Automatic control method and system for secondary granulation of high-voltage zinc oxide resistor disc

Abstract

The invention discloses an automatic control method and system for secondary granulation of a high-voltage zinc oxide resistor disc, relates to the technical field of intelligent manufacturing of power equipment, and solves the problems of out-of-control particle morphology caused by dynamic coupling parameter identification lag and control instability caused by multi-physical field parameter coupling in an existing method. According to the invention, a dynamic physical property parameter matrix is generated in real time based on multi-band dielectric relaxation spectrum analysis and terahertz wave tomography; predicting a fluidized phase change threshold value and an energy gathering area through multi-physics field coupling modeling; a time sequence attention deep reinforcement learning algorithm is adopted to generate a multi-field cooperative adjustment instruction; positioning a parameter conflict source and triggering decoupling compensation by combining a high-frequency vibration and acoustic emission combined monitoring module; performing closed-loop correction on the control network weight based on the laser spectrum data and a partial least squares regression model; the real-time performance of fluidization parameter identification, the stability of multi-field coupling control and the recovery efficiency of abnormal working conditions are remarkably improved, and meanwhile the batch consistency of the electrical performance of the resistor discs is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent manufacturing of electric power equipment, and more specifically to an automated control method and system for secondary granulation of high-voltage zinc oxide resistor sheets. Background Art

[0002] The secondary granulation process for high-voltage zinc oxide resistors produces a uniform granulated powder through wet mixing and spray drying. The core of its automated control lies in the coordinated regulation of powder dispersibility, atomization parameters, and drying dynamics. Existing technology relies on PLC / DCS systems to preset parameters for mixing and granulation equipment, and combines sensor feedback to achieve local process optimization.

[0003] Existing automated control technology for secondary granulation of high-voltage zinc oxide resistors focuses on mixing uniformity, granulation efficiency, and sintering process optimization, and improves the stability of a single process node through multi-stage equipment integration and sensor feedback. For example, patent CN202010491588A proposes a multi-stage mixed grinding process using a colloid mill and a high-speed stirred mill, combined with a DCS system to monitor the slurry state in real time, and uses continuous granulation equipment to reduce powder agglomeration and significantly improve batch consistency. Patent CN113149634B prepares composite powders through a chemical co-precipitation method, combining a PLC-controlled continuous ball milling system with a temperature and humidity sensor to close the drying loop to achieve nanoscale grain size control, while using staged cold sintering and high-temperature sintering processes to suppress abnormal grain growth. Regarding the granulation process, patent CN101880157B introduces ultrasonic dispersion and spray granulation technology, using a laser particle size analyzer to provide real-time feedback to adjust spray parameters and optimize particle distribution uniformity. Furthermore, patent CN101714439B utilizes automated coating equipment and a visual inspection system to achieve uniform coverage and thickness monitoring of the high-resistance layer slurry, enhancing the resistance of the resistor to contamination. While these technologies improve local process performance through equipment upgrades and single-parameter feedback mechanisms, their control logic is still limited to discrete parameter adjustment, resulting in significant deficiencies in dynamic parameter identification and multi-field coordinated control.

[0004] First, the problem of lag in dynamic coupling parameter identification is prominent: for example, patent CN202010491588A relies on a static model calibrated offline, but the natural fluctuations in raw material properties in actual production cause the fluidization parameters to deviate from the preset range, and the sensor system of patent CN101714439B has limited sampling capacity and cannot track the dynamic changes of the fluidization process in real time. The lag in control instructions leads to uncontrollable fluctuations in the surface morphology of the particles.

[0005] Secondly, parameter coupling causes control failure: for example, the fixed control algorithm of patent CN113149634B and the optimization strategy of patent CN202010491588A are difficult to cope with the interaction of multiple physical fields. For example, when the raw material feeding rate suddenly changes, the conflict between fluidization air pressure and temperature regulation causes the system response to become unstable; when the binder injection in patent CN101880157B is abnormal, the reverse coupling of hot air temperature and bed density is not effectively modeled, the recovery time is significantly prolonged, and particle structure defects are caused. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the present invention discloses an automated control method and system for secondary granulation of high-voltage zinc oxide resistors, aiming to solve the problems in the background technology.

[0007] In order to achieve the above technical effects, the present invention adopts the following technical solutions:

[0008] A method for automatically controlling secondary granulation of high-voltage zinc oxide resistor sheets, comprising:

[0009] Step 1: Based on multi-band dielectric relaxation spectrum analysis and terahertz wave tomography, the dielectric loss factor, surface charge density gradient and porosity are analyzed in real time to generate a dynamic physical property parameter matrix;

[0010] Step 2: Based on the multi-physics field coupling modeling method, the dynamic physical property parameter matrix is ​​coupled with the real-time process parameters to establish a fluidization process prediction model including thermal field, force field and electric field, and output fluidization state characteristic parameters, including the phase change critical threshold and energy accumulation area of ​​the fluidization process;

[0011] Step 3: Generate coordinated adjustment instructions for the airflow field, temperature field, and spray field according to the fluidization state characteristic parameters through a deep reinforcement learning control algorithm based on temporal attention;

[0012] Step 4: Using a high-frequency vibration and acoustic emission joint monitoring module to identify abnormal disturbance characteristics of the fluidized bed in real time and trigger a parameter decoupling and replanning mechanism; the high-frequency vibration and acoustic emission joint monitoring module uses wavelet envelope analysis to identify the characteristic frequency of binder blockage and locates the source of parameter conflict through a disturbance propagation chain reverse tracing model;

[0013] Step 5: Based on the porosity and laser-induced breakdown spectroscopy data of the granulated particles, a quantitative mapping between the control parameters and the residual pressure ratio and leakage current is established through a partial least squares regression model, and the reinforcement learning network weights of step 3 are corrected in real time.

[0014] As a further technical solution of the present invention, the fluidization process prediction model includes a physical property field mapping layer, a dynamic grid generation layer, a hybrid solution calculation layer, a phase change critical judgment layer, an energy aggregation analysis layer and a grid optimization feedback layer; the physical property field mapping layer is used to map the porosity and surface charge density gradient data in the dynamic physical property parameter matrix to the fluid domain grid nodes through the Radon back projection algorithm to establish an initial field distribution; the dynamic grid generation layer is used to construct an initial calculation grid based on real-time process parameters through the Delaunay triangulation algorithm, and trigger a local encryption mechanism when a porosity gradient mutation is detected to generate a field grid. Topological structure; the hybrid solution calculation layer is used to calculate the particle collision energy through the explicit discrete element solver using the DEM-CFD coupling method, and solve the flow field and heat conduction equations through the implicit finite volume method, and output the transient flow field, temperature field and particle motion trajectory data; the phase change critical judgment layer is used to identify the local gradient mutation area based on the void ratio gradient field data through the Canny edge detection algorithm, and mark the phase change critical point coordinates when the void ratio gradient value exceeds 0.25mm; the energy aggregation analysis layer is used to scan the particle collision energy field and locate the energy density higher than 1m / Jmm through the DBSCAN density clustering algorithm. 3 the grid optimization feedback layer is used to dynamically adjust the grid density through the unstructured grid dynamic encryption algorithm according to the coupling deviation of heat flow and electric potential gradient, and trigger an iterative optimization cycle when the deviation exceeds a preset threshold.

[0015] As a further technical solution of the present invention, an automated control system for secondary granulation of high-voltage zinc oxide resistors includes: a real-time physical property sensing module for analyzing the hydroxyl activity and humidity gradient on the surface of the material based on a dielectric relaxation spectrum decomposition algorithm, reconstructing the three-dimensional pore distribution in combination with terahertz wave inverse scattering tomography, and outputting a dynamic physical property parameter matrix to a dynamic multi-field modeling module, wherein the dynamic physical property parameter matrix includes dielectric loss factor, porosity, and surface charge density gradient;

[0016] The dynamic multi-field modeling module is used to construct a discrete element fluid-mechanical coupling model. It uses an explicit / implicit hybrid iterative algorithm to simultaneously solve the heat conduction equation, the particle collision energy equation, and the electric field potential energy equation. It predicts the bubble phase / emulsion phase transition threshold and the microscopic defect risk coordinates, generates a fluidization state map, and outputs it to the intelligent collaborative control module.

[0017] An intelligent collaborative control module uses a multi-head temporal attention mechanism to extract fluidization uniformity features, combines a rolling time domain optimization framework to generate collaborative adjustment instructions for airflow valve opening, infrared power, and atomization frequency, and outputs control instructions to the actuator and abnormal reverse suppression module;

[0018] The abnormal reverse suppression module is used to identify the characteristic frequency of adhesive blockage based on the wavelet envelope analysis algorithm, locate the source of parameter conflict through the disturbance propagation chain reverse tracking algorithm, dynamically decouple the pressure-temperature reverse coupling effect, output compensation instructions to the actuator, and synchronously feedback abnormal characteristics to the closed-loop feedback optimization module;

[0019] The closed-loop feedback optimization module is used to associate control parameters with electrical performance indicators through a partial least squares regression model. When grain boundary segregation is detected, the graph convolutional network is triggered to mine implicit control rules, update the reinforcement learning network weights and reward function constraints, and output the optimized parameters to the intelligent collaborative control module.

[0020] Based on the above technical solutions, the positive and beneficial effects of the present invention are:

[0021] 1. This solution uses multi-band dielectric relaxation spectroscopy analysis and terahertz tomography in step 1 to capture dynamic changes in the raw material's dielectric loss factor, surface charge density gradient, and porosity in real time. This overcomes the limitations of traditional static models that rely on offline calibration and enables dynamic tracking and adaptive modeling of fluidization parameters. Combined with the high-frequency vibration and acoustic emission monitoring module in step 4, this improves the sensor system's ability to capture transient disturbance characteristics, significantly shortens control command response time, and effectively suppresses uncontrollable fluctuations in particle surface morphology.

[0022] 2. Based on the multi-physics coupling modeling approach in step 2, a global prediction model for the interactions of heat, force, and electric fields is established. The sequential attention deep reinforcement learning control algorithm in step 3 dynamically analyzes the coupling relationships between parameters and generates coordinated adjustment instructions for the airflow, temperature, and spray fields, eliminating the response conflicts of traditional fixed control algorithms under multi-physics field interactions. Combined with the disturbance propagation chain reverse tracking model in step 4, the source of parameter conflicts is precisely located, triggering a decoupling replanning mechanism, significantly shortening the recovery time from abnormal operating conditions and preventing the occurrence of particle structure defects.

[0023] 3. A quantitative mapping between control parameters and resistor performance is established through the partial least squares regression model in step 5. The reinforcement learning network weights are modified in real time to enhance the robustness of the control strategy to raw material fluctuations and process disturbances. The multi-module collaborative "perception-modeling-control-verification" closed-loop architecture systematically optimizes the dynamic coupling characteristics of the granulation process, ensuring improved consistency in the residual voltage ratio and leakage current of resistors between batches, meeting the stringent reliability requirements of high-voltage power grids for zinc oxide resistors. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which:

[0025] Figure 1 This is a structural diagram of an automated control method for secondary granulation of high-voltage zinc oxide resistors according to the present invention;

[0026] Figure 2 This is a diagram showing the working principle and steps of the adaptive adjustment mechanism of the terahertz wave pulse repetition frequency of the present invention;

[0027] Figure 3 This is a working principle diagram of step 1 of the present invention;

[0028] Figure 4 This is a diagram showing the working principle of the fluidization process prediction model of the present invention;

[0029] Figure 5 A diagram showing the working method of the temporal attention-based deep reinforcement learning control algorithm of the present invention;

[0030] Figure 6 This is a schematic diagram of the automatic control system for secondary granulation of high-voltage zinc oxide resistors according to the present invention. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] In an embodiment, a method for automatically controlling secondary granulation of zinc oxide resistor sheets for high voltage is provided. Figure 1 As shown, it includes: step 1, based on multi-band dielectric relaxation spectrum analysis and terahertz wave tomography method, real-time analysis of dielectric loss factor, surface charge density gradient and porosity is performed to generate a dynamic physical property parameter matrix; Figure 3As shown in the figure, the specific working principle is as follows: the dielectric response signal of the material at the fluidized bed inlet is collected through a broadband dielectric sensor array, the frequency domain curves of the real and imaginary parts of the complex dielectric constant are decomposed based on the Debye multi-relaxation time model, and the surface hydroxyl activity and humidity gradient parameters are extracted; at the same time, a terahertz wave transmitting device is used to project a pulse beam in the frequency band of 0.1THz to 1THz to the material layer, and the scattered signal is tomographically imaged by Radon transform, and the three-dimensional pore distribution topology is reconstructed based on the inverse projection iterative algorithm; then, the surface charge density gradient tensor obtained by dielectric spectrum decomposition is tensor-fused with the pore connectivity matrix of terahertz tomography, and the eigenvalue weighting algorithm is used to eliminate the dimensional difference of cross-modal data to generate a dynamic physical property parameter matrix; during the detection process, if the pore connectivity rate suddenly changes beyond the preset threshold, the tomographic resolution is adjusted to dynamically match the material state change through the adaptive adjustment mechanism of the terahertz wave pulse repetition frequency;

[0033] Step 2: Based on the multi-physics field coupling modeling method, the dynamic physical property parameter matrix is ​​coupled with the real-time process parameters to establish a fluidization process prediction model including thermal field, force field and electric field, and output fluidization state characteristic parameters, including the phase change critical threshold and energy accumulation area of ​​the fluidization process;

[0034] Step 3: Generate coordinated adjustment instructions for the airflow field, temperature field, and spray field according to the fluidization state characteristic parameters through a deep reinforcement learning control algorithm based on temporal attention;

[0035] Step 4: Use the high-frequency vibration and acoustic emission joint monitoring module to identify abnormal disturbance characteristics of the fluidized bed in real time and trigger the parameter decoupling and replanning mechanism; the high-frequency vibration and acoustic emission joint monitoring module uses wavelet envelope analysis to identify the characteristic frequency of binder blockage and locates the source of parameter conflict through the disturbance propagation chain reverse tracking model; the working principle of the high-frequency vibration and acoustic emission joint monitoring module is as follows:

[0036] The vibration signal was decomposed into 32 sub-bands using wavelet packet transform, and the energy entropy of each sub-band was calculated to quantify the unevenness of the energy distribution in the frequency domain. Hilbert envelope detection was used to extract the time domain envelope signal of the high-frequency modulation component. Fast Fourier transform spectrum analysis was then used to identify the characteristic frequency of the adhesive blockage and generate the frequency domain abnormality feature vector.

[0037] The steady-state and non-steady-state noise components in the acoustic emission signal are separated using the least mean square adaptive noise cancellation algorithm. The time-domain waveform characteristics of effective acoustic emission events are extracted using the short-time energy threshold method. The propagation path of the acoustic emission event is inverted and the three-dimensional spatial coordinates are calculated based on the spatial coordinate distribution of the sensor array using the time difference positioning method to locate the abnormal disturbance source.

[0038] A reverse tracing model of the disturbance propagation chain is constructed to map the coupling relationship between parameters such as air pressure, temperature, and spray volume and the device status. The conditional probability table of the node is updated using the maximum likelihood estimation algorithm. The parameter conflict propagation path topology network is traversed based on the improved Dijkstra algorithm. The coupling coefficient defined by the partial derivatives between the parameters is used as the path weight, and a priority sequence of the parameter conflict sources is output.

[0039] If it is detected that the characteristic frequency in the vibration signal belongs to the binder blockage frequency band and the sound source coordinates deviate from the threshold distance of the fluidized bed geometric center, the parameter decoupling replanning instruction is triggered, and a conflict parameter identification code is sent to the control system and the independent adjustment mode is activated;

[0040] Step 5: Based on the porosity and laser-induced breakdown spectroscopy data of the granulated particles, a quantitative mapping between the control parameters and the residual pressure ratio and leakage current is established through a partial least squares regression model, and the reinforcement learning network weights of step 3 are corrected in real time.

[0041] In the above implementation step 1, the broadband dielectric sensor array collects the frequency domain response data of the complex dielectric constant of the fluidized bed inlet material by applying a sweep frequency excitation signal from 1kHz to 10MHz. The real part (ε') and imaginary part (ε") of the complex dielectric constant are nonlinearly fitted based on the Debye multi-relaxation time model. The Cole-Cole distribution function is used to separate the polarization mechanisms corresponding to different relaxation times (τ), and the surface hydroxyl activity (-OH group concentration) and humidity gradient (water distribution state) are quantitatively analyzed. The surface charge density gradient is obtained by calculating the relaxation peak shift dominated by the Maxwell-Wagner interface polarization effect. The terahertz wave tomography module uses a broadband pulse beam from 0.1THz to 1THz to penetrate the material layer, reconstructs the projection data of the scattered signal based on the Radon transform, and uses the filtered back projection algorithm (Filtered Back Projection Algorithm) to obtain the surface hydroxyl activity (-OH group concentration) and humidity gradient (water distribution state). The three-dimensional refractive index distribution field is iteratively solved by the Fourier transform Projection (FBP), and the imaging distortion caused by the sudden change of the refractive index at the particle boundary is corrected by the Mie scattering theory to reconstruct the pore connectivity matrix. In the multimodal data fusion stage, the charge density gradient tensor (three-dimensional) of the dielectric spectrum and the porosity matrix (two-dimensional) of the terahertz imaging are subjected to high-order singular value decomposition (HOSVD) by using the tensor decomposition technology. The dimensional difference between the dielectric parameters (dimension F / m) and the porosity (dimensionless) is eliminated by the eigenvalue weighting algorithm to construct a normalized dynamic physical property parameter matrix. When a sudden change in pore connectivity is detected (such as exceeding ±15%), the terahertz pulse repetition frequency adjustment mechanism based on proportional integral differential control is triggered. The spatiotemporal resolution of the tomographic imaging is optimized by dynamically adjusting the pulse emission interval (10Hz to 100Hz) to ensure imaging accuracy when the material state changes drastically.

[0042] like Figure 2As shown in the figure, the working principle of the adaptive adjustment mechanism of the terahertz wave pulse repetition frequency is as follows:

[0043] The spatial gradient value of the pore connectivity matrix is ​​calculated in real time through the back-projection iterative algorithm. Based on the PID closed-loop control frequency regulator, the error signal calculation module performs proportional-integral-differential operations on the deviation between the current pore connectivity and the target value to generate the pulse repetition frequency adjustment value.

[0044] Inputting the pulse repetition frequency adjustment amount into a direct digital frequency synthesizer, and reconstructing the terahertz wave pulse sequence through a phase accumulator and a waveform lookup table, so that the repetition frequency is adaptively switched within the range of 10 Hz to 100 Hz;

[0045] Dynamically adjust the integration time based on the sliding mode variable structure control algorithm, and control the sampling window width of the data acquisition card through FPGA to match the changes in pulse repetition frequency;

[0046] If the signal-to-noise ratio of the reconstructed pore image is lower than 40 dB, the Radon transform parameter optimizer is triggered, and the projection angle distribution is iteratively corrected using the conjugate gradient method. The optimized tomographic imaging parameters are then output to the pulse control unit, forming a closed-loop feedback regulation link.

[0047] During implementation, the back-projection iterative algorithm is based on the projection data reconstruction principle of Radon transform, and iteratively corrects the three-dimensional refractive index distribution field to calculate the spatial gradient value of the pore connectivity matrix in real time. The gradient amplitude reflects the intensity of the sudden change in the pore structure; when When a preset threshold is exceeded (e.g., gradient amplitude > 0.5 / mm), the PID closed-loop control module is triggered. The error signal calculation module performs a proportional-integral-differential operation on the deviation between the current pore connectivity (P_curr) and the target value (P_target) (e = P_curr - P_target), generating a pulse repetition frequency adjustment value Δf. Its mathematical expression is Δf = K_p·e + K_i·∫edt + K_d·de / dt, where K_p, K_i, and K_d are adjustable gain coefficients. This adjustment value Δf is input into a direct digital synthesizer (DDS), which accumulates the frequency control word (FTW) through a phase accumulator. This drives a waveform lookup table (LUT) to output a digital terahertz pulse train, achieving stepless switching of the repetition frequency between 10 Hz and 100 Hz. A sliding mode variable structure control algorithm is used to dynamically adjust the integral time constant (T_int). The sliding surface is designed using the Lyapunov stability criterion. The integral time is adjusted based on the repetition frequency change rate (df / dt) to ensure that the sampling window width (T_win) of the data acquisition card satisfies T_win∝1 / f to avoid spectral aliasing. If the signal-to-noise ratio (SNR) of the reconstructed pore image is lower than 40dB, the Radon transform parameter optimizer is triggered. The projection angle distribution matrix is ​​iteratively modified based on the conjugate gradient method to minimize the objective function J(θ) = ‖R(θ) - R_meas‖ 2 , where R(θ) is the theoretical projection matrix, R_meas is the measured projection data, and the optimized projection angle set θ_opt is output to the pulse control unit to form a closed-loop feedback link for adaptive imaging parameters.

[0048] In implementation, the broadband dielectric sensor array consists of 16 interdigitated electrode pairs, evenly distributed around the fluidized bed inlet. Communication with the host computer is via a PCIe data acquisition card (sampling rate 1MS / s). The terahertz wave transmitter utilizes a fiber-coupled photoconductive antenna (on an InGaAs substrate) with a pulse width of 100 fs and a peak power of 20 mW. The receiver is equipped with a high-sensitivity pyroelectric detector (sensitivity 0.1 nW / √Hz). Data is preprocessed in real time using an FPGA module (Xilinx Zynq-7000). In the software implementation, the dielectric spectrum analysis module runs a Debye model fitting algorithm on the MATLAB / Simulink platform, with 256 frequency domain data segments and an iterative convergence threshold of 1e-6. Terahertz tomography uses the GPU-accelerated (NVIDIA CUDA) FBP algorithm with a projection angle interval of 1°, 50 iterations, and a spatial resolution of 50 μm. For data fusion, HOSVD decomposition is implemented using the Python TensorLy library, and eigenvalue weighting coefficients are automatically calculated based on the principal component contributions. In the adaptive regulation module, the pulse repetition frequency PID controller parameters are set to K_p = 0.8, K_i = 0.2, and K_d = 0.05, with a response time of ≤ 10ms. The system is connected to the DCS system via the OPC UA protocol, and the dynamic physical property parameter matrix update period is 100ms.

[0049] When designing the experiment, in order to verify the advantages of the multimodal dynamic sensing technology in step 1 over the traditional static calibration method in parameter real-time and control accuracy. A comparative experiment was conducted, and the experimental control group (Group B) used the static dielectric sensor (single-point offline calibration) of patent CN202010491588A combined with the laser particle size analyzer of CN101714439B (sampling interval 5min). The experimental raw materials uniformly used ZnO main material (D50 = 5μm, moisture content 1.5% ± 0.2%); fluidized bed parameters: temperature 450 ± 5℃, air pressure 0.8MPa, spray rate 200L / h; detection cycle: continuous operation for 24 hours, recording data once every hour. The experimental records are shown in Table 1:

[0050] Table 1 Step 1 Experimental Record

[0051]

[0052] Experiments show that the hardware architecture of step 1 breaks through the hysteresis limitations of traditional static models through high-frequency dielectric-terahertz collaborative sensing, providing reliable technical support for the precise control of the high-voltage resistor granulation process.

[0053] In the above-mentioned implementation step 2, the fluidization process prediction model includes a physical property field mapping layer, a dynamic grid generation layer, a hybrid solution calculation layer, a phase change critical judgment layer, an energy aggregation analysis layer and a grid optimization feedback layer; the physical property field mapping layer is used to map the porosity and surface charge density gradient data in the dynamic physical property parameter matrix to the fluid domain grid nodes through the Radon back projection algorithm to establish an initial field distribution; the dynamic grid generation layer is used to construct an initial calculation grid based on real-time process parameters through the Delaunay triangulation algorithm, and trigger a local encryption mechanism when a porosity gradient mutation is detected to generate a field grid topology. structure; the hybrid solution calculation layer is used to calculate the particle collision energy through the explicit discrete element solver using the DEM-CFD coupling method, and solve the flow field and heat conduction equations through the implicit finite volume method, and output the transient flow field, temperature field and particle motion trajectory data; the phase change critical judgment layer is used to identify the local gradient mutation area based on the void ratio gradient field data through the Canny edge detection algorithm, and mark the phase change critical point coordinates when the void ratio gradient value exceeds 0.25mm; the energy aggregation analysis layer is used to scan the particle collision energy field and locate the energy density higher than 1m / Jmm through the DBSCAN density clustering algorithm. 3 the grid optimization feedback layer is used to dynamically adjust the grid density through the unstructured grid dynamic encryption algorithm according to the coupling deviation of heat flow and electric potential gradient, and trigger an iterative optimization cycle when the deviation exceeds a preset threshold.

[0054] Among them, such as Figure 4 As shown in the figure, the automated control process of the secondary granulation of high-voltage zinc oxide resistors starts with physical property field mapping. The dynamic physical property parameters (porosity, surface charge density gradient) are mapped to the fluid domain grid nodes through the Radon back projection algorithm to establish the initial field distribution. Then it enters the dynamic grid generation stage: the initial calculation grid is generated based on the real-time process parameters. If a sudden change in the porosity gradient is detected, the local grid encryption mechanism is triggered, otherwise the standard grid topology is maintained. The generated grid enters the hybrid solution calculation layer, and the DEM-CFD coupling method is used to simultaneously solve the particle collision energy (explicit discrete element) and the flow field heat conduction equation (implicit finite volume method), and outputs the transient flow field, temperature field and particle motion trajectory data.

[0055] The hybrid solution results are processed in two parallel ways: one is the phase transition critical judgment, which uses Canny edge detection to scan the porosity gradient field. If the local gradient value exceeds 0.25mm, it is marked as the phase transition critical point; the other is the energy concentration analysis, which uses the DBSCAN density clustering algorithm to scan the particle collision energy field. When the energy density is higher than 1mJ / mm 3The two-way detection results are aggregated to the grid optimization feedback layer. If the deviation of the coupling between heat flux and electric potential gradient exceeds the threshold, the unstructured grid dynamic encryption algorithm is triggered, the grid density is readjusted and fed back to the dynamic grid generation layer to form an iterative optimization loop; if the deviation does not exceed the standard, the current cycle ends. This process uses a closed-loop feedback mechanism triggered by multiple conditions such as porosity mutation and energy accumulation risk to achieve adaptive adjustment of grid accuracy during the fluidization process, ensuring the phase change prediction accuracy of zinc oxide particles and energy risk control capabilities under high-pressure conditions.

[0056] Among them, the physical field mapping layer is based on the Radon back projection algorithm. The porosity and surface charge density gradient data are projected along the multi-angle ray path to the fluid domain grid nodes through the integral projection theorem. The back projection filter function (such as Ram-Lak filter) is used to suppress Gibbs artifacts. Combined with cubic spline interpolation, the three-dimensional field distribution reconstruction is completed to ensure the conservation of physical parameters in the Euler grid. The dynamic grid generation layer uses the constrained Delaunay triangulation algorithm (CDT) to construct the initial unstructured grid. When the porosity gradient suddenly changes, the local gradient amplitude is used to generate the initial unstructured grid. Adaptive refinement is triggered, and Steiner points are inserted to reconstruct the topology using the Bowyer-Watson algorithm. The Ruppert refinement criterion is also introduced to eliminate slender elements, achieving mesh refinement from a baseline of 1 mm to 0.2 mm while maintaining a geometric quality constraint of a minimum internal angle of ≥30°. A hybrid solution layer employs a DEM-CFD bidirectional weak coupling strategy: an explicit discrete element solver calculates the particle collision force chain based on the Hertz-Mindlin nonlinear contact model, and a Verlet velocity integrator is used to update particle displacements (time step Δt = 1e-6s). An implicit finite volume method employs the PISO algorithm to solve the compressible Navier-Stokes equations at low Mach numbers. Rhie-Chow momentum interpolation is used to suppress pressure oscillations. An immersed boundary method (IBM) is embedded at the particle-fluid interface, and discrete Dirac functions are used to distribute the particle forces to the fluid grid, enabling bidirectional momentum exchange. The phase transition critical judgment layer uses the Canny edge detection operator to perform multi-scale analysis on the void fraction gradient field: first, a Gaussian filter (σ = 1.5) is used to smooth the noise, then the gradient amplitude and direction are calculated, and the edge is refined by non-maximum suppression (NMS). Finally, a double threshold segmentation (high threshold is 0.25mm, low threshold is 0.15mm) is used to extract the phase transition critical area, and the edge breaks are filled in by combining morphological closing operations. The energy concentration analysis layer uses the DBSCAN density clustering algorithm to scan the particle collision energy field, identifies the core points based on the neighborhood radius ε = 0.5mm and the minimum number of points minPts = 5, expands the cluster boundary using the connected component marking algorithm in graph theory, and accelerates the neighborhood search through the KD-Tree spatial index to locate the energy density exceeding 1mJ / mm3 The grid optimization feedback layer introduces the Zienkiewicz-Zhu a posteriori error estimator to calculate the heat flow With potential gradient Coupling deviation Based on anisotropic error index This system triggers dynamic refinement of the unstructured grid, employing the Rivara edge bisection method to split low-quality cells with aspect ratios greater than 10. Local refinement strength is adjusted via a gradient-driven weight function. Iterations are terminated when the global residual norm ‖R‖2 < 1e-4, achieving self-consistent mesh-physics convergence. This architecture systematically improves the spatiotemporal resolution and numerical stability of fluidization process predictions through a closed-loop mechanism involving parameter mapping, mesh adaptation, multi-field coupled solution, and feedback optimization.

[0057] During implementation, the hardware deployment of step 2 uses a multi-core heterogeneous computing cluster (CPU+GPU architecture): 1) The physical field mapping and dynamic mesh generation modules are deployed on the Intel Xeon Gold 6248R processor (48 cores / 3.0GHz), connected to 64GB DDR4 memory via the PCIe4.0 bus, supporting parallel calculations of Radon inverse projection and Delaunay partitioning; 2) The hybrid solution calculation layer uses NVIDIA A100 GPU acceleration, and optimizes the DEM collision force chain calculation (upper limit of 1e6 particles) and FVM flow field solution (upper limit of 5e6 grids) through the CUDA kernel; 3) The phase change criticality judgment and energy analysis module is integrated into the FPGA accelerator card (Xilinx Alveo U280) to realize the hardware pipeline processing of Canny edge detection and DBSCAN clustering (delay <10μs). In terms of software implementation, based on COMSOL The platform builds a multi-physics field coupling framework: 1) The physical field mapping module calls the MATLAB script to implement Radon inverse projection parameter configuration (projection angle step size 1°, iteration number 50 times); 2) The dynamic mesh generation module uses the Gmsh engine, sets the initial mesh size to 1mm, and the trigger condition of the local refinement area is the porosity gradient threshold of 0.05 / mm; 3) The DEM-CFD coupling module uses the EDEM-ANSYS Data was synchronized using the Fluent interface. The particle collision model parameters were set to a restitution coefficient of 0.8 and a friction coefficient of 0.3. The k-ε turbulence model and SIMPLE algorithm were used for fluid analysis. 4) The phase change determination module integrated the Canny operator (Gaussian kernel σ = 1.5, dual thresholds 0.15 / 0.25) from the OpenCV library. The energy clustering module was implemented using DBSCAN from Scikit-learn (ε = 0.5 mm, minPts = 5). 5) The mesh optimization module implemented dynamic mesh refinement using the PETSc library. The error estimator used the Zienkiewicz-Zhu a posteriori error criterion, and the iteration termination criterion was a global residual error of < 1e-4. Data synchronization between hardware was achieved via an InfiniBand network, with real-time feedback at a control cycle of 100ms.

[0058] Compared to traditional single-physics modeling methods, this solution achieves full-domain dynamic analysis of the fluidization process through a multi-level coupling architecture. Its advantages include: breaking through the limitations of traditional static grids and unidirectional coupling, significantly improving the spatial resolution of phase transition critical points and energy accumulation regions; effectively suppressing parameter transfer lag and numerical diffusion through DEM-CFD bidirectional coupling and adaptive grid optimization; and combining hardware acceleration and algorithm optimization to reduce resource consumption while maintaining computational accuracy, providing a high-confidence prediction model for the multi-physics coordinated control of high-voltage zinc oxide resistor granulation processes.

[0059] In the above implementation step 3, if Figure 5 As shown, the working method of the deep reinforcement learning control algorithm based on temporal attention is:

[0060] S1. Build a dual-channel deep Q network architecture to separate state evaluation and action decision paths;

[0061] S2. In the state evaluation channel, a multi-head temporal attention mechanism is used to encode the historical stream state sequence; the multi-head temporal attention mechanism calculates the self-attention weight matrix through formula (1):

[0062]

[0063] In formula (1), W attn(t, τ) is the self-attention weight of the current time step t to the historical time step τ. The domain is defined as t is a positive integer time step and τ∈[tN, t] (historical time window). It is used to measure the contribution of "the fluidization state at a certain moment in the past (such as particle dispersion, airflow stability)" to "the current state evaluation (whether the spray and temperature need to be adjusted)", allowing the system to focus on key historical states (such as the moment of large fluidization fluctuation in the recent period); Q t is the query vector of the current time step t, whose domain is a real number vector matching the process state dimension. It is used to actively encode the current core state of granulation (fluidization index, temperature, porosity, etc.), and serves as the “query end” to match the state pattern in the history that can guide the current decision (such as “when the current temperature is high, which moment in history is the most effective cooling strategy”); K τ is the key vector at the historical time step τ, defined as t A real vector of the same dimension (and τ∈[tN,t]) is the granulation state at the encoding historical moment τ, serving as the “key end” for Q t matching to measure temporal correlation; It's K τ The transposed matrix of K τ The matrix corresponding to the dimension is used to quantize Q through the inner product operation t With K τ The similarity provides a "basic value of relevance" for the attention weight; λ(t,τ) is a dynamic attenuation factor (expression λ(t,τ)=e -β|t-τ| , β is a learnable positive parameter), the domain is t, τ is a positive integer time step, β>0, |t-τ| is a non-negative time difference, which is used to control the "attenuation weight" of the historical state - the larger the time difference (the longer the history), the smaller λ is, reducing the interference of irrelevant states in the distant future (for example, the reference value of "the spray strategy three days ago" to "the current moment" decays over time because the granulation raw materials and environment may change); d k is the key vector dimension scaling factor used to The inner product result is scaled (divided by d k ), to avoid Softmax overflow or gradient disappearance caused by too high a dimension, and to ensure the stability of the attention weight distribution; N is the length of the time window, and its domain is a positive integer (set by the process timing correlation and computational cost, such as 10-20 steps), which is used to limit the attention to only look back to the most recent N steps of history, which not only reduces the amount of calculation but also adapts to the characteristic that "the recent state of granulation is more referenceable to the current adjustment" (such as the fluidization fluctuation within half an hour has a higher reference value for the current adjustment than one day ago); τ′ is a variable that traverses the historical time steps, and its domain is an integer of τ′∈[tN,t], which is used to traverse all historical moments in the time window, calculate the attention score of each moment and normalize it, and finally obtain the weight distribution of the current time step to each historical moment, ensuring that the attention mechanism can comprehensively evaluate the impact of the historical state of multiple moments.

[0064] S3, extracting process parameter temporal association rules through the gated recurrent unit network in the action decision channel;

[0065] S4. Perform tensor fusion on the output feature vectors of the state evaluation channel and the action decision channel using a dual-channel feature fusion formula to generate an initial control instruction set. The dual-channel feature fusion formula is expressed as follows:

[0066] Z t =Concat(H attn ,H gru )·W f +b f (2)

[0067] In formula (2), Z t As the initial control instruction set feature vector generated after fusion, the domain is a real number vector that matches the "state + action" feature dimension and the fusion matrix dimension. It is used to integrate "historical state attention information (such as which fluidization anomalies at historical moments are most critical to current decision-making)" and "process timing rules (such as the dynamic association law of spray → temperature → fluidization index)" to provide a unified feature representation for the subsequent generation of coordinated adjustment instructions for airflow, temperature, and spray field; Concat(·,·) is a feature splicing operation, and its domain is H with compatible input dimensions. attn (attention feature) and H gru (GRU features), used to preserve the complete information of "historical correlation of attention encoding" and "process timing rules extracted by GRU", to avoid information loss during feature fusion and provide a basis for subsequent linear weighted fusion; H attn is the attention feature vector, whose domain is a real number vector matching the output dimension of the attention network. It is used to encode the “importance distribution of fluidization states at different historical moments” in secondary granulation (for example, at which moments the fluidization index mutation and temperature abnormality fluctuation are most critical to the current state assessment, such as particle agglomeration risk and porosity uniformity), supporting the accurate judgment of the core state. gru It is a temporal correlation feature vector, whose domain is a real number vector matching the dimension of the hidden layer of the gated recurrent unit (GRU) network. It is used to capture the “temporal regularity” of the granulation process (e.g., how long does it take for the temperature to change after the spray flow rate is adjusted, and how the fluidization index responds to the dynamic correlation, such as “after the spray frequency is increased by 2 Hz, the infrared power needs to be adjusted 3 steps later to maintain temperature stability”), and provide an “empirical temporal pattern” for action decision-making; W f Is the fusion weight matrix, the domain of which is dimension satisfying (dim(H attn )+dim(H gru ))×dim(Zt)(real number matrix, used to learn the “historical state correlation (Hattn ) and the historical reference of H gru ) (e.g. more emphasis on H attn when the fluidization fluctuates greatly; more emphasis on H gru when the process is stable); b f is a bias term to compensate for the system bias in feature fusion (such as the inherent bias of the granulation equipment - airflow valve mechanical hysteresis, infrared heater thermal inertia), which calibrates the baseline of the "initial control instruction set" (such as the default spray frequency needs to be compensated by 1 Hz to compensate for equipment aging), to ensure the initial rationality of the control instruction.

[0068] S5, when the fluidization index deviates from the target interval, the instruction set is optimized by rolling horizon optimization of the objective function; the formula expression of the rolling horizon optimization objective function is:

[0069]

[0070] The constraint condition is:

[0071]

[0072] Wherein, represents the control increment of the current time , the definition domain is the control increment vector containing the airflow servo valve opening increment, infrared radiation power gradient value, etc., which is used to optimize the control increment in the secondary granulation to achieve the dual objectives of "fluidization index quickly approaching the target" and "control instruction smooth without mutation", avoiding equipment impact (such as particle breakage caused by rapid opening and closing of the airflow valve, and binder splashing); T p is the prediction horizon length, which is determined by the granulation process response delay and computing resources, and needs to cover the time from "control instruction taking effect to fluidization index stabilizing", such as taking 5-10 steps, which is used to predict the fluidization state of the future T p steps to plan the control strategy in advance, adapting to the characteristics of slow dynamics and strong inertia of granulation (such as the fluidization index needs to be stabilized for 3 steps after spray adjustment, and T p ≥ 3 to realize forward-looking control); γ k is a time decay factor (γ ∈ (0, 1) is a learnable or predefined parameter, and k ∈ [0, T p ] is the prediction step number), which is used to emphasize the control effect of "recent prediction steps (k is small)" on the current adjustment, and "far future steps (k is large)" has high uncertainty (such as equipment aging and raw material fluctuation), and γ k reduces the weight of far future punishment, focusing on the executable control instruction at the moment; α t+k∣tis the fluidization index prediction value of future k steps, defined as non-negative real number (such as fluidization index usually 0-100) that meets the requirements of the granulation process, used to quantify the "future k steps of the granulation flow state under the current control instruction", which is the core index to evaluate the control effect (stable fluidization index, uniform particle dispersion, conducive to the densification and electrical performance of secondary granulation); a target is the target fluidization index, a non-negative real number (such as the target value of the granulation stage is set to 60±5) determined by process experiments or expert experience, used as the core control target of granulation quality - when the fluidization index is stabilized near a target , the particle flowability and agglomeration degree are optimal, ensuring the densification, dielectric strength and other key performance of the subsequent tabletting process; is the quadratic norm of state penalty (Q is a positive definite diagonal matrix, the diagonal elements correspond to the penalty coefficients of fluidization index deviation, temperature deviation, porosity deviation), used to weight and punish the core state error (such as the diagonal element of fluidization index in Q is larger; the penalty coefficients of temperature deviation and porosity deviation are secondary (temperature affects the binder curing rate, and porosity affects the tabletting performance, but the disturbance to the immediate flow state is weaker than the fluidization index), through differentiated punishment to guide the control instruction to preferentially correct the "state deviation that most affects the quality". u t+k represents the control amount increment of future k steps (such as airflow servo valve opening correction amount, infrared power gradient value), whose domain is limited by the physical adjustment capability of the granulation equipment (such as airflow valve opening increment per step ≤5%). It is used to replace "absolute instruction mutation" with "increment adjustment" - if the airflow valve jumps from 10% to 50%, it will cause airflow impact, resulting in particle splashing and uneven distribution of binder; while "increment adjustment" can gradually approach the target state for flow field, temperature field and spraying field, ensuring the stability of the granulation process and the service life of the equipment. is the quadratic norm of control amount penalty, and the diagonal elements of R matrix correspond to the penalty coefficients of each control amount increment. In secondary granulation, the penalty coefficient of airflow valve increment is usually larger - because airflow field mutation will directly cause particle agglomeration or breakage (such as strong airflow blowing away the binder, leading to particle forming failure); the penalty coefficient of spraying frequency increment is secondary (it needs to respond dynamically but needs to avoid ultrasonic equipment mechanical loss), through this mechanism to inhibit "excessive aggressive control instruction", and match the adjustment amplitude with the process system inertia. p is the weight of control amount smoothing term (positive real number), which is used to enforce the "temporal continuity" of control instruction - the granulation system has strong inertia (such as after adjusting the airflow field, the fluidization state needs several seconds to minutes to stabilize), if the control amount increment jumps suddenly (such as the airflow valve opening increment jumps from +3% to +10%), it will amplify the process fluctuation (airflow mutation causes particle agglomeration, temperature shock leads to uneven binder curing). Therefore, p constrains the control amount increment to be small, and the control instruction is forced to be "temporally continuous". t - u t-1 || 2, allowing the control instructions to be "slowly adjusted and gradually advanced" to adapt to the process characteristics of granulation "slow dynamics and stable adjustment".

[0073] ||Δu t -Δu t-1 || 2 Quantify the "variation range of the control quantity increment between adjacent moments." The definition domain must match the safety threshold for equipment regulation (e.g., the change in airflow valve opening increment between adjacent steps is ≤3%). This prevents sudden "stops and starts" of control commands. For example, if the current spray frequency increment is +2Hz and the previous one was +1Hz, the change is legal; if it suddenly increases to +8Hz, a penalty is triggered, forcing the command to return to a smooth regulation mode, avoiding mechanical shock to the equipment and deteriorating process stability.

[0074] u min ,u max This is the physical limit of the actuator (e.g., airflow valve opening 0-100%, infrared power 0-50kW), used to prevent control instructions from exceeding the hardware limits of the equipment. If the airflow valve opening is too large, the fan will overload and burn out; if the infrared power is too high, the adhesive will instantly carbonize and fail. Limiting ensures equipment safety and process legitimacy. t |≤Δu max It is the maximum allowable rate of change of the control quantity increment (such as the air flow valve opening can change by a maximum of 5% per step), which is used to avoid the "butterfly effect" caused by excessive single-step adjustment - sudden changes in ultrasonic atomization frequency will cause equipment resonance, resulting in uneven atomized particle size; rapid adjustment of the air flow valve will cause flow field turbulence, which will increase the risk of particle agglomeration sharply. Therefore, by limiting the single-step increment, the "slow dynamic, stable adjustment" characteristics of the granulation process are guaranteed.

[0075] S6. Output the coordinated adjustment instruction set to the actuator.

[0076] In implementation, a dual-channel deep Q-network architecture employs a dual-path design of state evaluation and action decision-making, separating the value function from the policy generation process. The state evaluation channel encodes the historical stream state sequence through a multi-head temporal attention mechanism (Formula 1). Its dynamic decay factor λ = e-βΔt achieves exponential decay of the historical state weights, addressing the long-range dependency forgetting problem of traditional RNNs. The action decision channel uses a gated recurrent unit (GRU) to extract temporal association rules of process parameters. Reset and update gates dynamically adjust the memory unit to capture the nonlinear coupling characteristics between parameters.

[0077] The multi-head temporal attention mechanism, based on a self-attention model, maps the feature vector of each time step in the fluidized state sequence into a query vector, a key vector, and a value vector. A dynamic attenuation factor, introduced by the time difference Δt, suppresses the influence of outdated states and enhances the weighted focus on key events (such as phase transition critical points). This multi-head mechanism (h parallel attention heads in Equation 1) allows the model to capture the multiscale dynamic characteristics of the fluidized bed in different subspaces, for example, simultaneously focusing on the macroscopic vortex structure of the flow field and the microscopic particle collision energy distribution.

[0078] The rolling horizon optimization objective function (Equation 3) employs a model predictive control (MPC) framework to solve the optimal control sequence within a finite time window. The objective function incorporates weighted penalty terms for fluidization index deviation ΔF, temperature deviation ΔT, and porosity deviation ΔP, achieving multi-objective collaborative optimization via the state weighting matrix Q. Among the constraints, Δumax limits the actuator's motion amplitude, while δumax constrains the rate of change of the controlled variable to avoid mechanical shock and thermal stress overload, complying with the physical safety boundaries of industrial control systems.

[0079] During implementation, the hardware deployment consists of a state assessment channel and an action decision channel deployed separately on edge computing devices, interacting with the DCS system in real time via a PCIe interface. The actuators, including a proportional airflow servo valve (Festo MPYE-5), an infrared radiation heater (power range 0-20kW), and an ultrasonic atomizer nozzle (frequency 20-100kHz), receive control commands via the EtherCAT bus. In the software implementation, a multi-head temporal attention mechanism is built using the PyTorch framework, with query / key vector dimension dk = 64, number of attention heads h = 8, and time window length N = 10. The GRU network has a hidden layer dimension of 128, and the time step is synchronized with the DCS sampling period (100ms). The rolling horizon optimizer uses the OSQP solver, with a prediction horizon Tp = 5, a weight matrix Q = diag(1.0, 0.5, 0.3), R = 0.1I, a smoothing term ρ = 0.2, and a control period of 50ms. During the training phase, the reinforcement learning network is updated using the Proximal Policy Optimization (PPO) algorithm, with an initial learning rate of α = 0.001, a discount factor γ = 0.99, and an experience replay pool capacity of 1e5 data points. During real-time operation, if the fluidization index is detected outside the range [0.8, 0.9], the optimizer generates the airflow valve opening correction ΔV (±5%), the infrared power gradient ΔP (±500W), and the atomization frequency ΔF (±10kHz) within 10ms. A priority queue is used to ensure that critical instructions (such as air pressure adjustment) are executed first.

[0080] During the design of the scheme, in order to verify the performance advantages of the deep reinforcement learning control algorithm based on temporal attention over the traditional PID control in the coordinated regulation of multiple physical fields, a comparative experiment was conducted. Among them, the existing method (Group B) adopted a PID control algorithm combined with a multi-loop compensation strategy with fixed parameters. Experimental fluidized bed parameters: temperature 450±10℃, air pressure 0.8±0.02MPa, spray rate 200±5L / h; raw material batch: ZnO main material (D50=5.0±0.3μm, moisture content 1.5%±0.1%); sampling period: continuous operation for 24 hours, recording data every 10 seconds. Among them, Group A deployed a deep reinforcement learning control algorithm based on temporal attention to generate airflow, temperature and spray coordination instructions in real time; Group B adopted a PID controller (proportional coefficient Kp=2.0, integral time Ti=10s, differential time Td=0.5s) combined with feedforward compensation; the experimental records are shown in Table 2:

[0081] Table 2 Step 3 Application Experiment Record

[0082]

[0083] Experiments have shown that the deep reinforcement learning algorithm based on temporal attention effectively solves the parameter mismatch problem of traditional PID control under complex coupling conditions through dynamic state perception and multi-objective rolling optimization, providing a high-precision and high-robustness control solution for high-voltage resistor granulation. Therefore, through the collaborative design of the temporal attention mechanism and rolling optimization, this solution breaks through the limitations of traditional PID control in insufficient modeling of multi-field coupling effects and effectively solves the control instability problem caused by parameter conflicts; the dynamic attenuation factor strengthens the decision weight of the recent state, significantly improving the system's response speed to sudden changes in raw materials; the multi-objective optimization framework balances fluidization stability and actuator life to ensure process continuity under complex conditions.

[0084] In the above implementation step 4, when the improved Dijkstra algorithm is initialized, the conflict source node is set as the starting point based on the topological network node set and edge weights output by the parameter coupling graph construction layer; the priority queue is managed by the Fibonacci heap data structure, with the path cumulative weight as the key value, and if the current node is a strongly coupled path endpoint, the edge weight is updated by the dynamic weight correction coefficient; during the traversal process of the improved Dijkstra algorithm, the coupling coefficient updated by the path weight calculation layer is received in real time. If it is detected that the weight change of any edge exceeds the preset threshold, the current search is interrupted and the edge is reinserted into the priority queue; when the target node is searched, the path is traced back through the reverse pointer, and the conflict propagation chain with the minimum cumulative weight is output; at the same time, if there are multiple equal-weight paths to the same node, the posterior probability selection mechanism based on the conditional probability table is triggered, and the path with a higher conditional probability value is selected as the optimal solution to generate the conflict source positioning result.

[0085] The wavelet packet transform of high-frequency vibration signals maps the signal into 32 subbands through multi-scale decomposition, and uses energy entropy to quantify the differences in energy distribution within the frequency bands. This approach relies on the time-frequency localization of wavelet basis functions to capture the transient characteristics of non-stationary signals, while energy entropy, as an extended form of information entropy, effectively characterizes frequency domain complexity. Hilbert envelope detection, combined with fast Fourier transform (FFT) complex analysis, extracts the resonant characteristic frequencies of adhesive blockage by demodulating and refining the high-frequency modulated signal. Essentially, envelope demodulation separates the fault modulated signal from the carrier, followed by spectral analysis to extract fault features. The least mean square adaptive noise cancellation (LMS-ANC) algorithm for acoustic emission signals, based on Wiener filter theory, eliminates steady-state background noise and incoherent interference by adaptively adjusting filter weights, preserving the impulse waveform characteristics of sudden acoustic emission events. The short-term energy thresholding algorithm, based on the statistical characteristics of energy mutations in acoustic emission events, combines a predefined threshold to capture valid events. The time-difference positioning algorithm is based on a geometric acoustic model of elastic wave propagation. It uses the time difference of arrival (TDOA) of a sensor array to construct a system of hyperbolic equations and solves the three-dimensional coordinates through least-squares optimization. Its core is the assumption of wave velocity consistency and the geometric constraints of the sensor spatial distribution. The disturbance propagation chain backtracking model uses a Bayesian network to model parameter coupling relationships. Maximum likelihood estimation (MLE) iteratively updates the conditional probability table (CPT) using observation data to reflect the dynamic dependence strength between parameters. The improved Dijkstra algorithm introduces a dynamic weight correction mechanism, using the Jacobian matrix elements defined by the partial derivatives between parameters as coupling coefficients. Combined with the Fibonacci heap to optimize path search efficiency, it achieves optimal solution search for conflicting propagation paths in topological networks. Its improvements lie in the real-time weight update strategy and the posterior probability optimization mechanism, which address the lack of adaptability of traditional algorithms to dynamic coupling relationships.

[0086] During implementation, vibration monitoring uses PCB 352C03 piezoelectric accelerometers (frequency response range 0.5Hz-10kHz) installed on the four quadrants of the fluidized bed wall, and acoustic emission sensors use PAC Micro80 (center frequency 300kHz) arranged in a circular array on the upper flange of the bed. The signal acquisition module integrates the NIPXIe-6358 multi-channel synchronous acquisition card (sampling rate 5MS / s) and transmits it to the industrial computer via the PCIe bus. At the software implementation level, the vibration signal is decomposed into 32 sub-bands through a 6-layer wavelet packet (db10 wavelet basis). When calculating the energy entropy of each frequency band, a sliding window length of 1024 points is used, and the entropy threshold is set to 0.85 to trigger an abnormality judgment. The FFT analysis after Hilbert envelope demodulation focuses on the 50-80kHz frequency band, and the characteristic frequency matching uses the correlation coefficient method (threshold 0.75). In acoustic emission signal processing, the LMS-ANC filter order was set to 32, with a step size of μ = 0.01. The noise reference channel was derived from the vibration signal at the bottom of the bed. A 256-point Hamming window was used for short-term energy calculation, and the energy threshold was dynamically set based on the 3σ principle of the baseline noise level. The wave velocity in the time-of-day positioning algorithm was set to 2800 m / s (the longitudinal wave velocity of the steel bed), and the Levenberg-Marquardt nonlinear optimization algorithm was used for coordinate inversion. When constructing the parameter-coupled topology network, air pressure (0-0.6 MPa), temperature (80-120°C), and spray volume (5-20 mL / min) were defined as node variables. The coupling coefficient was obtained by obtaining the Jacobian matrix partial derivatives from offline step response experiments. In the implementation of the improved Dijkstra algorithm, the Fibonacci heap was initialized with 50 nodes, the dynamic weight modification coefficient k = 1.2 (strong coupling path), the weight change threshold was 0.05, and Bayesian inference was used to update the path confidence based on the posterior probability optimization. The control system receives the priority sequence of conflict sources via the OPC UA protocol and uses a PID controller (proportional band 40%, integral time 120s) to achieve decoupling control when triggering parameters to be adjusted independently.

[0087] This solution significantly improves the early detection capability of fluidized bed binder blockage and the positioning accuracy of parameter conflict sources through multi-physics field signal fusion analysis and dynamic coupling topology modeling. Compared with traditional single-modal monitoring and static decoupling methods, it can effectively avoid false alarms and missed alarms, achieve rapid response and autonomous control optimization of abnormal operating conditions, improve the stability of the secondary granulation process of zinc oxide resistors and product consistency, while reducing energy consumption and raw material loss caused by abnormal equipment shutdowns.

[0088] At the same time, in step 4, the disturbance propagation chain reverse tracing model includes a parameter coupling graph construction layer, a conditional probability table update layer, a path weight calculation layer, a path optimization layer, a conflict source priority sorting layer and a dynamic feedback correction layer; the parameter coupling graph construction layer is used to map the parameters into nodes based on the historical correlation data of air pressure, temperature, spray volume parameters and equipment status through graph theory methods, and map the physical coupling relationship between parameters into directed edges to construct an initial propagation path topology network; the conditional probability table update layer is used to adopt the maximum likelihood estimation algorithm to update the node conditional probability table based on the real-time monitored vibration and acoustic emission abnormality feature data; the path weight calculation layer is used to calculate the coupling coefficient through the partial derivative chain method, and normalize the coupling coefficient to the path path weight, generate a coupling coefficient matrix, if the absolute value of the partial derivative of the coupling coefficient is greater than 0.5, it is marked as a strong coupling path; the path optimization layer is used to traverse the propagation path topology network with the coupling coefficient matrix as the adjacency matrix, optimize the search efficiency through the relaxation factor dynamic adjustment strategy, and output the shortest conflict propagation chain; the conflict source priority sorting layer is used to calculate the conflict source confidence based on the product of the path cumulative weight and the node conditional probability, generate the conflict source priority sequence through the quick sorting algorithm, and mark the node with a confidence higher than 0.7 as a first-level conflict source; the dynamic feedback correction layer is used to feed back the compensated fluidization index deviation to the conditional probability table, iteratively optimize the coupling coefficient through the gradient descent method, and trigger the dynamic update of the topological network edge weight.

[0089] Among them, the conditional probability table update layer uses Markov chain Monte Carlo (MCMC) sampling combined with maximum likelihood estimation (MLE) to dynamically update the node conditional probability distribution through the Bayesian network. Its mathematical essence is to perform parameterized reconstruction of the joint probability distribution, using real-time vibration and acoustic emission abnormality feature data as observation evidence to iteratively optimize the posterior probability of the latent variable. The path weight calculation layer calculates the coupling coefficient between parameters based on the chain derivative rule and generates the path weight matrix through Sigmoid function normalization processing. The strong coupling path is determined by the threshold segmentation algorithm based on the gradient modulus ( This mechanism, derived from an analysis of the vanishing / exploding gradient phenomenon in deep learning, enhances path search stability by constraining the values ​​of high-order partial derivatives. The path optimization layer improves the Dijkstra algorithm by introducing a dynamic relaxation factor adjustment strategy. Fibonacci heaps are used to optimize priority queue operations. During topological network traversal, the path weight update step size is dynamically adjusted using a relaxation factor α∈(0,1), balancing the conflict between global search and local convergence. The conflict source prioritization layer employs a confidence fusion algorithm, taking the product of the cumulative path weight and the node conditional probability as the confidence metric. A divide-and-conquer strategy based on QuickSort is designed to achieve sorting with O(n log n) time complexity. Its physical significance lies in quantifying the multi-dimensional coupling contribution of parameter conflicts in the time-frequency-space domain. The dynamic feedback correction layer constructs a closed-loop control loop. Feedback signals from the fluidization index deviation drive a gradient descent optimizer (learning rate η = 0.01). The coupling coefficient matrix is ​​iteratively updated using the mean squared error (MSE) as the objective function, correcting the topological network edge weights in real time. This process is essentially an adaptive parameter identification based on Lyapunov stability theory.

[0090] In implementation, the parameter coupling data acquisition adopts the NI CompactDAQ-9185 chassis integrated PCIe-6353 data acquisition module (16-bit resolution, 1 MS / s sampling rate) to obtain the real-time data stream of gas pressure (0-0.6 MPa), temperature (80-120°C), and spray amount (5-20 mL / min) through the Modbus TCP protocol connection with the fluidized bed PLC controller (Siemens S7-1500), and synchronously accesses the analog signals of the PCB 356A01 three-axis vibration sensor (frequency band 0.5-10 kHz) and the PAC Micro-II acoustic emission sensor (center frequency 300 kHz). In software implementation, the parameter coupling graph construction layer constructs a dynamic graph data structure based on the PyTorch Geometric library, sets the node feature vector dimension to 128, and initializes the initial edge weight through the Jacobian matrix obtained by offline multiple regression analysis. The conditional probability table update layer uses the Pyro probabilistic programming framework to implement variational inference, receives the vibration signal energy entropy (window length 1024 points) and acoustic emission event time domain features (short-time energy threshold 3σ) every 5 seconds, updates the Bayesian network parameters through stochastic gradient Hamiltonian Monte Carlo (SGHMC) sampling. The path weight calculation layer embeds an automatic differentiation engine (PyTorch Autograd) to perform symbolic differentiation on the parameter coupling function f(x, y), generates a dense Jacobian matrix, and suppresses overfitting through L1 regularization (λ = 0.01). The path optimization layer uses an improved Dijkstra algorithm, with a Fibonacci heap capacity preset to 200 nodes and a relaxation factor α dynamically adjusted according to the path length (α = 0.8 × exp(-0.05L)). The conflict source priority sorting layer uses a CUDA-accelerated quicksort kernel function, with a confidence calculation that combines a weight factor w = 0.6 (path cumulative weight) and 1-w = 0.4 (node conditional probability). The dynamic feedback correction layer deploys an Adam optimizer (β1 = 0.9, β2 = 0.999), receives D50 particle size feedback from the laser diffraction particle size analyzer (Malvern Mastersizer 3000) every 30 seconds, updates the coupling coefficient matrix through backpropagation, and triggers dynamic re-planning of the topology network weight.

[0091] Compared with traditional static parameter decoupling methods, this model significantly improves the tracing accuracy and response speed of multi-parameter coupling conflicts through the deep integration of dynamic graph optimization and Bayesian probability reasoning, effectively solving the control lag problem caused by strong parameter coupling in the high-pressure zinc oxide resistor pelletization process.

[0092] To verify the effectiveness of this step, a comparative experiment was also designed. Group A (Step 4) used joint monitoring of high-frequency vibration and acoustic emission, combined with the improved Dijkstra algorithm to locate the parameter conflict source; Group B (existing method) was based on single-modal vibration monitoring + static Bayesian network parameter association model, without a dynamic weight correction mechanism.

[0093] In the experiment, the fluidized bed specifications were: diameter 1.2m, working pressure 0.4MPa, spray volume 15mL / min, and binder PVP-K30; each experiment was repeated 5 times, each lasting 30 minutes, with a sampling interval of 10ms. The experimental data are shown in Table 3:

[0094] Table 3 Step 4 Application Experiment Record

[0095]

[0096]

[0097] Experiments showed that Group A improved its detection rate by 16.3%. Multimodal signal fusion (vibration + acoustic emission) effectively identified weak blockage features, reducing the missed detection rate from 17.6% (Group B) to 1.8% (Group A). ​​Acoustic emission propagation path inversion based on the time difference positioning method reduced the average error from 13.5mm (Group B) to 3.8mm (Group A), relying on dynamic coupling coefficient topology network optimization. An improved Dijkstra algorithm reduced path search delay, shortening response time from 316ms to 123ms, ensuring real-time control. A parameter decoupling mechanism avoided ineffective adjustments, compressing the fluidization index fluctuation range to ±1.4% (Group A), better than the ±4.5% of Group B, and reducing the energy consumption of repeated granulation. In summary, Step 4 significantly improved the reliability and energy efficiency of secondary granulation of high-voltage zinc oxide resistors through multi-sensor collaboration and dynamic coupling analysis.

[0098] In the above implementation step 5, the partial least squares regression model extracts the spectral principal component vector through multimodal data normalization and kernel principal component dimensionality reduction, and inputs it into the latent variable space together with the control parameter time series data; then, the latent variable weights are iteratively calculated based on the covariance maximization criterion, and if the cross-validation error does not converge, the dynamic latent variable number reduction mechanism is triggered; when the regression residual exceeds the dynamically adjusted adaptive threshold, the regression coefficient matrix and the reinforcement learning weight are subjected to the Kronecker product operation to generate a gradient correction; if the detected gradient direction is consistent with the historical trajectory, the learning rate is increased through the momentum acceleration mechanism to update the network weights.

[0099] Among them, multimodal data normalization uses the Z-score method to eliminate the dimensional differences between laser-induced breakdown spectroscopy (LIBS) and porosity detection data. Its core lies in constructing a unified data distribution space to enhance feature comparability. Specifically, it normalizes the data by calculating the mean and standard deviation of each modal data to avoid model weight shift due to dimensional differences. Kernel principal component dimensionality reduction (Kernel PCA) uses a Gaussian kernel function to map high-dimensional spectral data to the reproducing kernel Hilbert space. The nonlinear principal components are extracted by solving the eigenvectors of the covariance matrix. Its essence is to enhance the correlation between spectral features and control parameters through nonlinear transformation while suppressing noise interference. The relevant theory refers to the mathematical framework of kernel partial least squares regression. The covariance maximization criterion iteratively solves the latent variable weights based on the NIPALS algorithm. The latent variable space is constructed by maximizing the covariance between the control parameter time series data and the spectral principal components. Its mathematical essence is to maximize the joint variance contribution of the independent variable and the dependent variable, forming an orthogonal latent variable sequence to capture the dominant dynamic characteristics of the system. The dynamic latent variable reduction mechanism uses the k-fold cross-validation error as the convergence criterion. If the validation error does not decrease after three consecutive iterations, a latent variable reduction strategy is triggered. By gradually eliminating low-contribution latent variables, the model complexity is dynamically balanced to avoid overfitting. An adaptive threshold for the regression residual is calculated based on the 3σ value of the residual root mean square (RMSE) using a sliding window. When the residual exceeds the threshold, a Kronecker product is performed on the regression coefficient matrix and the reinforcement learning weight matrix to generate a gradient correction. This operation leverages the fully connected nature of the tensor space to achieve global gradient propagation in the parameter space. A momentum acceleration mechanism uses an exponentially weighted moving average (EWMA) to update the learning rate. The weight update step size is adjusted by accumulating an adaptive momentum factor based on the historical gradient direction. The acceleration mechanism is triggered if the angle between the current gradient direction and the historical trajectory is less than 30°. Its mathematical basis is the leaky average principle of the momentum method, effectively suppressing gradient oscillations and accelerating convergence.

[0100] During implementation, LIBS spectra were acquired using a spectral detection module consisting of an Nd:YAG laser (wavelength 1064 nm, pulse energy 150 mJ) and an AndorICCD spectrometer (resolution 0.1 nm). This was optically coupled to the observation window of the granulation fluidized bed and synchronously connected to a NIPXIe-4499 high-speed acquisition card (sampling rate 2 MS / s) to acquire the plasma emission spectrum in real time. Porosity was determined using a Malvern Mastersizer 3000 laser particle size analyzer, and D50 particle size distribution data was input into an industrial computer via the Modbus TCP protocol. At the software level, a multimodal normalization module applied Savitzky-Golay smoothing filtering (window width 11 points, cubic polynomial) to the LIBS spectra (512 channels in the 200-980 nm range) and calculated the Z-score value. The porosity data was then subjected to a Box-Cox transformation to eliminate skewed distributions. Kernel principal component dimensionality reduction (KPCD) was performed using a Gaussian kernel function (bandwidth parameter γ = 0.01) to extract the first five principal components (cumulative contribution ≥ 85%). These components were aligned with the time series data of the control parameters (air pressure 0.4-0.6 MPa, temperature 80-120°C, and spray volume 10-20 mL / min) via dynamic time warping (DTW) and then entered into the latent variable space. Covariance maximization was iteratively calculated using the NIPALS algorithm (maximum number of iterations 100, convergence threshold 1e-6). The number of dynamic latent variables was initially set to 8, and cross-validation was performed using a 5-fold grouping (20% validation set) with an error tolerance of 0.05. The regression residual adaptive threshold module set a sliding window length of 50 sampling points. When the residual exceeded the threshold, a Kronecker product operation was triggered to generate a 32-dimensional gradient correction vector, which was written to the reinforcement learning network weight register of a Siemens S7-1500 PLC via the OPC UA protocol. The momentum acceleration mechanism initializes the momentum factor β = 0.9, the base learning rate value is 0.01, the maximum acceleration ratio is 5 times, and the gradient direction consistency detection is calculated using cosine similarity (the threshold value 0.866 corresponds to a 30° angle).

[0101] Compared with the traditional single-modal static regression model, this scheme effectively solves the problem of nonlinear coupling modeling of high-dimensional spectral data and control parameters through kernel principal component dimensionality reduction and dynamic latent variable mechanism. Combined with the gradient correction strategy of momentum acceleration, it significantly improves the convergence speed and stability of the reinforcement learning network, and realizes precise closed-loop control of the residual voltage ratio and leakage current in the granulation process of zinc oxide resistors.

[0102] like Figure 6 As shown in the figure, when implementing the automatic control system for secondary granulation of high voltage zinc oxide resistor sheets, it includes:

[0103] The real-time physical property sensing module is used to analyze the hydroxyl activity and humidity gradient on the material surface based on the dielectric relaxation spectrum decomposition algorithm, reconstruct the three-dimensional pore distribution by combining terahertz wave inverse scattering tomography, and output the dynamic physical property parameter matrix to the dynamic multi-field modeling module. The dynamic physical property parameter matrix includes dielectric loss factor, porosity, and surface charge density gradient;

[0104] The dynamic multi-field modeling module is used to construct a discrete element fluid-mechanical coupling model. It uses an explicit / implicit hybrid iterative algorithm to simultaneously solve the heat conduction equation, the particle collision energy equation, and the electric field potential energy equation. It predicts the bubble phase / emulsion phase transition threshold and the microscopic defect risk coordinates, generates a fluidization state map, and outputs it to the intelligent collaborative control module.

[0105] An intelligent collaborative control module uses a multi-head temporal attention mechanism to extract fluidization uniformity features, combines a rolling time domain optimization framework to generate collaborative adjustment instructions for airflow valve opening, infrared power, and atomization frequency, and outputs control instructions to the actuator and abnormal reverse suppression module;

[0106] The abnormal reverse suppression module is used to identify the characteristic frequency of adhesive blockage based on the wavelet envelope analysis algorithm, locate the source of parameter conflict through the disturbance propagation chain reverse tracking algorithm, dynamically decouple the pressure-temperature reverse coupling effect, output compensation instructions to the actuator, and synchronously feedback abnormal characteristics to the closed-loop feedback optimization module;

[0107] The closed-loop feedback optimization module is used to associate control parameters with electrical performance indicators through a partial least squares regression model. When grain boundary segregation is detected, the graph convolutional network is triggered to mine implicit control rules, update the reinforcement learning network weights and reward function constraints, and output the optimized parameters to the intelligent collaborative control module.

[0108] During implementation, the real-time physical property perception module analyzes the surface hydroxyl activity of zinc oxide particles through a dielectric relaxation spectrum decomposition algorithm. The core of this module is to quantify the dielectric response characteristics of the material by utilizing the correlation between the dielectric loss factor and the polarization relaxation time. Combined with the three-dimensional refractive index inversion algorithm of terahertz backscatter tomography, the pore distribution is reconstructed through the phase unwrapping technology of the frequency-domain scattering matrix, achieving subvoxel-level analysis of the porosity at the scale of 0.1-10μm. The dynamic multi-field modeling module adopts a discrete element-computational fluid dynamics bidirectional coupling model, introduces an explicit time-stepping method to solve the particle collision energy equation, combines it with the implicit finite element method to solve the unsteady-state heat conduction equation, and realizes multi-scale field coupling through field variable interpolation. The intelligent collaborative control module designs a multi-head temporal attention mechanism, uses a self-attention weight matrix to extract the dynamic flow field characteristics of the fluidized bed, and constructs a rolling time domain optimization framework to solve the Pareto optimal solution set of the airflow valve opening, infrared power, and atomization frequency through quadratic programming. The anomaly back-suppression module extracts the modulation frequency components of binder blockage based on wavelet envelope analysis. It locates the source of parameter conflict through backpropagation of the partial differential adjoint equation of the disturbance propagation chain, and employs singular value decomposition (SVD) to achieve real-time decoupling of the pressure-temperature coupling matrix. The closed-loop feedback optimization module establishes a partial least squares regression (PLSR) model, associating control parameters with the grain boundary segregation index SGB = XGB / X0 using the latent variable spatial projection weight wi. When SGB > 0.8 is detected, a graph convolutional network (GCN) is triggered to explore implicit topological rules between process parameters. Backpropagation is then used to update the Boltzmann policy gradient of the reinforcement learning network.

[0109] Among them, the physical property perception module uses Keysight N5227B vector network analyzer (10 MHz-67 GHz) to collect dielectric relaxation spectrum, integrates Terahertz Imaging System TI-1000 type terahertz imaging module (center frequency 0.3 THz) to realize three-dimensional module analysis scanning, and is connected with NIPXIe-8880 industrial computer through PCIe 4.0 interface. The dynamic modeling module deploys the ANSYS Twin Builder digital twin platform, configures a 128-core GPU cluster (NVIDIA A100) to solve the DEM-CFD coupled equation in parallel, and uses Octree grid self-adaptive encryption (minimum grid size 50 μm) for spatial discretization. The intelligent control module builds a distributed control system based on the ROS2 framework, uses Festo VUVG-L15-P76 proportional valve (linearity ±0.5%) for airflow valve, and realizes power closed-loop control of infrared heater through Omron E5CC temperature control module (PID parameter Kp=1.2, Ti=120 s). The anomaly suppression module realizes real-time wavelet envelope analysis on FPGA (Xilinx Zynq UltraScale+), uses Daubechies 9 wavelet basis for 8-layer decomposition, and sets the feature frequency identification threshold to ±5% of the fundamental frequency. The closed-loop optimization module uses Python / Matlab hybrid programming, sets the number of latent variables of the PLSR model to 5, and the GCN graph node contains 12-dimensional process parameters (air pressure 0.4-0.8 MPa, temperature 80-120℃, etc.), and the edge weight is defined by the Pearson correlation coefficient ρ>0.7.

[0110] Although the specific embodiments of the present application are described above, those skilled in the art should understand that these specific embodiments are only illustrative, and those skilled in the art can make various omissions, substitutions and changes to the details of the above-mentioned method and system without departing from the principles and essence of the present application. For example, combining the above method steps, performing substantially the same function in substantially the same manner to achieve substantially the same result according to the same method is within the scope of the present application. Therefore, the scope of the present application is only limited by the appended claims.

Claims

1. A method for automatically controlling secondary granulation of high voltage zinc oxide resistors; characterized by: include: Step 1: Based on multi-band dielectric relaxation spectrum analysis and terahertz wave tomography, the dielectric loss factor, surface charge density gradient and porosity are analyzed in real time to generate a dynamic physical property parameter matrix; Step 2: Based on the multi-physics field coupling modeling method, the dynamic physical property parameter matrix is ​​coupled with the real-time process parameters to establish a fluidization process prediction model including thermal field, force field and electric field, and output fluidization state characteristic parameters, including the phase change critical threshold and energy accumulation area of ​​the fluidization process; Step 3: Generate coordinated adjustment instructions for the airflow field, temperature field, and spray field according to the fluidization state characteristic parameters through a deep reinforcement learning control algorithm based on temporal attention; Step 4: Using a high-frequency vibration and acoustic emission joint monitoring module to identify abnormal disturbance characteristics of the fluidized bed in real time and trigger a parameter decoupling and replanning mechanism; the high-frequency vibration and acoustic emission joint monitoring module uses wavelet envelope analysis to identify the characteristic frequency of binder blockage and locates the source of parameter conflict through a disturbance propagation chain reverse tracing model; Step 5: Based on the porosity and laser-induced breakdown spectroscopy data of the granulated particles, a quantitative mapping between the control parameters and the residual pressure ratio and leakage current is established through a partial least squares regression model, and the reinforcement learning network weights of step 3 are corrected in real time.

2. The method for automatically controlling secondary granulation of zinc oxide resistor sheets for high voltage use according to claim 1, characterized in that: The working principle of step 1 is as follows: the dielectric response signal of the material at the fluidized bed inlet is collected through a broadband dielectric sensor array, the frequency domain curves of the real and imaginary parts of the complex dielectric constant are decomposed based on the Debye multi-relaxation time model, and the surface hydroxyl activity and humidity gradient parameters are extracted; at the same time, a terahertz wave transmitting device is used to project a pulse beam in the frequency band of 0.1THz to 1THz to the material layer, the scattered signal is tomographically imaged through Radon transform, and the three-dimensional pore distribution topology is reconstructed based on the back projection iterative algorithm; then, the surface charge density gradient tensor obtained by dielectric spectrum decomposition is tensor-fused with the pore connectivity matrix of the terahertz tomography, and the dimensional differences of cross-modal data are eliminated through the eigenvalue weighting algorithm to generate a dynamic physical property parameter matrix; During the detection process, if the pore connectivity rate suddenly changes beyond the preset threshold, the tomographic imaging resolution is adjusted through the terahertz wave pulse repetition frequency adaptive adjustment mechanism to dynamically match the material state changes.

3. The method for automatically controlling secondary granulation of a high voltage zinc oxide resistor according to claim 2, characterized in that: The working principle of the terahertz wave pulse repetition frequency adaptive adjustment mechanism is: The spatial gradient value of the pore connectivity matrix is ​​calculated in real time through the back-projection iterative algorithm. Based on the PID closed-loop control frequency regulator, the error signal calculation module performs proportional-integral-differential operations on the deviation between the current pore connectivity and the target value to generate the pulse repetition frequency adjustment value. Inputting the pulse repetition frequency adjustment amount into a direct digital frequency synthesizer, and reconstructing the terahertz wave pulse sequence through a phase accumulator and a waveform lookup table, so that the repetition frequency is adaptively switched within the range of 10 Hz to 100 Hz; Dynamically adjust the integration time based on the sliding mode variable structure control algorithm, and control the sampling window width of the data acquisition card through FPGA to match the changes in pulse repetition frequency; If the signal-to-noise ratio of the reconstructed pore image is lower than 40 dB, the Radon transform parameter optimizer is triggered, and the projection angle distribution is iteratively corrected using the conjugate gradient method. The optimized tomographic imaging parameters are then output to the pulse control unit, forming a closed-loop feedback regulation link.

4. The method for automatically controlling secondary granulation of a high voltage zinc oxide resistor according to claim 1, characterized in that: The fluidization process prediction model includes a physical property field mapping layer, a dynamic grid generation layer, a hybrid solution calculation layer, a phase change criticality judgment layer, an energy aggregation analysis layer, and a grid optimization feedback layer; the physical property field mapping layer is used to map the porosity and surface charge density gradient data in the dynamic physical property parameter matrix to the fluid domain grid nodes through the Radon back projection algorithm to establish the initial field distribution; The dynamic grid generation layer is used to construct the initial calculation grid based on real-time process parameters through the Delaunay triangulation algorithm. When a porosity gradient mutation is detected, the local encryption mechanism is triggered to generate a field grid topology structure; the hybrid solution calculation layer is used to adopt the DEM-CFD coupling method to calculate the particle collision energy through the explicit discrete element solver, and at the same time solve the flow field and heat conduction equations through the implicit finite volume method to output transient flow field, temperature field and particle motion trajectory data; the phase change critical judgment layer is used to identify the local gradient mutation area through the Canny edge detection algorithm based on the porosity gradient field data, and mark the phase change critical point coordinates when the porosity gradient value exceeds 0.25mm; the energy aggregation analysis layer is used to scan the particle collision energy field and locate the energy density higher than 1m / Jmm through the DBSCAN density clustering algorithm. 3 the grid optimization feedback layer is used to dynamically adjust the grid density through the unstructured grid dynamic encryption algorithm according to the coupling deviation of heat flow and electric potential gradient, and trigger an iterative optimization cycle when the deviation exceeds a preset threshold.

5. The method for automatically controlling secondary granulation of high voltage zinc oxide resistor sheets according to claim 1, characterized in that: The working method of the deep reinforcement learning control algorithm based on temporal attention is: S1. Build a dual-channel deep Q network architecture to separate state evaluation and action decision paths; S2. In the state evaluation channel, a multi-head temporal attention mechanism is used to encode the historical stream state sequence; the multi-head temporal attention mechanism calculates the self-attention weight matrix through formula (1): In formula (1), Q t represents the query vector at the current time step t; K τ is the key vector of the historical time step τ, used to compare with Q t Matching historical state correlation; λ(t,τ) represents the dynamic attenuation factor, expressed as λ(t,τ)=e -β|t-τ| , where β is a learnable parameter used to control the decay rate of the historical state weight; |t-τ| represents the time difference between the current time step and the historical time step; d k represents the key vector dimension scaling factor; N represents the time window length; K τ The transposed matrix of S3, extracting process parameter temporal association rules through the gated recurrent unit network in the action decision channel; S4. Perform tensor fusion on the output feature vectors of the state evaluation channel and the action decision channel using a dual-channel feature fusion formula to generate an initial control instruction set. The dual-channel feature fusion formula is expressed as follows: Z t =Concat(H attn ,H gru )·W f +b f (2) In formula (2), H attn represents the attention feature vector; H gru is the time series correlation feature vector; W f is the fusion weight matrix; b f is the bias term; S5. When it is detected that the fluidization index deviates from the target range, the instruction set is constrainedly optimized by using a rolling horizon optimization objective function; the formula expression of the rolling horizon optimization objective function is: The constraints are: In the rolling horizon optimization objective function, α t+k∣t represents the predicted value of the fluidization index at the kth step in the future; α target Indicates the target fluidization index; Δu t+k represents the control quantity increment of the kth step in the future, including the airflow servo valve opening correction, the infrared radiation power gradient value and the ultrasonic atomization frequency modulation coefficient; Q is the state weighting matrix, and the diagonal elements correspond to the penalty coefficients of fluidization index deviation, temperature deviation and porosity deviation respectively; R is the control quantity weighting matrix, which is used to suppress the amplitude of the control instruction; γ is the time attenuation factor; ρ is the control quantity smoothing term weight, which is used to prevent the control instruction from changing suddenly in adjacent time steps; u min / u max is the physical limit value of the actuator; Δu max The maximum permissible rate of change of the control quantity increment; S6. Output the coordinated adjustment instruction set to the actuator.

6. The method for automatically controlling secondary granulation of high voltage zinc oxide resistor sheets according to claim 1, characterized in that: The working principle of the high-frequency vibration and acoustic emission joint monitoring module is as follows: The vibration signal was decomposed into 32 sub-bands using wavelet packet transform, and the energy entropy of each sub-band was calculated to quantify the unevenness of the energy distribution in the frequency domain. Hilbert envelope detection was used to extract the time domain envelope signal of the high-frequency modulation component. Fast Fourier transform spectrum analysis was then used to identify the characteristic frequency of the adhesive blockage and generate the frequency domain abnormality feature vector. The steady-state and non-steady-state noise components in the acoustic emission signal are separated based on the least mean square adaptive noise cancellation algorithm, and the time domain waveform characteristics of the effective acoustic emission event are extracted by the short-time energy threshold method. The time difference positioning method is used to invert the propagation path of the acoustic emission event according to the spatial coordinate distribution of the sensor array and calculate the three-dimensional spatial coordinates to locate the position of the abnormal disturbance source; Construct a reverse tracking model of the disturbance propagation chain to map the coupling relationship between air pressure, temperature, spray volume parameters and equipment status, and update the conditional probability table of the node through the maximum likelihood estimation algorithm; Based on the improved Dijkstra algorithm, the parameter conflict propagation path topology network is traversed, the coupling coefficient defined by the partial derivatives between parameters is used as the path weight, and the priority sequence of the parameter conflict sources is output; If it is detected that the characteristic frequency in the vibration signal belongs to the binder blockage frequency band and the sound source coordinates deviate from the threshold distance of the fluidized bed geometric center, the parameter decoupling replanning instruction is triggered, the conflict parameter identification code is sent to the control system, and the independent adjustment mode is activated.

7. The method for automatically controlling secondary granulation of high voltage zinc oxide resistors according to claim 6, characterized in that: When the improved Dijkstra algorithm is initialized, the conflict source node is set as the starting point based on the topological network node set and edge weights output by the parameter coupling graph construction layer; The priority queue is managed through the Fibonacci heap data structure, with the path cumulative weight as the key value. If the current node is the endpoint of a strongly coupled path, the edge weight is updated through the dynamic weight correction coefficient; During the traversal process of the improved Dijkstra algorithm, the coupling coefficient updated by the path weight calculation layer is received in real time. If it is detected that the weight change of any edge exceeds the preset threshold, the current search is interrupted and the edge is reinserted into the priority queue. When the search reaches the target node, the path is traced back through the reverse pointer to output the conflict propagation chain with the minimum cumulative weight. At the same time, if there are multiple equal-weight paths to the same node, the posterior probability optimization mechanism based on the conditional probability table is triggered, and the path with a higher conditional probability value is selected as the optimal solution to generate the conflict source positioning result.

8. The method for automatically controlling secondary granulation of high voltage zinc oxide resistor sheets according to claim 1, characterized in that: The disturbance propagation chain reverse tracing model includes a parameter coupling graph construction layer, a conditional probability table update layer, a path weight calculation layer, a path optimization layer, a conflict source priority sorting layer and a dynamic feedback correction layer; The parameter coupling graph construction layer is used to map the parameters into nodes based on the historical correlation data of air pressure, temperature, spray volume parameters and equipment status through graph theory methods, and map the physical coupling relationship between parameters into directed edges to construct an initial propagation path topology network; the conditional probability table update layer is used to update the node conditional probability table based on the real-time monitored vibration and acoustic emission abnormality feature data using the maximum likelihood estimation algorithm; The path weight calculation layer is used to calculate the coupling coefficient by the partial derivative chain method, normalize the coupling coefficient to the path weight, and generate a coupling coefficient matrix. If the absolute value of the partial derivative of the coupling coefficient is greater than 0.5, it is marked as a strong coupling path; The path optimization layer is used to traverse the propagation path topology network using the coupling coefficient matrix as the adjacency matrix, optimize the search efficiency through the relaxation factor dynamic adjustment strategy, and output the shortest conflict propagation chain; the conflict source priority sorting layer is used to calculate the confidence of the conflict source based on the product of the path cumulative weight and the node conditional probability, generate the conflict source priority sequence through the quick sorting algorithm, and mark the node with a confidence higher than 0.7 as a first-level conflict source; The dynamic feedback correction layer is used to feed back the compensated fluidization index deviation to the conditional probability table, iteratively optimize the coupling coefficient through the gradient descent method, and trigger the dynamic update of the topological network edge weights.

9. The method for automatically controlling secondary granulation of high voltage zinc oxide resistor sheets according to claim 1, characterized in that: The partial least squares regression model extracts the spectral principal component vector through multimodal data standardization and kernel principal component dimensionality reduction, and inputs it into the latent variable space together with the control parameter time series data; then, the latent variable weights are iteratively calculated based on the covariance maximization criterion. If the cross-validation error does not converge, the dynamic latent variable number reduction mechanism is triggered; when the regression residual exceeds the dynamically adjusted adaptive threshold, the regression coefficient matrix and the reinforcement learning weight are subjected to the Kronecker product operation to generate a gradient correction; if the detected gradient direction is consistent with the historical trajectory, the learning rate is increased through the momentum acceleration mechanism to update the network weights.

10. An automatic control system for secondary granulation of high voltage zinc oxide resistors, characterized by: The method for automatically controlling secondary granulation of high-voltage zinc oxide resistor sheets as claimed in any one of claims 1 to 9 comprises: The real-time physical property sensing module is used to analyze the hydroxyl activity and humidity gradient on the material surface based on the dielectric relaxation spectrum decomposition algorithm, reconstruct the three-dimensional pore distribution by combining terahertz wave inverse scattering tomography, and output the dynamic physical property parameter matrix to the dynamic multi-field modeling module. The dynamic physical property parameter matrix includes dielectric loss factor, porosity, and surface charge density gradient; The dynamic multi-field modeling module is used to construct a discrete element fluid-mechanical coupling model. It uses an explicit / implicit hybrid iterative algorithm to simultaneously solve the heat conduction equation, the particle collision energy equation, and the electric field potential energy equation. It predicts the bubble phase / emulsion phase transition threshold and the microscopic defect risk coordinates, generates a fluidization state map, and outputs it to the intelligent collaborative control module. An intelligent collaborative control module uses a multi-head temporal attention mechanism to extract fluidization uniformity features, combines a rolling time domain optimization framework to generate collaborative adjustment instructions for airflow valve opening, infrared power, and atomization frequency, and outputs control instructions to the actuator and abnormal reverse suppression module; The abnormal reverse suppression module is used to identify the characteristic frequency of adhesive blockage based on the wavelet envelope analysis algorithm, locate the source of parameter conflict through the disturbance propagation chain reverse tracking algorithm, dynamically decouple the pressure-temperature reverse coupling effect, output compensation instructions to the actuator, and synchronously feedback abnormal characteristics to the closed-loop feedback optimization module; The closed-loop feedback optimization module is used to associate control parameters with electrical performance indicators through a partial least squares regression model. When grain boundary segregation is detected, the graph convolutional network is triggered to mine implicit control rules, update the reinforcement learning network weights and reward function constraints, and output the optimized parameters to the intelligent collaborative control module.

Images