A mold forming process optimization method based on machine learning

Abstract

The present application relates to the technical field of process optimization, and discloses a mold forming process optimization method based on machine learning, which comprises the following steps: performing mold flow simulation on benchmark process parameters to obtain a multivariate simulation data set of a target mold; mapping a simulation feature vector counted from the multivariate simulation data set and an actual forming defect detected in a mold forming test process to a shared feature space to generate a hybrid feature representation vector of the target mold; constructing a process optimization process of the target mold according to the hybrid feature representation vector, defining a combination of process parameters to be optimized, and performing a constrained multi-objective Bayesian optimization search on the process optimization process to obtain optimized process parameters of the target mold. The present application can improve the efficiency of a process optimization based on machine learning.

Description

Technical Field

[0001] This invention relates to the field of process optimization technology, and in particular to a method for optimizing mold forming process based on machine learning. Background Technology

[0002] In the field of mold forming and processing, traditional process parameter optimization methods rely heavily on technicians' repeated manual interpretation of mold flow simulation results and on-site mold trial experience. Because the simulation environment and the actual molding process have inherent deviations in machine response characteristics and workshop environment fluctuations, the process window established solely based on virtual defect cloud maps often fails to meet expectations in actual mass production for shrinkage control and warpage suppression. This necessitates frequent offline iterative corrections, resulting in an excessively long cycle from parameter design to stable production, and the optimization effect is difficult to reliably reproduce across different production batches.

[0003] Existing data-driven process optimization methods often directly utilize simulation data to construct surrogate models, lacking effective correction mechanisms for the distribution differences between simulation features and measured defects. The resulting hybrid representations struggle to accurately map the unsteady-state impacts of process variables on actual molding quality. Furthermore, conventional multi-objective optimization, when searching for the optimal process combination, fails to fully incorporate key mechanistic constraints such as upper limits for injection speed and lower limits for solidification layer growth rates. This leads to the generation of numerous candidate solutions exceeding the process feasibility boundaries, reducing search efficiency and weakening the optimal balance between comprehensive quality defect suppression and unit energy consumption control achieved by the obtained parameters. Therefore, improving process optimization efficiency has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a machine learning-based method for optimizing mold forming processes to solve the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides a machine learning-based method for optimizing mold forming processes, comprising: Mold flow simulation was performed on the baseline process parameters to obtain a multivariate simulation data set for the target mold; The simulation feature vectors statistically derived from the multivariate simulation data set are mapped together with the actual molding defects detected during the mold forming test to a shared feature space to generate a hybrid feature representation vector for the target mold. The process optimization process of the target mold is constructed based on the hybrid feature representation vector, and the combination of process parameters to be optimized is defined. The process optimization process is subjected to constrained multi-objective Bayesian optimization search to obtain the optimized process parameters of the target mold.

[0006] In a preferred embodiment, the step of performing mold flow simulation on the baseline process parameters to obtain a multivariate simulation data set for the target mold includes: Obtain the three-dimensional geometric model of the target mold and the physical property parameters of the plastic part material; Based on the benchmark process parameters, the three-dimensional geometric model, and the physical property parameters, the target mold is simulated for the injection, holding, and cooling stages to obtain a virtual molding defect cloud map, and the simulation energy consumption and filling time are recorded simultaneously. The virtual molding defect cloud map, the simulation energy consumption, and the filling time together constitute a multi-dimensional simulation data set.

[0007] In a preferred embodiment, the virtual forming defect cloud map includes: The virtual molding defect cloud map includes the distribution of shrinkage area and the distribution of warpage deformation area. The shrinkage area distribution is generated by marking the areas in each grid node during the cooling stage where the volume shrinkage rate exceeds a threshold. The warpage deformation area distribution is generated by marking the areas in each grid node after demolding where the displacement vector exceeds a threshold.

[0008] In a preferred embodiment, the step of mapping the simulation feature vector statistically derived from the multivariate simulation data set, together with the actual molding defects detected during the mold forming test, to a shared feature space to generate a hybrid feature representation vector for the target mold includes: Defect statistical features are extracted from the virtual molding defect cloud map and combined with the simulation energy consumption and the filling time to form a simulation feature vector. The actual molding defect features of the target mold obtained by visual inspection after actual injection molding are collected to form a measured feature vector. The simulated feature vector and the measured feature vector are aligned in distribution to obtain a hybrid feature representation vector.

[0009] In a preferred embodiment, the step of combining the defect statistical features and the actual molding defect features includes: The maximum depth and area ratio of the shrinkage mark region, as well as the maximum displacement and area ratio of the warp deformation region, are extracted as defect statistical features. The actual molding defect features of the target mold are obtained by visual inspection after actual injection molding. The actual molding defect features include: the shrinkage depth deviation and warpage displacement deviation of the actual injection molded part.

[0010] In a preferred embodiment, the process optimization process for constructing the target mold based on the hybrid feature representation vector includes: A numerical mapping is established between the hybrid feature representation vector and the corresponding comprehensive evaluation value of molding defects and unit energy consumption evaluation value to generate a process response surface. Mechanistic boundary constraints, constructed from the upper limit of injection speed and the lower limit of solidification layer growth rate, are applied to the process response surface to form a constrained process response surface, which serves as a process optimization process.

[0011] In a preferred embodiment, the unit energy consumption evaluation value and the comprehensive evaluation value of molding defects include: The volume ratio of the shrinkage area and the surface deviation of the warping area in the virtual molding defect cloud map are statistically analyzed and used as the evaluation values ​​for shrinkage severity and warping deformation, respectively. The severity evaluation value of the shrinkage mark and the warpage deformation evaluation value are weighted and fused to generate a comprehensive evaluation value of the molding defect; The ratio of the simulated energy consumption to the filling time is used as the unit energy consumption evaluation value.

[0012] In a preferred embodiment, defining the combination of process parameters to be optimized and performing constrained multi-objective Bayesian optimization search on the process optimization process to obtain the optimized process parameters of the target mold includes: Injection speed, holding pressure, melt temperature, and mold temperature are determined as the combination of process parameters to be optimized. On the constrained process response surface, a multi-point sampling is performed using the desired hypervolume improvement strategy to generate multiple sets of candidate process parameter combinations. Each set of candidate process parameters is fed back to the model flow simulation for verification. The constrained process response surface is updated based on the verification results and approximated to the Pareto front to obtain the optimized process parameters.

[0013] In a preferred embodiment, the multi-point sampling using the desired hypervolume improvement strategy includes: Within the feasible region defined by the mechanistic boundary constraints, the acquisition function of the desired hypervolume improvement strategy consists of the expected hypervolume increment of the current Pareto front and the distance penalty term between candidate points. Based on this acquisition function, the multiple sets of candidate process parameter combinations are selected within the feasible region. The calculation formula for the acquisition function is as follows: ; In the formula, The output value of the acquisition function. For the candidate process parameter combinations The expected increase in hypervolume at the current Pareto frontier, The preset distance penalty coefficient, For the candidate process parameter combinations The candidate process parameter combinations already exist in the candidate process parameter combinations. The Euclidean distance between them The candidate process parameter combinations are those that already exist in the candidate process parameter combinations. It is a set consisting of the selected candidate process parameter combinations. The candidate process parameter combinations are defined within the feasible domain of the mechanism boundary constraints.

[0014] In a preferred embodiment, when the ambient temperature and humidity of the molding workshop are detected to exceed the preset fluctuation range, or when a raw material batch change signal is received, it is determined that the environmental state has deviated. Obtain the measured shrinkage depth deviation and measured warpage displacement deviation of the current batch of re-injection molded test pieces as online correction samples; The constrained process response surface is locally weighted to generate an updated constrained process response surface; The revised process parameters adapted to the environmental offset are searched from the updated constraint process response surface and used as recommended values ​​for the process parameters adapted to the new batch. These values ​​are then written into the parameter register of the molding machine controller.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention maps the features of mold flow simulation and the measured features of molding defects to a shared feature space and generates a hybrid feature representation vector. This enables the constructed process response surface to more accurately describe the response relationship between process parameters and actual molding defects and energy consumption, thereby improving the accuracy of subsequent optimization. By injecting mechanistic boundary constraints consisting of an upper limit for injection speed and a lower limit for solidification layer growth rate into constrained multi-objective Bayesian optimization, it ensures that each iteration sampling is within the process feasible region, thus improving the search hit rate and obtaining optimized process parameters that achieve a better balance between defect comprehensive suppression and unit energy consumption.

[0016] 2. Employing a multi-point sampling strategy based on expected supervolume improvement with added distance penalties, multiple sets of high-potential candidate process parameters can be generated simultaneously in a single optimization iteration. After simulation verification, these parameters rapidly approach the Pareto front, significantly shortening the cycle from parameter design to obtaining the optimal process combination. By sensing workshop environment shifts or raw material batch changes, the constrained process response surface is locally weighted online and re-searched to generate revised process parameters adapted to the new operating conditions. These revised parameters are then directly written into the molding machine controller, ensuring the optimized process window has continuous and effective batch adaptability, guaranteeing the stability of molding quality and the consistency of energy consumption control during mass production. Attached Figure Description

[0017] Figure 1 A schematic flowchart illustrating a machine learning-based method for optimizing mold forming process is provided in one embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a machine learning-based method for optimizing mold forming processes. The execution subject of this machine learning-based mold forming process optimization method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the machine learning-based mold forming process optimization method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a machine learning-based mold forming process optimization method according to an embodiment of the present invention. In this embodiment, the machine learning-based mold forming process optimization method includes: Mold flow simulation was performed on the baseline process parameters to obtain a multivariate simulation data set for the target mold; In this embodiment of the invention, the step of performing mold flow simulation on the benchmark process parameters to obtain a multivariate simulation data set for the target mold includes: Obtain the three-dimensional geometric model of the target mold and the physical property parameters of the plastic part material; Based on the benchmark process parameters, the three-dimensional geometric model, and the physical property parameters, the target mold is simulated for the injection, holding, and cooling stages to obtain a virtual molding defect cloud map, and the simulation energy consumption and filling time are recorded simultaneously. The virtual molding defect cloud map, the simulation energy consumption, and the filling time together constitute a multi-dimensional simulation data set.

[0021] The virtual molding defect cloud map includes: The virtual molding defect cloud map includes the distribution of shrinkage area and the distribution of warpage deformation area. The shrinkage area distribution is generated by marking the areas in each grid node during the cooling stage where the volume shrinkage rate exceeds a threshold. The warpage deformation area distribution is generated by marking the areas in each grid node after demolding where the displacement vector exceeds a threshold.

[0022] The process involves obtaining the 3D geometric model of the target mold and the physical properties of the plastic part material. The 3D geometric model is obtained by directly reading the mold's computer-aided design file, which records all spatial geometric information about the mold cavity, gating system, and cooling channels. The physical properties of the plastic part material are extracted from the material supplier's property data sheets. These extracts include data on melt density, specific heat capacity, thermal conductivity, shear viscosity as a function of shear rate, PVT characteristics, and curing temperature. These values ​​are then structured and stored in the form of material cards for subsequent simulation use.

[0023] Based on the baseline process parameters, the three-dimensional geometric model, and the physical property parameters, the target mold is simulated for the injection, holding, and cooling stages to obtain a virtual molding defect cloud map, and the simulation energy consumption and filling time are recorded simultaneously. In the injection stage simulation, the cavity space is discretized into mesh elements. For each mesh element, the mass conservation equation, momentum conservation equation, and energy conservation equation are solved. The shear viscosity data table and PVT characteristic data table from the physical property parameters are used to describe the rheological behavior and density change of the melt. Boundary conditions are set according to the injection velocity curve in the baseline process parameters. The simulation starts from the melt front entering the gate until the cavity volume is filled, and the pressure field, temperature field, and velocity field at each position of the melt front reaching the cavity are recorded. In the holding stage simulation, the pressure field and temperature field at the end of the injection stage are used as the initial state. Driven by the holding pressure curve in the baseline process parameters, melt continues to be added to the cavity to compensate for the volume shrinkage caused by cooling. During the simulation, the changes in pressure, temperature, and density are continuously calculated for each mesh element until the gate solidifies and seals. The simulation of the cooling stage begins with the temperature field at the end of the holding pressure stage. The location information of the cooling channels in the 3D geometric model is transformed into thermal boundary conditions. The 3D transient heat conduction equation is solved, and the temperature decrease of each grid node within the cavity over time is calculated until the temperature of all nodes is lower than the heat deformation temperature in the physical property parameters. After completing the simulation of the injection, holding pressure, and cooling stages, a virtual molding defect cloud map is output, and the total energy consumption from the start of injection to the end of cooling, as well as the filling time for the melt to fill the cavity, are recorded simultaneously.

[0024] The virtual molding defect cloud map includes the distribution of shrinkage dent regions and the distribution of warpage deformation regions. The generation process of the shrinkage dent region distribution is as follows: at the end of each time step of the cooling stage simulation, each grid node in the cavity thickness direction is traversed, and the difference between the current specific volume of the node and the specific volume of the node when the gate is frozen is extracted. This difference is divided by the specific volume when the gate is frozen to obtain the volume shrinkage rate of the node. The volume shrinkage rate is compared with a preset volume shrinkage rate threshold. When the volume shrinkage rate exceeds the threshold, the grid node is marked as a shrinkage dent contributing node. The indentation depth of all marked nodes in the cavity surface normal direction is obtained by accumulating the shrinkage strain of the node in the thickness direction. The spatial coordinates of each marked node and its corresponding indentation depth value are set together to form the shrinkage dent region distribution. The generation process of the warped deformation region distribution involves extracting the cavity mesh after demolding, using the residual thermal stress in each mesh node at the end of the cooling stage as the initial internal stress, releasing the displacement constraint of the cavity surface on the nodes, and causing the mesh nodes to displace under the action of internal stress. The displacement vector of each mesh node is calculated, and the magnitude of this displacement vector is equal to the composite modulus of the displacement components of the node in the three spatial coordinate directions. This displacement vector modulus is compared with a preset displacement threshold. When the displacement vector modulus exceeds the threshold, the mesh node is marked as a warped deformation node. The spatial coordinates of all marked nodes and their displacement vectors together constitute the warped deformation region distribution. The virtual molding defect cloud map, the simulation energy consumption, and the filling time together constitute a multivariate simulation data set. The shrinkage area distribution data, warped deformation region distribution data, simulation energy consumption values, and filling time values ​​are packaged into a unified multivariate simulation data set. This set uses the mesh node index as the association key to store the spatial distribution information of the defect cloud map in correspondence with the process quantity information of energy consumption and time.

[0025] The beneficial effects include: obtaining complete spatial geometric information of the cavity, gating system, and cooling channels by directly reading the mold computer-aided design file; extracting melt density, specific heat capacity, thermal conductivity, shear viscosity, PVT characteristic data, and curing temperature from the material property data sheet; and storing the information in a structured format as material cards, ensuring the integrity and consistency of the simulation input information and eliminating information transmission deviations. In the three-stage simulation of injection, holding, and cooling, the mass, momentum, and energy conservation equations are solved grid-by-grid. The ability of the shear viscosity and PVT characteristic data sheets to realistically describe melt rheology and density changes is combined with the driving control of injection speed and holding pressure curves, completely recording the pressure field, temperature field, velocity field, total energy consumption, and filling time. This gives the virtual molding defect cloud map a basis for reproducing the real molding physical process. In the generation of shrinkage region distribution, the volumetric shrinkage rate is obtained by comparing the difference between the current specific volume of each grid node and the specific volume at the time of gate freezing, divided by the specific volume at the time of gate freezing. The volumetric shrinkage rate threshold is used as a boundary to mark the shrinkage contributing nodes, and the cumulative thickness-direction shrinkage strain is used to obtain the depression depth, forming a precisely located shrinkage region distribution. In the generation of warping deformation region distribution, the displacement constraints of the grid nodes after demolding are released, allowing residual thermal stress to drive node displacement. The displacement vector magnitude is compared with the displacement threshold to mark the warping deformation nodes, forming a warping deformation region distribution. The simulation energy consumption values ​​and filling time values ​​of the shrinkage region distribution and warping deformation region distribution are packaged into a multivariate simulation dataset using the grid node index as the association key. This achieves a complete correspondence between the spatial distribution of defects and process quantity information, providing a comprehensive data foundation for subsequent feature extraction and hybrid mapping.

[0026] The simulation feature vectors statistically derived from the multivariate simulation data set are mapped together with the actual molding defects detected during the mold forming test to a shared feature space to generate the hybrid feature representation vector of the target mold.

[0027] In this embodiment of the invention, the step of mapping the simulation feature vector statistically derived from the multivariate simulation data set, together with the actual molding defects detected during the mold forming test, to a shared feature space to generate the hybrid feature representation vector of the target mold includes: Defect statistical features are extracted from the virtual molding defect cloud map and combined with the simulation energy consumption and the filling time to form a simulation feature vector. The actual molding defect features of the target mold obtained by visual inspection after actual injection molding are collected to form a measured feature vector. The simulated feature vector and the measured feature vector are aligned in distribution to obtain a hybrid feature representation vector.

[0028] The elements derived from the defect statistical features and the actual molding defect features include: The maximum depth and area ratio of the shrinkage mark region, as well as the maximum displacement and area ratio of the warp deformation region, are extracted as defect statistical features. The actual molding defect features of the target mold are obtained by visual inspection after actual injection molding. The actual molding defect features include: the shrinkage depth deviation and warpage displacement deviation of the actual injection molded part.

[0029] Defect statistical features are extracted from the virtual molding defect cloud map and combined with the simulation energy consumption and filling time to form a simulation feature vector. The maximum depth and area ratio of the shrinkage mark region, and the maximum displacement and deformation area ratio of the warped deformation region are extracted as defect statistical features. The maximum depth of the shrinkage mark region is extracted by traversing the indentation depth values ​​of all marked nodes in the shrinkage mark region distribution and taking the maximum indentation depth value as the maximum depth of the shrinkage mark region. The area ratio of the shrinkage mark region is extracted by counting the total number of marked nodes in the shrinkage mark region distribution and the total number of mesh nodes on the cavity surface, and the ratio obtained by dividing the total number of marked nodes by the total number of mesh nodes on the cavity surface is taken as the area ratio of the shrinkage mark region. The maximum displacement of the warped deformation region is extracted by traversing the displacement vector magnitude of all marked nodes in the warped deformation region distribution and taking the maximum displacement vector magnitude as the maximum displacement of the warped deformation region. The method for extracting the deformation area ratio is to count the total number of marked nodes in the warped deformation region distribution and the total number of mesh nodes on the cavity surface. The ratio obtained by dividing the total number of marked nodes by the total number of mesh nodes on the cavity surface is taken as the deformation area ratio of the warped deformation region. The four values ​​of the maximum depth of the shrinkage area, the area ratio of the shrinkage area, the maximum displacement of the warped deformation region, and the deformation area ratio of the warped deformation region are taken as defect statistical features. The defect statistical features are concatenated with the simulation energy consumption value and the filling time value. The concatenation order is the maximum depth of the shrinkage area, the area ratio of the shrinkage area, the maximum displacement of the warped deformation region, the area ratio of the warped deformation region, the simulation energy consumption value, and the filling time value. After concatenation, a one-dimensional ordered arrangement containing six values ​​is obtained as the simulation feature vector.

[0030] The actual molding defect features obtained by visual inspection after the target mold is measured during injection molding are collected to form a measured feature vector. The measured injection molded part is obtained by performing actual injection molding production using the aforementioned benchmark process parameters on the same target mold. After a complete molding cycle is completed and the part is demolded, it is removed from the injection molding machine and placed at a visual inspection station after cooling to room temperature. The visual inspection station consists of an industrial camera and a structured light projection device. The industrial camera captures images of the part surface from multiple preset angles, and the structured light projection device projects a stripe grating onto the part surface. The industrial camera simultaneously acquires the stripe image modulated by the part surface morphology. The acquired stripe image undergoes phase unwrapping processing to obtain the three-dimensional point cloud data of the part surface. The three-dimensional point cloud data is registered and compared with the nominal shape in the part's computer-aided design file. During the comparison, the nominal shape of each measurement point on the part surface is extracted along the normal direction to the nominal shape. The distance to the nominal shape is calculated by subtracting the normal distance of the measurement point from the normal distance of the corresponding position of the nominal shape in the shrinkage-prone area to obtain the shrinkage depth deviation value of the measurement point. The maximum value of the shrinkage depth deviation values ​​of all measurement points in the shrinkage-prone area is taken as the shrinkage depth deviation of the actual injection molded part. In the warp-prone area, the difference between the normal distance of the measurement point and the normal distance of the corresponding position of the nominal shape is obtained to obtain the warp displacement deviation value of the measurement point. The maximum value of the warp displacement deviation values ​​of all measurement points in the warp-prone area is taken as the warp displacement deviation of the actual injection molded part. The shrinkage depth deviation value and the warp displacement deviation value of the actual injection molded part are concatenated into a one-dimensional ordered arrangement containing two values ​​as the measured feature vector.

[0031] The simulated feature vectors and the measured feature vectors are aligned in distribution to obtain a hybrid feature representation vector. The distribution alignment process involves multiple simulated feature vectors and multiple measured feature vectors obtained from multiple simulations and measurements under the same baseline process parameters. The entire set of simulated feature vectors is considered as the source domain dataset, and the entire set of measured feature vectors is considered as the target domain dataset. The goal of distribution alignment is to adjust the numerical distribution of the source domain dataset to be consistent with the numerical distribution of the target domain dataset. The specific implementation of distribution alignment is as follows: First, calculate the mean and standard deviation for each dimension of each simulated feature vector in the source domain dataset. Subtract the mean of each dimension from the original value of each simulated feature vector and divide by the standard deviation to obtain the standardized dataset of the source domain dataset. Similarly, calculate the mean and standard deviation for each dimension of each measured feature vector in the target domain dataset and standardize it in the same way to obtain the standardized dataset of the target domain dataset. Then, associate the standardized datasets of the source and target domains. The association method is to use canonical correlation analysis to find the projection direction with the highest correlation between each dimension of the standardized dataset of the source and the dimensions of the standardized dataset of the target domain. In the projection direction, the covariance structure of the standardized dataset of the target domain is used to correct the covariance structure of the standardized dataset of the source domain. The correction process is as follows: first, the covariance structure of the source domain dataset is adjusted by the covariance structure of the standardized dataset of the target domain. The covariance matrices of the standardized dataset and the target domain standardized dataset are calculated respectively. Eigenvalue decomposition is performed on the covariance matrix of the source domain standardized dataset to obtain eigenvectors and eigenvalues. The source domain standardized dataset is transformed into a space composed of eigenvectors. In this space, the distribution variance of the source domain standardized dataset is scaled to be consistent with that of the target domain standardized dataset. The scaling method is to replace the variance of each dimension of the source domain standardized dataset with the variance of the corresponding dimension of the target domain standardized dataset. Then, the scaled data is transformed back to the original dimension space through inverse transformation to obtain an aligned set of simulated feature vectors. The aligned simulated feature vector that is closest to the measured feature vector is selected from the aligned set of simulated feature vectors and concatenated with the measured feature vector. During concatenation, all dimensions of the two vectors are sequentially arranged and merged into a vector with a dimension equal to the sum of the dimensions of the two vectors. This vector is the hybrid feature representation vector.

[0032] The beneficial effects are that four statistical features of defects—the maximum depth and area ratio of the shrinkage area, the maximum displacement and deformation area ratio of the warp deformation area—are extracted from the virtual molding defect cloud map and combined with simulation energy consumption and filling time values ​​in a fixed order to form a six-dimensional simulation feature vector. This integrates the spatial distribution information of defects and process quantity information into a unified numerical sequence, enabling a comprehensive and structured representation of the molding state on the simulation side. Multi-angle image acquisition and stripe grating projection are performed on the measured injection-molded parts using an industrial camera and structured light projection device. Phase unwrapping processing is performed on the modulated stripe images to obtain three-dimensional point cloud data of the part surface. This data is then compared with the nominal shape and subtracted along the normal direction to extract the maximum shrinkage depth deviation of the shrinkage-prone area and the maximum warp displacement deviation of the warp-prone area, forming a measured feature vector. This makes the detection results of actual molding defects accurate and repeatable. Multiple simulation feature vectors and multiple measured feature vectors obtained from multiple simulations and measurements under the same baseline process parameters are used as the source domain dataset and the target domain dataset, respectively. First, the two datasets are standardized dimension by dimension to eliminate the difference in units. Then, canonical correlation analysis is used to find the projection direction with the greatest correlation between the source domain and the target domain. Eigenvalue decomposition is performed on the covariance matrix of the standardized source domain dataset. In the space formed by the feature vectors, the variance of each dimension of the source domain is replaced with the variance of the corresponding dimension of the target domain to complete the covariance structure correction. After inverse transformation, the aligned set of simulation feature vectors is obtained. The one that is closest to the measured feature vector is selected from the aligned set of simulation feature vectors and concatenated with the measured feature vector to generate a hybrid feature representation vector. This makes the numerical distribution of the simulation features and the measured features more consistent, eliminating the inherent distribution offset between the simulation environment and the real molding environment, and providing input information that is closer to the real molding state for the subsequent construction of the process response surface.

[0033] The process optimization process of the target mold is constructed based on the hybrid feature representation vector, and the combination of process parameters to be optimized is defined. The process optimization process is subjected to constrained multi-objective Bayesian optimization search to obtain the optimized process parameters of the target mold.

[0034] In this embodiment of the invention, the process optimization process for constructing the target mold based on the hybrid feature representation vector includes: A numerical mapping is established between the hybrid feature representation vector and the corresponding comprehensive evaluation value of molding defects and unit energy consumption evaluation value to generate a process response surface. Mechanistic boundary constraints, constructed from the upper limit of injection speed and the lower limit of solidification layer growth rate, are applied to the process response surface to form a constrained process response surface, which serves as a process optimization process.

[0035] The unit energy consumption evaluation value and the comprehensive evaluation value of molding defects include: The volume ratio of the shrinkage area and the surface deviation of the warping area in the virtual molding defect cloud map are statistically analyzed and used as the evaluation values ​​for shrinkage severity and warping deformation, respectively. The severity evaluation value of the shrinkage mark and the warpage deformation evaluation value are weighted and fused to generate a comprehensive evaluation value of the molding defect; The ratio of the simulated energy consumption to the filling time is used as the unit energy consumption evaluation value.

[0036] The process parameters to be optimized are defined, and a constrained multi-objective Bayesian optimization search is performed on the process optimization process to obtain the optimized process parameters of the target mold, including: Injection speed, holding pressure, melt temperature, and mold temperature are determined as the combination of process parameters to be optimized. On the constrained process response surface, a multi-point sampling is performed using the desired hypervolume improvement strategy to generate multiple sets of candidate process parameter combinations. Each set of candidate process parameters is fed back to the model flow simulation for verification. The constrained process response surface is updated based on the verification results and approximated to the Pareto front to obtain the optimized process parameters.

[0037] The multi-point sampling using the desired hypervolume improvement strategy includes: Within the feasible region defined by the mechanistic boundary constraints, the acquisition function of the desired hypervolume improvement strategy consists of the expected hypervolume increment of the current Pareto front and the distance penalty term between candidate points. Based on this acquisition function, the multiple sets of candidate process parameter combinations are selected within the feasible region. The calculation formula for the acquisition function is as follows: ; In the formula, The output value of the acquisition function. For the candidate process parameter combinations The expected increase in hypervolume at the current Pareto frontier, The preset distance penalty coefficient, For the candidate process parameter combinations The candidate process parameter combinations already exist in the candidate process parameter combinations. The Euclidean distance between them The candidate process parameter combinations are those that already exist in the candidate process parameter combinations. It is a set consisting of the selected candidate process parameter combinations. The candidate process parameter combinations are defined within the feasible domain of the mechanism boundary constraints.

[0038] After obtaining the optimized process parameters of the target mold, the process further includes: When the temperature and humidity of the molding workshop environment exceed the preset fluctuation range, or when a raw material batch change signal is received, it is determined that the environmental state has deviated. Obtain the measured shrinkage depth deviation and measured warpage displacement deviation of the current batch of re-injection molded test pieces as online correction samples; The constrained process response surface is locally weighted to generate an updated constrained process response surface; The revised process parameters adapted to the environmental offset are searched from the updated constraint process response surface and used as recommended values ​​for the process parameters adapted to the new batch. These values ​​are then written into the parameter register of the molding machine controller.

[0039] A numerical mapping is established between the hybrid feature representation vectors and the corresponding comprehensive evaluation values ​​of molding defects and unit energy consumption evaluation values ​​to generate a process response surface. Before generating the process response surface, it is necessary to obtain the comprehensive evaluation value of molding defects and unit energy consumption evaluation value corresponding to each set of hybrid feature representation vectors. The comprehensive evaluation value of molding defects is obtained by statistically analyzing the volume ratio of the shrinkage area in the virtual molding defect cloud map as the shrinkage severity evaluation value. The statistical method for the volume ratio of the shrinkage area is to divide the sum of the volumes of all marked nodes in the shrinkage area distribution by the sum of the volumes of all mesh nodes in the cavity. The surface shape deviation of the warped area in the virtual molding defect cloud map is statistically analyzed as the warping deformation degree evaluation value. The statistical method for the surface deviation of the warped area is to extract the displacement component of each marked node along the normal direction of the cavity surface after demolding in the distribution of the warped deformation area, and to calculate the root mean square value of the normal displacement component of all marked nodes. The weighted fusion of the shrinkage severity evaluation value and the warped deformation degree evaluation value generates a comprehensive evaluation value of molding defects. The weighted fusion method is to multiply the shrinkage severity evaluation value by a preset shrinkage weight coefficient and add the warped deformation degree evaluation value by a preset warping weight coefficient, and then sum the two to obtain the comprehensive evaluation value of molding defects. The unit energy consumption evaluation value is obtained by dividing the simulated energy consumption value by the filling time value. After obtaining multiple sets of hybrid feature representation vectors and their corresponding comprehensive evaluation values ​​of molding defects and unit energy consumption under the same baseline process parameters, a process response surface is established through numerical mapping. The numerical mapping method involves associating six values ​​from each set of hybrid feature representation vectors with the combination of process parameters to be optimized, consisting of injection speed, holding pressure, melt temperature, and mold temperature. The mapping relationship is constructed by using each dimension of the hybrid feature representation vector as the input and the comprehensive evaluation value of molding defects and the unit energy consumption evaluation value as the output. This mapping relationship is built by performing multidimensional interpolation on all existing input-output data pairs. During the multidimensional interpolation process, for any position of the process parameter combination to be optimized that does not appear in the existing data, several existing data points spatially closest to that position are found. The output values ​​of these existing data points are weighted and averaged using the reciprocal of the distance as the weight to obtain the estimated comprehensive evaluation value of molding defects and the estimated unit energy consumption evaluation value for that position. This method is used to generate output value estimates point-by-point across the entire range of values ​​for the process parameter combinations to be optimized, forming a continuous process response surface defined on the space of the process parameter combinations to be optimized.

[0040] A mechanistic boundary constraint, constructed from an upper limit for injection speed and a lower limit for solidification layer growth rate, is applied to the process response surface to form a constrained process response surface, which serves as the process optimization process. The upper limit for injection speed is a preset maximum injection speed, determined by the melt shear rate limit that the mold gating system can withstand. Regions in the process response surface whose corresponding injection speed values ​​exceed this preset maximum are marked as infeasible and removed from the process response surface. The lower limit for solidification layer growth rate is a preset minimum solidification layer thickening rate. The time-varying curve of the melt solidification layer thickness at the cavity wall is extracted from the mold flow simulation data during the cooling stage. The average slope of this curve from cavity filling to gate freezing is calculated as the solidification layer growth rate. Regions in the process response surface whose corresponding solidification layer growth rate values ​​are lower than this preset minimum solidification layer thickening rate are marked as infeasible and removed from the process response surface. The remaining process response surface region after removing the upper limit for injection speed and the lower limit for solidification layer growth rate is the constrained process response surface, which serves as the process optimization process for subsequent optimization searches.

[0041] Injection speed, holding pressure, melt temperature, and mold temperature are defined as the combination of process parameters to be optimized. The adjustment range of injection speed is determined by the screw feed speed specification of the injection molding machine; the adjustment range of holding pressure is jointly determined by the output capacity of the injection molding machine's hydraulic system and the cavity's pressure-bearing capacity; the adjustment range of melt temperature is determined by the recommended processing temperature range in the material properties of the plastic part; and the adjustment range of mold temperature is determined by the heat exchange capacity of the mold's cooling circuit. Lower and upper limits are set for the adjustment ranges of these four types of process parameters, forming the search boundary for the combination of process parameters to be optimized. The combination of process parameters to be optimized can take any value within the search boundary, forming a four-dimensional search space.

[0042] On the constrained process response surface, a multi-point sampling strategy is used to generate multiple sets of candidate process parameter combinations. The execution space of the multi-point sampling is the feasible region corresponding to the constrained process response surface after removing infeasible regions through mechanistic boundary constraints. The sampling function of the expected hypervolume improvement strategy consists of two parts: the expected hypervolume increment of the current Pareto front and the distance penalty term between candidate points. The Pareto front is determined by finding all process parameter combinations on the constrained process response surface that are not superior to each other in terms of both the comprehensive evaluation value of molding defects and the evaluation value of unit energy consumption. That is, there is no other point that is superior to this point in both objectives. These points that are not superior to each other constitute the current Pareto front. The expected hypervolume increment is obtained by calculating the expected expansion of the target space region enclosed by the Pareto front after adding any candidate process parameter combination point in the feasible region to the current Pareto front. The distance penalty term between candidate points is obtained by... Calculate the Euclidean distance between the candidate process parameter combination point and all points in the existing candidate process parameter combination set, and extract the minimum distance value. Multiply the minimum distance value by a preset distance penalty coefficient to obtain the distance penalty amount. The value of the acquisition function is equal to the expected hypervolume increment minus the distance penalty amount. The larger the value of the acquisition function, the higher the comprehensive value of the candidate point in improving the diversity of the solution set and expanding the coverage of the Pareto front. Search within the feasible region with the goal of maximizing the acquisition function value. Each time, select a point that maximizes the acquisition function as a new candidate process parameter combination. After adding the new candidate point to the existing candidate process parameter combination set, recalculate the acquisition function values ​​of the remaining points in the feasible region and continue to select the next candidate point. Repeat this selection process until a preset number of candidate process parameter combinations are generated.

[0043] Each set of candidate process parameters is fed back to the model flow simulation for verification. The constrained process response surface is updated based on the verification results and approximated to the Pareto front to obtain the optimized process parameters. Each generated candidate process parameter combination is sequentially substituted into the mold flow simulation process. Using the injection speed, holding pressure, melt temperature, and mold temperature from this candidate process parameter combination as new process conditions, the simulation of the injection, holding, and cooling stages is re-executed. The new virtual molding defect cloud map, new simulation energy consumption, and new filling time corresponding to this candidate process parameter combination are output. From the new virtual molding defect cloud map, the maximum depth and area ratio of the new shrinkage region, as well as the maximum displacement and deformation area ratio of the new warpage deformation region, are extracted and combined with the new simulation energy consumption and new filling time to form a new simulation feature vector. This vector is then aligned with the measured feature vector to obtain a new hybrid feature representation vector. This new hybrid feature representation vector is substituted into the mapping relationship of the process response surface to obtain the new comprehensive evaluation value of molding defects and the new unit energy consumption evaluation value corresponding to this candidate process parameter combination. This candidate process parameter combination, along with its new comprehensive evaluation value of molding defects and the new unit energy consumption evaluation value, are added as new data points to the input-output mapping dataset of the constrained process response surface. The constrained process response surface is then updated. The update method is as follows: To incorporate the new data points into the multidimensional interpolation data base, the output value estimate for the entire space is regenerated point by point. The updated constrained process response surface reflects a more accurate mapping relationship between process parameters and molding quality and energy consumption. The Pareto front is re-determined on the updated constrained process response surface. The coverage position of the Pareto front in the target space before and after the update is compared. If the updated Pareto front has shifted in the direction of lower comprehensive evaluation value of molding defects and lower unit energy consumption evaluation value compared to the previous update, it indicates that the Pareto front has been approximated and improved. The iterative process of using the expected hypervolume improvement strategy to perform multi-point sampling to generate candidate process parameter combinations and feeding them back to the mold flow simulation to verify the updated constrained process response surface is repeated until the Pareto front no longer shifts. At this point, all process parameter combination points on the Pareto front are the optimized process parameter set that achieves the best trade-off between the comprehensive evaluation value of molding defects and the unit energy consumption evaluation value for the target mold. The set with the lowest comprehensive evaluation value of molding defects or the set with the lowest unit energy consumption evaluation value is selected from this set as the optimized process parameters for the target mold.

[0044] When the ambient temperature and humidity in the molding workshop exceed the preset fluctuation range, or when a raw material batch change signal is received, an environmental state deviation is determined. The ambient temperature and humidity in the molding workshop are detected by installing temperature and humidity sensors. These sensors collect ambient temperature and humidity values ​​at fixed time intervals. The preset fluctuation range consists of an upper temperature limit, a lower temperature limit, an upper humidity limit, and a lower humidity limit. An environmental state deviation determination is triggered when the collected ambient temperature value exceeds the upper temperature limit or falls below the lower temperature limit, or when the ambient humidity value exceeds the upper humidity limit or falls below the lower humidity limit. The raw material batch change signal is received by establishing a data communication connection with the raw material management database. When the batch number field of the plastic part material used in the target mold in the raw material management database is updated, the raw material management database sends a raw material batch change signal to this process. Upon receiving this signal, an environmental state deviation is determined.

[0045] The measured shrinkage depth deviation and measured warpage displacement deviation of the current batch of re-injection molded test pieces are obtained as online correction samples. After determining that the environmental conditions have deviated, the test pieces are re-injected on the molding machine using the current combination of process parameters to be optimized. After the test pieces cool to room temperature, they are placed at the vision inspection station. The surface three-dimensional point cloud data of the test pieces are collected by an industrial camera and a structured light projection device and compared with the nominal shape registration. The maximum shrinkage depth deviation of the shrinkage-prone area and the maximum warpage displacement deviation of the warpage-prone area are extracted along the normal direction. The measured shrinkage depth deviation and measured warpage displacement deviation of the current batch of re-injection molded test pieces are used as online correction samples.

[0046] The constrained process response surface is locally weighted to generate an updated constrained process response surface. The scope of the local weight adjustment is limited to the region on the constrained process response surface adjacent to the position of the current combination of process parameters to be optimized. The adjacent region is defined by taking the coordinate point corresponding to the combination of process parameters to be optimized in the four-dimensional search space as the center and delineating a four-dimensional neighborhood. The weights used by all existing data points in this four-dimensional neighborhood when participating in multidimensional interpolation are adjusted. The adjustment method is to increase the weight value of the data point corresponding to the online correction sample while decreasing the weight value of the original data points in the neighborhood. The magnitude of the increase and decrease in weight is determined by the Euclidean distance between the data point and the corresponding data point of the online correction sample in the search space. The closer the distance, the larger the weight adjustment magnitude. After completing the local weight adjustment, the output value estimate of the constrained process response surface in this four-dimensional neighborhood is recalculated to obtain the updated constrained process response surface. The mapping relationship of the updated constrained process response surface in the neighborhood is closer to the actual forming state reflected by the online correction sample. The original mapping relationship is maintained in the region outside the neighborhood.

[0047] Revised process parameters adapted to the environmental offset are searched on the updated constrained process response surface and used as recommended values ​​for the new batch. These values ​​are then written into the parameter register of the molding machine controller. The Pareto front is redefined on the updated constrained process response surface. From the updated Pareto front, the set of process parameters with the lowest comprehensive evaluation value for molding defects is selected as the revised process parameters. The injection speed, holding pressure, melt temperature, and mold temperature values ​​in these revised process parameters are encoded according to the data format specified in the molding machine controller's parameter register. These encoded values ​​are then written into the parameter register of the molding machine controller via the data communication interface. The parameter register is a read-write storage area within the molding machine controller used to store the process parameter settings for the current molding cycle. After writing, the molding machine will execute injection, holding pressure, and cooling actions according to these revised process parameters in the next molding cycle, completing the adaptation of the process parameters for the new batch.

[0048] The expected hypervolume increment of the current Pareto front in the acquisition function is derived from the Pareto front set consisting of the points that are not dominant among all evaluated process parameter combinations on the constrained process response surface. The acquisition method is as follows: first, determine the hypervolume enclosed by the current Pareto front in the target space consisting of the comprehensive evaluation value of forming defects and the unit energy consumption evaluation value. Then, for any candidate process parameter combination in the feasible region, add the candidate point to the current Pareto front and recalculate the hypervolume enclosed by the Pareto front. Calculate the difference in hypervolume before and after the addition as the single hypervolume increment of the candidate point. Take the expected hypervolume increment of the candidate point as the distribution of possible output values ​​in the target space. This gives the expected hypervolume increment of the candidate point. The distance penalty term between candidate points in the acquisition function is derived from the set of selected candidate process parameter combinations. The method for obtaining the distance penalty term is as follows: for any candidate process parameter combination within the feasible region, calculate the Euclidean distance between the candidate point and each existing candidate point in the set of candidate process parameter combinations in four dimensions: injection speed, holding pressure, melt temperature, and mold temperature. The Euclidean distance is calculated by squaring the coordinate differences between the two points in the four dimensions, summing them, and then taking the square root. The minimum value among all Euclidean distances is taken as the minimum distance from the candidate point to the set of existing candidate points. Multiply the minimum distance by the preset distance penalty coefficient to obtain the distance penalty amount of the candidate point.

[0049] The acquisition function's formula signifies that when selecting candidate process parameter combinations within the feasible region, it simultaneously considers the contribution to the expansion of the Pareto front and the spatial dispersion between candidate points. The expected hypervolume increment measures the expansion of the target spatial coverage area brought about by adding a candidate point to the Pareto front; a larger value indicates a greater potential contribution from the candidate point to pushing the Pareto front towards a smaller overall defect evaluation value and a smaller unit energy consumption evaluation value. The distance penalty measures the proximity between a candidate point and selected candidate points in the search space; a larger value indicates a closer proximity between the candidate point and selected candidate points in the search space. The acquisition function's output is the expected hypervolume increment minus the distance penalty. A high expected hypervolume increment and sufficient distance between the candidate point and existing candidate points in the search space result in a high output value. Conversely, a high expected hypervolume increment but excessive proximity to existing candidate points in the search space leads to an increased distance penalty and a decreased output value. Similarly, a low expected hypervolume increment and a large distance between the candidate point and existing candidate points result in a decreased output value.

[0050] The trend of the acquisition function is that in the unexplored region of the feasible region, the output value of the acquisition function is mainly dominated by the expected hypervolume increment. As the region is gradually covered by candidate points, the expected hypervolume increment decreases due to the stabilization of the Pareto front. At this time, the distance penalty term causes the output value of the acquisition function to decrease further in the region where there are already dense candidate points. The distribution trend of the acquisition function in the feasible region is that a high value region is formed in the region where the Pareto front has great expansion potential and the candidate points have not yet been densely covered. This guides the selection of candidate process parameter combinations to shift from the fully explored region to the underexplored region. As the number of iterations increases, the Pareto front approaches the true optimal trade-off boundary and the expected hypervolume increment decreases overall. The high value region of the acquisition function gradually concentrates in the sparse segment of the Pareto front, so that the newly selected candidate points fill the gaps on the Pareto front and improve the uniformity and integrity of the Pareto front.

[0051] The beneficial effect is that by establishing a numerical mapping between the hybrid feature representation vector and the comprehensive evaluation value of molding defects and the unit energy consumption evaluation value, and generating output value estimates point by point in the entire space of process parameter combinations to be optimized using multidimensional interpolation, a process response surface is constructed, which allows for a continuous expression of the mapping relationship between process parameters and molding quality and energy consumption. Applying mechanistic boundary constraints consisting of an upper limit for injection speed and a lower limit for solidification layer growth rate to the process response surface eliminates infeasible regions, forming a constrained process response surface. This ensures that the optimization search is always performed within the process's achievable range, eliminating the generation of invalid candidate solutions. On the constrained process response surface, a desired hypervolume improvement strategy is used for multi-point sampling. The expected hypervolume increment is used to measure the Pareto front expansion potential, and a distance penalty term ensures the dispersion of candidate points in the search space. Multiple sets of candidate process parameter combinations are selected and fed back to the mold flow simulation for verification. New data points are added to the constrained process response surface, and the Pareto front is updated successively until the Pareto front no longer moves. This yields an optimized set of process parameters that achieves the optimal trade-off between the comprehensive evaluation value of molding defects and the unit energy consumption evaluation value, ensuring that the optimization results take into account both defect suppression and energy consumption control. When the ambient temperature and humidity exceed the preset fluctuation range or the raw material batch changes, the environmental state deviation is determined, online correction samples are obtained, and the local weights of the constraint process response surface are adjusted to make the mapping relationship in the neighborhood range closer to the actual molding state. An updated constraint process response surface is generated, and the revised process parameters are selected from the updated Pareto front and written into the molding machine controller parameter register to achieve rapid adaptation to the new batch.

[0052] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0053] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A machine learning based mold forming process optimization method, characterized by, The method includes: Mold flow simulation was performed on the baseline process parameters to obtain a multivariate simulation data set for the target mold; The simulation feature vectors statistically derived from the multivariate simulation data set are mapped together with the actual molding defects detected during the mold forming test to a shared feature space to generate a hybrid feature representation vector for the target mold. The process optimization process of the target mold is constructed based on the hybrid feature representation vector, and the combination of process parameters to be optimized is defined. The process optimization process is subjected to constrained multi-objective Bayesian optimization search to obtain the optimized process parameters of the target mold.

2. The machine learning based mold forming process optimization method of claim 1, wherein, The process of performing mold flow simulation on the baseline process parameters yields a multivariate simulation data set for the target mold, including: Obtain the three-dimensional geometric model of the target mold and the physical property parameters of the plastic part material; Based on the benchmark process parameters, the three-dimensional geometric model, and the physical property parameters, the target mold is simulated for the injection, holding, and cooling stages to obtain a virtual molding defect cloud map, and the simulation energy consumption and filling time are recorded simultaneously. The virtual molding defect cloud map, the simulation energy consumption, and the filling time together constitute a multi-dimensional simulation data set.

3. The method for optimizing mold forming process based on machine learning as described in claim 2, characterized in that, The virtual molding defect cloud map includes: The virtual molding defect cloud map includes the distribution of shrinkage area and the distribution of warpage deformation area. The shrinkage area distribution is generated by marking the areas in each grid node during the cooling stage where the volume shrinkage rate exceeds a threshold. The warpage deformation area distribution is generated by marking the areas in each grid node after demolding where the displacement vector exceeds a threshold.

4. The method for optimizing mold forming process based on machine learning as described in claim 1, characterized in that, The step of mapping the simulation feature vector statistically derived from the multivariate simulation data set, together with the actual molding defects detected during the mold forming test, to a shared feature space to generate the hybrid feature representation vector of the target mold includes: Defect statistical features are extracted from the virtual molding defect cloud map and combined with the simulation energy consumption and the filling time to form a simulation feature vector. The actual molding defect features of the target mold obtained by visual inspection after actual injection molding are collected to form a measured feature vector. The simulated feature vector and the measured feature vector are aligned in distribution to obtain a hybrid feature representation vector.

5. The method for optimizing mold forming process based on machine learning as described in claim 4, characterized in that, The elements derived from the defect statistical features and the actual molding defect features include: The maximum depth and area ratio of the shrinkage mark region, as well as the maximum displacement and area ratio of the warp deformation region, are extracted as defect statistical features. The actual molding defect features of the target mold are obtained by visual inspection after actual injection molding. The actual molding defect features include: the shrinkage depth deviation and warpage displacement deviation of the actual injection molded part.

6. The method for optimizing mold forming process based on machine learning as described in claim 1, characterized in that, The process optimization process for constructing the target mold based on the hybrid feature representation vector includes: A numerical mapping is established between the hybrid feature representation vector and the corresponding comprehensive evaluation value of molding defects and unit energy consumption evaluation value to generate a process response surface. Mechanistic boundary constraints, constructed from the upper limit of injection speed and the lower limit of solidification layer growth rate, are applied to the process response surface to form a constrained process response surface, which serves as a process optimization process.

7. The method for optimizing mold forming process based on machine learning as described in claim 6, characterized in that, The unit energy consumption evaluation value and the comprehensive evaluation value of molding defects include: The volume ratio of the shrinkage area and the surface deviation of the warping area in the virtual molding defect cloud map are statistically analyzed and used as the evaluation values ​​for shrinkage severity and warping deformation, respectively. The severity evaluation value of the shrinkage mark and the warpage deformation evaluation value are weighted and fused to generate a comprehensive evaluation value of the molding defect; The ratio of the simulated energy consumption to the filling time is used as the unit energy consumption evaluation value.

8. The method for optimizing mold forming process based on machine learning as described in claim 6, characterized in that, The process parameters to be optimized are defined, and a constrained multi-objective Bayesian optimization search is performed on the process optimization process to obtain the optimized process parameters of the target mold, including: Injection speed, holding pressure, melt temperature, and mold temperature are determined as the combination of process parameters to be optimized. On the constrained process response surface, a multi-point sampling is performed using the desired hypervolume improvement strategy to generate multiple sets of candidate process parameter combinations. Each set of candidate process parameters is fed back to the model flow simulation for verification. The constrained process response surface is updated based on the verification results and approximated to the Pareto front to obtain the optimized process parameters.

9. The method for optimizing mold forming process based on machine learning as described in claim 8, characterized in that, The multi-point sampling using the desired hypervolume improvement strategy includes: Within the feasible region defined by the mechanistic boundary constraints, the acquisition function of the desired hypervolume improvement strategy consists of the expected hypervolume increment of the current Pareto front and the distance penalty term between candidate points. Based on this acquisition function, the multiple sets of candidate process parameter combinations are selected within the feasible region. The calculation formula for the acquisition function is as follows: ； In the formula, The output value of the acquisition function. For the candidate process parameter combinations The expected increase in hypervolume at the current Pareto frontier, The preset distance penalty coefficient, For the candidate process parameter combinations The candidate process parameter combinations already exist in the candidate process parameter combinations. The Euclidean distance between them The candidate process parameter combinations are those that already exist in the candidate process parameter combinations. It is a set consisting of the selected candidate process parameter combinations. The candidate process parameter combinations are defined within the feasible domain of the mechanism boundary constraints.

10. The method for optimizing mold forming process based on machine learning as described in claim 6, characterized in that, After obtaining the optimized process parameters of the target mold, the process further includes: When the temperature and humidity of the molding workshop environment exceed the preset fluctuation range, or when a raw material batch change signal is received, it is determined that the environmental state has deviated. Obtain the measured shrinkage depth deviation and measured warpage displacement deviation of the current batch of re-injection molded test pieces as online correction samples; The constrained process response surface is locally weighted to generate an updated constrained process response surface; The revised process parameters adapted to the environmental offset are searched from the updated constraint process response surface and used as recommended values ​​for the process parameters adapted to the new batch. These values ​​are then written into the parameter register of the molding machine controller.

Images