mutation method with a mutation probability set to 0.05.

[0030] After 200 iterations or 50 consecutive generations when the fitness value no longer increases, the algorithm terminates and outputs the chromosome corresponding to the individual with the highest fitness value in the population. After decoding, this chromosome is the determined optimal combination of flotation process parameters, including the optimal sodium oleate dosage, the optimal water glass dosage, and the optimal pulp pH value.

[0031] Finally, based on the optimal combination of SEN process parameters determined by the genetic algorithm, a step-by-step flotation operation is performed in the physical flotation equipment.

[0032] The step-by-step flotation operation includes a first-stage roughing, a two-stage sweeping, and a three-stage cleaning.

[0033] In the primary roughing operation, the slurry, which has been ground and has been adjusted to a mass concentration of 30%, is pumped into the roughing flotation tank group. Reagents are added according to the determined optimal dosage of sodium oleate and water glass using a high-precision metering pump. At the same time, the pH value of the slurry is precisely controlled to the optimal value using an online pH meter and an automatic dosing system.

[0034] The roughing flotation time is 5 minutes, and the frothy product produced is rough concentrate, while the product in the tank is roughing tailings.

[0035] The roughing tailings are pumped into the first-stage scavenging flotation cell group and supplemented with sodium oleate equivalent to 20% of the roughing amount for the first-stage scavenging. The tailings are then fed into the second-stage scavenging flotation cell group and supplemented with sodium oleate equivalent to 10% of the roughing amount for the second-stage scavenging.

[0036] The foam products from the two-stage scavenging process are combined and returned to the grinding and classification system.

[0037] The tailings from the second stage of scavenging are discarded as final tailings.

[0038] The rough concentrate produced by the first-stage roughing process is pumped into the first-stage cleaning flotation tank group, and water is added to dilute the slurry mass concentration to 15%. No reagents are added, and the first-stage cleaning is carried out.

[0039] The first-level selected foam products enter the second-level selected products, the second-level selected foam products enter the third-level selected products, and the third-level selected foam products are the final fluorite concentrate.

[0040] The tailings from each stage of the flotation process are sequentially returned to the previous stage. This model-based prediction and global optimization approach enables precise, efficient, and adaptive control of the low-grade fluorite flotation process.

[0041] As one embodiment of the present invention, in the step of generating the ore feature input vector, the automated mineral analysis system sets its electron beam acceleration voltage to 20 kV, beam current to 3 nanoamps, and working distance to 13 mm when performing surface scanning, so as to ensure sufficient excitation energy and spatial resolution for fine-grained mineral particles.

[0042] As one embodiment of the present invention, the specific implementation of the crossover and mutation operations of the genetic algorithm during iterative optimization is as follows:

[0043] For the crossover operation, a crossover point on the chromosome is randomly selected, and all gene segments after the crossover point of the two parent individuals are exchanged to generate two offspring individuals. For the mutation operation, for the selected individual gene, a random perturbation amount that follows a normal distribution with a mean of 0 and a standard deviation of 1 / 10 of any of the three flotation process parameters optimized by the genetic algorithm is superimposed on its current value to generate a new gene value after mutation.

[0044] In one embodiment of the present invention, the reagent used to adjust the pH value of the pulp in the stepwise flotation operation is a sodium carbonate solution or a dilute sulfuric acid solution, which is added precisely by a peristaltic pump controlled in a closed loop with an online pH meter to ensure that the fluctuation range of the pH value is controlled within ±0.1.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] 1. This invention transforms the time-consuming and labor-intensive physical testing process into an efficient computer numerical simulation by constructing a high-precision deep neural network prediction model between ore characteristics and flotation results. This enables rapid optimization of flotation process parameters, shortens the process development cycle, and reduces related testing costs.

[0047] 2. The method proposed in this invention has the ability to adapt to changes in ore properties. When the mineralogical characteristics of the feed ore fluctuate, it is only necessary to re-characterize the ore characteristics and run the optimization algorithm to obtain the optimal process parameters for the current ore properties within a few minutes. This realizes the forward-looking and proactive control of the flotation process and improves the stability and robustness of the production process.

[0048] 3. This invention uses a genetic algorithm for global optimization search. Compared with traditional manual parameter tuning or orthogonal experimental methods, it can avoid getting trapped in local optima in a broad multidimensional parameter space and find the combination of process parameters that makes the comprehensive flotation index reach the global optimum. Thus, while ensuring the grade of concentrate, it maximizes the recovery rate of fluorite and improves the utilization efficiency of mineral resources.

[0049] 4. This invention solidifies complex flotation process knowledge into a calculable and iterative digital model, providing core technical support for the intelligent and digital transformation of flotation plants. By combining expert experience with data-driven models, it achieves scientific and precise control of the flotation process, reduces reliance on operator experience, and ensures consistency and high-level production operations. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the overall technical solution architecture of the stepwise flotation method for separating low-grade fluorite ore proposed in this invention;

[0051] Figure 2 This is a schematic diagram of the core principle framework of the digital twin prediction model based on ore depth characteristics and flotation process parameters in this invention.

[0052] Figure 3 This is a logical flowchart of the quantitative characterization of multi-dimensional mineralogical features and the construction of ore feature input vectors in this invention.

[0053] Figure 4 This is a diagram illustrating the structure and training logic framework of the multi-objective prediction deep neural network model for the flotation process in this invention.

[0054] Figure 5 This is a flowchart of the global optimization process of the genetic algorithm using a deep neural network as the fitness function in this invention.

[0055] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal flotation device and the digital twin system in this invention. Detailed Implementation

[0056] Please refer to Figures 1 to 6 This invention provides a stepwise flotation method for separating low-grade fluorite ore. Its core lies in constructing a digital twin prediction model based on ore depth characteristics and flotation process parameters, and combining it with a global optimization algorithm to solve for the optimal combination of process parameters, thereby guiding the precise execution of the physical flotation process.

[0057] This method completely eliminates the reliance on extensive physical experiments in traditional flotation process development, shortens the process debugging cycle, reduces testing costs, and possesses adaptive capabilities to fluctuations in ore properties. The specific implementation methods of this invention will be described in detail below, following the S-step sequence clearly defined in the invention description.

[0058] The method first performs step S1: quantitatively characterizing the selected low-grade fluorite ore using multi-dimensional mineralogical features to generate a set of accurate ore feature input vectors.

[0059] This step forms the data foundation of the entire method, and its accuracy directly determines the reliability of subsequent prediction models. In practice, representative raw ore samples are taken and prepared into standard analytical samples after crushing and fractionation.

[0060] The sample was sent into an X-ray diffractometer for full-spectrum scanning analysis. The scanning angle range was 5 degrees to 90 degrees, the step size was 0.02 degrees, and the counting time for each step was 1 second.

[0061] The obtained diffraction patterns were compared with the standard mineral diffraction database provided by the International Diffraction Data Center using the Bragg diffraction equation. The mass fractions of four key mineral phases in the ore—fluorite, calcite, quartz, and barite—were quantitatively determined using the Rietveld full-spectrum fitting refinement method, with the measurement uncertainty controlled within ±0.5%.

[0062] Meanwhile, another parallel sample was placed in a standard laboratory ball mill and ground at a fixed speed and grinding time to prepare a slurry with a mass concentration of 25%.

[0063] The slurry was injected into a laser particle size analyzer. The instrument uses the Mie scattering theory model and sets the refractive index and absorption rate as standard optical parameters for each mineral. The particle size distribution curve of the slurry was measured, and the median particle size D50 value was calculated, which is the particle size corresponding to a cumulative volume percentage of 20%.

[0064] In addition, another sample was taken, and after being inlaid with epoxy resin and polished, it was made into a smooth sheet and placed in an automated mineral analysis system.

[0065] The system integrates a field emission scanning electron microscope and an energy dispersive spectrometer, sets the electron beam acceleration voltage to 20 kV, the beam current to 3 nanoamps, and the working distance to 13 mm, to perform high-resolution surface scanning of the light sheet surface.

[0066] The system's built-in mineral identification software automatically identifies and delineates different mineral particles based on energy spectrum characteristics. It uses image processing algorithms to calculate the overall degree of dissociation between fluorite mineral particles and gangue mineral particles, as well as the degree of individual dissociation of fluorite minerals in three key particle sizes: less than 74 micrometers, 74 to 150 micrometers, and greater than 150 micrometers.

[0067] Finally, six of the most representative parameters—fluorite mass fraction, calcite mass fraction, quartz mass fraction, barite mass fraction, median grain size D50 value, and fluorite monomer liberation degree—were selected and combined in a fixed order to form a six-dimensional ore feature input vector, which served as the sole input data source for the subsequent prediction model.

[0068] After completing step S1, step S2 is executed: the ore feature input vector is input into the pre-constructed and trained multi-objective prediction deep neural network model for the flotation process.

[0069] This deep neural network model is a feedforward multilayer perceptron, and its structure has been determined to be optimal after extensive experimental verification.

[0070] The model's input layer contains six neurons, each receiving dimensional data from the ore feature input vector.

[0071] The first hidden layer contains 128 neurons, the second hidden layer contains 64 neurons, and the third hidden layer contains 32 neurons.

[0072] All hidden layers use the modified linear unit function as the activation function, which is defined as outputting the same value as the input when the input is greater than 0, and outputting 0 otherwise. Its mathematical expression is:

[0073] ;

[0074] The output layer contains three neurons, which correspond to the three key flotation indicators of predicted fluorite concentrate grade, fluorite recovery rate and tailings grade, respectively. The activation function is a linear function, that is, the output is equal to the weighted sum of the inputs plus the bias term.

[0075] The training data for this model comes from a dedicated database containing 3,500 sets of historical physical flotation test data.

[0076] Each set of data records the complete six-dimensional ore feature vector, the three flotation process parameters used (sodium oleate collector per ton of ore, water glass depressant per ton of ore, and pulp pH value), and the three flotation result indicators actually measured through chemical analysis.

[0077] The training process uses an adaptive moment estimation optimizer with an initial learning rate of 0.001. The loss function is defined as the weighted sum of the mean square errors between the predicted output and the actual measured output, and the weights are allocated according to the importance of each indicator.

[0078] The backpropagation algorithm is used to iteratively adjust the weights of all connections and the biases of neurons in the network until the loss function value on the validation set no longer decreases after 20 consecutive iterations and is less than the preset threshold of 0.005. At this point, the model is considered to have converged, and the final network parameters are saved.

[0079] In practical applications, the ore feature input vector generated in step S1 is directly fed into this trained model. After four layers of nonlinear transformation, the model outputs a three-dimensional prediction vector, which includes a prediction of the flotation performance that can be achieved for the current ore properties under any given process parameters.

[0080] Then, step S3 is executed: using the multi-objective prediction deep neural network model of the flotation process as the core fitness function, a genetic algorithm is applied to perform a global optimization search within a preset process parameter optimization space to determine an optimal combination of flotation process parameters. This step is crucial for achieving intelligent optimization of process parameters.

[0081] First, the three key adjustable process parameters in the flotation process are encoded into a real number chromosome. This chromosome consists of three gene loci, which represent the amount of sodium oleate collector per ton of ore, the amount of water glass inhibitor per ton of ore, and the pH value of the pulp, respectively.

[0082] Set the physical feasible search boundaries for each parameter: sodium oleate dosage ranges from 800g to 1500g per ton of ore, water glass dosage ranges from 800g to 2000g per ton of ore, and slurry pH value ranges from 0.8 to 10.5.

[0083] An initial population of 100 individuals is initialized, with each individual's chromosome generated by uniform random sampling within its respective parameter range.

[0084] Define the comprehensive fitness function Its mathematical expression is:

[0085] ;

[0086] G represents the fluorite concentrate grade predicted by the deep neural network model. For the predicted fluorite recovery rate, and These are preset weighting coefficients used to balance the relative importance of grade and recovery rate in production targets, and their sum is 1.

[0087] Penalty items Used to ensure that the concentrate grade meets industrial requirements, it is defined as: when When it is less than 97%, Equal to a large positive value, such as 100; otherwise It equals 0.

[0088] The evolutionary process of the genetic algorithm strictly follows preset rules:

[0089] The selection operation uses the roulette wheel selection method, which means that the probability of each individual being selected is proportional to the proportion of its fitness value to the total fitness of the population.

[0090] The crossover operation uses a single-point crossover method. Two parent individuals are randomly selected, and a crossover point is randomly chosen on the chromosome. All gene segments after that point are exchanged to generate two new offspring individuals. The crossover probability is fixed at 0.8.

[0091] The mutation operation uses the Gaussian mutation method. Each individual in the population is selected for mutation with a probability of 0.05. For each selected individual, each gene locus on its chromosome is subjected to a random perturbation with an independent probability, which follows a normal distribution with a mean of 0 and a standard deviation of 1 / 10 of any of the three flotation process parameters optimized by the genetic algorithm.

[0092] For example, for the gene on sodium oleate dosage, its range is 1000 grams, so the standard deviation is 100 grams.

[0093] The algorithm begins to iterate. In each iteration, for each individual in the new generation population, its chromosome is decoded into a specific combination of process parameters, which is then input into the deep neural network model in step S2 to obtain the predicted G and R values, and then its fitness value F is calculated.

[0094] The algorithm is considered to have converged and the iteration is terminated if it continues to run for 200 iterations, or if the improvement of the highest fitness value in the population is less than 0.001 for 50 consecutive generations.

[0095] The final output is the chromosome corresponding to the individual with the highest fitness value. After decoding, the optimal sodium oleate dosage, optimal water glass dosage, and optimal pulp pH value are obtained, forming a complete optimal combination of flotation process parameters.

[0096] Finally, step S4 is executed: based on the optimal combination of flotation process parameters determined by the genetic algorithm, a step-by-step flotation operation is performed in the physical flotation equipment.

[0097] This step involves translating the optimization results from the digital world into production practices in the physical world.

[0098] The specific operating procedure is as follows:

[0099] The raw ore slurry, after being ground using a standard ball milling process, is precisely adjusted to a mass concentration of 30% using a concentration regulating device before being pumped into the primary roughing flotation cell group. Before the slurry enters the flotation cell, two high-precision metering pumps continuously and stably inject the reagent solution into the slurry pipeline according to the optimal sodium oleate and water glass dosages determined in step S3, ensuring thorough mixing of the reagents and the slurry.

[0100] Meanwhile, an online pH meter is installed at the inlet of the roughing flotation cell. Its real-time measurement value is fed back to the central control system. The control system drives the peristaltic pump to accurately titrate sodium carbonate solution or dilute sulfuric acid solution into the slurry, dynamically regulating and stabilizing the pH value of the slurry at the optimal value determined in step S3. Its control accuracy ensures that the fluctuation range is within ±0.1.

[0101] The residence time for roughing flotation is 5 minutes, during which mechanical stirring is used to ensure that the air bubbles come into full contact with the target mineral.

[0102] The frothy product produced by the roughing operation is the rough concentrate, which is scraped off and sent to the cleaning system; the product that does not float in the tank is the roughing tailings, which is pumped into the first-stage scavenging flotation tank group.

[0103] In the first-stage scavenging operation, a collector equivalent to 20% of the optimal amount of sodium oleate used in the roughing stage is added to recover the residual fluorite in the roughing tailings.

[0104] The tailings from the first-stage scavenging are then pumped into the second-stage scavenging flotation tanks, supplemented with collector equivalent to 10% of the optimal amount of sodium oleate used in the roughing stage, for further recovery. The froth products from the two-stage scavenging operations are combined and returned to the front end of the grinding and classification system for regrinding and re-selection.

[0105] The tailings from the second-stage scavenging process, due to their extremely low fluorite content, are transported to the tailings dam as the final tailings.

[0106] The rough concentrate produced by the first-stage roughing process is pumped into the first-stage cleaning flotation tank group. Before this, water is added to dilute the pulp mass concentration to 15% in order to reduce impurity entrainment.

[0107] No chemicals are added during any of the sorting processes; purification is achieved solely through physical separation.

[0108] The froth product from the first stage of refining enters the second stage of refining flotation cells, and the froth product from the second stage of refining enters the third stage of refining flotation cells. The froth product produced in the third stage of refining is the final fluorite concentrate, the grade of which is predicted by the model in step S2 and guaranteed by the optimization in step S3.

[0109] The tailings generated from each level of beneficiation operation, i.e. the product in the tank, are returned sequentially to the feed end of the previous level operation to form a closed loop, so as to maximize the recovery of useful minerals.

[0110] Throughout the execution of the method, the data flow and processing logic are highly coordinated. The ore feature input vector generated in step S1 is the only external input, and all subsequent steps revolve around this vector.

[0111] The deep neural network model in step S2 acts as a black box mapper, and its internal weights and biases are fixed after training, ensuring the consistency and repeatability of predictions.

[0112] The genetic algorithm in step S3 uses the S2 model as its fitness evaluator to perform an efficient and global search in a large process parameter space, and its convergence is guaranteed by strict termination conditions.

[0113] Step S4 is a faithful execution of the optimal parameter combination output by step S3. Its operational details, such as the amount of reagent added, concentration control, and process structure, are clearly defined, ensuring a seamless connection from digital instructions to physical operations.

[0114] This method achieves a fundamental shift in flotation processes from experience-driven to data- and model-driven by abstracting complex physicochemical processes into computable mathematical models and solving them using intelligent optimization algorithms.

[0115] To ensure the robustness and reliability of the entire method, each step is designed with corresponding verification and exception handling mechanisms.

[0116] In step S1, if the goodness-of-fit factor Rwp of the X-ray diffraction analysis is greater than the preset threshold of 10, the analysis of the sample is deemed to have failed and the sample needs to be prepared and analyzed again.

[0117] In step S2, if any dimension of the input ore feature vector exceeds the statistical range of historical training data, the model will output a warning signal, prompting the operator to verify the ore's origin. In step S3, if the genetic algorithm fails to find a solution that satisfies a concentrate grade greater than 97% after the maximum number of iterations, it is determined that the current ore has extremely poor beneficiation properties, and it is recommended to abandon processing or adopt other beneficiation methods.

[0118] In step S4, if the online pH meter malfunctions, the system will automatically switch to the backup manual control mode and issue an alarm.

[0119] These mechanisms ensure the stable operation of the method under various working conditions.

[0120] In summary, the stepwise flotation method for separating low-grade fluorite ore described in this invention constructs a complete "sensing-prediction-optimization-execution" closed loop through the close connection and coordinated operation of the four steps S1 to S4. This method not only solves the problems of high development costs and long cycles of traditional processes, but also endows the flotation process with unprecedented adaptability and intelligence, providing a brand-new technical path for the efficient, green, and economical separation of complex and difficult-to-process ores.

Claims

1. A stepwise flotation method for separating low-grade fluorite ore, characterized in that, include: Quantitative characterization of multi-dimensional mineralogical features of the selected low-grade fluorite ore is performed to generate a six-dimensional ore feature input vector. The six-dimensional ore feature input vector is input into a pre-constructed and trained multi-objective prediction deep neural network model for the flotation process; Using the deep neural network model as the fitness function, a genetic algorithm is applied to perform a global optimization search within the preset process parameter optimization space to determine the optimal combination of flotation process parameters. The process parameters include the amount of sodium oleate collector per ton of ore, the amount of water glass inhibitor per ton of ore, and the pH value of the pulp. The genetic algorithm iteratively evolves through roulette wheel selection, single-point crossover, and Gaussian mutation operations, and evaluates the individual performance based on the comprehensive fitness function. According to the optimal combination of flotation process parameters, a step-by-step flotation operation is performed in the physical flotation equipment. The step-by-step flotation operation includes a first-stage roughing, a second-stage scavenging, and a third-stage cleaning. In the roughing operation, sodium oleate and water glass are added according to the optimal reagent dosage and the pH of the pulp is adjusted to the optimal value. In the scavenging operation, collectors are added in stages. In the cleaning operation, no reagents are added and the tailings from each stage are returned to the previous stage in sequence.

2. The stepwise flotation method for separating low-grade fluorite ore according to claim 1, characterized in that, The multi-objective prediction deep neural network model for the flotation process is a feedforward multilayer perceptron containing an input layer, three hidden layers, and an output layer. The input layer has 6 neurons, the first hidden layer contains 128 neurons, the second hidden layer contains 64 neurons, and the third hidden layer contains 32 neurons. All hidden layers use a modified linear unit function as the activation function. The output layer contains three neurons and uses a linear function as the activation function, outputting the predicted fluorite concentrate grade, fluorite recovery rate, and tailings grade.

3. The stepwise flotation method for separating low-grade fluorite ore according to claim 1, characterized in that, The quantitative characterization of the multi-dimensional mineralogical characteristics of the selected low-grade fluorite ore to generate a six-dimensional ore feature input vector includes: X-ray diffractometer was used to perform full-spectrum scanning analysis on ore samples. By comparing the Bragg diffraction equation with the standard mineral diffraction database and combining the Rietveld full-spectrum fitting and refinement method, the mass fraction of fluorite, calcite, quartz and barite was quantitatively analyzed. The slurry prepared under standard ball milling regime was measured using a laser particle size analyzer. The particle size distribution curve was obtained based on Mie scattering theory and the median particle size D50 value was calculated. An automated mineral analysis system was used to scan the surface of the ore section. Mineral particles were identified by combining scanning electron microscopy with energy dispersive spectroscopy. The overall degree of liberation of fluorite and gangue minerals and the degree of liberation of individual fluorite particles at different particle sizes were calculated. The mass fractions of fluorite, calcite, quartz, barite, median grain size (D50), and fluorite monomer dissociation are combined to form a six-dimensional ore feature input vector.

4. The stepwise flotation method for separating low-grade fluorite ore according to claim 3, characterized in that, When the automated mineral analysis system performs surface scanning, the electron beam accelerating voltage is set to 20 kV, the beam current is 3 nanoamps, and the working distance is 13 mm.

5. The stepwise flotation method for separating low-grade fluorite ore according to claim 1, characterized in that, The multi-objective prediction deep neural network model for the flotation process is trained under supervision based on a database containing 3,500 sets of historical physical flotation test data. Each set of data contains a complete six-dimensional ore feature vector, three flotation process parameters, and measured flotation result indicators. The training adopts an adaptive moment estimation optimizer, with the weighted mean square error between the predicted output and the measured output as the loss function. The network weights and biases are adjusted through the backpropagation algorithm until the validation set loss no longer decreases for 20 consecutive rounds and is less than the threshold of 0.

005.

6. The stepwise flotation method for separating low-grade fluorite ore according to claim 1, characterized in that, The initialization of the genetic algorithm includes a population of 100 individuals. The search range for sodium oleate usage per ton of ore is 800g to 1500g, the search range for water glass usage per ton of ore is 800g to 2000g, the search range for slurry pH value is 0.8 to 10.5, the crossover probability is 0.8, the mutation probability is 0.05, the maximum number of iterations is 200, or the algorithm terminates when the highest fitness value increases by less than 0.001 for 50 consecutive generations.

7. The stepwise flotation method for separating low-grade fluorite ore according to claim 6, characterized in that, The crossover operation of the genetic algorithm involves randomly selecting a crossover point on a chromosome and exchanging all gene segments after that crossover point between two parent individuals to generate two offspring individuals.

8. The stepwise flotation method for separating low-grade fluorite ore according to claim 7, characterized in that, The mutation operation of the genetic algorithm includes superimposing random perturbations on each gene locus of the selected individual, which follow a normal distribution with a mean of 0 and a standard deviation of 1 / 10 of any of the three flotation process parameters optimized by the genetic algorithm.

9. The stepwise flotation method for separating low-grade fluorite ore according to claim 1, characterized in that, In the stepwise flotation operation, the pulp concentration of the roughing operation is 30%, and the flotation time is 5 minutes; the first-stage scavenging is supplemented with collector equivalent to 20% of the amount of sodium oleate used in the roughing operation, and the second-stage scavenging is supplemented with collector equivalent to 10% of the amount of sodium oleate used in the roughing operation; the pulp concentration of the third-stage cleaning operation is diluted to 15%, and the tailings from each stage of cleaning are returned to the previous stage in sequence.

10. The stepwise flotation method for separating low-grade fluorite ore according to claim 1, characterized in that, In the stepwise flotation operation, the reagent used to adjust the pH value of the pulp is a sodium carbonate solution or a dilute sulfuric acid solution. It is added precisely by a peristaltic pump controlled in a closed loop with an online pH meter to ensure that the pH value fluctuation range is controlled within ±0.1.

Images