Multi-component proportioning intelligent proportioning method and system for anti-corrosion powder coating of drinking water pipeline

Abstract

The invention provides a drinking water pipeline anti-corrosion powder coating multi-component proportioning intelligent batching method and system, and relates to the technical field of intelligent manufacturing, and the method comprises the following steps: obtaining drinking water pipeline anti-corrosion powder coating raw material parameters and formula information; constructing a quantitative characteristic volume element numerical simulation unit, and performing multi-scale bridging calculation in combination with microscopic molecular dynamics and a macroscopic discrete element method; inputting parameters, calculating, correcting a formula, and converting into control parameters; dynamically adjusting parameters through a regression network self-learning feedback compensation algorithm; starting mixing equipment for batching, calculating the uniformity in real time, stopping mixing after the uniformity reaches the standard, detecting a finished product, and generating a report. According to the invention, through a self-learning feedback compensation algorithm based on a regression network, in combination with real-time weight detection and a material blockage identification result, intelligent dynamic adjustment of quantitative feeding control parameters is realized.

Description

Technical Field

[0001] The present invention relates to intelligent manufacturing technology, and in particular to a multi-component intelligent proportioning method and system for anti-corrosion powder coatings for drinking water pipelines. Background Art

[0002] The existing anti-corrosion powder coating technology for drinking water pipelines mainly has the following problems:

[0003] Formula design is highly dependent on experience: Traditional formula design mainly relies on empirical formulas and trial and error methods, lacking theoretical guidance, resulting in long formula optimization cycles, low efficiency, and difficulty in meeting the diverse needs of different application scenarios. Insufficient particle behavior prediction mechanism: Existing technologies lack understanding of the relationship between the microscopic forces and macroscopic flow behavior of powder particles, and are unable to accurately predict the physical and chemical behavior of different component particles during the mixing process, resulting in a large deviation between formula design and actual performance. The interaction effects of multiple components are difficult to quantify: Anti-corrosion powder coatings are usually composed of multiple components such as resins, curing agents, pigments and fillers, and functional additives. There are complex physical and chemical interactions between the components, and existing technologies find it difficult to accurately quantify these interaction effects.

[0004] Insufficient dosing accuracy: Traditional dosing equipment has limited control accuracy and is easily affected by material properties and environmental factors, resulting in deviations between the actual and target ratios, affecting product quality and stability. Inability to monitor mixing uniformity in real time: Lacking effective real-time monitoring of the mixing process, the mixing endpoint is primarily determined by empirical judgment, which can easily lead to insufficient or excessive mixing, impacting product performance and production efficiency.

[0005] Low level of intelligent batching: The existing batching process is not highly automated and lacks a data-driven intelligent decision-making mechanism, making it difficult to cope with raw material changes and environmental disturbances, resulting in insufficient production adaptability and stability. Lack of multi-scale design methods: The lack of multi-scale design methods from the molecular scale to macroscopic performance has led to an "experience gap" between formulation design and actual performance, making it difficult to achieve precise formulation design. Summary of the Invention

[0006] The embodiments of the present invention provide a multi-component intelligent proportioning method and system for anti-corrosion powder coatings for drinking water pipelines, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provides a multi-component intelligent dosing method for anti-corrosion powder coatings for drinking water pipelines, including:

[0009] Obtain raw material parameters and target formula information for anti-corrosion powder coatings for drinking water pipelines;

[0010] A numerical simulation unit based on quantitative characteristic volume elements is constructed. The microscopic force calculation module calculates the interaction forces between powder particles using a simplified molecular dynamics method and forms a parameterized database. The macroscopic flow prediction module uses an improved discrete element method to predict the flow and mixing behavior of powder particle groups within the target sampling area. The improved discrete element method includes an adaptive time step algorithm and a domain decomposition parallel algorithm. The microscopic force calculation module and the macroscopic flow prediction module perform data exchange and collaborative calculations based on a multi-scale bridging calculation method in an adaptive coupling region.

[0011] Inputting the raw material parameters, target formula information, and production environment parameters into the numerical simulation unit for calculation, correcting the weight ratio of each component in the target formula information based on the calculation results, and converting the corrected weight ratio into a quantitative feeding control parameter; utilizing the real-time weight detection and material blockage identification results, dynamically adjusting the quantitative feeding control parameter through a self-learning feedback compensation algorithm based on a regression network;

[0012] Based on the dynamically adjusted quantitative feeding control parameters, the mixing equipment is started to batch the ingredients, the mixing process data is collected and input into the numerical simulation unit to calculate the mixing uniformity index in real time, the mixing is automatically stopped when the target requirements are met, the mixed finished product is tested, the target formula information is compared and a batching process report is generated.

[0013] In an optional embodiment,

[0014] The construction of the quantitative characteristic volume element includes:

[0015] Measuring a powder particle sample to obtain size distribution data, morphological characteristic data, and spatial arrangement data of the powder particles; determining a characteristic radius of the powder particles based on the size distribution data, calculating a shape factor based on the morphological characteristic data, and calculating an interparticle void ratio based on the spatial arrangement data;

[0016] Calculate the volume of a quantitative eigenvalue:

[0017] ;

[0018] Among them, V QCVEV is the volume of the quantitative characteristic volume element, R i is the characteristic radius of the i-th type powder particles, f s is the shape factor, δ is the interparticle void ratio, and N is the number of types of powder particles;

[0019] Determining a minimum size threshold of the quantitative characteristic volume element so that the physical properties of the quantitative characteristic volume element do not change with the increase of volume when the quantitative characteristic volume element is above the minimum size threshold;

[0020] Based on the volume and minimum size threshold of the quantitative characteristic volume element, a three-dimensional digital representation of the quantitative characteristic volume element is generated and assigned physical property parameters. Boundary constraints and initial conditions are imposed on the three-dimensional digital representation so that the quantitative characteristic volume element can represent the structural characteristics and mechanical behavior of the overall powder system.

[0021] In an optional embodiment,

[0022] The microscopic force calculation module calculates the interaction forces between powder particles based on a simplified molecular dynamics method and forms a parameterized database. The macroscopic flow prediction module uses an improved discrete element method to predict the flow and mixing behavior of powder particle groups within the target sampling area, including:

[0023] The microscopic force calculation module uses a simplified molecular dynamics method to calculate the interaction force between powder particles. The simplified molecular dynamics method represents multiple molecular collections as a single interaction point and calculates the interparticle force through the spatial derivative of the effective potential energy function. The microscopic force calculation module calculates the van der Waals force, electrostatic force, and capillary force between powder particles. The van der Waals force is corrected using the Hamaker constant, the electrostatic force uses the modified Coulomb's law, and the capillary force takes into account the influence of ambient humidity.

[0024] The microscopic force calculation results are stored to form a parameterized database, which uses a multi-dimensional interpolation algorithm to achieve fast calculation under different process conditions;

[0025] The macro-flow prediction module simulates powder particle populations using an improved discrete element method (DEM), which includes an adaptive time step algorithm and a domain decomposition parallel algorithm. The adaptive time step adjusts the calculation step size based on the dynamic characteristics of the particles, and the domain decomposition parallel algorithm decomposes the calculation domain into multiple sub-regions.

[0026] Based on the force data in the parameterized database, an improved Verlet algorithm is applied to solve the powder particle motion equations, wherein the improved Verlet algorithm introduces velocity half-step iteration and position constraint correction to reduce numerical errors, and the motion equations include the particle linear motion equations and angular motion equations;

[0027] Based on the results of solving the motion equation, the macro flow prediction module calculates the flow velocity field and mixing uniformity in the target area, where the flow velocity field is calculated by particle mass, velocity and weight function, and the mixing uniformity is evaluated using the Lacey mixing index.

[0028] In an optional embodiment,

[0029] The microscopic force calculation module and the macroscopic flow prediction module perform data interaction and collaborative calculations based on the multi-scale bridging calculation method of the adaptive coupling region, including:

[0030] The microscopic force calculation results are transferred to the macroscopic flow prediction module through the parameter mapping method, where the microscopic force calculation results are converted into the forces used in the macroscopic calculation through the combination of mapping operators and transfer parameter sets;

[0031] An adaptive coupling region is set at the interface between the microscopic force calculation module and the macroscopic flow prediction module, wherein the adaptive coupling region includes all points that are no more than a preset critical distance from the interface;

[0032] The time scale difference between micro-scale and macro-scale calculations is reconciled through a multi-time-step approach, where the product of the number of micro-scale sub-steps and the micro-scale time step equals the macro-scale time step;

[0033] A spatial filtering method is used to smooth the spatial scale difference between the microscale and the macroscale, where the calculated physical quantity at the macroscale is obtained by integrating the calculated physical quantity at the microscale through the spatial filtering kernel function;

[0034] The collaborative calculation is performed in an iterative coupling manner, including: initializing a microscopic force calculation module and a macroscopic flow prediction module; calculating the interaction force between particles at the microscopic scale; mapping the calculated microscopic force to the macroscopic scale through the parameter mapping method; calculating the movement of powder particle groups at the macroscopic scale; feeding back the macroscopic calculation results to the microscopic calculation module to update the boundary conditions; calculating the relative error between two adjacent iterative results, and outputting the result if it is less than a preset error threshold, otherwise returning to perform the microscopic inter-particle interaction force calculation.

[0035] In an optional embodiment,

[0036] Inputting the raw material parameters, target formula information, and production environment parameters into the numerical simulation unit for calculation, correcting the weight ratio of each component in the target formula information according to the calculation results, and converting the corrected weight ratio into quantitative feeding control parameters includes:

[0037] Constructing a multidimensional dataset including the raw material parameters, target recipe information, and production environment parameters, inputting the multidimensional dataset into a multi-physics field coupled numerical simulation unit, establishing an objective function through the numerical simulation unit, the objective function including a relative deviation term between a performance indicator and a target value, a performance indicator weight coefficient, and a recipe cost factor; and obtaining an optimized ratio vector by minimizing the objective function;

[0038] An interaction effect correction coefficient matrix is ​​established, wherein the interaction effect correction coefficient matrix is ​​a symmetric matrix, and the matrix elements represent the interaction influence coefficients between the components; the matrix elements are determined by iterative fitting through the simulation results of the numerical simulation unit; the optimized ratio vector is corrected using the interaction effect correction coefficient matrix to obtain a corrected ratio vector; and the quantitative feeding control parameters are calculated based on the corrected ratio vector, the total feeding amount, the feeding time, and the feeding device calibration coefficient.

[0039] In an optional embodiment,

[0040] Using real-time weight detection and material blockage identification results, the self-learning feedback compensation algorithm based on the regression network dynamically adjusts the quantitative feeding control parameters including:

[0041] Obtain the real-time weight value and target weight value of the material; extract the silo pressure signal, material temperature signal and vibration signal collected during the material transportation process through the feature mapping function and perform weighted fusion to obtain the blockage status index;

[0042] A deep spatiotemporal regression network with a causal inference enhancement layer is constructed. The network takes the time series deviation between the real-time material weight and the target weight, as well as the congestion status indicator, as input features. The network captures dynamic characteristics through parallel temporal convolution branches and spatial relationship branches, and outputs control parameter compensation. The causal inference enhancement layer identifies the causal relationship structure between control factors and outcomes through structured intervention simulation and counterfactual reasoning mechanisms, and employs a bidirectional attention mechanism to adaptively fuse physical prior knowledge with data-driven features.

[0043] Construct a hierarchical multi-objective loss function of the main control objectives, process objectives, and constraints, dynamically optimize the control parameters based on the meta-learning framework, and dynamically adjust the update direction and step size of the parameters according to the control error characteristics and congestion risk;

[0044] The feedforward compensation control quantity is calculated based on the prediction results of the deep spatiotemporal regression network. Combined with the feedback correction term generated online, the total control quantity of the feeding system is generated through an adaptive fusion algorithm. The adaptive fusion algorithm can dynamically adjust the weight ratio of feedforward compensation and feedback correction according to the material's cumulative weight error, error change rate and blockage risk index; the total control quantity is executed to dynamically adjust the quantitative feeding control parameters, and the control effect is stored in the experience replay pool. The deep spatiotemporal regression network is continuously optimized through knowledge distillation and incremental learning.

[0045] In an optional embodiment,

[0046] Construct a hierarchical multi-objective loss function of the main control objectives, process objectives, and constraints, dynamically optimize the control parameters based on the meta-learning framework, and dynamically adjust the update direction and step size of the parameters according to the control error characteristics and congestion risk.

[0047] Constructing a hierarchical multi-objective loss function, the hierarchical multi-objective loss function includes a main control target layer loss function, a process target layer loss function and a constraint condition layer loss function, wherein the main control target layer loss function is used to characterize product ratio accuracy and mixing uniformity, the process target layer loss function is used to characterize weight control accuracy, anti-blocking capability and energy consumption optimization goals, and the constraint condition layer loss function is used to characterize equipment safety constraints and material property constraints. The main control target layer loss function, the process target layer loss function and the constraint condition layer loss function are combined to obtain a comprehensive loss function;

[0048] A meta-learning framework is used to optimize control parameters. In an inner loop, the control parameters are updated based on the gradient of the comprehensive loss function. In an outer loop, the meta-learner parameters are optimized, the sensitivity of each control parameter to the comprehensive loss function is calculated, and the learning rate of each control parameter is dynamically adjusted based on the sensitivity, weight control error, and congestion risk index.

[0049] The degree of conflict between objectives is detected and the optimization priority is determined. Based on the optimization priority, the weight of each objective in the comprehensive loss function is progressively adjusted. The update direction and step size of different parameters are dynamically adjusted according to the control error characteristics and congestion risk.

[0050] According to a second aspect of the embodiments of the present invention,

[0051] Provides a multi-component intelligent dosing system for anti-corrosion powder coatings for drinking water pipelines, including:

[0052] The first unit is used to obtain the raw material parameters and target formula information of the anti-corrosion powder coating for drinking water pipelines;

[0053] The second unit is used to construct a numerical simulation unit based on quantitative characteristic volume elements. The microscopic force calculation module calculates the interaction force between powder particles based on a simplified molecular dynamics method and forms a parameterized database. The macroscopic flow prediction module uses an improved discrete element method to predict the flow and mixing behavior of the powder particle group in the target sampling area. The improved discrete element method includes an adaptive time step algorithm and a domain decomposition parallel algorithm. The microscopic force calculation module and the macroscopic flow prediction module perform data interaction and collaborative calculation based on a multi-scale bridging calculation method of an adaptive coupling region.

[0054] a third unit configured to input the raw material parameters, target formula information, and production environment parameters into the numerical simulation unit for calculation, correct the weight ratio of each component in the target formula information based on the calculation results, and convert the corrected weight ratio into a quantitative feeding control parameter; and dynamically adjust the quantitative feeding control parameter through a self-learning feedback compensation algorithm based on a regression network using the results of real-time weight detection and material blockage identification;

[0055] The fourth unit is used to start the mixing equipment to batch the ingredients based on the dynamically adjusted quantitative feeding control parameters, collect the mixing process data and input it into the numerical simulation unit to calculate the mixing uniformity index in real time, automatically stop mixing when the target requirements are met, detect the mixed finished product, compare the target formula information and generate a batching process report.

[0056] According to a third aspect of the embodiments of the present invention,

[0057] An electronic device is provided, comprising:

[0058] processor;

[0059] a memory for storing processor-executable instructions;

[0060] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0061] According to a fourth aspect of the embodiments of the present invention,

[0062] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0063] The present invention realizes the scientific connection between the microscopic forces of powder particles and the macroscopic flow behavior by constructing a numerical simulation unit based on quantitative characteristic volume elements, transforming the formula design from experience-oriented to theory-driven, significantly improving the scientific nature and accuracy of the formula design; innovatively linking the microscopic force calculation module and the macroscopic flow prediction module through a multi-scale bridging method of adaptive coupling regions, realizing full-chain prediction from molecular scale to macroscopic performance, effectively filling the "scale gap" in formula design; based on the simplified molecular dynamics method, the interaction force between powder particles is calculated and a parameterized database is formed, combined with the improved discrete element method to efficiently predict the flow and mixing behavior of powder particle groups, greatly reducing the computational complexity and time cost while ensuring the calculation accuracy.

[0064] The present invention realizes intelligent dynamic adjustment of quantitative feeding control parameters through a self-learning feedback compensation algorithm based on a regression network, combined with real-time weight detection and material blockage identification results, significantly improving feeding accuracy and stability and reducing the need for human intervention; the present invention replaces part of the physical experiment with numerical simulation, greatly reducing trial and error costs and material waste; through precise batching, it reduces excessive use of raw materials and improves resource utilization efficiency; intelligent control reduces manpower requirements and operational errors, and comprehensively reduces R&D and production costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1This is a flow chart of a multi-component intelligent proportioning method for anti-corrosion powder coatings for drinking water pipelines according to an embodiment of the present invention;

[0066] Figure 2 This is a comparison chart of the convergence of the multi-scale bridging calculation method and the traditional single-scale calculation method in the prediction of mixing uniformity;

[0067] Figure 3 A simulation diagram of the multi-scale bridging calculation of the present invention;

[0068] Figure 4 It is a comparison chart of the real-time weight accumulation curve of material feeding;

[0069] Figure 5 This is a comparison chart of congestion state prediction and control parameter adjustment. DETAILED DESCRIPTION

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. 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 shall fall within the scope of protection of the present invention.

[0071] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0072] Figure 1 FIG. 1 is a flow chart of a multi-component intelligent batching method for anti-corrosion powder coatings for drinking water pipelines according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0073] Obtain raw material parameters and target formula information for anti-corrosion powder coatings for drinking water pipelines;

[0074] A numerical simulation unit based on quantitative characteristic volume elements is constructed. The microscopic force calculation module calculates the interaction forces between powder particles using a simplified molecular dynamics method and forms a parameterized database. The macroscopic flow prediction module uses an improved discrete element method to predict the flow and mixing behavior of powder particle groups within the target sampling area. The improved discrete element method includes an adaptive time step algorithm and a domain decomposition parallel algorithm. The microscopic force calculation module and the macroscopic flow prediction module perform data exchange and collaborative calculations based on a multi-scale bridging calculation method in an adaptive coupling region.

[0075] Inputting the raw material parameters, target formula information, and production environment parameters into the numerical simulation unit for calculation, correcting the weight ratio of each component in the target formula information based on the calculation results, and converting the corrected weight ratio into a quantitative feeding control parameter; utilizing the real-time weight detection and material blockage identification results, dynamically adjusting the quantitative feeding control parameter through a self-learning feedback compensation algorithm based on a regression network;

[0076] Based on the dynamically adjusted quantitative feeding control parameters, the mixing equipment is started to batch the ingredients, the mixing process data is collected and input into the numerical simulation unit to calculate the mixing uniformity index in real time, the mixing is automatically stopped when the target requirements are met, the mixed finished product is tested, the target formula information is compared and a batching process report is generated.

[0077] In an optional embodiment,

[0078] The construction of the quantitative characteristic volume element includes:

[0079] Measuring a powder particle sample to obtain size distribution data, morphological characteristic data, and spatial arrangement data of the powder particles; determining a characteristic radius of the powder particles based on the size distribution data, calculating a shape factor based on the morphological characteristic data, and calculating an interparticle void ratio based on the spatial arrangement data;

[0080] Calculate the volume of a quantitative eigenvalue:

[0081] ;

[0082] Among them, V QCVEV is the volume of the quantitative characteristic volume element, R i is the characteristic radius of the i-th type powder particles, f s is the shape factor, δ is the interparticle void ratio, and N is the number of types of powder particles;

[0083] Determining a minimum size threshold of the quantitative characteristic volume element so that the physical properties of the quantitative characteristic volume element do not change with the increase of volume when the quantitative characteristic volume element is above the minimum size threshold;

[0084] Based on the volume and minimum size threshold of the quantitative characteristic volume element, a three-dimensional digital representation of the quantitative characteristic volume element is generated and assigned physical property parameters. Boundary constraints and initial conditions are imposed on the three-dimensional digital representation so that the quantitative characteristic volume element can represent the structural characteristics and mechanical behavior of the overall powder system.

[0085] For example, when measuring powder particle samples, a laser diffraction particle size analyzer is used to obtain particle size distribution data. For epoxy resin powder, the measurement results show a particle size range of 20-120 microns, with a D50 value of 65 microns. A scanning electron microscope is used to observe and photograph the particle morphology, and image processing software is used to analyze and obtain morphological characteristics, including parameters such as roundness, roughness, and aspect ratio. Mercury intrusion is used to measure the spatial arrangement of the powder particles after packing, determining the particle packing density and pore distribution.

[0086] When determining the characteristic radius of powder particles based on size distribution data, the particles are divided into 3-5 categories by size, with each category accounting for a certain proportion. Taking epoxy resin powder as an example, these can be divided into three categories: small particles (20-50 microns, accounting for 30%), medium particles (50-90 microns, accounting for 55%), and large particles (90-120 microns, accounting for 15%). The characteristic radii of each category are 35 microns, 70 microns, and 105 microns, respectively. When calculating the shape factor based on morphological feature data, the particle's roundness, surface roughness, and aspect ratio are taken into consideration. The value range is generally between 0.8 and 1.2, where 1.0 represents a perfect sphere and values ​​greater than 1.0 indicate increased shape complexity. For epoxy resin powder, the calculated shape factor value is 1.15. When calculating the interparticle void ratio based on spatial arrangement data, the void ratio is obtained through porosity measurement. Typical interparticle void ratios in powder coatings range from 0.3 to 0.5. The measured void ratio of epoxy resin powder is 0.38.

[0087] When calculating the volume of a quantitative characteristic volume element, the volume contributions of all types of particles need to be accumulated while considering the effects of shape factor and void fraction.

[0088] To determine the minimum size threshold for a quantitative characteristic volume element, a statistically representative volume element (RVE) method is used, which is determined by gradually increasing the sampling volume and measuring changes in physical properties. Specifically, volume elements of different sizes (ranging from 50×50×50 microns to 500×500×500 microns) are selected, and key physical properties such as thermal conductivity, fluidity, and mechanical strength are calculated. When the volume increases to a certain value and the rate of change of these properties is less than 2%, the volume is considered to have reached the minimum size threshold. For epoxy resin powder, tests have shown that when the side length of the volume element reaches 350 microns, the rate of change of physical properties is less than 2%, so 350 microns is determined as the minimum size threshold.

[0089] When generating a three-dimensional digital representation of the quantitative characteristic volume element, voxel modeling technology is used to discretize the quantitative characteristic volume element into 100×100×100 voxel units. Each voxel is assigned corresponding physical property parameters according to its location, including density, elastic modulus, friction coefficient, thermal conductivity, and resistivity. The density of each voxel of epoxy resin powder is set to 1.2-1.3 g / cm³, elastic modulus is 2.5-3.5 GPa, friction coefficient is 0.35-0.45, thermal conductivity is 0.15-0.25 W / (m·K), and surface resistivity is 10 13 -10 15 When applying boundary constraints and initial conditions to the three-dimensional digital representation, periodic boundary conditions are set in the X, Y, and Z directions. The initial conditions include a temperature of 298 K, a pressure of 101.3 kPa, and no external force.

[0090] To verify the representativeness of the QVE, comparative tests were conducted. The constructed QVE was tested for flowability and thermal conductivity under the same conditions as a macroscopic powder system (100 times the volume of the QVE). The results showed that the flow angle of the QVE deviated from that of the macroscopic system by less than 3%, and the thermal conductivity measurement deviated by less than 4%, demonstrating that the constructed QVE effectively represents the structural characteristics and mechanical behavior of the entire powder system.

[0091] Existing powder system analysis technologies mainly use two methods: one is to simplify powder particles into ideal spheres, ignoring the influence of particle morphology and spatial arrangement, resulting in large deviations from actual system behavior; the other is to use high-precision direct numerical simulation, which has high computational complexity and is difficult to apply to actual industrial production environments. The quantitative characteristic volume element construction method proposed in this application comprehensively considers the size distribution, morphological characteristics, and spatial arrangement characteristics of powder particles, overcoming the limitations of the ideal spherical assumption; at the same time, by determining the minimum size threshold, the volume element is made as small as possible while ensuring representativeness, effectively reducing computational complexity. This method has wide applicability to different types of powders and provides an effective tool for powder coating formulation design and production process optimization.

[0092] In an optional embodiment,

[0093] The microscopic force calculation module calculates the interaction forces between powder particles based on a simplified molecular dynamics method and forms a parameterized database. The macroscopic flow prediction module uses an improved discrete element method to predict the flow and mixing behavior of powder particle groups within the target sampling area, including:

[0094] The microscopic force calculation module uses a simplified molecular dynamics method to calculate the interaction force between powder particles. The simplified molecular dynamics method represents multiple molecular collections as a single interaction point and calculates the interparticle force through the spatial derivative of the effective potential energy function. The microscopic force calculation module calculates the van der Waals force, electrostatic force, and capillary force between powder particles. The van der Waals force is corrected using the Hamaker constant, the electrostatic force uses the modified Coulomb's law, and the capillary force takes into account the influence of ambient humidity.

[0095] The microscopic force calculation results are stored to form a parameterized database, which uses a multi-dimensional interpolation algorithm to achieve fast calculation under different process conditions;

[0096] The macro-flow prediction module simulates powder particle populations using an improved discrete element method (DEM), which includes an adaptive time step algorithm and a domain decomposition parallel algorithm. The adaptive time step adjusts the calculation step size based on the dynamic characteristics of the particles, and the domain decomposition parallel algorithm decomposes the calculation domain into multiple sub-regions.

[0097] Based on the force data in the parameterized database, an improved Verlet algorithm is applied to solve the powder particle motion equations, wherein the improved Verlet algorithm introduces velocity half-step iteration and position constraint correction to reduce numerical errors, and the motion equations include the particle linear motion equations and angular motion equations;

[0098] Based on the results of solving the motion equation, the macro flow prediction module calculates the flow velocity field and mixing uniformity in the target area, where the flow velocity field is calculated by particle mass, velocity and weight function, and the mixing uniformity is evaluated using the Lacey mixing index.

[0099] For example, the simplified molecular dynamics method greatly simplifies the computational complexity by representing multiple molecular assemblies as a single interaction point. For example, a powder particle with a diameter of 65 microns contains about 10 12 Using simplified molecular dynamics, the particle is simplified into a single interaction point with equivalent physical properties, reducing computational complexity by approximately 10 orders of magnitude. The microscopic force calculation module calculates the interparticle forces using the spatial derivative of the effective potential energy function, specifically including the three main forces: van der Waals, electrostatic, and capillary forces.

[0100] The van der Waals force calculation uses the Hamaker constant correction method, which adjusts the Hamaker constant according to the particle material and size. Taking epoxy resin powder as an example, its Hamaker constant value is 6.5×10 -20Joules. When the distance between the two particle surfaces decreases from 100 nanometers to 10 nanometers, the van der Waals force increases from 0.12 nanonewtons to 12.5 nanonewtons. The electrostatic force calculation uses the modified Coulomb law, taking into account the dielectric constant, particle charge, and shielding effect. When the relative humidity is 45%, the particle charge of epoxy resin powder is about 3.5×10 -14 Coulomb, the electrostatic force between two particles with a distance of 50 microns is approximately 0.85 nanonewtons. The capillary force calculation takes into account the influence of ambient humidity. Using the liquid bridge model, it is calculated that when the relative humidity is 60%, the capillary force generated by the liquid bridge between the two particles is approximately 15 nanonewtons, which is much greater than the contributions of van der Waals and electrostatic forces.

[0101] The results of microscopic force calculations are stored in a parameterized database, which uses a five-dimensional structure to store data, including five key parameters: particle type, particle size, particle spacing, ambient temperature, and ambient humidity. The database records the inter-particle force values ​​under different parameter combinations. The parameterized database uses a multidimensional interpolation algorithm to achieve fast calculations under different process conditions. The particle type is mapped by classification, and the particle size, particle spacing, ambient temperature, and humidity are interpolated by cubic spline. A quadtree index structure is used to optimize query efficiency. The average relative error of the interpolation algorithm is less than 3%, and the query response time is less than 0.5 milliseconds, which is about 200 times faster than direct calculation.

[0102] The macro-flow prediction module uses an improved discrete element method (DEM) to simulate powder particle populations. This DEM includes two key technologies: an adaptive time-step algorithm and a domain decomposition parallel algorithm. The adaptive time-step algorithm dynamically adjusts the computational step size based on the dynamic characteristics of the particles. Specifically, it determines the optimal time step size by monitoring the particle's maximum velocity, acceleration, and the rate of change of the contact force. When the system is in a stable state, the time step can reach 10 microseconds; during the violent collision phase, the time step is automatically reduced to 0.1 microsecond to ensure numerical stability. The domain decomposition parallel algorithm decomposes the entire computational domain into multiple sub-regions. Using a meshing method, the 500×500×500 micron computational domain is divided into 5×5×5 sub-regions. Each sub-region independently calculates the motion state of the particles within it, and particles at the region boundaries exchange information via a message passing interface (MPI). When using 16-core parallel computing, computational efficiency is increased by approximately 13 times.

[0103] Based on the force data in the parameterized database, an improved Verlet algorithm is applied to solve the equations of motion for the powder particles. The improved Verlet algorithm introduces velocity half-step iteration and position constraint correction mechanisms to reduce numerical errors. Velocity half-step iteration significantly improves computational stability by updating the velocity at the midpoint of the entire step length; position constraint correction uses the SHAKE algorithm to address rigid constraints and prevent non-physical penetration of particles. The equations of motion include the linear and angular equations of motion for the particles. The linear equations of motion calculate the motion state of the particle's center of mass, while the angular equations of motion calculate the particle's rotational motion about its axis. Taking epoxy resin powder in a mixer as an example, at a rotational speed of 60 rpm, the average linear velocity of the powder particles is approximately 0.12 m / s, and the angular velocity is approximately 18 rad / s.

[0104] Based on the results of the equation of motion, the macroscopic flow prediction module calculates the flow velocity field and mixing uniformity within the target area. The flow velocity field is calculated using particle mass, velocity, and a weighting function. A Gaussian kernel function is used as the weighting function, and the kernel width parameter is set to 1.5 times the particle diameter. The velocity field is represented by a 10×10×10 grid with a spatial resolution of 50 microns. Mixing uniformity is assessed using the Lacey mixing index, which is determined by calculating the ratio of the component concentration variance to the variance of the ideal random mixing state. When the Lacey mixing index reaches 0.85 or above, it is considered fully mixed. During the mixing process of epoxy resin powder and curing agent in a V-type mixer, after 180 seconds of mixing, the Lacey mixing index increased from an initial 0.12 to 0.87, indicating a fully mixed state.

[0105] This method, by combining microscopic force calculations with macroscopic flow predictions, achieves multi-scale predictions of the mixing process for anti-corrosion powder coatings for drinking water pipelines. It accurately simulates the physical and chemical behavior of different components during mixing, providing a scientific basis for subsequent formulation optimization and process parameter adjustment. Compared to traditional full molecular dynamics simulations, the computational efficiency is improved by over five orders of magnitude; compared to simplified spherical particle models, the prediction accuracy is improved by approximately 40%, showing significant advantages in the case of irregular particles and high humidity environments.

[0106] In an optional embodiment,

[0107] The microscopic force calculation module and the macroscopic flow prediction module perform data interaction and collaborative calculations based on the multi-scale bridging calculation method of the adaptive coupling region, including:

[0108] The microscopic force calculation results are transferred to the macroscopic flow prediction module through the parameter mapping method, where the microscopic force calculation results are converted into the forces used in the macroscopic calculation through the combination of mapping operators and transfer parameter sets;

[0109] An adaptive coupling region is set at the interface between the microscopic force calculation module and the macroscopic flow prediction module, wherein the adaptive coupling region includes all points that are no more than a preset critical distance from the interface;

[0110] The time scale difference between micro-scale and macro-scale calculations is reconciled through a multi-time-step approach, where the product of the number of micro-scale sub-steps and the micro-scale time step equals the macro-scale time step;

[0111] A spatial filtering method is used to smooth the spatial scale difference between the microscale and the macroscale, where the calculated physical quantity at the macroscale is obtained by integrating the calculated physical quantity at the microscale through the spatial filtering kernel function;

[0112] The collaborative calculation is performed in an iterative coupling manner, including: initializing a microscopic force calculation module and a macroscopic flow prediction module; calculating the interaction force between particles at the microscopic scale; mapping the calculated microscopic force to the macroscopic scale through the parameter mapping method; calculating the movement of powder particle groups at the macroscopic scale; feeding back the macroscopic calculation results to the microscopic calculation module to update the boundary conditions; calculating the relative error between two adjacent iterative results, and outputting the result if it is less than a preset error threshold, otherwise returning to perform the microscopic inter-particle interaction force calculation.

[0113] Exemplarily, when the microscopic force calculation results are transferred hierarchically to the macroscopic flow prediction module via a parameter mapping method, a hybrid strategy combining physics-based mapping and data-driven mapping is employed. Physics-based mapping establishes a mapping relationship based on the physical nature of interparticle forces, while data-driven mapping trains a mapping model based on simulation and experimental data. In specific implementation, the microscopic force calculation results are converted into the forces used in the macroscopic calculation through a combination of a mapping operator and a transfer parameter set. The mapping operators include a physics-preserving operator that ensures force and energy conservation, and a statistical equivalence operator that ensures consistency in the statistical characteristics of the microscopic and macroscopic calculations. The transfer parameter set includes 12 key parameters, such as the values ​​of the force between epoxy resin powder and curing agent particles at different distances. When the microscopic calculation results in a van der Waals force of 8.5 nanonewtons, an electrostatic force of 2.1 nanonewtons, and a capillary force of 18.3 nanonewtons, the equivalent contact force in the macroscopic calculation is 28.9 nanonewtons after parameter mapping, achieving a mapping accuracy exceeding 95%.

[0114] When setting an adaptive coupling region at the interface between the microscopic force calculation module and the macroscopic flow prediction module, it is necessary to determine the boundaries and range of the adaptive coupling region. The adaptive coupling region includes all points that are no more than a preset critical distance from the interface, and the critical distance is dynamically adjusted according to the particle size and physical properties. Taking epoxy resin powder as an example, when the particle diameter is 65 microns, the initial critical distance is set to 3 times the particle diameter (195 microns), and then automatically adjusted according to the force gradient between the particles. When the force gradient exceeds the preset threshold (15 nanonewtons / micron), the critical distance is automatically expanded to 4 times the particle diameter (260 microns) to ensure calculation stability. 100 sampling points are set in the coupling region to form an information exchange network at the micro-macro interface.

[0115] When reconciling the timescale differences between microscale and macroscale calculations using a multi-time-step approach, a timescale bridging mechanism is established. The timestep for microscale calculations is typically much smaller than that for macroscale calculations. For the epoxy resin powder system, the timestep for microscale calculations is set to 0.1 microseconds, while the timestep for macroscale calculations is set to 10 microseconds. To reconcile this difference, a sub-loop strategy is employed, whereby multiple microscale timesteps are executed within each macroscale timestep. Specifically, the product of the number of microscale substeps and the microscale timestep equals the macroscale timestep. For this setting, 100 microscale timesteps are executed within each macroscale timestep. To improve computational efficiency, microscale calculations are performed only within the coupled region, and a time extrapolation technique is used to predict intermediate states.

[0116] Spatial filtering methods are used to smooth the spatial scale difference between the microscale and macroscale, addressing the issue of information transfer between different spatial resolutions. The spatial resolution of microscale calculations is typically in the nanometer range, while that of macroscale calculations is in the micrometer range. Taking epoxy resin powder as an example, the characteristic size of the molecular clusters in the microscale model is 10 nanometers, while the characteristic size of the particles in the macroscale model is 65 microns, resulting in a 6,500-fold difference in spatial scale. To smooth this difference, the physical quantities calculated at the macroscale are integrated over the physical quantities calculated at the microscale using a spatial filtering kernel function. The spatial filtering kernel uses a modified Gaussian kernel with a kernel width parameter set to 50 microns and a smoothing radius set to 150 microns. Different weights are assigned to different force components during the filtering process, with long-range forces (such as electrostatic forces) receiving a weight of 0.8 and short-range forces (such as van der Waals forces) receiving a weight of 0.4, comprehensively considering the distance characteristics of the forces.

[0117] The specific process for executing collaborative calculations using an iterative coupling approach involves multiple steps. Initialization begins by setting the initial parameters for the microscopic force calculation module and the macroscopic flow prediction module. Microscopic module initialization includes setting molecular dynamics force field parameters, thermodynamic parameters (temperature 298 K, pressure 101.3 kPa), and potential energy function parameters. Macroscopic module initialization includes setting particle properties (epoxy resin powder density 1.25 g / cm³, friction coefficient 0.38) and boundary conditions (mixing vessel dimensions 350 × 350 × 500 mm). After initialization, interparticle interaction forces are calculated at the microscale, including van der Waals, electrostatic, and capillary forces for each pair of particles within the coupling region. These calculated microscopic forces are then mapped to the macroscale using a parameter mapping method to generate the interparticle contact and non-contact forces required for macroscopic calculations. The powder particle group motion is then calculated at the macroscale, updating the particle positions, velocities, and force states. After the calculations are complete, the macroscale results are fed back to the microscopic calculation module to update the boundary conditions, primarily including information such as the new particle positions, local density, and velocity gradients. Finally, the relative error between two consecutive iterations is calculated, with a preset error threshold of 5%. If the error is less than this threshold, the result is output and the calculation for the current time step is completed. Otherwise, the calculation of the microscale interparticle interaction force is returned to the beginning of a new iteration. Testing has shown that this iterative scheme converges to below the preset error threshold in an average of 3-5 iterations.

[0118] like Figure 2 The convergence comparison results of the multi-scale bridging calculation method and the traditional single-scale calculation method in the prediction of mixing uniformity are shown as follows. The horizontal axis represents the number of iterations (from 0 to 1250 steps), and the vertical axis represents the relative error in logarithmic scale (from 10 0 to 10 -5 The solid line with black square marks in the figure represents the convergence curve of the multi-scale bridging calculation method, and the dotted line with hollow circle marks represents the convergence curve of the traditional single-scale calculation method. It can be clearly seen from the figure that the multi-scale bridging calculation method shows a faster convergence speed in the early stage of iteration, and the curve drops more steeply. It only takes about 420 iterations to reach 10 -5 In contrast, the traditional single-scale calculation method converges significantly slower, and even after 1250 iterations, its relative error remains at 10 -3 This comparison intuitively demonstrates the significant advantages of the multi-scale bridging method in terms of computational efficiency and convergence performance. Not only does it reduce the number of iterations by approximately 66.4%, but it also achieves higher computational accuracy.

[0119] Figure 3A simulation diagram of a multiscale bridging calculation is shown. The diagram is divided into three regions: the microscopic calculation region on the left, the adaptive coupling region in the middle, and the macroscopic prediction region on the right. The left region, composed of densely packed small particles, represents microscopic molecular dynamics calculations; the right region, composed of sparsely packed large particles, represents macroscopic discrete element calculations. The coupling region in the middle (marked by a white box) enables a smooth transition between the two calculation methods, with a gradient of particle size and density. Particles of varying grayscale represent different velocity states, and the horizontal arrow indicates the overall flow direction. This partitioned calculation strategy combines the accuracy of microscopic calculations with the efficiency of macroscopic calculations, enabling information transfer between different scales through adaptive coupling regions, effectively improving the simulation accuracy and computational efficiency of the powder coating mixing process.

[0120] There are two main types of existing multi-scale calculation methods: one uses a serial coupling strategy, which cannot capture the dynamic interaction between the micro and macro scales; the other uses a fixed region coupling strategy, performing microscopic fine calculations in a predetermined region. The coupling region is fixed and cannot adapt to the dynamically changing physical field. The main improvements of the adaptive coupling region multi-scale bridging method proposed in this invention over the existing technology are: introducing the concept of adaptive coupling region, dynamically adjusting the position and coupling strength of the micro-macro interface according to the physical field gradient; developing a hybrid mapping strategy based on physical preservation and statistical equivalence, which effectively solves the problem of accuracy loss in the process of information transmission at different scales; designing a multi-time step coordination mechanism and spatial filtering and smoothing technology, which systematically solves the problem of differences in time scales and spatial scales. The starting point of the improvement is to solve the "scale gap" problem that is prevalent in material calculations, that is, the problem of the connection between microscopic physical mechanisms and macroscopic performance.

[0121] In an optional embodiment,

[0122] Inputting the raw material parameters, target formula information, and production environment parameters into the numerical simulation unit for calculation, correcting the weight ratio of each component in the target formula information according to the calculation results, and converting the corrected weight ratio into quantitative feeding control parameters includes:

[0123] Constructing a multidimensional dataset including the raw material parameters, target recipe information, and production environment parameters, inputting the multidimensional dataset into a multi-physics field coupled numerical simulation unit, establishing an objective function through the numerical simulation unit, the objective function including a relative deviation term between a performance indicator and a target value, a performance indicator weight coefficient, and a recipe cost factor; and obtaining an optimized ratio vector by minimizing the objective function;

[0124] An interaction effect correction coefficient matrix is ​​established, wherein the interaction effect correction coefficient matrix is ​​a symmetric matrix, and the matrix elements represent the interaction influence coefficients between the components; the matrix elements are determined by iterative fitting through the simulation results of the numerical simulation unit; the optimized ratio vector is corrected using the interaction effect correction coefficient matrix to obtain a corrected ratio vector; and the quantitative feeding control parameters are calculated based on the corrected ratio vector, the total feeding amount, the feeding time, and the feeding device calibration coefficient.

[0125] For example, when constructing a multidimensional dataset containing raw material parameters, target formulation information, and production environment parameters, multiple types of data must be collected and organized. Raw material parameters include the physical and chemical properties and processing characteristics of the raw materials. For example, for anti-corrosion powder coatings for drinking water pipelines, the main components include epoxy resin matrix (55-65%), curing agent (15-20%), anti-corrosion filler (8-12%), flow aid (3-6%), and pigment (3-6%). Specific raw material parameters include the epoxy value (0.52-0.58 equivalent weight / 100g), melting point (85-95°C), particle size distribution (D50=65 microns), and particle shape factor (1.12-1.18) of the epoxy resin; the active hydrogen content (135-155 mg KOH / g), melting point (90-100°C), and particle size distribution (D50=55 microns) of the curing agent; the particle size (5-15 microns), specific surface area (2.5-3.5 m² / g), and porosity (0.35-0.42) of the anti-corrosion filler; the melting point (75-85°C) and thixotropic index (1.2-1.4) of the additive; and the hiding power (95-98%) and particle size (0.2-0.8 microns) of the pigment. The target formulation information includes target performance requirements such as adhesion (≥5MPa), impact resistance (≥50kg·cm), coating hardness (2-3 hours), water resistance (no abnormalities after immersion for 168 hours), and corrosion resistance (no rust after 1000 hours of salt spray testing). Production environment parameters include temperature (22-25°C), relative humidity (45-55%), air pressure (101±1kPa), and mixing equipment speed (30-60rpm). These parameters form a multidimensional dataset containing 32 variables.

[0126] The multidimensional dataset was input into the previously constructed numerical simulation unit based on quantitative eigenvalue elements. When establishing the objective function through the numerical simulation unit, the objective function consisted of three components: the relative deviation term between the performance indicator and the target value, the performance indicator weight coefficient, and the formulation cost factor. The relative deviation term was calculated by taking the normalized difference between the predicted value and the target value of each performance indicator. For example, the target adhesion value was 5.5 MPa, and the numerical simulation predicted a value of 5.2 MPa, with a relative deviation of 5.5%. The target corrosion resistance value was rust-free after 1000 hours in a salt spray test, and the predicted value was 950 hours, with a relative deviation of 5%. The performance indicator weight coefficients were set based on the importance of each indicator: the corrosion resistance weight was 0.35, the adhesion weight was 0.25, the water resistance weight was 0.20, the hardness weight was 0.10, and the impact resistance weight was 0.10. The formula cost factors take into account raw material and processing costs. The unit price of epoxy resin is 18 yuan / kg, the unit price of curing agent is 25 yuan / kg, the unit price of anti-corrosion filler is 32 yuan / kg, the unit price of additives is 48 yuan / kg, and the unit price of pigment is 60 yuan / kg. The optimized ratio vector is obtained by minimizing the objective function. The particle swarm optimization algorithm is used to solve the problem. The number of particles is set to 100, the inertia weight is 0.7, the individual learning factor is 1.5, the social learning factor is 1.8, and convergence is achieved after 150 iterations. The optimized ratio vector is 61.2% epoxy resin, 18.5% curing agent, 10.8% anti-corrosion filler, 4.2% flow aid, and 5.3% pigment.

[0127] To establish the interaction correction coefficient matrix, a 5×5 symmetric matrix (corresponding to the five main components) was first constructed. The matrix elements represent the interaction coefficients between the components. These coefficients reflect the physicochemical interactions between the components at the microscopic level and their impact on macroscopic properties. Interaction data for different component ratios were obtained through the collaborative calculation of the microscopic force calculations and the macroscopic flow prediction module within the aforementioned numerical simulation unit. For example, the interaction coefficient between epoxy resin and curing agent was 0.95, indicating good reactivity between the two. The interaction coefficient between epoxy resin and anticorrosive filler was 0.82, indicating that the filler had some hindrance to the epoxy resin's fluidity. The interaction coefficient between curing agent and flow aid was 1.12, indicating that the additive promoted the dispersion and reaction of the curing agent. The matrix elements were determined by iterative fitting based on the simulation results of the numerical simulation unit. Specifically, an orthogonal design experiment was used to design 20 different ratios. The performance indicators were predicted using the numerical simulation unit, and the interaction coefficients were fitted using nonlinear regression. After six iterations, the average relative error of the matrix elements was less than 2.5%, and the fitting accuracy met the requirements for industrial application.

[0128] When using the interaction effect correction coefficient matrix to correct the optimized mix ratio vector, the optimized mix ratio vector is multiplied by the interaction effect correction coefficient matrix to obtain the corrected mix ratio vector. The epoxy resin content before correction was 61.2%, which was adjusted to 60.5% after correction; the curing agent content before correction was 18.5%, which was adjusted to 18.8% after correction; the anti-corrosion filler content before correction was 10.8%, which was adjusted to 11.2% after correction; the flow aid content before correction was 4.2%, which was adjusted to 4.0% after correction; and the pigment content before correction was 5.3%, which was adjusted to 5.5% after correction. Numerical simulation unit verification shows that the corrected mix ratio significantly improves mixing uniformity and fluidity. The Lacey mixing index increases from 0.82 to 0.88, and the powder flow angle decreases from 36° to 33°, making it more conducive to subsequent spraying processing.

[0129] When calculating the quantitative feeding control parameters according to the corrected ratio vector, total feeding amount, feeding time and feeding device calibration coefficient, taking the total feeding amount of 200 kg and the feeding time of 40 minutes as an example, the epoxy resin feeding rate is set to 60.5%×200 kg÷40 min×1.02 (calibration coefficient)=3.09 kg / min; the curing agent feeding rate is 18.8%×200 kg÷40 min×0.98=0.92 kg / min; the anti-corrosion filler feeding rate is 11.2%×200 kg÷40 min×1.03=0.58 kg / min; the flow aid feeding rate is 4.0%×200 kg÷40 min×0.97=0.19 kg / min; and the pigment feeding rate is 5.5%×200 kg÷40 min×1.01=0.28 kg / min.

[0130] This invention introduces an interaction correction coefficient matrix to quantify the synergistic and antagonistic effects between components, overcoming the limitations of traditional linear additive models. It also establishes a comprehensive objective function encompassing performance, cost, and process adaptability to achieve multi-objective balanced optimization. This invention addresses the difficulty in accurately describing and predicting interactions between components in traditional formulation design, establishing a scientific formulation design method based on multi-scale computation.

[0131] In an optional embodiment,

[0132] Using real-time weight detection and material blockage identification results, the self-learning feedback compensation algorithm based on the regression network dynamically adjusts the quantitative feeding control parameters including:

[0133] Obtain the real-time weight value and target weight value of the material; extract the silo pressure signal, material temperature signal and vibration signal collected during the material transportation process through the feature mapping function and perform weighted fusion to obtain the blockage status index;

[0134] A deep spatiotemporal regression network with a causal inference enhancement layer is constructed. The network takes the time series deviation between the real-time material weight and the target weight, as well as the congestion status indicator, as input features. The network captures dynamic characteristics through parallel temporal convolution branches and spatial relationship branches, and outputs control parameter compensation. The causal inference enhancement layer identifies the causal relationship structure between control factors and outcomes through structured intervention simulation and counterfactual reasoning mechanisms, and employs a bidirectional attention mechanism to adaptively fuse physical prior knowledge with data-driven features.

[0135] Construct a hierarchical multi-objective loss function of the main control objectives, process objectives, and constraints, dynamically optimize the control parameters based on the meta-learning framework, and dynamically adjust the update direction and step size of the parameters according to the control error characteristics and congestion risk;

[0136] The feedforward compensation control quantity is calculated based on the prediction results of the deep spatiotemporal regression network. Combined with the feedback correction term generated online, the total control quantity of the feeding system is generated through an adaptive fusion algorithm. The adaptive fusion algorithm can dynamically adjust the weight ratio of feedforward compensation and feedback correction according to the material's cumulative weight error, error change rate and blockage risk index; the total control quantity is executed to dynamically adjust the quantitative feeding control parameters, and the control effect is stored in the experience replay pool. The deep spatiotemporal regression network is continuously optimized through knowledge distillation and incremental learning.

[0137] For example, when obtaining the real-time and target weight values ​​of the material, a high-precision weighing sensor is used to monitor the material output weight of the feeding equipment in real time. For the epoxy resin component of the anti-corrosion powder coating for drinking water pipelines, the target feeding rate is 3.09 kg / min. The system collects weight data every 100 milliseconds to form a weight time-series data stream. The real-time weight value is noise-filtered by the signal processing circuit. Using the Kalman filter algorithm, the process noise covariance is set to 0.02 and the observation noise covariance is set to 0.05. This significantly improves the signal-to-noise ratio of the weight signal, and the stability of the filtered weight data is improved by approximately 85%. At the same time, the silo pressure signal, material temperature signal, and vibration signal are collected during the material transportation process. For the epoxy resin component, the normal silo pressure value is 0.15-0.18 MPa, the temperature is 22-25°C, and the vibration frequency characteristic peak is 35-45 Hz.

[0138] The collected multi-source signals are extracted and weightedly fused using a feature mapping function to obtain a congestion status indicator. Feature extraction uses a method that combines wavelet packet decomposition with statistical features. Four features, namely mean, standard deviation, kurtosis, and skewness, are extracted from the silo pressure signal; two features, namely mean and rate of change, are extracted from the temperature signal; and three features, namely frequency domain energy distribution, main frequency, and root mean square value, are extracted from the vibration signal. Weighted feature fusion uses an adaptive weighting algorithm, with initial weights set to 0.5 for pressure features, 0.2 for temperature features, and 0.3 for vibration features. During system operation, weights are dynamically adjusted based on the correlation between each feature and the congestion event. The fused congestion status indicator ranges from 0 to 1, with values ​​less than 0.3 indicating normal, 0.3-0.7 indicating a slight congestion risk, and values ​​greater than 0.7 indicating a severe congestion risk.

[0139] The deep spatiotemporal regression network with a causal inference enhancement layer is constructed. The network architecture consists of an input layer, a temporal convolution branch, a spatial relationship branch, a causal inference enhancement layer, and an output layer. The input layer receives the time series deviation between the real-time material weight and the target weight (with a window length of 60 data points, approximately 6 seconds of data) and a congestion status indicator. The temporal convolution branch uses three one-dimensional convolution layers with kernel sizes of 5, 3, and 3, respectively, and the number of kernels is 32, 64, and 128, respectively, to capture the dynamic characteristics and trends of the time series data. The spatial relationship branch uses a structure combining fully connected layers with an attention mechanism, with a node count of 64, 32, and 16 in the fully connected layers, to capture the spatial correlation between different sensor signals. The causal inference enhancement layer uses structured intervention simulation and counterfactual reasoning to identify the causal relationship between control factors and outcomes. The structured intervention simulation uses a virtual manipulated variable method to construct a causal graph connecting feed rate, vibration frequency, and congestion risk. Edge weights are trained using historical data. The counterfactual reasoning mechanism, based on a dual machine learning model, uses a random forest algorithm (with a tree depth of 5 and a total of 100 trees) to construct a counterfactual predictor, inferring the causal relationship that "if the feed rate is adjusted to X, the blockage risk will change to Y." A bidirectional attention mechanism adaptively integrates physical prior knowledge with data-driven features. The physical prior knowledge includes powder flow characteristic models (e.g., the flow angle of epoxy resin powder is 33° and the angle of repose is 38°) and equipment characteristic parameters. Attention weights are calculated using a softmax function.

[0140] When constructing a hierarchical multi-objective loss function, control objectives are divided into three levels. The primary control objective layer includes minimizing the cumulative material weight error and feed rate stability, with target thresholds for cumulative material weight error of ±0.5% and feed rate fluctuation of ±2%. The process objective layer includes minimizing the unit-time weight error and controlling the jam risk, with target thresholds for unit-time weight error of ±1.5% and a jam risk index of 0.3. The constraint layer includes equipment safety constraints (e.g., a maximum feed rate of no more than 4.5 kg / min) and material property constraints (e.g., minimum flowability requirements). Control parameters are dynamically optimized using a meta-learning framework. The outer loop utilizes the MAML (Model-Agnostic Meta-Learning) algorithm, while the inner loop utilizes the Adam optimizer. The learning rate is initially set to 0.001 and dynamically adjusted based on the control error characteristics and jam risk. When the epoxy resin powder weight error continues to increase and exceeds 1.2%, the learning rate is automatically increased to 0.002. When the jam risk index exceeds 0.5, the optimization shifts toward reducing the feed rate, with the step size reduced to 0.0005.

[0141] Based on the prediction results of a deep spatiotemporal regression network, a feedforward compensation control variable is calculated. For epoxy resin powder, when the detected weight error is -2% (i.e., underfeeding) and the jam risk index is 0.25, the network predicts a 7.5% increase in the feed rate. Simultaneously, a feedback correction term is generated in real time via a PI controller with a proportional coefficient Kp of 0.65 and an integral coefficient Ki of 0.08. The feedforward compensation control variable and the feedback correction term are combined using an adaptive fusion algorithm to generate the total control variable for the feeding system. The adaptive fusion algorithm dynamically adjusts the weight ratio based on the material's cumulative weight error, error rate of change, and jam risk index. When the cumulative weight error is large (>2%) and the rate of change is low, the feedforward compensation weight is increased to 0.7, while the feedback correction weight is 0.3. When the jam risk is high (>0.6), the feedforward compensation weight is reduced to 0.3, while the feedback correction weight is increased to 0.7, prioritizing system safety.

[0142] The total control variable is dynamically adjusted to adjust the quantitative feeding control parameters, updating the feeding system's speed setpoint, vibration frequency, and silo pressure. For epoxy resin powder, the feed rate was 3.09 kg / min before adjustment, but after adjustment, it was changed to 3.32 kg / min based on the total control variable. The vibration frequency was adjusted from 40 Hz to 43 Hz, and the silo pressure was adjusted from 0.16 MPa to 0.17 MPa. The adjusted control parameters are transmitted to the actuator in real time, with a system response time of less than 200 milliseconds. The control results (including weight error before and after adjustment, change in blockage risk, and control parameter adjustment) are stored in an experience replay pool with a capacity of 10,000 records and a first-in, first-out policy. A deep spatiotemporal regression network is continuously optimized through knowledge distillation and incremental learning. The knowledge distillation uses a teacher-student network structure with a temperature coefficient of 2.0. Incremental learning updates the model every 500 new data points, using a gradient accumulation method (10 cumulative steps) for batch updates to maintain model adaptability to new data while avoiding catastrophic forgetting.

[0143] After processing by the above algorithm, the feeding weight error of epoxy resin powder was reduced from the initial ±2.8% to ±0.6%, the occurrence rate of blockage events was reduced from 5.2 times per 100 hours to 0.3 times, and the average processing time of a single blockage was reduced from 12 minutes to 2.5 minutes. The overall feeding accuracy of the system was improved by about 78%, and production efficiency was improved by about 25%.

[0144] like Figure 4 As shown in the comparison graph of the real-time weight accumulation curve of material feeding, the horizontal axis in the figure represents time (seconds), the vertical axis represents cumulative weight (kg), and the black horizontal dotted line represents the target weight (20kg). The three control methods distinguished by different line types are the method of the present invention (solid line), traditional PID control (dashed line) and deep reinforcement learning control (dotted line). It can be seen from the trend of the curve that the response speed of the method of the present invention is faster, and it quickly approaches the target weight in about 5 seconds and is accurate and stable with minimal error; while the traditional PID control has obvious overshoot phenomenon, and it takes a long time to stabilize near the target value; the performance of deep reinforcement learning control is between the two, and although it responds faster, there are still slight fluctuations. This shows that the method of the present invention has significant advantages in feeding accuracy and response speed.

[0145] like Figure 5The comparison chart of congestion state prediction and control parameter adjustment shows the performance of the three control methods from two dimensions. The left vertical axis represents the congestion risk index (0-100), which is presented in the form of a bar graph. Black represents the method of the present invention, dark gray represents traditional PID control, and light gray represents deep reinforcement learning control; the right vertical axis represents the control parameter adjustment amount (-1.0 to 1.0), which is presented in the form of a curve. It can be observed from the figure that: the method of the present invention can identify the congestion risk earlier and respond quickly at the initial stage of risk increase; after identifying the risk, the control parameter adjustment speed of the method of the present invention is faster and the amplitude is more reasonable, so that the congestion risk decreases rapidly; the traditional PID control has a significant lag in identifying the congestion risk, and the parameter adjustment is not timely; the method of the present invention shows better anti-interference ability and adaptability throughout the process.

[0146] Existing powder feeding control technologies mainly adopt: quantitative feeding systems based on PID control, which have fixed parameters and cannot cope with changes in material properties; learning control systems based on standard neural networks, which lack causal reasoning capabilities and have difficulty explaining the control decision-making process and responding to unseen abnormal situations. The self-learning feedback compensation algorithm based on regression networks proposed in this application introduces a causal inference enhancement layer, combines structured intervention simulation and counterfactual reasoning mechanisms, and enables the control system to have causal reasoning capabilities and be able to distinguish between correlation and causality; adopts a dual-branch network structure with time and space separation to capture the dynamic characteristics of time series and multi-sensor spatial relationships respectively; designs a hierarchical multi-objective loss function and a meta-learning framework, which can dynamically optimize the control strategy under multi-objective constraints. The present invention can solve the problems of unstable accuracy and frequent blockages caused by fluctuations in material properties, environmental changes and equipment wear in the powder material feeding process.

[0147] In an optional embodiment,

[0148] Construct a hierarchical multi-objective loss function of the main control objectives, process objectives, and constraints, dynamically optimize the control parameters based on the meta-learning framework, and dynamically adjust the update direction and step size of the parameters according to the control error characteristics and congestion risk.

[0149] Constructing a hierarchical multi-objective loss function, the hierarchical multi-objective loss function includes a main control target layer loss function, a process target layer loss function and a constraint condition layer loss function, wherein the main control target layer loss function is used to characterize product ratio accuracy and mixing uniformity, the process target layer loss function is used to characterize weight control accuracy, anti-blocking capability and energy consumption optimization goals, and the constraint condition layer loss function is used to characterize equipment safety constraints and material property constraints. The main control target layer loss function, the process target layer loss function and the constraint condition layer loss function are combined to obtain a comprehensive loss function;

[0150] A meta-learning framework is used to optimize control parameters. In an inner loop, the control parameters are updated based on the gradient of the comprehensive loss function. In an outer loop, the meta-learner parameters are optimized, the sensitivity of each control parameter to the comprehensive loss function is calculated, and the learning rate of each control parameter is dynamically adjusted based on the sensitivity, weight control error, and congestion risk index.

[0151] The degree of conflict between objectives is detected and the optimization priority is determined. Based on the optimization priority, the weight of each objective in the comprehensive loss function is progressively adjusted. The update direction and step size of different parameters are dynamically adjusted according to the control error characteristics and congestion risk.

[0152] For example, when constructing a hierarchical multi-objective loss function, the control objectives are divided into three hierarchical levels, forming a hierarchical multi-objective optimization structure. The master objective layer loss function is used to minimize the material cumulative weight error and feed rate stability. The target threshold for the material cumulative weight error is ±0.5%, and the target threshold for the feed rate fluctuation is ±2%. Taking the epoxy resin component as an example, the target ratio is 60.5%, with an allowable deviation range of ±0.5%. The master objective layer loss function combines the cumulative weight error term with the feed rate stability term. The process objective layer loss function is used to minimize the unit time weight error and control the jam risk. The target threshold for the unit time weight error is ±1.5%, and the target threshold for the jam risk indicator is 0.3. The unit time weight error is calculated as the root mean square deviation between the feed rate and the target rate at 10 consecutive sampling points; the jam risk is represented by the jam state indicator. The process objective layer loss function combines the deviations of these two indicators. The constraint layer loss function is used to represent equipment safety constraints and material property constraints. Equipment safety constraints include a maximum feed rate of no more than 85% of the rated value (4.5 kg / min) and a maximum vibration frequency of no more than 50 Hz. Material property constraints include minimum flowability requirements (flow angle no more than 38°) and a temperature limit (no more than 30°C). The constraint layer loss function uses a barrier function approach, where the loss value increases exponentially when a constraint is about to be violated. The three loss functions are combined to form a comprehensive loss function. Using a hierarchical weighting method, the initial weights assigned to the master control objective layer, process objective layer, and constraint layer are 0.5, 0.3, and 0.2, respectively.

[0153] When optimizing control parameters using a meta-learning framework, a two-layer optimization structure is constructed. The inner loop updates the control parameters based on the gradient of the comprehensive loss function, while the outer loop optimizes the meta-learner parameters. The inner loop uses the Adam optimizer to update control parameters such as the feed rate, vibration frequency, and silo pressure. The initial learning rate is set to 0.001, β1 is 0.9, and β2 is 0.999. For epoxy resin powder, for example, when a weight error of -1.8% and a blockage risk index of 0.28 are detected, the feed rate is initially set to 3.09 kg / min, which is adjusted to 3.25 kg / min after five iterations of the inner loop. The outer loop optimizes the meta-learner parameters, including the learning rate adjustment factor, the target weight adjustment coefficient, and the feature extraction parameters. The outer loop uses the MAML (Model-Agnostic Meta-Learning) algorithm, with outer loop updates performed every 200 inner loop iterations. The sensitivity of each control parameter to the comprehensive loss function was calculated, and the gradient was estimated using the finite difference method. For example, the sensitivity of the epoxy resin feed rate was 0.42, the sensitivity of the vibration frequency was 0.28, and the sensitivity of the silo pressure was 0.15. The learning rate of each control parameter was dynamically adjusted based on the sensitivity, weight control error, and the blockage risk index. When the epoxy resin weight error continued to increase and exceeded 1.2%, the learning rate of the feed rate parameter was automatically increased to 0.002. When the blockage risk index exceeded 0.5, the optimization direction shifted to reducing the feed rate, and the step size was reduced to 0.0005.

[0154] Correlation analysis and conflict detection algorithms are used to detect the degree of conflict between objectives and determine optimization priorities. The gradient cosine similarity between different objectives is calculated. A positive cosine similarity indicates consistent objective orientation, while a negative cosine similarity indicates conflict. For example, under specific operating conditions, the gradient cosine similarity between the weight control accuracy objective and the anti-blocking capability objective is -0.72, indicating a strong conflict. Optimization priorities are determined based on the degree of conflict and objective importance, and priority weights are calculated using the analytic hierarchy process. In the epoxy resin powder feeding process, when a conflict occurs between weight control and anti-blocking objectives, if the blocking risk index is less than 0.5, the weight control objective priority is 0.65, and the anti-blocking objective priority is 0.35. If the blocking risk index is greater than or equal to 0.5, the weight control objective priority is reduced to 0.25, and the anti-blocking objective priority is increased to 0.75. Based on the optimization priority, the weights of each objective in the comprehensive loss function are progressively adjusted, and a soft attention mechanism is used to dynamically assign weights. The update direction and step size of different parameters are dynamically adjusted according to the control error characteristics and blockage risk. When the blockage risk increases, the update direction of the feeding rate parameter tends to decrease, and the step size is reduced to 60% of the original; the update direction of the vibration frequency parameter tends to increase, and the step size is increased to 150% of the original, thereby ensuring material fluidity while reducing the blockage risk.

[0155] This invention achieves precise adaptive optimization of control parameters, significantly improving the proportioning accuracy and mixing uniformity of epoxy resin components, significantly reducing the incidence of blockage events, effectively reducing system energy consumption, and significantly improving production efficiency. Compared with traditional single PID control methods and fixed-weight multi-objective optimization methods, this method exhibits greater adaptability and robustness in the face of changes in material properties and environmental disturbances. It can effectively solve the problems of precision fluctuations and blockages in the powder material feeding process, improve system stability and reliability, and meet the technical requirements of high-quality coating production.

[0156] The present invention also provides that when the mixing equipment is started based on dynamically adjusted quantitative feeding control parameters, the system first transmits the optimized feeding rate, vibration frequency, and silo pressure parameters to the programmable logic controller. The controller controls the operating status of the screw feeder and vibrator through a proportional valve and a frequency converter, respectively. For the epoxy resin powder component, the feeding rate is set to 3.25 kg / min, the vibration frequency is set to 42 Hz, and the silo pressure is set to 0.17 MPa; for the curing agent component, the feeding rate is set to 1.3 kg / min, the vibration frequency is set to 38 Hz, and the silo pressure is set to 0.15 MPa. When the mixing equipment is started for batching, a dual-shaft paddle mixer is used with a rotation speed set to 75 rpm. A capacitive humidity sensor, an infrared temperature sensor, and a conductivity sensor are installed in the mixer, with a sampling frequency of 200 ms / time.

[0157] When collecting mixing process data, the system acquires parameters such as mixture temperature, mixing torque, material conductivity, and humidity, forming a multidimensional time-series data stream. The collected data is filtered and denoised by the signal processing module before being input into the numerical simulation unit in real time. Based on an improved discrete element method, the numerical simulation unit updates the currently collected physical parameters into the particle dynamics calculation model, calculates the spatial distribution of the different component particles in the mixture, and then evaluates the mixing uniformity index. The system uses the Lacey mixing index as a mixing uniformity indicator, calculated by the particle concentration variance ratio. For the mixing process of epoxy resin and curing agent, the target Lacey index is set to 0.85.

[0158] When the real-time calculated Lacey index is higher than the target value of 0.85 for 5 seconds, the system determines that the mixing has reached the target requirement, automatically slows down and stops the mixer. Under typical working conditions, the mixing time of epoxy resin and curing agent is about 120 seconds, which reduces energy consumption by about 30% compared with the traditional fixed-time mixing method. After the mixing is completed, the system starts the automatic sampling device and extracts samples from 5 different positions in the mixture for near-infrared spectroscopy analysis to detect the actual content of each component. The test results are automatically compared with the target formula information, the ratio deviation of each component is calculated, and a detailed batching process report is generated, including raw material parameters, ratio adjustment amount, mixing time, energy consumption data, quality test results and qualification judgment. The batching process report is uploaded to the production management system through the data interface to achieve full process data traceability.

[0159] According to a second aspect of the embodiments of the present invention,

[0160] Provides a multi-component intelligent dosing system for anti-corrosion powder coatings for drinking water pipelines, including:

[0161] The first unit is used to obtain the raw material parameters and target formula information of the anti-corrosion powder coating for drinking water pipelines;

[0162] The second unit is used to construct a numerical simulation unit based on quantitative characteristic volume elements. The microscopic force calculation module calculates the interaction force between powder particles based on a simplified molecular dynamics method and forms a parameterized database. The macroscopic flow prediction module uses an improved discrete element method to predict the flow and mixing behavior of the powder particle group in the target sampling area. The improved discrete element method includes an adaptive time step algorithm and a domain decomposition parallel algorithm. The microscopic force calculation module and the macroscopic flow prediction module perform data interaction and collaborative calculation based on a multi-scale bridging calculation method of an adaptive coupling region.

[0163] a third unit configured to input the raw material parameters, target formula information, and production environment parameters into the numerical simulation unit for calculation, correct the weight ratio of each component in the target formula information based on the calculation results, and convert the corrected weight ratio into a quantitative feeding control parameter; and dynamically adjust the quantitative feeding control parameter through a self-learning feedback compensation algorithm based on a regression network using the results of real-time weight detection and material blockage identification;

[0164] The fourth unit is used to start the mixing equipment to batch the ingredients based on the dynamically adjusted quantitative feeding control parameters, collect the mixing process data and input it into the numerical simulation unit to calculate the mixing uniformity index in real time, automatically stop mixing when the target requirements are met, detect the mixed finished product, compare the target formula information and generate a batching process report.

[0165] According to a third aspect of the embodiments of the present invention,

[0166] An electronic device is provided, comprising:

[0167] processor;

[0168] a memory for storing processor-executable instructions;

[0169] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0170] According to a fourth aspect of the embodiments of the present invention,

[0171] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0172] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0173] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-component intelligent batching method for anti-corrosion powder coatings for drinking water pipelines, characterized in that: include: Obtain raw material parameters and target formula information for anti-corrosion powder coatings for drinking water pipelines; A numerical simulation unit based on quantitative characteristic volume elements is constructed. The microscopic force calculation module calculates the interaction forces between powder particles using a simplified molecular dynamics method and forms a parameterized database. The macroscopic flow prediction module uses an improved discrete element method to predict the flow and mixing behavior of powder particle groups within the target sampling area. The improved discrete element method includes an adaptive time step algorithm and a domain decomposition parallel algorithm. The microscopic force calculation module and the macroscopic flow prediction module perform data exchange and collaborative calculations based on a multi-scale bridging calculation method in an adaptive coupling region. Inputting the raw material parameters, target formula information, and production environment parameters into the numerical simulation unit for calculation, correcting the weight ratio of each component in the target formula information according to the calculation results, and converting the corrected weight ratio into a quantitative feeding control parameter; Utilizing real-time weight detection and material blockage identification results, the quantitative feeding control parameters are dynamically adjusted through a self-learning feedback compensation algorithm based on a regression network; Based on the dynamically adjusted quantitative feeding control parameters, the mixing equipment is started to batch the ingredients, the mixing process data is collected and input into the numerical simulation unit to calculate the mixing uniformity index in real time, the mixing is automatically stopped when the target requirements are met, the mixed finished product is tested, the target formula information is compared and a batching process report is generated.

2. The method according to claim 1, characterized in that The construction of the quantitative characteristic volume element includes: Measuring a powder particle sample to obtain size distribution data, morphological characteristic data, and spatial arrangement data of the powder particles; determining a characteristic radius of the powder particles based on the size distribution data, calculating a shape factor based on the morphological characteristic data, and calculating an interparticle void ratio based on the spatial arrangement data; Calculate the volume of a quantitative eigenvalue: ； Among them, V QCVEV is the volume of the quantitative characteristic volume element, R i is the characteristic radius of the i-th type powder particles, f s is the shape factor, δ is the interparticle void ratio, and N is the number of types of powder particles; Determining a minimum size threshold of the quantitative characteristic volume element so that the physical properties of the quantitative characteristic volume element do not change with the increase of volume when the size is above the minimum size threshold; Based on the volume and minimum size threshold of the quantitative characteristic volume element, a three-dimensional digital representation of the quantitative characteristic volume element is generated and assigned physical property parameters. Boundary constraints and initial conditions are imposed on the three-dimensional digital representation so that the quantitative characteristic volume element can represent the structural characteristics and mechanical behavior of the overall powder system.

3. The method according to claim 1, characterized in that The microscopic force calculation module calculates the interaction forces between powder particles based on a simplified molecular dynamics method and forms a parameterized database. The macroscopic flow prediction module uses an improved discrete element method to predict the flow and mixing behavior of powder particle groups within the target sampling area, including: The microscopic force calculation module uses a simplified molecular dynamics method to calculate the interaction force between powder particles. The simplified molecular dynamics method represents multiple molecular collections as a single interaction point and calculates the interparticle force through the spatial derivative of the effective potential energy function. The microscopic force calculation module calculates the van der Waals force, electrostatic force, and capillary force between powder particles. The van der Waals force is corrected using the Hamaker constant, the electrostatic force uses the modified Coulomb's law, and the capillary force takes into account the influence of ambient humidity. The microscopic force calculation results are stored to form a parameterized database, which uses a multi-dimensional interpolation algorithm to achieve fast calculation under different process conditions; The macro-flow prediction module simulates powder particle populations using an improved discrete element method (DEM), which includes an adaptive time step algorithm and a domain decomposition parallel algorithm. The adaptive time step adjusts the calculation step size based on the dynamic characteristics of the particles, and the domain decomposition parallel algorithm decomposes the calculation domain into multiple sub-regions. Based on the force data in the parameterized database, an improved Verlet algorithm is applied to solve the powder particle motion equations, wherein the improved Verlet algorithm introduces velocity half-step iteration and position constraint correction to reduce numerical errors, and the motion equations include the particle linear motion equations and angular motion equations; Based on the results of solving the motion equation, the macro flow prediction module calculates the flow velocity field and mixing uniformity in the target area, where the flow velocity field is calculated by particle mass, velocity and weight function, and the mixing uniformity is evaluated using the Lacey mixing index.

4. The method according to claim 1, wherein The microscopic force calculation module and the macroscopic flow prediction module perform data interaction and collaborative calculations based on the multi-scale bridging calculation method of the adaptive coupling region, including: The microscopic force calculation results are transferred to the macroscopic flow prediction module through the parameter mapping method, where the microscopic force calculation results are converted into the forces used in the macroscopic calculation through the combination of mapping operators and transfer parameter sets; An adaptive coupling region is set at the interface between the microscopic force calculation module and the macroscopic flow prediction module, wherein the adaptive coupling region includes all points that are no more than a preset critical distance from the interface; The time scale difference between micro-scale and macro-scale calculations is reconciled through a multi-time-step method, where the product of the number of micro-scale sub-steps and the micro-scale time step is equal to the macro-scale time step; A spatial filtering method is used to smooth the spatial scale difference between the microscale and the macroscale, where the calculated physical quantity at the macroscale is obtained by integrating the calculated physical quantity at the microscale through the spatial filtering kernel function; The collaborative calculation is performed in an iterative coupling manner, including: initializing a microscopic force calculation module and a macroscopic flow prediction module; calculating the interaction force between particles at the microscopic scale; mapping the calculated microscopic force to the macroscopic scale through the parameter mapping method; calculating the movement of powder particle groups at the macroscopic scale; feeding back the macroscopic calculation results to the microscopic calculation module to update the boundary conditions; calculating the relative error between two adjacent iterative results, and outputting the result if it is less than a preset error threshold, otherwise returning to perform the microscopic inter-particle interaction force calculation.

5. The method according to claim 1, characterized in that Inputting the raw material parameters, target formula information, and production environment parameters into the numerical simulation unit for calculation, correcting the weight ratio of each component in the target formula information according to the calculation results, and converting the corrected weight ratio into quantitative feeding control parameters includes: Constructing a multidimensional dataset including the raw material parameters, target recipe information, and production environment parameters, inputting the multidimensional dataset into a multi-physics field coupled numerical simulation unit, establishing an objective function through the numerical simulation unit, the objective function including a relative deviation term between a performance indicator and a target value, a performance indicator weight coefficient, and a recipe cost factor; and obtaining an optimized ratio vector by minimizing the objective function; An interaction effect correction coefficient matrix is ​​established, wherein the interaction effect correction coefficient matrix is ​​a symmetric matrix, and the matrix elements represent the interaction influence coefficients between the components; the matrix elements are determined by iterative fitting through the simulation results of the numerical simulation unit; the optimized ratio vector is corrected using the interaction effect correction coefficient matrix to obtain a corrected ratio vector; and the quantitative feeding control parameters are calculated based on the corrected ratio vector, the total feeding amount, the feeding time, and the feeding device calibration coefficient.

6. The method according to claim 1, characterized in that Using real-time weight detection and material blockage identification results, the self-learning feedback compensation algorithm based on the regression network dynamically adjusts the quantitative feeding control parameters including: Obtain the real-time weight value and target weight value of the material; extract the silo pressure signal, material temperature signal and vibration signal collected during the material transportation process through the feature mapping function and perform weighted fusion to obtain the blockage status index; A deep spatiotemporal regression network with a causal inference enhancement layer is constructed. The network takes the time series deviation between the real-time material weight and the target weight, as well as the congestion status indicator, as input features. The network captures dynamic characteristics through parallel temporal convolution branches and spatial relationship branches, and outputs control parameter compensation. The causal inference enhancement layer identifies the causal relationship structure between control factors and outcomes through structured intervention simulation and counterfactual reasoning mechanisms, and employs a bidirectional attention mechanism to adaptively fuse physical prior knowledge with data-driven features. Construct a hierarchical multi-objective loss function of the main control objectives, process objectives, and constraints, dynamically optimize the control parameters based on the meta-learning framework, and dynamically adjust the update direction and step size of the parameters according to the control error characteristics and congestion risk; The feedforward compensation control quantity is calculated based on the prediction results of the deep spatiotemporal regression network. Combined with the feedback correction term generated online, the total control quantity of the feeding system is generated through an adaptive fusion algorithm. The adaptive fusion algorithm can dynamically adjust the weight ratio of feedforward compensation and feedback correction according to the material's cumulative weight error, error change rate and blockage risk index; the total control quantity is executed to dynamically adjust the quantitative feeding control parameters, and the control effect is stored in the experience replay pool. The deep spatiotemporal regression network is continuously optimized through knowledge distillation and incremental learning.

7. The method according to claim 6, characterized in that Construct a hierarchical multi-objective loss function of the main control objectives, process objectives, and constraints, dynamically optimize the control parameters based on the meta-learning framework, and dynamically adjust the update direction and step size of the parameters according to the control error characteristics and congestion risk. Constructing a hierarchical multi-objective loss function, the hierarchical multi-objective loss function includes a main control target layer loss function, a process target layer loss function and a constraint condition layer loss function, wherein the main control target layer loss function is used to characterize product ratio accuracy and mixing uniformity, the process target layer loss function is used to characterize weight control accuracy, anti-blocking capability and energy consumption optimization goals, and the constraint condition layer loss function is used to characterize equipment safety constraints and material property constraints. The main control target layer loss function, the process target layer loss function and the constraint condition layer loss function are combined to obtain a comprehensive loss function; A meta-learning framework is used to optimize control parameters. In an inner loop, the control parameters are updated based on the gradient of the comprehensive loss function. In an outer loop, the meta-learner parameters are optimized, the sensitivity of each control parameter to the comprehensive loss function is calculated, and the learning rate of each control parameter is dynamically adjusted based on the sensitivity, weight control error, and congestion risk index. The degree of conflict between objectives is detected and the optimization priority is determined. Based on the optimization priority, the weight of each objective in the comprehensive loss function is progressively adjusted. The update direction and step size of different parameters are dynamically adjusted according to the control error characteristics and congestion risk.

8. A multi-component intelligent proportioning system for anti-corrosion powder coatings for drinking water pipelines, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain the raw material parameters and target formula information of the anti-corrosion powder coating for drinking water pipelines; The second unit is used to construct a numerical simulation unit based on quantitative characteristic volume elements. The microscopic force calculation module calculates the interaction force between powder particles based on a simplified molecular dynamics method and forms a parameterized database. The macroscopic flow prediction module uses an improved discrete element method to predict the flow and mixing behavior of the powder particle group in the target sampling area. The improved discrete element method includes an adaptive time step algorithm and a domain decomposition parallel algorithm. The microscopic force calculation module and the macroscopic flow prediction module perform data interaction and collaborative calculation based on a multi-scale bridging calculation method of an adaptive coupling region. a third unit, configured to input the raw material parameters, target formula information, and production environment parameters into the numerical simulation unit for calculation, correct the weight ratio of each component in the target formula information according to the calculation results, and convert the corrected weight ratio into a quantitative feeding control parameter; Utilizing real-time weight detection and material blockage identification results, the quantitative feeding control parameters are dynamically adjusted through a self-learning feedback compensation algorithm based on a regression network; The fourth unit is used to start the mixing equipment to batch the ingredients based on the dynamically adjusted quantitative feeding control parameters, collect the mixing process data and input it into the numerical simulation unit to calculate the mixing uniformity index in real time, automatically stop mixing when the target requirements are met, detect the mixed finished product, compare the target formula information and generate a batching process report.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images