Glass grinding roughness prediction method based on discrete element simulation and neural network

Abstract

The application provides a glass grinding roughness prediction method based on discrete element simulation and a neural network, and comprises the following steps: constructing a fixed abrasive grinding disc model based on abrasive particle size and spatial distribution; constructing a motion trajectory model of abrasive particles on the surface of a current workpiece to be measured based on the grinding disc model and motion parameters of a double-sided grinding equipment; inputting abrasive particle size, a preset grinding pressure and grinding speed of each abrasive particle in the motion trajectory model into a pre-trained single abrasive particle scratch prediction model to obtain scratch damage parameters corresponding to each abrasive particle; the single abrasive particle scratch prediction model is trained based on a discrete element single abrasive particle scratch simulation dataset constructed based on a sample workpiece; and roughness of the surface of the workpiece is calculated based on the motion trajectory model and the scratch damage parameters to obtain a final surface roughness prediction result of the current workpiece to be measured. Based on the method, the surface roughness of double-sided grinding is predicted, and efficient optimization of processing parameters is effectively supported and experimental cost is reduced.

Description

Technical Field

[0001] This invention relates to the fields of precision machining and intelligent manufacturing technology, specifically to a method for predicting the roughness of glass grinding based on discrete element simulation and neural networks. Background Technology

[0002] With the rapid development of semiconductor packaging, precision instruments, aerospace, and high-end electronic devices, optical glass has become a key basic material due to its excellent optical transmittance, chemical stability, and mechanical properties. Especially in high-end manufacturing scenarios such as hard disk drive substrates and precision optical lenses, extremely stringent requirements are placed on the surface quality, dimensional accuracy, and subsurface damage control of glass materials. Double-sided grinding, as a core process in glass precision machining, can efficiently and uniformly grind both the upper and lower surfaces of a workpiece simultaneously, and has become the mainstream processing method. However, during double-sided grinding, the friction, grooving, and cutting behaviors between the abrasive grains and the workpiece surface can easily lead to problems such as brittle material removal, crack initiation and propagation, and groove accumulation, directly affecting the surface morphology, roughness, and subsurface damage depth of the workpiece. Therefore, accurately predicting the surface roughness of double-sided ground glass is of significant engineering importance for optimizing process parameters, ensuring processing quality, and improving production efficiency.

[0003] Currently, research on surface roughness in double-sided grinding mainly focuses on experimental statistical methods, numerical simulation methods, and intelligent algorithm prediction. Traditional empirical models are mostly constructed based on the Preston equation or regression analysis methods, fitting the relationship between roughness and process parameters through a large amount of experimental data. In terms of numerical simulation, the finite element method and the discrete element method are used to simulate the interaction between abrasive grains and the workpiece. Among them, the discrete element method has been widely used in recent years in the study of single abrasive grain scratching and grinding mechanisms because it can better describe the crack initiation, propagation, and particle failure process inside hard and brittle materials. In addition, intelligent algorithms such as neural networks have also been tried to predict grinding parameters and analyze surface quality, showing a certain data fitting ability.

[0004] However, existing technologies still have significant shortcomings. Traditional empirical models are highly dependent on specific experimental conditions and have poor adaptability to different abrasive grain sizes, grinding pressures, and processing speeds, making it difficult to characterize the dynamic behavior of cracks in the processing of hard and brittle materials. The finite element method faces problems such as mesh distortion and tracking difficulties when simulating crack propagation, and cannot effectively support parametric analysis of complex working conditions. Discrete element methods are mostly limited to the local scratching process of a single abrasive grain, and cannot correlate the composite motion trajectory of multiple abrasive grains and the superposition mechanism of surface morphology in actual double-sided grinding equipment, resulting in a disconnect between microscopic grinding behavior and macroscopic processing effects. Although intelligent prediction models have the advantage of data fitting, they generally lack physical characterization of the crack formation, material removal, and surface damage evolution processes in hard and brittle materials, limiting their generalization ability. These limitations make it difficult for existing methods to truly reflect the formation mechanism of double-sided grinding surfaces, restricting the efficient optimization of processing parameters and the reduction of experimental costs. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a method for predicting the roughness of glass grinding based on discrete element simulation and neural networks.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a method for predicting the roughness of glass grinding based on discrete element simulation and neural networks, comprising: Based on the size and spatial distribution of abrasive particles, a model of a bonded abrasive grinding disc is constructed. The particle size of the abrasive particles in the grinding disc model follows a normal distribution, and the spatial arrangement of the abrasive particles on the surface of the grinding disc is determined by the Poisson disk sampling method. Based on the grinding disc model and the motion parameters of the double-sided grinding equipment, a motion trajectory model of abrasive particles on the surface of the current workpiece to be tested is constructed. The abrasive grain size, preset grinding pressure, and grinding speed of each abrasive grain in the motion trajectory model are input into the pre-trained single abrasive grain scratch prediction model to obtain the scratch damage parameters corresponding to each abrasive grain; wherein, the single abrasive grain scratch prediction model is trained based on the discrete element single abrasive grain scratch simulation dataset constructed from sample workpieces. Based on the motion trajectory model and scratch damage parameters, the surface roughness of the workpiece is calculated by reconstructing the surface morphology of the current workpiece under test, and the final surface roughness prediction result of the current workpiece under test is obtained.

[0007] This invention provides a method for predicting the roughness of glass grinding based on discrete element simulation and neural networks, comprising: constructing a bonded abrasive grinding disc model based on abrasive particle size and spatial distribution; wherein the abrasive particle size in the grinding disc model follows a normal distribution, and the spatial arrangement of abrasive particles on the grinding disc surface is determined by a Poisson disk sampling method; constructing a motion trajectory model of abrasive particles on the surface of the current workpiece to be tested based on the grinding disc model and the motion parameters of the double-sided grinding equipment; inputting the abrasive particle size, preset grinding pressure, and grinding speed of each abrasive particle in the motion trajectory model into a pre-trained single abrasive scratch prediction model to obtain scratch damage parameters corresponding to each abrasive particle; wherein the single abrasive scratch prediction model is trained based on a discrete element single abrasive scratch simulation dataset constructed from sample workpieces; and calculating the surface roughness of the current workpiece by reconstructing the surface morphology of the current workpiece to be tested based on the motion trajectory model and scratch damage parameters to obtain the final surface roughness prediction result of the current workpiece to be tested. In this invention, firstly, a bonded abrasive grinding disc model is constructed with abrasive grains following a normal distribution and spatial arrangement sampled using Poisson disk sampling. An abrasive grain motion trajectory model is also established, and the grinding disc is reconstructed using probabilistic statistical methods, providing a foundation for coupling single-grain scratch damage with motion trajectory. Then, a single-grain scratch prediction model trained on a discrete element single-grain scratch simulation dataset is used. Combined with inputs of abrasive grain size, grinding pressure, and grinding speed, the model outputs scratch damage parameters corresponding to each abrasive grain. This solves the problems of traditional empirical models relying on specific experimental conditions and being difficult to adapt to different abrasive grain sizes, grinding pressures, and processing speeds, as well as the mesh distortion and tracking difficulties faced by the finite element method when simulating crack propagation. This effectively supports parametric analysis under complex working conditions. Finally, the surface morphology of the workpiece is reconstructed and roughness is calculated using the motion trajectory model and scratch damage parameters. This solves the problems of existing research being limited to the local scratch process of a single abrasive grain and unable to correlate the composite motion trajectory of multiple abrasive grains and the surface morphology superposition mechanism, achieving a connection between microscopic grinding behavior and macroscopic processing effects. Meanwhile, this invention addresses the limitation of intelligent prediction models in terms of physical characterization of material removal and surface damage evolution processes in hard and brittle materials, enabling prediction results to accurately reflect the formation mechanism of double-sided grinding surfaces. In summary, this invention achieves prediction of surface roughness in double-sided grinding processes, effectively supporting efficient optimization of processing parameters and reducing experimental costs.

[0008] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a method for predicting the roughness of glass grinding based on discrete element simulation and neural networks, provided in an embodiment of the present invention; Figure 2 A simplified structural diagram of a double-sided grinding apparatus is shown as an example. Figure 3 A simplified diagram of the grinding motion is shown as an example; Figure 4 A schematic diagram of the single abrasive scratch simulation model of the present invention is shown as an example; Figure 5 An exemplary complete execution block diagram of a glass grinding roughness prediction method based on discrete element simulation and neural network is shown; Figure 6 An exemplary histogram of surface roughness under different abrasive grain sizes is shown; Figure 7 A surface roughness histogram under different grinding pressure conditions is shown as an example; Figure 8 An exemplary histogram of surface roughness under different rotation-revolution ratios is shown. Detailed Implementation

[0010] This invention provides a method for predicting the roughness of glass grinding based on discrete element method (DEM) simulation and neural networks. A single diamond abrasive grain scratching glass model is established in DEM software. By setting parameters such as abrasive grain size, grinding pressure, and scratching speed, scratching simulation is conducted to obtain damage characteristic parameters such as crack depth, crack width, groove depth, and groove width, and a single abrasive grain scratch dataset is constructed. A neural network model is trained using this dataset to establish the mapping relationship between processing parameters and damage characteristic parameters. A bonded abrasive grinding disc model composed of octahedral abrasive grains is established. The abrasive grain size follows a normal distribution and is distributed on the grinding disc surface using a Poisson disk sampling algorithm. An abrasive grain motion trajectory model is constructed by combining the kinematic relationship of a double-sided grinding device. The single abrasive grain scratch features predicted by the neural network are coupled with the abrasive grain motion trajectory. Based on the superposition law of multiple abrasive grain scratches, a three-dimensional surface morphology of the workpiece is generated. The roughness of multiple measurement areas uniformly arranged using the golden ratio is statistically analyzed to obtain the workpiece surface roughness prediction result. This method establishes a correlation between the damage characteristics of a single abrasive grain and the overall surface quality of double-sided grinding, and can be used for surface roughness prediction and process parameter analysis in the double-sided grinding process of quartz glass and other hard and brittle materials.

[0011] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0012] To predict the surface roughness of double-sided grinding processes, effectively support efficient optimization of processing parameters, and reduce experimental costs, this invention provides a method for predicting the roughness of glass grinding based on discrete element simulation and neural networks. Figure 1 This is a flowchart illustrating a method for predicting the roughness of glass grinding based on discrete element simulation and neural networks, provided in an embodiment of the present invention. Figure 1 As shown, it includes: S101. Based on the size and spatial distribution of abrasive particles, construct a model of a bonded abrasive grinding disc.

[0013] In the grinding disc model, the abrasive grain size follows a normal distribution, and the spatial arrangement of the abrasive grains on the grinding disc surface is determined using the Poisson disk sampling method.

[0014] Optionally, the abrasive grains in the grinding disc model have an octahedral structure, and the grain size follows a normal probability density function, which is expressed as: ; in, The abrasive grain size is... Indicates abrasive grain size The corresponding probability density function, This represents the average abrasive grain size. This represents the standard deviation of the abrasive grain size.

[0015] Based on actual equipment parameters, the grinding disc is annular, and its geometric characteristics are as follows: ; in, This indicates the geometric region of the grinding wheel. Indicates the inner diameter of the grinding disc. Indicates the outer diameter of the grinding disc. This represents the length of a straight line from a point within the geometric region to the origin of the grinding wheel center. This indicates that the point is... The angle of the line relative to the horizontal reference axis.

[0016] The spatial arrangement of abrasive grains on the grinding disc surface is achieved using the Bridson two-dimensional Poisson disk sampling method. This method ensures that the distance between any two abrasive grains is not less than a predetermined repulsion radius. That is, any two abrasive grains satisfy:

[0017] Indicates the first The position corresponding to each abrasive grain Indicates the first The position corresponding to each abrasive grain.

[0018] This provides the size and spatial distribution of the abrasive grains on the grinding disc, laying the foundation for the kinematic model.

[0019] S102. Based on the grinding disc model and the motion parameters of the double-sided grinding equipment, construct a motion trajectory model of the abrasive particles on the surface of the workpiece to be tested.

[0020] Optionally, S102 includes: Obtain the motion parameters of the double-sided grinding equipment; the motion parameters should include at least the number of teeth on the sun gear, the number of teeth on the planet gears, the number of teeth on the external gear ring, the speed of the sun gear, the speed of the grinding disc, and the speed of the external gear ring. Based on the motion parameters and the coordinates of each abrasive grain in the grinding disc model, the motion trajectory matrix of each abrasive grain on the surface of the current workpiece is calculated through coordinate transformation and relative motion decomposition. The grinding speed of each abrasive grain relative to the current workpiece is calculated based on the motion trajectory matrix; The motion trajectory matrix and grinding speed are used together as the motion trajectory model.

[0021] Alternatively, the motion trajectory model can be represented as: ; ; in, Represents any abrasive grain Relative to the absolute x-coordinate in the planetary coordinate system, Represents any abrasive grain Relative to the absolute ordinate in the planetary gear coordinate system, Indicates abrasive grains Distance to the center of the millstone Indicates the grinding disc rotation speed. This represents the angular velocity of the planetary gear's rotation. This represents the angular velocity of the planetary gears' revolution. Indicates the rotational speed of the sun gear. Indicates the number of teeth on the sun gear. Indicates the number of teeth on the planetary gear. Indicates the number of teeth on the external gear ring. This indicates the rotational speed of the external gear ring. Indicates the grinding process time. This represents the initial phase angle of the abrasive grains. This indicates the lateral distance from the center of the millstone to the center of the planetary gears; Grinding speed is expressed as: ; in, This indicates the grinding speed of the abrasive particles relative to the workpiece being tested.

[0022] Figure 2 A simplified structural diagram of a double-sided grinding apparatus is shown as an example. Figure 3 An exemplary schematic diagram of the grinding motion is shown, such as... Figure 3 As shown, the left-hand diagram illustrates the overall kinematic structure of the double-sided grinder, including sun gears rotating around their respective centers. Planetary wheels , ) and grinding disc ( The upper middle figure shows the equivalent model of the grinding motion. The lower middle figure is a schematic diagram of the abrasive grain movement trajectory on the workpiece, and the right figure is a schematic diagram of the three-dimensional morphology of the microgrooves generated by the abrasive grains on the workpiece surface, which intuitively shows the micro-cutting trajectory and groove geometry of the abrasive grains as they move across the workpiece surface.

[0023] S103. Input the abrasive particle size, preset grinding pressure, and grinding speed of each abrasive particle in the motion trajectory model into the pre-trained single abrasive scratch prediction model to obtain the scratch damage parameters corresponding to each abrasive particle.

[0024] Among them, the single abrasive scratch prediction model is trained based on the discrete element single abrasive scratch simulation dataset constructed from sample workpieces.

[0025] Optionally, the process of constructing a single abrasive scratch prediction model includes: The Latin hypercube sampling method was used to perform multi-level sampling of the sample abrasive particle size, sample grinding pressure and sample grinding speed of the sample workpiece to obtain the sample input dataset; Discrete element single abrasive scratch simulation is performed based on sample workpieces and sample input datasets to obtain sample output datasets corresponding to the sample input datasets, which include sample crack depth, sample crack width, sample groove depth, and sample groove width; the sample input datasets and sample output datasets together constitute the discrete element single abrasive scratch simulation dataset. Using the sample input dataset as input and the sample output dataset as output, a neural network model is trained to obtain a single abrasive scratch prediction model.

[0026] It should be noted that the neural network model can be a BP (Back Propagation) neural network model, a fully connected deep neural network (DNN) model, or a support vector regression (SVR), random forest regression, or a lightweight MLP (Multi-Layer Perceptron) model. Considering that the input consists of a small number of continuous physical quantities such as abrasive particle size, grinding pressure, and grinding speed, and the output consists of continuous damage parameters such as scratch depth or scratch width, a fully connected BP / DNN (MLP) network with a nonlinear activation function is usually preferred. This type of model has a simple structure, low training cost, and can effectively fit the nonlinear mapping relationship existing in the abrasive scratch process, making it suitable as a single abrasive scratch prediction surrogate model embedded in a discrete element-neural network hybrid framework.

[0027] The generation process of the single abrasive scratch prediction model specifically includes: First, the microscopic parameters of the glass material are calibrated. The interactions between particles in the discrete element method are calibrated through uniaxial compression, direct tension, and three-point bending experiments to ensure that the mechanical response of the simulation model is consistent with that of the real glass material.

[0028] Secondly, a single abrasive scratch simulation of glass was established based on calibrated glass micro-parameters. Diamond abrasive grains were defined as rigid regular octahedrons. Given the grinding pressure, grinding speed, and abrasive grain diameter, the abrasive grains scratched at a given speed according to the initial penetration depth. A scripting language was written to identify the scratch depth and width of the particles and to extract the crack depth and width.

[0029] Furthermore, the Latin hypercube sampling method was used to sample the three-dimensional parameters of abrasive diameter, grinding pressure and grinding speed at multiple levels. Glass single abrasive scratch simulation was carried out under different processing parameter conditions to obtain datasets of crack depth, crack width, groove depth and groove width under different grinding processes.

[0030] Finally, the single abrasive scratch simulation dataset was imported to train the neural network. Grinding pressure, abrasive particle size, and grinding speed were used as input parameters, and crack depth, crack width, groove depth, and groove width were used as output parameters. The neural network model was trained, validated, and subjected to error analysis to obtain the single abrasive scratch neural network prediction model. ; In the formula, The depth of a single abrasive grain scratch channel. The width of a single abrasive scratch groove. The abrasive grain size is... For grinding pressure, This refers to the grinding speed.

[0031] By inputting the abrasive grain size, abrasive grain movement speed, and grinding pressure from the above steps into the trained neural network model, the crack depth, crack width, groove depth, and groove width after each abrasive grain scratch on the grinding disc can be predicted.

[0032] Figure 4 A schematic diagram of the single abrasive scratch simulation model of the present invention is shown as an example. Wherein, Figure 4 Figure (a) shows a schematic diagram of the initial state of the single abrasive scratch model; (b) shows a schematic diagram of the morphology of the single abrasive scratch model after simulation; (c) shows a schematic diagram used to identify the particle distribution in the channel width and depth after simulation; (d) shows a schematic diagram of the crack propagation inside the material after simulation. Figure 4 As shown, the model demonstrates the complete discrete element simulation process from initial indentation to scratch formation, geometric feature extraction, and crack identification.

[0033] S104. Based on the motion trajectory model and scratch damage parameters, the surface roughness of the workpiece is calculated by reconstructing the surface morphology of the current workpiece to be tested, and the final surface roughness prediction result of the current workpiece to be tested is obtained.

[0034] Optionally, S104 includes: The surface of the workpiece to be tested is divided into grids. Based on the motion trajectory model, the local depths generated by the motion trajectories of each abrasive grain at each grid are superimposed to obtain the reconstructed morphology of the workpiece surface. Multiple circular sampling zones are selected on the reconstructed morphology of the workpiece surface. The surface roughness of each circular sampling zone is calculated and then the weighted average value is taken to obtain the final surface roughness prediction result.

[0035] The workpiece surface reconstruction morphology mentioned above is based on the initial flat workpiece surface. On the divided grid cells, the motion trajectory of each abrasive grain on the workpiece is tracked according to the motion trajectory model. The corresponding scratch damage parameters (such as scratch width and groove depth) are accumulated / superimposed grid by grid according to the actual location. This simulates the effect of material being gradually removed during multiple abrasive grain reciprocating grinding. The final generated digital elevation surface includes scratch intersection, overlap and cumulative damage features, which is used to equivalently reflect the actual three-dimensional surface texture and micromorphology of the workpiece after double-sided grinding.

[0036] Optionally, multiple circular sampling partitions are selected on the reconstructed morphology of the workpiece surface. The surface roughness of each circular sampling partition is calculated, and a weighted average is taken to obtain the final surface roughness prediction result, including: Using the center of the workpiece to be tested as the origin, several circular sampling zones are uniformly arranged within the reconstructed morphology of the workpiece surface based on the golden ratio. Calculate the surface roughness of the corresponding circular sampling partition based on the surface height data within each circular sampling partition. The final surface roughness prediction result is obtained by taking a weighted average of multiple surface roughness values.

[0037] Optionally, the local depth generated by the motion trajectory of each abrasive grain at each grid is represented as: ; in, This represents the normal distance from the grid point to the center line of the scratch. Indicates in The local depth generated below This indicates the depth of the groove created by a single abrasive grain scratch. This indicates the width of the groove created by a single abrasive grain.

[0038] Alternatively, the circular sampling partition can be represented as: ; in, Indicates the first The radius coordinates of each circular sampling partition. Indicates the first The polar angle of a circular sampling partition. This indicates the radius of the workpiece currently being measured. This represents the outer radius of the circular sampling partition. Indicates the sequence number of the circular sampling partition. This indicates the total number of circular sampling partitions.

[0039] This invention discretizes the two-dimensional surface of the workpiece under test into regular, tiny, flat grid cells, which serve as the basic computational unit. The specific number of grid cells is dynamically determined based on the actual size of the workpiece and the step size of the abrasive grain movement trajectory. In practice, the grid cells must be dense enough to capture microscopic machining marks while also considering computational power. During reconstruction, each grid cell is traversed, and the local depths generated by the abrasive grain trajectories passing through that cell at the corresponding time are extracted, summed, and finally superimposed to generate a macroscopic three-dimensional morphology. Subsequently, multiple circular sampling partitions are uniformly selected on the morphology image to calculate the local roughness, and the final result is obtained by weighted averaging based on area or weight, thereby achieving accurate prediction from microscopic scratches to macroscopic roughness.

[0040] Optionally, the scratch damage parameters include at least one of the groove depth, groove width, crack depth, and crack width corresponding to the current workpiece under test.

[0041] For each detection region, extract the first [item] within the circular sampling partition. Surface height value of each grid The following calculations yield the first... Surface roughness of a single circular sampling partition : ; ; In the formula, This represents the average height of all grid surfaces within the current circular sampling partition. This represents the number of valid grid cells within the circular sampling partition.

[0042] Final weighted surface roughness of the entire substrate The expression is: ; This indicates the radius of the workpiece currently being measured. This represents the outer boundary radius of the circular sampling partition.

[0043] To fully illustrate the method of the present invention, Figure 5An exemplary complete execution block diagram of a glass grinding roughness prediction method based on discrete element simulation and neural networks is shown.

[0044] like Figure 5 As shown, the entire prediction process begins with "Start". First, Step 1 establishes a model of the bonded abrasive grinding disc, including the normal distribution and spatial arrangement characteristics of abrasive grain size. Then, Step 2 presents the kinematic model, which obtains the motion parameters of the equipment and the abrasive grains in contact with the substrate, and calculates the kinematic parameters. The system then processes the motion information of the contacting abrasive grains; next, it proceeds to Step 3, the single abrasive grain grinding physical model, and uses a pre-trained neural network parameter file to obtain scratch damage parameters; subsequently, it proceeds to Step 4 to output the three-dimensional surface morphology, generating the first... The surface morphology generated by contact abrasive particles is determined by whether contact abrasive particles are present, and the loop continues or ends. Finally, the loop logic of this parallel processing is combined with the independent parameter input and output module on the right (covering grinding disk size, grinding pressure, grinding speed and groove geometry information, etc.) to jointly guide the output of the final "surface roughness" result.

[0045] This invention proposes a method for predicting the roughness of glass grinding based on discrete element method (DEM) simulation and neural networks. By establishing a single-abrasive-grain DEM scratch model of glass material, microscopic damage characteristic parameters such as crack depth, crack width, groove depth, and subsurface damage depth are obtained. A neural network is then used to establish a mapping relationship between processing parameters and damage characteristics. Furthermore, by combining this with a kinematic trajectory model of a double-sided grinding equipment, the damage behavior of a single abrasive grain is coupled with the actual motion trajectory of multiple abrasive grains, thus achieving the prediction of surface roughness in double-sided grinding. In summary, this method has two advantages: First, it uses the DEM method to establish a single-abrasive-grain scratch model, which can characterize crack initiation, propagation, and particle failure behavior during glass material processing, providing a better physical mechanism basis compared to traditional empirical models. Second, by constructing a single-abrasive-grain scratch dataset and using a neural network to establish a nonlinear mapping relationship between processing parameters and damage characteristics, it is applicable to the rapid prediction of single-abrasive-grain damage characteristics under different abrasive grain sizes, grinding pressures, and processing speeds, reducing the need for extensive repetitive DEM simulations and physical processing experiments, and improving prediction efficiency. Third, by coupling the neural network prediction results with the motion trajectory of the double-sided grinding equipment, a multi-abrasive surface morphology generation model is established, realizing the extension from single-abrasive local damage to overall surface morphology prediction, and solving the problem that existing discrete element methods cannot directly predict overall surface roughness. Fourth, compared with traditional experimental methods, this invention can achieve surface roughness prediction without conducting a large number of physical double-sided grinding experiments, which can effectively reduce experimental cycle, material consumption and processing costs, and improve the efficiency of process parameter optimization. Fifth, this invention combines the mechanism analysis capability of discrete element simulation with the efficient prediction capability of neural network models, reducing the overall computational cost while ensuring prediction accuracy, and is suitable for process analysis and intelligent optimization in the double-sided grinding process of hard and brittle materials.

[0046] To establish a motion trajectory model and surface morphology prediction model for double-sided grinding, this embodiment focuses on a glass double-sided grinding device. The structural and motion parameters of this device are as follows: The sun gear has 160 teeth, the planet gears have 200 teeth, and the external gear ring has 560 teeth, with a gear module of 1 / 12 inch. The sun gear pitch circle diameter is 338 mm, the planet gear pitch circle diameter is 423 mm, and the external gear ring pitch circle diameter is 1186 mm, with a total of 5 planet gears. The grinding disc diameter is 1156 mm; the upper grinding disc speed range is 0~30 rpm, and the lower grinding disc speed range is -50~0 rpm; the speed range of both the sun gear and the external gear ring is 0~30 rpm. The equipment can load 8+2 aluminum substrates, each with a diameter of 97 mm.

[0047] Subsequent embodiments all use the above-described equipment structural parameters and motion parameters, and unless otherwise specified, all other conditions remain the same. The specific implementation process of the present invention will be described in detail below with reference to the accompanying drawings.

[0048] (1) Establish a single abrasive scratch model for glass material in discrete element method. The abrasive adopts a regular octahedral structure. By changing parameters such as abrasive particle size, grinding pressure and scratching speed, scratch simulation is carried out to obtain damage characteristic parameters such as crack depth, crack width, groove depth and groove width, and construct a single abrasive scratch dataset.

[0049] (2) Use the single abrasive scratch dataset to train a neural network model, establish the mapping relationship between processing parameters and single abrasive damage characteristics, and realize the rapid prediction of single abrasive scratch characteristics.

[0050] (3) Establish a double-sided grinding consolidation abrasive grinding disc model. The spatial arrangement of diamond abrasive grains on the grinding disc surface adopts the Poisson disk sampling algorithm. The abrasive grains adopt a regular octahedral structure, and the abrasive grain size follows a normal distribution. Based on the structural parameters of the double-sided grinding equipment, a kinematic model is established to obtain the motion trajectory of the abrasive grains on the workpiece surface.

[0051] (4) The single abrasive groove depth, groove width and subsurface damage characteristic parameters predicted by the neural network are mapped to the motion trajectory of the corresponding abrasive grains, and the workpiece surface morphology is generated based on the principle of multi-abrasive grain scratch superposition.

[0052] (5) Multiple measurement areas are evenly arranged on the generated three-dimensional surface morphology using the golden ratio, and the surface roughness parameters are calculated respectively. The weighted average of the roughness of each measurement area is used as the final surface roughness prediction result of the workpiece.

[0053] The effectiveness of this invention will be verified by changing process parameters such as pressure, abrasive particle size, and rotation speed.

[0054] Example 1: Surface roughness prediction under different abrasive particle sizes.

[0055] To verify the predictive ability of the model of this invention for the surface quality of double-sided grinding under different abrasive particle sizes, diamond abrasives with average particle sizes of 5 μm, 10 μm, 20 μm, and 30 μm were selected to establish a grinding disc model. All other parameters remained consistent: grinding pressure of 20 kPa, sun gear speed of 20 rpm, external gear ring speed of 10 rpm, grinding disc speed of -20 rpm, and processing time of 80 s.

[0056] Figure 6 A histogram of surface roughness under different abrasive grain sizes is shown as an example. Figure 6 As shown, the surface roughness of the workpiece generally increases with the increase of abrasive grain size.

[0057] Example 2: Surface roughness prediction under different grinding pressure conditions.

[0058] To verify the adaptability of the model of this invention to changes in grinding pressure, grinding pressures were set to 5 kPa, 10 kPa, 15 kPa, 20 kPa, 25 kPa, and 30 kPa, respectively. All other parameters remained consistent: average abrasive particle size of 10 μm, sun gear speed of 20 rpm, external gear ring speed of 10 rpm, grinding disc speed of -20 rpm, and processing time of 80 s.

[0059] Figure 7 A surface roughness histogram is shown for example under different grinding pressure conditions. Figure 7 As shown, the prediction results indicate that the surface roughness of the workpiece increases with increasing grinding pressure. This result is consistent with the brittle removal mechanism of hard and brittle materials during actual double-sided grinding, verifying that the method of this invention can accurately predict the effect of grinding pressure changes on surface roughness.

[0060] Example 3: Surface roughness prediction under different rotation-revolution ratios.

[0061] To investigate the influence of kinematic parameters on surface roughness during double-sided grinding, the planetary gear's revolution-to-rotation speed ratio was set to 0.2, 0.5, 0.8, 1.2, 1.5, and 2.0, respectively. All other parameters remained consistent: average abrasive grain size of 10 μm, grinding pressure of 20 kPa, revolution speed of 10 rpm, grinding disc rotation speed of -20 rpm, and processing time of 80 s.

[0062] Figure 8 The surface roughness histogram prediction results under different rotation-to-revolution ratios are shown as an example. Figure 8 As shown, the surface roughness of the workpiece generally decreases as the revolution-rotation ratio increases from 0.2 to 2.0.

[0063] The method provided in this embodiment of the invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc., and this embodiment of the invention does not limit the application to such devices.

[0064] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0065] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings and the disclosure, will understand and implement other variations of the disclosed embodiments in carrying out the claimed invention. In this description, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0066] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the roughness of glass grinding based on discrete element simulation and neural networks, characterized in that, include: Based on the size and spatial distribution of abrasive particles, a model of a bonded abrasive grinding disc is constructed; wherein, the particle size of the abrasive particles in the grinding disc model follows a normal distribution, and the spatial arrangement of the abrasive particles on the surface of the grinding disc is determined by the Poisson disk sampling method. Based on the grinding disc model and the motion parameters of the double-sided grinding equipment, a motion trajectory model of the abrasive grains on the surface of the current workpiece to be tested is constructed. The abrasive grain size, preset grinding pressure, and grinding speed of each abrasive grain in the motion trajectory model are input into a pre-trained single abrasive grain scratch prediction model to obtain scratch damage parameters corresponding to each abrasive grain; wherein, the single abrasive grain scratch prediction model is trained based on a discrete element single abrasive grain scratch simulation dataset constructed from sample workpieces. Based on the motion trajectory model and the scratch damage parameters, the surface roughness of the workpiece is calculated by reconstructing the surface morphology of the current workpiece under test, and the final surface roughness prediction result of the current workpiece under test is obtained.

2. The method for predicting glass grinding roughness based on discrete element simulation and neural networks according to claim 1, characterized in that, The construction process of the single abrasive scratch prediction model includes: The Latin hypercube sampling method was used to perform multi-level sampling of the sample abrasive particle size, sample grinding pressure and sample grinding speed of the sample workpiece to obtain the sample input dataset; Discrete element single abrasive scratch simulation is performed based on the sample workpiece and the sample input dataset to obtain a sample output dataset corresponding to the sample input dataset, which includes sample crack depth, sample crack width, sample groove depth, and sample groove width; wherein, the sample input dataset and the sample output dataset together constitute the discrete element single abrasive scratch simulation dataset. Using the sample input dataset as input and the sample output dataset as output, a neural network model is trained to obtain the single abrasive scratch prediction model.

3. The method for predicting glass grinding roughness based on discrete element simulation and neural networks according to claim 1, characterized in that, The abrasive grains in the grinding disc model have an octahedral structure, and the grain size follows a normal distribution probability density function, which is expressed as: ； in, The abrasive grain size is... Indicates abrasive grain size The corresponding probability density function, This represents the average abrasive grain size. This represents the standard deviation of the abrasive grain size.

4. The method for predicting glass grinding roughness based on discrete element simulation and neural networks according to claim 1, characterized in that, The step of constructing a motion trajectory model of abrasive grains on the surface of the workpiece under test based on the grinding disc model and the motion parameters of the double-sided grinding equipment includes: Obtain the motion parameters of the double-sided grinding equipment; the motion parameters include at least the number of teeth on the sun gear, the number of teeth on the planet gears, the number of teeth on the outer gear ring, the rotational speed of the sun gear, the rotational speed of the grinding disc, and the rotational speed of the outer gear ring; Based on the motion parameters and the coordinates of each abrasive grain in the grinding disc model, the motion trajectory matrix of each abrasive grain on the surface of the current workpiece under test is calculated through coordinate transformation and relative motion decomposition. The grinding speed of each abrasive grain relative to the current workpiece is calculated based on the motion trajectory matrix. The motion trajectory matrix and the grinding speed are used together as the motion trajectory model.

5. The method for predicting glass grinding roughness based on discrete element simulation and neural networks according to claim 4, characterized in that, The motion trajectory model is represented as follows: ； ； in, Represents any abrasive grain Relative to the absolute x-coordinate in the planetary coordinate system, Represents any abrasive grain Relative to the absolute ordinate in the planetary gear coordinate system, Indicates abrasive grains Distance to the center of the millstone Indicates the grinding disc rotation speed. This represents the angular velocity of the planetary gear's rotation. This represents the angular velocity of the planetary gears' revolution. Indicates the rotational speed of the sun gear. Indicates the number of teeth on the sun gear. Indicates the number of teeth on the planetary gear. Indicates the number of teeth on the external gear ring. Indicates the rotational speed of the external gear ring. Indicates the grinding process time. This represents the initial phase angle of the abrasive grains. This indicates the lateral distance from the center of the millstone to the center of the planetary gears; The grinding speed is expressed as: ； in, This indicates the grinding speed of the abrasive particles relative to the workpiece being tested.

6. The method for predicting glass grinding roughness based on discrete element simulation and neural networks according to claim 4, characterized in that, The step of calculating the surface roughness of the workpiece by reconstructing the surface morphology of the current workpiece based on the motion trajectory model and the scratch damage parameters, to obtain the final surface roughness prediction result of the current workpiece, includes: A grid is defined on the surface of the workpiece to be tested. Based on the motion trajectory model, the local depths generated by the motion trajectories of each abrasive grain at each grid are superimposed to obtain the reconstructed morphology of the workpiece surface. Multiple circular sampling partitions are selected on the reconstructed morphology of the workpiece surface. The surface roughness of each circular sampling partition is calculated and then a weighted average is taken to obtain the final surface roughness prediction result.

7. The method for predicting glass grinding roughness based on discrete element simulation and neural networks according to claim 6, characterized in that, The process involves selecting multiple circular sampling zones on the reconstructed morphology of the workpiece surface, calculating the surface roughness of each circular sampling zone, and then taking a weighted average to obtain the final surface roughness prediction result, including: With the center of the workpiece to be tested as the origin, several circular sampling zones are uniformly arranged within the reconstructed morphology of the workpiece surface based on the golden ratio. Calculate the surface roughness of the corresponding circular sampling partition based on the surface height data within each of the circular sampling partitions. The final surface roughness prediction result is obtained by taking a weighted average of the multiple surface roughness values.

8. The method for predicting glass grinding roughness based on discrete element simulation and neural networks according to claim 6, characterized in that, The local depth generated at each grid point by the motion trajectory of each abrasive grain is represented as follows: ； in, This represents the normal distance from the grid point to the center line of the scratch. Indicates in The local depth generated below This indicates the depth of the groove created by a single abrasive grain scratch. This indicates the width of the groove created by a single abrasive grain.

9. The method for predicting glass grinding roughness based on discrete element simulation and neural networks according to claim 6, characterized in that, The circular sampling partition is represented as follows: ； in, Indicates the first The radius coordinates of each circular sampling partition. Indicates the first The polar angle of a circular sampling partition. This indicates the radius of the workpiece currently being measured. This represents the outer radius of the circular sampling partition. Indicates the sequence number of the circular sampling partition. This indicates the total number of circular sampling partitions.

10. The method for predicting glass grinding roughness based on discrete element simulation and neural networks according to claim 1, characterized in that, The scratch damage parameters include at least one of the groove depth, groove width, crack depth, and crack width corresponding to the current workpiece under test.

Images