A smelting device for recycling and preparing periclase from magnesium ore waste

Abstract

This invention belongs to the field of specialized equipment for building material production, and discloses a smelting device for recycling and reprocessing periclase from magnesium ore waste. The device includes a waste pretreatment unit, a multi-channel centrifugal sedimentation density distribution acquisition unit, an intelligent sorting decision unit, a sorting execution mechanism, a smelting reactor, and a process feedback control module connected in sequence. The density distribution acquisition unit obtains the equivalent density of particles through centrifugal sedimentation; the intelligent sorting decision unit extracts a five-dimensional density difference feature vector, adaptively adjusts the classification boundary using a support vector machine, and combines improved hierarchical clustering to achieve dynamic particle grouping and screen for high-purity components; the process simulation interface embeds a thermodynamic-kinetic coupled model to pre-simulate the smelting quality, and inversely optimizes the feature weights when a batch is determined to be a challenging recycling batch. This invention solves the problem of insufficient sorting accuracy caused by the nonlinear mapping between density and chemical purity, and improves the purity stability of periclase products.

Description

Technical Field

[0001] This invention belongs to the field of solid material separation and special equipment for building material production, specifically relating to a smelting device for recycling magnesium ore waste and preparing periclase. Background Technology

[0002] Magnesia, as an important inorganic non-metallic mineral raw material, plays an irreplaceable role in industries such as refractory materials, metallurgical auxiliaries, and high-end ceramics. In the preparation of high-purity periclase, the quality of the raw materials directly determines the thermal stability, erosion resistance, and structural density of the final product. The large amounts of waste generated during magnesium mining and processing, such as tailings, low-grade slag, and roasting residues, are increasingly being considered potential secondary resource carriers. How to efficiently recover the useful components and achieve targeted conversion into high-value-added periclase products has become a core issue for promoting the sustainable development of the magnesium industry. Waste recycling and reuse not only concerns improving resource efficiency but also involves reducing environmental impact and optimizing the economics of the industrial chain, thus attracting dual attention from academia and industry.

[0003] Existing magnesium ore waste treatment technologies mainly follow traditional mineral processing pathways, generally employing single physicochemical methods such as acid-base leaching, flotation, or gravity separation for initial enrichment. Gravity separation is widely used in the initial impurity removal stage due to its simple operation and low energy consumption. Its basic principle lies in using the different settling rates of the components in the waste in the medium due to their density differences to achieve separation.

[0004] Existing technologies, such as jigs or spiral chutes, can effectively remove some high-density silicate or iron-aluminum oxide impurities from waste materials, thereby increasing the relative content of light magnesium components. These methods can indeed achieve a certain degree of extensive enrichment when processing waste materials with relatively homogeneous composition and large density dispersion, providing a basic raw material guarantee for subsequent smelting. However, the design logic of these existing technologies is essentially based on the simplified assumption that density equals purity, ignoring the inherent complex multiphase nature and component cross-contamination of the magnesium ore waste system. Furthermore, some existing technologies attempt to use centrifugal sedimentation equipment for sorting, but this also relies on a single density threshold, failing to solve the problem of identifying overlapping density ranges.

[0005] With the increasing depletion of magnesium ore resources and the refinement of processing techniques, existing sorting strategies based on a single density threshold have revealed deep-seated technical bottlenecks when dealing with waste materials that are highly mixed in composition, have a wide particle size distribution, and complex mineral intergrowth relationships. The density ranges of target components (such as magnesium-bearing minerals like magnesite and brucite) and common impurities (such as talc, chlorite, quartz, and iron-manganese oxides) in magnesium ore waste significantly overlap. For example, the density of some high-silica impurities can be as low as 2.6 g / cm³, while the density of some hydrous magnesium minerals can be as high as 2.8 g / cm³. These two types of minerals are difficult to distinguish effectively through sedimentation behavior in a conventional gravity field.

[0006] Density distribution spectra of waste samples were obtained through high-precision density measurement. However, without the ability to structurally analyze multidimensional density characteristics, it is still impossible to accurately map them to actual chemical purity, resulting in a large number of false positive components still being present in the sorted magnesium-rich segments. Misjudgments caused by the non-one-to-one correspondence between physical properties and chemical composition not only increase the ineffective energy consumption during subsequent high-temperature smelting but also induce quality problems such as periclase lattice distortion and increased porosity due to impurity residues, seriously restricting the consistency and reliability of product performance.

[0007] To compensate for the uncertainties in front-end sorting, companies are often forced to add multiple rounds of chemical testing and process parameter fine-tuning at the back end, forming an inefficient closed loop of "sorting-testing-rework," which significantly increases overall recycling costs and time delays. How to extract discriminative density difference feature vectors from raw waste, accurately define the boundaries between impurities and useful components using machine learning methods, and on this basis, achieve dynamic clustering of sorted components and quantitative prediction of purity ranges, thereby driving the pre-optimization of smelting parameters, has become a key technical problem in overcoming the current bottlenecks in both recycling efficiency and product quality. Summary of the Invention

[0008] The purpose of this invention is to provide a smelting device for recycling magnesium ore waste to produce periclase, which solves the technical problems of insufficient sorting accuracy, high smelting energy consumption, and unstable periclase product quality caused by the nonlinear mapping between density and chemical purity in existing magnesium ore waste recycling.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] This invention addresses the technical problems in existing technologies, such as insufficient sorting accuracy, high smelting energy consumption, and large fluctuations in product quality caused by the complex composition of magnesium ore waste and the non-one-to-one correspondence between density and chemical purity.

[0011] A smelting device for recycling magnesium ore waste into periclase includes a waste pretreatment unit, a density distribution acquisition unit, an intelligent sorting decision unit, a sorting execution mechanism, a smelting reactor body, and a process feedback control module.

[0012] The waste pretreatment unit is used to crush, screen, and homogenize the input magnesium ore waste to ensure that the particle size distribution is controlled within the range of 0.5mm-5mm. The waste is then fed into the density distribution acquisition unit at a constant rate via a vibrating feeder. The density distribution acquisition unit uses a multi-channel centrifugal sedimentation density measurement device, which has several concentric annular sedimentation chambers inside. The rotation speed of each chamber is independently controllable, ranging from 300rpm to 3000rpm. The chamber walls are equipped with a high-precision pressure sensor array and an infrared transmittance detection probe to record the sedimentation position and light transmission characteristics of particles in the centrifugal field at different rotation speeds in real time. This allows for the inversion of the equivalent density value of each particle in each batch of waste. The sampling frequency is not less than 10kHz, and the single measurement cycle does not exceed 120 seconds.

[0013] The intelligent sorting decision unit consists of a central coordinating controller, a density gradient analysis engine, a support vector machine classification model library, a clustering optimization module, and a process simulation interface. The central coordinating controller uses an industrial-grade embedded processor with a main frequency of no less than 1.8GHz, running a real-time operating system, and is responsible for coordinating the data flow and control commands of each submodule. The density gradient analysis engine receives the raw density data stream from the density distribution acquisition unit, first performs denoising and normalization on the data, and then calculates the first derivative, second derivative, and number of local extrema of the density distribution, generating a density difference feature vector containing at least five dimensions, specifically including: density mean μ, standard deviation σ, skewness coefficient γ1, kurtosis coefficient γ2, and density distribution inflection point. The point spacing Δd; the feature vector is fed into the support vector machine classification model library as input. The model library contains the initial classification hyperplane obtained by training with 500 batches of waste samples in history. Each batch has ≥8000 particles. The training data includes a five-dimensional density feature vector and the corresponding XRF chemical purity label (MgO purity ≥98% is useful component, <90% is impurity). Five-fold cross-validation is used for model selection. The kernel function is the radial basis function (RBF). The penalty factor C is initially set to 1.0, and the kernel width parameter γ is initially set to 0.5.

[0014] Before processing each new batch of waste, the support vector machine classification model library performs classification prediction based on the density difference feature vector of the current batch, outputting the preliminary boundary threshold T0 between impurity components and useful components. Subsequently, the system determines whether the boundary threshold T0 falls within the preset density overlap limit range [2.65g / cm³, 2.75g / cm³]. If it is determined to be yes, the classification parameter adaptive adjustment mechanism is activated: the central coordinating controller calls the gradient descent optimizer, with the objective function of minimizing the cross-validation error, and iteratively updates the penalty factor C and the kernel width parameter γ until the newly generated boundary threshold T1 leaves the density overlap limit range, or the error decrease is less than 0.001 after three consecutive iterations. The finally determined optimized component differentiation model is locked for the full prediction of the current batch.

[0015] Based on the optimized component differentiation model, the intelligent sorting decision unit classifies the density data of the entire batch of waste point by point and calculates the proportion of impurity components. Based on this, the purity prediction range is calculated. ,in , δ is the confidence half-width derived from historical batch variance statistics. δ is calculated based on the standard error of purity data from 500 historical batches of waste, with a confidence level of 95%. The calculation formula is as follows:

[0016] ;

[0017] in The standard deviation of the purity sample. The t-distribution quantile is calculated and fixed at 0.03; this purity prediction range is passed as a key input to the clustering optimization module; the clustering optimization module adopts an improved hierarchical clustering algorithm, using the equivalent density value as the only clustering dimension, initially dividing all particles into N single-element clusters, where N equals the total number of particles; then, adjacent clusters are merged in ascending density order if and only if the maximum density difference between two clusters is less than a preset crossover threshold. ( The merging operation is performed when the condition is met; the merging process continues until no adjacent cluster pairs that meet the conditions exist, and finally a simplified set of sorting components is output. Each component It contains a group of particles with similar densities, but the density varies greatly within the group. .

[0018] From the sorted component set S, the system automatically selects a high-purity subset with a predicted purity value higher than 92%. The data structure of the high-purity subset H includes the average density, particle number, mass percentage, and predicted purity value of each component. This data is fed into a process simulation interface, which embeds a thermodynamic-kinetic coupled model of the periclase smelting process. The model inputs include the MgO content of the raw material, impurity types (SiO2, Al2O3, Fe2O3, CaO), particle size distribution, and feed rate. The model output consists of three key quality impact assessment indicators: average grain size... Open porosity and flexural strength The model is based on the Arrhenius reaction rate equation and Fick's diffusion law and has been calibrated with 500 sets of industrial test data. The prediction error is measured by the root mean square error (RMSE): grain size RMSE≤4.2μm, porosity RMSE≤0.5%, flexural strength RMSE≤2.1MPa, and the relative error is controlled within ±5%.

[0019] Based on the aforementioned quality impact assessment indicators, the central collaborative controller determines whether the current waste batch belongs to the "recycling challenge" batch, and the determination criteria are as follows: or or If any condition is met, the iterative optimization process is triggered: the system calls the gradient descent algorithm to narrow the purity prediction range. To achieve the goal, the weights of each dimension in the density difference feature vector are adjusted in reverse. The weight update formula is: ;

[0020] Learning rate Fixed at 0.01, gradient Numerical calculations are performed using the finite difference method; the classification and clustering process is re-executed after each update, and the decrease of W is evaluated; if W continues to narrow after two consecutive updates, the resource utilization improvement threshold is confirmed to have been reached, the current weight configuration is written into the final waste recycling configuration file, and the feature extraction layer parameters in the support vector machine classification model library are updated synchronously.

[0021] The sorting actuator includes a multi-channel pneumatic diversion valve array, a high-speed electromagnetic push rod, and a material guide slide. The number of pneumatic diversion valves corresponds one-to-one with the number of components k in the sorting component set S. Each valve has a response time of ≤10ms and is controlled by a central coordinating controller via an RS485 bus to send switching commands. The high-speed electromagnetic push rod is installed at the end of the main conveyor belt and is used to complete the pushing action within 0.5 seconds after receiving the grouping signal, guiding the corresponding component into the designated collection bin. The material guide slide uses a wear-resistant ceramic liner and has an adjustable inclination angle of 15° to 45° to ensure that components of different densities do not cross-mix under gravity assistance.

[0022] The smelting reactor is a DC electric arc furnace with a rated power of 5MW. The furnace lining is made of chrome-magnesium bricks, and the operating temperature range is 1600℃-2000℃. The furnace is equipped with an oxygen potential control system, which monitors the CO / CO2 ratio in the furnace gas in real time using a mass spectrometer and adjusts the argon purging flow rate accordingly to maintain the oxygen partial pressure at 10. -6 atm to 10 -4 The feed inlet is directly connected to the high-purity subset H output of the sorting actuator to ensure that only high-purity components enter the smelting zone; the furnace bottom is equipped with an electromagnetic stirring device with an adjustable frequency range of 1Hz to 10Hz, which is used to promote the homogenization of the melt and the floating of inclusions.

[0023] The process feedback control module includes an online X-ray fluorescence spectrometer (XRF), an infrared thermal imager, and an acoustic emission sensor. The XRF is installed downstream of the smelting furnace outlet and collects melt composition data every 30 seconds, focusing on monitoring the contents of MgO, SiO2, and Al2O3. The infrared thermal imager covers the entire cooling ingot area with a spatial resolution of 0.5 mm, used to identify surface cracks and abnormal temperature gradients. The acoustic emission sensor is embedded in the furnace wall with a sampling rate of 1 MHz, used to capture microcrack initiation signals during melt solidification. All feedback data is transmitted back to the central coordinating controller in real time and compared with the quality impact assessment indicators. If the measured MgO purity is lower than 98.5% or the porosity exceeds 6%, the system automatically triggers the "smelting parameter compensation protocol": increasing the smelting temperature by 50°C, extending the holding time by 15 minutes, and increasing the electromagnetic stirring intensity to 8 Hz. This deviation event is recorded in the model training database for incremental learning of the next round of support vector machine model.

[0024] Furthermore, the density distribution acquisition unit and the intelligent sorting decision unit are connected via Gigabit Ethernet, and the data transmission protocol adopts the IEEE 1588 Precision Time Protocol (PTP) to ensure that the time synchronization error between the density measurement time and the sorting command issuance time is less than 1 millisecond; the control commands between the central coordinating controller and the sorting execution mechanism are transmitted via a dual-redundant CAN bus with a baud rate of 1 Mbps and CRC check and automatic retransmission mechanism; during the entire system operation, all intermediate data and decision logs are stored on a local solid-state drive in an encrypted binary stream compliant with the ISO / IEC 27001 standard, with a retention period of no less than two years.

[0025] In a preferred embodiment of the present invention, the support vector machine classification model library automatically initiates a transfer learning mechanism each time a new waste source is introduced: by using the labeled samples of the source domain (historical waste) and the unlabeled samples of the target domain (new waste), the feature space is aligned using the maximum mean difference (MMD) minimization criterion, thereby quickly adapting the density-purity mapping relationship of the new waste with only a small number of new sample labels, significantly reducing the cold start cost of the model.

[0026] In another preferred embodiment of the present invention, before performing hierarchical clustering, the clustering optimization module first performs wavelet denoising processing on the density data, using the Daubechies4 wavelet basis, with a decomposition level of 3, and a soft threshold function, in order to eliminate the interference of measurement noise on the clustering boundary and ensure the physical authenticity of the grouping results.

[0027] As another preferred embodiment of the present invention, the thermodynamic-kinetic coupling model in the process simulation interface adopts a modular architecture, wherein the impurity reaction sub-model, grain growth sub-model and porosity evolution sub-model are decoupled from each other and can be updated independently through plug-ins; when a new impurity (such as titanium-containing minerals) is detected in the waste, only the corresponding impurity reaction plug-in needs to be loaded, without reconstructing the entire model, thereby improving the adaptability and maintainability of the system.

[0028] In addition, the present invention also discloses a method for recycling and reprocessing periclase from magnesium ore waste, using the above-mentioned smelting apparatus for recycling and reprocessing periclase from magnesium ore waste, comprising the following steps:

[0029] Step 1: Crushing, screening and homogenizing the magnesium ore waste, controlling the particle size to be between 0.5 mm and 5 mm;

[0030] Step 2: Determine the equivalent density of each particle using a multi-channel centrifugal sedimentation device, with a sampling frequency ≥10kHz and a single measurement cycle ≤120 seconds;

[0031] Step 3: Generate a five-dimensional density difference feature vector based on the density data, input it into the support vector machine classification model library for component boundary prediction. When the initial boundary threshold falls into the range of [2.65, 2.75] g / cm³, iteratively update the penalty factor C and kernel width parameter γ through the gradient descent optimizer until the new boundary threshold leaves the range or the error decreases by less than 0.001 for three consecutive iterations.

[0032] Step 4: Classify the entire batch of particles point by point, count the percentage of impurities, and calculate the predicted purity range;

[0033] Step 5: The improved hierarchical clustering algorithm is used to dynamically group the particles according to their density. Initially, all particles are divided into single-element clusters. Adjacent clusters are merged in ascending order of density. Merging is performed only when the maximum density difference between two clusters is less than 0.08 g / cm³. The final output is a set of sorted groups with a density range within each group ≤ 0.08 g / cm³.

[0034] Step 6: Screen high-purity components with a predicted purity of ≥92% and input them into the thermodynamic-kinetic coupled model for smelting quality simulation;

[0035] Step 7: If the pre-simulation indicators fail to meet the standards, reverse-optimize the feature weights and re-execute classification and clustering; only send the high-purity components into the DC electric arc furnace for melting, with a furnace temperature of 1600–2000℃ and an oxygen partial pressure maintained at 10. -6 Up to 10 -4 atm;

[0036] Step 8: Monitor product quality in real time using XRF, infrared thermal imager and acoustic emission sensor. If MgO purity is <98.5% or porosity is >6%, automatically compensate smelting parameters and record deviation events for model iteration.

[0037] Furthermore, the density distribution acquisition unit is also equipped with a shape factor correction module, including at least two sets of orthogonally arranged high-speed linear array cameras, for synchronously capturing the two-dimensional projection contours of particles during centrifugal sedimentation.

[0038] Extracting particle projection area through image processing and projected perimeter Calculate the shape factor And based on the pre-calibrated piecewise linear correction function The Stokes equivalent diameter is corrected to obtain the corrected equivalent density value. ;

[0039] The correction function Defined as:

[0040] when hour ,when hour ,when hour ,when hour .

[0041] Furthermore, the support vector machine classification model library embeds an active learning module, which, when the initial boundary threshold... Falling into the density overlap limit range First, uncertainty sampling and representative sampling are performed: the 100 particles closest to the current classification hyperplane, and 30 particles each from the peak region and the two flanking regions of the density distribution histogram are selected, and their true MgO purity labels are obtained using an online X-ray fluorescence microprobe; the labeled samples are added to the training set to perform incremental learning of the support vector machine classification model library, and the boundary thresholds are recalculated. ;like If the error still falls within the overlapping region, restart the gradient descent optimizer to iteratively update the penalty factor. With kernel width parameter .

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] This invention, by constructing a complete technical path encompassing density measurement, feature extraction, intelligent classification, dynamic clustering, process simulation, and closed-loop feedback, abandons the traditional experience-based sorting model that relies on a single density threshold. It achieves high-precision identification and directional separation of multiphase components in magnesium ore waste. Simultaneously, it pre-couples the sorting results with smelting process parameters, making the preparation process of high-purity periclase predictable, controllable, and self-optimizing. While ensuring product quality stability, it significantly improves the resource utilization rate of magnesium ore waste, reduces the overall energy consumption and carbon emission intensity per unit product, and provides reliable technical support for the green and low-carbon transformation of the magnesium industry. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of the overall structure of the device described in this invention.

[0046] Figure 2 This is a flowchart illustrating the method of the present invention.

[0047] Figure label:

[0048] Jaw crusher 101, vibrating screen 102, homogenizing mixing bin 103, vibrating feeder 104, waste pretreatment unit 105, density distribution acquisition unit 106, intelligent sorting decision unit 107, sorting execution mechanism 108, smelting reaction furnace body 109, process feedback control module 110. Detailed Implementation

[0049] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0050] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0051] Example 1: See Figure 1 and Figure 2 This embodiment discloses a smelting apparatus for recycling and reprocessing magnesium ore waste into periclase. The smelting apparatus is composed of a waste pretreatment unit 105, a density distribution acquisition unit 106, an intelligent sorting decision unit 107, a sorting execution mechanism 108, a smelting reactor body 109, and a process feedback control module 110 connected in sequence. The waste pretreatment unit 105 includes a jaw crusher 101, a vibrating screen 102, and a homogenizing mixing bin 103.

[0052] The input magnesium ore waste is first coarsely crushed by a jaw crusher 101, with the particle size controlled below 5mm. It then enters a vibrating screen 102 with a screen aperture set to 0.5mm to ensure the final output particle size distribution is strictly limited to the range of 0.5mm to 5mm. The undersized material is fed into a homogenization mixing bin 103, where it is stirred at 15 rpm for 10 minutes to achieve uniform component distribution. The homogenized material is then fed into a density distribution acquisition unit 106 at a constant rate of 3.2kg / s via a frequency-controlled vibrating feeder 104.

[0053] The density distribution acquisition unit 106 employs a multi-channel centrifugal sedimentation density measurement device. Its main body is a horizontal cylindrical shell containing four concentric annular sedimentation chambers. The diameters of each chamber are 120mm, 180mm, 240mm, and 300mm, respectively, and their axial lengths are all 200mm. Each sedimentation chamber is equipped with an independent servo motor drive, with the rotation speed continuously adjustable from 300rpm to 3000rpm, with an adjustment accuracy of ±5rpm. Sixteen miniature piezoresistive pressure sensors are evenly embedded radially on the chamber walls, with a range of 0–10kPa, an accuracy of ±0.1%, and a sampling frequency of 10kHz. Simultaneously, infrared transmittance detection probes are symmetrically installed on both sides of the chamber, emitting at a wavelength of 940nm and using silicon photodiodes as receivers with a response time ≤1μs. When waste particles enter the sedimentation chamber with the carrier gas, they settle to different radii according to their density differences under centrifugal force. The pressure sensors record local static pressure changes, and the infrared probes simultaneously acquire the transmittance attenuation curve. The central coordinating controller calculates the equivalent density value of each particle based on the Stokes sedimentation equation. The formula is as follows:

[0054] ;

[0055] in, The carrier gas density is 1.2 kg / m³. The dynamic viscosity of the carrier gas (1.8 × 10⁻⁶) -5 The carrier gas was dry air with a fixed composition (volume fraction: N2 78.08%, O2 20.95%, Ar 0.93%, CO2 0.04%), the temperature was controlled at 25±2°C, and the relative humidity was below 5%. The radial settling velocity of the particles is calculated using the central difference method based on the static pressure versus time curve recorded by the pressure sensor array, with a sampling interval of 0.1 ms. The equivalent diameter of the particles is obtained by measuring a laser particle size analyzer. Angular velocity (rad / s) The settlement radius is (m).

[0056] A single measurement cycle is 90 seconds, which can complete the density measurement of approximately 8,000 particles. The data is transmitted in real time to the intelligent sorting decision unit 107 via gigabit Ethernet. The transmission protocol adopts the IEEE 1588 Precision Time Protocol (PTP), and the time synchronization error is less than 0.8 milliseconds.

[0057] The intelligent sorting decision unit 107 consists of a central coordinating controller, a density gradient analysis engine, a support vector machine classification model library, a clustering optimization module, and a process simulation interface. The central coordinating controller uses an NXP.MX8 QuadMax industrial-grade embedded processor with a main frequency of 1.8GHz, runs the VxWorks7 real-time operating system, and features dual-core lockstep functionality to ensure the determinism and timeliness of control commands. The density gradient analysis engine receives the raw density data stream from the density distribution acquisition unit 106. First, impulse noise is removed using a sliding window mid-range filter (window width = 51 points), followed by Min-Max normalization to the [0,1] interval. Next, the first derivative of the density distribution is calculated. Second derivative It identifies local maxima and minima, and counts the number of inflection points. Finally, it generates a five-dimensional density difference feature vector. ,in The density mean Standard deviation The skewness coefficient is defined as the third central moment divided by the cube of the standard deviation. It is the kurtosis coefficient (fourth central moment divided by the square of the variance minus 3). This represents the average density spacing between adjacent inflection points.

[0058] The feature vector F is fed into the support vector machine (SVM) classification model library. The library contains pre-stored initial classification hyperplanes trained on 500 historical batches of waste samples. The kernel function is a radial basis function (RBF), the penalty factor C is initially set to 1.0, and the kernel width parameter γ is initially set to 0.5. For the current batch, the model library performs classification prediction and outputs the preliminary boundary threshold T0 between impurity and useful components. The system then determines whether T0 falls within the preset density overlap limit range [2.65 g / cm³, 2.75 g / cm³]. If it does, an adaptive adjustment mechanism for classification parameters is initiated: the central coordinating controller calls the L-BFGS gradient descent optimizer, using the 5-fold cross-validation error E_cv as the optimization objective, to iteratively update C and γ.

[0059] The optimization process stops when one of the termination conditions is met:

[0060] (1) New boundary threshold ;

[0061] (2) After three consecutive iterations .

[0062] For example, in a certain batch, the initial After two iterations, C was updated to 2.3, γ was updated to 0.38, and T1 = 2.58 g / cm³, successfully escaping the overlapping region and the optimization ended.

[0063] Based on the optimized component differentiation model, the system classifies 8000 density data points of the entire batch of waste material point by point, labeling them as "useful" or "impurity". The percentage of impurity components is then statistically determined. The purity prediction range is:

[0064] .

[0065] This range is then passed to the clustering optimization module. The clustering optimization module first performs wavelet denoising on the density data: using the Daubechies4 wavelet basis, a decomposition level of 3, a soft thresholding function, and a threshold value of [value missing]. ,in Estimated using Median Absolute Deviation (MAD) from density data, the calculation formula is as follows: MAD is the median of the absolute deviation of the equivalent density value from the median.

[0066] The denoised data was used for hierarchical clustering: initially, 8000 particles were divided into 8000 single-element clusters; after sorting them in ascending density order, adjacent cluster pairs were checked sequentially. If the maximum density difference between two clusters was <0.08 g / cm³, they were merged into a new cluster; this process was repeated until no cluster pairs met the condition. The final output was a set of sorted components. In this example, k=12, and each group The density range within each group is ≤0.075g / cm³.

[0067] The system selects a high-purity subset H from S, where the predicted purity value is ≥0.92. The predicted purity value is taken as the midpoint of the range. In this example, H contains The sample consists of five components with average densities of 2.81, 2.85, 2.89, 2.93, and 2.97 g / cm³, respectively, representing a total mass percentage of 82.3%. These data were fed into the process simulation interface. The interface incorporates a thermodynamic-kinetic coupled model, which comprises three decoupled sub-models: an impurity reaction sub-model based on the FactSage database, considering the phase equilibrium of SiO₂, Al₂O₃, Fe₂O₃, CaO, and MgO at high temperatures; and a grain growth sub-model employing a modified Burke-Turnbull equation.

[0068] ;

[0069] in, The pre-exponential factor (taken as 1.2 × 10⁻⁶) -6 m), The activation energy is 420 kJ / mol. The gas constant is... Absolute temperature For heat preservation time, The time index is 0.4. The gas constant is 8.314 J / (mol·K); the stomatal evolution sub-model is based on the Kelvin equation and Ostwald ripening theory to calculate the open porosity. The model input also includes particle size distribution ( The model outputs three quality impact assessment indicators: feed rate (3.2 kg / s) and impurity content (calculated from the density-purity mapping relationship). .

[0070] The central coordinating controller determined that this batch belonged to the "recycling challenge type" (because...). and This triggers an iterative optimization process. The system narrows the predicted purity range. To achieve the goal, the weights of each dimension in the feature vector F are adjusted in reverse. ].

[0071] The initial weights are set to [0.2, 0.2, 0.2, 0.2, 0.2]. The learning rate η is fixed at 0.01, and the gradient... Calculated using the finite difference method: a perturbation is applied to each weight component. Then, re-perform classification and clustering, and observe the changes in W.

[0072] After the first update, w=[0.195,0.205,0.210,0.190,0.200], and W decreased to 0.058; after the second update, w=[0.190,0.210,0.215,0.185,0.200], and W further decreased to 0.055. After optimization, the purity prediction range was recalculated to [0.907,0.960]. The high-purity subset screening criteria are met. After two consecutive narrowings of W, the resource utilization improvement threshold is confirmed. The current weight configuration is written to the waste recycling configuration file config_batch_20240517.bin, and the feature extraction layer parameters in the support vector machine classification model library are updated simultaneously.

[0073] The sorting actuator 108 includes 12 pneumatic diversion valves, a high-speed electromagnetic push rod, and a material guide slide. The pneumatic diversion valves adopt a pilot-operated two-position three-way structure with a response time of 8ms. Switching commands are sent by the central coordinating controller via a dual-redundant CAN bus (1Mbps baud rate, CRC-16 checksum). The high-speed electromagnetic push rod is installed at the end of the main conveyor belt, with a stroke of 50mm, a thrust of 200N, and an action delay of ≤0.45 seconds. The material guide slide is made of alumina ceramic (Al2O3 content ≥95%), with an 8mm lining thickness. The tilt angle is adjusted by a servo motor; in this example, it is set to 32°. The high-purity subset H corresponds to the following components: The components are introduced into collection bins H1–H5 respectively, while the remaining low-purity components are introduced into waste bin R.

[0074] The smelting reactor body 109 is a DC electric arc furnace with a rated power of 5MW, a furnace volume of 8m³, and is lined with chrome-magnesium bricks (Cr2O3 content 18%, MgO content 80%). The working temperature is set at 1850℃.

[0075] The furnace is equipped with an oxygen potential control system: a mass spectrometer monitors the CO / CO2 ratio in the furnace gas in real time, with a sampling frequency of 1Hz; when the ratio > 0.8, the PLC adjusts the argon purging flow rate from 50 Nm³ / h to 70 Nm³ / h to maintain the oxygen partial pressure at 5 × 10⁻⁶. -5 ATM. The feed inlet is connected to the collection bins H1–H5 via airtight rotary valves to ensure that only high-purity components enter the smelting zone. An electromagnetic stirring device is installed at the furnace bottom, with 120 coil turns and an adjustable current frequency of 1–10Hz; in this example, it is initially set to 5Hz.

[0076] The process feedback control module 110 includes an online X-ray fluorescence spectrometer, an infrared thermal imager, and an acoustic emission sensor. The XRF is installed below the discharge chute, acquiring melt composition data every 30 seconds, with detection limits of MgO ± 0.1% and SiO2 ± 0.05%. The infrared thermal imager is a FLIRA8580sc, with a frame rate of 120Hz and a spatial resolution of 0.48mm, covering the entire ingot casting platform. The acoustic emission sensor is a resonant piezoelectric ceramic probe with a center frequency of 150kHz and a sampling rate of 1MHz, embedded within the refractory material layer of the furnace wall. In one melting operation, the XRF measured MgO purity at 98.1% (below the 98.5% threshold), and the infrared image showed an abnormal temperature gradient zone on the ingot surface (ΔT > 50℃), with a sudden increase in acoustic emission signal energy (> 50dB). The central coordinating controller triggered the "melting parameter compensation protocol": the melting temperature was increased to 1900℃, the holding time was extended from 45 minutes to 60 minutes, and the electromagnetic stirring frequency was increased to 8Hz. After compensation, the purity of MgO was retested and found to be 98.7%, while the porosity decreased to 5.8%. This deviation event was recorded in the SQLite database event_log.db, containing the original density data, classification parameters, smelting settings, and measured results, which will be used for incremental learning of the support vector machine model.

[0077] In a preferred embodiment of the present invention, when new waste materials are introduced (such as waste materials from a new mining area in Haicheng, Liaoning), the support vector machine classification model library automatically initiates a transfer learning mechanism. The system calls 2000 labeled samples from the source domain (historical Shandong waste) and 500 unlabeled samples from the target domain (new waste) to calculate the maximum mean difference (MMD):

[0078] ;

[0079] in For RBF kernel mapping, The model is a regenerated kernel Hilbert space. By minimizing the MMD and aligning the feature distributions of the two domains, the classification accuracy can be improved to over 92% with only 30 manually labeled samples from the new waste, significantly reducing the cold start cost.

[0080] In another preferred embodiment of the present invention, the wavelet denoising step of the cluster optimization module before hierarchical clustering effectively suppresses measurement noise. Comparative experiments show that the standard deviation of cluster boundary fluctuation is 0.021 g / cm³ without denoising, which is reduced to 0.008 g / cm³ after Daubechies4 wavelet processing, significantly improving the physical consistency of grouping.

[0081] In another preferred embodiment of the present invention, the modular architecture of the process simulation interface supports rapid adaptation to novel impurities. When titanium-containing minerals (TiO2 content 0.8%) are detected in the waste, the system automatically loads the "Titanium Impurity Reaction Plugin," which contains TiO2-MgO-Al2O3 ternary phase diagram data and magnesium titanate (MgTiO3) formation kinetic parameters. Without reconstructing the entire thermodynamic model, its inhibitory effect on grain growth can be accurately predicted.

[0082] In practice, a batch of magnesia ore waste from Dashiqiao Magnesia Mine in Liaoning Province (initial MgO content 86.2%) was processed. After pretreatment, the particle size was 2.1±0.8 mm. Density distribution samples were collected from 8000 particles, with a density range of 2.52–3.05 g / cm³. Adaptive SVM classification and dynamic clustering were used to select a high-purity subset H (78.5% by mass), with a predicted purity of 94.3%. Melting parameters: 1850℃, holding time 50 min, stirring at 6 Hz. Process feedback showed an MgO purity of 98.9%. .

[0083] Comparative Example 1: The same batch of waste was sorted using traditional heavy media separation with a fixed density threshold (2.70 g / cm³). The sorted material had an MgO content of 91.5%, but the composition fluctuated greatly (standard deviation ±2.1%). Under the same smelting conditions, the product had an MgO purity of 97.2%. , Furthermore, local inclusions were observed.

[0084] Comparative Example 2: Using the device of the present invention but disabling the adaptive adjustment and clustering optimization functions of the intelligent sorting decision unit 107, only using the initial SVM model ( Binary sorting was performed. The high-purity component accounted for 72.1%, with a predicted purity of 93.0%. The purity of the MgO product after smelting was 98.0%, but there was a large variation between batches (standard deviation of ±0.4% in three repeated experiments), while the standard deviation of Example 1 was only ±0.15%.

[0085] The test results are summarized in Table 1 below:

[0086] Table 1:

[0087]

[0088] This invention significantly improves sorting accuracy and product quality stability while reducing energy consumption through refined density gradient analysis and intelligent decision-making closed loop. Throughout the system's operation, all intermediate data (including original density sequences, feature vectors, classification boundaries, clustering results, smelting settings, and feedback signals) are stored in encrypted binary stream format on a local solid-state drive, conforming to the ISO / IEC 27001 standard, with a retention period of 24 months, supporting full-process traceability and continuous model optimization.

[0089] The data perception, intelligent decision-making, process simulation, and closed-loop execution technology paradigm constructed in this invention successfully achieves a paradigm shift from traditional experience-based sorting to precise and intelligent recycling. By deconstructing the traditional mapping relationship between density and purity, and constructing a causal relationship between multi-dimensional density gradient features and final smelting quality, this invention abandons the industry's conventional thinking of treating density as a single purity indicator. Instead, it starts from the physical essence, quantifying implicit information such as density distribution patterns, statistical moment characteristics, and local gradient changes into discriminable feature vectors. Furthermore, through the adaptive adjustment mechanism of the support vector machine classification model library, it dynamically identifies high-purity useful components under complex conditions where component densities severely overlap.

[0090] Furthermore, this invention breaks through the traditional sequential operation mode where sorting and smelting are isolated from each other. By using a mechanical-dynamic coupling model, the sorting results are pre-coupled and virtually verified with the smelting process parameters, achieving predictive control and pre-emptive risk mitigation of the final product quality. This forms a closed-loop iterative optimization capability from physical measurement to process execution, fundamentally ensuring the stability of periclase product grain development, densification, and mechanical properties. It also achieves a qualitative improvement in resource utilization and energy consumption indicators, providing a systematic solution for the high-value recycling of magnesium ore waste that combines theoretical rigor with engineering feasibility.

[0091] Example 2: Based on Example 1, this example optimizes the density distribution acquisition unit and the intelligent sorting decision unit to solve the problems of calculation deviation of equivalent density of non-spherical particles and classification boundary drift in density overlap interval.

[0092] In the multi-channel centrifugal sedimentation density measurement device described in Example 1, two sets of orthogonally arranged high-speed linear array cameras with a resolution of 2048 pixels and a sampling frequency of 2000 fps are added to the sidewall of each annular sedimentation chamber. When a particle passes through the measurement area of ​​the infrared transmittance detection probe, the linear array camera continuously acquires at least 20 frames of images, and the projected area of ​​the particle is extracted through background subtraction and connected component analysis. (Unit: mm²) and projected perimeter (Unit: mm)

[0093] Define shape factor :

[0094]

[0095] in For dimensionless numbers, for spherical particles For irregular particles .

[0096] The formula for correcting the equivalent diameter based on the shape factor is:

[0097]

[0098] In the formula:

[0099] The corrected equivalent diameter is in mm.

[0100] Stokes equivalent diameter calculated from Stokes equations, in mm;

[0101] This is a shape correction function, dimensionless.

[0102] Correction function By conducting sedimentation experiments and three-dimensional CT scan calibration on 500 sets of magnesium mineral standards with known shapes, a piecewise linear model was obtained through fitting:

[0103]

[0104] This leads to the correction of the equivalent density value:

[0105]

[0106] The meanings of the symbols in the formula are consistent with those of the Stokes equation in Example 1: The carrier gas density is (kg / m³). The dynamic viscosity of the carrier gas is (Pa·s). The radial settling velocity of the particles is (m / s). Angular velocity (rad / s) The settling radius is (m). This correction reduced the density measurement error from ±0.08 g / cm³ to ±0.02 g / cm³, verified by 500 repeated tests.

[0107] An active learning module is embedded in the support vector machine classification model library described in Example 1. When the initial boundary threshold... Falling into the density overlap limit range At that time, the following active sampling strategy will be implemented:

[0108] (1) Uncertainty sampling: Select the 100 particles closest to the current classification hyperplane (i.e., the samples with the lowest classification confidence) and obtain their true MgO purity labels using an online X-ray fluorescence microprobe. This microprobe is integrated upstream of the sorting actuator, with a spatial resolution of 0.5 mm and a single-point detection time of ≤3 seconds.

[0109] (2) Representative sampling: 30 particles were selected from the peak area and the two flanking areas of the density distribution histogram for labeling to ensure full coverage of the density distribution.

[0110] Add up to 160 newly labeled samples to the training set, perform incremental learning of the support vector machine classification model library, and use the sequential minimum optimization algorithm to update only the support vector set. After incremental learning, recalculate the boundary thresholds. ,like If the error still falls within the overlapping region, the gradient descent optimizer described in Example 1 will be restarted to iteratively update the penalty factor. With kernel width parameter .

[0111] This active learning mechanism reduces the number of labeled samples required for the model to adapt to new sources of waste from more than 500 in traditional methods to no more than 160, and shortens the cold start time from 4 hours to 30 minutes.

[0112] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0113] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smelting apparatus for recycling and reprocessing magnesium ore waste into periclase, characterized in that: It includes a waste pretreatment unit, a density distribution acquisition unit, an intelligent sorting decision unit, a sorting execution mechanism, a smelting reactor body, and a process feedback control module connected in sequence; The waste pretreatment unit is used to crush, screen and homogenize magnesium ore waste, so that the particle size distribution is controlled within 0.5mm-5mm, and then feeds it into the density distribution acquisition unit at a constant rate through a vibrating feeder. The density distribution acquisition unit is a multi-channel centrifugal sedimentation density measurement device. It has several concentric annular sedimentation chambers inside, and the rotation speed of each chamber is independently controllable, ranging from 300 rpm to 3000 rpm. The chamber walls are embedded with a high-precision pressure sensor array and an infrared transmittance detection probe, which are used to record the sedimentation position and light transmission characteristics of particles in the centrifugal field in real time, and to obtain the equivalent density value of each particle. The intelligent sorting decision unit includes a central collaborative controller, a density gradient analysis engine, a support vector machine classification model library, a clustering optimization module, and a process simulation interface. It is used to generate density difference feature vectors based on density data and perform adaptive classification, dynamic clustering, and smelting quality prediction. The sorting actuator includes a pneumatic diversion valve array, a high-speed electromagnetic push rod, and a material guide slide that correspond one-to-one with the sorting group number, used to introduce components of different purities into the corresponding collection bins; The smelting reactor body is a DC electric arc furnace structure. The smelting reactor body is equipped with an oxygen potential control system and an electromagnetic stirring device. The feed inlet only receives a high-purity subset. The process feedback control module includes an online X-ray fluorescence spectrometer, an infrared thermal imager, and an acoustic emission sensor, which are used to monitor the melt composition, ingot surface condition, and microcracks in the solidification process in real time, and to feed back the deviation signal to the central coordinating controller to trigger the smelting parameter compensation protocol.

2. The smelting apparatus for recycling and reprocessing magnesium ore waste into periclase according to claim 1, characterized in that: The density difference feature vector includes five dimensions: density mean, standard deviation, skewness coefficient, kurtosis coefficient, and density distribution inflection point spacing. The support vector machine classification model library adopts a radial basis function kernel and is configured with an adaptive adjustment mechanism for classification parameters. When the initial boundary threshold falls into the density overlap limit range [2.65g / cm³, 2.75g / cm³], the initial penalty factor and the initial kernel width parameter are iteratively updated through the gradient descent optimizer until the new boundary threshold leaves the range or the error decreases by less than 0.001 for three consecutive iterations.

3. A smelting apparatus for recycling and reprocessing magnesium ore waste into periclase according to claim 2, characterized in that: The clustering optimization module uses equivalent density as the sole clustering dimension and employs an improved hierarchical clustering algorithm: initially, all particles are divided into single-element clusters, and adjacent clusters are merged sequentially in ascending density order, merging only if and only if the maximum density difference between two clusters is less than a preset crossover threshold. The merging process is performed at the same time, and the final output shows the density range within each group. The set of sorted components S.

4. A smelting apparatus for recycling and reprocessing magnesium ore waste into periclase according to claim 3, characterized in that: The intelligent sorting decision unit selects a high-purity subset H from the sorting component set S, with a purity prediction value ≥ 0.

92. The purity prediction range... Based on the proportion of impurity components The calculation shows that: , .

5. A smelting apparatus for recycling and reprocessing magnesium ore waste into periclase according to claim 4, characterized in that: The process simulation interface incorporates a thermodynamic-kinetic coupled model. Inputs include the MgO content of the raw material, impurity types, particle size distribution, and feed rate. The output is the average grain size. Open porosity and flexural strength ;when or or When a batch is identified as a challenging batch to recycle, a weighted iterative optimization process is triggered to narrow the predicted purity range. To achieve the goal, the weights of each dimension of the density difference feature vector are adjusted in reverse. The learning rate η = 0.01, and the gradient is calculated using the finite difference method.

6. A smelting apparatus for recycling and reprocessing magnesium ore waste into periclase according to claim 5, characterized in that: The smelting parameter compensation protocol is automatically triggered when the measured MgO purity is below 98.5% or the porosity exceeds 6%. Specific measures include: increasing the smelting temperature by 50°C, extending the holding time by 15 minutes, increasing the electromagnetic stirring frequency to 8Hz, and recording the deviation event to the model training database for incremental learning of the support vector machine model.

7. A smelting apparatus for recycling and reprocessing magnesium ore waste into periclase according to claim 1, characterized in that: The density distribution acquisition unit and the intelligent sorting decision unit are connected via gigabit Ethernet and adopt the IEEE 1588 precision time protocol, with a time synchronization error of less than 1 millisecond. The central collaborative controller and the sorting execution mechanism use a dual-redundant CAN bus to transmit control commands at a baud rate of 1 Mbps, and have CRC check and automatic retransmission mechanisms.

8. A smelting apparatus for recycling and reprocessing magnesium ore waste into periclase according to claim 2, characterized in that: The support vector machine classification model library automatically initiates a transfer learning mechanism when introducing new waste materials. It uses labeled samples from the source domain and unlabeled samples from the target domain to align the feature space by minimizing the maximum mean difference. Only a small number of new samples are needed to adapt the density-purity mapping relationship of the new waste materials.

9. A smelting apparatus for recycling and reprocessing magnesium ore waste into periclase according to claim 1, characterized in that: The density distribution acquisition unit is also equipped with a shape factor correction module, including at least two sets of orthogonally arranged high-speed linear array cameras, which are used to synchronously capture the two-dimensional projection contours of particles during centrifugal sedimentation. Extracting particle projection area through image processing and projected perimeter Calculate the shape factor And based on the pre-calibrated piecewise linear correction function The Stokes equivalent diameter is corrected to obtain the corrected equivalent density value. ; The correction function Defined as: when hour ,when hour ,when hour ,when hour .

10. A smelting apparatus for recycling and reprocessing magnesium ore waste into periclase according to any one of claims 1-9, characterized in that, The smelting apparatus is used as follows: Step 1: Crush, screen and homogenize the magnesium ore waste, controlling the particle size to be between 0.5mm and 5mm; Step 2: Determine the equivalent density of each particle using a multi-channel centrifugal sedimentation device; Step 3: Generate a five-dimensional density difference feature vector based on the density data, input it into the support vector machine classification model library for component boundary prediction. When the initial boundary threshold falls into the range of [2.65, 2.75] g / cm³, iteratively update the penalty factor C and kernel width parameter γ through the gradient descent optimizer until the new boundary threshold leaves the range or the error decreases by less than 0.001 for three consecutive iterations. Step 4: Classify the entire batch of particles point by point, count the percentage of impurities, and calculate the predicted purity range; Step 5: The improved hierarchical clustering algorithm is used to dynamically group the particles according to their density. Initially, all particles are divided into single-element clusters. Adjacent clusters are merged in ascending order of density. Merging is performed only when the maximum density difference between two clusters is less than 0.08 g / cm³. The final output is a set of sorted groups with a density range within each group ≤ 0.08 g / cm³. Step 6: Screen high-purity components with a predicted purity of ≥92% and input them into the thermodynamic-kinetic coupled model for smelting quality simulation; Step 7: If the pre-simulation indicators fail to meet the standards, reverse-optimize the feature weights and re-execute classification and clustering; only send the high-purity components into the DC electric arc furnace for melting, with a furnace temperature of 1600–2000℃ and an oxygen partial pressure maintained at 10. -6 Up to 10 -4 atm; Step 8: Monitor product quality in real time using XRF, infrared thermal imager and acoustic emission sensor. If MgO purity is <98.5% or porosity is >6%, automatically compensate smelting parameters and record deviation events for model iteration.

Images