Deep learning based zinc leaching rate prediction and process optimization system

The deep learning-based zinc leaching rate prediction and process optimization system solves the problem of prediction deviation in existing technologies, achieves accurate prediction and risk control in the presence of iron interference, and improves the overall efficiency and safety of hydrometallurgical zinc refining.

CN123072525APending Publication Date: 2026-11-10HUNAN RUIXIANG NONFERROUS METAL MATERIALS CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611069447.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-17
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing methods for predicting zinc leaching rates are unable to capture the dynamic competitive relationship between zinc leaching and iron interference, leading to systematic biases in prediction results when raw materials and processes change, and process optimization has failed to effectively control the risk of iron leaching.

Method used

A deep learning-based zinc leaching rate prediction and process optimization system is adopted. The raw material element analysis data is divided into a zinc leaching driving parameter set and an iron interference driving parameter set through a parameter decoupling unit. A dual-gated recurrent network is used to extract the state latent vector. The iron alum deactivation evaluator and feature corrector are combined to eliminate the iron interference effect. An iron leaching risk assessment module is introduced into the process optimization.

Benefits of technology

It enables accurate prediction of zinc leaching rate under conditions of raw material composition fluctuations and process changes, improves prediction accuracy, and controls the risk of iron leaching during the optimization process, thus achieving synergistic optimization of the entire hydrometallurgical zinc refining process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123072525A_ABST
    Figure CN123072525A_ABST
Patent Text Reader

Abstract

The present application relates to the technical field of zinc hydrometallurgy process optimization, in particular to a zinc leaching rate prediction and process optimization system based on deep learning, a parameter decoupling unit divides raw material data into a zinc leaching driving parameter set and an iron interference driving parameter set according to the chemical mechanism of zinc leaching and iron interference, and calculates the total active oxygen molar ratio, a zinc leaching timing coding unit and an iron interference timing coding unit extract a zinc leaching state hidden vector and an iron interference state hidden vector respectively, an iron jarosite deactivation evaluator in a competition inhibition fusion unit generates a zinc leaching inhibition factor, a feature corrector uses the factor to attenuate and modulate the zinc leaching state hidden vector, and outputs a zinc leaching feature vector eliminating iron interference, a leaching rate prediction unit obtains a zinc leaching rate prediction value by regression on the feature vector, and a process optimization unit generates an iron dissolution risk index according to the iron interference state hidden vector, and searches and outputs an optimal process parameter combination when the index is lower than a safety threshold.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of hydrometallurgical zinc smelting process optimization technology, and more specifically, to a zinc leaching rate prediction and process optimization system based on deep learning. Background Technology

[0002] In the acid leaching process of hydrometallurgical zinc refining, the leaching rate is the core indicator for measuring the dissolution efficiency of zinc minerals. The valence state transformation of iron elements associated with the raw materials during the leaching process and the precipitation reaction of potassium ferric sulfate can inhibit the leaching rate of zinc by consuming hydrogen ions and oxidants and forming a mass transfer barrier on the surface of zinc mineral particles. Therefore, the accurate quantification of the iron interference effect is the key to accurately predicting the leaching rate. Most existing zinc leaching rate prediction methods are based on statistical modeling of the relationship between raw material elemental composition and leaching rate, establishing static mapping relationships through regression analysis or shallow neural networks. While these methods can achieve acceptable prediction accuracy under relatively stable raw material and process conditions, they struggle to capture the dynamic competition between zinc leaching and iron interference reactions. Furthermore, they fail to incorporate the time-varying cumulative effect of iron alum precipitation on the nonlinear inhibitory effect of the leaching process into the prediction framework. This leads to significant systematic biases in the prediction results when the iron content in the raw material fluctuates or leaching conditions change. At the process optimization level, existing methods primarily rely on zinc leaching rate... Parameter optimization with the single objective of maximizing or minimizing energy consumption rarely incorporates the risk of iron leaching as an independent safety constraint into the optimization model. This means that while optimized process parameters can improve the leaching rate, the surge in iron leaching may increase the burden on subsequent iron removal processes. Therefore, how to explicitly separate and eliminate the iron interference effect in zinc leaching rate prediction, and how to balance leaching efficiency improvement and iron leaching risk control in process optimization, is a key technical challenge for achieving synergistic optimization of the entire hydrometallurgical zinc refining process. To solve this technical problem, we provide a zinc leaching rate prediction and process optimization system based on deep learning. Summary of the Invention

[0003] The purpose of this invention is to provide a deep learning-based zinc leaching rate prediction and process optimization system to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, a deep learning-based zinc leaching rate prediction and process optimization system is provided, including: The parameter decoupling unit is used to receive the raw material element analysis data of the wet zinc leaching process. Based on the chemical mechanism of zinc leaching and iron interference, the raw material element analysis data is divided into the zinc leaching driving parameter set and the iron interference driving parameter set. The total active oxygen molar ratio is calculated based on the molar content of iron and zinc in the raw material. The zinc leaching timing coding unit adopts the first gated recurrent network, taking the time series of the zinc leaching driving parameter set as input to extract the zinc leaching state latent vector; The iron interference timing coding unit adopts a second gated recurrent network, taking the time series of the iron interference driving parameter set as input, and extracts the iron interference state latent vector. The competitive inhibition fusion unit includes an iron alum inactivation evaluator and a feature corrector. The iron alum inactivation evaluator takes the iron interference state latent vector and the total active oxygen molar ratio as inputs and generates a zinc leaching inhibition factor through a fully connected network. The feature corrector uses the zinc leaching inhibition factor to perform element-wise attenuation modulation on the zinc leaching state latent vector and outputs a zinc leaching feature vector that has eliminated the iron interference effect. The leaching rate prediction unit performs regression mapping on the zinc leaching feature vector to obtain the predicted zinc leaching rate value at the current moment. The process optimization unit is connected to the leaching rate prediction unit and the iron interference timing coding unit. It has an iron leaching risk assessment module, which is used to generate an iron leaching risk index based on the iron interference state latent vector. Under the condition that the iron leaching risk index is lower than the safety threshold, the module searches for and outputs the combination of leaching process parameters with the goal of maximizing the zinc leaching rate prediction value.

[0005] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention decouples raw material parameters into two types of driving sets based on the chemical mechanism of zinc leaching and iron interference. It uses a dual-gated recurrent network to extract the temporal latent vectors of zinc leaching state and iron interference state, respectively. A zinc leaching inhibition factor is generated through an iron alum deactivation evaluator. The zinc leaching characteristics are modulated element-wise to explicitly eliminate the iron interference effect. A hybrid expert regression network is used to achieve accurate prediction of leaching rate. At the same time, iron leaching risk constraints are introduced into the process optimization. Through multi-stage progressive search, the optimization is carried out within the safety threshold. The chemical mechanism of hydrometallurgy is deeply integrated with deep learning, which can adapt to the fluctuation of raw material composition and changes in operating conditions, and improve the accuracy of zinc leaching rate prediction. Attached Figure Description

[0006] Figure 1 This is an overall block diagram of the present invention. Detailed Implementation

[0007] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0008] This invention provides a deep learning-based zinc leaching rate prediction and process optimization system. Please refer to [link / reference]. Figure 1 As shown, it includes: The parameter decoupling unit is used to receive the raw material element analysis data of the wet zinc leaching process. Based on the chemical mechanism of zinc leaching and iron interference, the raw material element analysis data is divided into the zinc leaching driving parameter set and the iron interference driving parameter set. The total active oxygen molar ratio is calculated based on the molar content of iron and zinc in the raw material. The zinc leaching timing coding unit adopts the first gated recurrent network, taking the time series of the zinc leaching driving parameter set as input to extract the zinc leaching state latent vector; The iron interference timing coding unit adopts a second gated recurrent network, taking the time series of the iron interference driving parameter set as input, and extracts the iron interference state latent vector. The competitive inhibition fusion unit includes an iron alum inactivation evaluator and a feature corrector. The iron alum inactivation evaluator takes the iron interference state latent vector and the total active oxygen molar ratio as inputs and generates a zinc leaching inhibition factor through a fully connected network. The feature corrector uses the zinc leaching inhibition factor to perform element-wise attenuation modulation on the zinc leaching state latent vector and outputs a zinc leaching feature vector that has eliminated the iron interference effect. The leaching rate prediction unit performs regression mapping on the zinc leaching feature vector to obtain the predicted zinc leaching rate value at the current moment. The process optimization unit is connected to the leaching rate prediction unit and the iron interference timing coding unit. It has an iron leaching risk assessment module, which is used to generate an iron leaching risk index based on the iron interference state latent vector. Under the condition that the iron leaching risk index is lower than the safety threshold, the module searches for and outputs the combination of leaching process parameters with the goal of maximizing the zinc leaching rate prediction value.

[0009] The acidic leaching environment is a solid-liquid mixed reaction system in the wet zinc smelting process, which uses sulfuric acid aqueous solution as the leaching agent and zinc mineral particles as the dispersed solid phase. It is the spatial carrier in which the zinc leaching reaction and the iron interference reaction occur together. The specific settings include five quantifiable parameters, and the corresponding reaction conditions can be directly reproduced based on these settings.

[0010] The first is the liquid phase medium and acidity parameters. The system uses a sulfuric acid aqueous solution prepared from pure water and industrial concentrated sulfuric acid as the continuous liquid phase. Sulfuric acid is used as the only leaching acid agent. Hydrochloric acid, nitric acid and other acids are not introduced. The initial free sulfuric acid concentration is controlled within the range of 15 grams per liter to 60 grams per liter according to the phase composition of the raw materials. During the leaching reaction, sulfuric acid is continuously consumed by the minerals. At the end of the leaching, the free sulfuric acid concentration is maintained within the range of 2 grams per liter to 10 grams per liter. This acidity range can ensure the stable dissolution of zinc ions and avoid excessive acid from aggravating the indiscriminate dissolution of impurities such as iron and silicon.

[0011] Secondly, there are temperature and pressure conditions. In conventional atmospheric pressure acid leaching, the system temperature is controlled between 70 and 95 degrees Celsius, and a constant temperature is maintained by indirect heating through steam coils. The entire reaction is carried out in an open state under atmospheric pressure, with the liquid surface connected to the outside air. If a pressure leaching process is used for zinc sulfide concentrate, the system pressure is controlled between 0.6 MPa and 1.5 MPa, and the temperature is increased to 120 to 160 degrees Celsius to enhance the oxidation and dissolution efficiency of zinc sulfide minerals.

[0012] Thirdly, the slurry concentration and mass transfer conditions are important. After crushing and grinding, the proportion of zinc mineral particles with a diameter of less than 0.074 mm is not less than 80%, which meets the feed particle size requirements of the leaching process. After the mineral particles are added to the sulfuric acid solution, a suspended slurry system is formed. The liquid-solid mass ratio is controlled between 3:1 and 10:1, which is the ratio of the mass of the liquid phase to the mass of the solid phase ore. A mechanical stirring device is installed in the reaction tank, and the linear speed of the stirring blade is controlled between 2 meters per second and 5 meters per second to ensure that the mineral particles are uniformly suspended and do not settle, thereby enhancing the solid-liquid two-phase mass transfer efficiency.

[0013] Fourthly, the oxidation atmosphere parameter is used for the leaching system of sulfide mineral raw materials. By introducing compressed air or industrial oxygen into the bottom of the slurry, or by adding manganese dioxide and hydrogen peroxide in a certain amount, the oxidation environment required for the reaction is provided. The oxidation-reduction potential of the system is controlled between 350 mV and 650 mV to ensure the oxidation and destruction efficiency of the sulfide lattice.

[0014] Fifthly, the system consists of coexisting components. In addition to zinc minerals and sulfuric acid, the system also contains iron, calcium, magnesium, silicon, and aluminum, which are associated with the raw materials, as well as soluble sulfates and suspended insoluble gangue residues generated by the reaction. This constitutes a complex reaction system with multiple coexisting components. The valence state transformation related to iron interference and the precipitation reaction of iron alum both occur simultaneously in this coexisting system.

[0015] The process by which zinc oxides or sulfides react with sulfuric acid to form soluble zinc sulfate is the core reaction in acidic leaching systems, in which zinc is transferred from solid minerals to liquid solutions. Based on the phase type of zinc minerals, this process can be divided into two independent reaction pathways, which together constitute the chemical mechanism basis of zinc leaching.

[0016] The first type is the double decomposition reaction between zinc oxide and sulfuric acid, which belongs to the acid-base neutralization reaction and is the main leaching reaction for oxidized zinc minerals. Specifically, it is defined as follows: When the zinc oxide lattice comes into contact with hydrogen ions generated by the ionization of sulfuric acid, ionic bonds break and recombine, resulting in the reaction of soluble zinc sulfate with water molecules.

[0017] Taking the most abundant zinc oxide phase in zinc calcinate as an example, the quantitative relationship of its chemical reaction is that one mole of zinc oxide reacts with one mole of sulfuric acid to produce one mole of zinc sulfate and one mole of water. The reaction process is as follows: Hydrogen ions in the liquid phase diffuse through the slurry boundary layer and reach the solid surface of zinc oxide particles. The hydrogen ions react chemically with oxygen ions on the zinc oxide lattice surface to generate water molecules and break the oxide lattice structure, releasing free zinc ions. The generated zinc ions diffuse from the particle surface into the liquid phase and combine with sulfate ions in the solution to form soluble zinc sulfate. Water molecules diffuse into the liquid phase simultaneously.

[0018] This reaction does not require the participation of an oxidant, has a low activation energy, and a fast reaction rate. In an acidic environment above 70 degrees Celsius, the leaching rate of zinc oxide can reach more than 95% within 1 to 2 hours. It is the core reaction in the neutral and acidic leaching stages of zinc roasted sand.

[0019] In addition to zinc oxide, composite oxides such as zinc ferrite can also react with sulfuric acid under high acidity conditions, but the reaction rate and leaching efficiency are lower than those of zinc oxide. Usually, it is necessary to extend the leaching time or increase the acidity to achieve full dissolution.

[0020] The second type is the redox reaction of zinc sulfides with sulfuric acid, which is the main leaching reaction of sulfide-type zinc minerals. Specifically, it is defined as the reaction process in which sulfur loses electrons due to oxidation of the zinc sulfide lattice under the combined action of acidic environment and oxidant, and zinc ions are released and combine with sulfate to form soluble zinc sulfate.

[0021] Taking zinc sulfide, the main component of natural sphalerite, as an example, the direct reaction rate between zinc sulfide and sulfuric acid under normal atmospheric pressure and dilute sulfuric acid is extremely slow, making industrial-grade leaching almost impossible. It is necessary to use an oxidant to achieve efficient dissolution. Depending on the degree of oxidation, there are two reaction pathways. The first is the moderate oxidation pathway, in which sulfur is mainly converted into elemental sulfur. The quantitative reaction relationship is that one mole of zinc sulfide reacts with one mole of sulfuric acid and one mole of oxidant to produce one mole of zinc sulfate, one mole of elemental sulfur, and one mole of water. This pathway has low acid consumption, and the by-product elemental sulfur is easy to process in subsequent steps. It is the mainstream reaction pathway in the current pressure leaching process. The second path is the deep oxidation pathway, in which sulfur is completely oxidized to sulfate. The quantitative reaction relationship is that one mole of zinc sulfide reacts with two moles of oxygen to produce one mole of zinc sulfate in a sulfuric acid medium. This path requires higher temperatures and oxygen partial pressures to achieve complete decomposition of sulfides without any elemental sulfur residue.

[0022] The reaction process consists of four consecutive steps. The first step is that the oxidant molecules and hydrogen ions diffuse to the surface of the zinc sulfide particles. The second step is that the oxidant and sulfur ions on the surface of the zinc sulfide lattice undergo electron transfer. After the sulfur element is oxidized, it destroys the stability of the sulfide lattice. The third step is that the zinc ions in the lattice are released and combine with sulfate ions in the solution to form zinc sulfate. The fourth step is that the generated zinc sulfate diffuses into the liquid phase. The oxidation products, elemental sulfur or sulfate ions, migrate to the liquid phase simultaneously, or partially adhere to the surface of the mineral particles to form a product layer.

[0023] In actual industrial scenarios, the oxygen-enriched pressurized acid leaching process for zinc concentrate uses this type of reaction to extract zinc. Under an acidic environment of 150 degrees Celsius and an oxygen partial pressure of 1.0 MPa, zinc sulfide concentrate can achieve a zinc leaching rate of over 98% after 2 to 3 hours of reaction. Most of the sulfur element enters the leaching residue in the form of elemental sulfur. Both types of reactions directly depend on the hydrogen ion concentration and oxidant supply level in the system. These are the core influencing factors in the zinc leaching driving parameters. Differences in the zinc phase composition in the raw materials will directly change the proportion of the two types of reactions, ultimately affecting the overall zinc leaching rate and the upper limit of the leaching rate.

[0024] In an acidic leaching environment, the iron element that dissolves simultaneously with the zinc minerals first exists in the liquid phase system in the form of ferrous ions. The iron in the raw materials is mostly distributed with zinc minerals in the form of ferrous sulfide, sphalerite, hematite and other phases. Under acidic conditions, the crystal lattice of the iron minerals breaks, and the ferrous element dissociates into the slurry solution. Ferrous ions have thermodynamic stability in a weakly acidic environment at room temperature and will not spontaneously hydrolyze and precipitate.

[0025] As the leaching reaction progresses, the oxidant in the system comes into contact with ferrous ions and undergoes a redox reaction, oxidizing the ferrous ions to ferric ions. When oxygen is used as the oxidant, the quantitative reaction relationship is: four moles of ferrous ions react with one mole of oxygen and four moles of hydrogen ions to produce four moles of ferric ions and two moles of water. The reaction expression is as follows: The oxidation reaction process directly consumes free hydrogen ions and oxidant in the system. Both of these substances are core reactants in the zinc leaching reaction. The oxidation process of iron occupies the reaction substrate that could originally be used for zinc leaching, forming the first inhibitory effect from the reactant supply level. The rate of oxidation reaction is determined by the system temperature, oxidant concentration and ferrous ion concentration. For every 10 degrees Celsius increase in temperature, the oxidation reaction rate increases by about 1.5 times.

[0026] The generated ferric ions will undergo hydrolysis under specific operating conditions, combining with sulfate ions and monovalent cations in the solution to form precipitates of potassium ferric alum complex salts. These complex salts have the general formula: Alkaline ferric sulfate crystals, among which It contains potassium, sodium, or ammonium ions. The crystals are yellowish-brown solid particles that are sparingly soluble in acidic solutions. Taking potassium-type ferrous sulfate as an example, three moles of ferric sulfate react with one mole of potassium sulfate and six moles of water to produce two moles of ferrous sulfate precipitate and six moles of sulfuric acid. The reaction expression is as follows: The formation of potassium ferric alum requires three core conditions: a temperature above 70 degrees Celsius, a pH value between 1.0 and 3.0, and the presence of sufficient monovalent cations in the system. When all three conditions are met, ferric ions will rapidly hydrolyze and crystallize to form micron-sized potassium ferric alum microcrystals.

[0027] Newly formed potassium ferrous sulfate crystals preferentially undergo heterogeneous deposition on the surface of existing solid particles in the slurry. Heterogeneous deposition refers to the nucleation method in which new phase crystals grow on the existing solid surface as a substrate. The required reaction energy barrier is much lower than that of homogeneous nucleation in solution. Therefore, suspended zinc mineral particles become the main carriers for potassium ferrous sulfate deposition. As the reaction continues, the crystals grow and accumulate, gradually forming a continuous porous precipitate covering layer on the surface of zinc mineral particles, which is a physical barrier that hinders mass transfer.

[0028] The physical barrier hinders zinc leaching by increasing the mass transfer resistance of hydrogen ions. The zinc leaching reaction depends on the diffusion of hydrogen ions from the liquid phase to the mineral surface. Without a precipitate covering, hydrogen ions only need to pass through the liquid-solid boundary layer at the mineral surface to reach the reaction interface. The mass transfer process follows Fick's first law, and the expression for the hydrogen ion mass flux is: ,in This represents the hydrogen ion mass flux without precipitation, expressed in moles per square meter per second. denoted as the molecular diffusion coefficient of hydrogen ions in the bulk liquid phase, expressed in square meters per second. This represents the hydrogen ion concentration in the bulk liquid phase, expressed in moles per cubic meter. This represents the hydrogen ion concentration on the surface of zinc minerals, expressed in moles per cubic meter. The thickness of the liquid-solid boundary layer is expressed in meters.

[0029] When the zinc mineral surface forms a thickness of After the potassium ferric sulfate precipitate layer is formed, hydrogen ions need to pass through the boundary layer and the precipitate layer sequentially to reach the reaction interface. The mass transfer resistance of the two layers is in series. The effective diffusion coefficient in the precipitate layer is affected by porosity and tortuosity, and is much lower than the diffusion coefficient of the bulk liquid phase. The expression for the hydrogen ion mass flux at this time is: ,in This represents the hydrogen ion mass flux with a precipitate covering, expressed in moles per square meter per second. denoted as the intrinsic diffusion coefficient of hydrogen ions in the precipitated solid framework, expressed in square meters per second. Let be the porosity of the sediment layer, and be a dimensionless constant. Let be the tortuosity of the precipitate layer, and be a dimensionless constant used to describe the effect of the tortuosity of the precipitate channels on the extension of the diffusion path. The thickness of the potassium ferrous sulfate precipitate layer is expressed in meters.

[0030] Since the effective diffusion coefficient of the precipitate layer is only one-tenth to one-thirtieth of that of the bulk liquid phase, the total mass transfer resistance increases with the increase of the precipitate layer thickness, and the hydrogen ion mass transfer flux decreases synchronously. The number of hydrogen ions reaching the zinc mineral surface per unit time decreases, directly reducing the rate of zinc leaching reaction and forming a second inhibitory effect at the physical barrier level.

[0031] To quantify the overall inhibition intensity, the zinc leaching mass transfer inhibition coefficient can be defined as the ratio of the mass transfer flux with precipitation to the mass transfer flux without precipitation, expressed as: The value ranges from zero to one. The smaller the value, the stronger the mass transfer inhibition effect of iron alum precipitation on zinc leaching. The zinc leaching mass transfer inhibition coefficient can be used as the chemical mechanism benchmark for the zinc leaching inhibition factor in the subsequent competitive inhibition fusion unit.

[0032] In addition to the two core inhibition pathways of mass transfer obstruction and reactant consumption, the precipitation of potassium ferrous sulfate also completely encapsulates some unreacted zinc mineral particles, causing the encapsulated zinc mineral particles to completely detach from contact with the liquid phase, forming passivated particles, which further reduces the overall zinc leaching rate. Based on the complete iron interference reaction mechanism, the set of iron interference driving parameters can be divided, and the elemental components related to iron valence state transformation and ferrous sulfate crystallization can be modeled separately, accurately capturing the dynamic changes of iron interference, improving the accuracy of subsequent zinc leaching rate prediction, and providing a quantitative mechanistic basis for controlling the risk of iron leaching in the process optimization stage.

[0033] The elemental components related to the valence state transformation of iron and the precipitation of iron alum in the raw material elemental analysis data, including the content of iron, manganese, and aluminosilicates, are classified into the iron interference driving parameter set.

[0034] After receiving the raw material elemental analysis data from the hydrometallurgical zinc leaching process, the parameter decoupling unit performs parameter classification based on the chemical mechanism of zinc leaching and iron interference. The zinc leaching driving parameter set is used to collect all elemental components that directly affect the substrate supply, reaction rate, and dissolution equilibrium of the zinc leaching reaction. The classification process is based on the stoichiometric relationship of the chemical reaction to determine the classification, avoiding irrelevant parameters from being mixed in, which would lead to feature mixing in subsequent time-series coding. The classification determination is carried out around three core dimensions, corresponding to the influence of hydrogen ion concentration, the influence of oxidant supply, and the influence of zinc mineral solubility. Elemental components that satisfy the direct interaction relationship of any one of the dimensions and have no direct catalytic effect on the iron interference reaction are included in the zinc leaching driving parameter set.

[0035] The first parameter to be included in the zinc leaching driving parameter set is the zinc content. The zinc mass percentage in the raw material elemental analysis data is the most crucial leaching substrate parameter. The zinc phase exists in the raw material minerals, and its phase composition directly determines the solubility of the zinc minerals, making it a core component related to zinc mineral solubility. The parameter decoupling unit first decouples the zinc content... The conversion from mass percentage to molar content is calculated using the formula: Molar content of zinc equals the mass percentage of zinc multiplied by the total mass of raw material per unit weight, divided by the molar mass of zinc. The unit is moles per kilogram of raw material. The mathematical expression is: [Formula omitted for brevity]. This represents the zinc content by mass, and is a dimensionless value. The value for the unit mass of raw material is fixed at 1 kilogram. The molar mass of zinc is taken as 65.38 grams per mole. This represents the total molar content of zinc in a unit of raw material, expressed in moles per kilogram.

[0036] Subsequently, the parameter decoupling unit calculates the effective leaching zinc ratio based on phase analysis data. The effective leaching zinc ratio is the proportion of zinc phases soluble under acidic conditions to the total zinc. It is obtained by multiplying the proportions of zinc oxide, zinc sulfide, and zinc ferrite phases by the standard leaching efficiency coefficients of their respective phases and then summing them. The standard leaching efficiency coefficients for zinc oxide are 0.98, zinc sulfide is 0.95, and zinc ferrite is 0.30. The effective leaching zinc content is the total zinc molar content multiplied by the effective leaching zinc ratio. Finally, the total zinc element content and the effective leaching zinc ratio are jointly included in the zinc leaching driving parameter set as the core input variables for subsequent time-series coding, directly reflecting the total amount of reaction substrate and intrinsic solubility of the leaching system.

[0037] Secondly, the sulfur content is included in the zinc leaching driving parameter set. Sulfur in the raw material mainly exists in the form of sulfides and sulfates. Sulfide sulfur participates in redox reactions, directly consuming the oxidant in the system and changing the redox potential level; it is a component related to oxidant supply. Sulfate sulfur increases the total ionic strength of the solution, affecting the solubility equilibrium and activity coefficient of zinc sulfate; it is a component indirectly related to the solubility of zinc minerals. The parameter decoupling unit first calculates the molar content of reduced sulfur, which refers to sulfur element with a negative divalent oxidation state in sulfides. The calculation method is the total molar content of sulfur multiplied by the proportion of reduced sulfur phase. The conversion logic of the total molar content of sulfur is consistent with that of zinc, and the molar mass is taken as 32.07 grams per mole. Subsequently, based on the stoichiometric relationship of sulfide oxidation, the theoretical oxidant requirement corresponding to a unit of sulfur in the raw material is calculated. Using oxygen as the oxidant benchmark, 0.5 moles of oxygen are required to completely oxidize each mole of sulfide sulfur to elemental sulfur. The calculation formula is as follows: ,in This represents the molar content of reducing sulfur, expressed in moles per kilogram. The theoretical oxygen requirement corresponding to sulfur is expressed in moles per kilogram. Sulfur directly changes the redox potential of the leaching system through oxidant consumption, thereby affecting the oxidative leaching rate of zinc sulfide. Therefore, the total sulfur content and the proportion of reducing sulfur are included in the zinc leaching driving parameter set to characterize the consumption load on the oxidant supply side and assist the timing coding unit in capturing the dynamic influence of oxidation conditions on zinc leaching.

[0038] The third category of parameters categorized under zinc leaching driving parameters is the content of calcium and magnesium carbonates. Calcium and magnesium carbonates in the raw material mainly exist in the form of calcium carbonate, magnesium carbonate, and dolomite. Under acidic conditions, they undergo a neutralization reaction with sulfuric acid, directly consuming free hydrogen ions in the system and reducing the hydrogen ion concentration in the liquid phase. This is a core influencing component related to hydrogen ion concentration. The parameter decoupling unit calculates the molar content of carbonates corresponding to calcium and magnesium elements, and then converts it to the theoretical sulfuric acid consumption, i.e., the acid consumption value. Taking calcium carbonate as an example, one mole of calcium carbonate reacts with one mole of sulfuric acid to produce calcium sulfate, carbon dioxide, and water. The formula for calculating the total acid consumption per unit of calcium and magnesium carbonate in the raw material is... ,in This represents the molar content of calcium carbonate. The values ​​represent the molar content of magnesium carbonate, with each mole representing a kilogram. The theoretical sulfuric acid consumption corresponding to the carbonate is expressed in moles per kilogram. The acid consumption of calcium and magnesium carbonate directly reduces the concentration of free sulfuric acid in the system, slowing down the rate of zinc leaching. Excessive acid consumption can even lead to excessively low acidity at the leaching endpoint, causing zinc ions to hydrolyze and precipitate. Therefore, the total content of calcium and magnesium carbonate is included in the zinc leaching driving parameter set to characterize the hydrogen ion consumption load and reflect the dynamic trend of acidity changes in the system.

[0039] The classification process employs a dual-confirmation mechanism based on stoichiometric contribution verification. The parameter decoupling unit pre-establishes an influence weight matrix for the zinc leaching reaction. For each element, the influence weights of hydrogen ion concentration, oxidant supply, and zinc mineral solubility are calculated separately. Elements whose maximum value of the three weights exceeds a preset threshold of 0.1 are identified as having a direct driving effect on the zinc leaching process and are included in the zinc leaching driving parameter set. Elements whose three weights are all below the threshold and are related to iron valence state transformation and iron alum precipitation are included in the iron interference driving parameter set. The weight calculation is based on the stoichiometric ratio of the chemical reaction and the partial derivative coefficient of the reaction rate equation, ensuring that the classification results fully conform to the chemical mechanism and avoiding element misclassification.

[0040] Parameter decoupling classification based on chemical mechanisms can separate the core driving features of zinc leaching from the iron interference features at the source, avoiding the two types of features from mixing with each other in the subsequent encoding process, improving the purity of zinc leaching state encoding, and providing a basis for the accurate correction of the subsequent competition suppression fusion unit. After classification, the zinc leaching driving parameter set is arranged into a multi-dimensional time series matrix according to the sampling time order, and input into the zinc leaching time series encoding unit to perform subsequent time series feature extraction.

[0041] When performing parameter partitioning, the parameter decoupling unit adopts a feature allocation algorithm based on chemometrics. The feature allocation algorithm is a parameter classification method based on the stoichiometric relationship of chemical reaction. By calculating the contribution equivalent of each element to the zinc leaching process and the iron interference process, the parameter set is accurately partitioned, avoiding feature mismatch caused by traditional empirical classification and improving the feature purity of subsequent time-series coding.

[0042] The feature allocation algorithm pre-establishes an elemental composition comparison table for the zinc leaching driving parameter set and the iron interference driving parameter set. The elemental composition comparison table is constructed based on the thermodynamic database of zinc leaching reaction in hydrometallurgical refining. Each element entry includes the element name and molar mass, as well as zinc leaching reaction association tags and iron interference reaction association tags, along with the corresponding set of stoichiometric coefficients and reaction rate weighting coefficients. The zinc leaching reaction association tags are divided into three categories: zinc substrate, hydrogen ion consumption, and oxidant consumption, corresponding to the reaction types that directly participate in the zinc dissolution reaction and change the acidity of the system and affect the redox potential, respectively. The iron interference reaction association tags are divided into three categories: iron valence state transformation, iron alum nucleation, and iron ion complexation, corresponding to the reaction types that affect the oxidation rate of ferrous iron and participate in the growth of iron alum crystals and change the stability of ferric ions, respectively. The comparison table covers all common raw material associated elements such as zinc, sulfur, calcium, magnesium, iron, manganese, silicon, and aluminum. All stoichiometric coefficients and weighting parameters have been calibrated with industrial measured data to ensure the accuracy of the classification criteria.

[0043] During the partitioning process, the parameter decoupling unit first reads the mass percentage of each element from the raw material elemental analysis data, converts the mass percentage of each element into the molar content per unit of raw material, and uses the following conversion formula: ,in For the first The mass percentage of each element is a dimensionless value. For the first The molar mass of an element, expressed in grams per mole, is multiplied by a factor of 1000 for unit conversion. The first unit of raw material The total molar content of the elements, expressed in moles per kilogram.

[0044] After completing the molar content conversion, the parameter decoupling unit matches the chemical properties of each element with the classification labels in the elemental composition table, and calculates the total contribution equivalent of the corresponding element to the zinc leaching process and the total contribution equivalent to the iron interference process. The calculation method for the total contribution equivalent of zinc leaching is to iterate through all zinc leaching reactions associated with the corresponding element, multiply the stoichiometric coefficient of each reaction by the rate weighting coefficient of the corresponding reaction, and then sum all the products. The calculation formula is as follows: ,in This represents the total number of zinc leaching reactions associated with the corresponding element. For the first The stoichiometric coefficients of the corresponding elements in each zinc leaching reaction. For the first The rate weighting coefficients for each reaction range from zero to one, and are determined by the degree of influence of the corresponding reaction on the overall zinc leaching rate. Specifically, the weight for the zinc oxide dissolution reaction is 1.0, the weight for the zinc sulfide oxidation dissolution reaction is 0.9, the weight for the carbonate neutralization acid-consuming reaction is 0.8, and the weight for the sulfate ionic strength effect reaction is 0.3. For the first The total contribution equivalent of zinc leaching by all elements is a dimensionless value.

[0045] Similarly, the total contribution equivalent of iron interference is calculated using the following formula: ,in This represents the total number of iron-related interference reactions associated with the corresponding element. For the first The stoichiometric coefficients of the corresponding elements in each iron-interfering reaction. For the first The rate weighting coefficients for each reaction are as follows: ferrous oxidation reaction has a weight of 1.0, jaundice alum crystallization reaction has a weight of 0.9, ferric ion complexation stabilization reaction has a weight of 0.7, and heterogeneous nucleation catalysis reaction has a weight of 0.5. For the first The total contribution equivalent of iron interference from all elements is a dimensionless value.

[0046] After obtaining the two types of contribution equivalents, the two-dimensional attribution degree of each element is calculated. The zinc leaching attribution degree is the proportion of the total zinc leaching contribution equivalent to the sum of the two types of total contribution equivalents, and the iron interference attribution degree is the proportion of the total iron interference contribution equivalent to the sum of the two types of total contribution equivalents. The calculation formulas are as follows: , Both membership degrees are dimensionless values, ranging from zero to one, and their sum is always equal to one.

[0047] After calculating the attribution degree, classification matching is performed. If the zinc leaching attribution degree of a certain element is greater than or equal to 0.7, it indicates that the chemical properties of the corresponding element are highly consistent with the key reactants or products in the zinc leaching reaction mechanism. In this case, the entire molar content of the corresponding element is assigned to the zinc leaching driving parameter set. If the iron interference attribution degree of a certain element is greater than or equal to 0.7, it indicates that the chemical properties of the corresponding element are highly consistent with the key reactants or products in the iron interference reaction mechanism. In this case, the entire molar content of the corresponding element is assigned to the iron interference driving parameter set. If both attribution degrees of an element are between 0.3 and 0.7... The presence of elements in the parameter set indicates that the corresponding elements participate in two types of reaction processes simultaneously, belonging to cross-interaction elements. The algorithm uses a component splitting mechanism for allocation, splitting the molar content of the elements into two parts according to the degree of belonging, and assigning them to two parameter sets respectively. After splitting, the component obtained by the zinc leaching driving parameter set is the total molar content of the elements multiplied by the zinc leaching degree, and the component obtained by the iron interference driving parameter set is the total molar content of the elements multiplied by the iron interference degree. After splitting, the total element content is checked. If the difference between the sum of the components of the two parameter sets and the original total molar content of the elements is less than one-thousandth, the splitting is confirmed to be effective, ensuring the conservation of element mass.

[0048] For example, zinc only participates in the zinc leaching substrate reaction, with a zinc leaching attribution close to 1.0, and is entirely assigned to the zinc leaching driving parameter set. Iron dominates the iron valence state transformation and the formation of iron alum, with an iron interference attribution close to 1.0, and is entirely assigned to the iron interference driving parameter set. Manganese can act as an oxidant to promote the dissolution of zinc sulfide and also changes the redox potential of ferric ions, affecting the oxidation rate of ferrous iron. It is a typical cross-acting element, and after calculating the attribution, it is proportionally split into two parameter sets. Aluminosilicates mainly act as heterogeneous nucleation sites to accelerate the precipitation of iron alum, with an iron interference attribution higher than 0.7, and are entirely assigned to the iron interference driving parameter set. Calcium and magnesium carbonates only affect the zinc leaching acidity by consuming hydrogen ions through neutralization reactions, with a zinc leaching attribution higher than 0.7, and are entirely assigned to the zinc leaching driving parameter set.

[0049] After all elements are allocated, the parameter decoupling unit organizes all element component data of the two parameter sets and arranges them into multidimensional time series matrices according to the sampling time order. The time series matrix of the zinc leaching driving parameter set is input into the zinc leaching time series coding unit, and the time series matrix of the iron interference driving parameter set is input into the iron interference time series coding unit. The chemometrics-based quantization allocation algorithm can avoid the classification bias of traditional static reference tables. The component splitting mechanism for cross-acting elements can completely preserve the dual-effect characteristics of elements, achieving parameter decoupling of the two processes without losing the multidimensional influence information of elements, effectively improving the accuracy of subsequent time series coding and prediction models.

[0050] The total active oxygen molar ratio is calculated based on the molar content of iron and zinc in the raw materials, specifically including: The calculation of the total active oxygen molar ratio is based on the combined demand for active oxygen from both the iron ion oxidation reaction and the zinc ion leaching reaction. Active oxygen refers to the effective oxygen atoms participating in the redox reaction. The calculation process is as follows: After the parameter decoupling unit completes the division of the two types of driving parameter sets, it further calculates the total active oxygen molar ratio. The total active oxygen molar ratio is a quantitative indicator that measures the amount of oxidation equivalent required to complete the leaching reaction of a unit mass of raw material, and serves as a benchmark parameter for judging the sufficiency of oxidant supply in the iron alum deactivation evaluator.

[0051] First, the mass percentages of iron and zinc are extracted from the elemental analysis data of the raw materials and converted into molar content per unit mass of raw material. The conversion formula is: element molar content = element mass percentage multiplied by 1000 and then divided by element molar mass. The molar mass of iron is taken as 55.85 grams per mole and the molar mass of zinc is taken as 65.38 grams per mole. The calculated molar contents of iron and zinc are in moles per kilogram of raw material.

[0052] The parameter decoupling unit establishes an expression for the total demand for reactive oxygen species (ROS) based on the stoichiometric coefficient of active oxidants required to oxidize ferrous ions from ferrous ions to ferric ions, and the stoichiometric coefficient of active oxidants consumed in the dissolution of zinc minerals. ROS refers to oxygen atoms that participate in redox reactions and can provide an effective oxidative equivalent. One oxygen molecule can provide two ROS atoms. The stoichiometric coefficient is based on the mole number of ROS atoms. In the oxidation reaction of iron, each ferrous ion loses one electron to become a ferric ion, and each ROS atom gains two electrons to complete the reduction. Therefore, oxidizing one mole of ferrous ions requires 0.5 moles of ROS atoms, corresponding to an ROS of 0.5 for iron, with the dimension being moles of ROS per mole of iron.

[0053] The dissolution reaction of zinc minerals is divided into two pathways. The acid dissolution reaction of zinc oxide is a metathesis reaction, which does not involve electron transfer and does not consume active oxygen. The oxidative leaching reaction of zinc sulfide is a redox reaction, which requires the consumption of active oxygen to destroy the sulfide lattice. In the mainstream industrial medium oxidation pathway, one mole of zinc sulfide needs to consume one mole of active oxygen atoms to completely react and produce elemental sulfur and zinc sulfate. The stoichiometric coefficient of active oxygen of zinc sulfide is 1.0, and the dimension is molar active oxygen per mole of zinc sulfide.

[0054] The parameter decoupling unit, combining the zinc sulfide phase ratio from the raw material phase analysis data, calculates the theoretical reactive oxygen species (ROS) requirement on the zinc side. The zinc sulfide phase ratio is the molar proportion of zinc element in sulfide form to total zinc element. Based on the above stoichiometric relationship, the expression for the theoretical total ROS requirement is: ,in This refers to the total molar content of iron, expressed in moles per kilogram of raw material. This refers to the total molar content of zinc, expressed in moles per kilogram of raw material. The value represents the molar proportion of zinc sulfide phase to total zinc, and is a dimensionless value. The active stoichiometric coefficient for iron oxidation is 0.5 moles per mole. The active stoichiometric coefficient for zinc sulfide leaching is taken as 1.0 moles per mole. This represents the theoretical total reactive oxygen demand per unit mass of raw material, expressed in moles per kilogram of raw material.

[0055] To compensate for the utilization rate deviation caused by oxidant escape and side reaction consumption in actual production, the parameter decoupling unit introduces a dynamic balance correction coefficient. The dynamic balance correction coefficient is a dimensionless correction coefficient that is updated online based on historical batch operation data. It is used to correct the deviation between theoretical calculation values ​​and actual consumption, and improve the working condition adaptability of the total active oxygen molar ratio.

[0056] The dynamic balance correction coefficient is updated online using a sliding exponential averaging mechanism. After each leaching batch, the parameter decoupling unit reads the actual total oxidant input from the field monitoring instruments, converts the oxygen input into the total number of moles of active oxygen atoms, and obtains the actual active oxygen consumption for a single batch. Simultaneously, based on the raw material elemental analysis data for that batch, the corresponding theoretical total active oxygen demand is calculated, and the active oxygen utilization rate ratio for a single batch is calculated. This ratio equals the actual active oxygen consumption divided by the theoretical total active oxygen demand. The correction coefficient is then updated using an exponential smoothing formula. The update formula is as follows: ,in This represents the actual active oxygen consumption per batch, expressed in moles per kilogram of raw material. These are the dynamic balance correction factors before the update. These are the updated dynamic equilibrium correction coefficients, all of which are dimensionless values. The smoothing coefficient, set to 0.2, is a dimensionless constant used to balance the weights of new batch data and historical data, preventing significant jumps in the correction coefficient due to fluctuations in raw materials from a single batch. During system initialization, the initial value of the dynamic balance correction coefficient is set to 1.2, corresponding to the oxidant utilization rate level commonly found in industrial leaching systems, and is subsequently updated iteratively with each production batch.

[0057] Multiplying the theoretical total reactive oxygen species (ROS) requirement by the dynamic balance correction factor yields the final total ROS molar ratio. The calculation formula is as follows: ,in The total active oxygen molar ratio, expressed in moles per kilogram of raw material, represents the baseline value of the total equivalent active oxygen required to achieve the expected leaching effect under the current raw material conditions. After the total active oxygen molar ratio is calculated, the parameter decoupling unit transmits this value to the iron alum deactivation evaluator in the competitive inhibition fusion unit. Together with the iron interference state latent vector, it serves as the input parameter of the evaluator, used to quantify the sufficiency of oxidant supply and assist in judging the formation intensity of iron alum precipitation and its inhibition level on zinc leaching. The calculation method of dynamic equilibrium correction coefficient is introduced, which can adapt to the changes in oxidant utilization rate under different raw material batches and different operating conditions. Compared with the calculation method of fixed measurement coefficient, it has stronger on-site adaptability and can effectively improve the accuracy of subsequent zinc leaching inhibition factor calculation.

[0058] The zinc leaching timing coding unit employs a first-gated recurrent network, taking the time series of the zinc leaching driving parameter set as input, and extracts the zinc leaching state latent vector, specifically including: After the parameter decoupling unit completes the molar content conversion of the raw material element analysis data, it sequentially performs the sub-item demand calculation and total demand correction, and outputs the total active oxygen molar ratio. The total active oxygen molar ratio is the oxidation equivalent benchmark value required for a unit mass of raw material to complete the target leaching reaction. It serves as a quantitative indicator in the iron alum deactivation evaluator to measure whether the oxidant supply is sufficient, and provides a quantitative basis for the oxidation supply and demand level for the assessment of the iron interference inhibition intensity.

[0059] The calculation process first calculates the theoretical reactive oxygen species (ROS) requirements for iron and zinc. The stoichiometric coefficient for iron is determined by the electron transfer relationship in the oxidation of ferrous ions to ferric ions. Each mole of ferrous ions loses one mole of electrons during oxidation, while each mole of reactive oxygen atom accepts two moles of electrons for reduction. Therefore, oxidizing one mole of ferrous ions requires 0.5 moles of reactive oxygen atom. The corresponding ROS requirement for iron is 0.5 moles of reactive oxygen per mole of iron. The theoretical ROS requirement for iron is equal to the molar content of iron multiplied by the ROS requirement for iron. The calculation formula is as follows: ,in This refers to the total molar content of iron per unit mass of raw material, expressed in moles per kilogram. This represents the active stoichiometric coefficient for iron oxidation, with a value of 0.5 mol / mol. This represents the theoretical reactive oxygen demand corresponding to iron, expressed in moles per kilogram.

[0060] The active stoichiometric coefficient of zinc is determined by the electron transfer relationship in the oxidative leaching of zinc sulfide minerals. Under the mainstream moderate oxidation pathway in industry, sulfur in each mole of zinc sulfide is oxidized from -2 valence to zero valence elemental sulfur, losing two moles of electrons and consuming one mole of active oxygen atoms. Therefore, the active stoichiometric coefficient of zinc sulfide leaching is taken as 1.0 moles of active oxygen per mole of zinc. Since the dissolution of the zinc oxide phase in the raw material does not involve redox reactions and does not consume active oxygen, the molar proportion coefficient of the zinc sulfide phase to the total zinc needs to be introduced when calculating the zinc side. The theoretical active oxygen requirement on the zinc side is equal to the molar content of zinc multiplied by the proportion of the zinc sulfide phase and then multiplied by the active stoichiometric coefficient of zinc. The calculation formula is as follows: ,in This refers to the total molar content of zinc per unit mass of raw material, expressed in moles per kilogram. The active stoichiometric coefficient corresponding to zinc sulfide leaching is taken as 1.0 mol / mol. The value represents the molar proportion of zinc sulfide phase to total zinc, and is a dimensionless value. The theoretical reactive oxygen species (ROS) requirement corresponding to zinc is expressed in moles per kilogram. After completing the individual calculations, the two types of ROS requirements are summed to obtain the theoretical total ROS requirement. This is then multiplied by a dynamic equilibrium correction factor to compensate for oxidant dissipation and side reaction consumption in actual production, ultimately yielding the total ROS molar ratio. The calculation formula is as follows: ,in This is a dynamic equilibrium correction coefficient, a dimensionless value, updated online iteratively based on the ratio of actual oxidant consumption to theoretical consumption in historical batches. The final total active oxygen molar ratio, expressed in moles per kilogram, represents the oxidative equivalent baseline required to achieve the expected leaching depth under the current raw material conditions. After the total active oxygen molar ratio is calculated, it is transmitted to the iron alum deactivation evaluator within the competitive inhibition fusion unit. This evaluator serves as a quantitative benchmark for measuring oxidant supply sufficiency and participates in the calculation of the zinc leaching inhibition factor. The specific application logic is divided into three levels. The first level is the supply-demand ratio conversion. The iron alum deactivation evaluator first obtains the actual amount of active oxygen moles introduced into the current leaching system from the on-site detection data. It then calculates the ratio of the actual supply to the total active oxygen molar ratio to obtain the oxidant supply sufficiency coefficient. The calculation formula is as follows: ,in This represents the actual amount of active oxygen supplied per unit mass of raw material, expressed in moles per kilogram. The sufficiency coefficient for oxidant supply is a dimensionless value. A value less than one indicates insufficient supply, equal to one indicates a balance between supply and demand, and greater than one indicates excessive supply.

[0061] The second layer is the oxidation state feature mapping. The second fully connected layer in the iron alum inactivation evaluator performs a nonlinear transformation on the combination features of the total reactive oxygen molar ratio and the actual supply, extracting the oxidation supply feature representation directly related to the iron alum formation rate. The higher the sufficiency of oxidant supply, the faster the formation rate of ferric ions, the higher the steady-state concentration, the stronger the reaction driving force of jaundice iron alum crystallization, and the more significant the physical barrier inhibition effect of zinc leaching. The third layer is the fusion output inhibition factor. The oxidation supply feature and the iron interference feature obtained by the iron interference state latent vector transformation are fused element-wise and input together into the Sigmoid output layer to generate the final value. The introduction of the total active oxygen molar ratio as a zinc leaching inhibitor allows the calculation of the inhibitor to not only depend on the temporal state of iron itself, but also to combine the overall oxidation supply and demand balance of the system. This avoids evaluation bias caused by local operating conditions affecting a single detection parameter, and improves the accuracy of the quantification of inhibition intensity. The supply and demand assessment mechanism with the total active oxygen molar ratio as the core can unify the oxidation demand scale of the two processes of zinc leaching and iron interference from the perspective of chemical reaction stoichiometry, providing a traceable quantitative benchmark for the assessment of iron alum deactivation. Compared with the assessment method that relies solely on redox potential, it has stronger mechanistic consistency and is suitable for industrial production scenarios with large fluctuations in raw material composition.

[0062] The zinc leaching timing coding unit receives the zinc leaching driving parameter set output by the parameter decoupling unit and performs timing feature extraction using the first gated recurrent network. The first gated recurrent network adopts an attention-enhanced gated recurrent network structure, which introduces an adaptive time step selection mechanism on the basis of the standard gated recurrent network. It can adaptively focus on the time step features that affect the current leaching state, and dynamically adjust the temporal granularity of the encoding according to the feature distribution, taking into account both the capture of leaching details and the integration of long-term trends.

[0063] The zinc leaching timing encoding unit first arranges the elemental composition data of each sampling time in the zinc leaching driving parameter set into a multi-dimensional time series matrix in chronological order. Each sampling time corresponds to one row of the matrix, and the dimension of each row corresponds to the number of elemental components in the zinc leaching driving parameter set, including parameters such as the molar content of zinc, the molar content of sulfur, and the molar content of calcium and magnesium carbonates. The total length of the time series is the number of sampling steps within a preset sliding window, denoted as . The feature dimension is denoted as Therefore, the dimension of the input matrix is OK Before input, each feature dimension is subjected to min-max normalization to map each parameter to a value range of zero to one, eliminating the influence of differences in the dimensions of different parameters on network training. The normalization calculation formula is that the normalized value of each feature is equal to the original value minus the historical minimum value of the feature, and then divided by the difference between the historical maximum and minimum values ​​of the feature. The normalized multidimensional time series matrix is ​​input into the attention-enhanced gated recurrent network.

[0064] A standard gated loop network controls the iterative update of the hidden state through two gate control units: an update gate and a reset gate. The update gate controls the proportion of the hidden state from the previous time step that is carried over to the current time step, while the reset gate controls the proportion of the hidden state from the previous time step that is forgotten. The basic state update formula for each time step is to reset the gate. , update door Candidate hidden state Basic hidden state ,in For the first The input feature vector of the step, Let be the hidden state vector from the previous time step. , , , , as well as Here is the trainable weight matrix of the network. This represents element-wise multiplication. It is the Sigmoid activation function. For the first The candidate hidden states of each step are used as intermediate computational variables.

[0065] After the basic hidden state of the attention-enhanced gated recurrent network is updated at each time step, a time step attention weight calculation module is introduced to perform weighted aggregation of the hidden states of all historical time steps within the sliding window. The weight is determined by the contribution of the feature data of the time step to the current zinc leaching state. The contribution is obtained by calculating the similarity between the feature data of the time step and the previous hidden state. The specific calculation process is as follows: In the At each time step, calculate the values ​​from the 1st to the 6th time step within the window. The input features at each time step and the previous hidden state The cosine similarity is used to measure the degree of overlap between two vectors in a direction. The larger the value, the higher the correlation between the features at the corresponding time step and the current state, and the stronger their contribution. The formula for calculating the similarity at each time step is: ,in For the first The input feature vectors at each time step have numerators that are the inner product of two vectors and denominators that are the product of the L2 norms of the two vectors. For the first Step relative to the first The similarity of the steps ranges from negative one to one and is a dimensionless value.

[0066] After obtaining the similarity at all time steps, the similarity is mapped to attention weights summing to one using the Softmax function, thus converting contribution level into weight coefficients. The calculation formula is as follows: ,in For the first The attention weights at each time step range from zero to one, and the sum of all weights is one, making them dimensionless. The attention aggregate feature vector is obtained by weighted summation of the underlying hidden states at all historical time steps based on these attention weights, using the following formula: ,in For the first The attention-aggregated feature vector of the current step is then concatenated with the original hidden state of the current step through a channel concatenation. This concatenation is then mapped through a fully connected layer to an enhanced hidden state with the same dimension as the original hidden state. This enhanced hidden state is then used as the final hidden state output of the current step, replacing the original hidden state of the standard gated recurrent network, thus completing the attention-enhanced state update.

[0067] The adaptive time step selection mechanism adjusts the span of the time step based on the distribution of attention weights. Unlike traditional time-series coding with fixed sampling intervals, it can automatically adjust the coding granularity according to the smoothness of the leaching process. The specific execution logic is as follows: At the end of each encoding window, the information entropy of the attention weights for that window is calculated. Information entropy measures the dispersion of the weight distribution. A smaller information entropy indicates that the attention weights are concentrated in a few recent time steps, corresponding to a changing phase in the leaching process, requiring finer time granularity to capture these changes. A larger information entropy indicates a more dispersed weight distribution, corresponding to a stable phase in the leaching process, allowing for a longer time step to integrate the overall trend. The formula for calculating the information entropy of the attention weights is as follows: ,in The information entropy of the attention weights is expressed in bits and is a dimensionless value. Two information entropy thresholds are preset: a high stability threshold and a low stability threshold. When the information entropy is lower than the low stability threshold, it is determined that the attention is focused on the recent time step, and the leaching process fluctuates greatly. In this case, the step size is shortened, and the sampling step size of the next encoding window is adjusted to half of the original step size to increase the sampling point density and capture the dynamic changes in the leaching process. When the information entropy is higher than the high stability threshold, it is determined that the attention is distributed throughout the entire window, and the leaching process changes slowly. In this case, the step size is extended, and the sampling step size of the next encoding window is adjusted to twice the original step size to reduce redundant computation and integrate long-term leaching trends. When the information entropy is between the two thresholds, the original step size remains unchanged. After the step size adjustment, the input features of the corresponding time step are matched with the new step size interval through linear interpolation or mean aggregation to ensure that the input dimension is compatible with the network structure.

[0068] After time-step iterative processing and adaptive step-size adjustment by the attention-enhanced gated recurrent network, the enhanced hidden state output at the last time step of the sliding window is taken as the zinc leaching state latent vector. The zinc leaching state latent vector encodes the overall features of the evolution of zinc leaching driving parameters over time, including multi-dimensional temporal information such as zinc substrate supply, acidity changes, and oxidation conditions. Furthermore, the feature contribution of the corresponding time step is enhanced through the attention mechanism, and the adaptive step size adapts to different leaching dynamic rhythms. The zinc leaching state latent vector is output to the feature corrector in the competitive suppression fusion unit as the benchmark feature vector for subsequent iron interference correction. This temporal coding structure, which combines attention enhancement with adaptive step size, can improve the efficiency of capturing dynamic features corresponding to the leaching process and reduce interference from irrelevant historical information compared to the standard gated recurrent network. At the same time, dynamically adjusting the coding granularity can balance computational efficiency and feature accuracy, improve the coding quality of zinc leaching temporal features, and provide a feature basis for subsequent interference correction and leaching rate prediction.

[0069] The adaptive time step selection mechanism is embedded at the end of each encoding window of the attention-enhanced gated recurrent network. It dynamically adjusts the time step span of the next encoding window based on the distribution of attention weights in the current window, eliminating the need for manually setting a fixed sampling interval. It can adaptively match the dynamic change rate of the leaching process, balancing feature capture accuracy and computational efficiency. The core judgment criterion of the mechanism is the dispersion of the attention weight distribution in the current window. The dispersion is quantified by information entropy, which is a statistical indicator that measures the degree of dispersion of probability distribution. The smaller the entropy value, the more concentrated the weights are in a few time steps, corresponding to a rapid change phase in the leaching process, requiring a finer temporal granularity to capture the dynamic response. The larger the entropy value, the more evenly the weights are distributed throughout the time window, corresponding to a stable phase in the leaching process, allowing for a wider temporal granularity to integrate the overall change trend.

[0070] No. The formula for calculating the information entropy of the attention weights of each encoding window is as follows: ,in The total number of time steps contained in the current encoding window. For the current window The attention weights corresponding to each time step, with the sum of all weights being 1. The information entropy of the attention weights of the current window, in bits, is a dimensionless value ranging from 0 to 1. The mechanism pre-sets two information entropy judgment thresholds, namely a low entropy threshold. With high entropy threshold The low-entropy threshold corresponds to the judgment boundary where the weights are highly concentrated, while the high-entropy threshold corresponds to the judgment boundary where the weights are highly dispersed. Both thresholds are obtained through statistical calibration of historical leaching batch data. The low-entropy threshold is set to 0.3 times the maximum entropy value, and the high-entropy threshold is set to 0.8 times the maximum entropy value.

[0071] After each encoding window completes the attention weight calculation, the information entropy value of the current window is obtained by substituting it into the formula. This value is then compared with two thresholds. When the information entropy value is lower than the low entropy threshold, it is determined that the attention weights are concentrated in a few recent time steps, indicating that parameters such as acidity and substrate concentration in the current leaching system are in a rapidly changing phase. The mechanism then performs a time step shortening operation, adjusting the sampling step size of the next encoding window to half of the current step size. This doubles the number of sampling points per unit time, improving temporal resolution and capturing rapidly changing leaching dynamics. When the information entropy value is higher than the high entropy threshold, it is determined that the attention weights are evenly distributed. Within a preset time range, indicating that the reaction rate of the current leaching system is stable and the parameter changes are gradual, the mechanism extends the time step, adjusting the sampling step size of the next encoding window to twice the current step size. This halves the number of sampling points per unit time, compressing redundant computations and integrating long-term leaching trends. When the information entropy value is between two thresholds, the current time step size remains unchanged, maintaining the original encoding granularity. To avoid fluctuations in encoding features due to frequent jumps in step sizes between adjacent windows, the mechanism introduces a step size smoothing constraint. The target step size is updated using an exponential smoothing method during step size adjustment, with the update formula being: ,in This represents the actual step size of the current window, in minutes. The target step size is determined based on the entropy value, in minutes. This is a smoothing coefficient, with a value of 0.6, and is a dimensionless constant. This is the actual step size for the next window, in minutes. After smoothing, the step size changes smoothly without any abrupt jumps, ensuring the continuity of the timing coding.

[0072] After the step size adjustment is completed, the original sampled data needs to be resampled to match the new step size interval. When the step size is shortened, a linear interpolation method is used to insert new interpolation points between adjacent original sample points to generate denser time series data. When the step size is extended, a sliding window mean aggregation method is used to average the features of multiple consecutive original sample points and merge them into a feature vector of a coarse-grained time step. The resampled time series matrix keeps the total number of time steps within the window unchanged, ensuring that the network input dimension is compatible with the model structure and can adapt to different step sizes without adjusting the network parameters.

[0073] After time-step iterative processing by a gated recurrent network with attention weighting and adaptive step size adjustment, the enhanced hidden state output at the last time step of the encoding window is taken as the zinc leaching state latent vector. This latent vector is a high-dimensional condensed representation of the temporal evolution of the zinc leaching driving parameters. Each dimension of the vector corresponds to different leaching state features, including the current levels of core parameters such as zinc substrate supply, hydrogen ion concentration, and oxidation conditions, as well as encoding the rate and trend information of change of each parameter within the window. Furthermore, the feature contribution of key time steps is enhanced through an attention mechanism, and the adaptive step size adapts to the reaction rhythm of different stages. The final output is the zinc leaching state latent vector. The dimension of the latent vector in the leaching state is consistent with the dimension of the hidden layer in the network, usually set to 64 or 128 dimensions. The latent vector in the zinc leaching state is directly transmitted to the feature corrector in the competitive suppression fusion unit as the benchmark feature for subsequent correction of iron interference effect. This provides a temporal feature basis for eliminating the influence of iron interference and generating pure zinc leaching features. Compared with the fixed step size gated recurrent network, this encoding method can retain more refined dynamic features in the leaching fluctuation stage, reduce computational redundancy in the stable stage, and improve the representation ability of temporal encoding without significantly increasing the amount of computation. It is more in line with the multi-stage variable rate reaction characteristics of the hydrometallurgical zinc leaching process.

[0074] The iron interference timing coding unit receives the iron interference driving parameter set output by the parameter decoupling unit and performs timing feature extraction using a second gated recurrent network. The second gated recurrent network adopts a multi-scale gated recurrent network structure, which contains multiple parallel sub-gated recurrent networks. Each sub-gated recurrent network has a different expansion coefficient, which can simultaneously capture the iron ion valence state transformation and iron alum precipitation behavior characteristics at different time scales. This solves the problem that single-scale networks cannot take into account both rapid oxidation processes and slow precipitation processes, and improves the completeness of iron interference timing features.

[0075] The iron interference timing coding unit first arranges the elemental composition data of each sampling time in the iron interference driving parameter set in chronological order to form a multi-dimensional time series matrix. Each sampling time corresponds to a row in the matrix, and the dimension of each row corresponds to the number of elemental components in the iron interference driving parameter set, including the molar content of iron, manganese, and aluminosilicates. The total length of the time series is the number of sampling steps within a preset sliding window, denoted as . The feature dimension is denoted as Therefore, the dimension of the input matrix is OK Before input, each feature dimension is subjected to min-max normalization to map each parameter to a value range of zero to one, eliminating the influence of differences in the scale of different parameters on network training. The normalization calculation formula is that the normalized value of each feature is equal to the original value minus the historical minimum value of the feature, and then divided by the difference between the historical maximum and minimum values ​​of the feature. The normalized multidimensional time series matrix is ​​input in parallel into each sub-gated recurrent network.

[0076] Each sub-gated recurrent network downsamples the input time series at equal intervals using a set dilation coefficient. The dilation coefficient, a dimensionless positive integer, refers to the number of sampling intervals on the time axis and determines the width of the corresponding sub-network's temporal receptive field. A smaller dilation coefficient results in denser sampling points and higher temporal resolution, suitable for capturing rapidly changing iron ion valence state transformations. A larger dilation coefficient results in wider sampling intervals and a larger temporal receptive field, suitable for capturing slowly evolving trends in iron alum precipitation and accumulation. The multi-scale gated recurrent network consists of three parallel sub-gated recurrent networks, with corresponding dilation coefficients set to 1. 2 and 4 cover three time scales: short, medium, and long. The subnetwork with an expansion coefficient of 1 does not downsample the original sequence and retains all time steps, corresponding to a fast reaction scale of minutes. It is used to capture fast processes such as ferrous ion oxidation and manganese catalytic valence state conversion. The subnetwork with an expansion coefficient of 2 samples once every one original step, corresponding to a medium-speed change scale of ten minutes. It is used to capture medium-speed processes such as ferric ion hydrolysis and crystal nucleation. The subnetwork with an expansion coefficient of 4 samples once every three original steps, corresponding to a slow accumulation scale of hours. It is used to capture slow processes such as iron alum crystal growth and precipitation layer thickening.

[0077] The specific execution rule for downsampling is as follows: for an expansion coefficient of... The subnetwork, starting from the first time step of the original time series, sequentially selects indices 1, 1+, ... 1+2 Until the total number of steps is not exceeded The time steps constitute the downsampled subsequence, and the formula for calculating the length of the downsampled subsequence is: ,in Coefficient of expansion The corresponding subsequence length, For floor operations, The total length of the original input sequence.

[0078] Each sub-gated recurrent network adopts a standard gated recurrent unit structure, containing two gated units: a reset gate and an update gate. It independently processes the corresponding downsampled subsequence, iteratively updating the hidden state step-by-step to capture the dynamic features of iron disturbance at the corresponding scale. The hidden layer dimension of each sub-gated recurrent network is set to the same value, denoted as . To ensure dimensional consistency of multi-scale features and facilitate subsequent fusion processing, after each sub-network is processed, the hidden state of the last time step of the corresponding sub-sequence is taken as the output feature vector for that scale. The three sub-networks output three independent hidden state vectors, denoted as... All dimensions .

[0079] The iron interference temporal coding unit performs a concatenation operation on the feature channel dimension of the hidden states output by all sub-gated recurrent networks. This concatenation operation combines multiple vectors sequentially, end-to-end, into a longer vector, forming a high-dimensional combined hidden state. The dimension of the concatenated combined hidden state is the sum of the dimensions of the hidden states of each sub-network, calculated using the following formula: ,in To combine the hidden state vectors, the total dimension is .

[0080] Because the combined hidden state after concatenation has a high dimensionality, exhibits feature redundancy, and is detrimental to subsequent fusion calculations, the iron interference temporal coding unit further compresses the combined hidden state to a preset dimension through a fully connected dimensionality reduction layer. This fully connected dimensionality reduction layer consists of a single linear transformation network that linearly maps the high-dimensional features using trainable weight matrices and bias vectors, and completes feature compression and reconstruction using a non-linear activation function. The formula for dimensionality reduction is as follows: ,in This is the weight matrix of the dimensionality reduction layer, with dimensions equal to the preset output dimension multiplied by the combined hidden state dimension. This is the bias vector for the dimensionality reduction layer, with the dimension being the preset output dimension. This is a linear rectification activation function used to introduce nonlinear transformation capability. The output dimension is set to be consistent with the dimension of the zinc leaching state latent vector after dimensionality reduction, which facilitates the feature correction operation of the subsequent competitive suppression fusion unit.

[0081] The resulting iron interference state latent vector encodes the evolution characteristics of iron interference across multiple time scales. It includes rapid fluctuations in ferrous ion oxidation, long-term cumulative information on iron alum precipitation, and covers multi-dimensional interference effects such as manganese catalysis and heterogeneous nucleation of aluminosilicates. The iron interference state latent vector has two outputs: one input is to the iron alum deactivation evaluator in the competitive suppression fusion unit, used to calculate the zinc leaching inhibition factor in conjunction with the total active oxygen molar ratio; the other input is to the iron leaching risk assessment module in the process optimization unit, used to generate the iron leaching risk index. This multi-scale, parallel encoding structure, compared to a single-scale gated recurrent network, more comprehensively covers the multi-time scale characteristics of the iron interference process, avoiding the loss of fast-process details or insufficient capture of slow-process trends that can occur at a single scale. This effectively improves the characterization accuracy of the iron interference state and provides a more complete feature foundation for subsequent interference suppression correction and risk assessment.

[0082] The iron alum deactivation evaluator in the competitive suppression fusion unit uses a fully connected network with a dual-channel input fusion architecture to quantify the zinc leaching inhibition factor. The dual-channel input fusion architecture refers to encoding two types of input features with different dimensions and physical meanings through independent feature transformation channels, and then completing the feature combination through a specified fusion operator. This network structure can avoid the low-dimensional scalar features being overwhelmed by high-dimensional temporal features, while preserving the physical independence of the two types of features. One channel receives the iron interference state latent vector output by the iron interference temporal encoding unit, which corresponds to the temporal evolution feature of the iron interference intensity. The other channel receives the total active oxygen molar ratio output by the parameter decoupling unit, which corresponds to the baseline demand for oxidation supply in the system. After the two types of features undergo nonlinear transformation, a mechanism-aligned fusion operation is performed, and finally, a standardized zinc leaching inhibition factor is output.

[0083] The fully connected network first feeds the iron disturbance state latent vector into the first set of fully connected layers for nonlinear transformation to extract the deep feature representation of iron disturbance. A fully connected layer is a network layer where each neuron in the layer establishes a weighted connection with all neurons in the previous layer. Nonlinear mapping of the feature space can be achieved through trainable weights and biases. The first set of fully connected layers contains two layers combining linear transformation and ReLU activation functions. The iron disturbance state latent vector is denoted as... The dimension is the preset feature dimension. The first fully connected layer first maps the original latent vectors to a high-dimensional feature space, and the calculation formula is as follows: ,in This is the weight matrix of the first fully connected layer, with dimension 1. , The corresponding bias vector has a dimension of , This is a linear rectification activation function used to introduce nonlinear transformation capability. The high-dimensional features are those after the first layer transformation, with dimension . .

[0084] The second fully connected layer maps the high-dimensional features back to the feature dimension aligned with the second channel output, facilitating subsequent fusion operations. The calculation formula is as follows: ,in This is the weight matrix of the second fully connected layer, with dimension 1. , The corresponding bias vector has a dimension of , The extracted deep feature representation of iron interference has a dimension of [dimensionality missing]. This deep feature encodes deep interference features such as multi-scale iron ion valence state transformation and iron alum crystallization rate, and is a high-dimensional abstract representation of the intensity of iron interference.

[0085] Meanwhile, the fully connected network feeds the total reactive oxygen species (ROS) molar ratio into the second fully connected layer for nonlinear mapping, extracting a quantitative feature representation of ROS supply capacity. Since the ROS molar ratio is a scalar value with moles per kilogram, it cannot be directly fused with high-dimensional features. Therefore, dimensional normalization preprocessing is first performed, dividing the ROS molar ratio by the maximum historical batch value to obtain a dimensionless oxidation demand normalization coefficient. Simultaneously, the actual oxidant supply is introduced to calculate the oxidant supply-demand ratio, which is the ratio of the actual ROS supply per unit mass of raw material to the total ROS molar ratio, denoted as […]. The calculation formula is ,in This refers to the actual supply of reactive oxygen species, expressed in moles per kilogram. The molar ratio of total reactive oxygen species, expressed in moles per kilogram. This is a dimensionless supply-demand ratio coefficient, typically ranging from 0.5 to 1.5. A higher value indicates a more abundant supply of oxidant. Using this supply-demand ratio coefficient as the actual input to the second channel eliminates dimensional differences while directly reflecting the sufficiency of the oxidant supply, thus better aligning with the reaction mechanism of ferrous sulfate formation.

[0086] The second set of fully connected layers maps the scalar supply-demand ratio to a vector with the same dimension as the deep features of iron disturbance, achieving feature dimension alignment. The calculation formula is as follows: ,in This is the weight matrix of the second channel fully connected layer, with dimension 1. , The corresponding bias vector has a dimension of , The quantitative characteristic representation of reactive oxygen species supply capacity, with dimensions of Each dimension corresponds to the intensity of oxidation supply modulation for different reaction processes.

[0087] After completing the feature transformation of the two channels, the fully connected network performs an element-wise addition operation on the deep feature representation of iron interference and the reactive oxygen species quantification feature representation to form a fused feature vector. Element-wise addition refers to the operation of adding the values ​​at corresponding positions of two vectors with the same dimension. This fusion method differs from conventional channel splicing and can achieve dimension-wise modulation of the iron interference feature by the oxidation supply feature. This aligns with the chemical mechanism that "oxidant sufficiency directly amplifies or weakens the iron alum formation intensity," and possesses stronger physical interpretability. The fusion calculation formula is as follows: ,in To fuse feature vectors, the dimension is This fused feature vector carries information on both the intensity of iron interference and the sufficiency of active oxygen supply, and can comprehensively reflect the inhibitory potential of iron alum precipitation on zinc leaching under the current operating conditions.

[0088] The fully connected network inputs the fused feature vectors to the output layer, which has a Sigmoid activation function. The Sigmoid activation function is a non-linear function that maps any real value to a continuous interval between 0 and 1. Its function expression is: The output values ​​naturally conform to the required range of the strength coefficient.

[0089] The output layer maps the fused feature vector to a single variable through a linear transformation, and then obtains the final zinc leaching inhibition factor through Sigmoid activation. The calculation formula is as follows: ,in The weight vector of the output layer has a dimension of . , The output layer bias is denoted by , which is a scalar. It is the Sigmoid activation function. The zinc leaching inhibition factor is a dimensionless numerical value, strictly ranging between zero and one. The magnitude of the zinc leaching inhibition factor directly reflects the intensity of the inhibition of the zinc leaching process by the iron interference effect at the current moment. The closer the value is to one, the stronger the mass transfer hindrance and reactant consumption effect of iron alum precipitation, and the more significant the inhibition of the zinc leaching rate. The closer the value is to zero, the weaker the iron interference effect, with almost no impact on the zinc leaching process. The calculated zinc leaching inhibition factor is transmitted to the feature corrector to perform element-wise attenuation modulation on the latent vector of the zinc leaching state, eliminating the influence of the iron interference effect on the zinc leaching characteristics. This dual-channel mechanism fusion evaluation architecture, compared to a single-channel network structure that directly splices the input, can more accurately capture the coupling relationship between oxidation supply and iron interference. Dimensional normalization and supply-demand ratio preprocessing ensure the physical consistency of the input features. The element-wise addition fusion method improves the interpretability and convergence efficiency of the model. The final output zinc leaching inhibition factor has a clear physical meaning and can be directly used for subsequent feature correction and process control reference.

[0090] After obtaining the zinc leaching inhibition factor, the feature corrector within the competitive inhibition fusion unit performs interference cancellation processing on the latent vector of the zinc leaching state. The feature corrector adopts an element-wise attenuation modulation method based on an attention mechanism, combined with residual connections and gated linear units for secondary screening, to achieve layered stripping of the iron interference effect, and finally outputs the feature vector corresponding to the pure zinc leaching state. Element-wise attenuation modulation refers to the processing method of numerically attenuating each feature dimension of the latent vector of the zinc leaching state according to the corresponding inhibition ratio. The correction process first copies and expands the dimensionless scalar zinc leaching inhibition factor, whose value ranges from zero to one, into a zinc leaching inhibition factor vector with the same dimension as the latent vector of the zinc leaching state. The expansion process involves filling the scalar value into a vector with the same length as the feature dimension, where the element value at each position in the vector is equal to the value of the original zinc leaching inhibition factor. The feature dimension of the latent vector of the zinc leaching state is denoted as . The expanded zinc leaching inhibition factor vector also has a dimension of [missing value]. To ensure dimension matching for subsequent element-wise operations, the feature retention ratio corresponding to each dimension is calculated. The retention ratio is calculated by subtracting the value of the corresponding element in the zinc leaching inhibition factor vector from one, forming a retention ratio vector with the same dimension as the zinc leaching state latent vector. The physical meaning of the retention ratio vector is the proportion of pure zinc leaching features remaining after deducting the iron interference inhibition effect. The larger the value of the zinc leaching inhibition factor, the smaller the corresponding retention ratio value.

[0091] After calculating the retention ratio, each element in the zinc leaching state latent vector is multiplied element-wise with the corresponding element in the retention ratio vector to obtain the preliminary corrected intermediate feature vector. The calculation formula is as follows: ,in This is a vector of zinc leaching inhibition factors. This is the hidden vector for the zinc leaching state. For the whole vector, This represents element-wise multiplication. This is the intermediate feature vector after preliminary correction.

[0092] The initial correction directly strips away the feature components corresponding to the iron interference intensity using a linear proportional decay method, completing the first layer of coarse-grained correction. The corrected intermediate feature vector has removed most of the feature components directly related to iron interference. After the initial correction, the feature corrector introduces a residual connection mechanism. Residual connections are skip connection structures that add the original input features to the corrected output features element-wise. This avoids the loss of basic feature information about the original zinc leaching during the correction process, while also mitigating the vanishing gradient phenomenon during network training and improving the model's convergence stability. The specific operation is as follows: The original zinc leaching state latent vector is added element-wise to the pre-corrected intermediate eigenvector to form the residual-enhanced eigenvector. The calculation formula is as follows: ,in This is the feature vector after residual enhancement. The introduction of residual connections allows the correction process to retain the complete backbone information of the original zinc leaching features. The correction component is superimposed on the original features as a supplement, avoiding excessive feature distortion caused by deviations in the calculation of the zinc leaching inhibition factor, and improving the robustness of the correction process.

[0093] After residual enhancement, the feature corrector inputs the enhanced feature vector into a gated linear unit with a smaller parameter size. A gated linear unit is a linear transformation structure with a gating mechanism. It adaptively filters each dimension of the input features using learnable gating weights, specifically filtering residual nonlinear interference noise. Compared to fixed-ratio attenuation, it has stronger adaptability to different operating conditions. The gated linear unit consists of two parallel branches: a feature transformation branch and a gating branch. Both branches share the same input. The feature transformation branch spatially maps the input features using a linear weight matrix, extracting zinc leaching-related feature components. The gating branch outputs gating coefficients for each dimension using another set of linear weights combined with a sigmoid activation function. The gating coefficients range from zero to one, representing the degree of feature retention in the corresponding dimension. A gating coefficient closer to one indicates a higher degree of matching between the corresponding dimension's features and the pure zinc leaching features; a gating coefficient closer to zero indicates a higher proportion of interference components in the corresponding dimension's features.

[0094] The specific calculation formula for the gated linear unit is as follows: ,in and The weight matrix and bias vector of the feature transformation branch are given. and Here are the weight matrix and bias vector for the gated branch. This is the sigmoid activation function, used to map the output of the gated branch to the interval between zero and one. The output vector after gating, which is the final zinc leaching feature vector, is obtained by the gated linear unit learning the feature patterns of residual iron interference noise through a data-driven approach. It adaptively closes the feature channels corresponding to the interference dimensions, removes residual iron interference noise components from the residual enhanced features, and completes the second layer of fine-grained correction. Each element value in the final output zinc leaching feature vector represents the pure zinc leaching state characteristics unaffected by iron interference. It fully preserves the temporal characteristics of zinc leaching driving factors such as zinc substrate supply, acidity changes, and oxidation conditions, while simultaneously removing the interference caused by iron valence state transformation and iron alum precipitation. The disturbance component, the zinc leaching feature vector, is output to the leaching rate prediction unit as an input feature for regression prediction, used to generate the final zinc leaching rate prediction value. This hierarchical correction structure differs from the traditional single proportional attenuation correction method. Through three levels of processing—coarse-grained linear attenuation, residual information preservation, and fine-grained gating screening—it balances the consistency of the correction mechanism with data adaptability. It not only ensures the interpretability of the chemical mechanism of the zinc leaching inhibition factor, but also improves the accuracy of iron interference elimination by adapting to the nonlinear interference effect under complex working conditions through the gating unit, providing a feature basis for subsequent leaching rate prediction.

[0095] The leaching rate prediction unit receives the zinc leaching feature vector output by the competition suppression fusion unit and uses a hybrid expert regression network to perform regression prediction of the zinc leaching rate. The hybrid expert regression network is an integrated regression model composed of multiple independent regression expert branches and a gating assignment network. Different experts learn the leaching rate mapping pattern of different operating conditions. The gating network dynamically assigns the prediction weights of each expert according to the distribution of input features, which can adapt to the nonlinear mapping relationship of multiple stages of the leaching process and cover a wider range of raw materials and process parameters.

[0096] The hybrid expert regression network has a total of 100,000 members. An independent regression expert, The value of is determined to be 3 based on the stages of the leaching process, corresponding to three working conditions: rapid leaching, stable leaching, and saturated leaching. Each regression expert consists of a fully connected layer, which contains a trainable weight vector and a bias scalar. It is responsible for mapping the input zinc leaching feature vector to a preliminary zinc leaching rate prediction value. The parameters of the fully connected layers of different regression experts are different and are updated independently during training, learning the leaching rate mapping rules under the corresponding working conditions.

[0097] No. The forward computation process of the regression expert involves performing an inner product operation between the zinc leaching feature vector and the corresponding weight vector, and then superimposing the bias scalar to obtain the preliminary predicted value. The calculation formula is as follows: ,in The input is the zinc leaching feature vector, with the dimension being the preset feature dimension. , For the first The weight vectors of the regression experts have the same dimensions as the feature dimensions. , For the first The bias scalar of a regression expert. For the first The preliminary zinc leaching rate prediction output by each expert, expressed as a percentage, represents the zinc leaching ratio under the corresponding expert mode.

[0098] The gated network takes the same zinc leaching feature vector as input and outputs a set of weight coefficients through a fully connected transformation and a Softmax activation function. Each weight coefficient corresponds to the confidence of a regression expert. The sum of the weight coefficients is one, and the values ​​range from zero to one.

[0099] The gating network computation consists of two steps. The first step is raw score calculation, where the input features are mapped to a K-dimensional raw score vector through a fully connected layer. The score in each dimension corresponds to the degree of fit between the expert and the current working condition. The calculation formula is as follows: ,in Here is the weight matrix of the gated network, with dimension 1. OK List, Let be the bias vector of the gated network, with dimension . , The original score vector has dimensions of . The second step is weight normalization, which uses the Softmax activation function to normalize the original scores. The Softmax function maps any real value to a probability distribution with a sum of 1, thus converting the fit into weight coefficients. The formula for calculating the weighting coefficient of each expert is as follows: ,in The first element in the original score vector The value of each element. For the first The weight coefficients corresponding to each regression expert are dimensionless values, and the sum of the weight coefficients of all experts is always equal to one.

[0100] After obtaining the preliminary predictions and corresponding weights from each expert, the leaching rate prediction unit performs a weighted summation operation. It multiplies the preliminary zinc leaching rate predictions output by each regression expert by their corresponding weight coefficients and then sums them to obtain the final zinc leaching rate prediction. The calculation formula is as follows: ,in The final output is the predicted zinc leaching rate, expressed as a percentage. Its value range corresponds to the actual leaching rate interval, representing the proportion of zinc leached at the current moment. During training, this hybrid expert regression network simultaneously updates the parameters of all regression experts and the gated network via backpropagation. Supervised learning is used, and the training samples consist of zinc leaching feature vectors from historical leaching batches and corresponding measured leaching rate labels. The loss function uses mean squared error loss, and an expert diversity regularization term is introduced to avoid functional homogenization among multiple experts and ensure clear division of labor among them. The total loss function is calculated using the following formula: ,in This label shows the measured leaching rate, expressed as a percentage. The first term is the regularization coefficient, the second is a dimensionless constant, and the third term is the information entropy of the weight distribution. This entropy guides the gating network in assigning expert weights, preventing all samples from relying on a single expert. The backpropagation process is based on the chain rule. First, the gradient of the total loss function with respect to the final predicted value is calculated. Then, the gradient is propagated backward along two paths. One path is to each regression expert, calculating the gradient of the loss with respect to each expert's weight vector and bias scalar, and updating the corresponding parameters according to the gradient descent rule. The other path is to the gating network, calculating the gradient of the loss with respect to the gating weight matrix and bias vector, and updating the gating parameters synchronously. After multiple rounds of iterative training, the gating network can automatically identify the current stage of the leaching process based on the different feature distributions of the zinc leaching feature vector, dynamically increasing the weights of suitable experts and decreasing the weights of unsuitable experts, thus achieving automatic selection of the optimal prediction mode. The final predicted zinc leaching rate is transmitted to the process optimization unit as the objective function for process parameter optimization. It can also be output to the on-site monitoring terminal to provide operators with real-time predictions of the leaching effect.

[0101] The iron leaching risk assessment module inside the process optimization unit receives the iron interference state latent vector output by the iron interference timing coding unit, and performs risk quantification assessment using a multilayer perceptron classification network. The multilayer perceptron classification network is a feedforward neural network composed of multiple fully connected layers stacked together. It can extract deep features related to risk from the high-dimensional latent vector through multilayer nonlinear transformation, and output a quantifiable risk index to provide a safety constraint boundary for subsequent process parameter search.

[0102] The iron disturbance state latent vector is a high-dimensional temporal feature vector of a preset dimension, encoding dynamic information such as iron valence state transformation and iron alum crystallization accumulation at multiple time scales. It is directly used as the input feature of a multilayer perceptron (MLP) classification network. The MLP classification network consists of two fully connected hidden layers and one output layer. Each hidden layer is followed by a linear rectified activation function to introduce nonlinear transformation capability. The first hidden layer maps the input latent vector to a high-dimensional feature space, extracting basic risk features such as the trend of iron ion concentration change and the completion degree of iron ion valence state transformation. The trend of iron ion concentration change corresponds to the direction of increase or decrease in the iron dissolution rate within the sampling window, and the completion degree of iron ion valence state transformation corresponds to the reaction progress of ferrous ions oxidizing to ferric ions. The calculation expression for the first hidden layer is as follows: ,in Let be the input iron disturbance state latent vector. This is the weight matrix of the first hidden layer. This is the bias vector of the first hidden layer. It is a linear rectified activation function. This is the basic risk feature vector output by the first hidden layer.

[0103] The second hidden layer further applies a nonlinear mapping to the basic risk characteristics to extract deeper risk features such as the rate of iron alum precipitation. The rate of iron alum precipitation directly affects the thickening rate of the precipitation layer and the solid-liquid partition ratio of iron ions, and is a core indicator determining the level of iron leaching risk. The calculation expression for the second hidden layer is as follows: ,in This is the weight matrix of the second hidden layer. This is the bias vector for the second hidden layer. This is the deep risk feature vector output by the second hidden layer.

[0104] The output layer of the multilayer perceptron classification network uses a linear activation function. Linear activation functions do not truncate the output, allowing for direct output of continuous values. The output layer maps the deep risk feature vector to a single-valued original iron leaching risk index, calculated as follows: ,in The weight vector of the output layer. This is the bias scalar for the output layer. This is the original iron leaching risk index.

[0105] During network training, the risk level label corresponding to the measured iron leaching concentration is used as the monitoring signal. The label value range is preset to be between zero and one hundred. Therefore, the original risk index value output after training naturally falls between zero and one hundred. The higher the value, the higher the concentration of free iron ions in the liquid phase, the greater the risk of iron leaching exceeding the standard, and the higher the iron removal load in the subsequent process.

[0106] To eliminate systematic assessment bias caused by differences in elemental composition among different raw material batches, the iron leaching risk assessment module introduces a sliding window calibration mechanism based on historical data when generating the iron leaching risk index. This mechanism uses the risk assessment results of multiple consecutive leaching batches as statistical samples, and offsets baseline drift caused by fluctuations in the overall composition of the raw materials through benchmark offset correction between batches. The specific execution process is as follows: The iron leaching risk assessment module maintains a fixed-length sliding window queue, with a queue length of twenty. The queue stores the original iron leaching risk index values ​​for the most recent twenty normal production batches in chronological order. After the original risk index for each new batch is calculated, the average value of all historical samples within the current window is first calculated as the batch baseline value. The calculation formula is as follows: ,in The total length of the sliding window is 20. For the first in the window The original iron leaching risk index of each historical batch. The average baseline value for the sliding window is used, with the same dimensions as the original risk index, and its value ranges from zero to one hundred. The baseline offset is calculated as the difference between the current window's average baseline value and the system's preset standard baseline value. The standard baseline value is the average risk level corresponding to normal raw materials obtained from long-term production statistics, and is fixed at thirty. The offset calculation formula is as follows: ,in As the standard reference value, The offset is the baseline. A positive offset indicates that the overall iron content of the current batch of raw materials is too high, and the original risk index is systematically overestimated. A negative offset indicates that the iron content of the current raw materials is too low, and the original risk index is systematically underestimated.

[0107] Finally, offset correction is performed on the original risk index of the current batch to obtain the final iron leaching risk index. The correction calculation formula is as follows: ,in The final corrected iron leaching risk index is a dimensionless score, ranging from zero to one hundred. After correction, the iron leaching risk assessment module writes the original risk index of the current batch to the tail of the sliding window queue, while removing the oldest batch data from the head of the queue, completing the sliding update of the window to ensure that the window always covers the most recent production batch, adapting to the slow changing trend of raw material composition. The final iron leaching risk index serves as a safety constraint for process optimization and is input into a multi-stage progressive search engine. In each round of process parameter combination evaluation, the iron leaching risk index corresponding to the current parameter combination is first compared with the preset safety threshold. The safety threshold can be set according to the processing capacity of subsequent iron removal processes, typically set to forty. If the iron leaching risk index... If the leaching risk index is not lower than the safety threshold, it indicates that the iron leaching amount exceeds the standard under the corresponding process parameters, which will increase the processing burden of subsequent processes. Therefore, this combination of process parameters should be discarded directly. If the iron leaching risk index is lower than the safety threshold, it indicates that the corresponding process parameters meet the iron leaching safety constraints. This combination should be retained to participate in the subsequent optimization of leaching rate maximization. This risk assessment method, which combines sliding window baseline calibration, can adapt to the overall composition fluctuation of raw material batches compared with the evaluation method of fixed parameter models. It avoids risk misjudgment caused by systematic bias. The nonlinear feature extraction capability of the multilayer perceptron can accurately capture the mapping relationship between multi-dimensional iron interference features and leaching risk, providing a reliable safety boundary for process optimization and ensuring that the optimized process parameters simultaneously meet the dual requirements of improving leaching efficiency and controlling iron impurities.

[0108] The process optimization unit is equipped with a multi-stage progressive search engine. This multi-stage progressive search engine adopts a two-stage search strategy from coarse to fine. First, it narrows down the region where the optimal parameters are located through a large-scale discrete search, and then performs fine optimization within a small range. The multi-stage progressive search engine first reads a preset set of initial values ​​for process parameters from the storage space of the process optimization unit. The set of initial values ​​includes feasible ranges for leaching temperature, leaching time, sulfuric acid concentration, and oxidant flow rate. The feasible range for leaching temperature is 70°C to 95°C, the feasible range for leaching time is 60 minutes to 180 minutes, the feasible range for sulfuric acid concentration is 15 g / L to 60 g / L, and the feasible range for oxidant flow rate is 0.5 m / h to 2.0 m / h. The feasible ranges for all parameters are determined based on the conventional operating range of the hydrometallurgical zinc smelting industry, and all candidate combinations are practically feasible.

[0109] The multi-stage progressive search engine performs a first-stage coarse-grained search. This coarse-grained search traverses the main regions of the parameter space, locating parameter intervals with higher leaching rates, providing candidate seeds for subsequent fine-grained searches. The first-stage search uniformly divides the feasible range of each process parameter into multiple discrete candidate points, with the number of discrete points uniformly set to five. The parameter space coverage and computational complexity meet preset requirements. The first process parameter The formula for calculating the discrete candidate points is: ,in For the first The lower bound of the feasible range of each parameter. For the first The upper limit of the feasible range of each parameter. The number of discrete points for each parameter, taking a value of five. The value range is from one to five. For the first The parameter of the first There are 5 discrete candidate points, with dimensions consistent with the corresponding parameters. After completing the discrete point division of all parameters, the search engine performs full permutation and combination of all discrete candidate points of the four parameters. Each combination consists of one discrete point from each of the four parameters, resulting in a total of 5 to the power of 4, or 625 groups. Each group corresponds to an independent initial process parameter combination. For each initial process parameter combination, the process optimization unit calls the parameter decoupling unit, zinc leaching timing encoding unit, iron interference timing encoding unit, competition suppression fusion unit, and leaching rate prediction unit to perform forward calculation. Forward calculation refers to the forward operation process from input features to output results according to the data flow direction. The calculation process first generates the driving parameter timing sequence under the corresponding operating conditions based on the current process parameters. When generating the timing sequence, the elemental analysis data of the raw materials in the current batch is used as the initial substrate concentration. A reaction rate correction coefficient is introduced to simulate the influence of process parameters on the reaction process. The rate correction coefficient is obtained by coupling the temperature correction term, acidity correction term, and oxidant correction term. The temperature correction term is calculated based on the Arrhenius formula, with 85 degrees Celsius as the reference temperature. The calculation formula for the temperature correction coefficient is as follows: ,in The apparent activation energy for the zinc leaching reaction is expressed in joules per mole, and R is the ideal gas constant, which is 8.314 joules per mole per Kelvin. The base temperature is set at 358.15 Kelvin, which is 85 degrees Celsius. The leaching temperature corresponding to the current process combination is expressed in Kelvin. This is a dimensionless temperature correction factor. The acidity correction factor is calculated based on the first-order kinetic relationship between sulfuric acid concentration and reaction rate, using the following formula: ,in This represents the sulfuric acid concentration corresponding to the current process combination, in grams per liter. The baseline sulfuric acid concentration is 30 grams per liter. This is a dimensionless acidity correction factor. The oxidant correction factor is calculated based on the direct proportionality between oxidant flow rate and active oxygen supply rate, while also introducing a coupling coefficient to reflect the effect of temperature on oxidant utilization. The calculation formula is as follows: ,in This represents the oxidant flow rate corresponding to the current process combination, in cubic meters per hour. The baseline oxidant flow rate is set at 1.0 cubic meters per hour. This is a dimensionless correction factor for the oxidant.

[0110] The overall reaction rate correction factor is the product of the three partial factors, calculated using the following formula: The overall reaction rate correction factor is used to adjust the temporal change rate of the zinc leaching driving parameter set and the iron interference driving parameter set. The larger the rate factor, the faster the corresponding reaction process progresses, and the higher the substrate consumption and product generation within the same time period.

[0111] Based on the total reaction rate correction coefficient and leaching time parameter, a time series matrix of zinc leaching driving parameters and iron interference driving parameters under the corresponding process conditions is generated. The time step size of the time series matrix is ​​consistent with the original sampling step size, and the total number of steps is determined by the ratio of leaching time to step size. After the time series matrix is ​​generated, the data flows through each functional unit to perform forward operations. First, the parameter decoupling unit completes parameter verification and calculates the total active oxygen molar ratio. Then, the data is input into the zinc leaching time series encoding unit and the iron interference time series encoding unit to extract the corresponding state latent vectors. Subsequently, the data enters the competition suppression fusion unit to generate the zinc leaching feature vector, and one input is sent to the leaching rate prediction unit to obtain the current process data. The predicted zinc leaching rate for each process combination is input into the iron leaching risk assessment module to obtain the iron leaching risk index for the current process combination. After all combinations are calculated in the first stage, the search engine first performs a safety constraint screening, discarding combinations whose iron leaching risk index is not lower than the preset safety threshold, and retaining all combinations that meet the safety conditions. Then, the three combinations with the highest predicted zinc leaching rate are selected as candidate seeds and input into the second stage fine-grained search module. The two-stage progressive search method has a lower computational load than the global fine search. The coarse-grained stage achieves full coverage of the parameter space, avoids missing the optimal area, and provides a starting interval for subsequent local fine optimization.

[0112] After completing the forward calculation of all initial process parameter combinations in the first stage, the multi-stage progressive search engine performs a safety constraint screening operation. The safety threshold is pre-set to forty based on the processing capacity of the subsequent iron removal process. It is a dimensionless risk score, and the higher the value, the greater the risk of iron leaching exceeding the standard. If the iron leaching risk index corresponding to a certain initial process parameter combination is not lower than the safety threshold, it means that the liquid phase iron ion concentration under the corresponding parameters exceeds the processing load of the subsequent process, and the corresponding initial process parameter combination is directly discarded. If the iron leaching risk index is lower than the safety threshold, the corresponding initial process parameter combination and its corresponding zinc leaching rate prediction value are retained and enter the subsequent sorting stage.

[0113] After the first stage, the search engine sorts all remaining feasible combinations from highest to lowest zinc leaching rate prediction value and selects the three initial process parameter combinations with the highest zinc leaching rate prediction value as candidate seeds. Each candidate seed corresponds to a parameter neighborhood with a high leaching rate, which serves as the starting center point for the second stage of fine-grained search. The multi-starting point seed selection method can avoid the single-region search from getting stuck in local optima and improve the reliability of global optimization.

[0114] The multi-stage progressive search engine performs a second-stage fine-grained search. In this second stage, a particle swarm optimization algorithm is used to perform a local search within the neighborhood of each candidate seed's corresponding process parameter value. The particle swarm optimization algorithm is a stochastic optimization algorithm based on swarm intelligence. It searches for the optimal solution in the parameter space by simulating the cooperative behavior of the group. Each particle in the algorithm represents a set of four-dimensional process parameter combinations, including leaching temperature, leaching time, sulfuric acid concentration, and oxidant flow rate. The particle's position coordinates in the parameter space correspond to a specific set of process parameter values, and the particle's movement speed corresponds to the direction and magnitude of parameter adjustment.

[0115] To eliminate the interference of different parameter dimensions on the search process, the algorithm first maps all process parameters to a normalized parameter space of zero to one to perform the search operation. After the search is completed, it then reverses the normalization and maps them back to the original dimensions. The normalization adopts a minimum-maximum mapping method. The formula for calculating the normalized value of each parameter is as follows: ,in These are the original parameter values. and These represent the upper and lower limits of the feasible range for the corresponding parameters. The normalized parameter values ​​are dimensionless and range from zero to one. The neighborhood of each candidate seed is a hypercube region centered on the seed coordinates with a side length of 0.2 within the normalized space. All particles randomly initialize their initial positions and velocities within their corresponding hypercube regions. Each subpopulation contains twenty particles to balance search coverage and computational efficiency. The particle velocity update direction is guided by both the current particle's historical best position and the group's historical best position. The algorithm introduces linearly decreasing adaptive inertial weights to balance global exploration and local exploitation capabilities. Larger weights in the early iterations ensure global search breadth, while smaller weights in the later iterations improve local convergence accuracy. The update formula for the inertial weights is... ,in This represents the current iteration number. The maximum number of iterations is set to twenty. The initial inertia weight is set to 0.9. The final inertia weight is set to 0.4. For the first The inertia weight for each iteration is a dimensionless constant.

[0116] No. The particle in the first The velocity update formula for the next iteration is: ,in For the first During the nth iteration The position vectors of each particle. For the corresponding velocity vector, For the first The historical best position vector of a particle represents the parameter position with the highest fitness found during the search process of a single particle. This is the population's historical best position vector, representing the globally optimal parameter position found by the entire population. and The learning factor, with a value of two, controls the weights of individual experience and group experience on the speed update, respectively. and The random number between zero and one is generated independently for each iteration to increase the randomness of the search process and prevent the algorithm from converging to a local optimum too early. After the speed update is completed, the position update is performed, as shown in the formula. After the position is updated, boundary constraint processing is performed. If the particle position exceeds the zero-to-one range of the normalized parameter space, the value of the corresponding dimension is truncated to the boundary value, and the velocity of the corresponding dimension is multiplied by 0.5 in reverse to prevent particles from continuously flying out of the feasible region, ensuring that all candidate parameters are within the industrially feasible range. In each iteration, the process optimization unit denormalizes the normalized position of each particle to the original process parameter combination, and calls the parameter decoupling unit, zinc leaching timing coding unit, iron interference timing coding unit, competition suppression fusion unit, and leaching rate prediction unit to perform forward calculations, obtaining the predicted zinc leaching rate and iron leaching risk index of the corresponding particle. Then, safety constraint filtering is performed, retaining only particles with an iron leaching risk index lower than the safety threshold. For infeasible particles with an iron leaching risk index not lower than the safety threshold, their fitness value is set to zero to ensure that infeasible solutions are not selected as the optimal position. The particle fitness function directly uses the predicted zinc leaching rate value; the higher the value, the better the leaching effect of the corresponding process combination. The fitness calculation formula is as follows: ,in For the first Predicted zinc leaching rate for each particle To correspond to the iron leaching risk index, As a safety threshold, For the first The fitness value of each particle.

[0117] After calculating the fitness of all particles, the individual's best historical position and the population's best historical position are updated simultaneously. If the current particle's fitness is higher than its own best historical fitness, the current position is updated to the corresponding particle's new best historical position. If there is a particle in the current population with a fitness higher than the population's best historical fitness, the corresponding particle's position is updated to the new population's best historical position. This iteration continues until the preset maximum number of iterations is reached, or the leaching rate of the population's best historical position increases by less than one-thousandth in five consecutive iterations. At this point, the search is considered converged, and the iteration terminates. After this, each candidate seed's corresponding subpopulation outputs its own population's best historical position and corresponding leaching rate. The search engine compares the optimal results of all subpopulations and selects the best position. The set of leaching rates that meets the safety constraints is taken as the global optimal solution. The process parameter combination corresponding to the optimal position is the process parameter combination that maximizes the predicted zinc leaching rate while meeting the safety conditions for iron leaching risk. The process optimization unit finally outputs the process parameter combination that maximizes the predicted zinc leaching rate to an external display device for on-site operators to refer to and adjust the process parameters. This two-stage progressive search architecture, which combines coarse-grained global coverage and fine-grained local optimization, has both computational efficiency and optimization accuracy compared to a single intelligent algorithm search. The multi-starting-point particle swarm search can effectively avoid local optima problems, and the fitness mechanism with safety constraints ensures that all search results meet production safety requirements, achieving the dual goals of improving leaching efficiency and controlling iron impurities.

[0118] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based zinc leaching rate prediction and process optimization system, characterized in that, include: The parameter decoupling unit is used to receive the raw material element analysis data of the wet zinc leaching process. Based on the chemical mechanism of zinc leaching and iron interference, the raw material element analysis data is divided into the zinc leaching driving parameter set and the iron interference driving parameter set. The total active oxygen molar ratio is calculated based on the molar content of iron and zinc in the raw material. The zinc leaching timing coding unit uses a first gated recurrent network to extract the zinc leaching state latent vector by taking the time series of the zinc leaching driving parameter set as input. The iron interference timing coding unit uses a second gated recurrent network, taking the time series of the iron interference driving parameter set as input, to extract the iron interference state latent vector. The competitive inhibition fusion unit includes an iron alum inactivation evaluator and a feature corrector. The iron alum inactivation evaluator takes the iron interference state latent vector and the total active oxygen molar ratio as inputs and generates a zinc leaching inhibition factor through a fully connected network. The feature corrector uses the zinc leaching inhibition factor to perform element-wise attenuation modulation on the zinc leaching state latent vector and outputs a zinc leaching feature vector with the iron interference effect eliminated. The leaching rate prediction unit performs regression mapping on the zinc leaching feature vector to obtain the predicted zinc leaching rate value at the current moment. The process optimization unit is connected to the leaching rate prediction unit and the iron interference timing coding unit. It has an iron leaching risk assessment module, which is used to generate an iron leaching risk index based on the iron interference state latent vector. Under the condition that the iron leaching risk index is lower than the safety threshold, it searches for and outputs the combination of leaching process parameters with the goal of maximizing the zinc leaching rate prediction value.

2. The zinc leaching rate prediction and process optimization system based on deep learning according to claim 1, characterized in that: Based on the chemical mechanism of zinc leaching and iron interference, the raw material elemental analysis data are divided into a zinc leaching driving parameter set and an iron interference driving parameter set, specifically including: The chemical mechanisms of zinc leaching and iron interference include the zinc leaching reaction mechanism and the iron interference reaction mechanism. The zinc leaching reaction mechanism refers to the process in which zinc oxides or sulfides react with sulfuric acid to form soluble zinc sulfate under acidic leaching conditions. This process depends on the concentration of hydrogen ions and the supply of oxidant. The interference mechanism of iron refers to the fact that under the same acidic leaching environment, iron exists in the form of ferrous ions and is oxidized to ferric ions by oxidants. Subsequently, ferric ions combine with water or sulfate to form precipitates of potassium ferric alum. These precipitates adhere to the surface of zinc mineral particles, forming a physical barrier that hinders the contact between hydrogen ions and zinc minerals. At the same time, they consume hydrogen ions and oxidants to inhibit the leaching rate of zinc. The parameter decoupling unit, based on the chemical mechanism of zinc leaching and iron interference, classifies the elemental components related to hydrogen ion concentration, oxidant supply, and zinc mineral solubility in the raw material elemental analysis data, including the content of zinc, sulfur, and calcium and magnesium carbonates, into the zinc leaching driving parameter set. The elemental components related to the valence state transformation of iron and the precipitation of iron alum in the raw material elemental analysis data, including the content of iron, manganese and aluminosilicate, are classified into the iron interference driving parameter set. The classification method employs a chemometrics-based feature allocation algorithm. This algorithm pre-establishes an elemental composition comparison table for the zinc leaching driving parameter set and the iron interference driving parameter set. It matches the chemical properties of each element in the raw material elemental analysis data with the classification labels in the elemental composition comparison table. If the chemical properties of a certain element are consistent with the key reactants or products in the zinc leaching reaction mechanism, then the element is assigned to the zinc leaching driving parameter set. If the chemical properties of a certain element are consistent with the key reactants or products in the iron interference reaction mechanism, then the element is assigned to the iron interference driving parameter set. This process ultimately completes the classification of the zinc leaching driving parameter set and the iron interference driving parameter set.

3. The zinc leaching rate prediction and process optimization system based on deep learning according to claim 2, characterized in that: The calculation of the total active oxygen molar ratio based on the molar content of iron and zinc in the raw materials specifically includes: The calculation of the total active oxygen molar ratio is based on the combined demand for active oxygen from both the iron ion oxidation reaction and the zinc ion leaching reaction. Active oxygen refers to the effective oxygen atoms participating in the redox reaction. The calculation process is as follows: The molar contents of iron and zinc were extracted from the elemental analysis data of the raw materials. Based on the active oxidative stoichiometry required for iron ions to oxidize from ferrous ions to ferric ions and the active oxidative stoichiometry required for the dissolution of zinc minerals, an expression for the total demand for active oxygen was established. The parameter decoupling unit introduces a dynamic balance correction coefficient, which is updated online based on the ratio of actual active oxygen consumption to theoretical active oxygen consumption in historical leaching batches, in order to compensate for deviations in active oxygen utilization. The molar content of iron is multiplied by the corresponding active oxidant stoichiometry, then the molar content of zinc is multiplied by the corresponding active oxidant stoichiometry, and finally multiplied by the dynamic equilibrium correction coefficient to obtain the total active oxygen molar ratio. The total active oxygen molar ratio is used as a quantitative indicator to measure whether the oxidant supply is sufficient in the iron alum deactivation evaluator.

4. The zinc leaching rate prediction and process optimization system based on deep learning according to claim 2, characterized in that: The zinc leaching timing coding unit employs a first gated recurrent network, taking the time series of the zinc leaching driving parameter set as input, to extract the zinc leaching state latent vector, specifically including: The first gated recurrent network adopts an attention-enhanced gated recurrent network structure, which introduces an adaptive time step selection mechanism on the basis of the standard gated recurrent network. The zinc leaching timing coding unit arranges the elemental composition data of each sampling moment in the zinc leaching driving parameter set in chronological order to form a multidimensional time series matrix, and inputs the multidimensional time series matrix into the attention-enhanced gated recurrent network. The attention mechanism inside the attention-enhanced gated recurrent network weights the feature data of each time step in the time series. The weight is determined by the contribution of the feature data of the time step to the current zinc leaching state. The contribution is obtained by calculating the similarity between the feature data of the time step and the previous hidden state. The adaptive time step selection mechanism adjusts the span of the time step according to the distribution of attention weights. When the attention weights are concentrated in the recent time step, the time step is shortened to capture the leaching dynamics. When the attention weights are scattered over a preset time range, the time step is automatically extended to integrate the leaching trend. After iterative processing by the attention-enhanced gated recurrent network, the hidden state output at the last time step is the zinc leaching state hidden vector, which encodes the overall characteristics of the evolution of the zinc leaching driving parameters over time.

5. The zinc leaching rate prediction and process optimization system based on deep learning according to claim 4, characterized in that: The iron interference timing coding unit employs a second gated recurrent network, taking the time series of the iron interference driving parameter set as input, to extract the iron interference state latent vector, specifically including: The second gated recurrent network adopts a multi-scale gated recurrent network structure, which contains multiple parallel sub-gated recurrent networks, each with a different expansion coefficient. The iron interference timing coding unit arranges the element component data of each sampling time in the iron interference driving parameter set in chronological order to form a multidimensional time series matrix, and inputs the multidimensional time series matrix into each sub-gated recurrent network. Each sub-gated recurrent network downsamples the time series with a set expansion coefficient. The expansion coefficient determines the number of sampling intervals of the sub-gated recurrent network on the time axis, so that different sub-gated recurrent networks focus on the iron ion behavior patterns at different time scales. The iron interference temporal coding unit concatenates the hidden states output by all sub-gated recurrent networks to form a high-dimensional combined hidden state. Then, a fully connected dimensionality reduction layer compresses the combined hidden state to a preset dimension, finally obtaining the iron interference state latent vector.

6. The zinc leaching rate prediction and process optimization system based on deep learning according to claim 5, characterized in that: The fully connected network in the iron alum deactivation evaluator adopts a dual-channel input fusion architecture, with one channel receiving the iron interference state latent vector and the other channel receiving the total active oxygen molar ratio. The fully connected network performs a nonlinear transformation on the iron disturbance state latent vector through the first set of fully connected layers to extract the deep feature representation of iron disturbance. At the same time, it performs a nonlinear mapping on the total reactive oxygen species molar ratio through the second set of fully connected layers to extract the quantitative feature representation of reactive oxygen species supply capacity. The fully connected network performs an element-wise addition operation on the deep feature representation and the quantized feature representation to form a fused feature vector, which carries information on both the intensity of iron interference and the sufficiency of reactive oxygen supply. The fully connected network inputs the fused feature vector to the output layer with a Sigmoid activation function. The output layer maps the fused feature vector to a value between zero and one, which is the zinc leaching inhibition factor. The magnitude of the zinc leaching inhibition factor reflects the intensity of the inhibition of the zinc leaching process by the iron interference effect at the current moment.

7. The zinc leaching rate prediction and process optimization system based on deep learning according to claim 6, characterized in that: The feature corrector employs an element-wise attenuation modulation method based on an attention mechanism. This method replicates and expands the zinc leaching inhibition factor into a zinc leaching inhibition factor vector of the same dimension as the zinc leaching state latent vector. The difference between each element in the zinc leaching state latent vector and the corresponding element in the zinc leaching inhibition factor vector is calculated. The difference is calculated as follows: Subtract the difference between the values ​​of the corresponding elements in the zinc leaching inhibition factor vector, and then multiply each element in the zinc leaching state latent vector by the difference to obtain the intermediate feature vector after preliminary correction. The feature corrector further introduces a residual connection mechanism, which adds the zinc leaching state latent vector to the intermediate feature vector after preliminary correction element by element to form the feature vector after residual enhancement. Finally, the feature corrector inputs the residual-enhanced feature vector into a lightweight gated linear unit. This gated linear unit performs secondary filtering on each element of the feature vector through learnable gating weights to remove residual iron interference noise components. The final output is a zinc leaching feature vector with the iron interference effect eliminated. Each element value in the zinc leaching feature vector represents the pure zinc leaching state characteristics unaffected by iron interference.

8. The zinc leaching rate prediction and process optimization system based on deep learning according to claim 7, characterized in that: The leaching rate prediction unit employs a hybrid expert regression network, which consists of multiple independent regression experts and a gating network. Each regression expert consists of a fully connected layer, which is responsible for mapping the zinc leaching feature vector to the initial zinc leaching rate prediction value. The parameters of the fully connected layers of different regression experts are different, which are used to learn different zinc leaching rate prediction patterns. The gated network takes the zinc leaching feature vector as input and outputs a set of weight coefficients through the Softmax activation function. Each weight coefficient corresponds to the confidence of a regression expert, and the sum of the weight coefficients is one. The leaching rate prediction unit multiplies the preliminary zinc leaching rate prediction value output by each regression expert with the corresponding weight coefficient and then sums them to obtain the final zinc leaching rate prediction value. During training, the hybrid expert regression network updates the parameters of both the regression expert and the gated network simultaneously using the backpropagation algorithm. This allows the gated network to dynamically select the most suitable regression expert for prediction based on the different feature distributions of the zinc leaching feature vector.

9. The zinc leaching rate prediction and process optimization system based on deep learning according to claim 5, characterized in that: The iron leaching risk assessment module uses a multilayer perceptron classification network. The multilayer perceptron classification network takes the iron disturbance state latent vector as input and performs nonlinear feature extraction on the iron disturbance state latent vector through multiple fully connected hidden layers to extract key features related to iron leaching risk. Key features include the changing trend of iron ion concentration, the rate of iron alum precipitation, and the degree of completion of iron ion valence state transformation. The output layer of the multilayer perceptron classification network uses a linear activation function to map the extracted key features to an iron leaching risk index, with the value range of the iron leaching risk index set between zero and one hundred. When generating the iron leaching risk index, the iron leaching risk assessment module also introduces a sliding window calibration mechanism based on historical data. The sliding window calibration mechanism uses the average iron leaching risk index of the most recent leaching batches as the benchmark value to correct the offset of the currently generated iron leaching risk index, so as to eliminate the systematic deviation caused by the difference in raw material batches.

10. The zinc leaching rate prediction and process optimization system based on deep learning according to claim 9, characterized in that: The process optimization unit is equipped with a multi-stage progressive search engine. The multi-stage progressive search engine reads a preset set of initial values ​​for process parameters from the storage space of the process optimization unit. The set of initial values ​​for process parameters includes feasible ranges for leaching temperature, leaching time, sulfuric acid concentration, and oxidant flow rate. The multi-stage progressive search engine performs a first-stage coarse-grained search. In the first stage, the feasible range of each process parameter is evenly divided into multiple discrete candidate points. All discrete candidate points are combined into multiple initial process parameter combinations. For each initial process parameter combination, the process optimization unit calls the parameter decoupling unit, zinc leaching timing coding unit, iron interference timing coding unit, competition suppression fusion unit and leaching rate prediction unit to perform forward calculations and obtain the corresponding zinc leaching rate prediction value and iron leaching risk index. If the iron leaching risk index corresponding to a certain initial process parameter combination is not lower than the safety threshold, then the initial process parameter combination is discarded; if it is lower than the safety threshold, then the initial process parameter combination and its corresponding zinc leaching rate prediction value are retained. After the first stage, the initial process parameter combinations with the largest predicted zinc leaching rate are selected as candidate seeds. The multi-stage progressive search engine performs a second stage of fine-grained search. In the second stage, a particle swarm optimization algorithm is used to perform a local search in the neighborhood of the process parameter value corresponding to each candidate seed. Each particle in the particle swarm optimization algorithm represents a process parameter combination. The particle's velocity update direction is guided by the current particle's historical best position and the group's historical best position. In each iteration, the process optimization unit recalculates the predicted zinc leaching rate and iron leaching risk index corresponding to each particle, and only retains particles whose iron leaching risk index is lower than the safety threshold. After multiple iterations, the process parameter combination corresponding to the group's historical best position is the process parameter combination that maximizes the predicted zinc leaching rate and meets the iron leaching risk safety conditions. The process optimization unit finally outputs the process parameter combination that maximizes the predicted zinc leaching rate to an external display device.