Digital twin design of pneumatic classifier and its key technology

The design scheme for a digital twin of the pneumatic classifier is proposed, which is associated with the physical entity. The key technologies such as visualization design and data acquisition are analysed in order to realize the digital management of the pneumatic classifier and operating process and monitor flow field characteristics and particle motion based on data analysis according to the digital twin. It aims to realize real-time monitoring, simulation and prediction of the classification performance of the complex pneumatic classification systems, and provide practical guidance for the optimal design of pneumatic classifier.


Introduction
With the rapid development of modern engineering technology, the demands for powder raw materials and products increase yearly.The pneumatic classifier has the advantages of simple structure, convenient operation, adjustable parameters and controllable product particle size, and it is one of the key equipment of the powder preparation system.For a long time, domestic and foreign scholars have explored and researched on the pneumatic classification and obtained some achievements.The pneumatic classification is usually modelled using theoretical model and empirical model [1][2][3].The theoretical model is mainly established by analysing the forces acting on particles and the particle motion laws in the classifier [4][5][6].In order to derive theoretical models, researchers often assume that the particle concentration is very small and the particles do not interfere with each other.In fact, the classification process of particles not only includes the overall effect of particle group motion, but also the interaction between particles.Although the simplified theoretical model provides a certain basis for exploring the classification mechanism, the particles' sizes are mainly considered in study of particle motion law and the interactions between particles are not considered, such as inter-particle collision, inter-particle forces and aggregation and dispersion for the particles.Therefore, the current theoretical model cannot reflect the real classification process, and it is mostly used for qualitative analysis.In contrast, the empirical model is mainly based on experimental data, and the regression model is constructed based on the black-box theory [7][8][9][10][11].The fitting accuracy of the empirical model depends on numbers of the samples, and the more parameters to be determined and the better the fit of the model, the greater the workload of the experiment are.With the development of Computational Fluid Dynamics technology, numerical simulation has become a powerful tool to study the flow field distribution of pneumatic classifier and improve its structure [12][13][14][15][16][17][18][19][20][21].
Although theoretical models, empirical models and numerical simulations have theoretical guiding significance for the design, optimization and manufacture of the pneumatic classification systems, the pneumatic classification models of digital space and the pneumatic classification system in physical space do not form mapping relationship between virtual resources and real resources.Monitoring, simulating and predicting the performance of complex classification systems in real time cannot be implemented directly.In response to the above limitation, the paper focuses on the digital twin technology to realize the intelligent of pneumatic classification process.The combination of data-driven and model-driven digital twin technology has the characteristics of high integration between virtual and real resources, which provides new ideas for solving the above problems.In fact, digital twin technology is already widely used in various industries [22][23][24][25][26][27][28].In this paper, the digital twin of the pneumatic classifier is designed and the key technologies are discussed by combining the theoretical analysis, numerical simulation and experimental analysis of powder classification in order to realize the remote monitoring of the pneumatic classifier by the two-way collaboration of data between the physical model and the digital model.The digital model of the classifier can be verified and corrected and the accuracy of the model can be improved according to the monitoring information and material experiments of the physical system, which provides a new idea for the digitalization and intelligence of the pneumatic classifier.

Digital twin design of pneumatic classifier
The digital twin of the pneumatic classifier is a four-layer architecture, including the physical entity layer, the data transmission layer, the visual interaction layer, and the decision-making layer.These four layers are managed, maintained, expanded and human-computer interacted by the digital twin service platform, which can be in the form of tool groups, module engines, application software, APP, etc., or personalized design according to their own needs.The physical entity layer includes physical models, sensors, PIV (Particle Image Velocimetry) flow field test systems, dry laser particle size analysers and powder physical property testers.The physical model is the pneumatic classifier.Its main working principle and structure are described in the literatures [29,30].The data transmission layer includes the operating status data of the classifier, the sensor data and so on, and these multi-source heterogeneous data is integrated and processed to obtain the twin data and realize data sharing and interaction.The twin data is uploaded to the cloud server for storage.The visual interaction layer is connected to the server through network communication protocols such as WebSocket, which realizes the visualization of digital models and human-computer interaction.The visual interaction layer integrates the visual realtime status monitoring of classifier, the management of the classification process, the display of numerical simulation results, the motion of powder particles, and the display of experimental data results to achieve complete mapping of powder classification.The main body of the decision-making layer is people.Through the information obtained in the visual interaction layer, the decision-making layer predicts and regulates the classification performance according to the actual production demand.The designed digital twin model of the pneumatic classifier is shown in figure 1.The specific implementation process is as follows: the data obtained from the classifier, sensors and measurement equipment is transferred to the server after integration.The communication protocols are established and the data is transmitted to a visual interface in the browser.Gas phase numerical simulation, gas-solid two-phase numerical simulation, direct Monte Carlo particle collision model and particle population balance model are cascaded and integrated in a visual interface.When the obtained experimental data is fed back to the visual interface, the experimental data and the model data are compared.If the calculated value of the digital model is consistent with the experimental data, it indicates that the digital model describes the particle motion characteristics in the pneumatic classifier accurately.If the difference between them is large, the digital model should be corrected and optimized on the basis of analysing the reason of the errors.The visual interaction layer is regarded as a digital mirror of the physical entity layer, and the particles classification process is reproduced in the virtual space.The various data of the physical and visual interaction layers is saved in related databases.By monitoring the production process, the decision-making layer can predict and regulate the classification performance according to the actual production needs to optimize and improve the structure and system of the classifier.

Visual model creation and interaction
In order to realize the visual interactive interface of the model, it is necessary to build a scene that can realize the visualization of the three-dimensional model.At present, there are two categories of commonly used interactive software: one is represented by Unity3D, and the other is represented by WebGL.Due to the complex structure of the classifier, the mature commercial software SOLIDWORKS can be selected for three-dimensional modelling and assembly.However, the three-dimensional model file of the classifier is large and it is difficult to load directly in the Web page, and the model components need to be simplified.3DMAX can be selected to lightweight the model, so that the three-dimensional model can be simplified as much as possible under the condition of ensuring the realism, reducing the requirements for computer performance.After the scene is built and the classifier model is imported, it is necessary to interact with the twin, display the model part information, and control the digital model.
The parts are identified using Raycaster module.At the same time, in order to observe the characteristics of the internal flow field of the classifier intuitively, it is necessary to set the button to appear the parts as transparent and make the motion of particles more intuitive.

Acquisition of experimental data
In the visual interactive interface, the experimental data fed back from the physical entity layer is used to correct and improve the digital model, so that the numerical simulation and the physical experimental system are synergistic and integrated.The representative powder materials such as alumina powder, quartz powder and silicon carbide powder are selected for classification experiments.The bulk density, vibration density, packing angle and powder fluidity of the above powders can be measured using the powder comprehensive characteristics tester and the data is transmitted to the server to compare and analyse the influence of different powder properties on the classification performances and particle trajectories.The operating parameters of classification system in the laboratory, such as inlet air velocity, rotor cage speed and feeding speed, are adjusted to carry out material classification experiments.Under different operating conditions, fine powder products with different particle size distributions are obtained, and the particle size distributions of raw materials, coarse powder and fine powder are measured by laser particle size analyser.According to the measurement results of particle size distributions, the key performance indices of the classification experiment, such as the classification efficiency, classification precision, cut size, bypass value and fine powder yield, are calculated to obtain experimental data and upload the data to the server.PIV is used to measure the flow field in the pneumatic classifier to verify the results of numerical simulation, and to modify and improve the digital model, so as to prepare for the numerical calculation of gas-solid two-phase flow.The experimental devices for the flow field test platform can be built on the classification system for measuring the flow field in the classifier.Finally, the data is transmitted to the server for processing.

Acquisition of numerical simulation data
Computational Fluid Dynamic (CFD) is used to simulate the flow field in the pneumatic classifier.Comparing the measured flow field distribution with the simulation results, the calculation model and boundary conditions of continuous phase simulation and gas-solid two-phase flow simulation are corrected and improved, so as to improve the accuracy of digital model calculation.The motion process of particles can be decoupled into two independent processes, free motion and collision, under Lagrange coordinates.In the free motion of the sampled particles, the collision between the sampled particles is not considered, but the collision between the sampled particles and the wall is considered.The particles are mainly acted by the airflow drag force, centrifugal force and the gravity of themselves during this period.Thus, a large number of particles are separated according to the particle size.The discrete particle model using the Euler-Lagrangian method treats the particle phase as the discrete phase and the fluid phase as the continuous phase.The initial positions and velocities of the simulated particles and the time step required for simulation can be set; the calculation area can be meshed; and the particle motion equation within the time step can be solved to obtain the particles' motion trajectories.However, a huge number of particles will lead to an increase of the number of particle motion equations, resulting in an increase of the computations.In order to reduce the computation, the Monte Carlo method can be used to calculate the particle collision probability, simulate the evolution process of particle collision and agglomeration, and study the unsteady changes of particle size distribution, particle number and average particle size during the process of ultrafine particle agglomeration.The discrete phase model is used to simulate the motion trajectories of particles, the change of particle velocity and the average residence time of particles in the gas-solid two-phase flow field.The motion laws of the coarse and fine particles in the flow field can be obtained.The motion trajectories of particles at different moments in the flow field using CFD software are shown in figures 2 (a), (b), and (c).The numerical simulation data will be integrated into the visual interface.

Collection of operating data
The stable operation of the classifier is the most basic condition for the powder classification process, which requires monitoring of the key operating data of the classifier, including inlet air velocity, rotor cage speed, feeding speed and pressure drop between the air inlet and the air outlet.By measuring the pressure drop of the air inlet and air outlet, the energy loss of the classifier can be obtained.The pressure drop can be measured through a micro-differential pressure sensor.The sensor's data can be uploaded to a remote server via the 485communication protocol.The inlet air velocity is measured by the thermal wind velocity sensor and then it is transmitted to the server.The rotor cage speed can be controlled by a servo motor and the data will be transmitted to the server.The raw materials to be classified are fed into the classifier by the screw feeder, and the feeding speed can be converted from the frequency of the motor.A dry laser particle size analyser is used to measure the particle size.The measurement process does not require any dispersants and solvents.The particles are fully dispersed by the dispersion device.The physical entity layer composed of the above physical experiment and test system and the data acquisition system and the visual interaction layer composed of the virtual experimentation, numerical model, and random computing model are shown in figure 3.

Conclusions
The digital twin design for the pneumatic classifier is proposed combining theoretical analysis, numerical simulation and material classification experiment of powder classification, and its related key technologies of visualization model creation and data collection are discussed.On the one hand, the analysis of the motion characteristics of particles in the flow field, the analysis of particle collision and agglomeration behavior in the classifier, the agglomeration inhibition and the optimization of the equipment structure are the main research lines.On the other hand, the calibration of the physical properties of powder, particle size determination, the flow field measurement, and the material classification operation are used as the experimental solution.The digital model of the classifier is verified and corrected by the constructed pneumatic classification digital twin through the two-way collaboration of data between the digital model and the physical experiment.The high precision digital model and refined simulation are used for simulating the particle classification process and predicting the classification performance to design and optimize the classifier.The iterative optimization can be implemented based on the digital twin.

Figure 1 .
Figure 1.Digital twin model of the pneumatic classifier.The specific implementation process is as follows: the data obtained from the classifier, sensors and measurement equipment is transferred to the server after integration.The communication protocols are established and the data is transmitted to a visual interface in the browser.Gas phase numerical simulation, gas-solid two-phase numerical simulation, direct Monte Carlo particle collision model and particle population balance model are cascaded and integrated in a visual interface.When the obtained experimental data is fed back to the visual interface, the experimental data and the model data are compared.If the calculated value of the digital model is consistent with the experimental data, it indicates that the digital model describes the particle motion characteristics in the pneumatic classifier accurately.If the difference between them is large, the digital model should be corrected and optimized on the basis of analysing the reason of the errors.The visual interaction layer is regarded as a digital mirror of the physical entity layer, and the particles classification process is reproduced in the virtual space.The various data of the physical and visual interaction layers is saved in related databases.By monitoring the production process, the decision-making layer can predict and regulate the classification performance according to the actual production needs to optimize and improve the structure and system of the classifier.

Figure 2 .
Figure 2. The motion trajectories of the particles at different moments.

Figure 3 .
Figure 3. Virtual and real interaction of pneumatic classifier.