A Multi-Objective Debugging Method and System for Injection Molding Machine Process Based on 3D Simulation and Depth Modeling

Abstract

This invention provides a multi-objective debugging method for injection molding machine processes based on 3D simulation and depth modeling. The method includes: acquiring a 3D model of the machine and mold and collecting relevant characteristics; scanning the physical object and constructing a digital twin model; constructing a physical simulation model based on thermodynamics and kinetics and performing mold flow analysis; analyzing the effect of process parameters on corresponding machine-mold-plastic combinations based on the physical simulation data; constructing a transferable conditional variational autoencoder and generating multiple sets of initial process parameters according to the expected effect; constructing a multi-objective Bayesian optimization structure and exploring the parameter space through physical simulation to obtain the Pareto optimal solution. This method improves the efficiency and accuracy of multi-objective debugging of injection molding machine processes through the collaborative work of feature transfer layers, conditional variational autoencoders, Gaussian process regression, multi-objective Bayesian optimization, and physical simulation.

Description

Technical Field

[0001] This invention relates to the field of intelligent injection molding technology, and in particular to a multi-objective debugging method and system for injection molding machine processes based on 3D simulation and depth modeling. Background Technology

[0002] In injection molding production, setting process parameters is a crucial step that determines product quality, production efficiency, and cost. Traditional parameter tuning heavily relies on engineers' long-term experience, using a "trial and error" approach to repeatedly adjust, test molds, and measure on a real injection molding machine until a qualified product is produced. This method has the following significant drawbacks: 1. Long debugging cycle and high cost: Each trial molding requires a certain amount of raw materials, energy and manpower time. Various practical factors restrict the long cycle required for the formal use of new molds or new materials. 2. High dependence on experience: The quality of debugging is directly linked to the individual experience of engineers. Knowledge is difficult to standardize and pass on, and personnel changes will bring significant production risks. 3. Poor flexibility: When molds, materials or machines change, a lot of debugging work needs to be done again, making it difficult to respond quickly to changes; 4. Low predictability: It is difficult to accurately predict the defects of the final product (such as warping, shrinkage marks, weld lines, etc.) during the parameter setting stage. Problems are often exposed in the later stages of trial molding, resulting in greater waste of resources, time and manpower.

[0003] Furthermore, emerging technologies are actively addressing these challenges due to the limitations of traditional trial-and-error methods. For instance, simulation software for injection molding has been gradually applied in production, but it primarily serves as a tool for mold design and problem analysis, failing to effectively integrate real-time data with debugging parameters for in-depth analysis and decision-making. There are also methods for parameter analysis based on historical debugging big data, but these lack real-time data feedback, resulting in poor accuracy and limited generalization capabilities, making it difficult to meet the debugging needs of new machines and materials. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for multi-objective debugging of injection molding machine processes based on 3D simulation and depth modeling, so as to improve the efficiency and accuracy of multi-objective debugging of injection molding machine processes.

[0005] To achieve the above objectives, the present invention provides the following solution: A multi-objective debugging method for injection molding machine processes based on 3D simulation and depth modeling includes the following steps: Determine the process parameters and constraints that need to be modified for the injection molding machine, mold, and rate material, and optimize the process parameters based on the constraints. A 3D model is constructed based on the point cloud data of the injection molding machine and the mold, and the geometric information and material properties of the 3D model are extracted. The parameter effects of the three-dimensional model are evaluated based on the kinetic and thermodynamic models to obtain the three-dimensional evaluation results. A transferable conditional variational autoencoder is constructed, and the model is trained based on process parameters and 3D evaluation results to obtain candidate parameters. The transferable conditional variational autoencoder consists of a feature transfer layer, an encoder, and a decoder. Construct a proxy model and use the proxy model to perform multi-objective optimization on the candidate parameters to obtain the optimal parameters; The actual machine system is evaluated based on the optimal parameters, and the actual machine system is debugged based on the evaluation results.

[0006] Optionally, the process parameters that need to be modified include: barrel temperature, number of mold opening sections, number of mold closing sections, number of feeding sections, number of injection sections, number of holding pressure sections, pressure of each section, speed of each section, time of each section, cooling time, delay time and demolding time; the optimization objectives for optimizing the process parameters include: quality expressed as defect rate, efficiency expressed as cycle time and energy consumption expressed as cycle energy consumption.

[0007] Optionally, a 3D model is constructed based on the point cloud data of the injection molding machine and the mold, and the geometric information and material properties of the 3D model are extracted, including: Point cloud data of the injection molding machine and mold are obtained using a 3D scanner; Perform downsampling and noise reduction processing on point cloud data; The denoised point cloud data is smoothed and reconstructed using Poisson surfaces to obtain a 3D model. Obtain the geometric information of the 3D model by consulting relevant materials; Model flow analysis is performed on the 3D model to obtain material properties.

[0008] Optionally, the three-dimensional model is evaluated for parametric effects based on the kinetic and thermodynamic models to obtain three-dimensional evaluation results, including: The three-dimensional model was simulated using dynamic and thermodynamic models to obtain simulation results. Defect fingerprints are constructed based on defects in the finished product from the injection molding process, and the simulation results are evaluated based on the defect fingerprints to obtain the quality evaluation results. The efficiency of the simulation results is evaluated based on the time parameters to obtain the efficiency evaluation results. Energy consumption is predicted based on the simulation results using the constructed energy consumption model, and energy consumption assessment results are obtained.

[0009] Optionally, a transferable conditional variational autoencoder is constructed, and the model is trained based on process parameters and three-dimensional evaluation results to obtain candidate parameters, including: By binding the process parameters with the 3D evaluation results using feature relationships, the first dataset is obtained. A feature transfer layer is constructed using a domain adversarial neural network architecture, and the feature transfer layer is trained using a first dataset; the feature transfer layer includes a feature extraction network, a domain discriminator network, and a task prediction network; The output of the feature transfer layer is reparameterized and sampled by the encoder. The decoder performs feature fusion on the output of the encoder and the output of the feature transfer layer. The decoder output is filtered based on process parameter constraints to obtain candidate parameters.

[0010] Optionally, a feature transfer layer is constructed using a domain adversarial neural network architecture, and the feature transfer layer is trained using a first dataset, including: The process parameters and the three-dimensional evaluation results are used as the source domain dataset, and the first dataset is used as the target domain dataset. The source domain dataset and the target domain dataset are mapped to the latent feature space through a feature extraction network to obtain latent features; The feature sources of potential features are distinguished by a domain discriminator network; The task prediction network performs error prediction on latent features, minimizes the source domain prediction error, and maximizes the domain discriminator classification error.

[0011] Optionally, the process of generating the overall loss function of the transferable conditional variational autoencoder includes: Calculate the maximum mean difference loss of the feature extraction network to obtain the first loss function, which is expressed as: ;in For kernel function mapping, For the regenerating nucleus Hilbert space, and These are the latent features of the source domain dataset and the target domain dataset, respectively; The second loss function is obtained by calculating the loss function of the domain discriminator network using binary cross-entropy, and its expression is: ;in, The probability that a feature originates from the source domain. The vector input consists of the first dataset. The total number of data points in the source domain dataset. The total number of data points within the target domain dataset; The loss function of the task prediction network is calculated using the mean squared error, resulting in the third loss function, expressed as: ;in, The performance evaluation scores obtained within the dataset. The prediction score is used to evaluate the performance of the task prediction network output. The overall loss function is calculated based on the first loss function, the second loss function, and the third loss function.

[0012] Optionally, a surrogate model is constructed, and the candidate parameters are optimized using the surrogate model to obtain the optimal parameters, including: A second dataset was obtained through simulation and actual measurement; A Gaussian process model is constructed based on a Bayesian network structure, and the model is trained using a second dataset. The candidate parameters are used as the initial sample points for Bayesian optimization to perform simulation, and the first simulation result is obtained. The first simulation result is then simulated using a trained Gaussian process model to obtain the second simulation result. The trained Gaussian process model is updated based on the second simulation results until the model converges and the optimal parameters are obtained.

[0013] A multi-objective debugging system for injection molding machine processes based on 3D simulation and depth modeling, comprising: The human-machine interaction module is used to determine the process parameters and parameter constraints that need to be modified for the injection molding machine, mold, and rate material, and to optimize the process parameters based on the parameter constraints. The digital twin model building module is used to build a 3D model based on the point cloud data of the injection molding machine and the mold, and to extract the geometric information and material properties of the 3D model. The physics simulation module is used to evaluate the parametric effects of the 3D model based on the kinetic and thermodynamic models, and obtain the 3D evaluation results. The multi-objective parameter generation module is used to construct a transferable conditional variational autoencoder and train the model of the transferable conditional variational autoencoder based on process parameters and three-dimensional evaluation results to obtain candidate parameters. The multi-objective parameter optimization module is used to build a surrogate model and perform multi-objective optimization on candidate parameters through the surrogate model to obtain the optimal parameters; The parameter verification and migration module is used to evaluate the parameters of the actual machine system based on the optimal parameters, and to debug the actual machine system based on the evaluation results.

[0014] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The present invention provides a multi-objective debugging method for injection molding machine processes based on 3D simulation and depth models. This method includes: determining the process parameters and parameter constraints that need to be modified for the injection molding machine, mold, and rate material, and optimizing the process parameters according to the parameter constraints; constructing a three-dimensional model based on point cloud data of the injection molding machine and mold, and extracting the geometric information and material properties of the three-dimensional model; evaluating the parameter effects of the three-dimensional model based on a kinetic model and a thermodynamic model to obtain a three-dimensional evaluation result; constructing a transferable conditional variational autoencoder, and training the transferable conditional variational autoencoder model based on the process parameters and the three-dimensional evaluation result to obtain candidate parameters; constructing a surrogate model, and performing multi-objective optimization on the candidate parameters through the surrogate model to obtain the optimal parameters; evaluating the parameters of the actual machine system based on the optimal parameters, and debugging the actual machine system based on the evaluation results. This method achieves automated generation and optimization of process parameters through the collaborative work of feature transfer layer, conditional variational autoencoder, Gaussian process regression, multi-objective Bayesian optimization and physical simulation. It has good generalization and adaptability, can adapt to different injection molding scenarios and production needs, and improves the level of intelligence in injection molding production. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the multi-objective debugging method for injection molding machine processes according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the Bayesian optimization network structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-objective debugging system for injection molding machine processes according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this invention provides a multi-objective debugging method for injection molding machine processes based on 3D simulation and depth modeling, comprising the following steps: Step 100: Determine the process parameters and parameter constraints that need to be modified for the injection molding machine, mold, and rate material, and optimize the process parameters according to the parameter constraints; Step 200: Construct a 3D model based on the point cloud data of the injection molding machine and mold, and extract the geometric information and material properties of the 3D model; Step 300: Evaluate the parameter effects of the three-dimensional model based on the kinetic and thermodynamic models to obtain the three-dimensional evaluation results; Step 400: Construct a transferable conditional variational autoencoder and train the model of the transferable conditional variational autoencoder according to the process parameters and the three-dimensional evaluation results to obtain candidate parameters; the transferable conditional variational autoencoder consists of a feature transfer layer, an encoder and a decoder. Step 500: Construct a surrogate model and perform multi-objective optimization on the candidate parameters through the surrogate model to obtain the optimal parameters; Step 600: Evaluate the parameters of the actual machine system based on the optimal parameters, and debug the actual machine system based on the evaluation results.

[0020] In the specific implementation process, step 100 involves the user selecting the machine-mold-plastic combination through the human-machine interface, inputting specific data, importing the CAD model, and then processing it with numbers to obtain the feature value vector of the parameter usage scenario combination. Common process parameters for injection molding machines include: barrel temperature of each section, number of mold opening sections, number of mold closing sections, number of feeding sections, number of injection sections, number of holding pressure sections, pressure of each section, speed of each section, time of each section, cooling time, delay time, demolding time, etc. This embodiment specifically includes: barrel temperature, cooling time, demolding time, and pressure, speed, and time of mold opening, mold closing, feeding, injection, and holding pressure, totaling 18 parameters. Specific use requires adding or removing parameter types according to the actual needs of the machine. While determining the parameter types, the parameter value range is also determined as parameter constraints, and the optimization goals for parameter debugging are determined. Optimization goals include: quality expressed as defect rate, efficiency expressed as cycle time, and energy consumption expressed as cycle energy consumption.

[0021] In the specific implementation process, step 200 constructs a three-dimensional model including the mold, barrel, and screw based on the point cloud data of the injection molding machine and mold. The three-dimensional model is then analyzed using the instruction manual and relevant materials to obtain the specific heat capacity, density, and wall thickness of the barrel and mold. Simultaneously, the viscosity curve and PVT relationship of the plastic are obtained through a mold flow analysis material database. Specifically, the point cloud data is processed using a PCL point cloud library. First, statistical filtering and downsampling techniques are used to denoise the point cloud data. Then, smoothing and Poisson surface reconstruction are performed to generate a smooth, high-precision CAD three-dimensional model. The model is then placed into 3D modeling software to obtain geometric information such as the wall thickness and screw diameter of the mold and barrel. Material information such as the specific heat capacity of the barrel and mold is obtained through the instruction manual and relevant materials. Finally, material properties such as the viscosity curve and PVT relationship of the plastic are obtained through a mold flow analysis material database.

[0022] In the specific implementation process, step 300 uses mold flow software such as Moldflow and Moldex3D to perform mold flow analysis on the 3D model, simulating the mold opening and closing, material feeding, injection, holding pressure, and cooling processes during injection molding. The goal of physical simulation is mainly to verify parameters, therefore, the simulation results need to be evaluated, primarily in terms of quality, efficiency, and energy consumption. In the quality evaluation process, product quality is mainly reflected in the product's defect rate. Defects in injection molded products mainly include short shots, weld lines, burrs, poor demolding, cracks, shrinkage, bubbles, black spots, and color mixing. These defects have numerous causes and diverse manifestations; therefore, a defect fingerprint is established for each type of defect. Based on the simulation, for short shot defects, the ratio of material volume to mold cavity volume is calculated; for weld line defects, the ratio of the meeting angle at the melt front meeting point to the acceptable angle is used; and for burr defects, the degree of mold closure is used during the simulation. Based on actual machine production, short-shot defects are reflected by calculating the average ratio of multiple finished product volumes to the mold cavity volume; weld line defects are reflected by the ratio of the weld line merging angle to the acceptable angle; and burr defects are reflected by observing the burr formation on multiple finished products. Similarly, data from mold flow software and actual production are analyzed to determine whether parameters cause defects. When constructing the evaluation model, the scores obtained from simulation evaluation are consistent with the scoring standards of actual machine evaluation, and the scores for the same parameter should be as close as possible. This forms a one-dimensional vector representing the defect analysis results. ,in This represents the score of the i-th type of defect, thus further normalizing the quality assessment as follows: ,in This indicates the types of defects used for quality assessment. This represents the weight of the i-th type of defect score. Let represent the standard score of the i-th type of defect.

[0023] In efficiency evaluation, product production efficiency is directly linked to the time required to complete a process cycle. A significant portion of an injection molding cycle time consists of mold closing time, injection time, holding pressure time, cooling time, mold opening time, demolding time, and cycle interval time. These durations are all included in the process parameters, allowing the total cycle time to be directly obtained through parameter analysis. ,in, Indicates the types of time used for efficiency evaluation. Let represent the duration of the i-th type. By analyzing historical production data, a duration range can be defined, ensuring that the longest cycle time of the normal process flow does not exceed this range, thus normalizing the duration data as follows: ,in, This indicates the longest cycle time that a theoretically normal process flow can accept.

[0024] In the energy consumption assessment process, the energy consumption of injection molding machines accounts for a large proportion of the production process, mainly the energy consumption of power, heating, cooling, and monitoring and control equipment. For example, the energy consumption of power-providing equipment such as pumps and motors is mainly related to the set duration, required pressure, and flow rate of each action in the process. The energy consumption for heating the barrel is mainly related to the temperature of each section and the properties of the plastic itself, such as specific heat capacity. The energy consumption for cooling the mold is mainly related to the properties of the cooling medium and the cooling time. The energy consumption of control and monitoring equipment such as screens, controllers, and sensors is related to the machine's own configuration and is relatively fixed. To calibrate the energy consumption of simulation data, the configuration of the machines used is analyzed, and energy consumption models strongly correlated with the set parameters are built based on machine classification for energy consumption prediction. For the actual machine parameter data, the changes in electricity meter values ​​within a cycle are directly obtained, and the obtained energy consumption data is normalized. , ,in This indicates the total energy consumption per cycle in this process flow. This indicates the energy consumption generated by the power system. The energy consumed to heat the mold barrel The energy consumed to cool the mold. This indicates the maximum cycle energy consumption that can be accepted under normal theoretical process flow.

[0025] In the specific implementation process, step 400 constructs a transferable conditional variational autoencoder. It uses the determined type of process parameters as input, the quality, efficiency, and energy consumption of the parameters as output, the characteristics of the parameter usage scenarios as the feature transfer layer, and actual machine output data and simulation model generated data as the dataset for model training. Multiple sets of input values ​​are then inferred from the determined output values. Specifically, the transferable conditional variational autoencoder includes a feature transfer layer, an encoder, and a decoder. The feature transfer layer is responsible for indicating the applicable scenarios where the process parameters produce corresponding effects. Considering that new machine models, molds, and plastic combinations require a small amount of data to complete model construction, the feature transfer layer enables the new combination model to perform transfer learning on an older combination model trained with a large amount of data. In this embodiment, this is manifested by aligning the features of the target domain data containing the new combination and the source domain data consisting of the old combination, thereby outputting consistent input features. The encoder is used to learn the distribution of latent variables conditioned on effect evaluation. The decoder is used to reconstruct the output process parameters conditioned on effect evaluation.

[0026] Specifically, the construction process of a transferable conditional variational autoencoder includes: Step 401: Dataset Preparation. The dataset required for the transferable conditional variational autoencoder is classified by features, with process parameters as input and effect evaluation as output. Specifically, it consists of: Classification features: features that represent the operating environment of the parameters, such as machine features, mold parameters, and plastic properties. In this embodiment, seven data closely related to the effect of the parameters are used as feature vectors: machine model, screw diameter, mold volume, mold wall thickness, runner length, plastic density, and plastic viscosity. Process parameters: adjustable parameters that affect the process results. In this embodiment, 18 parameters are used as feature vectors, including barrel temperature, cooling time, demolding time, and pressure, speed, and time for mold opening, mold closing, feeding, injection, and holding. Effect evaluation: a feature vector composed of evaluation scores for quality, efficiency, and energy consumption. Each process parameter feature vector under a classification feature corresponds to a production target result, and a production target feature vector can correspond to multiple process parameter feature vectors. To convert the data into a computable quantity, a maximum standard value is set to normalize each data point and map the data to the [0,1] interval. The expression is: , ,in The vector consists of 18 parameters, including barrel temperature, cooling time, demolding time, and the pressure, speed, and time of mold opening, mold closing, material feeding, injection, and holding pressure. The evaluation scores for quality, efficiency, and energy consumption form a 3-dimensional vector. To illustrate the specific characteristics of the machine-mold-plastic combination, this embodiment uses seven parameters—machine model, screw diameter, mold volume, mold wall thickness, runner length, plastic density, and plastic viscosity—to form a seven-dimensional vector.

[0027] Step 402: Construct a feature transfer layer using a domain adversarial neural network architecture, and output the transfer features of the new combined dataset using the original machine-mold-plastic combined dataset. The feature transfer layer consists of a feature extraction network, a domain discriminator network, and a task prediction network, with each network having two fully connected hidden layers and one fully connected output layer. The historical machine-mold-plastic combined dataset is used as the source domain dataset. The newly combined dataset will be used as the target domain dataset. The original input is processed through a feature extraction network. x,c Mapping to the latent feature space, the expression is: ; ; ; in, express( x,c Potential characteristics of ).

[0028] The domain discriminator network is used to determine whether a feature comes from the source domain or the target domain. The expression is: ; ; ; in, , , These represent layers 1, 2, and 3 of the network, respectively. The weight matrix W and bias coefficient b represent the probability that a feature originates from the source domain. W and b represent the weight matrix and bias coefficients of each fully connected network layer, respectively. The weight matrix W and bias coefficient b of each fully connected network layer can be obtained by allocating weights to the input data of the activation function (e.g., increasing the weight of real data and decreasing the weight of simulation data) and adjusting the biases according to the model performance. The method for obtaining W and b in subsequent expressions is the same.

[0029] The task prediction network is used for feature prediction, and its expression is: ; ; ; It should be noted that, in order to achieve feature alignment, it is necessary to minimize the source domain prediction error while maximizing the classification error of the domain discriminator. The task prediction network uses mean squared error to calculate the loss, expressed as: ;in, The performance evaluation scores obtained within the dataset. The prediction score is used to evaluate the performance of the task prediction network output. The domain discrimination loss of the domain discriminator network is calculated using binary cross-entropy, and the expression is: ,in, The probability that a feature originates from the source domain. The vector input consists of the first dataset. The total number of data points in the source domain dataset. Let be the total number of data points in the target domain dataset; and to enhance feature alignment, the maximum mean difference (MMD) loss is introduced, expressed as: ,in For kernel function mapping, a Gaussian kernel is used for calculation. For the regenerating kernel Hilbert space, the kernel function expression is: Therefore, the overall loss function is: , of which For MMD loss weights, The parameters of the domain discriminator are weighed. Finally, the dataset processed in step 401 is input into the feature transfer layer for training. First, the data in the source and target domains are normalized, mapping the parameter values ​​to the [0,1] interval. Then, the network parameters of the feature extractor and the domain discriminator are randomly initialized, with a learning rate of 0.001 and a training batch size of 32. After forward propagation, the parameters are obtained ( x,c The prediction results are used, and the loss is calculated through backpropagation. When the total loss decreases by less than 10 consecutive training batches... The training is completed in time.

[0030] Step 403: Train the conditional variational autoencoder, using aligned features from the feature transfer layer output, conditioned on the evaluation results, and outputting the model with process parameters and machine-mold-plastic combination features. Specifically, the conditional variational autoencoder trains the model based on the aligned features from the feature transfer layer output, conditioned on the evaluation results, and outputting the model with process parameters and machine-mold-plastic combination features. y In this case, the learning conditions are distributed as follows It consists of an encoder and a decoder, wherein the encoder is composed of... Learning from given input and effectiveness evaluation objectives y The distribution of latent variables, the decoder is composed of The variables z and y are reconstructed from the latent variable z and the effect evaluation target y, and trained by maximizing the variational lower bound ELBO.

[0031] Furthermore, the encoder consists of an input layer, a fully connected hidden layer, and a latent variable output layer. The input layer is the aligned output of the feature transfer layer. , represented as The hidden layer consists of two fully connected layers, expressed as: ; ; The output implicit variable parameters are: ; ; Then, reparameterized sampling is performed to obtain: .

[0032] The decoder consists of an input layer, a fully connected hidden layer, and an output layer. The input layer takes the aligned input from the feature transfer layer. The vector formed by concatenating the latent variable Z generated by the encoder is represented as: The hidden layer consists of two fully connected layers, expressed as: ; ; The output of the output layer is represented as follows: ; The total loss of the variational autoencoder is then calculated using the variational lower bound as follows: ; In this embodiment, the total loss can be expressed as a combination of the MSE reconstruction loss and the KL divergence loss, with the following expression: ; ; ; Finally, the aligned feature vectors from the feature transfer layer are input to the encoder and decoder. The encoder outputs the latent variables and the aligned feature vectors. y To the decoder, through forward propagation of the output ( x,c After calculating the variational lower bound, backpropagation is performed to update the network until the total loss reaches the minimum value, thus completing the training.

[0033] Step 404: Using a conditional variational autoencoder, input features are used to call corresponding features and evaluate the expected results, generating multiple sets of process parameters that meet the requirements, and then filtering the constraints. Specifically, a conditional variational autoencoder trained on the target machine-mold-plastic combination dataset is selected, and multiple sets of latent variables z are randomly sampled from the prior distribution. The expected effect evaluation feature vector is input into the decoder to generate multiple sets of process parameters. These parameters are then filtered according to preset process parameter constraints. If too few process parameters meet the conditions, the hidden variables will be randomly sampled to output multiple sets of process parameters for further filtering.

[0034] In the specific implementation process, step 500 uses Gaussian process regression as a surrogate model. Parameters meeting the requirements are selected through parameter constraints. Multiple sets of qualified parameters are used as initial sample points for multi-objective optimization, and the results are verified through a simulation model until the Pareto front converges to obtain the optimal parameters. The specific steps of multi-objective Bayesian optimization include: Step 501: Construct a training set for training the Gaussian process regression model. This dataset directly uses input parameters obtained through simulation and those generated in practice, along with performance evaluation data. A Bayesian network structure is then built based on accumulated experience. For example... Figure 2 As shown, specifically, with efficiency as the optimization objective, the shorter the duration of the entire cycle, the higher the efficiency; with energy consumption as the optimization objective, the lower the pressure and flow rate used in the entire cycle, and the lower the heating temperature required, the lower the energy consumption; with quality as the optimization objective, defects such as short shots, weld lines, burrs, poor demolding, cracks, shrinkage, bubbles, black spots, and color mixing are specifically displayed as feature vectors. Based on experience, parameter modification suggestions are provided to reduce the defect rate. In this embodiment, if a short shot occurs, the defect is reduced by increasing the injection speed, heating temperature, and injection holding time; if a weld line appears, the defect is reduced by increasing the heating temperature and injection speed; if burrs appear, the defect is reduced by decreasing the heating temperature, injection speed, holding pressure, and holding time. Based on the corresponding parameter correction direction given for each objective, an initial Bayesian network structure is constructed, and the probability distribution of each parameter with respect to the specific optimization objective is determined through directed graphs and datasets.

[0035] Step 502: Construct a Gaussian process model based on a Bayesian network structure and train the model using the dataset. Specifically, a Gaussian process is a set of random variables, where any finite number of random variables follow a joint Gaussian distribution. The Gaussian process model assumes that the model follows a Gaussian process, expressed as: ,in This represents the normalized value of the process parameters. i Indicates the first i One goal, It is a mean function. Let covariance function be used. Let represent the evaluation effect value of the i-th target. By calculating the probability distribution of the process parameters and evaluation effects in the dataset one-to-one correspondence, the mean and variance of the above expression are obtained as the learning strategy. Then, based on the empirical strategy constructed by the Bayesian network structure, process parameter values ​​that have little impact on the target are eliminated, thereby completing the combination of Gaussian process model and Bayesian network structure.

[0036] Step 503: Use the multiple sets of parameters that meet the requirements obtained in Step 400 as the initial sample points for Bayesian optimization to perform simulation, and input the simulation results into the Gaussian process model for further simulation to obtain the dataset. And simulation results.

[0037] Step 504: Update the Gaussian process model based on the obtained simulation results, obtain the next set of evaluation point records, and select the sample points that have the largest trend of minimizing the objective function. Specifically, by inputting the simulation dataset into the GP model, the newly selected process parameters are calculated. Corresponding evaluation effect The posterior mean and variance, where the posterior mean is... Indicates to Best estimate, variance This indicates that the uncertainty of the model estimation decreases as the number of iterations increases. The process is repeated until convergence, all parameters and corresponding evaluation data are recorded and analyzed, and the Pareto optimal parameters are obtained by constructing an objective function based on the preset optimization direction.

[0038] Furthermore, based on the preset optimization direction, this embodiment defines the objective function for optimization as follows: ; in Indicates the number of defects. This indicates the weight of each objective. This represents the normalized score for each objective. Then, the acquisition function is used to guide the next step of parameter selection for each objective, resulting in the expected improvement function EI: ; in This is the best objective function value found so far. These are the process parameters to be simulated in the next step. Finally, record all the simulation results. The process continues until the Pareto front converges, the maximum number of iterations is reached, or the change in the acquisition function value is less than a threshold. Then, the process parameters that meet the requirements are selected through the expected process parameter constraints, and the process parameters that satisfy the optimization bias are found in the Pareto dominant solution as the Pareto optimal solution and the optimal parameters.

[0039] In the specific implementation process, step 600 deploys the optimal parameters in the actual machine and evaluates the performance of the parameters in terms of quality, efficiency and energy consumption. If the effect does not meet the expected goal, steps 400 and 500 are repeated until the expected effect is achieved (in this embodiment, the difference in evaluation results is less than 0.1%). The new data generated in the process is used as a dataset for transfer learning.

[0040] like Figure 3 As shown, the present invention also provides a multi-objective debugging system for injection molding machine processes based on 3D simulation and depth modeling, comprising: The human-machine interaction module is used to determine the process parameters and parameter constraints that need to be modified for the injection molding machine, mold, and rate material, and to optimize the process parameters based on the parameter constraints. The digital twin model building module is used to build a 3D model based on the point cloud data of the injection molding machine and the mold, and to extract the geometric information and material properties of the 3D model. The physics simulation module is used to evaluate the parametric effects of the 3D model based on the kinetic and thermodynamic models, and obtain the 3D evaluation results. The multi-objective parameter generation module is used to construct a transferable conditional variational autoencoder and train the model of the transferable conditional variational autoencoder based on process parameters and three-dimensional evaluation results to obtain candidate parameters. The multi-objective parameter optimization module is used to build a surrogate model and perform multi-objective optimization on candidate parameters through the surrogate model to obtain the optimal parameters; The parameter verification and migration module is used to evaluate the parameters of the actual machine system based on the optimal parameters, and to debug the actual machine system based on the evaluation results.

[0041] In some other embodiments, the standard workflow is divided into three stages according to the debugging system: "initialization and modeling", "parameter generation and optimization", and "real-world verification and evolution". The specific interaction steps are as follows: Phase 1: Initialization and Modeling 1. User command input On the human-machine interface, engineers select or input the injection molding machine model, mold model, and plastic grade for this debugging task. They can choose to import the initial CAD models of the mold and key injection molding machine components (barrel, screw). Engineers set the range of process parameters (constraints) and the optimization goals for this debugging task (prioritizing efficiency improvement, energy consumption reduction, or balanced optimization), and input specific expected values ​​for quality, efficiency, and energy consumption (expected defect rate below 1%, cycle time less than 30 seconds). After confirmation, the human-machine interface module decomposes the task instructions and distributes them to the relevant modules.

[0042] 2. Digital Twin Model Construction Upon receiving instructions, the digital twin model building module directly invokes an existing accurate model; otherwise, it initiates the building process. This module calls a 3D scanner control program to scan the actual machine and mold, acquiring point cloud data. Internally, the module performs point cloud denoising, smoothing, and surface reconstruction to generate a high-precision CAD model. Geometric parameters (mold wall thickness, screw diameter) are extracted from the model, and combined with external databases (material library, instruction manual) to obtain thermodynamic parameters (specific heat capacity, density) and material properties (viscosity curve of plastic, PVT relationship). Finally, the digital twin model building module encapsulates all geometric and physical parameters and sends them to the physical simulation module.

[0043] Phase Two: Parameter Generation and Optimization 1. Initial candidate parameter generation The human-computer interaction module sends the user-defined quality, efficiency, and energy consumption targets, as well as equipment combination characteristics (machine model, screw diameter, etc.), to the multi-target parameter generation module.

[0044] 2. Parameter Generation After receiving the request, the multi-objective parameter generation module first uses its feature transfer layer to map the newly combined features to a feature space aligned with the source domain, utilizing existing source domain knowledge. Then, it combines the user's objective and the sampled latent variables to generate multiple sets of candidate process parameters through the decoder. The module then performs preliminary constraint screening to remove parameters that clearly exceed the set range.

[0045] 3. Refined multi-objective optimization After receiving multiple sets of initial process parameters, the multi-objective parameter optimization module initiates an iterative optimization loop. The optimizer uses all currently known parameters and corresponding effect data to update its internal Gaussian process regression model, constructing a probabilistic surrogate model for the objective function. This model can predict the mean and variance of any new parameter. The optimizer calculates the acquisition function to find the next parameter point with the greatest potential to improve the target value. The optimizer sends the new parameters to the physical simulation module for simulation evaluation. The physical simulation module then returns the simulation results of the new parameters to the optimizer, which adds this pair of new data to the dataset and updates the surrogate model. This generation-simulation-model update process is repeated until Pareto front convergence, the number of iterations is exhausted, or the change in the acquisition function value is less than a threshold. Based on the user-defined optimization bias, the multi-objective parameter optimization module selects one or more sets of "optimal process parameters" from the found Pareto optimal solution set and sends them to the human-computer interaction module for display, while simultaneously sending them to the parameter verification and transfer module for real-world testing.

[0046] Phase Three: Real-world Validation and Evolution 1. Real-world deployment and verification After receiving the optimal parameters, the parameter verification and migration module prompts the engineer through the human-machine interaction module, requesting on-machine verification. Upon engineer confirmation, the system sends the parameters to the controller of the actual injection molding machine. The injection molding machine begins trial production. The parameter verification and migration module collects on-machine operating data using sensors (electricity meters, pressure sensors, and a vision inspection system) deployed on the machine, and calculates the machine's quality, efficiency, and energy consumption evaluation values ​​according to the effect evaluation criteria in step 300 of the method.

[0047] 2. Effect Evaluation and Model Evolution The parameter verification and migration module compares the actual evaluation values ​​with the user-defined target and calculates the deviation Δ. If Δ is less than a preset threshold (0.1%), the debugging is successful, and the parameter verification and migration module reports the final result to the human-computer interaction module, ending the process. If Δ is too large, a new round of optimization is triggered. The parameter verification and migration module sends a re-optimization instruction to the multi-objective parameter generation module and the multi-objective parameter optimization module. The new round of optimization will be based on a larger dataset, including the actual data from this test, to generate more accurate parameters.

[0048] 3. Perform model transfer learning The parameter validation and transfer module adds the complete data generated from this real-world validation to the training dataset, and then starts fine-tuning training of the parameter generation model (especially its feature transfer layer) in the multi-objective parameter generation module. This process usually uses a small learning rate to allow the model to quickly adapt to the real physical characteristics of the new combination and complete an "empirical evolution".

[0049] Through the close collaboration and data loop of the above six modules, a complete automated process from "user intent" to "optimal parameters" and then to "model evolution" is realized, which can intelligently, efficiently and adaptively solve multi-objective debugging problems in injection molding process.

[0050] The beneficial effects of this invention are as follows: 1) By replicating the actual process flow as closely as possible in a virtual simulation environment, engineers can obtain the best parameters and evaluate the effects without repeatedly testing on real machines. This greatly reduces the number of actual machine tests, saves time, materials and energy, improves debugging convenience, and reduces time and manpower costs. 2) By analyzing 3D simulation data and deep learning models, the parameter space can be automatically explored to find the optimal debugging results, reducing the reliance on debugging experience and realizing automated debugging; 3) By comprehensively considering indicators such as quality, efficiency, and energy consumption, and finding the Pareto optimal solution under multi-objective constraints, the process debugging was completed from multiple objectives, and the quality of the debugging results was improved. 4) Hybrid models combine physical laws and data laws, possessing both the interpretability and extrapolation capabilities of physical simulations and the powerful exploration and fitting capabilities of data-driven models. They can quickly adapt to new equipment and materials, and flexibly adapt to various scenarios, thus enhancing generalization capabilities.

[0051] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0052] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A multi-objective debugging method for injection molding machine processes based on 3D simulation and depth modeling, characterized in that, Includes the following steps: Determine the process parameters and parameter constraints that need to be modified for the injection molding machine, mold, and rate material, and optimize the process parameters according to the parameter constraints; A three-dimensional model is constructed based on the point cloud data of the injection molding machine and the mold, and the geometric information and material properties of the three-dimensional model are extracted. The three-dimensional model is evaluated based on the kinetic and thermodynamic models to obtain the three-dimensional evaluation results. A transferable conditional variational autoencoder is constructed, and the model of the transferable conditional variational autoencoder is trained according to the process parameters and the three-dimensional evaluation results to obtain candidate parameters; the transferable conditional variational autoencoder consists of a feature transfer layer, an encoder and a decoder. Construct a proxy model, and perform multi-objective optimization on the candidate parameters through the proxy model to obtain the optimal parameters; The actual machine system is evaluated based on the optimal parameters, and the actual machine system is debugged based on the evaluation results.

2. The multi-objective debugging method for injection molding machine processes based on 3D simulation and depth model as described in claim 1, characterized in that, The process parameters that need to be modified include: barrel temperature, number of mold opening sections, number of mold closing sections, number of feeding sections, number of injection sections, number of holding pressure sections, pressure of each section, speed of each section, time of each section, cooling time, delay time, and demolding time; the optimization objectives for optimizing the process parameters include: quality expressed as defect rate, efficiency expressed as cycle time, and energy consumption expressed as cycle energy consumption.

3. The multi-objective debugging method for injection molding machine processes based on 3D simulation and depth model as described in claim 1, characterized in that, A three-dimensional model is constructed based on the point cloud data of the injection molding machine and the mold, and the geometric information and material properties of the three-dimensional model are extracted, including: The point cloud data of the injection molding machine and the mold are obtained by a 3D scanner; The point cloud data is downsampled and denoised. The denoised point cloud data is smoothed and reconstructed using a Poisson surface to obtain the three-dimensional model. The geometric information of the three-dimensional model was obtained by consulting relevant materials; The material properties are obtained by performing model flow analysis on the three-dimensional model.

4. The multi-objective debugging method for injection molding machine processes based on 3D simulation and depth model according to claim 1, characterized in that, The three-dimensional model is evaluated for its parameter effects based on the kinetic and thermodynamic models, yielding three-dimensional evaluation results, including: Based on the kinetic model and the thermodynamic model, the three-dimensional model is subjected to flow simulation to obtain simulation results; A defect fingerprint is constructed based on the defects in the finished product from the injection molding process, and the simulation results are evaluated based on the defect fingerprint to obtain the quality evaluation result; The efficiency of the simulation results is evaluated based on the time parameters to obtain the efficiency evaluation results. The energy consumption is predicted by the constructed energy consumption model based on the simulation results, and the energy consumption assessment results are obtained.

5. The multi-objective debugging method for injection molding machine processes based on 3D simulation and depth model according to claim 1, characterized in that, A transferable conditional variational autoencoder is constructed, and the model is trained based on the process parameters and the three-dimensional evaluation results to obtain candidate parameters, including: The process parameters are bound to the three-dimensional evaluation results by feature relationships to obtain the first dataset; The feature transfer layer is constructed using a domain adversarial neural network architecture and trained using the first dataset; the feature transfer layer includes a feature extraction network, a domain discriminator network, and a task prediction network; The encoder is used to reparameterize and sample the output of the feature transfer layer. The decoder performs feature fusion on the output of the encoder and the output of the feature transfer layer. The output of the decoder is filtered based on process parameter constraints to obtain the candidate parameters.

6. The multi-objective debugging method for injection molding machine processes based on 3D simulation and depth model according to claim 5, characterized in that, The feature transfer layer is constructed using a domain adversarial neural network architecture, and trained using the first dataset, including: The process parameters and the three-dimensional evaluation results are used as the source domain dataset, and the first dataset is used as the target domain dataset. The feature extraction network maps both the source domain dataset and the target domain dataset to the latent feature space to obtain latent features. The domain discriminator network distinguishes the feature sources of the potential features; The task prediction network performs error prediction on the latent features, minimizing the source domain prediction error and maximizing the domain discriminator classification error.

7. The multi-objective debugging method for injection molding machine processes based on 3D simulation and depth model according to claim 5, characterized in that, The process of generating the overall loss function of the transferable conditional variational autoencoder includes: The maximum mean difference loss of the feature extraction network is calculated to obtain the first loss function, which is expressed as: ;in For kernel function mapping, For the regenerating nucleus Hilbert space, and These are the latent features of the source domain dataset and the target domain dataset, respectively; The loss function of the domain discriminator network is calculated using binary cross-entropy, resulting in the second loss function, expressed as: ;in, The probability that a feature originates from the source domain. The vector input consists of the first dataset. The total number of data points in the source domain dataset. The total number of data points within the target domain dataset; The loss function of the task prediction network is calculated using the mean squared error, resulting in the third loss function, expressed as: ;in, The performance evaluation scores obtained within the dataset. The prediction score is used to evaluate the performance of the task prediction network output. The overall loss function is calculated based on the first loss function, the second loss function, and the third loss function.

8. The multi-objective debugging method for injection molding machine processes based on 3D simulation and depth model according to claim 1, characterized in that, Construct a proxy model and perform multi-objective optimization on the candidate parameters using the proxy model to obtain the optimal parameters, including: A second dataset was obtained through simulation and actual measurement; A Gaussian process model is constructed based on a Bayesian network structure, and the Gaussian process model is trained using the second dataset. The candidate parameters are used as initial sample points for Bayesian optimization to perform simulation, resulting in a first simulation result. The first simulation result is then simulated using a trained Gaussian process model to obtain a second simulation result. The trained Gaussian process model is updated based on the second simulation results until the model converges, thus obtaining the optimal parameters.

9. A multi-objective debugging system for injection molding machine processes based on 3D simulation and depth modeling, characterized in that, include: The human-machine interaction module is used to determine the process parameters and parameter constraints that need to be modified for the injection molding machine, mold, and rate material, and to optimize the process parameters according to the parameter constraints. A digital twin model construction module is used to construct a three-dimensional model based on the point cloud data of the injection molding machine and the mold, and to extract the geometric information and material properties of the three-dimensional model; The physical simulation module is used to evaluate the parameter effects of the three-dimensional model based on the kinetic and thermodynamic models, and obtain the three-dimensional evaluation results. A multi-objective parameter generation module is used to construct a transferable conditional variational autoencoder and train the transferable conditional variational autoencoder model based on the process parameters and the three-dimensional evaluation results to obtain candidate parameters. A multi-objective parameter optimization module is used to construct a proxy model and perform multi-objective optimization on the candidate parameters through the proxy model to obtain the optimal parameters; The parameter verification and migration module is used to evaluate the parameters of the actual machine system based on the optimal parameters, and to debug the actual machine system based on the evaluation results.

Images