Vanadium ore smelting process multi-parameter coupling intelligent regulation method and system

Abstract

The application discloses a kind of vanadium ore smelting process multi-parameter coupling intelligent regulation and control method and system, it is related to non-ferrous metal smelting automation control technical field, it solves the existing heat balance model to ore impurity fluctuation response tardy, online sensing lag and the deviation caused by process island and resource chain break problem;Through LIBS / XRF combined probe and microwave dielectric spectrum, impurity concentration and dielectric constant are obtained in real time, dynamic correction rheological disturbance is carried out using extended Kalman filter, and dynamic coupling response model is input to calculate thermodynamic interference factor and generate leaching and precipitation strategy;Parallelly execute mother liquor tail gas material flow collaborative optimization, construct multidimensional control vector U, in digital twin through proximal strategy optimization iterative update and distributedly issue to edge PLC, finally with online XRD feedback correction model weight;The application significantly improves the adaptive ability to complex vanadium ore composition disturbance, response accuracy of online control and resource recycling efficiency between processes.

Description

Technical Field

[0001] This invention relates to the field of automation control technology for non-ferrous metal smelting, and more specifically to a multi-parameter coupled intelligent control method and system for vanadium ore smelting process. Background Technology

[0002] In vanadium ore smelting, multi-parameter coupled control is a core technology for ensuring stable furnace conditions and efficient resource utilization. Traditional methods rely on empirical operation, which is insufficient to cope with the synergistic effects of complex ore composition and dynamic operating conditions. In recent years, data-driven intelligent control technology has been gradually introduced, achieving real-time matching of key parameters such as temperature field and load distribution through dynamic modeling and optimization algorithms.

[0003] Currently, intelligent control of vanadium ore smelting mainly relies on the synergy of thermodynamic models and data-driven algorithms. Taking the blast furnace smelting process as an example, patent CN112342327B proposes a semi-empirical control strategy based on theoretical combustion temperature (TFT): by establishing a quantitative relationship between blast temperature, oxygen enrichment rate, pulverized coal injection ratio and TFT (e.g., a 1% increase in oxygen enrichment rate raises the temperature by 41-50℃), and combining historical data to fit a general formula, rapid suppression of sudden fluctuations in furnace temperature can be achieved. Patent CN116050085A further integrates a heat balance calculation model, constructs a heat balance evaluation rule library, and predicts furnace condition trends and generates control commands based on real-time parameters (blast pressure, gas composition, etc.) from the blast furnace data acquisition system. For the grinding-leaching process, patent CN113762386A develops an adaptive matching mechanism for operating conditions, which generates typical operating condition labels based on historical grinding and grading data clustering, dynamically adjusts parameters such as ball mill speed and acid leaching agent flow rate, and solves the problem of rigid operating conditions with fixed control limits. At the system coordination level, the CN119668109A patent adopts a recurrent neural network model, with target control parameters (calcination temperature, leaching pH) as the input layer and control point parameters (waste gas concentration, slag-iron ratio) as the output layer, to train an intelligent adjustment model for smelting load and optimize the multi-unit linkage efficiency.

[0004] However, the aforementioned technologies have revealed some problems in practical industrial applications. First, the thermal balance model and operating condition matching mechanism assume stable furnace charge composition, but the content of impurities such as silicon, aluminum, and chromium in vanadium ore fluctuates drastically. The static model does not embed a thermodynamic coupling response mechanism under impurity interference, resulting in the system's inability to correct key process parameters in real time when ore composition changes abruptly. For example, a surge in silicon and aluminum impurities causes silica gel to encapsulate vanadium oxides during leaching, leading to a significant decrease in leaching efficiency. Second, the high dependence of intelligent algorithms on real-time data and the lag of the sensing layer create a sharp contradiction. Key parameters such as gas composition and ion valence state rely on offline detection, and long-term analysis data is out of sync with model requirements. The filtering delay in multi-source signal transmission further amplifies execution deviations. A typical manifestation is that in vanadium precipitation wastewater treatment, the redox potential (ORP) electrode drifts due to high ammonia nitrogen interference, the reducing agent dosage deviates from the preset range, and chromium precipitation rate fluctuates drastically. More seriously, the independent control architecture between processes leads to the break in the material circulation chain. For example, the ammonia-rich nitrogen wastewater from vanadium precipitation mother liquor is difficult to link with the roasting tail gas purification unit, the ammonia resource recycling efficiency is far lower than the design expectation, and additional auxiliary materials are consumed and the environmental burden is aggravated. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention discloses a multi-parameter coupled intelligent control method and system for vanadium ore smelting processes, aiming to solve the problems in the background technology.

[0006] To achieve the above-mentioned technical effects, the present invention adopts the following technical solution:

[0007] A multi-parameter coupled intelligent control method for vanadium ore smelting process includes:

[0008] Step 1: In real time, the ore flow is scanned using a probe coupled with laser-induced breakdown spectroscopy and X-ray fluorescence to obtain the concentration matrix and valence distribution data of impurity elements;

[0009] Step 2: Based on the impurity element concentration matrix and the real-time measured dielectric constant of the slurry using microwave dielectric spectrum, the signal drift caused by the perturbation of the slurry rheological properties is dynamically corrected by the extended Kalman filter fusion algorithm, and a standardized impurity spectrum vector is output.

[0010] Step 3: Input the standardized impurity spectrum vector into the dynamic coupling response model, calculate the impurity thermodynamic interference factor, and generate acid concentration gradient, reaction time optimization command and precipitant addition strategy;

[0011] Step 4: Simultaneously collect the ammonia ion concentration of the vanadium precipitation mother liquor and the hydrogen chloride flow rate of the roasting tail gas. Solve the optimal solution of the ammonia-chlorine neutralization molar ratio through the material flow coordination method to obtain the coordinated adjustment command of the mother liquor spraying rate and the tail gas fan.

[0012] Step 5: Couple the output parameters of Step 3 and Step 4 to construct a multidimensional control vector U. Iteratively update U in the digital twin through a near-end strategy optimization algorithm to output the calcination oxygen partial pressure and vanadium precipitation pH setpoints.

[0013] Step 6: Distribute the multi-dimensional control vector to the edge PLC based on the Storm distributed architecture to execute dynamic dosing of leachate and variable frequency speed regulation of exhaust gas fan;

[0014] Step 7: Monitor the slag phase composition in real time using online XRD, compare it with the twin prediction value in Step 5, generate the model error, and feed it back to Step 3 to correct the impurity weight coefficient.

[0015] As a further technical solution of the present invention, a multi-parameter coupled intelligent control system for vanadium ore smelting process includes: a component acquisition module, which is used to receive LIBS / XRF probe data from a host computer, extract impurity concentration matrix and valence distribution through wavelet packet decomposition and principal component regression, and output the original spectrum data;

[0016] The Kalman filter module is used to fuse and compensate for slurry rheological disturbances based on the original spectrum data and the dielectric constant measured by microwave dielectric spectrum, and output a normalized impurity spectrum vector by means of extended Kalman filter algorithm;

[0017] The coupled response module is used to calculate the thermal disturbance factor online by using the improved Maxwell–Garnett and unsteady-state energy conservation equations, and to solve for the acid concentration gradient, reaction time and dosing strategy.

[0018] The material coordination module is used for parallel collection of mother liquor. A linear programming model with mass conservation constraints was constructed based on the concentration and flow rate of HCl in the exhaust gas, and solved using Lagrange duality and the interior-point method. Molar ratio, generating spray volume and fan adjustment commands;

[0019] The strategy optimization module is used to map the coupling response and the output parameters of the material coordination module into a multidimensional control vector U. The PPO proximal strategy optimization algorithm is applied in the digital twin for iterative updates, and the output oxygen partial pressure and vanadium precipitation pH settings are output.

[0020] The distributed delivery module is used to receive control vectors based on the Storm distributed architecture and deliver them to the edge PLCs through routing, OPC-UA to Modbus TCP conversion and backpressure mechanism.

[0021] The XRD feedback module is used to acquire XRD diffraction data of slag online, perform quantitative analysis using the Rietveld algorithm, compare the data with twin predictions, generate residuals, and feed back impurity weight update information to the coupled response module through a gradient correction algorithm.

[0022] Based on the above technical solution, the positive and beneficial effects of the present invention are as follows:

[0023] This scheme couples and integrates online impurity composition analysis and self-correcting thermodynamics, so that the concentration information of impurities such as silicon, aluminum, and chromium directly affects the heat balance weight and energy conservation equation. Sudden changes in ore composition trigger real-time reconstruction of dynamic heat flux and reaction parameters, thereby eliminating the blind spot of the static model's assumption of steady-state composition. This prevents the phenomenon of vanadium oxide encapsulation by silica gel caused by impurities in complex ores from causing a sharp drop in leaching efficiency.

[0024] Through multi-source signal fusion and time delay compensation mechanism, dielectric constant, ORP, potential and impurity spectrum are fused online. Time-domain complementarity and predictive compensation are introduced in extended Kalman filtering. The errors of sensing hysteresis and electrode drift are corrected in time. The intelligent algorithm always outputs the addition command based on the latest system state vector, so that the addition amount of reducing agent and precipitant is highly consistent with the preset strategy, eliminating the contradiction between long-term offline detection and online control.

[0025] The solution executes cross-process closed-loop scheduling and wastewater tail gas linkage resource reconstruction in parallel, enabling the calculation of ammonia nitrogen-HCl neutralization in vanadium precipitation mother liquor and the calcination tail gas purification parameters to be solved in synergy. The mother liquor spraying and tail gas recirculation ratio form a complete closed-loop material flow, breaking the island of single process. The solution coordinates the control of each link with the interior point method and the Lagrange dual decomposition algorithm, thereby improving the ammonia resource recovery rate and tail gas purification efficiency at the same time and significantly reducing the consumption of external auxiliary materials. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0027] Figure 1 This is an architecture diagram of a multi-parameter coupled intelligent control method for vanadium ore smelting process according to the present invention;

[0028] Figure 2 This is a flowchart of the working method of step 1 of the present invention;

[0029] Figure 3 This is a schematic diagram illustrating the working principle steps of the extended Kalman filter fusion method of the present invention.

[0030] Figure 4 This is a schematic diagram illustrating the working principle of the material flow coordination method of the present invention.

[0031] Figure 5This is a schematic diagram of the multi-parameter coupled intelligent control system for the vanadium ore smelting process of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] To facilitate understanding of this embodiment, a detailed description of a multi-parameter coupled intelligent control method for vanadium ore smelting process disclosed in this application embodiment will be provided first. Please refer to [link to relevant documentation]. Figure 1 The diagram shows the steps of a multi-parameter coupled intelligent control method for vanadium ore smelting. The steps of this method include:

[0034] Step 1: The ore flow is scanned in real time using a laser-induced breakdown spectroscopy (LAS) coupled with X-ray fluorescence probe to obtain the impurity element concentration matrix and valence state distribution data; please refer to [link to relevant documentation]. Figure 2 The specific working method is as follows: First, an optical beam splitter is used to coaxially align the high-energy pulsed laser with the output beam of the multi-target X-ray tube, and the signal acquisition time window of the high-energy pulsed laser and the multi-target X-ray tube is aligned through a timing control circuit to obtain the original pulse waveform; then, based on the ore flow velocity feedback signal obtained by the photoelectric encoder, the probe scanning step size and dwell time are adjusted in real time, and the X-ray tube voltage is automatically switched to adapt to the excitation depth requirements of ores with different particle sizes; finally, the characteristic X-ray energy spectrum of the atomic ion emission spectrum generated by the laser-induced breakdown spectrum and the micro-focal spot X-ray fluorescence is input into the multivariable spectral deconvolution module based on non-negative matrix factorization. By decoupling the overlapping peaks and correcting the matrix effect, the impurity concentration matrix and vanadium valence state distribution data containing spatial coordinates are obtained. The multivariate spectral deconvolution module based on nonnegative matrix factorization decomposes the original spectral matrix into a basis vector matrix and an abundance coefficient matrix using an alternating least squares iterative optimization method. The basis vector matrix contains a standard spectral line library of known elements. In each iteration of the alternating least squares iterative optimization method, a matrix effect correction mechanism is adopted to dynamically adjust the weight coefficient of the V spectral line according to the quenching factor of the Fe element-dominant spectral region. If Cr-V spectral peak overlap is detected, the overlapping peak decoupling protocol is triggered, and the overlapping peak area is separated using a Gaussian mixture model to obtain the pure elemental concentration matrix and vanadium valence state distribution data after matrix correction and peak decoupling.

[0035] It should be noted that the probe tip integrates a high-energy pulsed laser (wavelength 1064nm, pulse energy adjustable range 10–100mJ) and a multi-target X-ray tube (anode voltage 20–60kV, switchable target material Fe / Cr / Al). The "multi-target X-ray tube" refers to the switching of the emitted beam between different cathode targets within the same cavity via mechanical or electronic switching. In this application, the "combination" of the two spectral signals, "laser-induced breakdown spectroscopy (LIBS)" and "X-ray fluorescence (XRF)," refers to the spatial and temporal multiplexing of the two excitation and acquisition optical paths within the same measurement probe head using an optical beam splitter and reflecting prism, ensuring that the same batch of ore particles receives continuous excitation from both types of spectral signals at the same spatial location. The probe head is fixed to the side edge of the ore conveyor belt, and the probe scanning method is a transverse reciprocating scan, with the scanning speed coordinated with the ore conveying speed.

[0036] The excitation and acquisition process is synchronized by a programmable timing control unit (FPGA-based). Key parameters and triggering conditions are as follows: when the photoelectric encoder feeds back the conveyor belt speed... Exceeding the threshold (Typically 0.5 m / s) "Probe dwell time" = The FPGA determines the beam spot length. (Optional range 50–200μm) Real-time calculation, specifically: ;in and The measurement accuracy is determined by the laser jitter detection module and the encoder resolution, respectively. The value triggers the FPGA to acquire the laser pulse and X-ray tube output voltage within a specific window width. Alignment adjustments are made to ensure that the original pulse waveforms of LIBS and XRF signals are acquired sequentially at the same particle point. When When the signal exceeds the above range, the FPGA triggers either a "low-speed mode" or a "high-speed mode" and adjusts the excitation energy and dwell time accordingly to ensure sufficient signal strength for each excitation. The scan control logic is based on the FPGA's internal timer, precisely controlling the probe's lateral movement step size by calculating the number of sampling points and trigger interval for each scan cycle. ,in The lateral velocity, The trigger cycle is determined by the motion accuracy of the robotic arm and the belt speed fluctuation.

[0037] In terms of timing control and triggering mechanisms, the Field-Programmable Gate Array (FPGA) undertakes the core timing scheduling function. Based on the input photoelectric encoder signal, the FPGA first triggers the laser to emit a high-energy pulse when the probe moves to a new sampling point, and simultaneously opens the analog front-end acquisition channel to capture the atomic and ion emission lines formed by laser-induced breakdown. The acquisition window length... The pulse width and spot decay time are determined by the laser acquisition process, and are typically set to tens of nanoseconds. After laser acquisition, a delay is applied... A high-voltage pulse is triggered (less than 10 nanoseconds) to output the X-ray tube and activate the XRF acquisition channel to obtain the characteristic X-ray energy spectrum. This delay mechanism ensures that the two excitations do not interfere with each other and are time-correlated. The FPGA simultaneously monitors the laser energy and the X-ray tube voltage and current signals. If the monitored values ​​deviate from the calibration range (e.g., the laser pulse energy deviation exceeds ±5%), the acquisition is immediately paused and a correction alarm is issued, requiring intervention from the host computer or maintenance personnel for adjustment.

[0038] In the signal acquisition and preprocessing section, the probe end is equipped with a multi-channel high-speed analog-to-digital converter (ADC), with a sampling rate of no less than 100MS / s per channel and a quantization accuracy of 12-16 bits. The acquired raw waveform data is first stored in the FPGA ring buffer and then packaged according to timestamp, spatial coordinates, and trigger ID, before being transmitted to the back-end server via PCIe or Ethernet bus. Data preprocessing includes baseline correction, low-pass filtering, and power frequency noise suppression, using a finite impulse response (FIR) filter. The filter coefficients and cutoff frequency are flexibly configured according to the on-site electromagnetic environment and signal bandwidth characteristics, and are not limited thereto.

[0039] In the multivariate spectral deconvolution algorithm, the backend server organizes the processed time-domain waveform matrix into a two-dimensional spectral matrix S, where rows correspond to discrete wavelength channels (W channels) and columns correspond to N excitation sampling points. The algorithm employs an improved nonnegative matrix factorization (NMF) method. The objective function, in addition to the traditional squared error term, incorporates L1 norm sparsity regularization to improve the resolution for trace spectral lines. It uses alternating least squares (ALS) iteration: first, the abundance coefficient matrix H is fixed, and nonnegative optimization is performed only on the basis vector matrix W; then, W is fixed and H is updated in reverse. Each iterative subproblem can be solved using multiplicative updates or gradient projection methods. The iteration termination condition can be set to a relative change in the decomposition residual of less than 10⁻⁻⁶. 4 The maximum number of iterations can be reached, ranging from 50 to 100, which can be determined based on computational resources and accuracy requirements; no limit is imposed on this. The column vectors of the basis vector matrix W correspond to the calibrated standard spectral line library of elements, including major spectral lines of V, Si, Al, Cr, Fe, and their typical compound spectral regions.

[0040] Matrix effect correction, as a built-in subprocess of the NMF module, extracts the average attenuation factor η_Fe in the basis vector matrix W corresponding to the Fe-related spectral region (e.g., 640–660 nm) after each ALS iteration, and uses this to dynamically adjust the weights corresponding to the V spectral lines: a linear weighting mechanism is employed. κ is an empirical correction coefficient, ranging from 0.1 to 0.5. "The specific value can be determined based on on-site standard sample tests and is not limited thereto." This mechanism compensates for the suppression or enhancement effect of the slag matrix on the V emission peak, making the V concentration calculation more accurate.

[0041] When the overlap between the Cr and V peaks in a certain row of the abundance coefficient matrix H exceeds a set threshold δ (e.g., 30%) during the ALS iteration, the system automatically triggers the overlapping peak decoupling protocol. The protocol first extracts the spectral cross-section curves in the overlapping band and then calls a Gaussian mixture model (GMM) for multi-peak fitting. The GMM fitting estimates the peak separation parameters (weights π_l, center wavelength μ_l, variance σ_l) using the expectation-maximization (EM) algorithm and outputs the pure peak area after parameter convergence, replacing the abundance value for that band in the original ALS result. The decoupling iteration order and the number of EM iterations can be determined automatically based on the peak width and signal-to-noise ratio, without any restrictions.

[0042] After deconvolution and correction, the system obtains a clean impurity abundance coefficient matrix H_clean, which is then used to construct a spatial coordinate mapping. During the scanning process, the FPGA continuously records the probe's lateral movement position and trigger number, and sends these along with the spectral data. The backend server can then reconstruct the two-dimensional spatial coordinates (x, y) of each sample point based on the trigger number. Valence state distribution is determined by analyzing different valence state V spectral lines (such as V...). 2 ⁺At 480nm, V 3 ⁺At 492nm, V 4 The peak area ratio (around 504 nm) was calculated and combined with the impurity concentration matrix to form the final output dataset.

[0043] Finally, the impurity concentration matrix C_imp(x,y) containing spatial coordinates and the valence distribution vector C_valence are written to the central control unit via shared memory or cache, and then distributed by the host computer to the subsequent Kalman filter fusion and coupling response modules via the OPC-UA protocol. The OPC-UA session uses a secure channel and subscription mode, and its heartbeat cycle and historical data retransmission strategy can be flexibly adjusted according to the on-site network bandwidth and control cycle, without limitation. If the heartbeat times out or communication is interrupted, the system automatically caches the latest data and reconnects periodically until recovery.

[0044] The parameters and tools involved in each step of this embodiment include, but are not limited to: optical beam splitter insertion loss <1dB, FPGA timing jitter <5ns, ADC quantization accuracy 12bit, ALS maximum iteration count 100 times, and GMM convergence threshold 10⁻⁻⁶. 4 Features such as the OPC-UA heartbeat cycle of 1 second can be flexibly configured according to engineering needs, and there are no restrictions on this.

[0045] Step 2: Based on the impurity element concentration matrix and the real-time measured dielectric constant of the slurry using microwave dielectric spectrum, the signal drift caused by the perturbation of the slurry rheological properties is dynamically corrected by the extended Kalman filter fusion algorithm, and a standardized impurity spectrum vector is output.

[0046] It is important to clarify the context of step 2: Step 2 applies to the dynamic treatment of vanadium ore slurry (such as the slurry flow area in pre-leaching pipelines and stirred reaction tanks). In this scenario, the rheological properties (such as viscosity, shear rate, and flow velocity distribution) of the slurry continuously change dynamically due to variations in flow rate, particle agglomeration / dispersion, and temperature fluctuations. This leads to distortion of the laser-induced breakdown spectrum (LIBS) spot energy and changes in the electromagnetic field coupling efficiency of the microwave dielectric spectrum, causing the impurity concentration matrix and dielectric constant measurement to deviate from the true value (i.e., "signal drift"). If this drift is not corrected, the dynamic coupling response model in the subsequent step 3 will generate invalid process instructions based on erroneous inputs, resulting in problems such as decreased leaching efficiency and waste of precipitant.

[0047] It should be noted that Kalman filtering is a recursive algorithm for state estimation of dynamic systems, and the Extended Kalman Filter (EKF) fusion algorithm is an extension of it for nonlinear scenarios. It achieves iterative updates of state and covariance through local linearization (Jacobi matrix). "Fusion" refers to simultaneously utilizing LIBS impurity concentration and microwave dielectric spectrum dielectric properties, and using the EKF framework for complementary correction. Unexpected changes in slurry flow state (laminar / turbulent switching, changes in flow velocity gradient) or physical properties (viscosity, shear modulus) are the core causes of "signal drift" (e.g., a pipe flow velocity jumping from 1 m / s to 1.5 m / s, viscosity changing from 10 mPa·s to 15 mPa·s, etc.; specific thresholds are set according to process accuracy requirements and can be determined independently). The standardized impurity spectrum vector is the feature vector that integrates the impurity concentration matrix and dielectric properties according to a unified dimension (e.g., concentration normalization, dielectric constant standardization) after EKF correction. Its "standardization" is reflected in the unification of data dimensions and the elimination of drift errors, and it can be directly input into subsequent models.

[0048] The core input for step 2 is the impurity concentration matrix at time k. With respect to the dielectric constant of the slurry at time k :in Obtained by scanning the slurry flow using the LIBS system: LIBS excites plasma with a laser, collects the emission spectrum and deconvolves it (such as polynomial fitting, wavelet denoising, the algorithm is selected according to the equipment or process requirements, and can be determined by yourself), to obtain the concentration matrix of impurities such as silicon, aluminum, and chromium (the matrix dimension is determined by the spatial resolution of LIBS, such as 10×10 pixels, the number of pixels is adjusted according to the detection accuracy, and can be determined by yourself). Measurement using a microwave dielectric spectroscopy system: A microwave probe (coaxial / waveguide probe) is inserted into the pipe, emitting a signal at a specific frequency (e.g., 2.45 GHz, 9.15 GHz, selected based on the dielectric properties of the slurry and the penetration depth, which can be determined independently). The dielectric constant is calculated using the reflection / transmission coefficients (the calculation method depends on the probe type, such as transmission line model or resonant cavity model, which can be determined independently). Figure 3 As shown, the specific execution logic of extended Kalman filter fusion in the implementation is as follows:

[0049] s100, at time k, will and Organization as a measurement vector EKF iteration triggering and rheological disturbance linkage: When the slurry flow rate changes by more than 5% / s (e.g., the flow meter shows the flow rate changing from 1.2m / s to 1.3m / s, the change rate is approximately 8.3% / s) or the viscosity fluctuates by more than 10% (e.g., the viscometer reading changes from 12mPa·s to 14mPa·s, the fluctuation is approximately 16.7%) (the trigger threshold is set according to the process tolerance and can be determined by the user), the EKF encryption iteration frequency is increased (e.g., from 1Hz to 5Hz, the frequency strategy is implemented according to PID or rule base and can be determined by the user); if the disturbance is mild (flow rate fluctuation <2% / s, viscosity fluctuation <5%), the base frequency (e.g., 1Hz) is maintained to balance resources and accuracy.

[0050] s200, the prediction phase is based on the state of the previous time step. External disturbances Derivation of prior states With prior covariance: where, state vector Defined as the "impurity concentration-rheological perturbation coupling state", it includes the concentrations of silicon, aluminum, and chromium (cSi, cAl, cCr) and signal drift parameters (drift amplitude a, velocity v), i.e. = (The dimension can be adjusted according to the type of impurities and the complexity of the drift model, and can be determined by yourself).

[0051] External disturbances Measurable disturbances such as the change in slurry pump frequency Δf and the change in agitator speed Δn are collected by sensors (frequency converter, tachometer) and converted into disturbance factors (e.g., for every 1Hz increase in Δf, the predicted value of cSi increases by 0.05%, and the mapping relationship is determined by process modeling and can be determined by oneself).

[0052] The mapping function f(⋅) is used to describe the state evolution law. If it is assumed that the concentration drifts linearly with rheology, then Where A is the state transition matrix (diagonal elements such as 0.99 represent the concentration decay coefficient, and off-diagonal elements such as 0 represent state independence, determined based on historical data, which can be determined manually), and B is the perturbation input matrix (elements represent...). The weights of the influence on the state can be determined by the experiment (based on calibration). The process noise (covariance Qk is determined by LIBS signal wavelet packet decomposition: the spectrum is decomposed into n layers of wavelet packets (e.g., 3 layers, the number of decomposition layers is selected according to the signal bandwidth, which can be determined by yourself), the sub-band energy covariance is calculated, and it is used as the initial value of Qk and updated in real time).

[0053] To integrate "impurity concentration" and "dielectric constant", the MG dielectric mixing equation is introduced (Formula 1: ), and the "prior prediction value of impurity concentration" Mapped to "predicted dielectric constant value" ":in, The dielectric constant of the slag matrix (e.g., the dielectric constant of vanadium slag matrix at 25℃ and 1 atm, which can be determined by yourself based on offline test calibration); The dielectric constant of impurity particles (such as silicon) =11.7, Aluminum =9.0, Chromium =14.0, which can be determined by yourself based on material databases or experimental measurements); ff is the volume fraction of impurities (from...). Conversion, such as converting mass fraction c to volume fraction , (Silicon 2.33g / cm³) 3 Aluminum 2.70 g / cm³ 3 Chromium 7.19 g / cm³ 3 )and (e.g., 3.5g / cm) 3 (Based on experimental measurements, it can be determined independently.)

[0054] Jacobian matrix For the MG equation pair The partial derivative matrix, i.e. .because yes The function, Non-zero element corresponding to each impurity concentration pair Partial derivatives (such as) Indicates when the silicon concentration changes by 1%. The change in concentration (the amount of change). The Jacobian matrix is ​​achieved by numerical differentiation (such as the finite difference method, with a step size of 0.01% concentration change, the step size can be adjusted according to the accuracy and can be determined by the user) or analytical differentiation. Its function is to locally linearize the MG equation and provide a basis for updating EKF measurements.

[0055] s400. Through wavelet packet decomposition and sequence variance analysis, process noise covariance is realized. Covariance of measurement noise Adaptive adjustment: where process noise covariance That is, the LIBS spectral signal is decomposed into wavelet packets (e.g., using the db4 wavelet basis and decomposing into 3 layers; the wavelet basis and the number of layers are chosen according to the time-frequency resolution and computational cost, and can be determined by the user), to obtain the energy sequence of each frequency sub-band; the sub-band energy covariance matrix is ​​calculated, and the diagonal elements are taken as... Updated values ​​(high-frequency subband covariance is large, corresponding to fast-changing noise; low-frequency subband covariance is small, corresponding to slow-changing noise). Measure noise covariance. That is, to calculate the variance of the historical measurement sequence of microwave dielectric spectrum (e.g., using a sliding window to take the 10 most recent sampling points; the window size can be adjusted according to the signal stationarity and can be determined by the user), and then... The diagonal element (variance of dielectric constant measurement noise) is set as σε2; if the variance of the LIBS concentration sequence σc2 changes significantly (e.g., increases by 50%), it is adjusted accordingly. Medium concentration measurement noise variance elements ensure a comprehensive reflection of multi-source measurement noise.

[0056] In the S500 measurement update phase, the optimal fusion of "prior prediction value" and "actual measurement value" is achieved through Kalman gain.

[0057] Finally, the standardized impurity spectrum vector is obtained through the Kalman update equation.

[0058] In practical applications, the LIBS scanning frequency, microwave dielectric spectrum bandwidth, and EKF iteration cycle need to be adjusted according to the on-site operating conditions (e.g., for a pilot line commissioned for 3 months, the LIBS frequency is increased from 5Hz to 8Hz, and the concentration measurement delay is reduced from 200ms to 150ms; the optimization target is determined based on the production line response speed and can be determined independently). For special minerals (such as high-calcium vanadium ore) or extreme rheological scenarios (such as slurry containing fibrous impurities), customized EKF state equations and mapping functions are required (e.g., introducing a nonlinear transfer model and adding a fiber dielectric correction term; the direction of modification is determined based on the mineral characteristics analysis and can be determined independently).

[0059] Step 3: Input the standardized impurity spectrum vector into the dynamic coupling response model to calculate the impurity thermodynamic interference factor and generate acid concentration gradient, reaction time optimization instructions and precipitant addition strategy; the dynamic coupling response model includes an input preprocessing layer, a feature fusion layer, a thermodynamic factor calculation layer, a dynamic coupling analysis layer, a multi-objective optimization layer and an instruction synthesis layer.

[0060] In implementation, the dynamic coupled response model is deployed in the blast furnace smelting monitoring center, forming a closed loop with the upper-level EKF fusion module, the lower-level material coordination and strategy optimization module, and the edge execution layer. The standardized impurity spectrum vector and online dielectric constant time series output by the EKF fusion module are used as the data source for the input preprocessing layer. The "online dielectric constant measurement" in the system uses a coupled electrode array and a network analyzer to achieve dielectric measurement of the impurity-water-solid three-phase mixture, returning the dielectric constant in complex form. In this paper, only the real part is taken for subsequent processing. All parameters such as "frequency" and "sampling rate" can be set according to the capabilities of the on-site equipment and are not limited.

[0061] The input preprocessing layer first performs a normalization transformation on the original input. This transformation uses a minimum-maximum linear mapping method to map the vector components to the [0,1] interval to eliminate the influence of different physical quantities. The minimum and maximum values ​​are automatically read from field standard samples or historical databases. The system provides a real-time update interface to add new standard samples as needed. After normalization, the signal undergoes three-level wavelet packet denoising on the time series. This process uses the Daubechies wavelet basis and automatically sets a threshold based on the noise estimation index after each layer decomposition. Detail coefficients higher than the threshold are removed, and the approximate original signal is reconstructed through energy conservation. The threshold calculation uses a robust extreme value estimation method: the median of the absolute value of the high-frequency coefficients in the first layer is multiplied by an empirical factor of 1.4826 to obtain the noise standard deviation estimate, and then multiplied by a threshold factor (default 3). This factor can be adjusted through the background configuration interface. The trigger condition for this layer is to execute immediately after the input data acquisition is completed. The action is automatically scheduled by the preprocessing scheduler, and the output includes a noise confidence index, indicating the reliability of the denoised signal.

[0062] The preprocessed temporal feature vectors are written to high-speed shared memory, and the feature fusion layer is notified to load them. The feature fusion layer processes multimodal features within the temporal window in parallel based on a multi-head attention mechanism. The window length L defaults to five times the number of samples corresponding to the control period, and can be set between 5 and 20 in the parameter interface. The specific value can be determined according to system latency and real-time requirements, and is not limited. In each attention head, the system calculates a query-key-value matrix based on the fused features and historical window data. Its dimension is determined by the number of heads h set during model training and the hidden dimension of each head. The decision is made, with a typical value of h=8. =64. Attention weights are calculated using a scaled dot product mechanism and normalized using the softmax function. If the maximum attention weight calculated for any head is lower than a preset threshold τ=0.1, the system automatically invokes an adaptive recalibration mechanism. This mechanism collects the raw normalized data from the most recent Nr=10 windows, re-estimates the projection matrix parameters of the head using a regularized least squares method, and verifies whether the attention peak has rebounded within three fine-tuning iterations. The recalibration process is executed in parallel on the GPU, typically lasting no more than 50 milliseconds, to avoid blocking the real-time control flow. All head outputs are concatenated and a fused feature is generated through a linear mapping layer. Meanwhile, the attention heatmap is stored in the performance monitoring library for subsequent debugging and visualization.

[0063] Fusion features The data is passed to the thermodynamic factor calculation layer, which calls the built-in thermochemical data interface based on current operating parameters—primarily including the real-time measured furnace temperature. and pressure And the impurity concentration vector [c1,…,c5]—online retrieval of the corresponding standard enthalpy change. With standard entropy change Data. The enthalpy and entropy change database was obtained through experimental standards or literature calibration, with a data accuracy of up to one-thousandth. The indexes in the tables use impurity chemical formulas and operating condition ranges as keys, which can be determined according to the actual situation.

[0064] After retrieval, the thermodynamic factor calculation layer performs temperature correction, converting the standard-state enthalpy-entropy to the current operating condition. Enthalpy correction is approximated using a linear temperature difference:

[0065]

[0066] in For the first The specific heat at constant pressure of the impurities, The standard reference temperature is used; entropy correction employs the isobaric-isothermal approximation.

[0067]

[0068] These two calculations are implemented in parallel via vectorization in the numerical computation module, with latency controlled to within tens of milliseconds, meeting the requirements for second-level updates. After enthalpy-entropy correction, the system maps the contribution value of each impurity purification value according to the Gibbs free energy perturbation calculation formula:

[0069] The formula is executed with high-precision floating-point operations at the implementation layer, and the output vector [ΔG1,…,ΔG5] represents the thermodynamic interference factor of each impurity on the reaction system. The thermal perturbation gradient is calculated simultaneously within the layer, and the sensitivity matrix of adjacent concentration perturbations to ΔG is obtained through numerical difference, which is used for local mesh mapping in subsequent dynamic coupling analysis layers. The trigger condition is whenever the eigenvectors are fused... or online temperature Updates are executed immediately; action scheduling is completed by the hot factor scheduling unit, which broadcasts the calculation start signal through the event bus and writes the results to the shared memory of the coupled analysis layer after the calculation is completed.

[0070] The dynamic coupling analysis layer couples the thermal perturbation factor vector with a pre-established 3D reactor geometric mesh using the finite element method. The geometric model is generated offline using CAD and CFD tools and imported into the finite element solver, including mesh node coordinates, fluid and solid domain boundary conditions, and fluid dynamic parameters (such as viscosity and density). Within the layer, interpolation is first performed on the mesh node distribution based on the thermal perturbation factor, assigning a local energy perturbation value to each node. Subsequently, under boundary condition constraints including velocity and pressure fields, the partial differential equations are discretized using the software's built-in unsteady-state heat transfer and mass transfer weakly coupled solver, following the Galerkin method. During the solution process, the system monitors the residual convergence. If the iterative residual does not decrease for multiple iterations or reaches the maximum iteration count of 100, a degradation mechanism is triggered, automatically switching to a one-dimensional pipe model simulation to quickly obtain approximate results. After the simulation, the local acid concentration gradient ∇[H+] distribution map and the mass transfer resistance coefficient Φmix distribution map are saved as vector and matrix formats and written to the cache of the multi-objective optimization layer.

[0071] The multi-objective optimization layer links the coupling analysis results with the precipitant model to construct a three-objective optimization problem: minimizing the root mean square error of the acid concentration gradient, minimizing the reaction time, and minimizing precipitant consumption. The algorithm used is NSGA II, which is implemented in the optimization engine through the following process: First, an initial population is generated centered on the previous cycle instruction Uk−1, combined with a small perturbation. The population size is typically set to 50 to 150 individuals, with individual dimensions corresponding to the acid concentration rate, reaction time, and dosage, and can be automatically adjusted according to system load during runtime. The "perturbation amplitude" is determined by an adaptive noise scheduler based on system stability indicators, and the noise is generated based on a Gaussian distribution. During each generation of evolution, the algorithm stratifies the population through non-dominated sorting, classifying individuals into levels based on their dominance in the target space; then, crowding distance is used to select individuals within the same level to maintain diversity; the crossover operation uses simulated binary crossover (SBX), crossoverling the control variables of parent individuals according to probability; the mutation operation uses a multinomial mutation method, and automatically increases the mutation probability when the population convergence is below a threshold of 0.02 to prevent premature convergence. The entire iterative process is executed within a parallel computing framework, and crossover and mutation operations can be performed concurrently on multi-core CPUs or GPUs.

[0072] There are two termination conditions for the iteration: reaching the maximum number of generations (e.g., 100 generations) or the change in the main crowding of the Pareto front over consecutive g generations is less than a preset threshold. After termination, the algorithm outputs the Pareto front solution set to the policy synthesis unit and selects a representative solution based on the current production priority (e.g., prioritizing yield or minimizing reagent consumption). During the optimization process, the system continuously monitors the algorithm's convergence curve and performance indicators, and triggers a global restart mechanism when it detects that the population diversity is decreasing too rapidly, regenerating the population to restore the search range.

[0073] The instruction synthesis layer selects a representative solution from the Pareto solution set according to production priority rules, and performs a weighted linear combination of three control variables: acid concentration adjustment rate, reaction residence time, and precipitant dosage. The combination weights are issued by the production management module through the interface and can be set to different modes such as output priority, energy consumption priority, or resource recovery priority. The synthesized control sequence is framed according to the control cycle. Each frame is encapsulated by the OPC UA protocol and sent to the edge PLC for parallel execution with the mother liquor spraying and tail gas regulation instructions generated in step 4. The execution layer interfaces with the corresponding dosing pump station and valve actuator. After receiving the instruction, it drives the hardware in closed-loop PID mode. The execution feedback signal is uploaded in real time through the fieldbus and incorporated into the EKF fusion input of the next cycle.

[0074] Step 4: Simultaneously collect the ammonia ion concentration in the vanadium precipitation mother liquor and the hydrogen chloride flow rate in the roasting tail gas. Solve for the optimal ammonia-chlorine neutralization molar ratio using a material flow coordination method to obtain the coordinated adjustment command for the mother liquor spray rate and the tail gas blower. Please refer to [link to relevant documentation]. Figure 4Specifically, the material flow coordinated method monitors the ammonia ion concentration in the vanadium precipitation mother liquor online using an ion chromatograph, simultaneously measures the hydrogen chloride concentration in the roasting tail gas in real time using a tunable diode laser absorption spectroscopy, and collects the waste gas volumetric velocity using a turbine flow meter to construct a gas-liquid two-phase flow field model. Based on the Reynolds-Navier-Stokes equations, the trajectory of the spray droplets and the diffusion path of the waste gas are calculated. Then, a sequential quadratic programming algorithm is used to solve for the optimal ammonia-chlorine molar ratio, with the objective function set at a molar ratio of 1.05 ± 0.02. If the ammonia ion concentration fluctuates beyond ± 5%, the diaphragm pump frequency is adjusted in real time using an incremental PID control protocol. When the waste gas flow rate change rate exceeds a preset threshold, the induced draft fan speed is pre-adjusted through a feedforward compensation mechanism. It should be noted that "online monitoring" refers to continuous data collection and real-time output during the vanadium ore smelting process, which differs from periodic offline sampling; "ammonia ion concentration" specifically refers to the concentration of ammonia ions in the vanadium precipitation mother liquor. The content was obtained by quantitative analysis using ion chromatography; "Diode Laser Absorption Spectroscopy (TDLAS)" achieves quantitative monitoring by utilizing the relationship between absorption intensity and concentration at a specific wavelength. This scheme uses a tunable semiconductor laser to measure the HCl absorption spectrum in real time at the 1.72µm absorption band; "Turbine Flow Meter" is used to determine the exhaust gas volumetric flow rate, and its output is Nm³ at standard temperature and pressure. 3 / h; "Two-phase flow field model" refers to a numerical simulation environment that couples the motion of the gas phase (exhaust gas) and the liquid phase (spray droplets); "Sequential quadratic programming (SQP)" is an iterative algorithm for constrained nonlinear optimization problems, used to solve the optimal dosing strategy within the target molar ratio range; "Incremental PID control protocol" refers to adjusting the controller output according to the deviation increment to correct the pump frequency when the deviation exceeds the limit; "Feedforward compensation" refers to adjusting the induced draft fan speed in advance according to the measured rate of change of exhaust gas flow to reduce system response time delay.

[0075] In practical implementation, the system first continuously extracts vanadium precipitation mother liquor samples using an ion chromatograph, and then dilutes the mother liquor to a suitable concentration range for chromatographic detection using an automatic online diluent within the sampling flow path (common dilution ratios are 1:10 to 1:100, depending on the ion concentration range). The chromatograph employs gradient elution and a high-efficiency ion exchange column, operating at a temperature of 40 ℃ and a flow rate of 1.0 mL / min. A suppressor and conductivity detector are connected post-column, enabling one analysis and output per minute. Concentration C_NH4 (mg / L or mmol / L). When the change in C_NH4 exceeds ±5% relative to the previous sampling period, the system triggers the deviation monitoring submodule of the PID control protocol to adjust the diaphragm pump frequency in real time.

[0076] Simultaneously, a tunable diode laser absorption spectrometer was installed in the calcination exhaust gas pipeline. This instrument swept the frequency in the wavelength range of 1.712–1.716µm, and measured the light intensity. and Calculate HCl concentration by ratio (ppm or mg / m 3 The fundamental relationship follows the Beer-Lambert law:

[0077]

[0078] In the formula and These represent the incident and transmitted light intensities, respectively. This represents the absorption cross-section of HCl at that wavelength, typically taken as 1.2 × 10⁻⁻⁻⁶. 20 cm 2 / molecule, The optical path length is specified, and the probe is installed along the diameter of the exhaust pipe. The optical path length is selectable from 25 to 100 mm depending on the pipe diameter. The measurement is continuously updated at millisecond-level speeds, and the data is sent to the main control system via serial port or Ethernet.

[0079] Exhaust gas volumetric flow rate Qgas (Nm 3 The acquisition of the flow rate ( / h) is accomplished by a turbine flow meter. The impeller speed signal of the flow meter is converted into frequency by a photoelectric sensor, and then immediately converted into volumetric flow rate through the calibrated frequency-flow relationship. To ensure measurement accuracy, the flow meter is pre-installed with a heat exchanger and a straight pipe section interface. A straight pipe section of 10 times the pipe diameter is reserved at both the fluid inlet and outlet to reduce turbulence interference. The operating temperature and pressure compensation module converts the raw data to standard conditions.

[0080] The three data streams mentioned above are centralized in the main control system and merged into a synchronous sampling group via a high-speed data bus. The data set was first fed into a two-phase flow field model simulator. The model, based on Reynolds' time-averaged Navier–Stokes equations, describes the gas-phase flow field in the following form:

[0081]

[0082] The droplet motion is coupled using Lagrangian trajectory tracing, and the droplet equation is as follows:

[0083]

[0084] in The density is the gas phase density. Viscosity, For the gas phase velocity field, For droplet velocity, For the mass of the droplet, The drag coefficient, This includes volume forces and reaction forces. In this implementation, the reactor geometric model is discretized into a three-dimensional mesh, and the gas phase velocity field is solved using the finite volume method at the mesh nodes. Simultaneously, the initial positions and diameter distributions of the droplets are generated by random sampling (common diameters are 50–200 μm, and the distribution can be set according to the atomization characteristics of the sprayer). The droplet trajectory is then updated by integration along the time step. The model needs to run in parallel on a GPU or multi-core CPU cluster to achieve updates within 1 second.

[0085] After completing the two-phase flow field simulation, the system extracts data based on the current flow field. The constructed spray droplet-exhaust gas coupling field indices include the contact efficiency of HCl and NH4⁺ per unit volume, the average residence time of spray droplets in the pipe, and the residual concentration distribution of HCl in the exhaust gas. These indices serve as input to a sequential quadratic programming (SQP) algorithm. SQP formalizes the neutralization molar ratio optimization problem with the following constraints and objective function: Constraints: molar ratio target The value should fall within the range of 1.05 ± 0.02. SQP solves this by constructing a Lagrangian function and linearizing the constraints and quadratic approximation objective function in each iteration. The search direction is updated using Newton's method, and a new solution is obtained through step-size search. The iteration terminates when the change in the objective function is less than 10⁻⁶ or the number of iterations exceeds 50.

[0086] After the SQP algorithm completes its solution, the system obtains the optimal coefficients α∗ and β∗. The mother liquor spraying volume Qliq is calculated as follows: The exhaust gas fan extraction speed Qfan is calculated as follows: The two are then used as coordinated adjustment commands and sent to the diaphragm pump controller and the fan frequency converter.

[0087] If the NH4⁺ concentration fluctuates by more than ±5% (compared to the previous sampling value) during the above online monitoring, the system will immediately activate the incremental PID control protocol to adjust the frequency of the mother liquor diaphragm pump. The correction is made, and the incremental correction amount is defined as follows:

[0088]

[0089] in For the concentration increment deviation, Kp, Ki, and Kd are PID parameters. The PID controller is implemented differentially, directly outputting incremental commands to the pump frequency inverter. Simultaneously, when the rate of change of the exhaust gas volumetric velocity Qgas exceeds ±3%, a feedforward compensation mechanism is added to the fan control protocol: the velocity for the next cycle is predicted in advance based on the rate of change, and the required fan speed adjustment is calculated. Then, it is fused with the basic instruction Qfan obtained from SQP to eliminate system response lag, specifically as follows:

[0090]

[0091] in This is the feedforward gain coefficient, which is usually set between 0.5 and 1.5. The specific value can be determined according to the characteristics of the fan and the resistance of the pipeline network. This action is executed immediately by the feedforward control unit after receiving the flow rate change signal, and is issued after being combined with the incremental PID command.

[0092] Step 5: Couple the output parameters of Steps 3 and 4 to construct a multidimensional control vector U. Iteratively update U in the digital twin using a near-end strategy optimization algorithm, outputting the roasting oxygen partial pressure and vanadium precipitation pH setpoints. In implementation, the multidimensional control vector U includes a furnace condition control vector, a leaching parameter vector, a precipitation adjustment vector, a tail gas purification vector, a wastewater reuse vector, and a model calibration vector. The furnace condition control vector includes the roasting section blast flow rate, the furnace oxygen partial pressure setpoint, and the fuel injection rate. The leaching parameter vector includes the acid concentration gradient and reaction time at each stage in the leaching section. The precipitation adjustment vector includes the pH setting in the precipitation section, the precipitant addition acceleration rate at each stage, and the stage switching trigger threshold. The tail gas purification vector includes the tail gas extraction rate and the recirculation ratio. The wastewater reuse vector includes the mother liquor spray rate and the ammonia nitrogen recovery flow rate. The model calibration vector consists of the impurity weight coefficient update rate based on online XRD feedback, the adaptive gain of the heat balance model, and the rheological disturbance compensation factor. The above sub-vectors achieve data mapping and flow through a unified digital twin state interface. The outputs of the furnace condition control vector, leaching parameter vector, and tail gas purification vector are used as inputs to the sedimentation adjustment vector, wastewater reuse vector, and model correction vector. Finally, they are iteratively generated in the near-end strategy optimization algorithm to drive the PLC edge to perform multi-stage linkage control such as leaching dosing, tail gas speed regulation, and weight self-correction.

[0093] It should be noted that in this scheme, "near-end policy optimization algorithm" specifically refers to a reinforcement learning algorithm that iteratively updates policy network parameters based on the trust region concept in offline or online digital twin environments; "importance sampling mechanism" refers to weighting samples according to the probability ratio of actions before and after policy update to reduce variance when collecting policy gradients; "probability ratio exceeding confidence interval threshold" means that if the probability ratio of the same action output by the new and old policies deviates from the interval [1–ε, 1+ε], it needs to be pruned; "generalized dominance function" refers to smoothing the state-action value deviation through a temporal difference estimation algorithm; "Pareto optimal" in this scheme refers to the optimal solution between the dual objectives of maximizing vanadium recovery rate and minimizing energy consumption, where one cannot be improved without harming the other.

[0094] In implementation, the algorithm execution flow is triggered by the central control unit in offline or servo online mode: when the output parameters of step 3 or step 4 change significantly (i.e., the change in the impurity thermal disturbance factor vector ΔG exceeds 10% or the neutralization molar ratio deviates from the optimal ±0.02), the system triggers the strategy update sub-flow. First, data from k control cycles within the window are collected. Cached, of which This represents the current state of the digital twin (including simulation outputs such as ΔG, gradient field, and flow rate). This is the action vector (U component) executed at that time. The reward function value is given immediately. The combination of reward functions is defined as follows:

[0095] in , λ1 and λ2 are weighting coefficients set by the process management system to balance the dual objectives, serving as reference recovery rate and reference energy consumption (the specific weighting coefficients can be determined according to actual process requirements and are not limited). During the strategy improvement phase, the system follows the current strategy network. Perform importance sampling on the cached data and calculate the probability ratio for each trajectory:

[0096] like If the value exceeds the confidence interval [1−ϵ,1+ϵ] (typically ϵ=0.2), it is mapped back to this interval through a pruning operation. This pruning mechanism limits the magnitude of a single update, enhancing update stability. The policy loss function is...

[0097] in The generalized advantage estimate (GAE) is calculated as follows: ,in As a discount factor, The GAE attenuation coefficient. Value function. The value network is trained in parallel to estimate state values ​​and reduce the sampling variance of the policy network. This loss, combined with the value loss and entropy regularization, constitutes the total loss. The parameters θ of the policy network and the value network are... The learning rate can be set between 10^-4 and 10^-3 through iterative updates using the Adam optimizer (the specific value can be determined based on the model's convergence, and there is no limit to this). The batch size can be between 32 and 128 trajectory samples.

[0098] During the backpropagation phase, the system computes gradients in parallel across the GPU cluster and updates the network weights. After each update, the new policy network... This is used to perform a new round of simulation in the digital twin and generate the control vector U. The simulation outputs Rv and Ec are fed back to the main controller. The optimization process terminates when the preset number of iterations is completed or when the policy convergence curve shows that the average reward change is less than a threshold (e.g., 0.01), and the final policy network parameters and the corresponding optimal control vector are output.

[0099] The optimized results for roasting oxygen partial pressure and vanadium precipitation pH setpoints are mapped to specific physical control quantities: the oxygen partial pressure command is sent to the blower and oxygen valve execution units via the control system to precisely adjust the composition of the oxygen-enriched gas blown into the blast furnace or leaching tank; the vanadium precipitation pH setpoint command is sent to the dosing system to drive the alkaline precipitant or acidic regulator pump station to add the agent according to the given ratio and flow rate to achieve the target pH. Command issuance follows the OPC UA communication specification, and execution logs and feedback deviations are recorded. When the actual system feedback of pH or oxygen partial pressure deviates by more than ±0.02 units, the internal controller automatically calls the fine-tuning PID submodule to quickly compensate the local execution units, ensuring the overall stable operation of the system.

[0100] The entire execution process of step 5 is closely linked to steps 3 and 4: the acid concentration gradient and precipitation strategy provided in step 3, and the spray volume and fan commands in step 4 together constitute the state observation vector, which drives the strategy network to generate a new U; the output of step 5 is then fed back to the execution layer, and a closed loop is formed through online monitoring. This strategy optimization process iterates repeatedly in the control center at a rate of minutes or less, and combined with digital twin simulation, ensures that the effectiveness of each control vector update can be verified in the digital environment. The verified commands are then applied to actual production, thereby achieving multi-parameter coupled intelligent control of the entire vanadium ore smelting process.

[0101] Step 6: Distribute the multi-dimensional control vector to the edge PLC based on the Storm distributed architecture to execute dynamic dosing of leachate and variable frequency speed control of the exhaust fan; the working principle of the Storm distributed architecture is as follows:

[0102] The central control node encapsulates the multidimensional control vector into a data stream through a message queue and pushes it to the input component. The input component receives the data and injects it into the processing topology in the form of micro-batches or cell tuples.

[0103] The routing component partitions and maps the data stream based on the PLC identifier field;

[0104] The distribution component converts the partitioned control vector into Modbus TCP packets via the OPC-UA protocol and sends them to the mapped address of the edge controller. If no feedback is received within the set confirmation time limit, a retry mechanism is triggered to resend the packets with an exponential backoff strategy.

[0105] The status monitoring component subscribes to the PLC status report stream and automatically adjusts the data injection rate of the input components based on back pressure feedback;

[0106] The metrics aggregation component collects execution feedback and latency information from each edge node and writes it to a real-time database for the scheduling unit to retrieve and analyze.

[0107] It should be noted that in this embodiment, "Storm distributed architecture" refers to a real-time control distribution platform based on a streaming computing paradigm. It utilizes micro-batch or tuple-level data injection and multi-level flow control mechanisms to achieve high throughput, low latency, and high reliability distribution of industrial control commands. "Micro-batch injection" refers to assembling accumulated data in the message queue according to time windows and injecting it into the topology all at once to reduce network overhead. "Tuple injection" sends data one message at a time to achieve the lowest latency. "Partition mapping" means routing control vectors to corresponding sub-topologies based on the PLC's unique identifier field. "Exponential backoff" is a retry mechanism that progressively increases the retry interval to prevent network congestion. "Backpressure feedback" is used to dynamically control the upstream transmission rate based on downstream processing capacity. "Real-time index library" is a database that centrally stores execution latency and feedback results for further analysis and optimization by the scheduling unit.

[0108] In practical implementation, the central control node first obtains the latest multi-dimensional control vector U from the strategy optimization module. This vector includes the roasting oxygen partial pressure setting, vanadium precipitation pH setting, and all sub-vectors for leaching, precipitation, tail gas purification, and wastewater reuse. This vector is encapsulated into a Storm message structure by the message packaging component, containing a message header (including a timestamp, message ID, and a list of destination PLCs), a message body (a hexadecimal serialized representation of U), and a checksum field. The message size is typically 512 bytes to 2KB, with the message header being approximately 64 bytes long. The specific size can be determined based on the distribution strategy and network MTU, and is not limited thereto.

[0109] Message queues (such as Kafka or RabbitMQ) serve as the source for Storm Spouts. Spouts pull messages in micro-batches through a consumer client interface. Batch size and pull intervals can be configured between 100ms and 1s, or a cell tuple mode can be selected to inject each U (U) individually. The pulled messages are then converted into Storm Tuples by the Spout, with metadata fields appended, and then launched to the topology's routing Bolt via internal RPC calls.

[0110] The routing component (Fields Grouping Bolt) distributes the Stream to the corresponding sub-streams based on the PLC identifier field in the Tuple, using either a hash partitioning algorithm or a range partitioning algorithm. The PLC identifier field is a string or integer ID (such as 0x01, 0x02), and its mapping relationship with the physical PLC's IP address and Modbus address is stored in a central configuration library. This mapping can be updated by field maintenance personnel through the HMI interface, and there are no restrictions on the mapping scheme. Partition mapping ensures that all control commands from the same PLC arrive at the same sub-topology instance in sequence, thereby avoiding out-of-order execution.

[0111] After reading the tuple from its partition stream, the Instruction Dispatcher Bolt converts the control vector U in the message body into a Modbus TCP message according to the OPC UA client protocol stack. This includes function code 0x10 (write multiple holding registers) or 0x06 (write a single holding register), with the register address specified by the mapping table. The specific conversion rule is as follows: each component of U is encoded as a floating-point number or integer and occupies one or more registers. The encoding method supports IEEE 754, 32-bit floating-point, or 16-bit integer. After the message is sent, the Instruction Dispatcher Bolt starts a timer to monitor for ACK confirmation. The timeout period can be configured between 100–500ms. If no Modbus response with the corresponding identifier is received, exponential backoff retries are initiated: an initial delay of 250ms, multiplied by 1.5 after each failure, with a maximum of 5 retries. If the retries still fail, the PLC is marked as "offline," and the event information is written to the operation log.

[0112] The Status Monitor Bolt subscribes to status update topics reported by the OPC UA server or PLC, including command execution results, I / O channel status, alarm information, and heartbeat signals. This component calculates downstream processing latency. It then compares the current latency with a configurable maximum latency threshold (e.g., 200ms) to generate a backpressure feedback signal. The backpressure feedback notifies the upstream Spout via a flow control API call, instructing it to adjust the send rate r_send.

[0113]

[0114] When Δt exceeds the threshold, r_send will be compressed to prevent PLC overload; when Δt falls back, the rate will recover proportionally.

[0115] The Metrics Aggregator Bolt collects execution feedback and latency statistics from each distributing component and status monitoring component in real time at the topology end. It maintains execution success rate, average latency, and maximum latency data for each PLC within a sliding window and writes them to a real-time database (such as InfluxDB or TimescaleDB). The database time-series key consists of the PLCID, control instruction type, and timestamp, while the values ​​are multi-dimensional execution metrics. The scheduling unit can retrieve these metrics via SQL or a RESTful API for fault diagnosis or rescheduling logic decisions.

[0116] The entire Storm topology is deployed in cluster mode, containing multiple Supervisor nodes and Nimbus nodes. Nimbus is responsible for topology submission and monitoring, while Supervisor is responsible for running the Bolt / Spout processes. Resource allocation is containerized and scheduled using Apache Mesos or Kubernetes to ensure that CPU / memory resources can be dynamically increased according to the load. The topology parallelism hint is determined by the production management system at the time of submission and can be adjusted between 1 and 10 based on the number of PLCs and the control frequency.

[0117] Step 7: Monitor the slag phase composition in real time using online XRD, compare it with the twin prediction value in Step 5, generate the model error, and feed it back to Step 3 to correct the impurity weight coefficient.

[0118] It should be noted that online XRD refers to "high-temperature online X-ray diffractometer," which differs from laboratory offline XRD. It must withstand the high temperatures of the production line (e.g., the temperature of calcined slag often reaches 600-800°C; the specific temperature threshold depends on the process and can be determined independently), dust, and other harsh environments. Non-contact real-time detection is achieved by opening a detection window (often Φ50-100 mm in diameter, which can be determined independently) in the conveying pipeline. The sampling frequency is generally set to 5-15 min / time (adjusted according to production line rhythm and model update requirements, which can be determined independently). The Rietveld full-spectrum fitting algorithm refers to a nonlinear least squares method based on the full powder diffraction spectrum (2θ range is usually 10∘-90∘, the specific range depends on the phase diffraction characteristics and can be determined independently). By fitting the residual between the observed diffraction spectrum and the theoretically calculated spectrum, it analyzes the phase type, content, and crystal structure parameters (such as lattice constant and microstrain). Compared to the traditional "single-peak integration method," it can eliminate overlapping peak interference and improve the quantitative accuracy of low-content phases (such as vanadium-iron spinel with a content of <5%) by over 30% (data is for illustrative purposes only; actual verification depends on equipment and process). The error vector ΔP refers to the element-wise difference between the actual slag phase composition vector Pactual and the digital twin prediction vector Ppred (i.e., ΔP = Pactual − Ppred). Its dimension is consistent with the number of phases (e.g., 3-dimensional, corresponding to the volume fraction deviation of φFeV2O4, φCaSiO3, and φCr2O3), and it is used to quantify the direction and magnitude of the model prediction error.

[0119] The impurity weight coefficient matrix is ​​the set of contribution weights of each impurity (silicon, aluminum, chromium) to thermodynamic disturbance factors (entropy, enthalpy, free energy perturbation) in step 3, "Dynamic Coupled Response Model". For example, the weight wSi of silicon to silica gel formation and the weight wCr of chromium to redox potential are determined according to the type of impurity and the disturbance dimension (commonly 33 (impurities) × 3 × 3 (disturbance dimension), which can be determined by the user). It needs to be dynamically updated with slag phase feedback to match the actual process.

[0120] In implementation, this step consists of five major modules: diffraction acquisition subsystem, full-spectrum quantitative analysis subsystem, error calculation subsystem, weight allocation subsystem, and feedback update subsystem. Its application scenario is to compare the actual crystal phase composition with the digital twin predicted composition by performing real-time diffraction measurements on the slag flow or tray surface during the blast furnace roasting to slag tapping stage, and to dynamically correct the impurity weight parameters of the dynamic coupling response model based on this comparison.

[0121] In the diffraction acquisition subsystem, a high-temperature window-type X-ray diffraction probe is installed on the slag discharge pipeline or pallet conveyor belt. The probe window temperature is maintained between 350 and 600°C via a water-cooling or air-cooling system. The detector array covers an angle range of 10 to 90 degrees (2θ), continuously acquiring powder diffraction patterns in 0.02-degree increments. The geometric parameters between the probe and detector, the incident light intensity, and the exposure time can be adaptively adjusted in real time according to the slag emission intensity, ensuring that the signal-to-noise ratio always meets the requirements for quantitative analysis. After detection, the raw angle intensity data stream is packaged and transmitted to the analysis server via real-time Ethernet.

[0122] After receiving the diffraction data, the full-spectrum quantitative analysis subsystem first performs background subtraction, using polynomial curve fitting to remove the broad substrate and scattering background. Then, the Rietveld full-spectrum fitting algorithm is applied to co-fit the remaining diffraction peaks. A hybrid pseudo-Voigt function is selected for the peak shape to accommodate peak broadening and shifting effects caused by high temperatures. The fitting process uses nonlinear least squares iteration until the residuals converge. The software loads crystal structure models and relative atomic coordinates for three phases: spinel (FeV₂O₄), wollastonite (CaSiO₃), and chromite (Cr₂O₃). After fitting, the volume fraction of each phase is directly output. Each analysis includes a fitting residual index and R-factor to evaluate quantitative accuracy; the threshold can be set by the system administrator, typically not exceeding 5%.

[0123] The error calculation subsystem synchronizes and differentially calculates the actual crystal phase composition vector obtained from the analysis with the phase composition vector predicted by the digital twin in step 5 to obtain the phase composition deviation vector. The alignment is based on the shared timestamp and sampling sequence number of the two. The system ensures that the actual and predicted data at the same event point are synchronized. If the time difference between the two exceeds the preset tolerance (e.g., 200ms), the alignment pair is marked as abnormal and skipped in this feedback.

[0124] The weighting subsystem assigns weights to the deviation vector based on a three-phase covariance matrix pre-calculated from historical samples. The covariance matrix reflects the statistical correlation between each crystalline phase under normal operating conditions; for example, an increase in FeV₂O₄ is often accompanied by a decrease in CaSiO₃. This subsystem generates an error weight vector by performing matrix operations on the deviation vector and the covariance matrix. Each component represents the adjustment value for the impurity weights corresponding to the three phases, and after standardization, it is passed to the next module. The covariance matrix can be updated by re-statistically analyzing monthly or batch-by-batch production samples to adapt to changes in raw ore batches. The relevant cycle is determined by the process engineer based on the stability of the production line.

[0125] Upon receiving the error weight vector, the feedback update subsystem transforms it into incremental updates of the impurity weight parameters in the dynamically coupled response model based on a predefined mapping relationship. This mapping relationship is defined by a set of weight mapping coefficient matrices, typically obtained during model development through sensitivity analysis, describing the influence of each crystal phase deviation on different impurity weights. The update process employs a gradient descent strategy, controlling the update magnitude with a learning rate parameter, and records the weight change trajectory and convergence status after each iteration. After each weight update, the system writes the updated weights into the response model and triggers model reconstruction, thereby influencing the calculation of the thermodynamic factor in step 3 and the generation of subsequent control commands.

[0126] The entire online XRD feedback process is automatically driven by the event bus of the central control system. When any link times out or fails to resolve (such as excessive fitting residuals, abnormal data alignment, or covariance calculation errors), the system will mark the current feedback cycle as abnormal. Depending on the redundancy configuration, it can choose to use the weight of the previous cycle or enter the manual intervention mode. At the same time, it will record the abnormal details in the log for subsequent analysis.

[0127] Please see Figure 5 A multi-parameter coupled intelligent control system for vanadium ore smelting process includes: a composition acquisition module, used to receive LIBS / XRF probe data from a host computer, extract impurity concentration matrix and valence distribution through wavelet packet decomposition and principal component regression, and output the original spectrum data;

[0128] The Kalman filter module is used to fuse and compensate for slurry rheological disturbances based on the original spectrum data and the dielectric constant measured by microwave dielectric spectrum, and output a normalized impurity spectrum vector by means of extended Kalman filter algorithm;

[0129] The coupled response module is used to calculate the thermal disturbance factor online by using the improved Maxwell–Garnett and unsteady-state energy conservation equations, and to solve for the acid concentration gradient, reaction time and dosing strategy.

[0130] The material coordination module is used for parallel collection of mother liquor. A linear programming model with mass conservation constraints was constructed based on the concentration and flow rate of HCl in the exhaust gas, and solved using Lagrange duality and the interior-point method. Molar ratio, generating spray volume and fan adjustment commands;

[0131] The strategy optimization module is used to map the coupling response and the output parameters of the material coordination module into a multidimensional control vector U. The PPO proximal strategy optimization algorithm is applied in the digital twin for iterative updates, and the output oxygen partial pressure and vanadium precipitation pH settings are output.

[0132] The distributed delivery module is used to receive control vectors based on the Storm distributed architecture and deliver them to the edge PLCs through routing, OPC-UA to Modbus TCP conversion and backpressure mechanism.

[0133] The XRD feedback module is used to acquire XRD diffraction data of slag online, perform quantitative analysis using the Rietveld algorithm, compare the data with twin predictions, generate residuals, and feed back impurity weight update information to the coupled response module through a gradient correction algorithm.

[0134] In one implementation, the system automatically linked the batch information of vanadium-titanium magnetite ore delivered to the plant that day, loading the corresponding ore composition calibration parameters and historical operating condition models. After the ore was fed into the drying and preheating section by a bucket elevator, it flowed into the LIBS / XRF coupled probe scanning area via a conveyor belt. The composition acquisition module first triggered parallel excitation of laser and X-ray at a frequency of 20Hz. The probe acquired the spectrum in real time and mapped the signal into a matrix of five impurity concentrations (SiO2, Al2O3, Cr2O3, Fe2O3, and V2O5) and vanadium valence state ratios through wavelet packet decomposition and principal component regression model. The results were pushed to the Kalman filter module within 500ms.

[0135] The Kalman filter module fuses the raw spectrum data with the slurry dielectric constant obtained from an online microwave dielectric spectrometer. In the first 1-second control cycle, the EKF predicts the prior state based on the control disturbance input from the previous cycle, then uses the online dielectric constant to correct the drift term, generates a normalized impurity spectrum vector, and writes it into the coupled response module.

[0136] The coupled response module receives the temperature field, air volume, and the normalized vector. Based on the improved Maxwell–Garnett equation and real-time thermal measurement data of the furnace, it calculates the thermal interference factor online. Then, in the partitioned finite element model, it calculates the local acid concentration gradient and mass transfer resistance of each tank area in the leaching section, and generates preliminary acid concentration distribution and reaction time suggestions.

[0137] The material synergy module starts in parallel within the same control cycle: the ion chromatograph extracts a sample of the vanadium precipitation mother liquor every minute and outputs the NH4⁺ concentration, while the TDLAS instrument measures the HCl concentration in the exhaust gas at a rate of 10 Hz, and the Nm³ is provided by the turbine flow meter. 3 The system constructs a mass-conservation-constrained linear programming model for these three factors, and solves the optimal molar ratio of NH4⁺–HCl and the corresponding spray rate and fan speed commands within 2 seconds using Lagrange duality and interior point method.

[0138] The strategy optimization module then integrates the acid concentration gradient mapping given by the coupled response with parameters such as spray volume and wind speed generated by material co-generation into an 18-dimensional control vector U, and performs three generations of iterative optimization using the PPO algorithm in the digital twin. The simulation environment fully reproduces the dynamics of blast furnace top temperature, leaching tank stirring rate, and settling tank pH. The vanadium recovery rate and energy consumption ratio are evaluated for each control strategy, and the final oxygen partial pressure setpoint in the roasting zone and the pH target in the settling section are output after the tenth generation of optimization.

[0139] The distributed delivery module utilizes the Storm framework to stream this multi-dimensional vector data. It injects the routing topology in 200ms micro-batch injections via Kafka Spout, and uses Fields Grouping to partition data based on PLC IDs. The delivery component then encapsulates the control data into OPC UA packets and pushes them to the edge PLCs. Within 300ms, it loads the parameters for the diaphragm pump frequency and the fan inverter. The status monitoring component continuously subscribes to PLC execution feedback and adjusts the transmission rate based on latency to ensure both delivery pressure and execution stability.

[0140] After the high-temperature XRD probe installed in the slag discharge section acquires the first batch of slag diffraction patterns in real time, the XRD feedback module immediately performs Rietveld full-spectrum fitting to obtain the difference vector between the actual crystal phase composition and the digital twin prediction value. The error calculation subsystem uses a preset phase composition covariance matrix to weight the difference, and the generated weight increment is then applied to the impurity weight coefficient in the coupled response module by the gradient update algorithm to complete the first online correction.

[0141] The entire process operates at a high frequency with a 10-second cycle, and the system completes its first closed loop within 100 seconds. Subsequently, all modules iterate continuously at this cycle, allowing the control strategy and model parameters to constantly adapt to the actual operating conditions. Throughout production, the system successfully increased vanadium recovery by approximately 3%, reduced energy consumption by approximately 2%, stabilized the HCl emission concentration in the tail gas below 50 ppm, and increased the NH4⁺ reuse rate of the mother liquor by 15%. The above data are for illustrative purposes only; specific results should be determined based on the actual ore batch and equipment scale.

[0142] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A multi-parameter coupled intelligent control method for vanadium ore smelting process; characterized in that: include: Step 1: In real time, the ore flow is scanned using a probe coupled with laser-induced breakdown spectroscopy and X-ray fluorescence to obtain the concentration matrix and valence distribution data of impurity elements; Step 2: Based on the impurity element concentration matrix and the real-time measured dielectric constant of the slurry using microwave dielectric spectrum, the signal drift caused by the perturbation of the slurry rheological properties is dynamically corrected by the extended Kalman filter fusion algorithm, and a standardized impurity spectrum vector is output. Step 3: Input the standardized impurity spectrum vector into the dynamic coupling response model, calculate the impurity thermodynamic interference factor, and generate acid concentration gradient, reaction time optimization command and precipitant addition strategy; Step 4: Simultaneously collect the ammonia ion concentration of the vanadium precipitation mother liquor and the hydrogen chloride flow rate of the roasting tail gas. Solve the optimal solution of the ammonia-chlorine neutralization molar ratio through the material flow coordination method to obtain the coordinated adjustment command of the mother liquor spraying rate and the tail gas fan. Step 5: Couple the output parameters of Step 3 and Step 4 to construct a multidimensional control vector U. Iteratively update U in the digital twin through a near-end strategy optimization algorithm to output the calcination oxygen partial pressure and vanadium precipitation pH setpoints. Step 6: Distribute the multi-dimensional control vector to the edge PLC based on the Storm distributed architecture to execute dynamic dosing of leachate and variable frequency speed regulation of exhaust gas fan; Step 7: Monitor the slag phase composition in real time using online XRD, compare it with the twin prediction value in Step 5, generate the model error, and feed it back to Step 3 to correct the impurity weight coefficient.

2. The multi-parameter coupled intelligent control method for vanadium ore smelting process according to claim 1, characterized in that: The working method of step 1 is as follows: First, an optical beam splitter is used to coaxially align the output beam of the high-energy pulsed laser and the multi-target X-ray tube, and the signal acquisition time window of the high-energy pulsed laser and the multi-target X-ray tube is aligned through a timing control circuit to obtain the original pulse waveform; then, based on the ore flow velocity feedback signal obtained by the photoelectric encoder, the probe scanning step size and dwell time are adjusted in real time, and the X-ray tube voltage is automatically switched to adapt to the excitation depth requirements of ores with different particle sizes; finally, the characteristic X-ray energy spectrum of the atomic ion emission spectrum generated by the laser-induced breakdown spectrum and the micro-focal spot X-ray fluorescence is input into the multivariable spectral deconvolution module based on non-negative matrix factorization. By decoupling the overlapping peaks and correcting the matrix effect, the impurity concentration matrix and vanadium valence state distribution data containing spatial coordinates are obtained.

3. The multi-parameter coupled intelligent control method for vanadium ore smelting process according to claim 2, characterized in that: The multivariate spectral deconvolution module based on nonnegative matrix factorization decomposes the original spectral matrix into a basis vector matrix and an abundance coefficient matrix using an alternating least squares iterative optimization method. The basis vector matrix contains a standard spectral line library of known elements. In each iteration of the alternating least squares iterative optimization method, a matrix effect correction mechanism is adopted to dynamically adjust the weight coefficient of the V spectral line according to the quenching factor of the Fe element-dominant spectral region. If Cr-V spectral peak overlap is detected, the overlapping peak decoupling protocol is triggered, and the overlapping peak area is separated using a Gaussian mixture model to obtain the pure elemental concentration matrix and vanadium valence state distribution data after matrix correction and peak decoupling.

4. The intelligent control method for multi-parameter coupling in a vanadium ore smelting process according to claim 1, characterized in that: The working principle of the extended Kalman filter fusion algorithm is as follows: s100, the first Impurity concentration matrix derived from laser-induced breakdown spectrum at any time Dielectric constant measured by microwave dielectric spectroscopy Organization as a measurement vector Online fusion is achieved using an EKF iterative structure; s200, In the prediction phase, the previous impurity concentration is mapped using a mapping function. External disturbance input Combined to generate the current concentration and drift state The predicted covariance is updated using the Jacobian matrix and the process noise covariance. s300. Based on the measurement mapping function, the MG dielectric mixing equation maps the predicted concentration to the predicted dielectric constant value, as expressed by: （1） In formula (1), The dielectric constant of the slag matrix is... The dielectric constant of the impurity particles. This refers to the volume fraction of impurities, used to predict concentrations. The values ​​are mapped to predicted dielectric constants, thus constructing the Jacobian matrix; s400. The process noise covariance is obtained based on the energy component covariance after wavelet packet decomposition of the LIBS signal, and the measurement noise covariance is adjusted based on the dielectric constant sequence variance. S500 introduces a forgetting factor into the Kalman gain to enhance the response to new data. When the measurement residual exceeds a preset threshold, the covariance matrix is ​​dynamically recalibrated through a recalibration mechanism. s600, obtain the normalized impurity spectrum vector through the Kalman update equation.

5. The intelligent control method for multi-parameter coupling in a vanadium ore smelting process according to claim 1, characterized in that: The dynamic coupling response model includes an input preprocessing layer, a feature fusion layer, a thermodynamic factor calculation layer, a dynamic coupling analysis layer, a multi-objective optimization layer, and an instruction synthesis layer. The input preprocessing layer preprocesses the standardized impurity spectrum vector through normalization transformation and wavelet packet denoising. The feature fusion layer integrates the time series sequences of impurity concentration and dielectric constant in parallel through a multi-head attention mechanism, and triggers an adaptive recalibration mechanism to compensate for time-domain coupling information when the attention weight is lower than a preset threshold. The thermodynamic factor calculation layer generates a thermodynamic disturbance factor matrix for each impurity based on a nonlinear mapping function calculated using Gibbs free energy, which reflects the contribution of different species to the entropy and enthalpy disturbances of the reaction system. The dynamic coupling analysis layer uses the finite element method to couple the thermodynamic disturbance factor matrix with the reactor geometric model to solve the local acid concentration gradient and mass transfer resistance. The multi-objective optimization layer uses the NSGA-II algorithm to perform Pareto optimality search among the three objectives of acid concentration gradient, reaction time, and precipitant dosage, and updates the crossover and mutation operator when the population convergence is lower than a preset threshold. The instruction synthesis layer uses a weighted linear combination to translate the optimization results into a control sequence of acid concentration adjustment rate, reaction residence time, and precipitant dosage strategy.

6. The multi-parameter coupled intelligent control method for vanadium ore smelting process according to claim 1, characterized in that: The material flow coordinated method monitors the ammonia ion concentration in the vanadium precipitation mother liquor online using an ion chromatograph, simultaneously measures the hydrogen chloride concentration in the roasting tail gas in real time using a tunable diode laser absorption spectroscopy, and collects the volumetric flow velocity of the exhaust gas using a turbine flow meter to construct a gas-liquid two-phase flow field model. Based on the Reynolds-Neuschmann equation, the trajectory of the spray droplets and the diffusion path of the exhaust gas are calculated. Then, the optimal solution for the ammonia-chlorine molar ratio is solved using a sequential quadratic programming algorithm, with the objective function set at a molar ratio of 1.05 ± 0.

02. If the ammonia ion concentration fluctuates beyond ± 5%, the frequency of the diaphragm pump is adjusted in real time using an incremental PID control protocol. When the rate of change of the exhaust gas flow exceeds a preset threshold, the material flow coordinated method pre-adjusts the induced draft fan speed through a feedforward compensation mechanism.

7. The intelligent control method for multi-parameter coupling in a vanadium ore smelting process according to claim 1, characterized in that: Step 5 constructs a multidimensional control vector state space based on the output parameters of steps 3 and 4, and iteratively updates it using a near-end strategy optimization algorithm. The multidimensional control vector state space includes furnace condition control vector, leaching parameter vector, precipitation adjustment vector, tail gas purification vector, wastewater reuse vector, and model calibration vector. The near-end strategy optimization algorithm is based on digital twin simulation of the roasting-vanadium precipitation coupled process, and sets a dual objective function of maximizing vanadium recovery rate and minimizing energy consumption. An importance sampling mechanism is used to collect policy gradient data, calculate the probability ratio between the old and new policies, and constrain the policy update step size through a gradient pruning mechanism when the probability ratio exceeds a preset confidence interval threshold. Then, the near-end policy optimization algorithm estimates the time-series difference error through a generalized dominance function and backpropagates to update the policy network parameters. The Adam optimizer is used to iterate the policy network weights and output Pareto optimal roasting oxygen partial pressure setpoint and vanadium precipitation pH setpoint.

8. The multi-parameter coupled intelligent control method for vanadium ore smelting process according to claim 1, characterized in that: The working principle of the Storm distributed architecture is as follows: The central control node encapsulates the multidimensional control vector into a data stream through a message queue and pushes it to the input component. The input component receives the data and injects it into the processing topology in the form of micro-batches or cell tuples. The routing component partitions and maps the data stream based on the PLC identifier field; The distribution component converts the partitioned control vector into Modbus TCP packets via the OPC-UA protocol and sends them to the mapped address of the edge controller. If no feedback is received within the set confirmation time limit, a retry mechanism is triggered to resend the packets with an exponential backoff strategy. The status monitoring component subscribes to the PLC status report stream and automatically adjusts the data injection rate of the input components based on back pressure feedback; The metrics aggregation component collects execution feedback and latency information from each edge node and writes it to a real-time database for the scheduling unit to retrieve and analyze.

9. The multi-parameter coupled intelligent control method for vanadium ore smelting process according to claim 1, characterized in that: The working principle of step 7 is as follows: Slag phase diffraction patterns were acquired in real time using a high-temperature online X-ray diffractometer, and the crystal phase composition vector was obtained by analysis using the Rietveld full-spectrum fitting algorithm. ,in This represents the volume fraction of the vanadium-iron spinel phase. This represents the volume fraction of the wollastonite phase. This represents the volume fraction of the chromite phase. Synchronously call the digital twin prediction phase composition vector generated in step 5 ; Performed through the error vector calculation module ;like If the L2 norm exceeds the preset tolerance threshold, an error weighting mechanism based on Mahalanobis distance is triggered, as shown in the formula: （2） In formula (2), Σ is the covariance matrix of the phases. This is the inverse covariance matrix, used to eliminate inter-phase correlation interference; then, The input feedback correction channel is used to update the impurity weight coefficient matrix in the dynamic coupling response model of step 3 using the gradient descent method, and finally the corrected impurity thermodynamic interference factor calculation parameters are generated.

10. A multi-parameter coupled intelligent control system for vanadium ore smelting process, characterized in that: The multi-parameter coupled intelligent control method for vanadium ore smelting process, applied to any one of claims 1-9, comprises: The component acquisition module is used to receive LIBS / XRF probe data from the host computer, extract the impurity concentration matrix and valence distribution through wavelet packet decomposition and principal component regression, and output the original spectral data. The Kalman filter module is used to fuse and compensate for slurry rheological disturbances based on the original spectrum data and the dielectric constant measured by microwave dielectric spectrum, and output a normalized impurity spectrum vector by means of extended Kalman filter algorithm; The coupled response module is used to calculate the thermal disturbance factor online by using the improved Maxwell–Garnett and unsteady-state energy conservation equations, and to solve for the acid concentration gradient, reaction time and dosing strategy. The material coordination module is used for parallel collection of mother liquor. A linear programming model with mass conservation constraints was constructed based on the concentration and flow rate of HCl in the exhaust gas, and solved using Lagrange duality and the interior-point method. Molar ratio, generating spray volume and fan adjustment commands; The strategy optimization module is used to map the coupling response and the output parameters of the material coordination module into a multidimensional control vector U. The PPO proximal strategy optimization algorithm is applied in the digital twin for iterative updates, and the output oxygen partial pressure and vanadium precipitation pH settings are output. The distributed delivery module is used to receive control vectors based on the Storm distributed architecture and deliver them to the edge PLCs through routing, OPC-UA to Modbus TCP conversion and backpressure mechanism. The XRD feedback module is used to acquire XRD diffraction data of slag online, perform quantitative analysis using the Rietveld algorithm, compare the data with twin predictions, generate residuals, and feed back impurity weight update information to the coupled response module through a gradient correction algorithm.

Images