Mould foaming control system based on image import and model simulation

Abstract

The present application provides a mold foaming control system based on image import and model simulation, which relates to the field of mold forming technology, including: an acquisition module, a processing module, a conversion module and a control module, which can construct a target three-dimensional model based on physical image data, and generate control instructions after standardized processing. The conversion module also calls the offset prediction model trained by the knowledge distillation mechanism, integrates the structural coding, thermal field data, assembly parameters and material parameters, generates control instruction correction items, and realizes online compensation of the control path. The system of the present application can improve the control accuracy and molding consistency of the mold under complex thermal conditions and assembly error conditions, and is suitable for mold foaming application scenarios with high requirements for structural contour control accuracy.

Description

Technical Field

[0001] The present application relates to the field of mold forming technology, and in particular to a mold foaming control system based on image import and model simulation. Background Art

[0002] Mold foaming technology has been widely used in the molding of lightweight structural parts, large composite parts and complex curved products, especially in the automotive, home appliance, electronic products and other industries, where the requirements for structural consistency and dimensional accuracy are constantly increasing. In such applications, the structural matching accuracy of the mold cavity directly affects the surface quality, functional stability and assembly adaptability of the molded product. Existing mold control systems are mostly based on static geometric modeling and rule-driven open-loop control strategies. Their control paths are usually generated offline by standard models and lack the ability to respond to dynamic disturbance factors in the actual processing process. For example, in the hot foam molding process, the thermal expansion of the mold under heating, the stress accumulation caused by material non-uniformity, and the offset effect introduced by assembly errors may cause deviations between the mold space contour and the target model, affecting the molding accuracy of the product.

[0003] Some technical solutions attempt to incorporate methods such as 3D image reconstruction, computer-aided design (CAD) comparison, and point cloud sampling to assess and compensate for mold geometric errors. However, these approaches still primarily rely on surface-level geometric information and struggle to capture the underlying structural responses of multiple physical fields. Furthermore, existing data-driven models face challenges in processing heterogeneous inputs from multiple sources, such as structural connectivity information, thermal field data, assembly status, and material properties, resulting in complex modeling, insufficient prediction accuracy, and limited real-time performance.

[0004] Therefore, in the mold forming control process, how to accurately predict the mold offset trend under the joint action of multiple factors while taking into account modeling efficiency, and effectively generate correction instructions that can be used for actual control execution, remains a key technical challenge facing this field. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present application provides a mold foaming control system based on image import and model simulation, comprising: an acquisition module for acquiring image data of a target object under a preset perspective layout, and performing three-dimensional reconstruction based on the image data to generate a corresponding target three-dimensional model;

[0006] a processing module, configured to perform morphological processing on the target three-dimensional model to generate a standard geometric model suitable for mold cavity control;

[0007] a conversion module, configured to extract control position information for characterizing the spatial contour of the mold based on the standard geometric model, and convert the control position information into a control instruction sequence matching the cavity drive structure;

[0008] a control module, configured to input the control instruction sequence into a controller, and drive the mold cavity assembly to adjust its spatial structure so that the mold cavity matches the target three-dimensional model in contour during the molding stage;

[0009] The conversion module is further configured to:

[0010] The offset prediction model trained through the knowledge distillation mechanism is called. The offset prediction model takes the structural code, thermal field data, assembly parameters, and material parameters as inputs, and outputs a compensation vector corresponding to the control position information. The compensation vector is used to perform vector superposition with the control position information to generate a control instruction correction term.

[0011] Among them, the offset prediction model is obtained by fitting the offset vector and intermediate behavior characteristics generated by a teacher model with thermal-stress coupling simulation capabilities during the training stage, and is called in real time in the molding control process for online compensation.

[0012] As an optional implementation, the acquisition module is further configured to:

[0013] Acquire a structural data source, and identify mold boundary features, connection location information, and component identification data based on the structural data source to construct a structural recognition result;

[0014] Based on the structure recognition result, topological relationship extraction is performed to generate a structure code for describing the hierarchical relationship of the mold structure;

[0015] Wherein, the structural data source includes: geometric structure and connection information of mold components.

[0016] As an optional implementation, the acquisition module is further configured to:

[0017] Obtain the basic physical property database records corresponding to the mold material, extract the material attribute fields, and construct the material property table;

[0018] performing a standardized encoding operation on the material property table to generate material parameters for describing physical properties of the mold material;

[0019] The material property fields include: thermal expansion coefficient, thermal conductivity, and elastic modulus.

[0020] As an optional implementation, the acquisition module is further configured to:

[0021] Acquire a data stream from thermal sensors placed in a target area of ​​the mold, and reconstruct a time series of real-time temperature values ​​of each node in the data stream to generate a multi-node temperature time series graph;

[0022] Spatial interpolation and partition classification are performed on the temperature time series diagram to generate thermal field data for reflecting the thermal distribution state of the mold.

[0023] As an optional implementation, the acquisition module is further configured to:

[0024] Obtaining an assembly sequence log and component pairing parameters from a mold assembly process control system, parsing component dependencies based on the assembly sequence log, parsing boundary connection modes based on the component pairing parameters, and constructing an assembly constraint matrix;

[0025] A graph structure encoding operation is performed on the assembly constraint matrix to generate assembly parameters for describing assembly path information.

[0026] As an optional implementation, the teacher model is used to:

[0027] Receive the structural code, extract the topological structure relationship, boundary connection position and node spatial layout between the mold parts, and generate the structural feature tensor;

[0028] Receive thermal field data, construct a spatial temperature matrix, and perform temperature-structure mapping in combination with the structural feature tensor to generate a temperature tensor sequence;

[0029] Receive assembly parameters, analyze component assembly sequence and connection constraints, and generate assembly drawing structure;

[0030] receiving material parameters, extracting the thermal expansion coefficient, thermal conductivity, and elastic modulus of each region of the mold, and generating a material property vector;

[0031] Based on the structural feature tensor, temperature tensor sequence, assembly drawing structure and material property vector, a thermal-stress coupling simulation is performed in the teacher model to calculate the three-dimensional spatial offset of the target control point under the set thermal working condition and generate a training offset vector;

[0032] During the simulation process, intermediate stress response characteristics, thermal distribution change paths, and local structural response indicators are further extracted to form intermediate behavior characteristics. The training offset vector and the intermediate behavior characteristics are used together as supervisory information for training the offset prediction model.

[0033] As an optional implementation, the knowledge distillation mechanism includes:

[0034] Using the training offset vector as a first supervision target, calculating a first supervision loss value between the output result of the student model and the training offset vector;

[0035] Taking the intermediate behavior features as the second supervision target, perform feature alignment on the feature expression of the intermediate layer of the student model and calculate the second supervision loss value;

[0036] Performing a weighted fusion of the first supervised loss value and the second supervised loss value to generate a total loss function as an optimization target for parameter updating in the training phase;

[0037] During the training process, multiple sets of training samples are constructed based on the structural encoding, thermal field data, assembly parameters, and material parameters, and the model parameter update operation is repeatedly performed to obtain the student model;

[0038] During the molding control process operation phase, the current structural code, thermal field data, assembly parameters, and material parameters are input into the student model, and a compensation vector corresponding to the current control position information is calculated and output. The compensation vector and the control position information are vector-superimposed to generate the control instruction sequence for the control module to call.

[0039] As an optional implementation, based on the standard geometric model, extracting control position information for characterizing the mold space contour, and converting the control position information into a control instruction sequence matching the cavity drive structure includes:

[0040] Performing boundary curvature extraction on the standard geometric model, combining the local structural response index in the intermediate behavior feature to identify boundary abnormal areas with high curvature change, sparse point cloud or connection jump features;

[0041] Constructing a boundary buffer mask for the abnormal boundary area, and performing trajectory interpolation and weighted smoothing on the control position information within the buffer mask to generate a continuous control point sequence;

[0042] Perform vector superposition on the continuous control point sequence and the compensation vector in the order of control point indexes to generate a compensation control path sequence;

[0043] Performing trajectory segment analysis on the compensation control path sequence, and performing continuity check on the angle mutations and path jump amplitudes between adjacent control segments to construct a smooth transition segment;

[0044] The smooth transition section is merged with the compensation control path sequence to generate a complete control path, and according to the degree of freedom configuration of the cavity drive structure, dimension projection and instruction format conversion are performed on the complete control path to generate a control instruction sequence.

[0045] As an optional implementation, performing morphological processing on the target three-dimensional model to generate a standard geometric model suitable for mold cavity control includes:

[0046] Performing a boundary contour extraction operation on the target three-dimensional model, identifying a set of boundary points at the outer edge of the model, and constructing a boundary closure inspection area;

[0047] Performing density consistency analysis on the boundary point set, identifying distribution anomaly areas in the boundary point set, and performing boundary point interpolation correction operations to generate a boundary continuity model;

[0048] Performing a scale normalization operation on the boundary continuity model, constructing a scale conversion matrix based on the main dimension vector of the model, and generating a unit scale model;

[0049] Performing a curvature smoothing operation on the unit-scale model, performing facet reconstruction and mesh adjustment based on a local curvature threshold, and generating a target surface model with a continuous curvature distribution;

[0050] The target surface model is fused with the boundary continuity model to generate a standard geometric model suitable for mold cavity control.

[0051] As an optional embodiment, inputting the control instruction sequence into a controller, and having the controller drive the mold cavity assembly to adjust its spatial structure so that the mold cavity matches the target three-dimensional model in contour during the molding stage includes:

[0052] Performing degree of freedom decomposition processing on the control position information in the control instruction sequence, and mapping each control instruction to a corresponding driving submodule according to the structural layout of the mold cavity assembly;

[0053] Performing instruction format conversion on the mapping control instruction generated by the driving submodule, converting the position information into a corresponding displacement target value, motor rotation angle value or hydraulic stroke value, and generating an execution instruction stream;

[0054] The execution instruction stream is synchronously distributed according to a preset time step, and the instruction buffer is set according to the response parameters of each driver submodule to control the signal beat;

[0055] During the deformation process of the cavity component, the execution status information fed back by the position encoder, strain sensor or displacement sensor is collected to generate feedback data;

[0056] The difference between the feedback data and the target value in the execution instruction stream is calculated, and if the difference exceeds a tolerance threshold, a compensation correction instruction is output to the corresponding driving submodule.

[0057] Compared with the prior art, the mold foaming control system based on image import and model simulation provided by the embodiment of the present application has significant improvements in the real-time adaptability of structural control. The actual workpiece contour is obtained through the image acquisition and three-dimensional reconstruction module, and the standard geometric model constructed by the processing module is combined to enable the mold cavity control process to be close to the physical form, which has advantages in contour reconstruction accuracy and flexibility compared to traditional methods that rely on static design models. The present application also introduces an offset prediction model that integrates structural coding, thermal field data, assembly parameters and material parameters, so that the generation process of the mold control path can dynamically consider the interference of multiple sources under actual working conditions, and to a certain extent alleviates the influence of spatial deformation caused by thermal stress, assembly deviation and material property differences on molding accuracy. The model is deployed on the edge computing node with a lightweight network structure and is trained by a knowledge distillation mechanism. It has good industrial application deployment performance, which not only maintains prediction accuracy but also takes into account computational efficiency.

[0058] Furthermore, the molding control process utilizes a method that superimposes compensation vectors and control position information to generate the final control command, enabling online correction of the initial path. This approach can address offset errors caused by dynamic factors such as thermal field fluctuations during the molding process, enhancing the closed-loop adaptive capabilities of the control. While ensuring real-time system response and control accuracy, the overall system design improves the accuracy of mold structure adjustment and molding consistency, making it suitable for a variety of foam molding scenarios with complex contours or high molding requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 Schematic diagram of the mold foaming control system based on image import and model simulation provided in this application;

[0060] Figure 2 A flowchart of a knowledge distillation method provided in this application;

[0061] Figure 3 A flowchart of a method for converting control position information into a control instruction sequence matching a cavity drive structure provided in this application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments.

[0063] See also Figure 1 , which is a schematic diagram of a mold foaming control system based on image import and model simulation provided by this application, including:

[0064] An acquisition module 10 is configured to acquire image data of a target object under a preset viewing angle layout, and perform 3D reconstruction based on the image data to generate a corresponding 3D model of the target object;

[0065] a processing module 20 for performing morphological processing on the target three-dimensional model to generate a standard geometric model suitable for mold cavity control;

[0066] A conversion module 30 is used to extract control position information used to characterize the mold space contour based on the standard geometric model, and convert the control position information into a control instruction sequence that matches the cavity drive structure;

[0067] a control module 40 for inputting the control instruction sequence into a controller, and for driving the mold cavity assembly by the controller to adjust its spatial structure so that the mold cavity matches the target three-dimensional model in contour during the molding stage;

[0068] The conversion module 30 is further configured to:

[0069] The offset prediction model trained through the knowledge distillation mechanism is called. The offset prediction model takes the structural code, thermal field data, assembly parameters, and material parameters as inputs, and outputs a compensation vector corresponding to the control position information. The compensation vector is used to perform vector superposition with the control position information to generate a control instruction correction term.

[0070] Among them, the offset prediction model is obtained by fitting the offset vector and intermediate behavior characteristics generated by a teacher model with thermal-stress coupling simulation capabilities during the training stage, and is called in real time in the molding control process for online compensation.

[0071] First, some expressions in this application are explained: the standard geometric model refers to a three-dimensional geometric model that has been preprocessed and can be used for subsequent mold control; the structural encoding is data used to describe the topological relationship and connection characteristics between mold components; the thermal field data is the temperature distribution data on the mold surface; the assembly parameters are information on the order and method of component connection during the mold assembly process; the material parameters are the thermal and mechanical property data of the mold material, such as the thermal expansion coefficient, thermal conductivity and elastic modulus; the compensation vector is used to correct the deviation of the mold control position information; the offset prediction model is a prediction model based on machine learning, which is used to generate the compensation vector; the teacher model is a complex simulation model that obtains supervision data through thermal-stress coupling simulation.

[0072] Research has found that during the mold foaming process, the actual structure of the mold cavity will undergo slight deformation due to uneven temperature changes, assembly errors and differences in material properties. This slight deformation is difficult to accurately predict and compensate for using traditional static or linear methods.

[0073] Therefore, this embodiment proposes a knowledge distillation mechanism based on thermal-stress simulation: first, a complex teacher model is established through simulation to obtain the structural offset and local response characteristics of the mold under hot working conditions; then, the rich knowledge contained in the teacher model is compressed into a lightweight offset prediction model through the knowledge distillation method; finally, the offset prediction model is called in real time during the actual foaming molding process to output a compensation vector for online correction of control instructions, thereby improving the structural control accuracy of the mold during foaming molding.

[0074] For example, during the image acquisition stage, several industrial-grade color cameras are used and installed at multiple fixed viewing angles around the mold to obtain multi-view two-dimensional images of the target object; the multi-view images are reconstructed into a three-dimensional model through a computer vision algorithm, which can be achieved using a general open source three-dimensional reconstruction tool chain, such as the open source Structure from Motion (SfM) tool.

[0075] After obtaining the three-dimensional model, the processing module 20 first performs model preprocessing, including boundary contour extraction and inspection, point cloud density analysis and interpolation completion, scale normalization processing, and curvature smoothing processing of local areas, and finally obtains a standard geometric model with continuous boundaries, unified scale and continuous curvature to meet subsequent control requirements.

[0076] Based on the standard geometric model, the conversion module 30 extracts control position information representing the mold's spatial profile and generates a preliminary control instruction sequence based on this control position information. Subsequently, the pre-trained offset prediction model is invoked in real time via an edge computing device. Using the structural code, thermal field data, assembly parameters, and material parameters as real-time input, the module outputs a compensation vector corresponding to the control position information. This compensation vector is superimposed on the initial control instructions in real time to produce a corrected instruction sequence for actual control.

[0077] The offset prediction model is trained using knowledge distillation technology. First, a teacher model is built using common finite element analysis simulation software. Thermal-stress coupled simulations are performed to obtain offset vectors of key mold control points under different operating conditions. This supervised data is then integrated into a training dataset. Model training is performed using a conventional machine learning framework, employing a dual-loss function structure, such as predicting output deviation loss and feature consistency loss. After training, the model is converted to a standard universal format and deployed on industrial edge computing devices for real-time access.

[0078] During the actual molding control stage of the mold, the Programmable Logic Controller (PLC) is responsible for receiving the control command sequence from the edge computing device and driving the mold cavity drive component to adjust the spatial structure. At the same time, the control module 40 collects real-time feedback information such as displacement and strain, and compares the feedback information with the command target value in real time. If it is found to exceed the set tolerance, the offset prediction model is called again to make a new compensation prediction, ensuring real-time dynamic correction of the control command and building a closed-loop mold control system.

[0079] The equipment, computing platforms, and sensors involved in the implementation of the above-mentioned system are all general industry products with no special model restrictions. They only need to meet the accuracy and real-time requirements of industrial-grade data acquisition, image processing, simulation calculation, and motion control.

[0080] This embodiment helps improve the mold foaming control system's ability to predict structural changes under complex operating conditions, facilitating real-time and accurate dynamic correction of mold structural control instructions, thereby improving structural accuracy and dimensional consistency during the mold foaming process. The control system described in this embodiment helps reduce structural errors in foamed products caused by factors such as thermal deformation, assembly errors, and material differences, improving actual production efficiency and product quality, and reducing mold commissioning cycles and production costs.

[0081] In addition, this embodiment is applicable to mold control scenarios of various complex or special-shaped curved surface structures, has good versatility and flexibility, and can be widely used in fields with high requirements on molding precision, such as packaging products, automotive interior parts, and electronic components.

[0082] Furthermore, with respect to the above-mentioned acquisition module 10, in a specific implementation, the acquisition module 10 is used to obtain image data of the target object to be processed, and perform high-quality three-dimensional reconstruction based on the image data to generate an accurate target three-dimensional model, thereby providing reliable three-dimensional geometric input data for subsequent control processes.

[0083] First, regarding the image acquisition process, in this embodiment, the acquisition module 10 may include several industrial-grade digital cameras, such as those with a commonly used Gigabit Ethernet (GigE) interface, mounted at various preset viewing angles around the target object. The number of cameras typically ranges from three to six, depending on specific needs and the complexity of the target. The camera's mounting angle and position are pre-calculated to ensure that the captured two-dimensional image captures all key structural details on the target object's surface, minimizing image loss due to occlusion or blind spots.

[0084] Exemplarily, the camera layout method includes: first, using CAD or other 3D design tools to create a preliminary 3D model of the target object to simulate the camera's field of view; then, using simulation tools to simulate the imaging effect under a multi-camera arrangement to determine the optimal viewing angle layout; finally, based on the simulation results, the camera is fixed in the corresponding position and accurately positioned and leveled using an industrial fixed bracket.

[0085] Secondly, during the image data acquisition phase, each camera is triggered synchronously for acquisition via an industrial camera software development kit (SDK), such as a general-purpose industrial camera software development kit (SDK). This ensures that image data from all viewpoints is acquired at the same time, avoiding misalignment of image details due to delayed or asynchronous capture. During the actual acquisition process, camera images are typically stored as high-resolution standard red, green, and blue (RGB) images or grayscale images to meet the image quality requirements of the subsequent 3D reconstruction process.

[0086] The acquisition module 10 also includes a light source to provide a uniform and stable lighting environment for the target object, ensuring stable and clear image details. Specifically, a shadowless ring-shaped light emitting diode (LED) or LED surface light source can be used to achieve uniform surface illumination, avoiding shadows and bright spots in the image and reducing 3D reconstruction errors caused by uneven lighting.

[0087] After image data acquisition is completed, the acquisition module 10 performs a high-precision 3D reconstruction process based on the acquired multi-view image data. This process can be implemented using common open-source 3D reconstruction tools, such as the common Structure from Motion (SfM) algorithm and Multi-View Stereo (MVS) method. A typical engineering tool chain combination may include OpenMVG and OpenMVS, where:

[0088] OpenMVG is used to perform preliminary camera pose estimation, feature point extraction and matching, and generate the initial sparse 3D point cloud;

[0089] OpenMVS further refines the sparse point cloud to generate dense and precise point cloud data, and then builds an accurate 3D model of the target.

[0090] For example, OpenMVG is used to perform feature detection and matching such as Scale-Invariant Feature Transform (SIFT) or Oriented FAST and Rotated BRIEF (ORB) on all view images to preliminarily estimate the camera's pose parameters and obtain an initial sparse point cloud. Based on the sparse point cloud and camera pose information, OpenMVS is used to perform dense point cloud expansion and surface reconstruction to obtain an initial high-precision 3D model. The generated 3D model is further subjected to surface meshing and texture mapping to obtain a target 3D model with complete surface details. Point cloud processing tools such as open source CloudCompare and Open3D are used to perform a preliminary check on the model to eliminate abnormal points or outliers in the reconstruction process, ensuring that the quality of the generated target 3D model meets the strict requirements of the subsequent control process.

[0091] Regarding the above processing module 20:

[0092] As an optional implementation, performing morphological processing on the target three-dimensional model to generate a standard geometric model suitable for mold cavity control includes:

[0093] Performing a boundary contour extraction operation on the target three-dimensional model, identifying a set of boundary points at the outer edge of the model, and constructing a boundary closure inspection area;

[0094] Performing density consistency analysis on the boundary point set, identifying distribution anomaly areas in the boundary point set, and performing boundary point interpolation correction operations to generate a boundary continuity model;

[0095] Performing a scale normalization operation on the boundary continuity model, constructing a scale conversion matrix based on the main dimension vector of the model, and generating a unit scale model;

[0096] Performing a curvature smoothing operation on the unit-scale model, performing facet reconstruction and mesh adjustment based on a local curvature threshold, and generating a target surface model with a continuous curvature distribution;

[0097] The target surface model is fused with the boundary continuity model to generate a standard geometric model suitable for mold cavity control.

[0098] In actual production environments, target 3D models reconstructed directly from image data often suffer from issues such as discontinuous boundary contours, uneven point cloud density, local geometric anomalies, and inconsistent scales, making them difficult to meet the requirements for precise mold cavity control. Therefore, the processing module 20 of this embodiment effectively optimizes and processes the target 3D model's geometry, generating a high-precision standard geometric model suitable for subsequent mold cavity control.

[0099] During implementation, the target 3D model undergoes boundary contour extraction and closure checks. This includes using point cloud processing tools, such as the contour extraction function in CloudCompare or Open3D, to identify the target model's external edge feature point set. An initial closed boundary contour line is then generated based on these contour feature points. A contour closure algorithm is then used to check boundary closure. If discontinuities or gaps are detected, interpolation and boundary contour optimization and correction algorithms are implemented. For example, a spherical interpolation completion algorithm based on the Random Sample Consensus (RANSAC) algorithm or a Poisson surface reconstruction algorithm are used to repair and ensure the continuity and closure of the model boundary.

[0100] Secondly, for areas of the target 3D model with uneven point cloud density or areas with excessively high or low density, local density analysis and optimization are performed. Specific implementations include: using statistical methods to analyze the local point cloud density distribution of the target model, using surface fitting interpolation to complete the point cloud in areas with density below a set threshold; and using uniform downsampling algorithms, such as voxel grid filtering or the farthest point sampling algorithm, to balance the point cloud density in areas with excessively high density, resulting in an optimized model with uniform point cloud density.

[0101] The optimized model is then scaled normalized. This can include using principal component analysis (PCA) or an oriented bounding box (OBB) algorithm to calculate the model's principal dimension vectors, and then calculating a scale transformation matrix based on the principal dimension vectors. This allows the entire model to be scaled uniformly to a preset unit scale range, such as 1m×1m×1m, eliminating scale differences between different model sizes and facilitating subsequent precise control path generation and comparison between models.

[0102] Next, the processing module 20 further performs curvature analysis and smoothing processing on the scale-normalized model to improve the continuity and smoothness of the model surface.

[0103] For example, open source software such as Trimesh is used to calculate the curvature characteristic values ​​of each region of the model to identify areas with high curvature mutations. Surface subdivision and local mesh optimization algorithms are then performed on the areas with curvature mutations, such as Laplace smoothing or quadratic B-spline interpolation, to obtain a refined surface model with a continuous curvature distribution.

[0104] Finally, the processing module 20 fuses the above-mentioned local optimization models after boundary correction, density optimization, scale normalization and curvature smoothing. Specifically, it can generate a standard geometric model through mature three-dimensional model fusion algorithms such as voxel grid fusion or Poisson reconstruction.

[0105] In addition, the above processing can be automatically executed through standardized Python scripts, automating pipeline operations to facilitate rapid deployment and consistent implementation in industrial environments.

[0106] It is understood that the above example software or tools are only applicable to the method disclosed in this application and should not be regarded as any limitation.

[0107] Regarding the above conversion module 30:

[0108] During the mold foaming process, the standard geometric model must be further converted into a precise control instruction sequence that matches the mold cavity drive structure to achieve accurate control of the mold cavity's spatial structure. Furthermore, considering complex factors such as temperature variations, material differences, and assembly errors in actual working conditions, this implementation also proposes an offset prediction model trained using a knowledge distillation mechanism to compensate for position errors in control instructions in real time, improving instruction accuracy and control effectiveness.

[0109] In specific implementations, control position information used to characterize the mold's spatial contour is extracted based on a standard geometric model. This involves performing spatial contour feature extraction on the model to generate a discretized set of spatial control points. Possible implementation tools include open-source 3D processing tools such as Open3D, MeshLab, and CloudCompare.

[0110] For example, the surface of the standard geometric model is first uniformly sampled (such as Poisson disk sampling) to obtain the initial distribution of spatial control points; then, curvature or normal vector analysis is performed on the set of spatial control points through custom scripts or built-in software functions to optimize the spatial distribution density of the control points, ensure that the control points evenly and reasonably cover the model surface, and avoid the problem of insufficient local control accuracy.

[0111] Secondly, the conversion module 30 converts the extracted spatial control position information into a control instruction sequence that matches the specific cavity drive structure.

[0112] In the specific implementation, the degree of freedom layout of the mold cavity drive structure is first clarified, such as the X, Y, and Z axis translational degrees of freedom and possible rotational degrees of freedom, and the corresponding degree of freedom mapping rules and conversion matrices are established according to the spatial layout and degree of freedom allocation of the drive structure; then, by writing an automated Python script or industrial control software script, the position of each spatial control point is converted into the target control instruction value of the specific control axis according to the degree of freedom mapping rule, such as a specific displacement stroke value or rotation angle value, to form an industrial control instruction format that matches the actual drive device, such as a PLC-compatible format or a G-code format.

[0113] Furthermore, in order to effectively deal with the problem of slight deviations caused by thermal field fluctuations, assembly errors or material property differences during the actual molding process of the mold, the conversion module 30 obtains an offset prediction model through knowledge distillation mechanism training, and calls it in real time in the actual control stage to generate a compensation vector.

[0114] In the specific implementation, the offset prediction model is obtained by fitting the supervision data generated by the teacher model, i.e., the offset vector and the intermediate behavior features, during the training phase.

[0115] In the specific implementation, a teacher model is first built through general finite element simulation software such as Ansys and COMSOL, and the heat-stress coupling simulation of the mold under different thermal field distribution and assembly error scenarios is carried out to obtain the spatial offset vector data of the control point and the intermediate behavior characteristics of the temperature gradient and stress change; then these supervision data are integrated to construct a training data set.

[0116] During the training of the offset prediction model, a lightweight student model is built using common machine learning frameworks such as PyTorch or TensorFlow. Either a convolutional neural network (CNN) or a multi-layer perceptron (MLP) architecture can be used. A dual loss function structure is designed during training, including a prediction output deviation loss, such as the mean squared error (MSE) loss, and a feature consistency loss, such as the L2 norm loss. After training, the student model is converted to a common model format supported by industrial computing platforms and deployed on edge computing devices for real-time access and prediction.

[0117] In an actual production environment, the conversion module 30 receives real-time structural encoding, thermal field data, assembly parameters, and material parameters as inputs to the offset prediction model. This input data can be provided by system-provided sensors and a real-time status database, such as industrial configuration software. It includes data collected by real-time temperature sensors, assembly parameter data derived from an assembly log database, and thermodynamic parameters from a material property database. The offset prediction model outputs a compensation vector in real time, which is used to perform vector superposition on the original control position information, thereby generating real-time control instruction correction terms.

[0118] The conversion module 30 further performs the following process: It fuses and superimposes the preliminary control instruction sequence with the real-time generated control instruction correction terms. The specific fusion method can be direct vector superposition or weighted summation to form a precise control instruction sequence for the control module 40. This instruction sequence is sent to the control module 40 in real time using a standard industrial communication protocol to achieve precise control of the mold cavity assembly.

[0119] Regarding the above control module 40:

[0120] As an optional embodiment, inputting the control instruction sequence into a controller, and having the controller drive the mold cavity assembly to adjust its spatial structure so that the mold cavity matches the target three-dimensional model in contour during the molding stage includes:

[0121] Performing degree of freedom decomposition processing on the control position information in the control instruction sequence, and mapping each control instruction to a corresponding driving submodule according to the structural layout of the mold cavity assembly;

[0122] Performing instruction format conversion on the mapping control instruction generated by the driving submodule, converting the position information into a corresponding displacement target value, motor rotation angle value or hydraulic stroke value, and generating an execution instruction stream;

[0123] The execution instruction stream is synchronously distributed according to a preset time step, and the instruction buffer is set according to the response parameters of each driver submodule to control the signal beat;

[0124] During the deformation process of the cavity component, the execution status information fed back by the position encoder, strain sensor or displacement sensor is collected to generate feedback data;

[0125] The difference between the feedback data and the target value in the execution instruction stream is calculated, and if the difference exceeds a tolerance threshold, a compensation correction instruction is output to the corresponding driving submodule.

[0126] The control module 40 is responsible for executing the control command sequence within the entire mold foaming control system. It drives the mold cavity components to precisely adjust the spatial structure to ensure that the mold cavity accurately matches the contours of the target 3D model during the molding process. Furthermore, the control module 40 provides real-time status feedback and a closed-loop compensation control mechanism, effectively improving control accuracy and structural consistency.

[0127] In a specific implementation, control module 40 receives a precise control instruction sequence generated by conversion module 30, which already takes into account online compensation vector corrections. To achieve precise motion control, the control instruction sequence can adopt an instruction format commonly used in the industrial control field, such as an industrial control format acceptable to PLCs, G-code instructions commonly used in numerical control (CNC) machining, or industrial communication protocols such as OPC UA, EtherCAT, or Modbus TCP.

[0128] Next, the control module 40 inputs the received control command sequence into an industrial controller, such as a general-purpose industrial PLC controller. Specifically, an industrial control platform with real-time response and high-precision control capabilities can be used. After receiving the control command sequence, the PLC controller interprets the command sequence into specific servo drive signals, such as position control commands or speed control commands, based on the control system's predefined degree-of-freedom allocation rules and control objectives.

[0129] The PLC controller then sends the parsed drive signal to the mold cavity assembly drive mechanism in real time. The specific implementation of the drive mechanism can adopt a sophisticated precision motion control unit from the industrial field, such as a servo motor with a ball screw, linear motor, hydraulic or pneumatic actuator. The specific selection is determined by the required accuracy, load, response speed, and other requirements of the mold cavity. For example, the PLC sends control instructions synchronously to multiple servo drive modules via an industrial real-time communication protocol. The servo drive modules execute position or speed control instructions in real time, driving the mold cavity to accurately complete spatial position adjustment, ensuring that the spatial contour of the cavity is strictly consistent with the target three-dimensional model.

[0130] Furthermore, to accurately monitor the real-time position of the mold cavity and provide closed-loop control feedback, control module 40 also incorporates a real-time position feedback and closed-loop control mechanism. This can be implemented using position encoders, displacement sensors, laser rangefinders, or strain sensors. Real-time sensors are installed at key locations within the mold cavity assembly to measure the cavity's actual spatial position and structural changes in real time. Real-time position feedback data is directly collected by the PLC controller, which performs real-time error calculation and judgment within the controller and compares it with the command target value in real time, forming a closed-loop control mechanism.

[0131] In practice, the PLC controller compares the real-time feedback data it collects with the preset control command target value, performing error analysis. If the real-time position deviation exceeds the preset tolerance threshold, the PLC controller automatically triggers the real-time compensation mechanism of the offset prediction model. This mechanism notifies the edge computing device via the industrial communication protocol to recalculate the compensation vector and make real-time corrections to the control command. This approach builds a high-precision closed-loop feedback control system, allowing for real-time adjustment of the mold cavity spatial position and rapid, accurate dynamic error correction.

[0132] Furthermore, the control module 40 described in this embodiment also includes a detailed mapping process for the degrees of freedom of the control instruction sequence during its execution. In practice, a degree of freedom allocation matrix can be established based on the specific structural layout of the mold cavity assembly, automatically mapping the control position information of the spatial control instruction sequence to the degrees of freedom of each driver submodule. For example, the degree of freedom allocation matrix can be implemented using a numerical matrix or a graphical interface. Specifically, automatic parsing and mapping can be performed using an automatic mapping program or script built into the industrial PLC controller.

[0133] Furthermore, the control module 40 specifically performs an instruction format conversion process, employing real-time calculation methods to explicitly convert spatial position information into target values ​​executable by specific drive submodules. For example, the control module 40 may employ a real-time coordinate transformation algorithm based on space vectors, such as an inverse kinematics solution, to convert the three-dimensional spatial control target point into the corresponding displacement target value, motor target angle, or hydraulic target stroke for each drive unit. This process can be automated using pre-programmed industrial control scripts within a PLC controller or real-time industrial PC, ensuring conversion accuracy and real-time performance.

[0134] Furthermore, the control module 40 further implements a synchronized control command distribution mechanism, sending control commands incrementally at predetermined time steps, such as millisecond-level fixed intervals. It also specifies a command buffer corresponding to the response parameters of each driver submodule. The specific size of the command buffer can be dynamically adjusted through the PLC control program based on the measured response delay characteristics of each driver module to ensure highly synchronized operation of the overall control system and avoid degradation of structural control accuracy due to asynchronous execution timing.

[0135] Furthermore, the control module 40 also specifically implements the process of trajectory segment analysis and continuity verification. During specific implementation, angle difference and displacement difference calculations are automatically performed on adjacent control segments in the control path to identify abnormal segments where angle changes or path transitions exceed preset thresholds. Subsequently, interpolation supplementation and weighted smoothing algorithms are used on the abnormal segments, such as cubic spline interpolation and Gaussian smoothing, to automatically construct smooth transition segments, thereby generating a continuous control path with good kinematic characteristics. The complete control path after this trajectory optimization process is further converted into an actual execution instruction sequence for the corresponding drive module, thereby effectively reducing the problems of sudden action and impact in actual control and improving the stability and accuracy of the control process.

[0136] As an optional implementation, the acquisition module 10 is further configured to:

[0137] Acquire a structural data source, and identify mold boundary features, connection location information, and component identification data based on the structural data source to construct a structural recognition result;

[0138] Based on the structure recognition result, topological relationship extraction is performed to generate a structure code for describing the hierarchical relationship of the mold structure;

[0139] Wherein, the structural data source includes: geometric structure and connection information of mold components.

[0140] As an optional implementation, the acquisition module 10 is further configured to:

[0141] Obtain the basic physical property database records corresponding to the mold material, extract the material attribute fields, and construct the material property table;

[0142] performing a standardized encoding operation on the material property table to generate material parameters for describing physical properties of the mold material;

[0143] The material property fields include: thermal expansion coefficient, thermal conductivity, and elastic modulus.

[0144] As an optional implementation, the acquisition module 10 is further configured to:

[0145] Acquire a data stream from thermal sensors placed in a target area of ​​the mold, and reconstruct a time series of real-time temperature values ​​of each node in the data stream to generate a multi-node temperature time series graph;

[0146] Spatial interpolation and partition classification are performed on the temperature time series diagram to generate thermal field data for reflecting the thermal distribution state of the mold.

[0147] As an optional implementation, the acquisition module 10 is further configured to:

[0148] Obtaining an assembly sequence log and component pairing parameters from a mold assembly process control system, parsing component dependencies based on the assembly sequence log, parsing boundary connection modes based on the component pairing parameters, and constructing an assembly constraint matrix;

[0149] A graph structure encoding operation is performed on the assembly constraint matrix to generate assembly parameters for describing assembly path information.

[0150] In a specific implementation, the acquisition module 10 further acquires a mold structure data source. This data source may include 3D geometry files of mold components, such as CAD files in the Standard for the Exchange of Product Data (STEP) or Initial Graphics Exchange Specification (IGES) formats, as well as detailed assembly information for connection points. This data is typically stored in an enterprise database or engineering data management platform. The acquisition module 10 obtains this data by accessing the enterprise's PLM system or engineering database.

[0151] After obtaining the structural data source, the acquisition module 10 performs boundary feature recognition based on the three-dimensional geometric file, and uses commonly used industrial CAD data processing tools, such as open source CAD processing tools such as FreeCAD or commercial software such as SolidWorks, to extract the boundary contour lines of each mold component, key connection features such as hole positions, pin slots or joint features, and component identification numbers; and generates structural recognition results based on this information.

[0152] Furthermore, based on the structural recognition results, the acquisition module 10 extracts topological relationships using automated scripts, such as Python or MATLAB. In specific implementations, a graph data structure can be used to represent the connection relationships between mold components, generating graph structure topology data that describes the structural hierarchy and connection topology relationships between mold components, thereby generating a structural code. This structural code is subsequently used as one of the input data for the offset prediction model, enhancing the accuracy and specificity of the model's predictions.

[0153] The acquisition module 10 further obtains basic physical property data corresponding to the material used in the mold from an internal material database of the enterprise or an external public material property database, such as a standard physical property database.

[0154] Exemplarily, the acquisition module 10 extracts relevant material property fields, including but not limited to thermal expansion coefficient, thermal conductivity, and elastic modulus, based on a database query interface. Furthermore, the acquisition module 10 encodes the extracted material property data in a standardized format, such as by converting it into a unified XML or JSON format data file, to form a standardized material property table for efficient reading by the subsequent offset prediction model.

[0155] Furthermore, the acquisition module 10 uses an array of industrial-grade thermal sensors to collect real-time temperature data streams from various areas of the mold. In specific implementations, these thermal sensors, such as commonly used industrial thermocouples or RTDs, are evenly distributed throughout the mold cavity and key structural locations based on the mold's actual structure, capturing real-time surface temperature data streams. This real-time temperature data is recorded in real time by the data acquisition system.

[0156] Subsequently, the acquisition module 10 reconstructs the time series of the temperature data stream through an automated data processing script to generate a multi-node temperature time series diagram in which the temperature of each node changes with time; further, spatial interpolation is performed on the time series diagram data, for example, Kriging interpolation or radial basis function interpolation and smoothing filtering processing, the temperature data is partitioned and classified, and a more representative thermal field data diagram is generated for the input data of the subsequent real-time prediction model, thereby improving the model's ability to predict thermal state errors.

[0157] Furthermore, the acquisition module 10 can also obtain the mold assembly process sequence log and component pairing parameters from the mold assembly process control system or the enterprise's Manufacturing Execution System (MES). In implementation, assembly log data, including information such as assembly sequence, assembly timestamp, component identification number, and boundary connection method, is automatically acquired through enterprise information platform interfaces, such as MES interfaces or product lifecycle management (PLM) database application programming interfaces (APIs).

[0158] Exemplarily, the acquisition module 10 automatically analyzes the assembly dependencies between components based on the assembly sequence log and, based on the component pairing parameter data, identifies the specific connection forms between component boundaries, such as bolted connections, pinned connections, and glued connections. Furthermore, the acquisition module 10 uses a graph data processing algorithm, such as the graph theory library NetworkX, to convert the dependencies and boundary connection information into an assembly constraint matrix. It then performs a graph structure encoding operation on the matrix to generate assembly parameters that describe the assembly path and dependencies.

[0159] As an optional implementation, the teacher model is used to:

[0160] Receive the structural code, extract the topological structure relationship, boundary connection position and node spatial layout between the mold parts, and generate the structural feature tensor;

[0161] Receive thermal field data, construct a spatial temperature matrix, and perform temperature-structure mapping in combination with the structural feature tensor to generate a temperature tensor sequence;

[0162] Receive assembly parameters, analyze component assembly sequence and connection constraints, and generate assembly drawing structure;

[0163] receiving material parameters, extracting the thermal expansion coefficient, thermal conductivity, and elastic modulus of each region of the mold, and generating a material property vector;

[0164] Based on the structural feature tensor, temperature tensor sequence, assembly drawing structure and material property vector, a thermal-stress coupling simulation is performed in the teacher model to calculate the three-dimensional spatial offset of the target control point under the set thermal working condition and generate a training offset vector;

[0165] During the simulation process, intermediate stress response characteristics, thermal distribution change paths, and local structural response indicators are further extracted to form intermediate behavior characteristics. The training offset vector and the intermediate behavior characteristics are used together as supervisory information for training the offset prediction model.

[0166] During the mold foaming process, in order to accurately predict the tiny structural deviations of the mold under complex thermal conditions and assembly errors, this implementation uses a teacher model based on thermal-stress coupling simulation to generate high-quality supervision data for training a lightweight deviation prediction model.

[0167] Specifically, the teacher model of this embodiment generates supervision information, including training offset vectors and intermediate behavior features, through multi-source data fusion and fine simulation processes. In a specific implementation, the teacher model receives structural coding, wherein the structural coding includes the topological relationship between the various components of the mold, the boundary connection position, and the spatial layout information. These data are usually generated by the aforementioned structure recognition process and input into the teacher model simulation platform in a structured data format, such as eXtensible Markup Language (XML) or JavaScript Object Notation (JSON). The teacher model can use general finite element simulation software, such as Ansys, COMSOL, etc., to import and parse structural data and construct a structural feature tensor. The structural feature tensor represents the spatial layout of the mold components and their connection characteristics in the form of a three-dimensional matrix or a multi-dimensional array.

[0168] Secondly, the teacher model receives thermal field data and performs temperature-structure mapping operations. In specific implementations, thermal field data is obtained through a temperature sensor array collected in real time and imported into the simulation model in the form of a standardized time series or matrix. The teacher model simulation platform first constructs a spatial temperature matrix and performs spatial interpolation based on the node positions in the structural feature tensor to form complete temperature field distribution data. Subsequently, based on the spatial coordinate mapping relationship between the structural feature tensor and the temperature matrix, accurate mapping of the temperature field to the structural nodes is achieved, and a temperature tensor sequence is obtained, specifically the temperature distribution tensor at multiple moments.

[0169] Furthermore, the teacher model also receives assembly parameter information. Specifically, these assembly parameters include the assembly sequence logs and connection constraints of the mold components. These data are typically imported into the teacher model simulation platform in a standard database format or structured text file. Based on the assembly sequence logs, the teacher model automatically constructs the mold's assembly drawing structure. For example, a graph data structure is used to represent the dependencies and connection constraints between components. This assembly drawing structure explicitly describes the order and connection constraints between components during the assembly process, providing accurate assembly status and constraint information for subsequent simulations.

[0170] The teacher model then receives material parameters, including physical properties such as the coefficient of thermal expansion, thermal conductivity, and elastic modulus for each region of the mold. These parameters are imported into the simulation platform using a standard database query or a unified format to construct a standardized material property vector, which is used to describe the material's thermodynamic properties in the simulation model.

[0171] Based on the above structural feature tensors, temperature tensor sequences, assembly drawing structures and material property vectors, the teacher model implements high-precision thermal-stress coupling simulation in the finite element simulation platform.

[0172] For example, a 3D mold simulation model is automatically constructed within a finite element simulation platform using structural feature tensors and assembly drawing structures, completing the assembly sequence and applying constraints. A temperature tensor sequence is applied to the structural model to perform heat conduction simulation and coupled thermal stress calculations. During the simulation, the finite element simulation platform's built-in multiphysics coupling module accurately simulates the stress response and 3D deformation of the mold structure caused by temperature changes under preset thermal conditions. The teacher model uses post-processing tools within the simulation platform to extract the actual 3D offsets of key control points of the mold during simulation and generate accurate training offset vector data. Specific control points are selected, including key locations within the mold structure, boundary connections, and specific nodes in areas requiring high precision. Furthermore, during the simulation, the teacher model extracts a wealth of intermediate behavioral feature data, including stress field distribution characteristics, such as principal stress and shear stress field distribution; temperature field distribution paths, i.e., the evolution of temperature gradient distribution; and local structural response indicators, such as stress and deformation data at local nodes. This data is automatically extracted and integrated into an intermediate behavioral feature dataset using standard post-processing scripts.

[0173] Through this refined, multi-source data fusion simulation process, the teacher model generates precise training offset vectors and detailed intermediate behavioral feature data. This high-quality supervisory data guides the knowledge distillation process of the lightweight offset prediction model, effectively improving its predictive power and generalization performance in real-world thermal environments.

[0174] In this way, this embodiment can effectively solve the problem that traditional pure empirical prediction or simple linear model cannot accurately handle the structural deviation problem under complex thermal conditions and assembly errors of the mold, significantly improve the training quality and prediction accuracy of the offset prediction model, and provide a powerful prediction and control foundation for the mold foaming control system.

[0175] For example, the offset prediction model can adopt a "multi-input-fusion-multi-task regression" architecture. First, the structural code is fed into a multi-layer graph convolutional network (GCN), each layer containing dozens to hundreds of hidden nodes to extract component topology and connectivity features. Thermal field data is fed into a three-dimensional convolutional network (3D-CNN), consisting of three to five layers of convolution and pooling units, to reflect temperature distribution and gradient changes. Assembly parameters are fed into a lightweight sequence model, such as a two-layer gated recurrent network or a simplified Transformer, to capture assembly order and constraint information. Material parameters are mapped into a low-dimensional embedding using a two-layer fully connected network. After the output features of each sub-network are aggregated along the channel dimension, they enter a fusion layer based on a multi-head attention mechanism. The number of heads can be configured based on hardware capabilities. After fusion, the model undergoes step-by-step dimensionality reduction through a four- to six-layer fully connected regression network. The final output is a three-dimensional offset vector and several dimensional intermediate response features, such as local maximum stress and temperature gradient indicators. To meet the needs of industrial edge deployment, batch normalization and dropout can be added after each layer, and weight pruning or quantization can be performed after network training.

[0176] For example, during training, the simulation platform batch generates several data sets of different structures, thermal conditions, and assembly scenarios, including the actual offsets and intermediate behavior features of target control points, for constructing training samples.

[0177] After training is completed, the model is exported to ONNX or TensorRT format and deployed on industrial PC or PLC edge computing nodes to achieve online prediction and compensation calls.

[0178] As an optional implementation, see Figure 2 , is a flow chart of a knowledge distillation method provided in this application, including steps S101 to S105, wherein:

[0179] S101: Using the training offset vector as a first supervision target, calculating a first supervision loss value between an output result of the student model and the training offset vector;

[0180] S102: Using the intermediate behavior features as the second supervision target, perform feature alignment on the feature expression of the intermediate layer of the student model, and calculate the second supervision loss value;

[0181] S103: Perform weighted fusion on the first supervised loss value and the second supervised loss value to generate a total loss function as an optimization target for parameter update in the training phase;

[0182] S104: During the training process, multiple sets of training samples are constructed based on the structural code, thermal field data, assembly parameters, and material parameters, and the model parameter update operation is repeatedly performed to obtain a student model;

[0183] S105: During the molding control process operation phase, the current structural code, thermal field data, assembly parameters, and material parameters are input into the student model, a compensation vector corresponding to the current control position information is calculated and output, the compensation vector is vector-superimposed with the control position information, and the control instruction sequence is generated for the control module 40 to call.

[0184] This application further discloses a knowledge distillation mechanism for training the above-mentioned offset prediction model to effectively improve the prediction accuracy and generalization performance of the student model.

[0185] In practice, during the training phase, the training offset vector extracted from the teacher model's simulation results is used as the first supervised target, meaning the training offset vector serves as the true prediction target for the student model. In practice, a regression loss function, such as the mean squared error loss function, implemented in deep learning frameworks like PyTorch or TensorFlow, can be used to measure the difference between the student model's predicted output and the offset vector generated by the teacher model. This first supervised loss value is then calculated to directly constrain the accuracy of the student model's predicted output.

[0186] Secondly, the intermediate behavioral features obtained by the teacher model during the simulation process are used as the second supervision target to assist in training the intermediate layer feature expression of the student model. In the specific implementation process, a feature alignment operation is performed between the intermediate layer outputs of the student model and the intermediate behavioral features of the teacher model. For example, the L2 norm loss function of the feature layer is used to achieve dimension-by-dimension alignment between the intermediate layer outputs of the student model and the intermediate behavioral features of the teacher model, thereby calculating the second supervision loss value.

[0187] Furthermore, this embodiment performs a weighted fusion of the two supervised loss values ​​mentioned above: the first supervised loss value for offset prediction and the second supervised loss value for intermediate feature consistency. In specific implementations, a weight coefficient can be introduced to adjust the contribution ratio between the two losses, for example, a weight of 0.7 is assigned to the first supervised loss and a weight of 0.3 is assigned to the second supervised loss. The fused losses constitute the total loss function, which serves as the training target for optimizing the student model parameters.

[0188] During the training process, a large number of training samples are constructed based on a pre-prepared training dataset, which includes various combinations of structural encoding, thermal field data, assembly parameters, and material parameters. Specifically, the training dataset can be generated in batches using industrial automation scripts, such as Python or MATLAB scripts. Within the deep learning framework, batch training is then used to repeatedly update the student model parameters to ensure stability and convergence of the model training.

[0189] During training, the Adam optimization algorithm can be used to update the parameters of the student model. The initial learning rate can be set to 0.001, and the learning rate gradually decays during training to achieve stable convergence. The training batch size can be set to 32 to 64. Each training sample includes the input structure encoding, thermal field data, assembly parameters, material parameters, and the corresponding supervision target, namely the training offset vector and intermediate behavioral features. The training process is performed for 100 to 150 rounds until the training loss reaches a stable convergence state.

[0190] During the actual operation stage of the molding control process, this implementation method calls the trained student model in real time through industrial edge computing equipment or industrial PC, uses the current structural code, thermal field data, assembly parameters and material parameters as real-time input data, and calculates and outputs the compensation vector corresponding to the current control position information in real time.

[0191] In the specific implementation, the calculated compensation vector is directly superimposed with the current control position information to automatically generate a control instruction sequence that has been corrected in real time, and is sent to the control module 40 through the industrial real-time communication protocol to achieve real-time and high-precision online compensation control of the mold cavity.

[0192] In this way, the student model can effectively improve its prediction accuracy and robustness for complex thermal structure deviations under low computing resource conditions, thereby solving the problem that traditional models are difficult to respond quickly and accurately in actual production, and providing a more efficient and reliable real-time compensation method for mold precision foaming control.

[0193] As an optional implementation, see Figure 3 , which is a flow chart of a method for converting control position information into a control instruction sequence matching a cavity drive structure provided by the present application, including steps S201 to S205, wherein:

[0194] S201: performing boundary curvature extraction on the standard geometric model, combining the local structural response index in the intermediate behavior feature to identify abnormal boundary regions with high curvature variation, sparse point cloud or connection jump features;

[0195] S202: constructing a boundary buffer mask for the abnormal boundary area, and performing trajectory interpolation and weighted smoothing on the control position information within the buffer mask to generate a continuous control point sequence;

[0196] S203: performing vector superposition on the continuous control point sequence and the compensation vector in the order of control point indices to generate a compensation control path sequence;

[0197] S204: performing trajectory segment analysis on the compensation control path sequence, and performing continuity check on the angle mutations and path transition amplitudes between adjacent control segments to construct a smooth transition segment;

[0198] S205: The smooth transition section is integrated with the compensation control path sequence to generate a complete control path, and according to the degree of freedom configuration of the cavity drive structure, dimension projection and instruction format conversion are performed on the complete control path to generate a control instruction sequence.

[0199] Regarding the above S201:

[0200] In the specific implementation, the standard geometric model is imported into a 3D geometry processing tool, such as Open3D or MeshLab, and the principal curvature and Gaussian curvature values ​​are calculated for each vertex on the model surface. Specifically, the curvature estimation algorithm in the tool library can be called, for example, a curvature estimation module based on quadratic surface fitting or normal difference, to obtain the local curvature characteristics of each vertex. Subsequently, a predetermined curvature threshold is set. If the local curvature change rate is greater than a certain threshold, combined with the point cloud density information, by statistically analyzing the point cloud distribution density in the neighborhood of each vertex, the vertex sets in the curvature mutation area and the sparse area are automatically identified.

[0201] Next, the vertex set with curvature and density anomalies is fused with the local structural response indicators in the intermediate behavioral features for determination. Specifically, the node stress or strain response values ​​derived from the teacher model simulation phase can be mapped to the corresponding geometric vertices. If the stress / strain of the node exceeds the set response threshold and simultaneously meets the curvature or density anomaly conditions, it is marked as a boundary anomaly region. This fusion determination can be automated using Python scripts: first, the response indicators are read from the intermediate feature data file (JSON, CSV, etc.), then compared one by one based on the vertex index, and finally, the anomaly region mask is generated in the geometry processing environment for subsequent buffering and interpolation operations.

[0202] Regarding the above S202:

[0203] In practice, a boundary buffer mask is first generated around the set of outlier vertices identified in step S201. Specifically, a point cloud dilation algorithm in Open3D or PCL is used to spatially expand the outlier vertices with a dilation radius of 2mm to 5mm, constructing a buffer zone encompassing the outlier region and adjacent control points. Control points within this buffer mask are then filtered from the initial control point sequence, and their 3D coordinates are extracted in index order.

[0204] For the control points extracted within the buffer zone, SciPy's cubic spline interpolation (scipy.interpolate.CubicSpline) or radial basis function interpolation (scipy.interpolate.Rbf) methods can be used in the Python environment to resample the original discrete points and generate equally spaced interpolated points. A weighted smoothing algorithm (such as a Gaussian weighted moving average based on the curvature or distance of the control points) is then applied to smooth the interpolated results to remove trajectory jitter caused by sudden boundary changes. The resulting continuous control point sequence fills gaps caused by discontinuous boundaries or sparse point clouds while ensuring a smooth transition in the motion path, providing a high-quality spatial trajectory for subsequent vector overlay and command generation.

[0205] Regarding the above S203:

[0206] In practice, the continuous control point sequence generated in step S202 is imported into the trajectory calculation unit. This sequence can be stored in a standard 3D point set format, such as a JSON array or CSV table. Each control point contains its 3D coordinate information within the standard geometric model. Simultaneously, a corresponding compensation vector sequence is imported from the offset prediction model. Each vector represents the spatial position correction that should be applied to that control point in the current hot and assembled states.

[0207] The trajectory correction module then runs on the edge computing device, using vector addition processing logic to sequentially add each control point to its corresponding compensation vector. This operation can be performed using a Python script in the NumPy environment, or automated batch coordinate correction can be achieved using the PLC's motion control library.

[0208] After compensating all control points, the system generates a new control path, known as a compensated control path sequence. This sequence is structurally complete, maintains good continuity between control points, and accurately reflects the corrected differences between the target model and the field conditions. This path file can be saved in common industrial control formats, such as .Gcode, .NC, or custom structured tables, for subsequent use in trajectory smoothing and command format conversion modules.

[0209] This step completes the path correction operation without affecting the control cycle, ensuring that each control point incorporates real-time predicted error compensation information, thereby improving the response accuracy of subsequent cavity control and the physical consistency of path execution. This solution is suitable for actual deployment in automatic molding equipment control systems at the edge with medium computing power, and can be implemented by embedding it in edge computing nodes in MES / PLC / RTU systems.

[0210] Regarding the above S204:

[0211] The goal of step S204 is to analyze and verify the continuity of the compensated control path sequence to construct a smooth transition section of the trajectory, thereby avoiding mechanical shock or control instability caused by sudden trajectory changes during cavity execution.

[0212] During implementation, the compensation control path sequence is first segmented and analyzed according to the control point index order. This is typically done using fixed point spacing or path length as the unit of division, forming multiple adjacent control segments. Between each pair of adjacent segments, the system calculates the vector angle and inter-point displacement changes to identify angle mutation points and path transition points.

[0213] For example, on an edge computing platform, industrial motion control algorithm libraries, such as Beckhoff TwinCAT Motion or the trajectory detection function module within NI LabVIEW Motion, are called to calculate the trajectory angle and interpolated velocity gradient for every two segments. If the angle between adjacent segments exceeds a set threshold, such as 10°, or the displacement change exceeds a threshold, such as 2mm, the position is marked as a trajectory discontinuity.

[0214] For the above-mentioned discontinuities, this application automatically inserts a smooth transition segment within its neighborhood. This segment can be constructed using a spline interpolation algorithm, such as using a cubic B-spline, Bezier curve, or Catmull-Rom spline interpolation method, to generate a smooth curve based on the coordinates and direction information of the front and rear control points to avoid sharp line changes in the trajectory. The interpolation operation can be implemented through the interpolation module provided by the industrial control platform, or through the SciPy or Spline library in the Python environment to automatically generate a spline interpolation path.

[0215] The resulting smooth transition segment will be merged back into the compensation control path sequence to form a complete path with continuous curvature, velocity, and acceleration distribution. This helps improve the structural stability and control accuracy during mold cavity movement and is suitable for mold structure control tasks with complex degrees of freedom or linkage characteristics.

[0216] Regarding the above S205:

[0217] Step S205 is used to fuse the smooth transition segment constructed by interpolation with the original compensation control path sequence, and convert the fused complete path into a control instruction sequence that can be directly executed by the cavity drive structure.

[0218] During implementation, the smooth transition segments constructed in step S204 are first merged with the original compensation control path sequence based on control point index or spatial order to form a seamless, complete control path. This fusion operation is accomplished by splicing the data structure and verifying the continuity between the interpolated segments and the original path segments, ensuring that the spatial coordinates, interpolation order, and timestamps of the control points remain consistent and coherent after fusion.

[0219] Subsequently, dimensional projection processing is performed on the complete control path based on the actual mold cavity's drive structure type and degree of freedom configuration. In specific implementation, if the cavity execution system is a three-axis linear platform with X, Y, and Z degrees of freedom, the system automatically extracts the three-dimensional coordinate information of each control point in the path and maps it to the target values ​​of the three drive axes respectively. If the system has redundant degrees of freedom or rotational degrees of freedom, such as with an articulated arm or tilt angle control, the path is required to perform degree of freedom decomposition and constraint calculation based on a preset kinematic model or inverse solution model to obtain the actual feasible control value of each control point under the current degree of freedom configuration.

[0220] The control system can perform dimensional decomposition and projection calculations through the motion control library built into the PLC or edge controller, and automatically organize the multi-dimensional control values ​​corresponding to each control point into a complete control instruction set.

[0221] Finally, the projected control values ​​are formatted and converted to generate a control instruction sequence that meets industrial execution standards. These control instructions can be output in G-code, commonly used in CNC machining, PLC action instruction formats commonly used in industrial motion control, or custom table formats. These instructions are then sent to the execution layer device via standard industrial communication protocols, enabling precise control of the spatial trajectory of the mold cavity components.

[0222] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

Claims

1. The mold foaming control system based on image import and model simulation is characterized by: include: An acquisition module is used to obtain image data of a target object under a preset viewing angle layout, and perform three-dimensional reconstruction based on the image data to generate a corresponding three-dimensional model of the target; a processing module, configured to perform morphological processing on the target three-dimensional model to generate a standard geometric model suitable for mold cavity control; a conversion module, configured to extract control position information for characterizing the spatial contour of the mold based on the standard geometric model, and convert the control position information into a control instruction sequence matching the cavity drive structure; a control module, configured to input the control instruction sequence into a controller, and drive the mold cavity assembly to adjust its spatial structure so that the mold cavity matches the target three-dimensional model in contour during the molding stage; The conversion module is further configured to: The offset prediction model trained through the knowledge distillation mechanism is called. The offset prediction model takes the structural code, thermal field data, assembly parameters, and material parameters as inputs, and outputs a compensation vector corresponding to the control position information. The compensation vector is used to perform vector superposition with the control position information to generate a control instruction correction term. Among them, the offset prediction model is obtained by fitting the offset vector and intermediate behavior characteristics generated by a teacher model with thermal-stress coupling simulation capabilities during the training stage, and is called in real time in the molding control process for online compensation.

2. The mold foaming control system based on image import and model simulation according to claim 1 is characterized in that: The acquisition module is also used for: Acquire a structural data source, and identify mold boundary features, connection location information, and component identification data based on the structural data source to construct a structural recognition result; Based on the structure recognition result, topological relationship extraction is performed to generate a structure code for describing the hierarchical relationship of the mold structure; Wherein, the structural data source includes: geometric structure and connection information of mold components.

3. The mold foaming control system based on image import and model simulation according to claim 1 is characterized in that: The acquisition module is also used for: Obtain the basic physical property database records corresponding to the mold material, extract the material attribute fields, and construct the material property table; performing a standardized encoding operation on the material property table to generate material parameters for describing physical properties of the mold material; The material property fields include: thermal expansion coefficient, thermal conductivity, and elastic modulus.

4. The mold foaming control system based on image import and model simulation according to claim 1, characterized in that: The acquisition module is also used for: Acquire a data stream from thermal sensors placed in a target area of ​​the mold, and reconstruct a time series of real-time temperature values ​​of each node in the data stream to generate a multi-node temperature time series graph; Spatial interpolation and partition classification are performed on the temperature time series diagram to generate thermal field data for reflecting the thermal distribution state of the mold.

5. The mold foaming control system based on image import and model simulation according to claim 1, characterized in that: The acquisition module is also used for: Obtaining an assembly sequence log and component pairing parameters from a mold assembly process control system, parsing component dependencies based on the assembly sequence log, parsing boundary connection modes based on the component pairing parameters, and constructing an assembly constraint matrix; A graph structure encoding operation is performed on the assembly constraint matrix to generate assembly parameters for describing assembly path information.

6. The mold foaming control system based on image import and model simulation according to claim 1, characterized in that: The teacher model is used to: Receive the structural code, extract the topological structure relationship, boundary connection position and node spatial layout between the mold parts, and generate the structural feature tensor; Receive thermal field data, construct a spatial temperature matrix, and perform temperature-structure mapping in combination with the structural feature tensor to generate a temperature tensor sequence; Receive assembly parameters, analyze component assembly sequence and connection constraints, and generate assembly drawing structure; receiving material parameters, extracting the thermal expansion coefficient, thermal conductivity, and elastic modulus of each region of the mold, and generating a material property vector; Based on the structural feature tensor, temperature tensor sequence, assembly drawing structure and material property vector, a thermal-stress coupling simulation is performed in the teacher model to calculate the three-dimensional spatial offset of the target control point under the set thermal working condition and generate a training offset vector; During the simulation process, intermediate stress response characteristics, thermal distribution change paths, and local structural response indicators are further extracted to form intermediate behavior characteristics. The training offset vector and the intermediate behavior characteristics are used together as supervisory information for training the offset prediction model.

7. The mold foaming control system based on image import and model simulation according to claim 6, characterized in that: The knowledge distillation mechanism includes: Using the training offset vector as a first supervision target, calculating a first supervision loss value between the output result of the student model and the training offset vector; Taking the intermediate behavior features as the second supervision target, perform feature alignment on the feature expression of the intermediate layer of the student model and calculate the second supervision loss value; Performing a weighted fusion of the first supervised loss value and the second supervised loss value to generate a total loss function as an optimization target for parameter updating in the training phase; During the training process, multiple sets of training samples are constructed based on the structural encoding, thermal field data, assembly parameters, and material parameters, and the model parameter update operation is repeatedly performed to obtain the student model; During the molding control process operation phase, the current structural code, thermal field data, assembly parameters, and material parameters are input into the student model, and a compensation vector corresponding to the current control position information is calculated and output. The compensation vector and the control position information are vector-superimposed to generate the control instruction sequence for the control module to call.

8. The mold foaming control system based on image import and model simulation according to claim 7, characterized in that: Extracting control position information for characterizing the mold space contour based on the standard geometric model and converting the control position information into a control instruction sequence matching the cavity drive structure includes: Performing boundary curvature extraction on the standard geometric model, combining the local structural response index in the intermediate behavior feature to identify boundary abnormal areas with high curvature change, sparse point cloud or connection jump features; Constructing a boundary buffer mask for the abnormal boundary area, and performing trajectory interpolation and weighted smoothing on the control position information within the buffer mask to generate a continuous control point sequence; Perform vector superposition on the continuous control point sequence and the compensation vector in the order of control point indexes to generate a compensation control path sequence; Performing trajectory segment analysis on the compensation control path sequence, and performing continuity check on the angle mutations and path jump amplitudes between adjacent control segments to construct a smooth transition segment; The smooth transition section is merged with the compensation control path sequence to generate a complete control path, and according to the degree of freedom configuration of the cavity drive structure, dimension projection and instruction format conversion are performed on the complete control path to generate a control instruction sequence.

9. The mold foaming control system based on image import and model simulation according to claim 1, characterized in that: Performing morphological processing on the target three-dimensional model to generate a standard geometric model suitable for mold cavity control includes: Performing a boundary contour extraction operation on the target three-dimensional model, identifying a set of boundary points at the outer edge of the model, and constructing a boundary closure inspection area; Performing density consistency analysis on the boundary point set, identifying distribution anomaly areas in the boundary point set, and performing boundary point interpolation correction operations to generate a boundary continuity model; Performing a scale normalization operation on the boundary continuity model, constructing a scale conversion matrix based on the main scale vector of the model, and generating a unit scale model; Performing a curvature smoothing operation on the unit-scale model, performing facet reconstruction and mesh adjustment based on a local curvature threshold, and generating a target surface model with a continuous curvature distribution; The target surface model is fused with the boundary continuity model to generate a standard geometric model suitable for mold cavity control.

10. The mold foaming control system based on image import and model simulation according to claim 1, characterized in that: Inputting the control instruction sequence into a controller, and having the controller drive the mold cavity assembly to adjust its spatial structure so that the mold cavity matches the target three-dimensional model in contour during the molding stage includes: Performing degree of freedom decomposition processing on the control position information in the control instruction sequence, and mapping each control instruction to a corresponding driving submodule according to the structural layout of the mold cavity assembly; Performing instruction format conversion on the mapping control instruction generated by the driving submodule, converting the position information into a corresponding displacement target value, motor rotation angle value or hydraulic stroke value, and generating an execution instruction stream; The execution instruction stream is synchronously distributed according to a preset time step, and the instruction buffer is set according to the response parameters of each driver submodule to control the signal beat; During the deformation process of the cavity component, the execution status information fed back by the position encoder, strain sensor or displacement sensor is collected to generate feedback data; The difference between the feedback data and the target value in the execution instruction stream is calculated, and if the difference exceeds a tolerance threshold, a compensation correction instruction is output to the corresponding driving submodule.

Images