A Machine Learning-Based Vacuum Glue Filling Control Method and System

Abstract

The present invention provides a vacuum potting control method and system based on machine learning, which relates to the field of intelligent control technology. The method includes: defining the geometric shape, material properties, boundary conditions and initial conditions of a finite element model to simulate a real potting environment, where the physical properties include the rheology and heat conductivity of the glue; simulating the potting process at different potting speeds, glue temperatures and vacuum degrees through a three-dimensional geometric model of the potting equipment to obtain simulation results, which include glue flow behavior, temperature distribution and stress distribution; training the simulation results to construct a prediction model for predicting potting results; and obtaining real-time parameters during the potting process, where the real-time parameters include real-time potting speed, glue temperature, vacuum degree and potting time. The present invention helps to accurately predict and control the flow and filling of the glue during the potting process, thereby improving the accuracy and consistency of potting.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control, and particularly to a vacuum gluing control method and system based on machine learning. Background Art

[0002] In traditional vacuum gluing control methods, although the experience and skills of operators are referred to to a certain extent, they mainly rely on preset parameters and fixed control logics. In this method, operators will set a relatively fixed set of key parameters such as gluing speed, glue temperature, and vacuum degree according to the standard working procedures of gluing equipment, the recommended usage conditions of glue, and common environmental factors. However, due to the complexity and variability of the actual gluing environment, this control method based on preset parameters and fixed logics often fails to ensure the consistency and stability of gluing.

[0003] Specifically, when facing different gluing requirements, changes in glue characteristics, or fluctuations in environmental factors, the preset parameters and fixed control logics may not be able to adapt to these changes, resulting in problems such as glue overflow, uneven filling, or bubble residue during the gluing process. Although operators will try to adjust according to the actual situation, limited by the accuracy and timeliness of the adjustment, it is still difficult to completely avoid the occurrence of the above problems. These problems not only affect the quality and performance of products, but may also cause the shutdown of the production line and waste of resources. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a vacuum gluing control method and system based on machine learning, which helps to accurately predict and control the flow and filling of glue during the gluing process, thereby improving the accuracy and consistency of gluing.

[0005] To solve the above technical problem, the technical solution of the present invention is as follows:

[0006] In the first aspect, a vacuum gluing control method based on machine learning, the method includes:

[0007] Construct a three-dimensional geometric model of the gluing equipment according to the physical characteristics of the vacuum gluing process, and perform mesh division on the three-dimensional geometric model; define the geometric shape, material properties, boundary conditions, and initial conditions of the finite element model to simulate the real gluing environment, and the physical characteristics include the rheology and heat conductivity of the glue;

[0008] Simulate the gluing process under different gluing speeds, glue temperatures, and vacuum degrees through the three-dimensional geometric model of the gluing equipment to obtain simulation results, and the simulation results include glue flow behavior, temperature distribution, and stress distribution;

[0009] Train the simulation results to construct a prediction model for predicting gluing results;

[0010] Obtain real-time parameters during the potting process. The real-time parameters include real-time potting speed, glue temperature, vacuum degree, and potting time;

[0011] According to the real-time potting speed, glue temperature, vacuum degree, potting time, and the prediction model, obtain the predicted potting result;

[0012] Use the three-dimensional geometric model of the potting equipment to perform simulation analysis on the real-time potting speed, glue temperature, vacuum degree, and potting time to obtain the analysis result;

[0013] Evaluate the key indicators of the glue flow behavior and temperature distribution under the current potting conditions according to the analysis result;

[0014] Dynamically adjust the control parameters during the potting process according to the predicted potting result and the key indicators.

[0015] Furthermore, construct a finite element model according to the physical characteristics of the vacuum potting process, including:

[0016] Determine the physical phenomena, geometric dimensions, and simulation time to be simulated. The physical phenomena include glue flow behavior, temperature distribution, and stress distribution;

[0017] Determine the physical parameters and material properties to be simulated. The material properties include the density, viscosity, thermal conductivity, and specific heat capacity of the glue; the physical parameters include the vacuum degree, ambient temperature, pressure, and the geometric shape of the potting equipment;

[0018] Construct a three-dimensional geometric model of the potting equipment according to the physical phenomena, geometric dimensions, simulation time, physical parameters, and material properties. Among them, the three-dimensional geometric model of the potting equipment is a finite element model.

[0019] Furthermore, train the simulation results to construct a prediction model for predicting the potting result, including:

[0020] Preprocess the simulation result data to obtain the preprocessed simulation data set;

[0021] Divide the simulation data set into a training set and a test set;

[0022] According to the characteristics of the simulation data set, determine the neural network structure, activation function, and loss function. The neural network structure includes an input layer, a hidden layer, and an output layer;

[0023] Initialize the weights and biases of the neural network structure, and set the learning rate, batch size, and number of training epochs during the training process;

[0024] According to the simulation data set and the neural network structure, calculate the output data through forward propagation;

[0025] Calculate the loss value of the neural network structure according to the loss function, and update the weights and biases of the neural network structure through the backpropagation algorithm. Repeat the steps until the preset number of training epochs is reached to obtain the neural network model;

[0026] Use the test set to evaluate the neural network model and evaluate the performance metrics of the neural network model on the test set to obtain the evaluation results;

[0027] Optimize the neural network model according to the evaluation results to obtain the trained neural network model, where the trained neural network model is a prediction model.

[0028] Further, divide the simulated dataset into a training set and a test set, including:

[0029] Set the size of the particle swarm. Each particle represents a dataset partitioning scheme, including the ratio or specific indices of the training set and the test set; Initialize the position and velocity of each particle;

[0030] Determine the fitness function for evaluating the quality of the data partitioning scheme represented by each particle;

[0031] For each particle, update its position according to its current position and velocity to generate a new data partitioning scheme;

[0032] Update the individual position and global position of each particle, and adjust the velocity of each particle according to the individual position and global position until the preset number of iterations is reached to obtain the final partitioning scheme;

[0033] According to the final partitioning scheme, divide the simulated dataset into a training set and a test set.

[0034] Further, according to the simulated dataset and the neural network structure, calculate the output data through forward propagation, including:

[0035] Pass the input data in the simulated dataset to the input layer of the neural network;

[0036] For each neuron in the input layer, pass the input data to the first hidden layer. In this process, the input data is weighted and summed with the weights between the input data and the neurons in the first hidden layer, and the bias term is added;

[0037] For each hidden layer, receive the output data of the previous layer neurons as the input signal;

[0038] Perform weighted summation on the input signal and add the bias term to obtain the processing result, and process the processing result through the activation function to generate the output data of the current hidden layer;

[0039] The output data of the current hidden layer is passed to the next layer until the last hidden layer is reached;

[0040] The output data of the last hidden layer is passed to the output layer;

[0041] At the output layer, the output data undergoes weighted summation and adjustment of the bias term again, and the final prediction result of the neural network is calculated through the activation function. The final prediction result is the prediction result for the simulated data set.

[0042] Furthermore, the control parameters include the glue injection speed, glue temperature, and vacuum degree.

[0043] Furthermore, the calculation formula of the fitness function is:

[0044] ;

[0045] Wherein, is the weighted accuracy rate of the th particle; represents the number of different classes in the simulated data set; is the weight of the th class; is the number of true positives for the th particle, indicating the number of samples correctly predicted as the th class; is the number of false positives for the th particle, indicating the number of samples actually of the th class but wrongly predicted as the th class, where = ; is the cost of misclassifying the th class as the th class; represents the number of false negatives for the th particle, indicating the number of samples actually of the th class but not predicted; represents the index of the particle; represents the index of the class.

[0046] In a second aspect, a vacuum glue injection control system based on machine learning, applied to the method described above, includes:

[0047] A partitioning module, configured to construct a three-dimensional geometric model of the glue injection device according to the physical characteristics of the vacuum glue injection process, perform mesh partitioning on the three-dimensional geometric model; define the geometric shape, material properties, boundary conditions, and initial conditions of the finite element model to simulate the real glue injection environment, and the physical characteristics include the rheology and heat conductivity of the glue;

[0048] A simulation module, which is used to simulate the dispensing process under different dispensing speeds, glue temperatures, and vacuum degrees through a three-dimensional geometric model of a dispensing device to obtain simulation results. The simulation results include glue flow behavior, temperature distribution, and stress distribution;

[0049] A construction module, which is used to train the simulation results to construct a prediction model for predicting the dispensing results;

[0050] An acquisition module, which is used to acquire real-time parameters during the dispensing process. The real-time parameters include real-time dispensing speed, glue temperature, vacuum degree, and dispensing time;

[0051] A prediction module, which is used to obtain a predicted dispensing result according to the real-time dispensing speed, glue temperature, vacuum degree, dispensing time, and the prediction model;

[0052] An analysis module, which is used to perform simulation analysis on the real-time dispensing speed, glue temperature, vacuum degree, and dispensing time by using the three-dimensional geometric model of the dispensing device to obtain analysis results;

[0053] An evaluation module, which is used to evaluate the key indicators of the glue flow behavior and temperature distribution under the current dispensing conditions according to the analysis results;

[0054] An adjustment module, which is used to dynamically adjust the control parameters during the dispensing process according to the predicted dispensing result and the key indicators.

[0055] In a third aspect, a computing device includes:

[0056] One or more processors;

[0057] A storage device, which is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the above method.

[0058] In a fourth aspect, a computer-readable storage medium stores a program, and when the program is executed by a processor, the above method is implemented.

[0059] The above solution of the present invention has at least the following beneficial effects:

[0060] By constructing a three-dimensional geometric model of the dispensing device and combining machine learning technology, the present invention can accurately simulate the glue flow behavior, temperature distribution, and stress distribution under different dispensing conditions. This helps to accurately predict and control the flow and filling of the glue during the dispensing process, thereby improving the dispensing accuracy and consistency and reducing the occurrence of problems such as glue overflow and uneven filling.

[0061] The present invention uses a prediction model to predict and analyze the real-time glue filling speed, glue temperature, vacuum degree, and glue filling time, can timely detect potential glue filling problems, and dynamically adjust control parameters according to the analysis results. This optimization method can not only improve the glue filling efficiency, but also effectively avoid risks such as resource waste and production line shutdown.

[0062] Through machine learning and simulation analysis technologies, the present invention reduces the dependence on the experience of operators, enabling even inexperienced operators to complete high-quality glue filling work under the guidance of the system. Since the present invention can precisely control various parameters in the glue filling process to ensure that the glue is filled and cured in the best state, the quality and performance of the product can be significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a schematic flow chart of a vacuum glue filling control method based on machine learning provided by an embodiment of the present invention.

[0064] Figure 2 is a schematic diagram of a vacuum glue filling control system based on machine learning provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] Hereinafter, exemplary embodiments of the present disclosure will be described in more detail with reference to the drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.

[0066] As Figure 1 shown, an embodiment of the present invention provides a vacuum glue filling control method based on machine learning, and the method includes the following steps:

[0067] Step 11: Construct a three-dimensional geometric model of the glue filling device according to the physical characteristics of the vacuum glue filling process, and perform mesh division on the three-dimensional geometric model; define the geometric shape, material properties, boundary conditions, and initial conditions of the finite element model to simulate the real glue filling environment, and the physical characteristics include the rheology and heat conductivity of the glue.

[0068] Step 12: Simulate the glue filling process at different glue filling speeds, glue temperatures, and vacuum degrees through the three-dimensional geometric model of the glue filling device to obtain simulation results, and the simulation results include glue flow behavior, temperature distribution, and stress distribution.

[0069] Step 13: Train the simulation results to construct a prediction model for predicting the glue filling results.

[0070] Step 14, obtain the real-time parameters during the potting process. The real-time parameters include the real-time potting speed, glue temperature, vacuum degree, and potting time.

[0071] Step 15, based on the real-time potting speed, glue temperature, vacuum degree, potting time, and the prediction model, to obtain the predicted potting result.

[0072] Step 16, use the three-dimensional geometric model of the potting equipment to conduct a simulation analysis on the real-time potting speed, glue temperature, vacuum degree, and potting time, so as to obtain the analysis result.

[0073] Step 17, evaluate the key indicators of the glue flow behavior and temperature distribution under the current potting conditions according to the analysis result.

[0074] Step 18, dynamically adjust the control parameters during the potting process according to the predicted potting result and the key indicators. The control parameters include the potting speed, glue temperature, and vacuum degree.

[0075] In the embodiment of the present invention, in Step 11, by constructing an accurate three-dimensional geometric model and performing mesh division on it, the real potting environment can be more accurately simulated. At the same time, considering the physical properties such as the rheology and heat conduction of the glue, the simulation results are closer to the actual situation, improving the accuracy and credibility of the simulation. In Step 12, by simulating the potting process under different potting conditions, rich simulation results can be obtained, including glue flow behavior, temperature distribution, stress distribution, etc. In Step 13, through training the simulation results, a prediction model capable of predicting the potting result is constructed. This step enables the rapid prediction of the potting result according to the current parameters during the subsequent real-time potting process, so as to timely adjust the control parameters and ensure the potting quality and efficiency. In Step 14, the key parameters during the potting process are obtained in real time, including the potting speed, glue temperature, vacuum degree, potting time, etc., providing real-time data support for subsequent prediction and analysis. This helps to promptly detect abnormal situations during the potting process and take corresponding measures for intervention. In Step 15, by combining the real-time parameters and the prediction model, the predicted potting result can be obtained quickly. This enables the operator to timely understand the possible potting result during the potting process, so as to make corresponding adjustments and improve the accuracy and stability of the potting.

[0076] Step 16: Using the three-dimensional geometric model to conduct simulation analysis on real-time parameters can obtain more detailed analysis results, which helps to understand more deeply the characteristics such as the glue flow behavior and temperature distribution under the current potting conditions. Step 17: Evaluating key indicators based on the analysis results can quantitatively evaluate the characteristics such as the glue flow behavior and temperature distribution under the current potting conditions. This provides clear evaluation criteria for operators, enabling them to judge whether the potting process meets the expected requirements according to these indicators and take corresponding adjustment measures in a timely manner. Step 18: Dynamically adjusting control parameters according to the predicted potting results and key indicators can ensure that the potting process is always carried out in the best state. This not only improves the quality and efficiency of potting, but also reduces the dependence on the experience of operators, making the entire potting process more intelligent and automated. At the same time, by adjusting control parameters in real time, various emergencies and environmental changes can also be effectively dealt with to ensure the stability and reliability of the potting process.

[0077] In a preferred embodiment of the present invention, the above-mentioned step 11 of constructing a finite element model according to the physical characteristics of the vacuum potting process may include:

[0078] Step 111: Determine the physical phenomena, geometric dimensions, and simulation time to be simulated. The physical phenomena include glue flow behavior, temperature distribution, and stress distribution;

[0079] Step 112: Determine the physical parameters and material properties to be simulated. The material properties include the density, viscosity, thermal conductivity, and specific heat capacity of the glue; the physical parameters include the vacuum degree, ambient temperature, pressure, and geometric shape of the potting equipment;

[0080] Step 113: Construct a three-dimensional geometric model of the potting equipment according to the physical phenomena, geometric dimensions, simulation time, physical parameters, and material properties. Among them, the three-dimensional geometric model of the potting equipment is a finite element model.

[0081] In the embodiment of the present invention, in step 111, by determining the physical phenomena to be simulated (such as glue flow behavior, temperature distribution, and stress distribution), geometric dimensions, and simulation time, a clear direction and scope are provided for constructing an accurate finite element model. This helps to ensure the pertinence and accuracy of the simulation, making the simulation results more consistent with the physical characteristics of the actual glue filling process. In step 112, determining the physical parameters and material properties to be simulated is a key step in constructing the finite element model. By carefully considering material properties such as the density, viscosity, thermal conductivity, and specific heat capacity of the glue, as well as physical parameters such as the vacuum degree, ambient temperature, pressure, and geometric shape of the glue filling equipment, a more realistic and comprehensive model of the glue filling environment can be constructed. This helps to improve the accuracy and credibility of the simulation, thereby better predicting and controlling the actual glue filling process. In step 113, based on the physical phenomena, geometric dimensions, simulation time, physical parameters, and material properties determined in the previous steps, a three-dimensional geometric model of the glue filling equipment is constructed. This step is the basis of the entire simulation process. By accurately constructing the three-dimensional geometric model, the actual structure and working environment of the glue filling equipment can be more realistically reflected, which helps to more deeply understand the dynamic behavior of the glue filling process.

[0082] The following is a detailed description of the implementation process from step 111 to step 113:

[0083] In step 111, analyze the main physical phenomena involved in the vacuum glue filling process, which have a direct impact on the glue filling quality and efficiency; select the glue flow behavior, temperature distribution, and stress distribution as the key physical phenomena to be simulated because these phenomena determine the filling uniformity, curing effect, and structural strength of the glue. Measure or obtain the detailed geometric dimension data of the glue filling equipment (such as the glue filling chamber, rubber tube, nozzle, etc.); according to the actual size of the equipment, set the corresponding geometric parameters in the simulation software to ensure that the simulated geometric environment is consistent with the actual situation. Evaluate the duration of the glue filling process, which is based on the fluidity of the glue, the amount of glue filled, and the working efficiency of the equipment; set the total time range of the simulation to cover the entire process from the start to the end of the glue filling to ensure that all key dynamic changes can be captured.

[0084] In step 112, obtain the key material properties of the glue such as density, viscosity, thermal conductivity, and specific heat capacity through experimental measurement or by referring to relevant literature materials. These properties will directly affect the flow performance and heat transfer characteristics of the glue and must be accurately set in the simulation; evaluate the external physical conditions such as the vacuum degree, ambient temperature, and pressure involved in the glue filling process; according to the actual process requirements, set the specific values or change ranges of these physical parameters in the simulation; the geometric shape of the glue filling equipment will also be input as an important physical parameter into the simulation model.

[0085] Step 113: Collect detailed design data of the potting equipment, including CAD drawings, equipment specifications, etc.; select a suitable 3D modeling software (such as SolidWorks, AutoCAD, etc.) for model construction; in the modeling software, gradually construct each component of the potting equipment according to the geometric dimensions, physical parameters, and material properties determined in Steps 111 and 112; ensure the accuracy and detail integrity of the model, especially for key parts that affect glue flow and temperature distribution (such as the corners of the potting chamber, the structure of the nozzle, etc.); after completing the preliminary modeling, conduct model verification work, check for errors or unreasonable designs, and optimize and adjust the model according to the verification results until the simulation requirements are met.

[0086] For example, in the potting process of a certain electronic device, the physical phenomena to be simulated are the flow behavior of the glue in a vacuum environment, the change in temperature distribution during the potting process, and the stress distribution generated after the glue cures. The geometric dimensions of the equipment include the length, width, and height of the potting chamber, the diameter and length of the nozzle, etc. The simulation time is set to the time required from the start of potting to the complete curing of the glue, such as 30 minutes.

[0087] Through experimental measurement, the density of the glue is obtained as 1.2 g / cm³, the viscosity is 500 cP, the thermal conductivity is 0.2 W / (m·K), and the specific heat capacity is 1.5 J / (g·K). At the same time, according to the process requirements, the vacuum degree in the simulation is set to -0.1 MPa, the ambient temperature is 25 °C, and the pressure in the potting chamber is normal pressure. The geometric shape of the equipment is imported into the simulation software through accurate CAD drawings. Use SolidWorks software to construct a 3D geometric model of the potting equipment according to the CAD drawings and equipment specifications. The key components such as the potting chamber, nozzle, vacuum system, and temperature control system are presented in detail in the model. After verification and optimization, the model is used for subsequent simulation analysis.

[0088] In another preferred embodiment of the present invention, in the above Step 11, mesh generation is performed on the 3D geometric model; the geometric shape, material properties, boundary conditions, and initial conditions of the finite element model are defined to simulate the real potting environment. The physical properties include the rheology and thermal conductivity of the glue, and may include:

[0089] Open the mesh generation software, such as ANSYS Meshing or HyperMesh; import the previously constructed 3D geometric model file, ensuring that the file format is compatible with the software. Analyze the complexity of the model to determine the areas that require fine meshing, such as the glue flow path, temperature change sensitive areas, etc.; according to the simulation accuracy requirements, set the global mesh size and perform local mesh refinement on key areas to improve the simulation accuracy; select the mesh shape and type, such as tetrahedron, hexahedron, etc., to adapt to the geometric characteristics of the model; start the mesh generation program and monitor the generation process to ensure that no error or warning messages appear; after completion, check the quality of the generated mesh to ensure that there are no deformed meshes or excessive mesh aspect ratios; save and export the finite element mesh model for subsequent finite element analysis.

[0090] Define the geometric shape, material properties, boundary conditions, and initial conditions of the finite element model:

[0091] Open the finite element analysis software, such as ANSYS Workbench or Abaqus; import the previously generated finite element mesh model, ensuring that the geometric shape is accurately maintained; based on experimental data or literature, obtain key material properties of the glue, such as density, viscosity, thermal conductivity, specific heat capacity, etc.; create a new material property set for the glue material in the software and input the corresponding property values to ensure that the material properties are correctly applied to the corresponding mesh elements. According to the actual situation of the glue filling equipment, set fixed constraints, such as the fixed support of the equipment shell, etc.; define the inlet and outlet conditions of the glue, including flow rate, velocity distribution, or pressure distribution, etc.; set the boundary conditions of the vacuum degree to simulate the real vacuum glue filling environment, and specify external boundary conditions such as environmental temperature and pressure.

[0092] Determine the initial temperature of the glue at the start of the simulation and apply it to the corresponding mesh elements, set the initial flow velocity of the glue to reflect the state at the start of the glue filling process, and check and confirm whether all the initial conditions and boundary conditions are set correctly. After completing the above steps, a complete and accurate finite element model can be obtained for simulating the real glue filling environment and its physical properties, including the rheology and thermal conductivity of the glue. Next, simulation calculations can be performed to analyze the dynamic behavior and performance of the glue filling process.

[0093] In another preferred embodiment of the present invention, in step 12 above, the glue filling process under different glue filling speeds, glue temperatures, and vacuum degrees is simulated through the 3D geometric model of the glue filling equipment to obtain simulation results, and the simulation results include glue flow behavior, temperature distribution, and stress distribution, which may include:

[0094] Set different dispensing speeds according to experimental requirements or actual production scenarios. This can be achieved by adjusting the speed value or flow rate value in the inlet boundary conditions. Modify the initial temperature value of the glue to simulate the dispensing process at different temperatures. Ensure that this temperature value is within the tolerable range of the material. Adjust the vacuum degree setting in the simulation environment. This involves modifying the environmental pressure boundary conditions to reflect the influence of different vacuum degrees on the dispensing process.

[0095] Select a solver dedicated to fluid dynamics simulation according to the physical characteristics of the dispensing process. For example, if the simulation involves the viscous flow of glue, select a solver that can handle non-Newtonian fluids. Confirm whether the selected solver supports all physical effects involved in the simulation, such as heat conduction, convection, etc. Enter the parameter setting interface of the solver and set an appropriate time step according to the accuracy requirements of the simulation and the computing resources. A smaller time step can improve the accuracy of the simulation. Set the number of iterations. For complex nonlinear problems, multiple iterations are required to obtain a convergent solution. Set an upper limit for the number of iterations according to the complexity of the problem and the recommended value of the solver. Check and adjust other relevant parameters, such as convergence criteria, relaxation factors, etc., to ensure the stability and convergence of the solution process.

[0096] After confirming that all solver parameters have been set correctly and without errors, click the "Start" button or a similar button to start the simulation program. During the simulation run, closely monitor the output information of the software, including the calculation progress, residual convergence situation, etc. This information can help determine whether the simulation is running normally and whether the expected accuracy requirements have been met. If warning or error messages appear during the simulation, immediately pause the simulation and view the detailed information. Take corresponding measures for troubleshooting and repair according to the type and content of the warning or error. For example, it may be necessary to adjust the mesh generation, modify the boundary conditions, or re-set the solver parameters, etc. Ensure continuous monitoring of the software running status throughout the simulation until the simulation is completed and the final result data is output.

[0097] After the simulation is completed, extract the simulation result data. These data include the flow behavior of the glue at different time points (velocity field, pressure field, etc.), temperature distribution map, and stress distribution map; use the post-processing tools provided by the software to visually and quantitatively analyze the simulation results. For example, streamline diagrams can be created to show the flow path of the glue, or temperature contour maps can be generated to display the temperature distribution inside the device; according to the analysis results, evaluate the effects of the dispensing process under different parameter combinations. Pay special attention to whether the glue is evenly filled, whether there are potential problems such as excessive temperature gradients or stress concentrations.

[0098] According to the problems or deficiencies found in the simulation results, adjust the design parameters or process conditions of the potting equipment to optimize the potting process. Compare and verify the simulation results with the experimental results to ensure the accuracy and reliability of the simulation. This can be accomplished by designing corresponding experimental schemes and conducting potting experiments in the actual environment. According to the comparison results, further adjust and improve the simulation model and parameter settings to enhance the prediction ability of the simulation.

[0099] In another preferred embodiment of the present invention, step 13 of training the simulation results to construct a prediction model for predicting potting results may include:

[0100] Step 131, preprocess the simulation result data to obtain a preprocessed simulation data set;

[0101] Step 132, divide the simulation data set into a training set and a test set;

[0102] Step 133, determine the neural network structure, activation function, and loss function according to the characteristics of the simulation data set. The neural network structure includes an input layer, a hidden layer, and an output layer;

[0103] Step 134, initialize the weights and biases of the neural network structure, and set the learning rate, batch size, and number of training epochs during the training process;

[0104] Step 135, calculate the output data through forward propagation according to the simulation data set and the neural network structure;

[0105] Step 136, calculate the loss value of the neural network structure according to the loss function, and update the weights and biases of the neural network structure through the backpropagation algorithm. Repeat the steps until the preset number of training epochs is reached to obtain a neural network model;

[0106] Step 137, evaluate the neural network model using the test set, and evaluate the performance metrics of the neural network model on the test set to obtain an evaluation result;

[0107] Step 138, optimize the neural network model according to the evaluation result to obtain a trained neural network model, where the trained neural network model is a prediction model.

[0108] In an embodiment of the present invention, in step 131, data preprocessing can clean and organize the original data, eliminate outliers, missing values, and duplicate values, and improve the quality and consistency of the data. In step 132, by dividing the training set and the test set, it can be ensured that there is independent validation data during the training process of the model, thereby avoiding the overfitting phenomenon and improving the generalization ability of the model. In step 133, designing a suitable neural network structure, activation function, and loss function for the characteristics of the data set can enable the model to better capture the complex relationships in the data and improve the prediction accuracy and efficiency of the model. In step 134, initializing the weights and biases and setting the training parameters can accelerate the convergence speed of the model, reduce the training time, and at the same time avoid problems such as gradient disappearance or gradient explosion during the training process of the model. In step 135, forward propagation is the basic step of neural network training, and the output data is calculated. In step 136, by continuously optimizing the weights and biases of the neural network through the backpropagation algorithm, the prediction error of the model can be gradually reduced and the prediction performance of the model can be improved. At the same time, setting the number of training epochs can ensure that the model stops training after reaching a certain performance and avoid the performance degradation caused by overtraining. In step 137, using an independent test set to evaluate the model can objectively measure the performance of the model and ensure the reliability and effectiveness of the model in practical applications. In step 138, optimizing the model according to the evaluation results can further improve the performance of the model and make it better adapt to the requirements of the actual application scenario. At the same time, the optimization process can also help to discover potential problems in the model.

[0109] When specifically applied, each of the above steps includes the following processes:

[0110] In step 131, remove the invalid values, missing values, and outliers in the simulation result data; convert the simulation result data to the same dimension, and use z-score normalization or min-max normalization; according to the characteristics of the simulation data, select the features with high correlation with the glue injection result; perform transformations on the non-linear features, such as logarithmic transformation, polynomial transformation, etc., to improve the fitting ability of the model; encode the categorical features, such as one-hot encoding.

[0111] In step 132, randomly divide the simulation data set into a training set and a test set;

[0112] Step 133: Analyze the simulated dataset, including the type of data (such as numerical, categorical, etc.), range, and distribution; identify the key features in the dataset; determine the number of neurons in the input layer based on the number of features in the simulated dataset, with each neuron corresponding to a feature value; the number of hidden layers and the number of neurons in each layer are parameters when designing the neural network, and these parameters are determined through experience (such as gradually increasing the number of layers or neurons until the performance no longer improves); the number of neurons in the output layer is related to the complexity of the prediction task. In the prediction of the potting result, it may be one or more continuous values (such as potting amount, position, etc.), so the output layer can be set with the corresponding number of neurons. Select an activation function, for example, the ReLU function. Determine the loss function. For a regression task (such as potting amount prediction), the loss function includes the Mean Squared Error (MSE).

[0113] Step 134: Use a random initialization method (such as He initialization, Xavier initialization) to assign values to the weights and biases of the neural network; determine the learning rate (such as 0.01, 0.001, etc.), batch size (such as 32, 64, 128, etc.), and number of training epochs (such as 100, 200 epochs, etc.).

[0114] Step 135: Input the data in the training set into the neural network in batches; calculate the output value of each neuron layer by layer according to the hierarchical structure of the neural network until the result of the final output layer is obtained.

[0115] Step 136: Measure the difference between the network prediction and the true label through the loss function (such as the mean squared error). For the data in the current batch, compare the output of the network with the true label and calculate the loss value through the loss function. This loss value reflects the degree of mismatch between the model prediction and the true data under the current network parameters. Use the backpropagation algorithm to calculate the gradients of the loss value with respect to each weight and bias:

[0116] Backpropagation is an effective method for calculating the gradients of weights and biases in a neural network. In this process, first calculate the gradient of the output layer with respect to the loss function, and then propagate these gradients layer by layer forward until reaching the input layer of the network. For each layer, calculate the gradient of the loss function with respect to the weights and biases of that layer according to the chain rule. These gradients indicate how the network parameters should be adjusted to reduce the loss.

[0117] Update the weights and biases of the neural network according to the learning rate and the calculated gradients:

[0118] After obtaining the gradient information, gradient descent or its variants (such as Adam, RMSprop, etc.) can be used to update the weights and biases of the network. The learning rate is an important hyperparameter that controls the step size of parameter updates. A larger learning rate may lead to unstable training, while a smaller learning rate may result in slow training. During the update process, each weight and bias is subtracted by the product of its corresponding gradient and the learning rate. In this way, the parameters of the network will be adjusted in the direction of reducing the loss. Repeat the above steps until the preset number of training epochs is reached. Training a neural network requires multiple epochs of iteration. Each epoch will iterate through the entire dataset once (or multiple times, depending on the batch size); by continuously repeating the above steps, the parameters of the network will be gradually optimized, improving the performance of the model on the training data; when the preset number of training epochs is reached, the training process will stop to obtain the trained neural network model. At this time, the performance of the model on the validation set or test set can be evaluated to judge its generalization ability.

[0119] Step 137, input the data in the test set into the trained neural network; according to the true labels of the test set and the prediction results of the neural network, calculate performance metrics (such as accuracy, recall rate, F1 score, etc.).

[0120] Step 138, analyze the evaluation results of the test set to find problems existing in the neural network model (such as overfitting, underfitting, etc.); according to the analysis results, adjust the structure of the neural network (such as increasing or decreasing the number of hidden layers, the number of neurons, etc.), change the activation function or loss function, adjust the training parameters, etc., to optimize the performance of the model, and use the optimized model to repeat the training and evaluation steps until satisfactory performance is achieved.

[0121] In another preferred embodiment of the present invention, the above step 132 of dividing the simulated dataset into a training set and a test set may include:

[0122] Step 1321, set the size of the particle swarm. Each particle represents a dataset division scheme, including the ratio or specific indices of the training set and the test set; initialize the position and velocity of each particle;

[0123] Step 1322, determine the fitness function for evaluating the quality of the data division scheme represented by each particle;

[0124] Step 1323, for each particle, update its position according to its current position and velocity to generate a new data division scheme;

[0125] Step 1324, update the individual position and global position of each particle, and adjust the velocity of each particle according to the individual position and global position until the preset number of iterations is reached to obtain the final division scheme;

[0126] Step 1325, divide the simulated data set into a training set and a test set according to the final division scheme.

[0127] In the embodiment of the present invention, through the particle swarm optimization algorithm, the best scheme for data set division can be systematically searched. The design of the fitness function can ensure that the division scheme not only meets the basic training-test ratio requirements, but also takes into account more complex factors such as data distribution and class balance, thereby optimizing the effect of model training and verification. The particles in the particle swarm algorithm represent different division schemes, which means that the algorithm can flexibly explore various possible division methods. By adjusting the initial positions, velocities and update strategies of the particles, the algorithm can easily adapt to different data set characteristics and division requirements. The particle swarm optimization algorithm has a powerful global search ability. Through the update mechanism of individual positions and global positions, the algorithm can search for the optimal solution in the entire search space, avoid falling into local optima, and thus obtain a more reliable data set division scheme.

[0128] When specifically applied, the above step 1321 may include:

[0129] Determine the number of particles in the particle swarm, which is a preset parameter. Among them, too many particles may increase the computational burden, while too few particles may lead to insufficient search space and affect the optimization effect. Each particle represents a data set division scheme, which can be an array representing the indices of the training set and the test set, or a numerical value representing the division ratio. For example, if the data set has 1000 samples, a particle may be represented as a binary array containing 1000 elements, where 1 indicates that the sample corresponding to the index is in the training set and 0 indicates that it is in the test set. Initialize the particle positions and velocities. The position initialization can be random to ensure the diversity of the initial division scheme.

[0130] Step 1322, define a fitness function, which is used to evaluate the quality of the data division scheme represented by each particle.

[0131] Step 1323, according to the current positions and velocities of the particles, use the update formula of particle swarm optimization to calculate the new positions of the particles. Convert the updated particle positions into a new data set division scheme.

[0132] Step 1324, use the fitness function to evaluate the quality of the newly generated partitioning scheme for each particle. If the fitness of the new scheme is higher than the historical best fitness of the particle individual, update the individual optimal position of the particle. If the fitness of the new scheme is higher than the historical best fitness of the entire particle swarm, update the global optimal position. According to the individual optimal position and the global optimal position, use the velocity update formula of particle swarm optimization to adjust the velocity of each particle. Repeat steps 1323 and 1324 until the preset number of iterations is reached.

[0133] Step 1325, after the iteration ends, select the particle corresponding to the global optimal position as the final partitioning scheme, and divide the simulated data set into a training set and a test set according to this scheme.

[0134] In another preferred embodiment of the present invention, the calculation formula of the fitness function is:

[0135] ;

[0136] where is the weighted accuracy of the th particle; represents the number of different classes in the simulated data set; is the weight of the th class; is the number of true positives for the th particle, indicating the number of samples correctly predicted as the th class; is the number of false positives for the th particle, indicating the number of samples that are actually the th class but are wrongly predicted as the th class, where = ; is the cost of misclassifying the th class as the th class; represents the number of false negatives for the th particle, indicating the number of samples that are actually the th class but are not predicted; represents the index of the particle; represents the index of the class.

[0137] In the embodiment of the present invention, through the weight , the function can handle unbalanced data sets, assign higher importance to minority classes, and thus increase the attention of the model to these classes. Through the misclassification cost , this function not only considers the accuracy of the prediction but also takes into account the costs of different types of mispredictions. This is very useful in many practical applications. The function comprehensively considers true positives ( ), false positives ( ), and false negatives ( ), and can evaluate the performance of the model more comprehensively. This fitness function is particularly suitable for the Particle Swarm Optimization (PSO) algorithm, where each particle represents a potential solution. By calculating the weighted accuracy of each particle, the PSO algorithm can effectively search for the optimal solution in the solution space. By considering various types of prediction errors, this function helps to promote the generalization ability of the model, enabling the model to not only perform well on the training data.

[0138] For the -th particle, its velocity in the -dimensional space is updated according to the formula:

[0139] ;

[0140] where and are the velocities of the -th particle in the -dimensional space at times and , respectively. is the inertia weight, which is used to balance the global search and local search capabilities. and are the learning factors, represents the individual learning factor, represents the social learning factor. and are random numbers between [0, 1], which are used to increase the randomness of the search. is the individual best position of the -th particle in the -dimensional space. is the global best position of all particles in the -dimensional space. is the position of the -th particle in the -dimensional space at time .

[0141] The position update formula is:

[0142] For the -th particle, its position in the -dimensional space is updated based on the velocity, specifically:

[0143] ;

[0144] Among them, and are respectively the th particle in the -dimensional space, at time and position.

[0145] In another preferred embodiment of the present invention, in step 135, according to the simulation data set and the neural network structure, the output data is calculated through forward propagation, including:

[0146] Step 1351, passing the input data in the simulation data set to the input layer of the neural network, specifically including: reading the simulation data set, which is a structured data matrix, where each row represents a sample and each column represents a feature; initializing the input layer of the neural network to ensure that the number of neurons in the input layer matches the number of features in the simulation data set; taking each sample in the simulation data set (i.e., each row in the data matrix) as the input data and passing it to the corresponding neurons in the input layer;

[0147] Step 1352, for each neuron in the input layer, passing the input data to the first hidden layer. In this process, the input data is weighted and summed with the weights between the input data and the neurons in the first hidden layer, and a bias term is added, specifically including: for each neuron in the input layer, weighting and summing the input data it receives with the weights connected to the neurons in the first hidden layer; adding a bias term (bias) to the result of the weighted sum, which is a learnable parameter used to adjust the output of the neuron, and completing the weighted sum and bias term addition operations from all neurons in the input layer to the neurons in the first hidden layer;

[0148] Step 1353, for each hidden layer, receiving the output data of the neurons in the previous layer as the input signal, specifically including: for each hidden layer in the neural network (starting from the first hidden layer), its input signal is the output data of the neurons in the previous layer (i.e., the upper layer); each neuron in the current hidden layer will receive the output data from all neurons (or specific connected neurons, depending on the network structure) in the previous layer;

[0149] Step 1354: Perform weighted summation on the input signal and add a bias term to obtain a processing result, and process the processing result through an activation function to generate the output data of the current hidden layer. Specifically, it includes: in the current hidden layer, perform a weighted summation operation on the received input signal (i.e., the output data of the neurons in the previous layer), using the connection weights between the neurons in the current layer and the neurons in the previous layer; add the bias term of the neurons in the current layer to the result of the weighted summation; pass the result of adding the weighted summation and the bias term to an activation function (such as ReLU, Sigmoid, etc.), and the activation function will perform a non-linear transformation on this result to generate the output data of the neurons in the current hidden layer;

[0150] Step 1355: The output data of the current hidden layer is passed to the next layer until the last hidden layer is reached. Specifically, it includes: after the current hidden layer processes the input signal and generates output data, these output data will be passed to the next layer (i.e., a deeper hidden layer or the output layer), and this passing process is a continuation of the forward propagation, and the data flows forward layer by layer in the neural network;

[0151] Step 1356: The output data of the last hidden layer is passed to the output layer. Specifically, it includes: when the data is passed to the last hidden layer and processed, the output data of this layer will be passed to the output layer of the neural network, and the output layer is responsible for receiving the output of the last hidden layer and preparing for the final prediction or decision;

[0152] Step 1357: In the output layer, the output data undergoes weighted summation and bias term adjustment again, and the final prediction result of the neural network is calculated through an activation function. The final prediction result is the prediction result for the simulated dataset. Specifically, it includes: in the output layer, perform a weighted summation operation on the received output data of the last hidden layer, using the connection weights between the neurons in the output layer and the neurons in the last hidden layer, and add the bias term of the neurons in the output layer to the result of the weighted summation; pass the result of adding the weighted summation and the bias term to the activation function of the output layer (in some cases, such as in regression tasks, an activation function may not be used or a linear activation function may be used); the result after being processed by the activation function is the final prediction result of the neural network, and this result is output and compared with the true labels in the simulated dataset to evaluate the performance of the network.

[0153] In an embodiment of the present invention, in step 1351, which is the starting point for the neural network to process data, it ensures that the original data can be received by the network and further processed. In step 1352, this step realizes the preliminary conversion and feature extraction of the data. The introduction of weighted summation and bias terms enables the network to learn complex patterns in the data. In step 1353, the design of the hidden layer enables the network to learn deep features of the data, and each layer abstracts and extracts higher-level information. In step 1354, the introduction of the activation function brings non-linear characteristics to the network, enabling the network to learn and represent complex non-linear relationships. In step 1355, through multi-layer transmission, the network can gradually extract and process the features in the data, realizing the recognition and learning of complex patterns. In step 1356, this step marks the end of the feature extraction stage, and the output layer is ready to receive the final feature representation for decision-making or prediction. In step 1357, the processing of the output layer enables the network to make a final prediction based on the learned feature representation. The activation function may be used for tasks such as classification and regression here, achieving the functional goals of the neural network.

[0154] When specifically applied, a certain manufacturing enterprise decides to use a neural network model to predict and control the potting process in order to improve the automation level and product quality of the vacuum potting process. First, a series of simulation data is collected for training and testing the neural network model.

[0155] The simulation dataset contains 1000 samples, and each sample has 5 characteristic parameters, namely: glue temperature (°C), vacuum pressure (kPa), glue flow rate ( ml / s ), potting time (s), and glue viscosity (cP). These parameters are organized into a 1000×5 data matrix, where each row represents a sample and each column represents a characteristic.

[0156] For prediction, a neural network with two hidden layers is designed, and each hidden layer has 8 neurons. The input layer has 5 neurons (matching the number of characteristics), and the output layer has 1 neuron for predicting the potting quality index (a value between 0 and 1, and the closer it is to 1, the higher the quality).

[0157] Specific steps of forward propagation:

[0158] In step 1351, take out the first sample from the simulation dataset (i.e., the first row of the data matrix), which is a vector containing 5 characteristic values: [glue temperature, vacuum pressure, glue flow rate, potting time, glue viscosity]. Pass this vector to the input layer of the neural network, and each characteristic value corresponds to an input neuron.

[0159] Step 1352: Each neuron in the input layer performs a weighted sum of the feature values it receives with the weights connected to the first hidden layer and adds a bias term. For example, the first neuron in the first hidden layer calculates: w1 × glue temperature + w2 × vacuum pressure + w3 × glue flow rate + w4 × potting time + w5 × glue viscosity + b, where w1 - w5 are the weights and b is the bias term. All neurons in the first hidden layer perform similar calculations.

[0160] Steps 1353 to 1354: The first hidden layer passes the processed data to the second hidden layer. Each neuron in the second hidden layer again performs a weighted sum, bias term adjustment, and activation function processing. An activation function (such as ReLU) is used to introduce non-linearity so that the neural network can learn complex patterns.

[0161] Steps 1355 to 1356: After the second hidden layer processes the data, it passes the data to the output layer. In this example, the output layer has only one neuron, which is responsible for predicting the potting quality index.

[0162] Step 1357: The neuron in the output layer performs a final weighted sum, bias term adjustment, and activation function processing on the received data. This processed value is the prediction result of the neural network for the potting quality index of the first sample. Repeat the above steps until all samples in the simulation dataset are processed. Each sample will obtain a predicted value of the potting quality index.

[0163] In another preferred embodiment of the present invention, step 14 is to obtain real-time parameters during the potting process. The following is the specific implementation process of this step:

[0164] Install sensors on the potting equipment to monitor the potting speed, glue temperature, vacuum degree, and potting time in real time. Continuously collect real-time data during the potting process through the sensors. This data can be transmitted to the data processing system through data lines or wirelessly. Perform preprocessing on the collected real-time data, including filtering, denoising, unit conversion, etc., to ensure the quality and consistency of the data. Store the preprocessed real-time data in the database for subsequent analysis and use. Display the real-time parameters through a user interface (such as a dashboard, curve graph, etc.) so that the operator can monitor the status of the potting process in real time.

[0165] Step 15 is to obtain the predicted potting result based on the real-time parameters and the prediction model. The following is the specific implementation process of this step:

[0166] Take the real-time parameters (glue filling speed, glue temperature, vacuum degree, and glue filling time) obtained in step 14 as input data, and prepare to input them into the prediction model. Ensure that the format of the input data is consistent with the data format used during the training of the prediction model. This includes processing such as data normalization and encoding. Input the formatted real-time parameters into the prediction model for forward propagation calculation to obtain the predicted glue filling results. The prediction results may include key indicators such as glue filling volume and glue filling position. Output the prediction results to the user interface or store them in the database for subsequent analysis and use.

[0167] Step 16 is to perform simulation analysis on the real-time parameters using a three-dimensional geometric model. The following is the specific implementation process of this step:

[0168] Based on the actual structure and dimensions of the glue filling equipment, establish an accurate three-dimensional geometric model. This can be accomplished using professional modeling software (such as SolidWorks, AutoCAD, etc.). Set the simulation parameters corresponding to the real-time parameters in the three-dimensional geometric model, including glue filling speed, glue temperature, vacuum degree, and glue filling time, etc. Use fluid dynamics simulation software (such as ANSYS Fluent, COMSOL Multiphysics, etc.) to perform simulation analysis on the three-dimensional geometric model with the set simulation parameters, and pay attention to key indicators such as the flow behavior of the glue and temperature distribution during the analysis process. After the simulation analysis is completed, extract and save the analysis results. These results include key data such as the glue flow velocity field, temperature field, and pressure field.

[0169] Step 17 is to evaluate the key indicators based on the analysis results. The following is the specific implementation process of this step:

[0170] Visualize the analysis results obtained in step 16, such as displaying key indicators such as the flow behavior of the glue and temperature distribution through cloud maps, vector maps, curve graphs, etc. Based on the visualized analysis results, evaluate the key indicators such as the flow behavior of the glue and temperature distribution. During the evaluation process, preset standards or historical data can be referred to determine whether the current glue filling conditions meet the requirements. If problems with key indicators (such as poor glue flow or uneven temperature distribution) are found during the evaluation process, the specific causes and locations of the problems need to be identified. Record the evaluation results and the identified problems for subsequent parameter adjustment and optimization.

[0171] Step 18 is to dynamically adjust the control parameters based on the predicted glue filling results and key indicators. The following is the specific implementation process of this step:

[0172] Based on the predicted potting results obtained in step 15 and the key indicators evaluated in step 17, determine the control parameters to be adjusted and their adjustment directions (such as increasing the potting speed, decreasing the glue temperature, etc.). Through the control system of the potting equipment, the control parameters determined to be adjusted are adjusted in real time. During the adjustment process, the response speed and stability of the equipment should be concerned to ensure that the adjustment effect meets the expectations. After the adjustment is completed, perform the potting operation again, and monitor the key indicators during the potting process and the final potting results in real time. By comparing the data before and after the adjustment, verify the effect of the parameter adjustment. If the verification result shows that the parameter adjustment is effective, iterative optimization can be continued to further improve the efficiency and stability of the potting process. If the verification effect is not good, the reasons need to be re-analyzed and the strategy adjusted.

[0173] As Figure 2 shown, an embodiment of the present invention further provides a vacuum potting control system 20 based on machine learning, including:

[0174] A partitioning module 21 for constructing a three-dimensional geometric model of the potting equipment according to the physical characteristics of the vacuum potting process, and performing mesh partitioning on the three-dimensional geometric model; defining the geometric shape, material properties, boundary conditions and initial conditions of the finite element model to simulate the real potting environment, and the physical characteristics include the rheology and heat conductivity of the glue;

[0175] A simulation module 22 for simulating the potting process at different potting speeds, glue temperatures, and vacuum degrees through the three-dimensional geometric model of the potting equipment to obtain simulation results, and the simulation results include glue flow behavior, temperature distribution and stress distribution;

[0176] A construction module 23 for training the simulation results to construct a prediction model for predicting potting results;

[0177] An acquisition module 24 for acquiring real-time parameters during the potting process, and the real-time parameters include real-time potting speed, glue temperature, vacuum degree and potting time;

[0178] A prediction module 25 for obtaining predicted potting results according to the real-time potting speed, glue temperature, vacuum degree and potting time and the prediction model;

[0179] An analysis module 26 for performing simulation analysis on the real-time potting speed, glue temperature, vacuum degree and potting time by using the three-dimensional geometric model of the potting equipment to obtain analysis results;

[0180] An evaluation module 27 for evaluating the key indicators of the glue flow behavior and temperature distribution under the current potting conditions according to the analysis results;

[0181] An adjustment module 28 for dynamically adjusting the control parameters during the potting process according to the predicted potting results and key indicators.

[0182] It should be noted that this system corresponds to the above method, and all implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0183] An embodiment of the present invention further provides a computing device, including: a processor and a memory storing a computer program. When the computer program is run by the processor, it executes the method as described above. All implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0184] An embodiment of the present invention further provides a computer-readable storage medium storing instructions. When the instructions are run on a computer, the computer is caused to execute the method as described above. All implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0185] The above is the preferred embodiment of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A vacuum potting control method based on machine learning, characterized in that, The method includes: Construct a three-dimensional geometric model of the potting equipment according to the physical characteristics of the vacuum potting process, and perform mesh division on the three-dimensional geometric model; define the geometric shape, material properties, boundary conditions and initial conditions of the finite element model to simulate the real potting environment, and the physical characteristics include the rheology and heat conductivity of the glue; Simulate the potting process at different potting speeds, glue temperatures, and vacuum degrees through the three-dimensional geometric model of the potting equipment to obtain simulation results, and the simulation results include glue flow behavior, temperature distribution, and stress distribution; Training the simulation results to build a prediction model for predicting the potting results, including: preprocessing the simulation result data to obtain a preprocessed simulation data set; dividing the simulation data set into a training set and a test set, including: setting the size of the particle swarm, where each particle represents a data set division scheme, including the ratio or specific indices of the training set and the test set; initializing the position and velocity of each particle; calculating the difference between the number of samples correctly predicted as the class and the cost of the number of samples that are actually the class but are wrongly predicted as the class, calculating the sum of the number of samples correctly predicted as the class and the number of samples that are actually the class but are not wrongly predicted as the class; determining a fitness function for evaluating the quality of the data division scheme represented by each particle according to the ratio of the difference to the sum; for each particle, updating its position according to its current position and velocity to generate a new data division scheme; updating the individual position and global position of each particle according to the fitness result of the new data division scheme, and adjusting the velocity of each particle according to the individual position and global position until a preset number of iterations is reached to obtain a final division scheme; dividing the simulation data set into a training set and a test set according to the final division scheme; Obtain real-time parameters during the potting process, and the real-time parameters include real-time potting speed, glue temperature, vacuum degree, and potting time; Obtain the predicted potting result according to the real-time potting speed, glue temperature, vacuum degree, potting time and the prediction model; Use the three-dimensional geometric model of the potting equipment to perform simulation analysis on the real-time potting speed, glue temperature, vacuum degree, and potting time to obtain analysis results; Evaluate the key indicators of the glue flow behavior and temperature distribution under the current potting conditions according to the analysis results; Dynamically adjust the control parameters during the potting process according to the predicted potting result and the key indicators.

2. The vacuum encapsulation control method based on machine learning according to claim 1, wherein Construct a finite element model according to the physical characteristics of the vacuum potting process, including: Determine the physical phenomena, geometric dimensions, and simulation time to be simulated, and the physical phenomena include glue flow behavior, temperature distribution, and stress distribution; Determine the physical parameters and material properties to be simulated, and the material properties include the density, viscosity, thermal conductivity coefficient, and specific heat capacity of the glue; the physical parameters include the vacuum degree, ambient temperature, pressure, and geometric shape of the potting equipment; Construct a three-dimensional geometric model of the potting equipment according to the physical phenomena, geometric dimensions, simulation time, physical parameters, and material properties, where the three-dimensional geometric model of the potting equipment is a finite element model.

3. The vacuum encapsulation control method based on machine learning according to claim 2, wherein Train the simulation results to construct a prediction model for predicting the potting result, including: Determine the neural network structure, activation function, and loss function according to the characteristics of the simulation dataset. The neural network structure includes an input layer, a hidden layer, and an output layer; Initialize the weights and biases of the neural network structure, and set the learning rate, batch size, and number of training epochs during the training process; Calculate the output data through forward propagation according to the simulation dataset and the neural network structure; Calculate the loss value of the neural network structure according to the loss function, and update the weights and biases of the neural network structure through the backpropagation algorithm. Repeat the steps until the preset number of training epochs is reached to obtain the neural network model; Evaluate the neural network model using the test set, and evaluate the performance indicators of the neural network model on the test set to obtain the evaluation result; Optimize the neural network model according to the evaluation result to obtain the trained neural network model, where the trained neural network model is the prediction model.

4. The vacuum encapsulation control method based on machine learning according to claim 3, wherein Calculate the output data through forward propagation according to the simulation dataset and the neural network structure, including: Transfer the input data in the simulation dataset to the input layer of the neural network; For each neuron in the input layer, the input data is passed to the first hidden layer. During this process, the input data is weighted and summed with the weights between the input data and the neurons in the first hidden layer, and a bias term is added. For each hidden layer, the output data of the neurons in the previous layer is received as the input signal. The input signal is weighted and summed and a bias term is added to obtain the processing result, and the processing result is processed through an activation function to generate the output data of the current hidden layer. The output data of the current hidden layer is passed to the next layer until the last hidden layer is reached. The output data of the last hidden layer is passed to the output layer. In the output layer, the output data is again adjusted by weighted summation and bias terms, and the final prediction result of the neural network is calculated through an activation function. The final prediction result is the prediction result for the simulated dataset.

5. A vacuum potting control method based on machine learning according to claim 4, characterized in that The control parameters include the glue injection speed, glue temperature, and vacuum degree.

6. A vacuum potting control method based on machine learning according to claim 5, characterized in that, The calculation formula of the fitness function is: ； Among them, is the weighted accuracy of the th particle; represents the number of different classes in the simulated dataset; is the weight of the th class; is the number of true positives for the th particle, representing the number of samples correctly predicted as the th class; is the number of false positives for the th particle, representing the number of samples that are incorrectly predicted as the th class when they are actually of the th class. Among them, ; is the cost of misclassifying the th class as the th class represents the number of false negatives for the th particle, representing the number of samples that are incorrectly not predicted as the th class when they are actually of the th class; represents the index of the particle; represents the index of the class.

7. A vacuum potting control system based on machine learning, characterized in that, Applied to the method described in any one of claims 1 to 6, it includes: A partitioning module for constructing a three-dimensional geometric model of the glue injection device according to the physical characteristics of the vacuum glue injection process, and performing mesh partitioning on the three-dimensional geometric model; defining the geometric shape, material properties, boundary conditions, and initial conditions of the finite element model to simulate the real glue injection environment, and the physical characteristics include the rheology and heat conductivity of the glue. A simulation module for simulating the glue injection process at different glue injection speeds, glue temperatures, and vacuum degrees through the three-dimensional geometric model of the glue injection device to obtain simulation results, and the simulation results include glue flow behavior, temperature distribution, and stress distribution. A construction module for training the simulation results to construct a prediction model for predicting the glue injection results. An acquisition module for acquiring real-time parameters during the glue injection process, and the real-time parameters include the real-time glue injection speed, glue temperature, vacuum degree, and glue injection time. A prediction module for obtaining the predicted glue injection results based on the real-time glue injection speed, glue temperature, vacuum degree, glue injection time, and the prediction model. An analysis module for performing simulation analysis on the real-time glue injection speed, glue temperature, vacuum degree, and glue injection time using the three-dimensional geometric model of the glue injection device to obtain analysis results. An evaluation module for evaluating the key indicators of the glue flow behavior and temperature distribution under the current glue injection conditions according to the analysis results. An adjustment module for dynamically adjusting the control parameters during the glue injection process according to the predicted glue injection results and the key indicators.

8. A computing device, characterized in that, It includes: One or more processors; A storage device for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, A program is stored in the computer-readable storage medium, and when the program is executed by a processor, the method described in any one of claims 1 to 6 is implemented.

Images