A test device and method for testing multiphase particle collision in a spray fluidized bed

Abstract

The present invention discloses a test device and method for testing multiphase particle collisions in a spray fluidized bed. The test device comprises a quasi-two-dimensional spray particle fluidization system, a high-speed photography system, and a control system. A multiphase particle collision test is performed in the quasi-two-dimensional spray particle fluidization system. Images are captured using the high-speed photography system. By establishing and verifying a CFD-DEM spray fluidized bed simulation, preparing a neural network training data set and training image segmentation, and performing event classification, a deep neural network is established. Finally, the particle collision velocity and collision frequency in the entire fluidized bed are accurately measured from the high-speed photography particle images.

Description

Technical Field

[0001] The present invention belongs to the field of spray fluidized bed particle coating and agglomeration, and in particular relates to a test device and method for testing the collision velocity and frequency of particles in a spray fluidized bed multiphase system, specifically to a test device and method for testing the particle collision velocity, particle collision frequency, particle velocity and acceleration, and particle concentration in a spray fluidized bed multiphase system, as well as a training method for the required neural network. Background Art

[0002] The spray fluidized bed is a complex, multi-scale, gas-solid-liquid multiphase process involving particle fluidization, liquid atomization, and drying. Particle formation within a spray fluidized bed is widely used in agglomeration granulation and particle coating (coating and encapsulation) processes in the pharmaceutical, food, agriculture, powder catalysts, cosmetics, and energy particle fields, and is crucial to food production safety, food and drug safety, and energy security. A deep understanding of the multiphase particle collision characteristics of a spray fluidized bed is key to improving the quality of particle-formed products, reducing energy consumption during the spray fluidized bed production process, and achieving intelligent production. Therefore, experimental testing of particle collision velocity and collision frequency is of great significance.

[0003] At present, non-invasive testing of particles in a spray fluidized bed multiphase system mainly focuses on the velocity and concentration of multiple particles or the motion trajectory of a single particle. Testing methods include particle image velocimetry (PIV), particle tracking velocimetry (PTV), magnetic particle tracking (MPT), positron emission particle tracking (PEPT), magnetic resonance imaging (MRI), and X-ray particle tracking velocimetry (XPTV). Among them, particle tracking velocimetry (PTV) based on the results of high-speed photography of a visualized fluidized bed can restore the particle motion trajectory, obtain particle collision events, and analyze the collision velocity and collision frequency of particles. However, its algorithm is complex and has low applicability under different flow states. Experimental measurement of particle collision velocity and collision frequency has always been a bottleneck in multiphase flow testing. Fast and efficient measurement methods are in urgent need of development, and measurement accuracy needs to be improved. Summary of the Invention

[0004] In response to the shortcomings of existing testing technologies and methods, the present invention provides a test device and method for testing multiphase particle collisions in a spray fluidized bed. The technical purpose is to capture particle collisions in a multi-type spray fluidized bed by high-speed photography at multiple positions, establish and verify CFD-DEM spray fluidized bed simulation in real time, prepare neural network training data sets and training image segmentation and event classification, and finally accurately measure the particle collision velocity and collision frequency in the entire fluidized bed from high-speed photography particle images.

[0005] The technical solutions adopted in the present invention are as follows:

[0006] A method for testing multiphase particle collision in a spray fluidized bed comprises the following steps:

[0007] S1. Establish a CFD-DEM simulation model and use it to simulate particle motion in a multiphase system. The simulation predicts the particle motion characteristics, particle collision velocity, and spatial distribution of collision frequency, and adjusts the camera position based on the spatial distribution. Simulation data on particle motion in a multiphase system is obtained through the simulation.

[0008] S2. Using simulation data, train and test a deep neural network for pairing discrete particles from image data, a classification deep neural network for determining the number of particle collisions from image data, and a regression deep neural network for calculating collision velocity from image data.

[0009] S3. Collecting image data of actual particle motion in a multiphase system; and training and testing a convolutional neural network for particle segmentation based on the collected image data; using the trained convolutional neural network for particle segmentation, predicting the center position and particle size of each particle in each frame of the high-speed photography image, and further calculating the particle volume fraction;

[0010] S4, using the deep neural network for paired discrete particles trained in S2 and the particle center position and particle size obtained in S3, predict the pairing information of the particles in each frame of the high-speed photography image, and further calculate the instantaneous velocity, acceleration and motion trajectory of the particles;

[0011] S5. Using the classification deep neural network trained in S2 to determine the number of particle collisions from image data and the regression deep neural network trained to calculate the collision velocity from image data, as well as the instantaneous velocity and acceleration of the particles obtained in S4, the number of collisions of each particle and the particle collision velocity of each collision in each frame of high-speed photography images are predicted, and the collision frequency and collision velocity distribution of the particles are calculated.

[0012] Furthermore, S4 uses a segmentation convolutional neural network to separate the center position and particle size of different types of particles from image data, and a deep neural network to pair discrete particles from image data to predict the discrete particles P at adjacent moments. i The relative displacement of the discrete particles is calculated based on the shooting interval; further, the acceleration of the discrete particles is calculated by the change in the instantaneous velocity of the discrete particles during the shooting interval; the same particles are connected at multiple time positions to reconstruct the particle motion trajectory.

[0013] Furthermore, the particle collision frequency is defined as the number of collisions experienced by a single particle per unit time, so the particle collision frequency is calculated in S5:

[0014] f c，i =N c，i / (t i+1 -t i )

[0015] Among them, N c，i is the number of collisions per particle, t i and t i+1 are adjacent moments; through each particle collision velocity v c，j Create a histogram and calculate the collision velocity distribution.

[0016] Furthermore, the method for training and testing a deep neural network for pairing discrete particles from image data is:

[0017] Step 1) Extract particle motion characteristics at different times from the CFD-DEM simulation data in the xy dimension, including each particle number and particle center position; the extracted time interval is the interval between the two images, and the units are converted based on the actual size of each pixel;

[0018] Step 2) Establish a Voronoi diagram based on the center position of the discrete particles and determine the P of each discrete particle. i The corresponding Thiessen polygons and the adjacent discrete particles P corresponding to their adjacent Thiessen polygons i，j ;

[0019] Step 3) Use the particle number information in DEM to determine the same particle P at different times i , extract t i discrete particles P i and P i，j The relative position of i+1 discrete particles P i and P i，j The relative position of t i and t i+1 Time P i The data were normalized by the equivalent particle diameter (the geometric mean of all extracted particle sizes) and set as the matching success category; extract t i discrete particles P i and P i，j The relative position of the relative position, t i+1 discrete particles P i，j and the relative position of its neighboring particles, and t i and t i+1 Time P i and P i，j The relative position of the particles is determined, and the data is standardized using the equivalent particle diameter and set as the unsuccessful matching category; a sufficient amount of data is extracted to form a training data set;

[0020] Step 4) randomly selecting a portion of the training data set obtained in step 3) and training a deep classification neural network for paired discrete particles. The number of hidden layers and the number of neurons in each layer can be adjusted as needed;

[0021] Step 5) Using the remaining data in the training data set obtained in step 4), the deep neural network of step 4) is tested;

[0022] Step 6) Repeat steps 4) and 5) until the prediction accuracy of the deep neural network reaches the set requirements.

[0023] Furthermore, the method for training and testing a classification deep neural network for determining the number of particle collisions from image data is as follows:

[0024] Step 1) Extract particle motion characteristics at different moments from the CFD-DEM simulation data in the xy dimension, including each particle number, particle center position, particle instantaneous velocity, and particle acceleration. The extracted time interval is the interval between the two images, and the unit is converted based on the actual size of each pixel. The number of particle-particle collisions is extracted, and the extracted time interval is 1 / 20 of the shortest particle collision duration. A single particle collision is defined as the period from physical contact to physical separation of the particles.

[0025] Step 2) Establish a Voronoi diagram based on the center position of the discrete particles and determine the P of each discrete particle. i The corresponding Thiessen polygons and the adjacent discrete particles P corresponding to their adjacent Thiessen polygons i，j ;

[0026] Step 3) Use the particle number information in DEM to extract t i discrete particles P i and P i，j The relative position, relative velocity and relative acceleration of i+1 discrete particles P i and P i，j The relative position, relative velocity and relative acceleration of the particles are normalized by the equivalent particle diameter; i to t i+1 Time P i The number of collisions that occurred and the number of particles that collided with them; different categories were set according to the number of collisions, and sufficient data was extracted to form a training data set;

[0027] Step 4) Randomly select a portion of the training data set obtained in step 3) to train a classification deep neural network that measures the number of particle collisions. The number of hidden layers and the number of neurons in each layer of each neural network can be adjusted according to needs.

[0028] Step 5) Using the remaining data in the training data set obtained in step 3) to test the deep neural network in step 4);

[0029] Step 6) Repeat steps 4) and 5) until the prediction accuracy of the deep neural network reaches the set requirements.

[0030] Furthermore, the method for training and testing a regression deep neural network for calculating collision velocity from image data is:

[0031] Step 1) In the xy dimension, extract the particle motion characteristics at different moments from the CFD-DEM simulation data, including each particle number, particle center position, particle instantaneous velocity, and particle acceleration, and convert the units using the actual size of each pixel point; extract the velocity of each particle-particle collision (i.e., the relative velocity when two particles come into physical contact) (32), and the extracted time interval is 1 / 20 of the shortest particle collision duration, and convert the collision velocity into units using the actual size of each pixel point;

[0032] Step 2) Establish a Voronoi diagram based on the center position of the discrete particles and determine the P of each discrete particle. i The corresponding Thiessen polygons and the adjacent discrete particles P corresponding to their adjacent Thiessen polygons i，j ;

[0033] Step 3) Use the particle number information in DEM to extract t i discrete particles P i and P i，j The relative position, relative velocity and relative acceleration of i+1 discrete particles P i and P i，j The relative position, relative velocity and relative acceleration of the particles are normalized by the equivalent particle diameter; i to t i+1 Moment and P i The collision particle number and the collision speed of each particle are extracted, and a sufficient amount of data is generated to form a training data set;

[0034] Step 4) Randomly select a portion of the training data set obtained in step 3) to train a regression deep neural network for measuring particle collision velocity. The number of hidden layers and the number of neurons in each layer of each neural network can be adjusted according to needs.

[0035] Step 5) Using the remaining data in the training data set obtained in step 3) to test the deep neural network in step 4);

[0036] Step 6) Repeat steps 4) and 5) until the prediction accuracy of the deep neural network reaches the set requirements.

[0037] Furthermore, the method for training and testing a convolutional neural network for particle segmentation from image data is:

[0038] Step 1) converting the raw image data of particle fluidization captured by a high-speed camera into a single-channel grayscale image;

[0039] Step 2) Segmenting the image data of discrete particles of the same type from the grayscale image obtained in step 1), including the particle image mask, the particle rectangular border, the pixel matrix within the particle border, and the particle equivalent particle size;

[0040] Step 3) performing normalized cross-correlation operations on the pixel matrix within the particle frame obtained in step 2) and the two-dimensional matrix of the multi-frame grayscale image data obtained in step 1) to obtain a corresponding correlation coefficient matrix;

[0041] Step 4) determining a plurality of peak positions corresponding to the center positions of the particles in the original image from the correlation coefficient matrix obtained in step 3) by setting a dynamic threshold;

[0042] Step 5) automatically segmenting different types of particles in each frame of the particle fluidization original image based on the particle center position obtained in step 3) and the discrete particle image data in step 2), and manually removing interfering particles;

[0043] Step 6) writing the qualified particles segmented in each frame of the original image in step 5) into training data, including the particle mask corresponding to the original image data, the particle type based on the equivalent particle size, and the particle rectangular border;

[0044] Step 7) Repeat steps 2) to 6) to obtain training data of different types of particles in each frame of the original image, save the corresponding names, and form the final training data set;

[0045] Step 8) randomly selecting a portion of the data from the training dataset obtained in step 7) and training an instance segmentation convolutional neural network that can achieve particle segmentation;

[0046] Step 9) Using the remaining data in the training data set obtained in step 7), testing the instance segmentation convolutional neural network in step 8);

[0047] Step 10) Repeat steps 8) and 9) until the prediction accuracy of the instance segmentation convolutional neural network reaches the set requirements.

[0048] A test device for testing multiphase particle collision in a spray fluidized bed, comprising a quasi-two-dimensional spray particle fluidization system, a high-speed photography system and a control system;

[0049] The two-dimensional spray particle fluidization system includes: a quasi-two-dimensional visualized particle fluidization chamber, a fluidizing gas pressure stabilizing chamber, and a particle collector; a spliced ​​distribution plate is provided between the fluidizing gas pressure stabilizing chamber and the quasi-two-dimensional visualized particle fluidization chamber; a fluidization chamber baffle is provided within the quasi-two-dimensional visualized particle fluidization chamber, and the fluidization chamber baffle divides the quasi-two-dimensional visualized particle fluidization chamber into multiple cavities that do not interfere with each other; a dual-fluid nozzle and a sensor are provided within the quasi-two-dimensional visualized particle fluidization chamber;

[0050] The high-speed photography system includes two high-speed cameras, a camera moving unit, and a high-speed camera synchronizer. The high-speed cameras are mounted on the camera moving unit, and the signals of the two high-speed cameras are connected to the high-speed camera synchronizer.

[0051] The control system uses a computer, which is connected to the two-dimensional spray particle fluidization system and the high-speed photography system by signal. The control system adjusts the position of the high-speed photography system according to the particle velocity, particle volume fraction, particle collision velocity and particle collision frequency obtained by the spray fluidized bed multiphase system particle collision test method.

[0052] Furthermore, when the time average and standard deviation of various variables in a certain window, such as particle velocity, particle volume fraction, particle collision velocity and particle collision frequency, change less than the preset value, the camera moving unit is adjusted to obtain a new window position until the position where particles move in the fluidization chamber is covered.

[0053] Furthermore, the high-speed cameras are two HCCD high-speed cameras, and the high-speed camera synchronizer is connected to the first HCCD high-speed camera and the second HCCD high-speed camera respectively.

[0054] Beneficial effects of the present invention:

[0055] 1) The present invention provides a test method for testing the collision velocity and frequency of multiphase particles in a spray fluidized bed, and proposes a new method for quickly and effectively extracting particle masks, which can ensure the accuracy of training instance segmentation convolutional neural networks and segmenting particles from images;

[0056] 3) The present invention provides a test method for measuring the collision velocity and frequency of multiphase particles in a spray fluidized bed, using verified CFD-DEM simulation data to provide accurate information such as particle pairing, particle position, particle velocity, and particle collision events for training classification deep neural networks, ensuring the accuracy of the final measurement results.

[0057] 3) The experimental device and method for testing the collision velocity and frequency of multiphase particles in a spray fluidized bed described in the present invention can comprehensively study the particle collision characteristics under different types of spray fluidized beds and different spray conditions, and establish a quantitative relationship between the collision velocity and frequency and the fluidization and spray characteristic parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic structural diagram of the quasi-two-dimensional spray particle fluidization system provided by the present invention;

[0059] Figure 2 It is a schematic structural diagram of the high-speed photography system provided by the present invention;

[0060] Figure 3This is a schematic diagram of various quasi-two-dimensional visualized multiphase system structures provided by the present invention;

[0061] Figure 4 This is a schematic diagram of the structure of the stepped three-dimensional dot matrix ruler provided by the present invention;

[0062] Figure 5 is a schematic diagram of discrete particle image data provided by the present invention;

[0063] Figure 6 It is a schematic diagram of the correlation coefficient matrix provided by the present invention;

[0064] Figure 7 The Voronoi diagram provided by the present invention and the schematic diagram of the particles therein;

[0065] Figure 8 It is a schematic diagram of relevant variables in the particle motion process provided by the present invention;

[0066] Figure 9 Schematic diagram of a deep neural network for predictive classification provided by the present invention;

[0067] Figure 10 Schematic diagram of a deep neural network for predictive regression provided by the present invention;

[0068] In the figure, 1. Two-dimensional visualization particle fluidization chamber, 2. Fluidizing gas pressure stabilizing chamber, 3. Particle collector, 4. Particles, 5. Nozzle, 6. Spliced ​​distribution plate, 7. Baffle in fluidization chamber, 8. Flow controller, 9. Sensor, 10. Control system, 11. First HCCD high-speed camera, 12. xy linear electric linear guide, 13. Second HCCD high-speed camera, 14. y linear electric linear guide, 15. High-speed camera synchronizer, 16. First servo motor, 17. Second servo motor, 18. LED light source, 19. Spouted fluidized bed, 20. Wurster fluidized bed, 21. Horizontal fluidized bed, 22. Stepped three-dimensional dot matrix, 23. Particle image mask, 24. Particle rectangular frame, 25. Pixel matrix in particle frame and particle equivalent particle size, 26. Correlation coefficient matrix, 27. Voronoi diagram, 28. Discrete particle P i and P i，j The relative position of particles P i Relative position, 30. Instantaneous velocity of particles, 31. Particle-particle collision, 32. Particle-particle collision velocity, 33. Deep classification neural network, 34. Regression deep neural network. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0070] This application proposes a method for testing the collision of multiphase particles in a spray fluidized bed, comprising the following steps:

[0071] S1. Establish a CFD-DEM simulation model and use it to simulate the motion of particles in a multiphase system. This simulation can predict the particle motion characteristics, particle collision velocity, and the approximate spatial distribution of the collision frequency, which can be used as a basis for adjusting the camera position. It can also obtain simulation data of particle motion in a multiphase system for subsequent deep neural network training and testing.

[0072] More specifically, the CFD-DEM simulation model is a coupling of computational fluid dynamics (CFD) and particle-based discrete element method (DEM). Specifically, the classic Gidaspow drag model and the classic Herztian soft-sphere collision model are employed, with a capillary force model accounting for the liquid bridge forces between particles generated by spray droplets. In this embodiment, the CFD grid is a regular hexahedron with a grid edge length to particle diameter ratio of approximately 2.5. The CFD calculation time step ensures that the maximum Courant criterion number is less than 0.5. Particle physical properties such as the coefficient of restitution, Young's modulus, Poisson's coefficient, and friction coefficient are experimentally measured, and the DEM calculation time step is less than 1 / 20 of the minimum collision duration. The minimum collision duration is calculated using the Hertz collision law, with the maximum collision velocity preset to twice the fluidization velocity. The simulation time is 1 second, and the simulation results are compared with the velocity field measured by PIV. If the comparison error does not meet the set 10% requirement, the grid size and particle physical parameters are readjusted and the calculation is repeated until the required results are met. More specifically, the velocity field measured by the PIV is a particle velocity field obtained by reducing the velocity field of the particle using the existing PIV testing technology.

[0073] S2, using the simulation data obtained in S1, respectively train and test a deep neural network for pairing discrete particles from image data, a classification deep neural network for determining the number of particle collisions from image data, and a regression deep neural network for calculating collision velocity from image data, such as Figure 9 、 10 The specific methods for training and testing deep neural networks are as follows:

[0074] S2.1. A method for training and testing a deep neural network for pairing discrete particles from image data, comprising the following steps:

[0075] Step 1) Extract particle motion characteristics at different moments from the CFD-DEM simulation data based on the camera's viewport position and lens depth, including each particle number and particle center position; the extracted time interval is the interval between the two images, and the units are converted based on the actual size of each pixel;

[0076] Step 2) Create a Voronoi diagram (27) based on the center position of the discrete particles to determine the position of each discrete particle P i The corresponding Thiessen polygons and the adjacent discrete particles P corresponding to their adjacent Thiessen polygons i，j ;

[0077] Step 3) In the xy dimension, use the particle number information in the DEM to determine the same particle P at different times i , extract t i discrete particles P i and P i，j The relative position of (28), t i+1 discrete particles P i and P i，j The relative position of t i and t i+1 Time P i The relative position of (29) was used to standardize the data using the equivalent particle diameter (equivalent particle diameter refers to the geometric mean of all extracted particle sizes) and set it as the matching success category; extract t i discrete particles P i，j The relative position of the particle to its neighbor, t i+1 discrete particles P i，j and the relative position of its neighboring particles, and t i and t i+1 Time P i，j The relative position of the particles is determined, and the data is standardized using the equivalent particle diameter and set as the unsuccessful matching category; a sufficient amount of data is extracted to form a training data set;

[0078] Step 4) Randomly select 70% of the data in the training data set obtained in step 3) to train a deep classification neural network (33) for paired discrete particles. The number of hidden layers and the number of neurons in each layer can be adjusted according to needs; in this embodiment, the number of hidden layers is selected to be 10 layers.

[0079] Step 5) Using the remaining 30% of the data in the training data set obtained in step 3), the deep neural network of step 4) is tested;

[0080] Step 6) Repeat steps 4) and 5) until the prediction accuracy of the deep neural network meets the set requirements; in this embodiment, 500 consecutive time steps are randomly selected from the obtained CFD-DEM simulation data, and 1000 particles are selected in each time step, which can achieve the training accuracy requirements of the classification deep neural network, and the successful pairing rate of training and test particles reaches 98%.

[0081] Preferably, for sparse particles in the near-side wall region, two HCCD cameras can be integrated to obtain the three-dimensional coordinates of discrete particles. Step (2) can use the three-dimensional Voronoi diagram, and step (3) can select the xyz dimension.

[0082] S2.2. Training and testing a classification deep neural network method for determining the number of particle collisions from image data, comprising the following steps:

[0083] Step 1) According to the camera's viewport position and camera lens depth, the particle motion characteristics at different moments are extracted from the CFD-DEM simulation data, including each particle number, particle center position, particle instantaneous velocity (30), and particle acceleration. The extracted time interval is the shooting interval between two pictures, and the unit is converted by the actual size of each pixel point; the number of particle-particle collisions (31) is extracted, and the extracted time interval is 1 / 20 of the shortest particle collision duration. A single particle collision refers to the period from physical contact of the particles to physical separation of the particles.

[0084] Step 2) Establish a Voronoi diagram based on the center position of the discrete particles, determine the Thiessen polygon corresponding to each discrete particle Pi, and the adjacent discrete particles P corresponding to its adjacent Thiessen polygon i，j ;

[0085] Step 3) In the xy dimension, use the particle number information in the DEM to extract t i discrete particles P i and P i，j The relative position, relative velocity and relative acceleration of i+1 discrete particles P i and P i，j The relative position, relative velocity and relative acceleration of the particles are normalized by the equivalent particle diameter; i to t i+1 Time P i The number of collisions that occurred and the number of particles that collided with them; different categories were set according to the number of collisions, and sufficient data was extracted to form a training data set;

[0086] Step 4) Randomly select 70% of the data in the training data set obtained in step 3) and train a classification deep neural network (33) for measuring the number of particle collisions. The number of hidden layers in each neural network and the number of neurons in each layer can be adjusted according to needs; in this embodiment, a deep neural network for prediction classification includes 20 hidden layers.

[0087] Step 5) Using the remaining 30% of the data in the training data set obtained in step 3), the deep neural network of step 4) is tested;

[0088] Step 6) Repeat step 4) and step 5) until the prediction accuracy of the deep neural network meets the set requirements; in this embodiment, 1000 consecutive time steps are randomly selected from the obtained CFD-DEM simulation data, and 1000 particles are selected in each time step, which can achieve the training accuracy requirements of the classification deep neural network, and the successful recognition rates of training and test particle collisions reach 95% and 93%, respectively.

[0089] Preferably, the particle collision frequency can be defined as the number of collisions experienced by a single particle per unit time;

[0090] Preferably, for sparse particles in the near-side wall region, two HCCD cameras can be integrated to obtain the three-dimensional coordinates of discrete particles. Step (2) can use the three-dimensional Voronoi diagram, and step (3) can select the xyz dimension.

[0091] S2.3. Training and testing a regression deep neural network method for calculating collision velocity from image data, comprising the following steps:

[0092] Step 1) According to the camera's window position and camera lens depth, from the CFD-DEM simulation data, such as Figure 8 Extract the particle motion characteristics at different moments, including each particle number, particle center position, particle instantaneous velocity (30), and particle acceleration, and convert the units using the actual size of each pixel point; extract the velocity of each particle-particle collision (i.e., the relative velocity when two particles come into physical contact) (32), and the extracted time interval is 1 / 20 of the shortest particle collision duration, and convert the collision velocity into units using the actual size of each pixel point;

[0093] Step 2) Establish a Voronoi diagram based on the center position of the discrete particles, determine the Thiessen polygon corresponding to each discrete particle Pi, and the adjacent discrete particles P corresponding to its adjacent Thiessen polygon i，j ;

[0094] Step 3) In the xy dimension, use the particle number information in the DEM to extract t i discrete particles P i and P i，jThe relative position, relative velocity and relative acceleration of i+1 discrete particles P i and P i，j The relative position, relative velocity and relative acceleration of the particles are normalized by the equivalent particle diameter; i to t i+1 Moment and P i The collision particle number and the collision speed of each particle are extracted, and a sufficient amount of data is generated to form a training data set;

[0095] Step 4) Randomly select 70% of the data in the training data set obtained in step 3) to train a regression deep neural network (34) for measuring particle collision velocity. The number of hidden layers in each neural network and the number of neurons in each layer can be adjusted according to needs; in this embodiment, a prediction regression deep neural network with 10 hidden layers is used.

[0096] Step 5) Using the remaining 30% of the data in the training data set obtained in step 3), the deep neural network of step 4) is tested;

[0097] Step 6) Repeat step 4) and step 5) until the prediction accuracy of the deep neural network meets the set requirements; in this embodiment, 1000 consecutive time steps are randomly selected from the obtained CFD-DEM simulation data, and 1000 particles are selected in each time step, which can achieve the training accuracy requirements of the classification deep neural network, and the average error of the training and test particle collision velocities is less than 4%.

[0098] Preferably, for sparse particles in the near-side wall region, two HCCD cameras can be integrated to obtain the three-dimensional coordinates of discrete particles. Step (2) can use the three-dimensional Voronoi diagram, and step (3) can select the xyz dimension.

[0099] S3. Based on the experimental apparatus constructed above, image data of actual particle motion within the experimental apparatus is collected. A convolutional neural network for particle segmentation is trained and tested based on this collected image data. Using the trained convolutional neural network for particle segmentation, the center position and particle size of each particle in each frame of the high-speed photography are predicted, and the particle volume fraction can be further calculated.

[0100] More specifically, the instance segmentation convolutional neural network can be used to separate the center positions, equivalent particle sizes and volumes of different types of particles from each frame of image data, and further calculate the global and local concentrations (particle volume fractions) and particle size distributions of different types of particles in the quasi-two-dimensional multiphase system at each moment.

[0101] The method for training and testing a convolutional neural network for segmenting particles from image data is as follows:

[0102] Step 1) converting the raw image data of particle fluidization captured by the HCCD high-speed camera into a single-channel grayscale image;

[0103] Step 2) manually segmenting the image data of discrete particles of the same type from the grayscale image obtained in step 1), including the particle image mask (23), the particle rectangular border (24), the pixel matrix within the particle border, and the particle equivalent particle size (25), as shown in FIG. Figure 5 ;

[0104] Step 3) Perform normalized cross-correlation operations on the pixel matrix within the particle frame obtained in step 2) and the two-dimensional matrix of the multi-frame grayscale image data obtained in step 1) to obtain the corresponding correlation coefficient matrix (26), as shown in FIG. Figure 6 ;

[0105] Step 4) determining a plurality of peak positions corresponding to the center positions of the particles in the original image from the correlation coefficient matrix obtained in step 3) by setting a dynamic threshold;

[0106] Step 5) automatically segmenting different types of particles in each frame of the particle fluidization original image based on the particle center position obtained in step 4) and the discrete particle image data in step 2), and manually removing interfering particles;

[0107] Step 6) writing the qualified particles segmented in each frame of the original image in step 5) into training data, including the particle mask corresponding to the original image data, the particle type based on the equivalent particle size, and the particle rectangular border;

[0108] Step 7) Repeat steps 2) to 6) to obtain training data of different types of particles in each frame of the original image, save the corresponding names, and form the final training data set;

[0109] Step 8) randomly selecting 70% of the data in the training dataset obtained in step 7) and training an instance segmentation convolutional neural network that can achieve particle segmentation;

[0110] Step 9) Using the remaining 30% of the data in the training dataset obtained in step 7), test the instance segmentation convolutional neural network in step 8);

[0111] Step 10) Repeat steps 8) and 9) until the prediction accuracy of the instance segmentation convolutional neural network reaches the set requirement;

[0112] In this embodiment, a Mask R-convolutional neural network was selected to segment particles from image data. 200 frames of high-speed photography images were randomly selected, and 100 particles were extracted from each frame to form a training data set. This was able to meet the training accuracy requirements of the Mask R-convolutional neural network, with the training and test particle segmentation accuracies reaching 95% and 93%, respectively, and the average errors in the center positions of the training and test particles were both less than 5%.

[0113] The equivalent particle size d of each particle e，i and volume Further, according to different query areas (global or local), the particles in the area are counted and the equivalent particle size d e，i Create a histogram and calculate the particle size distribution;

[0114] Further calculate the particle volume fraction (particle concentration) according to different query areas (global or local) Where A is the area of ​​the region and N is the total number of particles contained in the region.

[0115] S4. Using the deep neural network of paired discrete particles trained in S2 and the particle center position and particle size obtained in S3, the pairing information of the particles in each frame of high-speed photography is predicted, and the instantaneous velocity, acceleration and motion trajectory of the particles are further calculated.

[0116] The segmentation convolutional neural network is used to separate the center position and particle size of different types of particles from the experimental image data, and the deep neural network can predict the discrete particles P at adjacent moments by pairing discrete particles from the image data. i The relative displacement of the discrete particles is used to calculate the instantaneous velocity of the discrete particles according to the shooting interval; further, the acceleration of the discrete particles is calculated by the change in the instantaneous velocity of the discrete particles during the shooting interval; by connecting the same particles at multiple time positions, the particle motion trajectory can be reconstructed.

[0117] S5. Using the classification deep neural network trained in S2 to determine the number of particle collisions and the regression deep neural network that calculates the collision velocity from the image data, as well as the instantaneous velocity and acceleration of the particles obtained in S4, the number of collisions N for each particle in each frame of the high-speed photography image is predicted. c，i and the particle collision velocity v of each collision c，j , the particle collision frequency f can be further calculated c，i =N c，i / (t i+1 -t i ), the particle collision frequency is defined as the number of collisions experienced by a single particle per unit time; the particle collision velocity v c，j Establish a histogram and calculate the collision velocity distribution; t i and t i+1 are adjacent moments.

[0118] A test device for testing the collision of multiphase particles in a spray fluidized bed Figure 1 and Figure 2 As shown, it includes a quasi-two-dimensional spray particle fluidization system, a high-speed photography system, a control system 10 and a computing and analysis system;

[0119] The two-dimensional spray particle fluidization system includes: a quasi-two-dimensional visual particle fluidization chamber 1, a fluidizing gas pressure stabilization chamber 2, and a particle collector 3. More specifically:

[0120] like Figure 1 As shown, a coordinate system is established with the quasi-two-dimensional visualized particle fluidization chamber 1, with the two planes of the quasi-two-dimensional visualized particle fluidization chamber 1 parallel to the xy plane as the main planes, and the two planes parallel to the yz plane as the side planes.

[0121] The particle collector 3 is arranged on the top of the quasi-two-dimensional visualization particle fluidization chamber 1 . The particle collector 3 is communicated with the interior of the quasi-two-dimensional visualization particle fluidization chamber 1 and is used to capture particles 4 escaping from the quasi-two-dimensional visualization particle fluidization chamber 1 .

[0122] The fluidizing gas pressure-stabilizing chamber 2 is disposed at the bottom of the quasi-two-dimensional visualized particle fluidizing chamber 1. The fluidizing gas pressure-stabilizing chamber 2 is connected to the fluidizing gas supply unit. A spliced ​​distribution plate 6 is disposed between the fluidizing gas pressure-stabilizing chamber 2 and the quasi-two-dimensional visualized particle fluidizing chamber 1. In this embodiment, the spliced ​​distribution plate 6 is formed by splicing a sintered plate and a fish-scale plate. The sintered plate is a porous medium. The fish-scale plate also has holes arranged in a fish-scale pattern. The pore size on the sintered plate is smaller than that on the fish-scale plate, and the pores on the sintered plate are more densely distributed. By splicing the sintered plate and the fish-scale plate, the airflow can be vented at different speeds and directions, so that the particles can be blown to different heights and angles by the airflow, forming a circular motion. Figure 1 The denser ones are metal sintered plates, which have porous media inside and can adjust the gas outflow speed by adjusting the porosity; the sparser fish scale plates can control the opening size and angle to adjust the airflow speed.

[0123] A fluidization chamber baffle 7 can be provided in the quasi-two-dimensional visualization particle fluidization chamber 1 to divide the quasi-two-dimensional visualization particle fluidization chamber 1 into multiple cavities that do not interfere with each other, and to isolate the particles and gases in the separated cavities from each other.

[0124] Furthermore, in this embodiment, a baffle 7 is provided in the fluidization chamber on the yz plane. The vertically arranged baffle 7 divides the middle portion of the quasi-two-dimensional visualization particle fluidization chamber 1 into two parallel chambers.

[0125] Furthermore, in this embodiment, since the baffle 7 in the fluidization chamber divides the quasi-two-dimensional visualized particle fluidization chamber 1 into two non-interfering chambers, if the bottoms of the two separated chambers correspond to the sintered plate and the fish-scale plate respectively, since the porosity, opening size and direction of the holes corresponding to the sintered plate and the fish-scale plate are different, the fluidization degree of the particles corresponding to the two chambers, the collision intensity, etc. are changed.

[0126] A dual-fluid nozzle 5 is disposed at the top of the quasi-two-dimensional visualization particle fluidization chamber 1. Dual-fluid nozzle 5 is connected to a spray gas supply unit and a spray liquid supply unit via pipes. In this embodiment, the ambient humidity within the quasi-two-dimensional visualization particle fluidization chamber 1 is adjusted by adjusting the dual-fluid nozzle 5.

[0127] Furthermore, in the present application, the type of fluidized bed can be changed by changing the positions of the two-fluid nozzle 5, the spliced ​​distribution plate 6, and the baffle 7 in the fluidization chamber 1 in the quasi-two-dimensional visualization particle fluidization chamber; Figure 3 A spouted fluidized bed 19, a Wurster fluidized bed 20 and a horizontal fluidized bed 21 are shown.

[0128] The sensor 9 is arranged on the fluidizing gas pressure stabilizing chamber 2 and the quasi-two-dimensional visualization particle fluidizing chamber 1. The sensor 9 includes a temperature sensor, a pressure sensor, and a humidity sensor, and a sealing mechanism is required to be used for sealing at the installation position of the sensor 9.

[0129] In this embodiment, the fluidizing gas supply unit, the spray gas supply unit and the spray liquid supply unit include corresponding storage units, which are connected to the fluidizing gas pressure stabilizing chamber 2 and the dual-fluid nozzle 5 through pipelines; and a flow controller 8 is installed on the pipeline for correspondingly adjusting the flow rates of the fluidizing gas, spray gas and spray liquid.

[0130] The high-speed photography system includes: a high-speed camera, a camera moving unit, and a high-speed camera synchronizer 15; in this embodiment, the high-speed camera uses two HCCD high-speed cameras, namely a first HCCD high-speed camera 11 and a second HCCD high-speed camera 13;

[0131] The camera moving unit includes an xy linear electrically controlled linear guide 12 and a y linear electrically controlled linear guide 14; a first HCCD high-speed camera 11 is installed on the xy linear electrically controlled linear guide 12, and a second HCCD high-speed camera 13 is installed on the y linear electrically controlled linear guide 14; and the xy linear electrically controlled linear guide 12 and the y linear electrically controlled linear guide 14 are respectively provided with a first servo motor 16 and a second servo motor 17; the servo motors are used to drive the corresponding high-speed cameras to move along the guide rails, thereby adjusting the positions of the two high-speed cameras.

[0132] The high-speed camera synchronizer 15 is signal-connected to the first HCCD high-speed camera 11 and the second HCCD high-speed camera 13 , and synchronously controls the shooting of the two HCCD cameras.

[0133] The LED light source 18 disposed on the xy linear electric-controlled linear guide rail 12 and the y linear electric-controlled linear guide rail 14 provides uniform lighting for the entire system.

[0134] In this embodiment, the position of the first HCCD high-speed camera 11 can be adjusted in the x and y directions, and its window is on the xy plane, and the necessary dot matrix LED light source is used on the same side of the camera for shooting; the position of the second HCCD high-speed camera 13 can be adjusted in the y direction, and its window is on the yz plane, and the necessary dot matrix LED light source is used on the same side of the camera for shooting; in the near wall area, the three-dimensional motion trajectory of sparse particles can be measured through the synchronous control of the two HCCD cameras; the two cameras use the same time resolution for continuous shooting.

[0135] Preferably, the particle velocity measurement requires that the interval between two pictures taken by the HCCD camera be less than the ratio of 10 to the estimated maximum particle velocity;

[0136] Preferably, the particle collision velocity measurement requires that the interval between two pictures taken by the HCCD camera be approximately equal to the shortest particle collision duration.

[0137] Preferably, if Figure 4 The two HCCD cameras are calibrated using a stepped three-dimensional dot matrix scale (22) determined by dot diameter, dot spacing, and step height to obtain the actual size Sf of each pixel;

[0138] Preferably, the actual dot matrix scale midpoint diameter is similar to the particle equivalent diameter;

[0139] Preferably, the ratio of the particle equivalent diameter to the actual size of a single pixel should be at least greater than 8;

[0140] Preferably, the spherical particles are fixed on different steps of the dot matrix scale, and the relationship between the light intensity on the particle surface and the distance in the z direction can be calibrated.

[0141] The control system 10 is respectively connected to the flow controller 8, the sensor 9, the high-speed camera, the high-speed camera synchronizer 15, the first servo motor 16 and the second servo motor 17 by signals; on the one hand, the information collected by the sensor 9 and the image collected by the high-speed camera are input into the control system 10, and the calculation and analysis module in the control system 10 outputs the corresponding control instructions after calculation; on the other hand, the control system 10 outputs the corresponding control instructions to the flow controller 8, the high-speed camera synchronizer 15, the first servo motor 16 and the second servo motor 17.

[0142] In this application, the control system 10 is implemented as a computer. The computer hardware and program codes related to the experimental method can be used to implement and adjust the control system. Specifically, the computer calculates the corresponding control instructions based on the particle velocity distribution, particle volume fraction distribution, particle collision velocity distribution, and particle collision frequency distribution obtained by the above-mentioned experimental method, as follows:

[0143] When the time average and standard deviation of each variable (particle velocity, particle volume fraction, particle collision velocity, and particle collision frequency) in a certain window change less than the preset value, the camera movement unit is adjusted to obtain a new window position until it covers the position where particles are moving in the fluidization chamber.

[0144] Adjust the spray gas velocity to determine its effect on particle collisions. Obtain the average values ​​of particle collision frequency and particle collision velocity at different spray flow rates, and establish a quantitative relationship between the spray gas flow rate and the average values ​​of particle collision frequency and particle collision velocity at different distances from the nozzle.

[0145] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.

Claims

1. A method for testing the collision of multiphase particles in a spray fluidized bed, characterized in that: The steps include: S1. Establish a CFD-DEM simulation model and use it to simulate particle motion in a multiphase system. The simulation predicts the particle motion characteristics, particle collision velocity, and spatial distribution of collision frequency, and adjusts the camera position based on the spatial distribution. Simulation data on particle motion in a multiphase system is obtained through the simulation. S2. Using simulation data, train and test a deep neural network for pairing discrete particles from image data, a classification deep neural network for determining the number of particle collisions from image data, and a regression deep neural network for calculating collision velocity from image data. S3, collecting image data of the actual movement of particles in the multiphase system; Based on the acquired image data, a convolutional neural network for particle segmentation was trained and tested. The trained convolutional neural network for particle segmentation was used to predict the center position and particle size of each particle in each frame of the high-speed photography image, and further calculate the particle volume fraction. S4, using the deep neural network for paired discrete particles trained in S2 and the particle center position and particle size obtained in S3, predict the pairing information of the particles in each frame of the high-speed photography image, and further calculate the instantaneous velocity, acceleration and motion trajectory of the particles; S5. Using the classification deep neural network trained in S2 to determine the number of particle collisions from image data and the regression deep neural network trained to calculate the collision velocity from image data, as well as the instantaneous velocity and acceleration of the particles obtained in S4, the number of collisions of each particle and the particle collision velocity of each collision in each frame of high-speed photography images are predicted, and the collision frequency and collision velocity distribution of the particles are calculated.

2. A method for testing spray fluidized bed multiphase particle collision according to claim 1, characterized in that: S4 uses segmentation convolutional neural networks to separate the center position and particle size of different types of particles from image data, and pairs discrete particles from image data with deep neural networks to predict discrete particles P at adjacent moments. i The relative displacement of the discrete particles is calculated based on the shooting interval; further, the acceleration of the discrete particles is calculated by the change in the instantaneous velocity of the discrete particles during the shooting interval; the same particles are connected at multiple time positions to reconstruct the particle motion trajectory.

3. A method for testing spray fluidized bed multiphase particle collision according to claim 1, characterized in that: The particle collision frequency is defined as the number of collisions experienced by a single particle per unit time. Therefore, the particle collision frequency is calculated in S5: f c,i =N c,i / (t i+1 -t i ) Among them, N c,i is the number of collisions per particle, t i and t i+1 are adjacent moments; through each particle collision velocity v c,j Create a histogram and calculate the collision velocity distribution.

4. A method for testing the collision of multiphase particles in a spray fluidized bed according to claim 1, characterized in that: The method for training and testing a deep neural network for pairing discrete particles from image data is: Step 1) Extract particle motion characteristics at different times from the CFD-DEM simulation data in the xy dimension, including each particle number and particle center position; the extracted time interval is the interval between the two images, and the units are converted based on the actual size of each pixel; Step 2) Establish the Voronoi diagram (27) based on the center position of the discrete particles and determine the position of each discrete particle P i The corresponding Thiessen polygons and the adjacent discrete particles P corresponding to their adjacent Thiessen polygons i,j ; Step 3) Use the particle number information in DEM to determine the same particle P at different times i , extract t i discrete particles P i and P i,j The relative position of i+1 discrete particles P i and P i,j The relative position of t i and t i+1 Time P i The relative position of the particle was normalized with the equivalent particle diameter and set as the matching success category; t i discrete particles P i and P i,j The relative position of i+1 discrete particles P i,j and the relative position of its neighboring particles, and t i and t i+1 Time P i and P i,j The relative position of the particles is determined, and the data is standardized using the equivalent particle diameter and set as the unsuccessful matching category; a sufficient amount of data is extracted to form a training data set; Step 4) randomly selecting a portion of the training data set obtained in step 3) to train a deep classification neural network for paired discrete particles, where the number of hidden layers and the number of neurons in each layer are adjusted as needed; Step 5) Using the remaining data in the training data set obtained in step 4), the deep neural network of step 4) is tested; Step 6) Repeat steps 4) and 5) until the prediction accuracy of the deep neural network reaches the set requirements.

5. The method for testing the collision of multiphase particles in a spray fluidized bed according to claim 1, characterized in that: The method for training and testing a classification deep neural network for determining the number of particle collisions from image data is: Step 1) Extract particle motion characteristics at different moments from the CFD-DEM simulation data in the xy dimension, including each particle number, particle center position, particle instantaneous velocity, and particle acceleration. The extracted time interval is the interval between the two images, and the unit is converted based on the actual size of each pixel. The number of particle-particle collisions is extracted, and the extracted time interval is 1 / 20 of the shortest particle collision duration. A single particle collision is defined as the period from physical contact to physical separation of the particles. Step 2) Establish a Voronoi diagram based on the center position of the discrete particles and determine the P of each discrete particle. i The corresponding Thiessen polygons and the adjacent discrete particles P corresponding to their adjacent Thiessen polygons i,j ; Step 3) Use the particle number information in DEM to extract t i discrete particles P i and P i,j The relative position, relative velocity and relative acceleration of i+1 discrete particles P i and P i,j The relative position, relative velocity and relative acceleration of the particles are normalized by the equivalent particle diameter; i to t i+1 Time P i The number of collisions that occurred and the number of particles that collided with them; different categories were set according to the number of collisions, and sufficient data was extracted to form a training data set; Step 4) randomly selecting a portion of the training data set obtained in step 3) and training a classification deep neural network (33) for measuring the number of particle collisions, wherein the number of hidden layers of each neural network and the number of neurons in each layer are adjusted according to the needs; Step 5) Using the remaining data in the training data set obtained in step 3) to test the deep neural network in step 4); Step 6) Repeat steps 4) and 5) until the prediction accuracy of the deep neural network reaches the set requirements.

6. A method for testing the collision of multiphase particles in a spray fluidized bed according to claim 1, characterized in that: The method for training and testing a regression deep neural network for computing collision velocities from image data is: Step 1) Extract particle motion characteristics at different moments from the CFD-DEM simulation data in the xy dimension, including each particle number, particle center position, particle instantaneous velocity, and particle acceleration, and convert the unit to the actual size of each pixel. Extract the velocity of each particle-particle collision, that is, the relative velocity of two particles when they come into physical contact, with the extracted time interval being 1 / 20 of the shortest particle collision duration, and convert the collision velocity to the unit using the actual size of each pixel. Step 2) Establish a Voronoi diagram based on the center position of the discrete particles and determine the P of each discrete particle. i The corresponding Thiessen polygons and the adjacent discrete particles P corresponding to their adjacent Thiessen polygons i,j ; Step 3) Use the particle number information in DEM to extract t i discrete particles P i and P i,j The relative position, relative velocity and relative acceleration of i+1 discrete particles P i and P i,j The relative position, relative velocity and relative acceleration of the particles are normalized by the equivalent particle diameter; i to t i+1 Moment and P i The collision particle number and the collision speed of each particle are extracted, and a sufficient amount of data is generated to form a training data set; Step 4) randomly selecting a portion of the training data set obtained in step 3) to train a regression deep neural network (34) for measuring particle collision velocity, wherein the number of hidden layers of each neural network and the number of neurons in each layer are adjusted according to requirements; Step 5) Using the remaining data in the training data set obtained in step 3) to test the deep neural network in step 4); Step 6) Repeat steps 4) and 5) until the prediction accuracy of the deep neural network reaches the set requirements.

7. A method for testing the collision of multiphase particles in a spray fluidized bed according to claim 1, characterized in that: The method for training and testing a convolutional neural network for segmenting particles from image data is: Step 1) converting the raw image data of particle fluidization captured by a high-speed camera into a single-channel grayscale image; Step 2) Segmenting the image data of discrete particles of the same type from the grayscale image obtained in step 1), including a particle image mask (23), a particle rectangular border (24), a pixel matrix within the particle border, and an equivalent particle size (25); Step 3) performing normalized cross-correlation operations on the pixel matrix within the particle frame obtained in step 2) and the two-dimensional matrix of the multi-frame grayscale image data obtained in step 1) to obtain a corresponding correlation coefficient matrix (26); Step 4) determining a plurality of peak positions corresponding to the center positions of the particles in the original image from the correlation coefficient matrix obtained in step 3) by setting a dynamic threshold; Step 5) automatically segmenting different types of particles in each frame of the particle fluidization original image based on the particle center position obtained in step 3) and the discrete particle image data in step 2), and manually removing interfering particles; Step 6) writing the qualified particles segmented in each frame of the original image in step 5) into training data, including the particle mask corresponding to the original image data, the particle type based on the equivalent particle size, and the particle rectangular border; Step 7) Repeat steps 2) to 6) to obtain training data of different types of particles in each frame of the original image, save the corresponding names, and form the final training data set; Step 8) randomly selecting a portion of the training data set obtained in step 7) and training an instance segmentation convolutional neural network for particle segmentation; Step 9) Using the remaining data in the training data set obtained in step 7), testing the instance segmentation convolutional neural network in step 8); Step 10) Repeat steps 8) and 9) until the prediction accuracy of the instance segmentation convolutional neural network reaches the set requirements.

8. A test device for testing multiphase particle collision in a spray fluidized bed, characterized in that: It includes a quasi-two-dimensional spray particle fluidization system, a high-speed photography system and a control system (10); The quasi-two-dimensional spray particle fluidization system comprises: a quasi-two-dimensional visualized particle fluidization chamber (1), a fluidizing gas pressure stabilizing chamber (2), and a particle collector (3); a spliced ​​distribution plate (6) is provided between the fluidizing gas pressure stabilizing chamber (2) and the quasi-two-dimensional visualized particle fluidization chamber (1); a fluidizing chamber inner baffle (7) is provided in the quasi-two-dimensional visualized particle fluidization chamber (1), and the fluidizing chamber inner baffle (7) divides the quasi-two-dimensional visualized particle fluidization chamber (1) into a plurality of cavities that do not interfere with each other; a dual-fluid nozzle (5) and a sensor (9) are provided in the quasi-two-dimensional visualized particle fluidization chamber (1); The high-speed photography system includes two high-speed cameras, a camera moving unit, and a high-speed camera synchronizer (15). The high-speed cameras are mounted on the camera moving unit, and the signals of the two high-speed cameras are connected to the high-speed camera synchronizer (15). The control system (10) uses a computer, and the computer is respectively connected to the two-dimensional spray particle fluidization system and the high-speed photography system by signal; the control system (10) adjusts the position of the high-speed photography system according to the particle velocity, particle volume fraction, particle collision velocity and particle collision frequency obtained by the spray fluidized bed multiphase system particle collision test method as described in claim 1.

9. The test device for testing multiphase particle collision in a spray fluidized bed according to claim 8, characterized in that: When the time average and standard deviation of various variables in a certain window, including particle velocity, particle volume fraction, particle collision velocity and particle collision frequency, change less than the preset value, the camera moving unit is adjusted to obtain a new window position until it covers the position where particles are moving in the fluidization chamber.

10. The test device for testing multiphase particle collision in a spray fluidized bed according to claim 8, characterized in that: The high-speed cameras are two HCCD high-speed cameras, and the high-speed camera synchronizer (15) is respectively connected to the first HCCD high-speed camera (11) and the second HCCD high-speed camera (13) by signals.

Images