A chalcopyrite recovery prediction method

Abstract

The present application relates to a chalcopyrite recovery prediction method. The method comprises generating a multi-dimensional feature vector using the operating parameters of chalcopyrite flotation, and constructing the phase transition characteristics of the mineral interface using a continuous activation function; estimating the theoretical recovery rate according to the type of collector used, the type and concentration of ions in the water environment, and the pH value; mapping the theoretical recovery rate to a Logit anchor point in the unbounded real number space through a logarithmic probability Logit function; using a table prior fitting network TabPFN based on the Transformer architecture, taking the multi-dimensional feature vector and the phase transition characteristics as the input of the network TabPFN, and taking the Logit anchor point as the condition to predict the nonlinear residual in the unbounded real number space; generating a controlled residual by asymmetrically constraining the nonlinear residual; and obtaining a corrected recovery rate by superimposing the controlled residual and the theoretical recovery rate. The corrected recovery rate obtained by the method is more accurate than the recovery rates predicted by the two types of models currently available, thus solving the problems of the prior art.

Description

Technical Field

[0001] This invention relates to the field of mineral processing technology, and in particular to a method for predicting the recovery rate of chalcopyrite. Background Technology

[0002] Copper, comprising over 80% of all minerals, is found in chalcopyrite, and its enrichment and recovery are currently primarily achieved through froth flotation. The core of froth flotation lies in two key steps: pulp preparation and reagent addition. Specifically, the crushed and ground ore is first placed in an aqueous solution to form a pulp. Then, a pH adjuster is added to regulate the pulp's pH to the alkaline range (typically 8.0–11.0). Subsequently, a collector (such as xanthate ionic reagents or thiocyanate nonionic reagents) is introduced into the pulp. Utilizing the collector's selective adsorption on the mineral surface, the hydrophobicity of the chalcopyrite surface is altered, causing it to adhere to air bubbles and float to the surface for separation.

[0003] In recent years, with the increasing scarcity of global freshwater resources and the rising standards for green mine construction, utilizing unconventional water sources such as seawater and high-salinity circulating water to replace freshwater for flotation slurry preparation has become an inevitable trend in the mineral processing industry. However, these unconventional water sources are rich in high concentrations of unavoidable ions such as magnesium and calcium, and their complex chemical composition severely interferes with the physicochemical processes at the mineral surface. Particularly in the reagent dosing stage, different types of collectors exhibit significant differences in their "physical shield" performance under high-salinity environments. For example, ionic collectors are susceptible to the double-layer compression effect, leading to adsorption attenuation on the mineral surface, while non-ionic collectors, although providing steric hindrance protection, are more expensive and their dosage is sensitive. Therefore, accurately predicting the recovery rate changes of different reagent regimes under high-salinity environments and dynamically adjusting the dosing strategy accordingly is a key control challenge for the industrial-scale implementation of flotation using seawater and other unconventional water sources.

[0004] Currently, the industry mainly relies on two types of models to predict the recovery rate of chalcopyrite flotation. The first is the traditional empirical kinetic model, which is based on the reaction rate formula under ideal conditions. Its basic assumption is that the slurry environment is homogeneous and stable. However, in high-salt systems such as seawater, inevitable ions such as magnesium and calcium will cause complex precipitation, covering, and competitive adsorption, making the prediction results of this empirical kinetic model prone to deviate from reality. The second is the pure data-driven machine learning model. However, existing machine learning models predict the recovery rate based on the regularity summary of historical data, and the principle and accuracy of their prediction are also questionable. Summary of the Invention

[0005] Therefore, it is necessary to provide a method for predicting chalcopyrite recovery rate to address the problem of low accuracy in predicting chalcopyrite recovery rate in existing technologies.

[0006] To address the above problems, this invention provides a method for predicting chalcopyrite recovery rate, comprising: Multidimensional feature vectors are generated using the operating parameters of chalcopyrite flotation, and phase transition features of mineral interfaces are constructed using continuous activation functions. The theoretical recovery rate is estimated based on the type of collector used, the type and concentration of ions in the water environment, and the pH value. The theoretical recovery rate is mapped to Logit anchor points in an unbounded real space using the logit function; Using the TabPFN, a table prior fitting network based on the Transformer architecture, the multidimensional feature vector and the phase transition feature are used as inputs to the TabPFN network, and the nonlinear residuals are predicted in the unbounded real space with the Logit anchor points as conditions. Controlled residuals are generated by applying asymmetric constraints to the nonlinear residuals. The controlled residual is superimposed with the theoretical recovery rate to obtain the corrected recovery rate.

[0007] In one possible implementation, the method further includes: performing an inverse transformation mapping on the corrected recovery rate to obtain the final predicted recovery rate.

[0008] In one possible implementation, the method further includes: The final predicted recovery rate is compared with the set cliff warning threshold. If the final predicted recovery rate is lower than the cliff warning threshold, a reagent compensation strategy for extreme conditions is generated and executed, wherein the reagent compensation strategy includes any one or more of the following strategies: switching to the use of a non-ionic collector; increasing the dosage of the collector.

[0009] In one possible implementation, controlled residuals are generated by applying asymmetric constraints to the nonlinear residuals, specifically including: Determine the numerical range to which the theoretical recovery rate belongs; Obtain the attenuation coefficient reset coefficient corresponding to the numerical region; Based on the attenuation coefficient reset coefficient, using the formula ε final = ε raw × λ generates controlled residuals, where ε final Let ε be the controlled residual. raw Let λ be the nonlinear residual, and λ be the attenuation coefficient reset coefficient.

[0010] In one possible implementation, the method further includes: The recovery rate was pre-divided into three numerical regions: less than 50%, greater than or equal to 50% and less than or equal to 85%, and greater than 85%. Assign a first attenuation coefficient and reset coefficient equal to 1 to the numerical region that is less than 50%; For the numerical range of 50 or greater and 85% or less, assign a second attenuation coefficient reset coefficient less than 1; For the numerical range greater than 85%, assign a third attenuation coefficient reset factor that is smaller than the second attenuation coefficient reset factor.

[0011] In one possible implementation, a continuous activation function is used to construct the phase transition characteristics of the mineral interface. Specifically, this includes using the following formula to construct the iron ion oxidation phase transition characteristics of the mineral interface: ; Wherein: pH is the pH value of the chalcopyrite flotation pulp; Feature Fe This is a characteristic phase transition of iron ion oxidation; This represents the nonlinear inflection point of iron ion oxidation in the slurry.

[0012] In one possible implementation, a continuous activation function is used to construct the phase transition characteristics of the mineral interface. Specifically, this includes using the following formula to construct the precipitation-covered phase transition characteristics of the mineral interface: ; Among them: Feature Mg The precipitation-covered phase transition characteristics; pH refers to the pH value of the chalcopyrite flotation pulp; C Mg The mass concentration of magnesium ions in the slurry; This is the critical threshold at which magnesium ions begin to undergo macroscopic precipitation.

[0013] In one possible implementation, the theoretical recovery rate is estimated based on the type of collector used, specifically including: The theoretical recovery rate is calculated by selecting the corresponding calculation rule based on the type of collector, wherein: If the collector is an ionic agent, the basic score corresponding to the double-layer compression effect is deducted, and the precipitation coverage reduction is fully included to calculate the theoretical recovery rate. If the collector is a nonionic agent, the base score is retained and a portion of the precipitation cover reduction is included to calculate the theoretical recovery rate. If the collector is absent, a double base score deduction is performed to calculate the theoretical recovery rate.

[0014] In one possible implementation, the method further includes replacing the lower limit of the extreme value with the theoretical recovery rate if the estimated theoretical recovery rate is less than the lower limit of the extreme value.

[0015] In one possible implementation, the theoretical recovery rate is mapped to a Logit anchor in an unbounded real space via a log-odds function, specifically including: The theoretical recovery rate is converted into a normalized probability value p in the open interval of 0 to 1; The normalized probability value p is converted into a Logit anchor point in the unbounded real space by the formula Logit(p) = ln(p / (1-p)).

[0016] The beneficial effects of this invention are: Compared with existing technologies, the chalcopyrite recovery prediction method provided in this embodiment includes generating a multidimensional feature vector using the operating parameters of chalcopyrite flotation and constructing phase transition features of the mineral interface using a continuous activation function; estimating the theoretical recovery rate based on the selected collector type, ion type and concentration of the water environment, and pH value; mapping the theoretical recovery rate to Logit anchor points in unbounded real space using a log-probability function; using a TabPFN network based on the Transformer architecture, taking the multidimensional feature vector and phase transition features as inputs to the TabPFN network, and predicting nonlinear residuals in unbounded real space using Logit anchor points as conditions; generating controlled residuals by applying asymmetric constraints to the nonlinear residuals; and superimposing the controlled residuals with the theoretical recovery rate to obtain the corrected recovery rate. This method utilizes a continuous activation function to construct phase transition features of the mineral interface and uses the multidimensional feature vector and these phase transition features as inputs to the TabPFN network. It predicts nonlinear residuals by using the Logit anchor points mapped from the theoretical recovery rate via a log-odds function as conditions. Furthermore, it generates controlled residuals by applying asymmetric constraints to these nonlinear residuals. Finally, it superimposes these controlled residuals with the theoretical recovery rate to obtain a corrected recovery rate. This corrected recovery rate is more accurate than the recovery rates predicted by the two existing models, thus solving the problems of the prior art. Attached Figure Description

[0017] Figure 1 A flowchart of a chalcopyrite recovery prediction method provided by the present invention; Figure 2 A schematic diagram of the chalcopyrite recovery rate prediction device provided by the present invention; Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0018] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0019] As mentioned earlier, in the flotation process of chalcopyrite, especially in high-salinity water environments such as seawater and high-recirculation backflow, the pulp composition is complex and rich in unavoidable ions such as magnesium and iron ions. Existing recovery prediction technologies mainly rely on kinetic equations constructed from human experience or purely data-driven machine learning models. However, traditional mathematical modeling is difficult to handle nonlinear phase transition conditions with strong coupling of multiple variables; purely data-driven models, due to a lack of understanding of the physicochemical mechanisms of mineral interfaces, are prone to causing prediction results to deviate significantly from the true values.

[0020] Therefore, embodiments of this application provide a method for predicting chalcopyrite recovery rate, which can be used to solve the problems in the prior art. For example... Figure 1 The diagram shows the specific flow chart of this chalcopyrite recovery prediction method, which includes the following steps: Step S11: Generate a multidimensional feature vector using the operating parameters of chalcopyrite flotation, and construct the phase transition features of the mineral interface using a continuous activation function.

[0021] Among them, operating parameters refer to basic data reflecting the physicochemical state of the chalcopyrite pulp during flotation. In practical applications, these operating parameters may specifically include the pH value of the pulp, water quality type (e.g., seawater or freshwater), and the mass concentration of key scaling ions (e.g., magnesium ion mass concentration). The multidimensional feature vector transforms the aforementioned discrete, dimensionlessly variable operating parameters into a sequence vector (Token) in a high-dimensional space, enabling subsequent unified processing by the network.

[0022] In this application, the continuous activation function can be the Sigmoid function, which maps the input variable to a smooth interval of (0, 1) to simulate the gradual process of chemical reactions at the mineral interface, avoiding the abrupt errors caused by traditional step functions. It is important to note that the phase transition features in this application can specifically include iron ion oxidation phase transition features and / or precipitation-covered phase transition features. The mathematical expression of the Sigmoid function is f(x) = 1 / (1 + e^(-x)). On a coordinate system, this function is a smooth S-shaped curve, and it possesses a remarkable mathematical property: no matter how extreme the input independent variable x (such as unlimited acidity / alkalinity or ion concentration) is, it can perfectly compress and map it to the closed interval (0, 1).

[0023] For the specific implementation of step S11, for example, the operating parameters of the chalcopyrite flotation slurry can be collected first. For example, the pH value of the slurry can be measured by a pH meter, and the water quality type used to prepare the slurry can be identified (e.g., seawater). The mass concentration of key scale-forming ions such as magnesium ions and iron ions can be detected by chromatography or other methods, and the type of collector used can be identified, thereby obtaining the operating parameters. The operating parameters can then be used to generate a multidimensional feature vector.

[0024] For specific methods of constructing phase transition characteristics of mineral interfaces using continuous activation functions, one approach is to use Formula 1 as shown below to construct the iron ion oxidation phase transition characteristics of mineral interfaces: Formula 1 In Formula 1, pH refers to the pH value of the chalcopyrite flotation pulp, which can be measured using a pH meter or other methods. Fe This is a characteristic phase transition of iron ion oxidation; The nonlinear inflection point of iron ion oxidation in the slurry is actually the pre-calibrated critical threshold for the macroscopic oxidation and precipitation of iron ions, which is preferably set to 9.0 in this embodiment.

[0025] In a real mineral processing system, iron ions on the surface of unprotected chalcopyrite begin to undergo slight "rusting" (i.e., oxidative hydrolysis) at near-neutral pH (pH≈7.0). This invention does not employ the traditional hard-cut-off threshold of black and white, but instead introduces the aforementioned S-shaped smooth gradient sigmoid function for feature extraction. Typically, the nonlinear inflection point of iron ion oxidation in the slurry (i.e., the midpoint where 50% macroscopic oxide layer coverage occurs) can be pre-calibrated, for example, calibrated as... If the possible reaction starting point is pH=7.0, the calculated iron ion oxidation phase transition characteristic is 0.247%. This iron ion oxidation phase transition characteristic effectively describes, from a mathematical perspective, that an oxidation reaction has already occurred on the mineral surface at pH=7.0, and that precipitation increases rapidly when the alkalinity rises to the inflection point of pH=9.0, thus increasing the weight of its influence on the recovery rate.

[0026] In Formula 1 above, the slope is set to 3.0 because the oxidation and hydrolysis of iron ions on the crystal lattice surface (similar to metal rusting) is essentially a gradual electrochemical process accompanied by electron transfer. It requires surface adsorption, reaction generation, and gradual spread from point to surface. The reaction is relatively slow, so setting the slope to 3.0 makes the S-shaped curve relatively flat, creating a wider physical span (allowing the reaction to start at pH 7.0 and only reach 50% macroscopic coverage at pH 9.0). This simulates the thermodynamic gradual change characteristics of metal oxidation, characterized by "early start and slow process."

[0027] This allows for the determination of pH value and pre-calibrated parameters in the chalcopyrite flotation pulp. Substituting these values ​​into Formula 1, the iron ion oxidation phase transition characteristics at the mineral interface can be calculated. Fe .

[0028] As mentioned above, the phase transition characteristics in step S11 can also be precipitation-covered phase transition characteristics at the mineral interface. Therefore, step S11 can also specifically be constructed using Formula 2 as shown below: Formula 2 In Formula 2, Feature Mg Precipitation-covered phase transition characteristics; its function is to quantify the degree of magnesium ion precipitation and coverage on the chalcopyrite surface under high salinity conditions, and its value directly characterizes the inhibitory strength of the precipitate on the floatability of the mineral; pH is the pH value of the chalcopyrite flotation pulp, which can be measured by a pH meter or other methods; C Mg This represents the mass concentration of magnesium ions in the slurry. This is the critical threshold at which magnesium ions begin to undergo macroscopic precipitation. The magnitude of this critical threshold can usually be set according to the solubility product constant of magnesium hydroxide, for example, it can be set to about 9.5 in seawater systems.

[0029] In Formula 2, A nonlinear gating factor triggered by the pH of the slurry was used, employing a continuous activation function with a slope of 4.0. Compared to the gradual change in the iron ion oxidation characteristic (whose slope was set to 3.0), the larger slope coefficient replicates the highly sensitive abrupt change characteristic unique to precipitation reactions, such as the reaction occurring once the pH slightly exceeds a certain threshold. The value of this nonlinear gating factor rapidly increases from near 0 to near 1, simulating the explosive growth of precipitation nucleation.

[0030] Formula 2 This is a mass concentration scaling factor used to transform probabilistic events triggered by pH values ​​into coverage with physical dimensions, reflecting the physical law that higher concentrations result in thicker coverage. Multiplying the nonlinear gating factor and the mass concentration scaling factor means that even at high pH values, if the magnesium ion concentration is extremely low (e.g., in freshwater environments), the Feature... Mg The value remains close to zero; conversely, if the concentration is extremely high but the pH value does not reach the threshold, the Feature... Mg The value is also suppressed, and this dual-condition constraint mechanism effectively avoids misjudgment caused by a single variable.

[0031] Step S12: Estimate the theoretical recovery rate based on the type of collector used, the type and concentration of ions in the water environment, and the pH value.

[0032] The theoretical recovery rate is a baseline value derived from prior knowledge of mineral processing physicochemical processes, aiming to provide an anchor point consistent with physical common sense for subsequent predictions. This step is specifically implemented by calculating the recovery rate along a pathway based on the molecular structure characteristics of the collector. Therefore, step S12 can be achieved by estimating the theoretical recovery rate based on the selected collector type, the ion type and concentration in the water environment, and the pH value.

[0033] If the selected collector is an ionic reagent, it should be noted that in a conventional freshwater flotation system, a double electric layer of a certain thickness (i.e., an electrostatic buffer at the solid-liquid interface) will form on the surface of chalcopyrite particles. However, in seawater or high-salinity circulating water systems, the high concentration of basic salt ions (such as Na+) in the solution will lead to a higher concentration of Na+. + Electrostatic shielding occurs when ionic agents (such as those formed by the mineral) are compressed. This forces the dispersion layer on the mineral surface to be compressed, resulting in a thinner double layer. Therefore, the double layer compression effect caused by the high salinity of seawater causes severe electrostatic repulsion of charged groups of ionic agents when they approach the mineral, leading to a decrease in adsorption activity. Because the theoretical recovery rate is suppressed by seawater, the initial baseline value (e.g., from 96% to 93%) can be deducted from the first estimation module in the model. Furthermore, because the hydrophobic film formed is thin and cannot prevent sediment coverage in seawater, it should be subject to 100% sediment reduction in subsequent calculations, meaning the full amount of sediment coverage reduction should be included to ultimately calculate the theoretical recovery rate.

[0034] If a nonionic collector is selected, it relies on coordination bonds to bind copper atoms on the mineral surface, rather than electrostatics. Therefore, nonionic collectors are unaffected by the electrostatic interference of seawater (i.e., no double-layer compression). Furthermore, the larger molecular size of nonionic collectors creates a protective layer (i.e., a physical shield) after adsorption onto the mineral surface. Because it is also unaffected by seawater electrostatics, the initial baseline score is not reduced (retaining a baseline value of 96%). More importantly, this protective layer effectively prevents precipitates from directly contacting the ore. Therefore, when calculating the theoretical recovery rate, the reduction caused by precipitation can be reduced using the second estimation module in the model (e.g., the reduction coefficient is reduced from 10.0 for ionic collectors to 8.0), thus obtaining the theoretical recovery rate.

[0035] Of course, in the absence of a collector, the ore is completely unprotected. In high-salinity seawater, water molecules coat the mineral surface, forming a highly saline layer, making the mineral extremely hydrophilic (i.e., it sinks to the bottom and cannot float). Simultaneously, without the aid of a reagent, iron oxidation and rusting, as well as magnesium and calcium precipitation, will occur. Therefore, when calculating the theoretical recovery rate, the third estimation module in the model can perform a double base score deduction. This means that while the initial base value is deducted (from 96% to 93%), the maximum weighted reductions for "oxidation" and "precipitation" are applied to obtain the theoretical recovery rate.

[0036] Therefore, in step S12, the specific module in the model (first estimation module, second estimation module, or third estimation module) can be determined first based on the type of collector used. Then, the ion type, ion concentration, and pH value of the water environment are substituted into the corresponding estimation module to estimate the theoretical recovery rate. In practical applications, this model can utilize existing technologies, such as traditional mathematical modeling or purely data-driven models; there are no specific limitations here.

[0037] In addition, to avoid the theoretical recovery rate being negative or extremely small, an extreme lower limit can be set for the theoretical recovery rate. When the theoretical recovery rate calculated by the above method is less than the extreme lower limit, the extreme lower limit can be used as the theoretical recovery rate (that is, the extreme lower limit replaces the theoretical recovery rate estimated by the above method). In practical applications, the extreme lower limit can be, for example, 2%.

[0038] Step S13: Map the theoretical recovery rate to a Logit anchor point in the unbounded real space using the logit function.

[0039] Step S14: Use the TabPFN network based on the Transformer architecture to fit the table prior. Use the multidimensional feature vector and phase transition features as the input of the TabPFN network, and use the Logit anchor point as a condition to predict the nonlinear residual in the unbounded real space.

[0040] Here we can provide a unified explanation of steps S13 and S14.

[0041] In this embodiment, the TabPFN network is a table prior fitting network based on the Transformer architecture. It incorporates a multi-head self-attention mechanism, enabling it to learn causal topological patterns from massive synthetic data during pre-training and perform fine-tuning-free inference through context learning during the application phase. The nonlinear residual can refer to the deviation between the actual flotation recovery and the theoretical recovery. This deviation is mainly caused by implicit nonlinear interference resulting from strong coupling of multiple variables (such as the interference between pH, ion concentration, and reagent type). In this embodiment, the nonlinear residual can be, for example, the Logit residual.

[0042] In practice, the theoretical recovery rate obtained above can be mapped to an unbounded real value through the logit function in step S13, and used as the Logit anchor point in the unbounded real space. In step S14, the multidimensional feature vector and phase transition feature generated above can be input into the TabPFN network. At the same time, the nonlinear residual is predicted in the unbounded real space based on the Logit anchor point.

[0043] Since the multidimensional feature vector incorporates chalcopy flotation parameters, it can be used along with mineral interface phase transition features (including iron ion oxidation phase transition and precipitation-covered phase transition) as input to the TabPFN network. The logit function is a monotonically increasing function that maps probability or percentage values ​​defined in the open interval (0, 1) to the entire real number domain (-∞, +∞).

[0044] Unbounded real space can refer to a mathematical operation space formed after a log-probability function transformation, no longer constrained by the physical boundaries of 0 and 100. In this application, predicting nonlinear residuals within this unbounded real space involves utilizing the self-attention mechanism of the TabPFN network to capture the strong coupling relationships between variables in multidimensional feature vectors and phase transition features, calculating the deviation between the actual flotation result and the theoretical recovery rate. In this process, because the operation space is infinitely stretched into an unbounded real space, the model can sensitively perceive and express minute nonlinear fluctuations in extreme regions (such as when the recovery rate is extremely high or low).

[0045] For example, in a conventional bounded space, the relative difference between the recovery rate and 99% is small, which can easily lead to the vanishing of the model gradient and the neglect of this change; however, in an unbounded real space, the difference in the Logit function value corresponding to this change is significantly amplified, enabling the model to accurately fit the dramatic dynamics of a sudden drop or slight increase in the recovery rate under high alkalinity and high salinity conditions.

[0046] Therefore, this application uses the log-odds function mapping value of the theoretical recovery rate as the anchor point. The TabPFN network no longer directly predicts the absolute recovery rate value, but focuses on correcting complex interference terms that cannot be covered by the physical mechanism model. This solves the problem of prediction failure of traditional data-driven models in the physical extreme value region. Through spatial reshaping, the sensitivity and expression space of the model to small changes in signals under extreme conditions are greatly expanded.

[0047] Regarding the specific implementation of step S13 above, in practical applications, the theoretical recovery rate can first be converted into a normalized probability value p in the open interval of 0 to 1, so that the theoretical recovery rate can be normalized first. For example, if the theoretical recovery rate is 67%, it can be directly converted into 0.67. Then, the normalized probability value p can be converted into a Logit anchor point in the unbounded real space by using the Logit function formula Logit(p) = ln(p / (1-p)) as shown below. In this formula, Logit(p) is the Logit anchor point.

[0048] After calculating the Logit anchor points through step S13, in step S14, a Tabular Prior-Data Fitted Network (TabPFN) based on the Transformer architecture can be used as the prediction basis. Specifically, a massive amount of synthetic tabular data can be generated offline. For example, a structural causal model (SCM) and a Bayesian network can be introduced, and a large amount of synthetic tabular data can be generated by superimposing various nonlinear mathematical functions (such as the Sigmoid activation function). This makes the synthetic pre-training data approximate the real microscopic physicochemical evolution law of chalcopyrite in terms of mathematical topological features.

[0049] Next, a generative Transformer network containing a multi-head self-attention layer can be constructed to transform the various features in the tabular data (such as pH, ion concentration, and reagent type) into sequence vectors (Tokens) in a high-dimensional space. After pre-training on the synthetic data, the internal weights of the Transformer network are frozen, and a small amount of historical chalcopyrite data with real recovery rate labels (as a context set) is concatenated with the features of the new flotation conditions to be predicted into a sequence. This sequence is then input into the feedforward network for end-to-end feedforward inference, ultimately obtaining the TabPFN network. The multi-dimensional feature vector and phase transition features obtained in step S11 can then be used as inputs to the TabPFN network, and the Logit anchor points obtained in step S13 can be used as conditions to predict nonlinear residuals in an unbounded real space.

[0050] Step S15: Generate controlled residuals by applying asymmetric constraints to the nonlinear residuals.

[0051] Among them, asymmetric constraints are a dynamic control mechanism that aims to apply different degrees of attenuation or release to the residuals predicted by the model according to the numerical range of the theoretical recovery rate, so as to eliminate algorithmic illusion and overcome the boundary central tendency effect.

[0052] In this application, the principle of asymmetric constraints is that when the theoretical recovery rate is in a high steady-state region (e.g., greater than 85%), due to physical limits, the system applies a strong attenuation coefficient to the positive residual (predictions that attempt to continue to rise), blocking false predictions that violate common sense, such as model output exceeding 100%. When the theoretical recovery rate falls into the phase transition warning region (e.g., less than 50%) or is in a reagent-free state, it indicates that the mineral surface protection mechanism has been severely deteriorated. At this time, the system resets the attenuation coefficient to 1.0, thereby removing the suppression of the negative residual, enabling the model to fully adopt the predicted sharp drop, truly reflecting the risk of a cliff-like decline. Thus, through this asymmetric dynamic control, both overly optimistic estimations in the high segment and keen capture of extreme risks in the low segment are prevented, generating controlled residuals that conform to the laws of physical evolution.

[0053] Based on this fundamental principle, this application can pre-divide the numerical range of recovery rate into three numerical ranges: less than 50%, greater than or equal to 50% and less than or equal to 85%, and greater than 85%. A first attenuation coefficient reset factor equal to 1 is assigned to the numerical range less than 50%, a second attenuation coefficient reset factor less than 1 is assigned to the numerical range greater than or equal to 50% and less than or equal to 85% (for example, the value of the second attenuation coefficient reset factor can be around 0.2), and a third attenuation coefficient reset factor less than the second attenuation coefficient reset factor is assigned to the numerical range greater than 85% (for example, the value of the third attenuation coefficient reset factor can be 0.1). The classification is based on the flotation kinetics of chalcopyrite in a high-salt environment. For example, the range of less than 50% corresponds to the early warning zone of a sudden phase change, indicating that the protective shield on the mineral surface has been severely deteriorated and the flotation system is at risk of collapse. The range of greater than or equal to 50% and less than or equal to 85% corresponds to the normal steady-state zone, indicating that the reagent protection is working normally and no serious phase change has occurred. The range of greater than 85% is defined as the high-level saturation zone, indicating that the recovery rate is close to the physical limit and there is very little room for further improvement.

[0054] For example, when the calculated theoretical recovery rate is 42.5%, it is determined to fall into the first numerical region (<50%), which is the region with a recovery rate less than 50%. When the theoretical recovery rate is 72.0%, it is determined to fall into the second numerical region, which is the region with a recovery rate greater than or equal to 50% and less than or equal to 85%. When the theoretical recovery rate is 91.3%, it is determined to fall into the third numerical region, which is the region with a recovery rate greater than 85%. This three-stage division quantifies the abstract flotation state into clear logical judgment conditions, providing a basis for the subsequent differentiated allocation of attenuation coefficients.

[0055] For the specific implementation of step S15, one could determine the numerical range to which the theoretical recovery rate belongs, and then obtain the attenuation coefficient reset coefficient corresponding to that numerical range. After obtaining the attenuation coefficient reset coefficient for that numerical range, one could then use the formula ε... final = ε raw × λ generates controlled residuals, where ε is the value in the formula. final For the calculated controlled residual, ε raw The residual is nonlinear, and λ is the attenuation coefficient and reset coefficient.

[0056] In other words, the nonlinear residual ε obtained from the above steps can be... raw And the attenuation coefficient reset coefficient λ corresponding to the numerical region to which the obtained theoretical recovery rate belongs, are substituted into the formula ε final = ε raw × λ, thus calculating the controlled residual ε final.

[0057] Step S16: Superimpose the controlled residual with the theoretical recovery rate to obtain the corrected recovery rate.

[0058] The superposition operation can be a linear summation operation performed in Logit space or after mapping back to physical space. It aims to fuse a robust benchmark provided by physical priors (i.e., the theoretical recovery rate) with data-driven nonlinear corrections (i.e., controlled residuals) to obtain a more accurate corrected recovery rate. For example, the generated controlled residuals can be directly added to the determined theoretical recovery rate to obtain a corrected recovery rate that conforms to both macroscopic physical laws and microscopic nonlinear characteristics. This result provides a high-precision data foundation for subsequent inverse transformation mapping and industrial control. This benchmark + residual fusion strategy improves generalization ability in complex high-salinity environments, resulting in a global coefficient of determination (R²) of over 0.96 for the final prediction, and significantly reduces the mean absolute error in extreme cliff regions, effectively solving the technical problems of large prediction bias and poor universality in existing technologies.

[0059] The chalcopyrite recovery prediction method provided in this application includes: generating a multidimensional feature vector using the operating parameters of chalcopyrite flotation; constructing phase transition features of the mineral interface using a continuous activation function; estimating the theoretical recovery rate based on the selected collector type, ion type and concentration of the water environment, and pH value; mapping the theoretical recovery rate to Logit anchor points in an unbounded real space using a log-probability function; using a TabPFN network based on a Transformer architecture, taking the multidimensional feature vector and phase transition features as inputs to the TabPFN network, and predicting nonlinear residuals in the unbounded real space using Logit anchor points as conditions; generating controlled residuals by applying asymmetric constraints to the nonlinear residuals; and superimposing the controlled residuals with the theoretical recovery rate to obtain the corrected recovery rate. This method utilizes a continuous activation function to construct phase transition features of the mineral interface and uses the multidimensional feature vector and these phase transition features as inputs to the TabPFN network. It predicts nonlinear residuals by using the Logit anchor points mapped from the theoretical recovery rate via a log-odds function as conditions. Furthermore, it generates controlled residuals by applying asymmetric constraints to these nonlinear residuals. Finally, it superimposes these controlled residuals with the theoretical recovery rate to obtain a corrected recovery rate. This corrected recovery rate is more accurate than the recovery rates predicted by the two existing models, thus solving the problems of the prior art.

[0060] Of course, after obtaining the corrected recovery rate through step S16 above, to further improve accuracy, an inverse transformation mapping can be performed on the corrected recovery rate to obtain the final predicted recovery rate. For example, the formula can be used... The corrected recovery rate is then subjected to an inverse transformation mapping to obtain the final predicted recovery rate. In this formula... Among them, Logit final For the corrected recovery rate, y final To ultimately predict the recovery rate, that is, to be able to logit the corrected recovery rate final Substituting into the formula, the final predicted recovery rate y can be calculated. final The final predicted recovery rate y obtained in this way final It can effectively eliminate over-range output caused by residual prediction bias or extreme working conditions, ensuring that the generated final predicted recovery rate is always within the legal physical boundary (i.e., 0~100%).

[0061] After obtaining the final predicted recovery rate, the method provided in this application embodiment can further perform feedback control based on the final predicted recovery rate. Therefore, the method can further include: comparing the final predicted recovery rate with a set cliff warning threshold. The cliff warning threshold is a pre-set numerical boundary based on the phase transition critical point of chalcopyrite in a high-salt environment, used to identify the dangerous region where the flotation system is about to enter a non-linear drop in recovery rate. In practical applications, the cliff warning threshold is set based on the critical state of failure of the physical shield on the mineral surface. Generally, when the actual recovery rate is below 30.0%, it usually means that the precipitates produced by scale-forming ions (such as magnesium and iron) in seawater have largely covered the mineral surface, or that the double-layer compression effect in the absence of a collector has caused the mineral to become completely hydrophilic. Therefore, the cliff warning threshold can be set to 30.0%, or it can be adjusted to values ​​such as 25.0% or 35.0% depending on the specific ore properties and water quality conditions.

[0062] In this way, the final predicted recovery rate can be compared with the set cliff warning threshold. If the final predicted recovery rate is greater than or equal to the cliff warning threshold, no adjustment is needed. However, if the final predicted recovery rate is lower than the cliff warning threshold, it indicates that an extreme condition has occurred. In this case, the conventional steady-state operating parameters cannot maintain effective mineral flotation. Therefore, a reagent compensation strategy for extreme conditions is further generated and executed. This reagent compensation strategy includes any one or more of the following strategies: switching to the use of a non-ionic collector; increasing the dosage of the collector.

[0063] For example, the agent compensation strategy could be to switch from using ionic collectors (such as xanthates) to nonionic collectors (such as thiocyanates) when ionic collectors are currently being used. This is because in high-salinity seawater environments, high concentrations of cations compress the electric double layer on the mineral surface, generating a strong electrostatic shielding effect, making it difficult for negatively charged ionic collectors to adsorb. Nonionic collectors, on the other hand, bind to copper atoms on the mineral surface through coordination bonds, are not affected by the electrostatic interference of electric double layer compression, and their larger molecular volume can form a dense physical space shield on the mineral surface, effectively blocking the coverage of precipitates such as magnesium hydroxide. Therefore, it is possible to switch to using nonionic collectors.

[0064] Of course, if a nonionic collector is currently being used, the reagent compensation strategy can be to increase the collector dosage, that is, to instantaneously increase the amount of collector added on top of the original dosage. This increases the absolute concentration of the reagent in the pulp, and uses the law of mass action to force the adsorption equilibrium of the reagent on the mineral surface to shift to the right, thus counteracting the competitive adsorption of highly saline layers and precipitates. For example, if the predicted recovery rate is detected to drop to 28% (the cliff warning threshold is set at 30.0%) and thiourethanes are currently being used, the stroke frequency of the dosing pump can be increased by 20%-50%, thereby increasing the collector dosage.

[0065] It should be noted that, after obtaining the aforementioned final predicted recovery rate, the method provided in this application embodiment may further include verifying the final predicted recovery rate, for example, by using statistical standard calculations to determine the global coefficient of determination of the model. The global root mean square error (RMSE) and global mean absolute error (MAE) were used to verify it.

[0066] Based on the same inventive concept as the chalcopyrite recovery rate prediction method provided in the embodiments of this application, the embodiments of this application can also provide a chalcopyrite recovery rate prediction device (hereinafter referred to as device 20). For any unclear aspects regarding device 20, please refer to the relevant content of the method embodiments. Figure 2 As shown, the device 20 may include: a construction unit 201, an estimation unit 202, a mapping unit 203, a prediction unit 204, a generation unit 205, and a correction unit 206, wherein: Construction unit 201 is used to generate multidimensional feature vectors using the operating parameters of chalcopyrite flotation and to construct phase transition features of mineral interfaces using a continuous activation function. The estimation unit 202 is used to estimate the theoretical recovery rate based on the type of collector used, the type and concentration of ions in the water environment, and the pH value. Mapping unit 203 is used to map the theoretical recovery rate to a Logit anchor point in an unbounded real space using the logit function; The prediction unit 204 is used to use the TabPFN network based on the Transformer architecture to fit the multidimensional feature vector and the phase transition feature as the input of the TabPFN network, and predict the nonlinear residual in the unbounded real space with the Logit anchor as the condition. The generation unit 205 is used to generate controlled residuals by applying asymmetric constraints to the nonlinear residuals; The correction unit 206 is used to superimpose the controlled residual with the theoretical recovery rate to obtain the corrected recovery rate.

[0067] Since the device 20 adopts the same inventive concept as the method provided in the embodiments of this application, and the device 20 can also solve the technical problem if the method can solve the prior art problem, this will not be elaborated here.

[0068] The device 20 may also include a mapping unit for performing an inverse transformation mapping on the corrected recovery rate to obtain the final predicted recovery rate.

[0069] The device 20 may further include a feedback control unit for comparing the final predicted recovery rate with a set cliff warning threshold; and, if the final predicted recovery rate is lower than the cliff warning threshold, generating and executing a reagent compensation strategy for extreme conditions, wherein the reagent compensation strategy includes any one or more of the following strategies: switching to the use of a non-ionic collector; increasing the dosage of the collector.

[0070] Specifically, generating controlled residuals by applying asymmetric constraints to the nonlinear residuals includes: Determine the numerical range to which the theoretical recovery rate belongs; Obtain the attenuation coefficient reset coefficient corresponding to the numerical region; Based on the attenuation coefficient reset coefficient, using the formula ε final = ε raw × λ generates controlled residuals, where ε final Let ε be the controlled residual. raw Let λ be the nonlinear residual, and λ be the attenuation coefficient reset coefficient.

[0071] The device 20 may further include a division and allocation unit for pre-dividing the numerical region of recovery rate into three numerical regions: less than 50%, greater than or equal to 50% and less than or equal to 85%, and greater than 85%; allocating a first attenuation coefficient reset coefficient equal to 1 to the numerical region less than 50%; allocating a second attenuation coefficient reset coefficient less than 1 to the numerical region greater than or equal to 50% and less than or equal to 85%; and allocating a third attenuation coefficient reset coefficient less than the second attenuation coefficient reset coefficient to the numerical region greater than 85%.

[0072] Specifically, the phase transition characteristics of mineral interfaces are constructed using continuous activation functions, including constructing the iron ion oxidation phase transition characteristics of mineral interfaces using the following formula: ; Wherein: pH is the pH value of the chalcopyrite flotation pulp; Feature Fe This is a characteristic phase transition of iron ion oxidation; This represents the nonlinear inflection point of iron ion oxidation in the slurry.

[0073] Specifically, the phase transition characteristics of mineral interfaces are constructed using continuous activation functions, including constructing precipitation-covered phase transition characteristics at mineral interfaces using the following formula: ; Among them: Feature Mg The precipitation-covered phase transition characteristics; pH refers to the pH value of the chalcopyrite flotation pulp; C Mg The mass concentration of magnesium ions in the slurry; This is the critical threshold at which magnesium ions begin to undergo macroscopic precipitation.

[0074] The estimation of theoretical recovery rate based on the selected collector type specifically includes: The theoretical recovery rate is calculated by selecting the corresponding calculation rule based on the type of collector, wherein: If the collector is an ionic agent, the basic score corresponding to the double-layer compression effect is deducted, and the precipitation coverage reduction is fully included to calculate the theoretical recovery rate. If the collector is a nonionic agent, the base score is retained and a portion of the precipitation cover reduction is included to calculate the theoretical recovery rate. If the collector is absent, a double base score deduction is performed to calculate the theoretical recovery rate.

[0075] The device 20 may also include a replacement unit for replacing the theoretical recovery rate with the extreme lower limit if the estimated theoretical recovery rate is less than the extreme lower limit.

[0076] Specifically, mapping the theoretical recovery rate to a Logit anchor point in an unbounded real space using the logit function includes: The theoretical recovery rate is converted into a normalized probability value p in the open interval of 0 to 1; The normalized probability value p is converted into a Logit anchor point in the unbounded real space by the formula Logit(p) = ln(p / (1-p)).

[0077] like Figure 3 As shown, the present invention also provides an electronic device 300, which includes a processor 301 and a memory 302. Figure 3 Only some of the components of the electronic device 300 are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0078] In some embodiments, processor 301 may be a central processing unit (CPU), microprocessor, or other data processing chip for running program code stored in memory 802 or processing data, such as the method of the present invention.

[0079] In some embodiments, processor 301 may be a single server or a group of servers. The server group may be centralized or distributed. In some embodiments, processor 801 may be local or remote. In some embodiments, processor 301 may be implemented on a cloud platform. In one embodiment, the cloud platform may include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, inter-cloud, multi-cloud, or any combination thereof.

[0080] In some embodiments, memory 302 may be an internal storage unit of electronic device 300, such as a hard disk or memory of electronic device 300. In other embodiments, memory 302 may be an external storage device of electronic device 300, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on electronic device 300. Furthermore, memory 302 may include both internal storage units and external storage devices of electronic device 300. Memory 302 is used to store application software and various types of data installed on electronic device 300.

[0081] The electronic device 300 also includes a computer program stored in a memory 301 and executable on a processor 302, which, when executed by the processor 302, implements the method described in this application.

[0082] Of course, the electronic device 300 may also include a display 303, which in some embodiments may be an LED display, a liquid crystal display, a touch-screen liquid crystal display, or an OLED (Organic Light) display. Organic light-emitting diode (OLED) touchscreen, etc. Display 303 is used to display information from electronic device 300 and to display a visual user interface. Component 301 of electronic device 300 The 303s communicate with each other via the system bus.

[0083] This application also provides a computer-readable storage medium for storing a computer-readable program or instruction. When the program or instruction is executed by a processor, it can implement the steps or functions provided in the methods described above.

[0084] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0085] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting chalcopyrite recovery rate, characterized in that, include: Multidimensional feature vectors are generated using the operating parameters of chalcopyrite flotation, and phase transition features of mineral interfaces are constructed using continuous activation functions. The theoretical recovery rate is estimated based on the type of collector used, the type and concentration of ions in the water environment, and the pH value. The theoretical recovery rate is mapped to Logit anchor points in an unbounded real space using the logit function; Using the TabPFN, a table prior fitting network based on the Transformer architecture, the multidimensional feature vector and the phase transition feature are used as inputs to the TabPFN network, and the nonlinear residuals are predicted in the unbounded real space with the Logit anchor points as conditions. Controlled residuals are generated by applying asymmetric constraints to the nonlinear residuals. The controlled residual is superimposed with the theoretical recovery rate to obtain the corrected recovery rate.

2. The method according to claim 1, characterized in that, The method further includes: performing an inverse transformation mapping on the corrected recovery rate to obtain the final predicted recovery rate.

3. The method according to claim 2, characterized in that, The method further includes: The final predicted recovery rate is compared with the set cliff warning threshold. If the final predicted recovery rate is lower than the cliff warning threshold, a reagent compensation strategy for extreme conditions is generated and executed, wherein the reagent compensation strategy includes any one or more of the following strategies: switching to the use of a non-ionic collector; increasing the dosage of the collector.

4. The method according to claim 1, characterized in that, Controlled residuals are generated by applying asymmetric constraints to the nonlinear residuals, specifically including: Determine the numerical range to which the theoretical recovery rate belongs; Obtain the attenuation coefficient reset coefficient corresponding to the numerical region; Based on the attenuation coefficient reset coefficient, using the formula ε final = ε raw × λ generates controlled residuals, where ε final Let ε be the controlled residual. raw Let λ be the nonlinear residual, and λ be the attenuation coefficient reset coefficient.

5. The method according to claim 4, characterized in that, The method further includes: The recovery rate was pre-divided into three numerical regions: less than 50%, greater than or equal to 50% and less than or equal to 85%, and greater than 85%. Assign a first attenuation coefficient and reset coefficient equal to 1 to the numerical region that is less than 50%; For the numerical range of 50 or greater and 85% or less, assign a second attenuation coefficient reset coefficient less than 1; For the numerical range greater than 85%, assign a third attenuation coefficient reset factor that is smaller than the second attenuation coefficient reset factor.

6. The method according to claim 1, characterized in that, Phase transition characteristics of mineral interfaces are constructed using continuous activation functions, specifically including constructing the iron ion oxidation phase transition characteristics of mineral interfaces using the following formula: ； Wherein: pH is the pH value of the chalcopyrite flotation pulp; Feature Fe This is a characteristic phase transition of iron ion oxidation; This represents the nonlinear inflection point of iron ion oxidation in the slurry.

7. The method according to claim 1, characterized in that, Phase transition characteristics of mineral interfaces are constructed using continuous activation functions, specifically including constructing precipitation-covered phase transition characteristics at mineral interfaces using the following formula: ； Among them: Feature Mg The precipitation-covered phase transition characteristics; pH refers to the pH value of the chalcopyrite flotation pulp; C Mg The mass concentration of magnesium ions in the slurry; This is the critical threshold at which magnesium ions begin to undergo macroscopic precipitation.

8. The method according to claim 1, characterized in that, The theoretical recovery rate is estimated based on the type of collector used, specifically including: The theoretical recovery rate is calculated by selecting the corresponding calculation rule based on the type of collector, wherein: If the collector is an ionic agent, the basic score corresponding to the double-layer compression effect is deducted, and the precipitation coverage reduction is fully included to calculate the theoretical recovery rate. If the collector is a nonionic agent, the base score is retained and a portion of the precipitation cover reduction is included to calculate the theoretical recovery rate. If the collector is absent, a double base score deduction is performed to calculate the theoretical recovery rate.

9. The method according to claim 8, characterized in that, The method further includes replacing the theoretical recovery rate with the extreme lower limit if the estimated theoretical recovery rate is less than the extreme lower limit.

10. The method according to claim 1, characterized in that, The theoretical recovery rate is mapped to a Logit anchor point in an unbounded real space using the logit function, specifically including: The theoretical recovery rate is converted into a normalized probability value p in the open interval of 0 to 1; The normalized probability value p is converted into a Logit anchor point in the unbounded real space by the formula Logit(p) = ln(p / (1-p)).

Images